update

app.py

@@ -1,285 +1,147 @@
-from langchain.agents.initialize import initialize_agent
-from langchain.agents.tools import Tool
-from langchain.chains.conversation.memory import ConversationBufferMemory
-from langchain.llms.openai import OpenAI
-from audio_foundation_models import *
+import torch
+import numpy as np
 import gradio as gr
-
-_DESCRIPTION = '# [AudioGPT](https://github.com/AIGC-Audio/AudioGPT)'
-_DESCRIPTION += '\n<p>This is a demo to the work <a href="https://github.com/AIGC-Audio/AudioGPT" style="text-decoration: underline;" target="_blank">AudioGPT: Understanding and Generating Speech, Music, Sound, and Talking Head</a>. </p>'
-_DESCRIPTION += '\n<p>This model can only be used for non-commercial purposes.'
-if (SPACE_ID := os.getenv('SPACE_ID')) is not None:
-    _DESCRIPTION += f'\n<p>For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings. <a href="https://huggingface.co/spaces/{SPACE_ID}?duplicate=true"><img style="display: inline; margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space" /></a></p>'
-
-
-AUDIO_CHATGPT_PREFIX = """AudioGPT
-AudioGPT can not directly read audios, but it has a list of tools to finish different speech, audio, and singing voice tasks. Each audio will have a file name formed as "audio/xxx.wav". When talking about audios, AudioGPT is very strict to the file name and will never fabricate nonexistent files. 
-AudioGPT is able to use tools in a sequence, and is loyal to the tool observation outputs rather than faking the audio content and audio file name. It will remember to provide the file name from the last tool observation, if a new audio is generated.
-Human may provide new audios to AudioGPT with a description. The description helps AudioGPT to understand this audio, but AudioGPT should use tools to finish following tasks, rather than directly imagine from the description.
-Overall, AudioGPT is a powerful audio dialogue assistant tool that can help with a wide range of tasks and provide valuable insights and information on a wide range of topics. 
-TOOLS:
-------
-AudioGPT has access to the following tools:"""
-
-AUDIO_CHATGPT_FORMAT_INSTRUCTIONS = """To use a tool, please use the following format:
-```
-Thought: Do I need to use a tool? Yes
-Action: the action to take, should be one of [{tool_names}]
-Action Input: the input to the action
-Observation: the result of the action
-```
-When you have a response to say to the Human, or if you do not need to use a tool, you MUST use the format:
-```
-Thought: Do I need to use a tool? No
-{ai_prefix}: [your response here]
-```
-"""
-
-AUDIO_CHATGPT_SUFFIX = """You are very strict to the filename correctness and will never fake a file name if not exists.
-You will remember to provide the audio file name loyally if it's provided in the last tool observation.
-Begin!
-Previous conversation history:
-{chat_history}
-New input: {input}
-Thought: Do I need to use a tool? {agent_scratchpad}"""
-
-def cut_dialogue_history(history_memory, keep_last_n_words = 500):
-    tokens = history_memory.split()
-    n_tokens = len(tokens)
-    print(f"history_memory:{history_memory}, n_tokens: {n_tokens}")
-    if n_tokens < keep_last_n_words:
-        return history_memory
-    else:
-        paragraphs = history_memory.split('\n')
-        last_n_tokens = n_tokens
-        while last_n_tokens >= keep_last_n_words:
-            last_n_tokens = last_n_tokens - len(paragraphs[0].split(' '))
-            paragraphs = paragraphs[1:]
-        return '\n' + '\n'.join(paragraphs)
-
-class ConversationBot:
-    def __init__(self, load_dict):
-        print("Initializing AudioGPT")
-        self.tools = []
-        self.memory = ConversationBufferMemory(memory_key="chat_history", output_key='output')
-        self.models = dict()
-        for class_name, device in load_dict.items():
-            self.models[class_name] = globals()[class_name](device=device)
-        for class_name, instance in self.models.items():
-            for e in dir(instance):
-                if e.startswith('inference'):
-                    func = getattr(instance, e)
-                    self.tools.append(Tool(name=func.name, description=func.description, func=func))
-
-    def run_text(self, text, state):
-        print("===============Running run_text =============")
-        print("Inputs:", text, state)
-        print("======>Previous memory:\n %s" % self.agent.memory)
-        self.agent.memory.buffer = cut_dialogue_history(self.agent.memory.buffer, keep_last_n_words=500)
-        res = self.agent({"input": text})
-        if res['intermediate_steps'] == []:
-            print("======>Current memory:\n %s" % self.agent.memory)
-            response = res['output']
-            state = state + [(text, response)]
-            print("Outputs:", state)
-            return state, state, gr.Audio.update(visible=False), gr.Image.update(visible=False), gr.Button.update(visible=False)
-        else:
-            tool = res['intermediate_steps'][0][0].tool
-            if tool == "Generate Image From User Input Text":
-                res['output'] = res['output'].replace("\\", "/")
-                response = re.sub('(image/\S*png)', lambda m: f'![](/file={m.group(0)})*{m.group(0)}*', res['output'])
-                state = state + [(text, response)]
-                print(f"\nProcessed run_text, Input text: {text}\nCurrent state: {state}\n"
-                      f"Current Memory: {self.agent.memory.buffer}")
-                return state, state, gr.Audio.update(visible=False), gr.Image.update(visible=False), gr.Button.update(visible=False)
-            elif tool == "Detect The Sound Event From The Audio":
-                image_filename = res['intermediate_steps'][0][1]
-                response = res['output'] + f"![](/file={image_filename})*{image_filename}*"
-                state = state + [(text, response)]
-                print(f"\nProcessed run_text, Input text: {text}\nCurrent state: {state}\n"
-                      f"Current Memory: {self.agent.memory.buffer}")
-                return state, state, gr.Audio.update(visible=False), gr.Image.update(visible=False), gr.Button.update(visible=False)        
-            elif tool == "Generate Text From The Audio" or tool == "Transcribe speech" or tool == "Target Sound Detection":
-                print("======>Current memory:\n %s" % self.agent.memory)
-                response = re.sub('(image/\S*png)', lambda m: f'![](/file={m.group(0)})*{m.group(0)}*', res['output'])
-                image_filename = res['intermediate_steps'][0][1]
-                #response = res['output'] + f"![](/file={image_filename})*{image_filename}*"
-                state = state + [(text, response)]
-                print("Outputs:", state)
-                return state, state, gr.Audio.update(visible=False), gr.Image.update(visible=False), gr.Button.update(visible=False)
-            elif tool == "Audio Inpainting":
-                audio_filename = res['intermediate_steps'][0][0].tool_input
-                image_filename = res['intermediate_steps'][0][1]
-                print("======>Current memory:\n %s" % self.agent.memory)
-                print(res)
-                response = res['output']
-                state = state + [(text, response)]
-                print("Outputs:", state)
-                return state, state, gr.Audio.update(value=audio_filename,visible=True), gr.Image.update(value=image_filename,visible=True), gr.Button.update(visible=True)
-            print("======>Current memory:\n %s" % self.agent.memory)
-            response = re.sub('(image/\S*png)', lambda m: f'![](/file={m.group(0)})*{m.group(0)}*', res['output'])
-            audio_filename = res['intermediate_steps'][0][1]
-            state = state + [(text, response)]
-            print("Outputs:", state)
-            return state, state, gr.Audio.update(value=audio_filename,visible=True), gr.Image.update(visible=False), gr.Button.update(visible=False)
-
-    def run_image_or_audio(self, file, state, txt):
-        file_type = file.name[-3:]
-        if file_type == "wav":
-            print("===============Running run_audio =============")
-            print("Inputs:", file, state)
-            print("======>Previous memory:\n %s" % self.agent.memory)
-            audio_filename = os.path.join('audio', str(uuid.uuid4())[0:8] + ".wav")
-            audio_load = whisper.load_audio(file.name)
-            soundfile.write(audio_filename, audio_load, samplerate = 16000)
-            description = self.models['A2T'].inference(audio_filename)
-            Human_prompt = "\nHuman: provide an audio named {}. The description is: {}. This information helps you to understand this audio, but you should use tools to finish following tasks, " \
-                           "rather than directly imagine from my description. If you understand, say \"Received\". \n".format(audio_filename, description)
-            AI_prompt = "Received.  "
-            self.agent.memory.buffer = self.agent.memory.buffer + Human_prompt + 'AI: ' + AI_prompt
-            # AI_prompt = "Received.  "
-            # self.agent.memory.buffer = self.agent.memory.buffer + 'AI: ' + AI_prompt
-            print("======>Current memory:\n %s" % self.agent.memory)
-            #state = state + [(f"<audio src=audio_filename controls=controls></audio>*{audio_filename}*", AI_prompt)]
-            state = state + [(f"*{audio_filename}*", AI_prompt)]
-            print("Outputs:", state)
-            return state, state, txt + ' ' + audio_filename + ' ', gr.Audio.update(value=audio_filename,visible=True)
-        else:
-            # print("===============Running run_image =============")
-            # print("Inputs:", file, state)
-            # print("======>Previous memory:\n %s" % self.agent.memory)
-            image_filename = os.path.join('image', str(uuid.uuid4())[0:8] + ".png")
-            print("======>Auto Resize Image...")
-            img = Image.open(file.name)
-            width, height = img.size
-            ratio = min(512 / width, 512 / height)
-            width_new, height_new = (round(width * ratio), round(height * ratio))
-            width_new = int(np.round(width_new / 64.0)) * 64
-            height_new = int(np.round(height_new / 64.0)) * 64
-            img = img.resize((width_new, height_new))
-            img = img.convert('RGB')
-            img.save(image_filename, "PNG")
-            print(f"Resize image form {width}x{height} to {width_new}x{height_new}")
-            description = self.models['ImageCaptioning'].inference(image_filename)
-            Human_prompt = "\nHuman: provide an audio named {}. The description is: {}. This information helps you to understand this audio, but you should use tools to finish following tasks, " \
-                           "rather than directly imagine from my description. If you understand, say \"Received\". \n".format(image_filename, description)
-            AI_prompt = "Received.  "
-            self.agent.memory.buffer = self.agent.memory.buffer + Human_prompt + 'AI: ' + AI_prompt
-            print("======>Current memory:\n %s" % self.agent.memory)
-            state = state + [(f"![](/file={image_filename})*{image_filename}*", AI_prompt)]
-            print(f"\nProcessed run_image, Input image: {image_filename}\nCurrent state: {state}\n"
-                  f"Current Memory: {self.agent.memory.buffer}")
-            return state, state, txt + f'{txt} {image_filename} ', gr.Audio.update(visible=False)
-
-    def inpainting(self, state, audio_filename, image_filename):
-        print("===============Running inpainting =============")
-        print("Inputs:", state)
-        print("======>Previous memory:\n %s" % self.agent.memory)
-        # inpaint = Inpaint(device="cpu")
-        new_image_filename, new_audio_filename = self.models['Inpaint'].predict(audio_filename, image_filename)
-        AI_prompt = "Here are the predict audio and the mel spectrum." + f"*{new_audio_filename}*" + f"![](/file={new_image_filename})*{new_image_filename}*"
-        self.agent.memory.buffer = self.agent.memory.buffer + 'AI: ' + AI_prompt
-        print("======>Current memory:\n %s" % self.agent.memory)
-        state = state + [(f"Audio Inpainting", AI_prompt)]
-        print("Outputs:", state)
-        return state, state, gr.Image.update(visible=False), gr.Audio.update(value=new_audio_filename, visible=True), gr.Button.update(visible=False)
-    def clear_audio(self):
-        return gr.Audio.update(value=None, visible=False)
-    def clear_image(self):
-        return gr.Image.update(value=None, visible=False)
-    def clear_button(self):
-        return gr.Button.update(visible=False)
-    def init_agent(self, openai_api_key):
-        os.system('nvidia-smi')
-        self.llm = OpenAI(temperature=0, openai_api_key=openai_api_key)
-        self.agent = initialize_agent(
-            self.tools,
-            self.llm,
-            agent="conversational-react-description",
-            verbose=True,
-            memory=self.memory,
-            return_intermediate_steps=True,
-            agent_kwargs={'prefix': AUDIO_CHATGPT_PREFIX, 'format_instructions': AUDIO_CHATGPT_FORMAT_INSTRUCTIONS, 'suffix': AUDIO_CHATGPT_SUFFIX}, )
-        return gr.update(visible = True)
-
-
-
-if __name__ == '__main__':
-    bot = ConversationBot({#'ImageCaptioning': 'cuda:0',
-                           'T2A': 'cuda:0',
-                           #'I2A': 'cuda:0'
-                           #'TTS_OOD':'cuda:0'
-                           'TTS': 'cuda:0',
-                           #'T2S': 'cuda:0'
-                           'ASR': 'cuda:0',
-                           'A2T': 'cuda:0',
-                           #'Inpaint': 'cuda:0',
-                           #'SoundDetection': 'cuda:0'
-                           'Binaural': 'cuda:0'
-                           #'SoundExtraction': 'cuda:0'
-                           #'TargetSoundDetection': 'cuda:0',
-                           #'Speech_Enh_SC': 'cuda:0'
-                           #'Speech_SS': 'cuda:0'
-                           })
-    with gr.Blocks(css="#chatbot {overflow:auto; height:500px;}") as demo:
-        gr.Markdown(_DESCRIPTION)
-
-        with gr.Row():
-            openai_api_key_textbox = gr.Textbox(
-                placeholder="Paste your OpenAI API key here to start Audio ChatGPT(sk-...) and press Enter ↵️",
-                show_label=False,
-                lines=1,
-                type="password",
-            )
-
-        chatbot = gr.Chatbot(elem_id="chatbot", label="Audio ChatGPT")
-        state = gr.State([])
-        with gr.Row(visible = False) as input_raws:
-            with gr.Column(scale=0.7):
-                txt = gr.Textbox(show_label=False, placeholder="Enter text and press enter, or upload an image").style(container=False)
-            with gr.Column(scale=0.1, min_width=0):
-                run = gr.Button("🏃‍♂️Run")
-            with gr.Column(scale=0.1, min_width=0):
-                clear = gr.Button("🔄Clear️")
-            with gr.Column(scale=0.1, min_width=0):
-                btn = gr.UploadButton("🖼️/🎙️ Upload", file_types=["image","audio"])
-        with gr.Row():        
-            with gr.Column():
-                outaudio = gr.Audio(visible=False)
-        with gr.Row():           
-            with gr.Column():
-                show_mel = gr.Image(type="filepath",tool='sketch',visible=False)
-        with gr.Row():           
-            with gr.Column():        
-                run_button = gr.Button("Predict Masked Place",visible=False)
-        gr.Examples(
-            examples=["Generate a speech with text 'here we go'",
-                      "Transcribe this speech",
-                      "Transfer the mono speech to a binaural one",
-                      "Generate an audio of a dog barking",
-                      "Generate an audio of this uploaded image",
-                      "Give me the description of this audio",
-                      "I want to inpaint it",
-                      "What events does this audio include?",
-                      "When did the thunder happen in this audio?",
-                      "Extract the thunder event from this audio",
-                      "Generate a piece of singing voice. Text sequence is 小酒窝长睫毛AP是你最美的记号. Note sequence is C#4/Db4 | F#4/Gb4 | G#4/Ab4 | A#4/Bb4 F#4/Gb4 | F#4/Gb4 C#4/Db4 | C#4/Db4 | rest | C#4/Db4 | A#4/Bb4 | G#4/Ab4 | A#4/Bb4 | G#4/Ab4 | F4 | C#4/Db4. Note duration sequence is 0.407140 | 0.376190 | 0.242180 | 0.509550 0.183420 | 0.315400 0.235020 | 0.361660 | 0.223070 | 0.377270 | 0.340550 | 0.299620 | 0.344510 | 0.283770 | 0.323390 | 0.360340.",
-                      ],
-            inputs=txt
+from PIL import Image
+from omegaconf import OmegaConf
+from pathlib import Path
+from vocoder.bigvgan.models import VocoderBigVGAN
+from ldm.models.diffusion.ddim import DDIMSampler
+from ldm.util import instantiate_from_config
+from wav_evaluation.models.CLAPWrapper import CLAPWrapper
+
+SAMPLE_RATE = 16000
+
+torch.set_grad_enabled(False)
+device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+
+def dur_to_size(duration):
+    latent_width = int(duration * 7.8)
+    if latent_width % 4 != 0:
+        latent_width = (latent_width // 4 + 1) * 4
+    return latent_width
+
+def initialize_model(config, ckpt):
+    config = OmegaConf.load(config)
+    model = instantiate_from_config(config.model)
+    model.load_state_dict(torch.load(ckpt,map_location='cpu')["state_dict"], strict=False)
+
+    model = model.to(device)
+    model.cond_stage_model.to(model.device)
+    model.cond_stage_model.device = model.device
+    print(model.device,device,model.cond_stage_model.device)
+    sampler = DDIMSampler(model)
+
+    return sampler
+
+sampler = initialize_model('configs/text_to_audio/txt2audio_args.yaml', 'useful_ckpts/maa1_caps.ckpt')
+vocoder = VocoderBigVGAN('vocoder/logs/bigvnat',device=device)
+clap_model = CLAPWrapper('useful_ckpts/CLAP/CLAP_weights_2022.pth','useful_ckpts/CLAP/config.yml',use_cuda=torch.cuda.is_available())
+
+def select_best_audio(prompt,wav_list):
+    text_embeddings = clap_model.get_text_embeddings([prompt])
+    score_list = []
+    for data in wav_list:
+        sr,wav = data
+        audio_embeddings = clap_model.get_audio_embeddings([(torch.FloatTensor(wav),sr)], resample=True)
+        score = clap_model.compute_similarity(audio_embeddings, text_embeddings,use_logit_scale=False).squeeze().cpu().numpy()
+        score_list.append(score)
+    max_index = np.array(score_list).argmax()
+    print(score_list,max_index)
+    return wav_list[max_index]
+
+def txt2audio(sampler,vocoder,prompt, seed, scale, ddim_steps, n_samples=1, W=624, H=80):
+    prng = np.random.RandomState(seed)
+    start_code = prng.randn(n_samples, sampler.model.first_stage_model.embed_dim, H // 8, W // 8)
+    start_code = torch.from_numpy(start_code).to(device=device, dtype=torch.float32)
+    
+    uc = None
+    if scale != 1.0:
+        uc = sampler.model.get_learned_conditioning(n_samples * [""])
+    c = sampler.model.get_learned_conditioning(n_samples * [prompt])# shape:[1,77,1280],即还没有变成句子embedding，仍是每个单词的embedding
+    shape = [sampler.model.first_stage_model.embed_dim, H//8, W//8]  # (z_dim, 80//2^x, 848//2^x)
+    samples_ddim, _ = sampler.sample(S=ddim_steps,
+                                        conditioning=c,
+                                        batch_size=n_samples,
+                                        shape=shape,
+                                        verbose=False,
+                                        unconditional_guidance_scale=scale,
+                                        unconditional_conditioning=uc,
+                                        x_T=start_code)
+
+    x_samples_ddim = sampler.model.decode_first_stage(samples_ddim)
+
+    wav_list = []
+    for idx,spec in enumerate(x_samples_ddim):
+        wav = vocoder.vocode(spec)
+        wav_list.append((SAMPLE_RATE,wav))
+    best_wav = select_best_audio(prompt,wav_list)
+    return best_wav
+
+
+def predict(prompt, ddim_steps, num_samples, scale, seed):
+    melbins,mel_len = 80,624
+    with torch.no_grad():
+        result = txt2audio(
+            sampler=sampler,
+            vocoder=vocoder,
+            prompt=prompt,
+            seed=seed,
+            scale=scale,
+            ddim_steps=ddim_steps,
+            n_samples=num_samples,
+            H=melbins, W=mel_len
         )
 
-        openai_api_key_textbox.submit(bot.init_agent, [openai_api_key_textbox], [input_raws])    
-        txt.submit(bot.run_text, [txt, state], [chatbot, state, outaudio, show_mel, run_button])
-        txt.submit(lambda: "", None, txt)
-        run.click(bot.run_text, [txt, state], [chatbot, state, outaudio, show_mel, run_button])
-        run.click(lambda: "", None, txt)
-        btn.upload(bot.run_image_or_audio, [btn, state, txt], [chatbot, state, txt, outaudio])
-        run_button.click(bot.inpainting, [state, outaudio, show_mel], [chatbot, state, show_mel, outaudio, run_button])
-        clear.click(bot.memory.clear)
-        clear.click(lambda: [], None, chatbot)
-        clear.click(lambda: [], None, state)
-        clear.click(lambda:None, None, txt)
-        clear.click(bot.clear_button, None, run_button)
-        clear.click(bot.clear_image, None, show_mel)
-        clear.click(bot.clear_audio, None, outaudio)
-        demo.launch(server_name="0.0.0.0", server_port=7860)
+    return result
+
+
+with gr.Blocks() as demo:
+    with gr.Row():
+        gr.Markdown("## Make-An-Audio: Text-to-Audio Generation")
+
+    with gr.Row():
+        with gr.Column():
+            prompt = gr.Textbox(label="Prompt: Input your text here.        ")
+            run_button = gr.Button(label="Run")
+
+            
+            with gr.Accordion("Advanced options", open=False):
+                num_samples = gr.Slider(
+                    label="Select from audios num.This number control the number of candidates \
+                        (e.g., generate three audios and choose the best to show you). A Larger value usually lead to \
+                        better quality with heavier computation", minimum=1, maximum=10, value=3, step=1)
+                # num_samples = 1
+                ddim_steps = gr.Slider(label="Steps", minimum=1,
+                                       maximum=150, value=100, step=1)
+                scale = gr.Slider(
+                    label="Guidance Scale:(Large => more relevant to text but the quality may drop)", minimum=0.1, maximum=4.0, value=1.5, step=0.1
+                )
+                seed = gr.Slider(
+                    label="Seed:Change this value (any integer number) will lead to a different generation result.",
+                    minimum=0,
+                    maximum=2147483647,
+                    step=1,
+                    value=44,
+                )
+
+        with gr.Column():
+            # audio_list = []
+            # for i in range(int(num_samples)):
+            #     audio_list.append(gr.outputs.Audio())
+            outaudio = gr.Audio()
+
+
+    run_button.click(fn=predict, inputs=[
+                    prompt,ddim_steps, num_samples, scale, seed], outputs=[outaudio])# inputs的参数只能传gr.xxx
+    with gr.Row():
+        with gr.Column():
+            gr.Examples(
+                        examples = [['a dog barking and a bird chirping',100,3,2,55],['fireworks pop and explode',100,3,2,55],
+                                        ['piano and violin plays',100,3,2,55],['wind thunder and rain falling',100,3,2,55],['music made by drum kit',100,3,2,55]],
+                        inputs = [prompt,ddim_steps, num_samples, scale, seed],
+                        outputs = [outaudio]
+                        )
+        with gr.Column():
+            pass
+
+demo.launch()

configs/text_to_audio/bigvgan_args.yaml

@@ -0,0 +1,63 @@
+resblock: '1'
+num_gpus: 0
+batch_size: 64
+num_mels: 80
+learning_rate: 0.0001
+adam_b1: 0.8
+adam_b2: 0.99
+lr_decay: 0.999
+seed: 1234
+upsample_rates:
+- 4
+- 4
+- 2
+- 2
+- 2
+- 2
+upsample_kernel_sizes:
+- 8
+- 8
+- 4
+- 4
+- 4
+- 4
+upsample_initial_channel: 1536
+resblock_kernel_sizes:
+- 3
+- 7
+- 11
+resblock_dilation_sizes:
+- - 1
+  - 3
+  - 5
+- - 1
+  - 3
+  - 5
+- - 1
+  - 3
+  - 5
+activation: snakebeta
+snake_logscale: true
+resolutions:
+- - 1024
+  - 120
+  - 600
+- - 2048
+  - 240
+  - 1200
+- - 512
+  - 50
+  - 240
+mpd_reshapes:
+- 2
+- 3
+- 5
+- 7
+- 11
+use_spectral_norm: false
+discriminator_channel_mult: 1
+num_workers: 4
+dist_config:
+  dist_backend: nccl
+  dist_url: tcp://localhost:54341
+  world_size: 1

configs/text_to_audio/clap_args.yaml

@@ -0,0 +1,26 @@
+# TEXT ENCODER CONFIG
+text_model: 'bert-base-uncased'
+text_len: 100
+transformer_embed_dim: 768
+freeze_text_encoder_weights: True
+
+# AUDIO ENCODER CONFIG
+audioenc_name: 'Cnn14'
+out_emb: 2048
+sampling_rate: 44100
+duration: 9
+fmin: 50
+fmax: 14000
+n_fft: 1028
+hop_size: 320
+mel_bins: 64
+window_size: 1024
+
+# PROJECTION SPACE CONFIG 
+d_proj: 1024
+temperature: 0.003
+
+# TRAINING AND EVALUATION CONFIG
+num_classes: 527
+batch_size: 1024
+demo: False

configs/text_to_audio/txt2audio_args.yaml

@@ -0,0 +1,78 @@
+model:
+  base_learning_rate: 1.0e-05
+  target: ldm.models.diffusion.ddpm_audio.LatentDiffusion_audio
+  params:
+    linear_start: 0.00085
+    linear_end: 0.0120
+    num_timesteps_cond: 1
+    log_every_t: 200
+    timesteps: 1000
+    first_stage_key: image
+    cond_stage_key: caption
+    image_size: 32      # unused
+    mel_dim: 10         # 80 // 2^3
+    mel_length: 78     # 624 // 2^3
+    channels: 4
+    cond_stage_trainable: false
+    conditioning_key: crossattn
+    monitor: val/loss_simple_ema
+    scale_by_std: True
+    use_ema: False
+
+    scheduler_config: # 10000 warmup steps
+      target: ldm.lr_scheduler.LambdaLinearScheduler
+      params:
+        warm_up_steps: [10000]
+        cycle_lengths: [10000000000000]
+        f_start: [1.e-6]
+        f_max: [1.]
+        f_min: [ 1.]
+
+    unet_config:
+      target: ldm.modules.diffusionmodules.openaimodel.UNetModel
+      params:
+        image_size: 32 # ununsed
+        in_channels: 4
+        out_channels: 4
+        model_channels: 320
+        attention_resolutions:
+        - 1
+        - 2
+        num_res_blocks: 2
+        channel_mult:  # num_down = len(ch_mult)-1
+        - 1
+        - 2
+        num_heads: 8
+        use_spatial_transformer: true
+        transformer_depth: 1
+        context_dim: 1024
+        use_checkpoint: true
+        legacy: False
+
+    first_stage_config:
+      target: ldm.models.autoencoder.AutoencoderKL
+      params:
+        embed_dim: 4
+        monitor: val/rec_loss
+        ckpt_path: 
+        ddconfig:
+          double_z: true
+          z_channels: 4
+          resolution: 624
+          in_channels: 1
+          out_ch: 1
+          ch: 128
+          ch_mult: [ 1, 2, 2, 4 ]  # num_down = len(ch_mult)-1
+          num_res_blocks: 2
+          attn_resolutions: [78, 156]
+          dropout: 0.0
+        lossconfig:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: ldm.modules.encoders.modules.FrozenCLAPEmbedder
+      params:
+        weights_path: useful_ckpts/CLAP/CLAP_weights_2022.pth
+
+ckpt_path: useful_ckpts/maa1_caps.ckpt
+

ldm/lr_scheduler.py

@@ -0,0 +1,98 @@
+import numpy as np
+
+
+class LambdaWarmUpCosineScheduler:
+    """
+    note: use with a base_lr of 1.0
+    """
+    def __init__(self, warm_up_steps, lr_min, lr_max, lr_start, max_decay_steps, verbosity_interval=0):
+        self.lr_warm_up_steps = warm_up_steps
+        self.lr_start = lr_start
+        self.lr_min = lr_min
+        self.lr_max = lr_max
+        self.lr_max_decay_steps = max_decay_steps
+        self.last_lr = 0.
+        self.verbosity_interval = verbosity_interval
+
+    def schedule(self, n, **kwargs):
+        if self.verbosity_interval > 0:
+            if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_lr}")
+        if n < self.lr_warm_up_steps:
+            lr = (self.lr_max - self.lr_start) / self.lr_warm_up_steps * n + self.lr_start
+            self.last_lr = lr
+            return lr
+        else:
+            t = (n - self.lr_warm_up_steps) / (self.lr_max_decay_steps - self.lr_warm_up_steps)
+            t = min(t, 1.0)
+            lr = self.lr_min + 0.5 * (self.lr_max - self.lr_min) * (
+                    1 + np.cos(t * np.pi))
+            self.last_lr = lr
+            return lr
+
+    def __call__(self, n, **kwargs):
+        return self.schedule(n,**kwargs)
+
+
+class LambdaWarmUpCosineScheduler2:
+    """
+    supports repeated iterations, configurable via lists
+    note: use with a base_lr of 1.0.
+    """
+    def __init__(self, warm_up_steps, f_min, f_max, f_start, cycle_lengths, verbosity_interval=0):
+        assert len(warm_up_steps) == len(f_min) == len(f_max) == len(f_start) == len(cycle_lengths)
+        self.lr_warm_up_steps = warm_up_steps
+        self.f_start = f_start
+        self.f_min = f_min
+        self.f_max = f_max
+        self.cycle_lengths = cycle_lengths
+        self.cum_cycles = np.cumsum([0] + list(self.cycle_lengths))
+        self.last_f = 0.
+        self.verbosity_interval = verbosity_interval
+
+    def find_in_interval(self, n):
+        interval = 0
+        for cl in self.cum_cycles[1:]:
+            if n <= cl:
+                return interval
+            interval += 1
+
+    def schedule(self, n, **kwargs):
+        cycle = self.find_in_interval(n)
+        n = n - self.cum_cycles[cycle]
+        if self.verbosity_interval > 0:
+            if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
+                                                       f"current cycle {cycle}")
+        if n < self.lr_warm_up_steps[cycle]:
+            f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
+            self.last_f = f
+            return f
+        else:
+            t = (n - self.lr_warm_up_steps[cycle]) / (self.cycle_lengths[cycle] - self.lr_warm_up_steps[cycle])
+            t = min(t, 1.0)
+            f = self.f_min[cycle] + 0.5 * (self.f_max[cycle] - self.f_min[cycle]) * (
+                    1 + np.cos(t * np.pi))
+            self.last_f = f
+            return f
+
+    def __call__(self, n, **kwargs):
+        return self.schedule(n, **kwargs)
+
+
+class LambdaLinearScheduler(LambdaWarmUpCosineScheduler2):
+
+    def schedule(self, n, **kwargs):
+        cycle = self.find_in_interval(n)
+        n = n - self.cum_cycles[cycle]
+        if self.verbosity_interval > 0:
+            if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, "
+                                                       f"current cycle {cycle}")
+
+        if n < self.lr_warm_up_steps[cycle]:
+            f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle]
+            self.last_f = f
+            return f
+        else:
+            f = self.f_min[cycle] + (self.f_max[cycle] - self.f_min[cycle]) * (self.cycle_lengths[cycle] - n) / (self.cycle_lengths[cycle])
+            self.last_f = f
+            return f
+

ldm/models/autoencoder.py

@@ -0,0 +1,474 @@
+import os
+import torch
+import pytorch_lightning as pl
+import torch.nn.functional as F
+from contextlib import contextmanager
+from packaging import version
+import numpy as np
+from ldm.modules.diffusionmodules.model import Encoder, Decoder
+from ldm.modules.distributions.distributions import DiagonalGaussianDistribution
+from torch.optim.lr_scheduler import LambdaLR
+from ldm.util import instantiate_from_config
+# from icecream import ic
+
+class VQModel(pl.LightningModule):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 n_embed,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 batch_resize_range=None,
+                 scheduler_config=None,
+                 lr_g_factor=1.0,
+                 remap=None,
+                 sane_index_shape=False, # tell vector quantizer to return indices as bhw
+                 use_ema=False
+                 ):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.n_embed = n_embed
+        self.image_key = image_key
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        self.loss = instantiate_from_config(lossconfig)
+        self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25,
+                                        remap=remap,
+                                        sane_index_shape=sane_index_shape)
+        self.quant_conv = torch.nn.Conv2d(ddconfig["z_channels"], embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        if colorize_nlabels is not None:
+            assert type(colorize_nlabels)==int
+            self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
+        if monitor is not None:
+            self.monitor = monitor
+        self.batch_resize_range = batch_resize_range
+        if self.batch_resize_range is not None:
+            print(f"{self.__class__.__name__}: Using per-batch resizing in range {batch_resize_range}.")
+
+        self.use_ema = use_ema
+        if self.use_ema:
+            self.model_ema = LitEma(self)
+            print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+        self.scheduler_config = scheduler_config
+        self.lr_g_factor = lr_g_factor
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.parameters())
+            self.model_ema.copy_to(self)
+            if context is not None:
+                print(f"{context}: Switched to EMA weights")
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.parameters())
+                if context is not None:
+                    print(f"{context}: Restored training weights")
+
+    def init_from_ckpt(self, path, ignore_keys=list()):
+        sd = torch.load(path, map_location="cpu")["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        missing, unexpected = self.load_state_dict(sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+            print(f"Unexpected Keys: {unexpected}")
+
+    def on_train_batch_end(self, *args, **kwargs):
+        if self.use_ema:
+            self.model_ema(self)
+
+    def encode(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        quant, emb_loss, info = self.quantize(h)
+        return quant, emb_loss, info
+
+    def encode_to_prequant(self, x):
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        return h
+
+    def decode(self, quant):
+        quant = self.post_quant_conv(quant)
+        dec = self.decoder(quant)
+        return dec
+
+    def decode_code(self, code_b):
+        quant_b = self.quantize.embed_code(code_b)
+        dec = self.decode(quant_b)
+        return dec
+
+    def forward(self, input, return_pred_indices=False):
+        quant, diff, (_,_,ind) = self.encode(input)
+        dec = self.decode(quant)
+        if return_pred_indices:
+            return dec, diff, ind
+        return dec, diff
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
+        if self.batch_resize_range is not None:
+            lower_size = self.batch_resize_range[0]
+            upper_size = self.batch_resize_range[1]
+            if self.global_step <= 4:
+                # do the first few batches with max size to avoid later oom
+                new_resize = upper_size
+            else:
+                new_resize = np.random.choice(np.arange(lower_size, upper_size+16, 16))
+            if new_resize != x.shape[2]:
+                x = F.interpolate(x, size=new_resize, mode="bicubic")
+            x = x.detach()
+        return x
+
+    def training_step(self, batch, batch_idx, optimizer_idx):
+        # https://github.com/pytorch/pytorch/issues/37142
+        # try not to fool the heuristics
+        x = self.get_input(batch, self.image_key)
+        xrec, qloss, ind = self(x, return_pred_indices=True)
+
+        if optimizer_idx == 0:
+            # autoencode
+            aeloss, log_dict_ae = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train",
+                                            predicted_indices=ind)
+
+            self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True)
+            return aeloss
+
+        if optimizer_idx == 1:
+            # discriminator
+            discloss, log_dict_disc = self.loss(qloss, x, xrec, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train")
+            self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=True)
+            return discloss
+
+    def validation_step(self, batch, batch_idx):
+        log_dict = self._validation_step(batch, batch_idx)
+        with self.ema_scope():
+            log_dict_ema = self._validation_step(batch, batch_idx, suffix="_ema")
+        return log_dict
+
+    def _validation_step(self, batch, batch_idx, suffix=""):
+        x = self.get_input(batch, self.image_key)
+        xrec, qloss, ind = self(x, return_pred_indices=True)
+        aeloss, log_dict_ae = self.loss(qloss, x, xrec, 0,
+                                        self.global_step,
+                                        last_layer=self.get_last_layer(),
+                                        split="val"+suffix,
+                                        predicted_indices=ind
+                                        )
+
+        discloss, log_dict_disc = self.loss(qloss, x, xrec, 1,
+                                            self.global_step,
+                                            last_layer=self.get_last_layer(),
+                                            split="val"+suffix,
+                                            predicted_indices=ind
+                                            )
+        rec_loss = log_dict_ae[f"val{suffix}/rec_loss"]
+        self.log(f"val{suffix}/rec_loss", rec_loss,
+                   prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
+        self.log(f"val{suffix}/aeloss", aeloss,
+                   prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True)
+        if version.parse(pl.__version__) >= version.parse('1.4.0'):
+            del log_dict_ae[f"val{suffix}/rec_loss"]
+        self.log_dict(log_dict_ae)
+        self.log_dict(log_dict_disc)
+        return self.log_dict
+
+    def test_step(self, batch, batch_idx):
+        x = self.get_input(batch, self.image_key)
+        xrec, qloss, ind = self(x, return_pred_indices=True)
+        reconstructions = (xrec + 1)/2 # to mel scale
+        test_ckpt_path = os.path.basename(self.trainer.tested_ckpt_path)
+        savedir = os.path.join(self.trainer.log_dir,f'output_imgs_{test_ckpt_path}','fake_class')
+        if not os.path.exists(savedir):
+            os.makedirs(savedir)
+
+        file_names = batch['f_name']
+        # print(f"reconstructions.shape:{reconstructions.shape}",file_names)
+        reconstructions = reconstructions.cpu().numpy().squeeze(1) # squuze channel dim
+        for b in range(reconstructions.shape[0]):
+            vname_num_split_index = file_names[b].rfind('_')# file_names[b]:video_name+'_'+num
+            v_n,num = file_names[b][:vname_num_split_index],file_names[b][vname_num_split_index+1:]
+            save_img_path = os.path.join(savedir,f'{v_n}_sample_{num}.npy')
+            np.save(save_img_path,reconstructions[b])
+        
+        return None
+
+    def configure_optimizers(self):
+        lr_d = self.learning_rate
+        lr_g = self.lr_g_factor*self.learning_rate
+        print("lr_d", lr_d)
+        print("lr_g", lr_g)
+        opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
+                                  list(self.decoder.parameters())+
+                                  list(self.quantize.parameters())+
+                                  list(self.quant_conv.parameters())+
+                                  list(self.post_quant_conv.parameters()),
+                                  lr=lr_g, betas=(0.5, 0.9))
+        opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
+                                    lr=lr_d, betas=(0.5, 0.9))
+
+        if self.scheduler_config is not None:
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(opt_ae, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                },
+                {
+                    'scheduler': LambdaLR(opt_disc, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                },
+            ]
+            return [opt_ae, opt_disc], scheduler
+        return [opt_ae, opt_disc], []
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    def log_images(self, batch, only_inputs=False, plot_ema=False, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.image_key)
+        x = x.to(self.device)
+        if only_inputs:
+            log["inputs"] = x
+            return log
+        xrec, _ = self(x)
+        if x.shape[1] > 3:
+            # colorize with random projection
+            assert xrec.shape[1] > 3
+            x = self.to_rgb(x)
+            xrec = self.to_rgb(xrec)
+        log["inputs"] = x
+        log["reconstructions"] = xrec
+        if plot_ema:
+            with self.ema_scope():
+                xrec_ema, _ = self(x)
+                if x.shape[1] > 3: xrec_ema = self.to_rgb(xrec_ema)
+                log["reconstructions_ema"] = xrec_ema
+        return log
+
+    def to_rgb(self, x):
+        assert self.image_key == "segmentation"
+        if not hasattr(self, "colorize"):
+            self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
+        x = F.conv2d(x, weight=self.colorize)
+        x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
+        return x
+
+
+class VQModelInterface(VQModel):
+    def __init__(self, embed_dim, *args, **kwargs):
+        super().__init__(embed_dim=embed_dim, *args, **kwargs)
+        self.embed_dim = embed_dim
+
+    def encode(self, x):# VQModel的quantize写在encoder里,VQModelInterface则将其写在decoder里
+        h = self.encoder(x)
+        h = self.quant_conv(h)
+        return h
+
+    def decode(self, h, force_not_quantize=False):
+        # also go through quantization layer
+        if not force_not_quantize:
+            quant, emb_loss, info = self.quantize(h)
+        else:
+            quant = h
+        quant = self.post_quant_conv(quant)
+        dec = self.decoder(quant)
+        return dec
+
+
+class AutoencoderKL(pl.LightningModule):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 ):
+        super().__init__()
+        self.image_key = image_key
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        self.loss = instantiate_from_config(lossconfig)
+        assert ddconfig["double_z"]
+        self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        self.embed_dim = embed_dim
+        if colorize_nlabels is not None:
+            assert type(colorize_nlabels)==int
+            self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
+        if monitor is not None:
+            self.monitor = monitor
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+        # self.automatic_optimization = False # hjw for debug
+
+    def init_from_ckpt(self, path, ignore_keys=list()):
+        sd = torch.load(path, map_location="cpu")["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        self.load_state_dict(sd, strict=False)
+        print(f"Restored from {path}")
+
+    def encode(self, x):
+        h = self.encoder(x)
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+        return posterior
+
+    def decode(self, z):
+        z = self.post_quant_conv(z)
+        dec = self.decoder(z)
+        return dec
+
+    def forward(self, input, sample_posterior=True):
+        posterior = self.encode(input)
+        if sample_posterior:
+            z = posterior.sample()
+        else:
+            z = posterior.mode()
+        dec = self.decode(z)
+        return dec, posterior
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
+        return x
+
+    def training_step(self, batch, batch_idx, optimizer_idx):
+        inputs = self.get_input(batch, self.image_key)
+        reconstructions, posterior = self(inputs)
+
+        if optimizer_idx == 0:
+            # train encoder+decoder+logvar
+            aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train")
+            self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return aeloss
+
+        if optimizer_idx == 1:
+            # train the discriminator
+            discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                                last_layer=self.get_last_layer(), split="train")
+
+            self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return discloss
+
+    def validation_step(self, batch, batch_idx):
+        # self.log_images(batch,only_inputs=False,save_dir='mel_result_ae13_26/fake_class')
+        return self.log_dict
+
+    def test_step(self, batch, batch_idx):
+        test_ckpt_path = os.path.basename(self.trainer.tested_ckpt_path)
+        savedir = os.path.join(self.trainer.log_dir,f'output_imgs_{test_ckpt_path}','fake_class')
+        os.makedirs(savedir,exist_ok=True)
+        inputs = self.get_input(batch, self.image_key)# inputs shape:(b,c,mel_len,T) or (b,c,h,w)
+        # ic(inputs.shape)
+        # inputs = inputs[...,:624]
+        # ic(inputs.shape)
+        xrec, posterior = self(inputs)# reconstructions:(b,c,mel_len,T) or (b,c,h,w)
+        file_names = batch['f_name']
+        # print(f"reconstructions.shape:{reconstructions.shape}",file_names)
+        for b in range(len(file_names)):
+            rcon = (xrec[b].squeeze().detach().cpu().numpy() + 1) / 2 # to mel scale,squeeze channel dim
+            vname_num_split_index = file_names[b].rfind('_')# file_names[b]:video_name+'_'+num
+            v_n,num = file_names[b][:vname_num_split_index],file_names[b][vname_num_split_index+1:]
+            save_img_path = os.path.join(savedir,f'{v_n}_sample_{num}.npy')
+            np.save(save_img_path,rcon)
+        
+        return None
+        
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
+                                  list(self.decoder.parameters())+
+                                  list(self.quant_conv.parameters())+
+                                  list(self.post_quant_conv.parameters()),
+                                  lr=lr, betas=(0.5, 0.9))
+        opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
+                                    lr=lr, betas=(0.5, 0.9))
+        return [opt_ae, opt_disc], []
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    @torch.no_grad()
+    def log_images(self, batch, only_inputs=False,save_dir = 'mel_result_ae13_26_debug/fake_class', **kwargs): # 在main.py的on_validation_batch_end中调用
+        log = dict()
+        x = self.get_input(batch, self.image_key)
+        x = x.to(self.device)
+        if not only_inputs:
+            xrec, posterior = self(x)
+            if x.shape[1] > 3:
+                # colorize with random projection
+                assert xrec.shape[1] > 3
+                x = self.to_rgb(x)
+                xrec = self.to_rgb(xrec)
+            log["samples"] = self.decode(torch.randn_like(posterior.sample()))
+            log["reconstructions"] = xrec
+        log["inputs"] = x
+        return log
+
+    def to_rgb(self, x):
+        assert self.image_key == "segmentation"
+        if not hasattr(self, "colorize"):
+            self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
+        x = F.conv2d(x, weight=self.colorize)
+        x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
+        return x
+
+
+class IdentityFirstStage(torch.nn.Module):
+    def __init__(self, *args, vq_interface=False, **kwargs):
+        self.vq_interface = vq_interface  # TODO: Should be true by default but check to not break older stuff
+        super().__init__()
+
+    def encode(self, x, *args, **kwargs):
+        return x
+
+    def decode(self, x, *args, **kwargs):
+        return x
+
+    def quantize(self, x, *args, **kwargs):
+        if self.vq_interface:
+            return x, None, [None, None, None]
+        return x
+
+    def forward(self, x, *args, **kwargs):
+        return x

ldm/models/autoencoder_multi.py

@@ -0,0 +1,201 @@
+"""
+与autoencoder.py的区别在于，autoencoder.py计算loss时只有一个discriminator，而此处又多了个multiwindowDiscriminator，所以优化器
+优化的参数改为：
+opt_disc = torch.optim.Adam(list(self.loss.discriminator.parameters()) + list(self.loss.discriminator_multi.parameters()),
+                            lr=lr, betas=(0.5, 0.9))
+"""
+
+import os
+import torch
+import pytorch_lightning as pl
+import torch.nn.functional as F
+from contextlib import contextmanager
+
+from packaging import version
+import numpy as np
+from ldm.modules.diffusionmodules.model import Encoder, Decoder
+from ldm.modules.distributions.distributions import DiagonalGaussianDistribution
+from torch.optim.lr_scheduler import LambdaLR
+from ldm.util import instantiate_from_config
+
+
+
+class AutoencoderKL(pl.LightningModule):
+    def __init__(self,
+                 ddconfig,
+                 lossconfig,
+                 embed_dim,
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 image_key="image",
+                 colorize_nlabels=None,
+                 monitor=None,
+                 ):
+        super().__init__()
+        self.image_key = image_key
+        self.encoder = Encoder(**ddconfig)
+        self.decoder = Decoder(**ddconfig)
+        self.loss = instantiate_from_config(lossconfig)
+        assert ddconfig["double_z"]
+        self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
+        self.embed_dim = embed_dim
+        if colorize_nlabels is not None:
+            assert type(colorize_nlabels)==int
+            self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
+        if monitor is not None:
+            self.monitor = monitor
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+
+    def init_from_ckpt(self, path, ignore_keys=list()):
+        sd = torch.load(path, map_location="cpu")["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        self.load_state_dict(sd, strict=False)
+        print(f"Restored from {path}")
+
+    def encode(self, x):
+        h = self.encoder(x)
+        moments = self.quant_conv(h)
+        posterior = DiagonalGaussianDistribution(moments)
+        return posterior
+
+    def decode(self, z):
+        z = self.post_quant_conv(z)
+        dec = self.decoder(z)
+        return dec
+
+    def forward(self, input, sample_posterior=True):
+        posterior = self.encode(input)
+        if sample_posterior:
+            z = posterior.sample()
+        else:
+            z = posterior.mode()
+        dec = self.decode(z)
+        return dec, posterior
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float()
+        return x
+
+    def training_step(self, batch, batch_idx, optimizer_idx):
+        inputs = self.get_input(batch, self.image_key)
+        reconstructions, posterior = self(inputs)
+
+        if optimizer_idx == 0:
+            # train encoder+decoder+logvar
+            aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                            last_layer=self.get_last_layer(), split="train")
+            self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return aeloss
+
+        if optimizer_idx == 1:
+            # train the discriminator
+            discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
+                                                last_layer=self.get_last_layer(), split="train")
+
+            self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
+            self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False)
+            return discloss
+
+    def validation_step(self, batch, batch_idx):
+        inputs = self.get_input(batch, self.image_key)
+        reconstructions, posterior = self(inputs)
+        aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step,
+                                        last_layer=self.get_last_layer(), split="val")
+
+        discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step,
+                                            last_layer=self.get_last_layer(), split="val")
+
+        self.log("val/rec_loss", log_dict_ae["val/rec_loss"])
+        self.log_dict(log_dict_ae)
+        self.log_dict(log_dict_disc)
+        return self.log_dict
+
+    def test_step(self, batch, batch_idx):
+        inputs = self.get_input(batch, self.image_key)# inputs shape:(b,c,mel_len,T) or (b,c,h,w)
+        reconstructions, posterior = self(inputs)# reconstructions:(b,c,mel_len,T) or (b,c,h,w)
+        reconstructions = (reconstructions + 1)/2 # to mel scale
+        test_ckpt_path = os.path.basename(self.trainer.tested_ckpt_path)
+        savedir = os.path.join(self.trainer.log_dir,f'output_imgs_{test_ckpt_path}','fake_class')
+        if not os.path.exists(savedir):
+            os.makedirs(savedir)
+
+        file_names = batch['f_name']
+        # print(f"reconstructions.shape:{reconstructions.shape}",file_names)
+        reconstructions = reconstructions.cpu().numpy().squeeze(1) # squuze channel dim
+        for b in range(reconstructions.shape[0]):
+            vname_num_split_index = file_names[b].rfind('_')# file_names[b]:video_name+'_'+num
+            v_n,num = file_names[b][:vname_num_split_index],file_names[b][vname_num_split_index+1:]
+            save_img_path = os.path.join(savedir,f'{v_n}_sample_{num}.npy')
+            np.save(save_img_path,reconstructions[b])
+        
+        return None
+        
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
+                                  list(self.decoder.parameters())+
+                                  list(self.quant_conv.parameters())+
+                                  list(self.post_quant_conv.parameters()),
+                                  lr=lr, betas=(0.5, 0.9))
+        opt_disc = torch.optim.Adam(list(self.loss.discriminator.parameters()) + list(self.loss.discriminator_multi.parameters()),
+                                    lr=lr, betas=(0.5, 0.9))
+        return [opt_ae, opt_disc], []
+
+    def get_last_layer(self):
+        return self.decoder.conv_out.weight
+
+    @torch.no_grad()
+    def log_images(self, batch, only_inputs=False, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.image_key)
+        x = x.to(self.device)
+        if not only_inputs:
+            xrec, posterior = self(x)
+            if x.shape[1] > 3:
+                # colorize with random projection
+                assert xrec.shape[1] > 3
+                x = self.to_rgb(x)
+                xrec = self.to_rgb(xrec)
+            log["samples"] = self.decode(torch.randn_like(posterior.sample()))
+            log["reconstructions"] = xrec
+        log["inputs"] = x
+        return log
+
+    def to_rgb(self, x):
+        assert self.image_key == "segmentation"
+        if not hasattr(self, "colorize"):
+            self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
+        x = F.conv2d(x, weight=self.colorize)
+        x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
+        return x
+
+
+class IdentityFirstStage(torch.nn.Module):
+    def __init__(self, *args, vq_interface=False, **kwargs):
+        self.vq_interface = vq_interface  # TODO: Should be true by default but check to not break older stuff
+        super().__init__()
+
+    def encode(self, x, *args, **kwargs):
+        return x
+
+    def decode(self, x, *args, **kwargs):
+        return x
+
+    def quantize(self, x, *args, **kwargs):
+        if self.vq_interface:
+            return x, None, [None, None, None]
+        return x
+
+    def forward(self, x, *args, **kwargs):
+        return x

ldm/models/diffusion/classifier.py

@@ -0,0 +1,267 @@
+import os
+import torch
+import pytorch_lightning as pl
+from omegaconf import OmegaConf
+from torch.nn import functional as F
+from torch.optim import AdamW
+from torch.optim.lr_scheduler import LambdaLR
+from copy import deepcopy
+from einops import rearrange
+from glob import glob
+from natsort import natsorted
+
+from ldm.modules.diffusionmodules.openaimodel import EncoderUNetModel, UNetModel
+from ldm.util import log_txt_as_img, default, ismap, instantiate_from_config
+
+__models__ = {
+    'class_label': EncoderUNetModel,
+    'segmentation': UNetModel
+}
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+class NoisyLatentImageClassifier(pl.LightningModule):
+
+    def __init__(self,
+                 diffusion_path,
+                 num_classes,
+                 ckpt_path=None,
+                 pool='attention',
+                 label_key=None,
+                 diffusion_ckpt_path=None,
+                 scheduler_config=None,
+                 weight_decay=1.e-2,
+                 log_steps=10,
+                 monitor='val/loss',
+                 *args,
+                 **kwargs):
+        super().__init__(*args, **kwargs)
+        self.num_classes = num_classes
+        # get latest config of diffusion model
+        diffusion_config = natsorted(glob(os.path.join(diffusion_path, 'configs', '*-project.yaml')))[-1]
+        self.diffusion_config = OmegaConf.load(diffusion_config).model
+        self.diffusion_config.params.ckpt_path = diffusion_ckpt_path
+        self.load_diffusion()
+
+        self.monitor = monitor
+        self.numd = self.diffusion_model.first_stage_model.encoder.num_resolutions - 1
+        self.log_time_interval = self.diffusion_model.num_timesteps // log_steps
+        self.log_steps = log_steps
+
+        self.label_key = label_key if not hasattr(self.diffusion_model, 'cond_stage_key') \
+            else self.diffusion_model.cond_stage_key
+
+        assert self.label_key is not None, 'label_key neither in diffusion model nor in model.params'
+
+        if self.label_key not in __models__:
+            raise NotImplementedError()
+
+        self.load_classifier(ckpt_path, pool)
+
+        self.scheduler_config = scheduler_config
+        self.use_scheduler = self.scheduler_config is not None
+        self.weight_decay = weight_decay
+
+    def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
+        sd = torch.load(path, map_location="cpu")
+        if "state_dict" in list(sd.keys()):
+            sd = sd["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
+            sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+        if len(unexpected) > 0:
+            print(f"Unexpected Keys: {unexpected}")
+
+    def load_diffusion(self):
+        model = instantiate_from_config(self.diffusion_config)
+        self.diffusion_model = model.eval()
+        self.diffusion_model.train = disabled_train
+        for param in self.diffusion_model.parameters():
+            param.requires_grad = False
+
+    def load_classifier(self, ckpt_path, pool):
+        model_config = deepcopy(self.diffusion_config.params.unet_config.params)
+        model_config.in_channels = self.diffusion_config.params.unet_config.params.out_channels
+        model_config.out_channels = self.num_classes
+        if self.label_key == 'class_label':
+            model_config.pool = pool
+
+        self.model = __models__[self.label_key](**model_config)
+        if ckpt_path is not None:
+            print('#####################################################################')
+            print(f'load from ckpt "{ckpt_path}"')
+            print('#####################################################################')
+            self.init_from_ckpt(ckpt_path)
+
+    @torch.no_grad()
+    def get_x_noisy(self, x, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x))
+        continuous_sqrt_alpha_cumprod = None
+        if self.diffusion_model.use_continuous_noise:
+            continuous_sqrt_alpha_cumprod = self.diffusion_model.sample_continuous_noise_level(x.shape[0], t + 1)
+            # todo: make sure t+1 is correct here
+
+        return self.diffusion_model.q_sample(x_start=x, t=t, noise=noise,
+                                             continuous_sqrt_alpha_cumprod=continuous_sqrt_alpha_cumprod)
+
+    def forward(self, x_noisy, t, *args, **kwargs):
+        return self.model(x_noisy, t)
+
+    @torch.no_grad()
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = rearrange(x, 'b h w c -> b c h w')
+        x = x.to(memory_format=torch.contiguous_format).float()
+        return x
+
+    @torch.no_grad()
+    def get_conditioning(self, batch, k=None):
+        if k is None:
+            k = self.label_key
+        assert k is not None, 'Needs to provide label key'
+
+        targets = batch[k].to(self.device)
+
+        if self.label_key == 'segmentation':
+            targets = rearrange(targets, 'b h w c -> b c h w')
+            for down in range(self.numd):
+                h, w = targets.shape[-2:]
+                targets = F.interpolate(targets, size=(h // 2, w // 2), mode='nearest')
+
+            # targets = rearrange(targets,'b c h w -> b h w c')
+
+        return targets
+
+    def compute_top_k(self, logits, labels, k, reduction="mean"):
+        _, top_ks = torch.topk(logits, k, dim=1)
+        if reduction == "mean":
+            return (top_ks == labels[:, None]).float().sum(dim=-1).mean().item()
+        elif reduction == "none":
+            return (top_ks == labels[:, None]).float().sum(dim=-1)
+
+    def on_train_epoch_start(self):
+        # save some memory
+        self.diffusion_model.model.to('cpu')
+
+    @torch.no_grad()
+    def write_logs(self, loss, logits, targets):
+        log_prefix = 'train' if self.training else 'val'
+        log = {}
+        log[f"{log_prefix}/loss"] = loss.mean()
+        log[f"{log_prefix}/acc@1"] = self.compute_top_k(
+            logits, targets, k=1, reduction="mean"
+        )
+        log[f"{log_prefix}/acc@5"] = self.compute_top_k(
+            logits, targets, k=5, reduction="mean"
+        )
+
+        self.log_dict(log, prog_bar=False, logger=True, on_step=self.training, on_epoch=True)
+        self.log('loss', log[f"{log_prefix}/loss"], prog_bar=True, logger=False)
+        self.log('global_step', self.global_step, logger=False, on_epoch=False, prog_bar=True)
+        lr = self.optimizers().param_groups[0]['lr']
+        self.log('lr_abs', lr, on_step=True, logger=True, on_epoch=False, prog_bar=True)
+
+    def shared_step(self, batch, t=None):
+        x, *_ = self.diffusion_model.get_input(batch, k=self.diffusion_model.first_stage_key)
+        targets = self.get_conditioning(batch)
+        if targets.dim() == 4:
+            targets = targets.argmax(dim=1)
+        if t is None:
+            t = torch.randint(0, self.diffusion_model.num_timesteps, (x.shape[0],), device=self.device).long()
+        else:
+            t = torch.full(size=(x.shape[0],), fill_value=t, device=self.device).long()
+        x_noisy = self.get_x_noisy(x, t)
+        logits = self(x_noisy, t)
+
+        loss = F.cross_entropy(logits, targets, reduction='none')
+
+        self.write_logs(loss.detach(), logits.detach(), targets.detach())
+
+        loss = loss.mean()
+        return loss, logits, x_noisy, targets
+
+    def training_step(self, batch, batch_idx):
+        loss, *_ = self.shared_step(batch)
+        return loss
+
+    def reset_noise_accs(self):
+        self.noisy_acc = {t: {'acc@1': [], 'acc@5': []} for t in
+                          range(0, self.diffusion_model.num_timesteps, self.diffusion_model.log_every_t)}
+
+    def on_validation_start(self):
+        self.reset_noise_accs()
+
+    @torch.no_grad()
+    def validation_step(self, batch, batch_idx):
+        loss, *_ = self.shared_step(batch)
+
+        for t in self.noisy_acc:
+            _, logits, _, targets = self.shared_step(batch, t)
+            self.noisy_acc[t]['acc@1'].append(self.compute_top_k(logits, targets, k=1, reduction='mean'))
+            self.noisy_acc[t]['acc@5'].append(self.compute_top_k(logits, targets, k=5, reduction='mean'))
+
+        return loss
+
+    def configure_optimizers(self):
+        optimizer = AdamW(self.model.parameters(), lr=self.learning_rate, weight_decay=self.weight_decay)
+
+        if self.use_scheduler:
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(optimizer, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                }]
+            return [optimizer], scheduler
+
+        return optimizer
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, *args, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.diffusion_model.first_stage_key)
+        log['inputs'] = x
+
+        y = self.get_conditioning(batch)
+
+        if self.label_key == 'class_label':
+            y = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
+            log['labels'] = y
+
+        if ismap(y):
+            log['labels'] = self.diffusion_model.to_rgb(y)
+
+            for step in range(self.log_steps):
+                current_time = step * self.log_time_interval
+
+                _, logits, x_noisy, _ = self.shared_step(batch, t=current_time)
+
+                log[f'inputs@t{current_time}'] = x_noisy
+
+                pred = F.one_hot(logits.argmax(dim=1), num_classes=self.num_classes)
+                pred = rearrange(pred, 'b h w c -> b c h w')
+
+                log[f'pred@t{current_time}'] = self.diffusion_model.to_rgb(pred)
+
+        for key in log:
+            log[key] = log[key][:N]
+
+        return log

ldm/models/diffusion/ddim.py

@@ -0,0 +1,262 @@
+"""SAMPLING ONLY."""
+
+import torch
+import numpy as np
+from tqdm import tqdm
+from functools import partial
+
+from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like, \
+    extract_into_tensor
+
+
+class DDIMSampler(object):
+    def __init__(self, model, schedule="linear", **kwargs):
+        super().__init__()
+        self.model = model
+        self.device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.schedule = schedule
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            # if attr.device != torch.device("cuda"):
+            #     attr = attr.to(torch.device("cuda"))
+            attr = attr.to(self.device)
+        setattr(self, name, attr)
+
+    def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
+        self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
+                                                  num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
+        alphas_cumprod = self.model.alphas_cumprod
+        assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
+
+        self.register_buffer('betas', to_torch(self.model.betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
+
+        # ddim sampling parameters
+        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
+                                                                                   ddim_timesteps=self.ddim_timesteps,
+                                                                                   eta=ddim_eta,verbose=verbose)
+        self.register_buffer('ddim_sigmas', ddim_sigmas)
+        self.register_buffer('ddim_alphas', ddim_alphas)
+        self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
+        self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
+        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
+            (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
+                        1 - self.alphas_cumprod / self.alphas_cumprod_prev))
+        self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
+
+    @torch.no_grad()
+    def sample(self,
+               S,
+               batch_size,
+               shape,
+               conditioning=None,
+               callback=None,
+               normals_sequence=None,
+               img_callback=None,
+               quantize_x0=False,
+               eta=0.,
+               mask=None,
+               x0=None,
+               temperature=1.,
+               noise_dropout=0.,
+               score_corrector=None,
+               corrector_kwargs=None,
+               verbose=True,
+               x_T=None,
+               log_every_t=100,
+               unconditional_guidance_scale=1.,
+               unconditional_conditioning=None,
+               # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+               **kwargs
+               ):
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                ctmp = conditioning[list(conditioning.keys())[0]]
+                while isinstance(ctmp, list): ctmp = ctmp[0]
+                cbs = ctmp.shape[0]
+                if cbs != batch_size:
+                    print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
+
+        self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
+        # sampling
+        C, H, W = shape
+        size = (batch_size, C, H, W)
+        # print(f'Data shape for DDIM sampling is {size}, eta {eta}')
+
+        samples, intermediates = self.ddim_sampling(conditioning, size,
+                                                    callback=callback,
+                                                    img_callback=img_callback,
+                                                    quantize_denoised=quantize_x0,
+                                                    mask=mask, x0=x0,
+                                                    ddim_use_original_steps=False,
+                                                    noise_dropout=noise_dropout,
+                                                    temperature=temperature,
+                                                    score_corrector=score_corrector,
+                                                    corrector_kwargs=corrector_kwargs,
+                                                    x_T=x_T,
+                                                    log_every_t=log_every_t,
+                                                    unconditional_guidance_scale=unconditional_guidance_scale,
+                                                    unconditional_conditioning=unconditional_conditioning,
+                                                    )
+        return samples, intermediates
+
+    @torch.no_grad()
+    def ddim_sampling(self, cond, shape,
+                      x_T=None, ddim_use_original_steps=False,
+                      callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, log_every_t=100,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None,):
+        device = self.model.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        if timesteps is None:
+            timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
+        elif timesteps is not None and not ddim_use_original_steps:
+            subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
+            timesteps = self.ddim_timesteps[:subset_end]
+
+        intermediates = {'x_inter': [img], 'pred_x0': [img]}
+        time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
+        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
+
+        # iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
+
+        for i, step in enumerate(time_range):
+            index = total_steps - i - 1
+            ts = torch.full((b,), step, device=device, dtype=torch.long)
+
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.model.q_sample(x0, ts)  # TODO: deterministic forward pass?
+                img = img_orig * mask + (1. - mask) * img
+
+            outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
+                                      quantize_denoised=quantize_denoised, temperature=temperature,
+                                      noise_dropout=noise_dropout, score_corrector=score_corrector,
+                                      corrector_kwargs=corrector_kwargs,
+                                      unconditional_guidance_scale=unconditional_guidance_scale,
+                                      unconditional_conditioning=unconditional_conditioning)
+            img, pred_x0 = outs
+            if callback: callback(i)
+            if img_callback: img_callback(pred_x0, i)
+
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates['x_inter'].append(img)
+                intermediates['pred_x0'].append(pred_x0)
+
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None):
+        b, *_, device = *x.shape, x.device
+
+        if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
+            e_t = self.model.apply_model(x, t, c)
+        else:
+            x_in = torch.cat([x] * 2)
+            t_in = torch.cat([t] * 2)
+            if isinstance(c, dict):
+                assert isinstance(unconditional_conditioning, dict)
+                c_in = dict()
+                for k in c:
+                    if isinstance(c[k], list):
+                        c_in[k] = [torch.cat([
+                            unconditional_conditioning[k][i],
+                            c[k][i]]) for i in range(len(c[k]))]
+                    else:
+                        c_in[k] = torch.cat([
+                            unconditional_conditioning[k],
+                            c[k]])
+            elif isinstance(c, list):
+                c_in = list()
+                assert isinstance(unconditional_conditioning, list)
+                for i in range(len(c)):
+                    c_in.append(torch.cat([unconditional_conditioning[i], c[i]]))
+            else:
+                c_in = torch.cat([unconditional_conditioning, c])# c/uc shape [b,seq_len=77,dim=1024],c_in shape [b*2,seq_len,dim]
+            e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
+            e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
+
+        if score_corrector is not None:
+            assert self.model.parameterization == "eps"
+            e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
+
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
+        sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
+        sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
+        # select parameters corresponding to the currently considered timestep
+        a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
+        a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
+        sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
+        sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
+
+        # current prediction for x_0
+        pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
+        if quantize_denoised:
+            pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+        # direction pointing to x_t
+        dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
+        noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
+        return x_prev, pred_x0
+
+    @torch.no_grad()
+    def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
+        # fast, but does not allow for exact reconstruction
+        # t serves as an index to gather the correct alphas
+        if use_original_steps:
+            sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
+            sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
+        else:
+            sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
+            sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
+
+        if noise is None:
+            noise = torch.randn_like(x0)
+        return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
+                extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise)
+
+    @torch.no_grad()
+    def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None,
+               use_original_steps=False):
+
+        timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps
+        timesteps = timesteps[:t_start]
+
+        time_range = np.flip(timesteps)
+        total_steps = timesteps.shape[0]
+        # print(f"Running DDIM Sampling with {total_steps} timesteps")
+
+        # iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
+        x_dec = x_latent
+        for i, step in enumerate(time_range):
+            index = total_steps - i - 1
+            ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long)
+            x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps,
+                                          unconditional_guidance_scale=unconditional_guidance_scale,
+                                          unconditional_conditioning=unconditional_conditioning)
+        return x_dec

ldm/models/diffusion/ddpm.py

@@ -0,0 +1,1444 @@
+"""
+wild mixture of
+https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
+https://github.com/CompVis/taming-transformers
+-- merci
+"""
+import torch
+import torch.nn as nn
+import numpy as np
+import pytorch_lightning as pl
+from torch.optim.lr_scheduler import LambdaLR
+from einops import rearrange, repeat
+from contextlib import contextmanager
+from functools import partial
+from tqdm import tqdm
+from torchvision.utils import make_grid
+from pytorch_lightning.utilities.distributed import rank_zero_only
+
+from ldm.util import log_txt_as_img, exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config
+from ldm.modules.ema import LitEma
+from ldm.modules.distributions.distributions import normal_kl, DiagonalGaussianDistribution
+from ldm.models.autoencoder import IdentityFirstStage, AutoencoderKL
+from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like
+from ldm.models.diffusion.ddim import DDIMSampler
+
+
+__conditioning_keys__ = {'concat': 'c_concat',
+                         'crossattn': 'c_crossattn',
+                         'adm': 'y'}
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+def uniform_on_device(r1, r2, shape, device):
+    return (r1 - r2) * torch.rand(*shape, device=device) + r2
+
+
+class DDPM(pl.LightningModule):
+    # classic DDPM with Gaussian diffusion, in image space
+    def __init__(self,
+                 unet_config,
+                 timesteps=1000,
+                 beta_schedule="linear",
+                 loss_type="l2",
+                 ckpt_path=None,
+                 ignore_keys=[],
+                 load_only_unet=False,
+                 monitor="val/loss",
+                 use_ema=True,
+                 first_stage_key="image",
+                 image_size=256,
+                 channels=3,
+                 log_every_t=100,
+                 clip_denoised=True,
+                 linear_start=1e-4,
+                 linear_end=2e-2,
+                 cosine_s=8e-3,
+                 given_betas=None,
+                 original_elbo_weight=0.,
+                 v_posterior=0.,  # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
+                 l_simple_weight=1.,
+                 conditioning_key=None,
+                 parameterization="eps",  # all config files uses "eps"
+                 scheduler_config=None,
+                 use_positional_encodings=False,
+                 learn_logvar=False,
+                 logvar_init=0.,
+                 ):
+        super().__init__()
+        assert parameterization in ["eps", "x0"], 'currently only supporting "eps" and "x0"'
+        self.parameterization = parameterization
+        print(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode")
+        self.cond_stage_model = None
+        self.clip_denoised = clip_denoised
+        self.log_every_t = log_every_t
+        self.first_stage_key = first_stage_key
+        self.image_size = image_size  # try conv?
+        self.channels = channels
+        self.use_positional_encodings = use_positional_encodings
+        self.model = DiffusionWrapper(unet_config, conditioning_key)
+        count_params(self.model, verbose=True)
+        self.use_ema = use_ema
+        if self.use_ema:
+            self.model_ema = LitEma(self.model)
+            print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
+
+        self.use_scheduler = scheduler_config is not None
+        if self.use_scheduler:
+            self.scheduler_config = scheduler_config
+
+        self.v_posterior = v_posterior
+        self.original_elbo_weight = original_elbo_weight
+        self.l_simple_weight = l_simple_weight
+
+        if monitor is not None:
+            self.monitor = monitor
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet)
+
+        self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps,
+                               linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
+
+        self.loss_type = loss_type
+
+        self.learn_logvar = learn_logvar
+        self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,))
+        if self.learn_logvar:
+            self.logvar = nn.Parameter(self.logvar, requires_grad=True)
+
+    def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        if exists(given_betas):
+            betas = given_betas
+        else:
+            betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
+                                       cosine_s=cosine_s)
+        alphas = 1. - betas
+        alphas_cumprod = np.cumprod(alphas, axis=0)
+        alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
+
+        timesteps, = betas.shape
+        self.num_timesteps = int(timesteps)
+        self.linear_start = linear_start
+        self.linear_end = linear_end
+        assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
+
+        to_torch = partial(torch.tensor, dtype=torch.float32)
+
+        self.register_buffer('betas', to_torch(betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
+
+        # calculations for posterior q(x_{t-1} | x_t, x_0)
+        posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / (
+                    1. - alphas_cumprod) + self.v_posterior * betas
+        # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
+        self.register_buffer('posterior_variance', to_torch(posterior_variance))
+        # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
+        self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
+        self.register_buffer('posterior_mean_coef1', to_torch(
+            betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
+        self.register_buffer('posterior_mean_coef2', to_torch(
+            (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
+
+        if self.parameterization == "eps":
+            lvlb_weights = self.betas ** 2 / (
+                        2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))
+        elif self.parameterization == "x0":
+            lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod))
+        else:
+            raise NotImplementedError("mu not supported")
+        # TODO how to choose this term
+        lvlb_weights[0] = lvlb_weights[1]
+        self.register_buffer('lvlb_weights', lvlb_weights, persistent=False)
+        assert not torch.isnan(self.lvlb_weights).all()
+
+    @contextmanager
+    def ema_scope(self, context=None):
+        if self.use_ema:
+            self.model_ema.store(self.model.parameters())
+            self.model_ema.copy_to(self.model)
+            if context is not None:
+                print(f"{context}: Switched to EMA weights")
+        try:
+            yield None
+        finally:
+            if self.use_ema:
+                self.model_ema.restore(self.model.parameters())
+                if context is not None:
+                    print(f"{context}: Restored training weights")
+
+    def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
+        sd = torch.load(path, map_location="cpu")
+        if "state_dict" in list(sd.keys()):
+            sd = sd["state_dict"]
+        keys = list(sd.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+        missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
+            sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+        if len(unexpected) > 0:
+            print(f"Unexpected Keys: {unexpected}")
+
+    def q_mean_variance(self, x_start, t):
+        """
+        Get the distribution q(x_t | x_0).
+        :param x_start: the [N x C x ...] tensor of noiseless inputs.
+        :param t: the number of diffusion steps (minus 1). Here, 0 means one step.
+        :return: A tuple (mean, variance, log_variance), all of x_start's shape.
+        """
+        mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start)
+        variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
+        log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape)
+        return mean, variance, log_variance
+
+    def predict_start_from_noise(self, x_t, t, noise):
+        return (
+                extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
+                extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
+        )
+
+    def q_posterior(self, x_start, x_t, t):
+        posterior_mean = (
+                extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start +
+                extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
+        )
+        posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape)
+        posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape)
+        return posterior_mean, posterior_variance, posterior_log_variance_clipped
+
+    def p_mean_variance(self, x, t, clip_denoised: bool):
+        model_out = self.model(x, t)
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        if clip_denoised:
+            x_recon.clamp_(-1., 1.)
+
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+        return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(self, x, t, clip_denoised=True, repeat_noise=False):
+        b, *_, device = *x.shape, x.device
+        model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised)
+        noise = noise_like(x.shape, device, repeat_noise)
+        # no noise when t == 0
+        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
+        return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def p_sample_loop(self, shape, return_intermediates=False):
+        device = self.betas.device
+        b = shape[0]
+        img = torch.randn(shape, device=device)
+        intermediates = [img]
+        for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps):
+            img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long),
+                                clip_denoised=self.clip_denoised)
+            if i % self.log_every_t == 0 or i == self.num_timesteps - 1:
+                intermediates.append(img)
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+    @torch.no_grad()
+    def sample(self, batch_size=16, return_intermediates=False):
+        image_size = self.image_size
+        channels = self.channels
+        return self.p_sample_loop((batch_size, channels, image_size, image_size),
+                                  return_intermediates=return_intermediates)
+
+    def q_sample(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
+                extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
+
+    def get_loss(self, pred, target, mean=True):
+        if self.loss_type == 'l1':
+            loss = (target - pred).abs()
+            if mean:
+                loss = loss.mean()
+        elif self.loss_type == 'l2':
+            if mean:
+                loss = torch.nn.functional.mse_loss(target, pred)
+            else:
+                loss = torch.nn.functional.mse_loss(target, pred, reduction='none')
+        else:
+            raise NotImplementedError("unknown loss type '{loss_type}'")
+
+        return loss
+
+    def p_losses(self, x_start, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+        model_out = self.model(x_noisy, t)
+
+        loss_dict = {}
+        if self.parameterization == "eps":
+            target = noise
+        elif self.parameterization == "x0":
+            target = x_start
+        else:
+            raise NotImplementedError(f"Paramterization {self.parameterization} not yet supported")
+
+        loss = self.get_loss(model_out, target, mean=False).mean(dim=[1, 2, 3])
+
+        log_prefix = 'train' if self.training else 'val'
+
+        loss_dict.update({f'{log_prefix}/loss_simple': loss.mean()})
+        loss_simple = loss.mean() * self.l_simple_weight
+
+        loss_vlb = (self.lvlb_weights[t] * loss).mean()
+        loss_dict.update({f'{log_prefix}/loss_vlb': loss_vlb})
+
+        loss = loss_simple + self.original_elbo_weight * loss_vlb
+
+        loss_dict.update({f'{log_prefix}/loss': loss})
+
+        return loss, loss_dict
+
+    def forward(self, x, *args, **kwargs):
+        # b, c, h, w, device, img_size, = *x.shape, x.device, self.image_size
+        # assert h == img_size and w == img_size, f'height and width of image must be {img_size}'
+        t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
+        return self.p_losses(x, t, *args, **kwargs)
+
+    def get_input(self, batch, k):
+        x = batch[k]
+        if len(x.shape) == 3:
+            x = x[..., None]
+        x = rearrange(x, 'b h w c -> b c h w')
+        x = x.to(memory_format=torch.contiguous_format).float()
+        return x
+
+    def shared_step(self, batch):
+        x = self.get_input(batch, self.first_stage_key)
+        loss, loss_dict = self(x)
+        return loss, loss_dict
+
+    def training_step(self, batch, batch_idx):
+        loss, loss_dict = self.shared_step(batch)
+
+        self.log_dict(loss_dict, prog_bar=True,
+                      logger=True, on_step=True, on_epoch=True)
+
+        self.log("global_step", self.global_step,
+                 prog_bar=True, logger=True, on_step=True, on_epoch=False)
+
+        if self.use_scheduler:
+            lr = self.optimizers().param_groups[0]['lr']
+            self.log('lr_abs', lr, prog_bar=True, logger=True, on_step=True, on_epoch=False)
+
+        return loss
+
+    @torch.no_grad()
+    def validation_step(self, batch, batch_idx):
+        _, loss_dict_no_ema = self.shared_step(batch)
+        with self.ema_scope():
+            _, loss_dict_ema = self.shared_step(batch)
+            loss_dict_ema = {key + '_ema': loss_dict_ema[key] for key in loss_dict_ema}
+        self.log_dict(loss_dict_no_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
+        self.log_dict(loss_dict_ema, prog_bar=False, logger=True, on_step=False, on_epoch=True)
+
+    def on_train_batch_end(self, *args, **kwargs):
+        if self.use_ema:
+            self.model_ema(self.model)
+
+    def _get_rows_from_list(self, samples):
+        n_imgs_per_row = len(samples)
+        denoise_grid = rearrange(samples, 'n b c h w -> b n c h w')
+        denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
+        denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
+        return denoise_grid
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs):
+        log = dict()
+        x = self.get_input(batch, self.first_stage_key)
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        x = x.to(self.device)[:N]
+        log["inputs"] = x
+
+        # get diffusion row
+        diffusion_row = list()
+        x_start = x[:n_row]
+
+        for t in range(self.num_timesteps):
+            if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
+                t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
+                t = t.to(self.device).long()
+                noise = torch.randn_like(x_start)
+                x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+                diffusion_row.append(x_noisy)
+
+        log["diffusion_row"] = self._get_rows_from_list(diffusion_row)
+
+        if sample:
+            # get denoise row
+            with self.ema_scope("Plotting"):
+                samples, denoise_row = self.sample(batch_size=N, return_intermediates=True)
+
+            log["samples"] = samples
+            log["denoise_row"] = self._get_rows_from_list(denoise_row)
+
+        if return_keys:
+            if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
+                return log
+            else:
+                return {key: log[key] for key in return_keys}
+        return log
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        params = list(self.model.parameters())
+        if self.learn_logvar:
+            params = params + [self.logvar]
+        opt = torch.optim.AdamW(params, lr=lr)
+        return opt
+
+
+class LatentDiffusion(DDPM):
+    """main class"""
+    def __init__(self,
+                 first_stage_config,
+                 cond_stage_config,
+                 num_timesteps_cond=None,
+                 cond_stage_key="image",# 'caption' for txt2image, 'masked_image' for inpainting 
+                 cond_stage_trainable=False,
+                 concat_mode=True,# true for inpainting
+                 cond_stage_forward=None,
+                 conditioning_key=None, # 'crossattn' for txt2image, None for inpainting
+                 scale_factor=1.0,
+                 scale_by_std=False,
+                 *args, **kwargs):
+        self.num_timesteps_cond = default(num_timesteps_cond, 1)
+        self.scale_by_std = scale_by_std
+        assert self.num_timesteps_cond <= kwargs['timesteps']
+        # for backwards compatibility after implementation of DiffusionWrapper
+        if conditioning_key is None:
+            conditioning_key = 'concat' if concat_mode else 'crossattn'
+        if cond_stage_config == '__is_unconditional__':
+            conditioning_key = None
+        ckpt_path = kwargs.pop("ckpt_path", None)
+        ignore_keys = kwargs.pop("ignore_keys", [])
+        super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
+        self.concat_mode = concat_mode
+        self.cond_stage_trainable = cond_stage_trainable
+        self.cond_stage_key = cond_stage_key 
+        try:
+            self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
+        except:
+            self.num_downs = 0
+        if not scale_by_std:
+            self.scale_factor = scale_factor
+        else:
+            self.register_buffer('scale_factor', torch.tensor(scale_factor))
+        self.instantiate_first_stage(first_stage_config)
+        self.instantiate_cond_stage(cond_stage_config)
+        self.cond_stage_forward = cond_stage_forward
+        self.clip_denoised = False
+        self.bbox_tokenizer = None  
+
+        self.restarted_from_ckpt = False
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys)
+            self.restarted_from_ckpt = True
+
+    def make_cond_schedule(self, ):
+        self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
+        ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
+        self.cond_ids[:self.num_timesteps_cond] = ids
+
+    @rank_zero_only
+    @torch.no_grad()
+    def on_train_batch_start(self, batch, batch_idx, dataloader_idx):
+        # only for very first batch
+        if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt:
+            assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
+            # set rescale weight to 1./std of encodings
+            print("### USING STD-RESCALING ###")
+            x = super().get_input(batch, self.first_stage_key)
+            x = x.to(self.device)
+            encoder_posterior = self.encode_first_stage(x)
+            z = self.get_first_stage_encoding(encoder_posterior).detach()
+            del self.scale_factor
+            self.register_buffer('scale_factor', 1. / z.flatten().std())
+            print(f"setting self.scale_factor to {self.scale_factor}")
+            print("### USING STD-RESCALING ###")
+
+    def register_schedule(self,
+                          given_betas=None, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s)
+
+        self.shorten_cond_schedule = self.num_timesteps_cond > 1
+        if self.shorten_cond_schedule:
+            self.make_cond_schedule()
+
+    def instantiate_first_stage(self, config):
+        model = instantiate_from_config(config)
+        self.first_stage_model = model.eval()
+        self.first_stage_model.train = disabled_train
+        for param in self.first_stage_model.parameters():
+            param.requires_grad = False
+
+    def instantiate_cond_stage(self, config):
+        if not self.cond_stage_trainable:
+            if config == "__is_first_stage__":# inpaint
+                print("Using first stage also as cond stage.")
+                self.cond_stage_model = self.first_stage_model
+            elif config == "__is_unconditional__":
+                print(f"Training {self.__class__.__name__} as an unconditional model.")
+                self.cond_stage_model = None
+                # self.be_unconditional = True
+            else:
+                model = instantiate_from_config(config)
+                self.cond_stage_model = model.eval()
+                self.cond_stage_model.train = disabled_train
+                for param in self.cond_stage_model.parameters():
+                    param.requires_grad = False
+        else:
+            assert config != '__is_first_stage__'
+            assert config != '__is_unconditional__'
+            model = instantiate_from_config(config)
+            self.cond_stage_model = model
+
+    def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False):
+        denoise_row = []
+        for zd in tqdm(samples, desc=desc):
+            denoise_row.append(self.decode_first_stage(zd.to(self.device),
+                                                            force_not_quantize=force_no_decoder_quantization))
+        n_imgs_per_row = len(denoise_row)
+        denoise_row = torch.stack(denoise_row)  # n_log_step, n_row, C, H, W
+        denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
+        denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
+        denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
+        return denoise_grid
+
+    def get_first_stage_encoding(self, encoder_posterior):
+        if isinstance(encoder_posterior, DiagonalGaussianDistribution):
+            z = encoder_posterior.sample()
+        elif isinstance(encoder_posterior, torch.Tensor):
+            z = encoder_posterior
+        else:
+            raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
+        return self.scale_factor * z
+
+    def get_learned_conditioning(self, c):
+        if self.cond_stage_forward is None:
+            if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
+                c = self.cond_stage_model.encode(c)
+                if isinstance(c, DiagonalGaussianDistribution):
+                    c = c.mode()
+            else:
+                c = self.cond_stage_model(c)
+        else:
+            assert hasattr(self.cond_stage_model, self.cond_stage_forward)
+            c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
+        return c
+
+    def meshgrid(self, h, w):
+        y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1)
+        x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1)
+
+        arr = torch.cat([y, x], dim=-1)
+        return arr
+
+    def delta_border(self, h, w):
+        """
+        :param h: height
+        :param w: width
+        :return: normalized distance to image border,
+         wtith min distance = 0 at border and max dist = 0.5 at image center
+        """
+        lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2)
+        arr = self.meshgrid(h, w) / lower_right_corner
+        dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0]
+        dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0]
+        edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0]
+        return edge_dist
+
+    def get_weighting(self, h, w, Ly, Lx, device):
+        weighting = self.delta_border(h, w)
+        weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"],
+                               self.split_input_params["clip_max_weight"], )
+        weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device)
+
+        if self.split_input_params["tie_braker"]:
+            L_weighting = self.delta_border(Ly, Lx)
+            L_weighting = torch.clip(L_weighting,
+                                     self.split_input_params["clip_min_tie_weight"],
+                                     self.split_input_params["clip_max_tie_weight"])
+
+            L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device)
+            weighting = weighting * L_weighting
+        return weighting
+
+    def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1):  # todo load once not every time, shorten code
+        """
+        :param x: img of size (bs, c, h, w)
+        :return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1])
+        """
+        bs, nc, h, w = x.shape
+
+        # number of crops in image
+        Ly = (h - kernel_size[0]) // stride[0] + 1
+        Lx = (w - kernel_size[1]) // stride[1] + 1
+
+        if uf == 1 and df == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params)
+
+            weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h, w)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx))
+
+        elif uf > 1 and df == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf),
+                                dilation=1, padding=0,
+                                stride=(stride[0] * uf, stride[1] * uf))
+            fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2)
+
+            weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h * uf, w * uf)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx))
+
+        elif df > 1 and uf == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df),
+                                dilation=1, padding=0,
+                                stride=(stride[0] // df, stride[1] // df))
+            fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2)
+
+            weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h // df, w // df)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx))
+
+        else:
+            raise NotImplementedError
+
+        return fold, unfold, normalization, weighting
+
+    @torch.no_grad()
+    def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False,
+                  cond_key=None, return_original_cond=False, bs=None):
+        x = super().get_input(batch, k)
+        if bs is not None:
+            x = x[:bs]
+        x = x.to(self.device)
+        encoder_posterior = self.encode_first_stage(x)
+        z = self.get_first_stage_encoding(encoder_posterior).detach()
+
+        if self.model.conditioning_key is not None:
+            if cond_key is None:
+                cond_key = self.cond_stage_key
+            if cond_key != self.first_stage_key:# cond_key is not image. for inapint it's masked_img
+                if cond_key in ['caption', 'coordinates_bbox']:
+                    xc = batch[cond_key]
+                elif cond_key == 'class_label':
+                    xc = batch
+                else:
+                    xc = super().get_input(batch, cond_key).to(self.device)
+            else:
+                xc = x
+            if not self.cond_stage_trainable or force_c_encode:
+                if isinstance(xc, dict) or isinstance(xc, list):
+                    # import pudb; pudb.set_trace()
+                    c = self.get_learned_conditioning(xc)
+                else:
+                    c = self.get_learned_conditioning(xc.to(self.device))
+            else:
+                c = xc
+            if bs is not None:
+                c = c[:bs]
+
+            if self.use_positional_encodings:
+                pos_x, pos_y = self.compute_latent_shifts(batch)
+                ckey = __conditioning_keys__[self.model.conditioning_key]
+                c = {ckey: c, 'pos_x': pos_x, 'pos_y': pos_y}
+
+        else:
+            c = None
+            xc = None
+            if self.use_positional_encodings:
+                pos_x, pos_y = self.compute_latent_shifts(batch)
+                c = {'pos_x': pos_x, 'pos_y': pos_y}
+        out = [z, c]
+        if return_first_stage_outputs:
+            xrec = self.decode_first_stage(z)
+            out.extend([x, xrec])
+        if return_original_cond:
+            out.append(xc)
+        return out
+
+    @torch.no_grad()
+    def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
+        if predict_cids:
+            if z.dim() == 4:
+                z = torch.argmax(z.exp(), dim=1).long()
+            z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
+            z = rearrange(z, 'b h w c -> b c h w').contiguous()
+
+        z = 1. / self.scale_factor * z
+
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                uf = self.split_input_params["vqf"]
+                bs, nc, h, w = z.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
+
+                z = unfold(z)  # (bn, nc * prod(**ks), L)
+                # 1. Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                # 2. apply model loop over last dim
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
+                                                                 force_not_quantize=predict_cids or force_not_quantize)
+                                   for i in range(z.shape[-1])]
+                else:
+
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
+                                   for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)  # # (bn, nc, ks[0], ks[1], L)
+                o = o * weighting
+                # Reverse 1. reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization  # norm is shape (1, 1, h, w)
+                return decoded
+            else:
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+                else:
+                    return self.first_stage_model.decode(z)
+
+        else:
+            if isinstance(self.first_stage_model, VQModelInterface):
+                return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+            else:
+                return self.first_stage_model.decode(z)
+
+    # same as above but without decorator
+    def differentiable_decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
+        if predict_cids:
+            if z.dim() == 4:
+                z = torch.argmax(z.exp(), dim=1).long()
+            z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
+            z = rearrange(z, 'b h w c -> b c h w').contiguous()
+
+        z = 1. / self.scale_factor * z
+
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                uf = self.split_input_params["vqf"]
+                bs, nc, h, w = z.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
+
+                z = unfold(z)  # (bn, nc * prod(**ks), L)
+                # 1. Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                # 2. apply model loop over last dim
+                if isinstance(self.first_stage_model, VQModelInterface):  
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
+                                                                 force_not_quantize=predict_cids or force_not_quantize)
+                                   for i in range(z.shape[-1])]
+                else:
+
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
+                                   for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)  # # (bn, nc, ks[0], ks[1], L)
+                o = o * weighting
+                # Reverse 1. reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization  # norm is shape (1, 1, h, w)
+                return decoded
+            else:
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+                else:
+                    return self.first_stage_model.decode(z)
+
+        else:
+            if isinstance(self.first_stage_model, VQModelInterface):
+                return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+            else:
+                return self.first_stage_model.decode(z)
+
+    @torch.no_grad()
+    def encode_first_stage(self, x):
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                df = self.split_input_params["vqf"]
+                self.split_input_params['original_image_size'] = x.shape[-2:]
+                bs, nc, h, w = x.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(x, ks, stride, df=df)
+                z = unfold(x)  # (bn, nc * prod(**ks), L)
+                # Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                output_list = [self.first_stage_model.encode(z[:, :, :, :, i])
+                               for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)
+                o = o * weighting
+
+                # Reverse reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization
+                return decoded
+
+            else:
+                return self.first_stage_model.encode(x)
+        else:
+            return self.first_stage_model.encode(x)
+
+    def shared_step(self, batch, **kwargs):
+        x, c = self.get_input(batch, self.first_stage_key)
+        loss = self(x, c)
+        return loss
+
+    def forward(self, x, c, *args, **kwargs):
+        t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
+        if self.model.conditioning_key is not None:
+            assert c is not None
+            if self.cond_stage_trainable:# true when use text
+                c = self.get_learned_conditioning(c) # c: string list -> [B, T, Context_dim]
+            if self.shorten_cond_schedule:  # TODO: drop this option
+                tc = self.cond_ids[t].to(self.device)
+                c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float()))
+        return self.p_losses(x, c, t, *args, **kwargs)
+
+    def _rescale_annotations(self, bboxes, crop_coordinates):  # TODO: move to dataset
+        def rescale_bbox(bbox):
+            x0 = clamp((bbox[0] - crop_coordinates[0]) / crop_coordinates[2])
+            y0 = clamp((bbox[1] - crop_coordinates[1]) / crop_coordinates[3])
+            w = min(bbox[2] / crop_coordinates[2], 1 - x0)
+            h = min(bbox[3] / crop_coordinates[3], 1 - y0)
+            return x0, y0, w, h
+
+        return [rescale_bbox(b) for b in bboxes]
+
+    def apply_model(self, x_noisy, t, cond, return_ids=False):
+
+        if isinstance(cond, dict):
+            # hybrid case, cond is exptected to be a dict
+            pass
+        else:
+            if not isinstance(cond, list):
+                cond = [cond]
+            key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn'
+            cond = {key: cond}
+
+        if hasattr(self, "split_input_params"):
+            assert len(cond) == 1  # todo can only deal with one conditioning atm
+            assert not return_ids  
+            ks = self.split_input_params["ks"]  # eg. (128, 128)
+            stride = self.split_input_params["stride"]  # eg. (64, 64)
+
+            h, w = x_noisy.shape[-2:]
+
+            fold, unfold, normalization, weighting = self.get_fold_unfold(x_noisy, ks, stride)
+
+            z = unfold(x_noisy)  # (bn, nc * prod(**ks), L)
+            # Reshape to img shape
+            z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+            z_list = [z[:, :, :, :, i] for i in range(z.shape[-1])]
+
+            if self.cond_stage_key in ["image", "LR_image", "segmentation",
+                                       'bbox_img'] and self.model.conditioning_key:  # todo check for completeness
+                c_key = next(iter(cond.keys()))  # get key
+                c = next(iter(cond.values()))  # get value
+                assert (len(c) == 1)  # todo extend to list with more than one elem
+                c = c[0]  # get element
+
+                c = unfold(c)
+                c = c.view((c.shape[0], -1, ks[0], ks[1], c.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                cond_list = [{c_key: [c[:, :, :, :, i]]} for i in range(c.shape[-1])]
+
+            elif self.cond_stage_key == 'coordinates_bbox':
+                assert 'original_image_size' in self.split_input_params, 'BoudingBoxRescaling is missing original_image_size'
+
+                # assuming padding of unfold is always 0 and its dilation is always 1
+                n_patches_per_row = int((w - ks[0]) / stride[0] + 1)
+                full_img_h, full_img_w = self.split_input_params['original_image_size']
+                # as we are operating on latents, we need the factor from the original image size to the
+                # spatial latent size to properly rescale the crops for regenerating the bbox annotations
+                num_downs = self.first_stage_model.encoder.num_resolutions - 1
+                rescale_latent = 2 ** (num_downs)
+
+                # get top left postions of patches as conforming for the bbbox tokenizer, therefore we
+                # need to rescale the tl patch coordinates to be in between (0,1)
+                tl_patch_coordinates = [(rescale_latent * stride[0] * (patch_nr % n_patches_per_row) / full_img_w,
+                                         rescale_latent * stride[1] * (patch_nr // n_patches_per_row) / full_img_h)
+                                        for patch_nr in range(z.shape[-1])]
+
+                # patch_limits are tl_coord, width and height coordinates as (x_tl, y_tl, h, w)
+                patch_limits = [(x_tl, y_tl,
+                                 rescale_latent * ks[0] / full_img_w,
+                                 rescale_latent * ks[1] / full_img_h) for x_tl, y_tl in tl_patch_coordinates]
+                # patch_values = [(np.arange(x_tl,min(x_tl+ks, 1.)),np.arange(y_tl,min(y_tl+ks, 1.))) for x_tl, y_tl in tl_patch_coordinates]
+
+                # tokenize crop coordinates for the bounding boxes of the respective patches
+                patch_limits_tknzd = [torch.LongTensor(self.bbox_tokenizer._crop_encoder(bbox))[None].to(self.device)
+                                      for bbox in patch_limits]  # list of length l with tensors of shape (1, 2)
+                print(patch_limits_tknzd[0].shape)
+                # cut tknzd crop position from conditioning
+                assert isinstance(cond, dict), 'cond must be dict to be fed into model'
+                cut_cond = cond['c_crossattn'][0][..., :-2].to(self.device)
+                print(cut_cond.shape)
+
+                adapted_cond = torch.stack([torch.cat([cut_cond, p], dim=1) for p in patch_limits_tknzd])
+                adapted_cond = rearrange(adapted_cond, 'l b n -> (l b) n')
+                print(adapted_cond.shape)
+                adapted_cond = self.get_learned_conditioning(adapted_cond)
+                print(adapted_cond.shape)
+                adapted_cond = rearrange(adapted_cond, '(l b) n d -> l b n d', l=z.shape[-1])
+                print(adapted_cond.shape)
+
+                cond_list = [{'c_crossattn': [e]} for e in adapted_cond]
+
+            else:
+                cond_list = [cond for i in range(z.shape[-1])]  # Todo make this more efficient
+
+            # apply model by loop over crops
+            output_list = [self.model(z_list[i], t, **cond_list[i]) for i in range(z.shape[-1])]
+            assert not isinstance(output_list[0],
+                                  tuple)  # todo cant deal with multiple model outputs check this never happens
+
+            o = torch.stack(output_list, axis=-1)
+            o = o * weighting
+            # Reverse reshape to img shape
+            o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+            # stitch crops together
+            x_recon = fold(o) / normalization
+
+        else:
+            x_recon = self.model(x_noisy, t, **cond)
+
+        if isinstance(x_recon, tuple) and not return_ids:
+            return x_recon[0]
+        else:
+            return x_recon
+
+    def _predict_eps_from_xstart(self, x_t, t, pred_xstart):
+        return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \
+               extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
+
+    def _prior_bpd(self, x_start):
+        """
+        Get the prior KL term for the variational lower-bound, measured in
+        bits-per-dim.
+        This term can't be optimized, as it only depends on the encoder.
+        :param x_start: the [N x C x ...] tensor of inputs.
+        :return: a batch of [N] KL values (in bits), one per batch element.
+        """
+        batch_size = x_start.shape[0]
+        t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device)
+        qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t)
+        kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0)
+        return mean_flat(kl_prior) / np.log(2.0)
+
+    def p_losses(self, x_start, cond, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+        model_output = self.apply_model(x_noisy, t, cond)
+
+        loss_dict = {}
+        prefix = 'train' if self.training else 'val'
+
+        if self.parameterization == "x0":
+            target = x_start
+        elif self.parameterization == "eps":
+            target = noise
+        else:
+            raise NotImplementedError()
+
+        loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3])
+        loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()})
+
+        logvar_t = self.logvar[t].to(self.device)
+        loss = loss_simple / torch.exp(logvar_t) + logvar_t
+        # loss = loss_simple / torch.exp(self.logvar) + self.logvar
+        if self.learn_logvar:
+            loss_dict.update({f'{prefix}/loss_gamma': loss.mean()})
+            loss_dict.update({'logvar': self.logvar.data.mean()})
+
+        loss = self.l_simple_weight * loss.mean()
+
+        loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3))
+        loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean()
+        loss_dict.update({f'{prefix}/loss_vlb': loss_vlb})
+        loss += (self.original_elbo_weight * loss_vlb)
+        loss_dict.update({f'{prefix}/loss': loss})
+
+        return loss, loss_dict
+
+    def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False,
+                        return_x0=False, score_corrector=None, corrector_kwargs=None):
+        t_in = t
+        model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids)
+
+        if score_corrector is not None:
+            assert self.parameterization == "eps"
+            model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs)
+
+        if return_codebook_ids:
+            model_out, logits = model_out
+
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        else:
+            raise NotImplementedError()
+
+        if clip_denoised:
+            x_recon.clamp_(-1., 1.)
+        if quantize_denoised:
+            x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon)
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+        if return_codebook_ids:
+            return model_mean, posterior_variance, posterior_log_variance, logits
+        elif return_x0:
+            return model_mean, posterior_variance, posterior_log_variance, x_recon
+        else:
+            return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False,
+                 return_codebook_ids=False, quantize_denoised=False, return_x0=False,
+                 temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None):
+        b, *_, device = *x.shape, x.device
+        outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised,
+                                       return_codebook_ids=return_codebook_ids,
+                                       quantize_denoised=quantize_denoised,
+                                       return_x0=return_x0,
+                                       score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
+        if return_codebook_ids:
+            raise DeprecationWarning("Support dropped.")
+            model_mean, _, model_log_variance, logits = outputs
+        elif return_x0:
+            model_mean, _, model_log_variance, x0 = outputs
+        else:
+            model_mean, _, model_log_variance = outputs
+
+        noise = noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        # no noise when t == 0
+        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
+
+        if return_codebook_ids:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1)
+        if return_x0:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0
+        else:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False,
+                              img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0.,
+                              score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None,
+                              log_every_t=None):
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        timesteps = self.num_timesteps
+        if batch_size is not None:
+            b = batch_size if batch_size is not None else shape[0]
+            shape = [batch_size] + list(shape)
+        else:
+            b = batch_size = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=self.device)
+        else:
+            img = x_T
+        intermediates = []
+        if cond is not None:
+            if isinstance(cond, dict):
+                cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
+                list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
+            else:
+                cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
+
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+        iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation',
+                        total=timesteps) if verbose else reversed(
+            range(0, timesteps))
+        if type(temperature) == float:
+            temperature = [temperature] * timesteps
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=self.device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != 'hybrid'
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img, x0_partial = self.p_sample(img, cond, ts,
+                                            clip_denoised=self.clip_denoised,
+                                            quantize_denoised=quantize_denoised, return_x0=True,
+                                            temperature=temperature[i], noise_dropout=noise_dropout,
+                                            score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1. - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(x0_partial)
+            if callback: callback(i)
+            if img_callback: img_callback(img, i)
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_loop(self, cond, shape, return_intermediates=False,
+                      x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, start_T=None,
+                      log_every_t=None):
+
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        device = self.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        intermediates = [img]
+        if timesteps is None:
+            timesteps = self.num_timesteps
+
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+        iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(
+            range(0, timesteps))
+
+        if mask is not None:
+            assert x0 is not None
+            assert x0.shape[2:3] == mask.shape[2:3]  # spatial size has to match
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != 'hybrid'
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img = self.p_sample(img, cond, ts,
+                                clip_denoised=self.clip_denoised,
+                                quantize_denoised=quantize_denoised)
+            if mask is not None:
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1. - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(img)
+            if callback: callback(i)
+            if img_callback: img_callback(img, i)
+
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+    @torch.no_grad()
+    def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None,
+               verbose=True, timesteps=None, quantize_denoised=False,
+               mask=None, x0=None, shape=None,**kwargs):
+        if shape is None:
+            shape = (batch_size, self.channels, self.image_size, self.image_size)
+        if cond is not None:
+            if isinstance(cond, dict):
+                cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
+                list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
+            else:
+                cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
+        return self.p_sample_loop(cond,
+                                  shape,
+                                  return_intermediates=return_intermediates, x_T=x_T,
+                                  verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised,
+                                  mask=mask, x0=x0)
+
+    @torch.no_grad()
+    def sample_log(self,cond,batch_size,ddim, ddim_steps,**kwargs):
+
+        if ddim:
+            ddim_sampler = DDIMSampler(self)
+            shape = (self.channels, self.image_size, self.image_size)
+            samples, intermediates =ddim_sampler.sample(ddim_steps,batch_size,
+                                                        shape,cond,verbose=False,**kwargs)
+
+        else:
+            samples, intermediates = self.sample(cond=cond, batch_size=batch_size,
+                                                 return_intermediates=True,**kwargs)
+
+        return samples, intermediates
+
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
+                   quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
+                   plot_diffusion_rows=True, **kwargs):
+
+        use_ddim = ddim_steps is not None
+
+        log = dict()
+        z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key,
+                                           return_first_stage_outputs=True,
+                                           force_c_encode=True,
+                                           return_original_cond=True,
+                                           bs=N)
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        log["inputs"] = x
+        log["reconstruction"] = xrec
+        if self.model.conditioning_key is not None:
+            if hasattr(self.cond_stage_model, "decode"):
+                xc = self.cond_stage_model.decode(c)
+                log["conditioning"] = xc
+            elif self.cond_stage_key in ["caption"]:
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["caption"])
+                log["conditioning"] = xc
+            elif self.cond_stage_key == 'class_label':
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
+                log['conditioning'] = xc
+            elif isimage(xc):
+                log["conditioning"] = xc
+            if ismap(xc):
+                log["original_conditioning"] = self.to_rgb(xc)
+
+        if plot_diffusion_rows:
+            # get diffusion row
+            diffusion_row = list()
+            z_start = z[:n_row]
+            for t in range(self.num_timesteps):
+                if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
+                    t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
+                    t = t.to(self.device).long()
+                    noise = torch.randn_like(z_start)
+                    z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
+                    diffusion_row.append(self.decode_first_stage(z_noisy))
+
+            diffusion_row = torch.stack(diffusion_row)  # n_log_step, n_row, C, H, W
+            diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
+            diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
+            diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
+            log["diffusion_row"] = diffusion_grid
+
+        if sample:
+            # get denoise row
+            with self.ema_scope("Plotting"):
+                samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
+                                                         ddim_steps=ddim_steps,eta=ddim_eta)
+                # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
+            x_samples = self.decode_first_stage(samples)
+            log["samples"] = x_samples
+            if plot_denoise_rows:
+                denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
+                log["denoise_row"] = denoise_grid
+
+            if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance(
+                    self.first_stage_model, IdentityFirstStage):
+                # also display when quantizing x0 while sampling
+                with self.ema_scope("Plotting Quantized Denoised"):
+                    samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
+                                                             ddim_steps=ddim_steps,eta=ddim_eta,
+                                                             quantize_denoised=True)
+                    # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True,
+                    #                                      quantize_denoised=True)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_x0_quantized"] = x_samples
+
+            if inpaint:
+                # make a simple center square
+                b, h, w = z.shape[0], z.shape[2], z.shape[3]
+                mask = torch.ones(N, h, w).to(self.device)
+                # zeros will be filled in
+                mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0.
+                mask = mask[:, None, ...]
+                with self.ema_scope("Plotting Inpaint"):
+
+                    samples, _ = self.sample_log(cond=c,batch_size=N,ddim=use_ddim, eta=ddim_eta,
+                                                ddim_steps=ddim_steps, x0=z[:N], mask=mask)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_inpainting"] = x_samples
+                log["mask"] = mask
+
+                # outpaint
+                with self.ema_scope("Plotting Outpaint"):
+                    samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,eta=ddim_eta,
+                                                ddim_steps=ddim_steps, x0=z[:N], mask=mask)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_outpainting"] = x_samples
+
+        if plot_progressive_rows:
+            with self.ema_scope("Plotting Progressives"):
+                img, progressives = self.progressive_denoising(c,
+                                                               shape=(self.channels, self.image_size, self.image_size),
+                                                               batch_size=N)
+            prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation")
+            log["progressive_row"] = prog_row
+
+        if return_keys:
+            if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
+                return log
+            else:
+                return {key: log[key] for key in return_keys}
+        return log
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        params = list(self.model.parameters())
+        if self.cond_stage_trainable:
+            print(f"{self.__class__.__name__}: Also optimizing conditioner params!")
+            params = params + list(self.cond_stage_model.parameters())
+        if self.learn_logvar:
+            print('Diffusion model optimizing logvar')
+            params.append(self.logvar)
+        opt = torch.optim.AdamW(params, lr=lr)
+        if self.use_scheduler:
+            assert 'target' in self.scheduler_config
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                }]
+            return [opt], scheduler
+        return opt
+
+    @torch.no_grad()
+    def to_rgb(self, x):
+        x = x.float()
+        if not hasattr(self, "colorize"):
+            self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x)
+        x = nn.functional.conv2d(x, weight=self.colorize)
+        x = 2. * (x - x.min()) / (x.max() - x.min()) - 1.
+        return x
+
+
+class DiffusionWrapper(pl.LightningModule):
+    def __init__(self, diff_model_config, conditioning_key):
+        super().__init__()
+        self.diffusion_model = instantiate_from_config(diff_model_config)
+        self.conditioning_key = conditioning_key # 'crossattn' for txt2image, concat for inpainting
+        assert self.conditioning_key in [None, 'concat', 'crossattn', 'hybrid', 'adm']
+
+    def forward(self, x, t, c_concat: list = None, c_crossattn: list = None):
+        """param x: tensor with shape:[B,C,mel_len,T]"""
+        if self.conditioning_key is None:
+            out = self.diffusion_model(x, t)
+        elif self.conditioning_key == 'concat':
+            xc = torch.cat([x] + c_concat, dim=1)# channel dim,x shape (b,3,64,64) c_concat shape(b,4,64,64)
+            out = self.diffusion_model(xc, t)
+        elif self.conditioning_key == 'crossattn':
+            cc = torch.cat(c_crossattn, 1)# [b,seq_len,dim]
+            out = self.diffusion_model(x, t, context=cc)
+        elif self.conditioning_key == 'hybrid':# not implemented in the LatentDiffusion
+            xc = torch.cat([x] + c_concat, dim=1)
+            cc = torch.cat(c_crossattn, 1)
+            out = self.diffusion_model(xc, t, context=cc)
+        elif self.conditioning_key == 'adm':
+            cc = c_crossattn[0]
+            out = self.diffusion_model(x, t, y=cc)
+        else:
+            raise NotImplementedError()
+
+        return out
+
+
+class Layout2ImgDiffusion(LatentDiffusion):
+    # TODO: move all layout-specific hacks to this class
+    def __init__(self, cond_stage_key, *args, **kwargs):
+        assert cond_stage_key == 'coordinates_bbox', 'Layout2ImgDiffusion only for cond_stage_key="coordinates_bbox"'
+        super().__init__(cond_stage_key=cond_stage_key, *args, **kwargs)
+
+    def log_images(self, batch, N=8, *args, **kwargs):
+        logs = super().log_images(batch=batch, N=N, *args, **kwargs)
+
+        key = 'train' if self.training else 'validation'
+        dset = self.trainer.datamodule.datasets[key]
+        mapper = dset.conditional_builders[self.cond_stage_key]
+
+        bbox_imgs = []
+        map_fn = lambda catno: dset.get_textual_label(dset.get_category_id(catno))
+        for tknzd_bbox in batch[self.cond_stage_key][:N]:
+            bboximg = mapper.plot(tknzd_bbox.detach().cpu(), map_fn, (256, 256))
+            bbox_imgs.append(bboximg)
+
+        cond_img = torch.stack(bbox_imgs, dim=0)
+        logs['bbox_image'] = cond_img
+        return logs

ldm/models/diffusion/ddpm_audio.py

@@ -0,0 +1,1262 @@
+"""
+wild mixture of
+https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
+https://github.com/CompVis/taming-transformers
+-- merci
+"""
+import os
+import torch
+import torch.nn as nn
+import numpy as np
+import pytorch_lightning as pl
+from torch.optim.lr_scheduler import LambdaLR
+from einops import rearrange, repeat
+from contextlib import contextmanager
+from functools import partial
+from tqdm import tqdm
+from torchvision.utils import make_grid
+from pytorch_lightning.utilities.distributed import rank_zero_only
+
+from ldm.util import log_txt_as_img, exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config
+from ldm.modules.ema import LitEma
+from ldm.modules.distributions.distributions import normal_kl, DiagonalGaussianDistribution
+from ldm.models.autoencoder import VQModelInterface, IdentityFirstStage, AutoencoderKL
+from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like
+from ldm.models.diffusion.ddim import DDIMSampler
+from ldm.models.diffusion.ddpm import DDPM, disabled_train
+from omegaconf import ListConfig
+
+__conditioning_keys__ = {'concat': 'c_concat',
+                         'crossattn': 'c_crossattn',
+                         'adm': 'y'}
+
+
+class LatentDiffusion_audio(DDPM):
+    """main class"""
+    def __init__(self,
+                 first_stage_config,
+                 cond_stage_config,
+                 num_timesteps_cond=None,
+                 mel_dim=80,
+                 mel_length=848,
+                 cond_stage_key="image",
+                 cond_stage_trainable=False,
+                 concat_mode=True,
+                 cond_stage_forward=None,
+                 conditioning_key=None,
+                 scale_factor=1.0,
+                 scale_by_std=False,
+                 *args, **kwargs):
+        self.num_timesteps_cond = default(num_timesteps_cond, 1)
+        self.scale_by_std = scale_by_std
+        assert self.num_timesteps_cond <= kwargs['timesteps']
+        # for backwards compatibility after implementation of DiffusionWrapper
+        if conditioning_key is None:
+            conditioning_key = 'concat' if concat_mode else 'crossattn'
+        if cond_stage_config == '__is_unconditional__':
+            conditioning_key = None
+        ckpt_path = kwargs.pop("ckpt_path", None)
+        ignore_keys = kwargs.pop("ignore_keys", [])
+        super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
+        self.concat_mode = concat_mode
+        self.mel_dim = mel_dim
+        self.mel_length = mel_length
+        self.cond_stage_trainable = cond_stage_trainable
+        self.cond_stage_key = cond_stage_key
+        try:
+            self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
+        except:
+            self.num_downs = 0
+        if not scale_by_std:
+            self.scale_factor = scale_factor
+        else:
+            self.register_buffer('scale_factor', torch.tensor(scale_factor))
+        self.instantiate_first_stage(first_stage_config)
+        self.instantiate_cond_stage(cond_stage_config)
+        self.cond_stage_forward = cond_stage_forward
+        self.clip_denoised = False
+        self.bbox_tokenizer = None  
+
+        self.restarted_from_ckpt = False
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys)
+            self.restarted_from_ckpt = True
+
+    def make_cond_schedule(self, ):
+        self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
+        ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
+        self.cond_ids[:self.num_timesteps_cond] = ids
+
+    @rank_zero_only
+    @torch.no_grad()
+    def on_train_batch_start(self, batch, batch_idx, dataloader_idx):
+        # only for very first batch
+        if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt:
+            assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
+            # set rescale weight to 1./std of encodings
+            print("### USING STD-RESCALING ###")
+            x = super().get_input(batch, self.first_stage_key)
+            x = x.to(self.device)
+            encoder_posterior = self.encode_first_stage(x)
+            z = self.get_first_stage_encoding(encoder_posterior).detach()
+            del self.scale_factor
+            self.register_buffer('scale_factor', 1. / z.flatten().std())
+            print(f"setting self.scale_factor to {self.scale_factor}")
+            print("### USING STD-RESCALING ###")
+
+    def register_schedule(self,
+                          given_betas=None, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s)
+
+        self.shorten_cond_schedule = self.num_timesteps_cond > 1
+        if self.shorten_cond_schedule:
+            self.make_cond_schedule()
+
+    def instantiate_first_stage(self, config):
+        model = instantiate_from_config(config)
+        self.first_stage_model = model.eval()
+        self.first_stage_model.train = disabled_train
+        for param in self.first_stage_model.parameters():
+            param.requires_grad = False
+
+    def instantiate_cond_stage(self, config):
+        if not self.cond_stage_trainable:
+            if config == "__is_first_stage__":
+                print("Using first stage also as cond stage.")
+                self.cond_stage_model = self.first_stage_model
+            elif config == "__is_unconditional__":
+                print(f"Training {self.__class__.__name__} as an unconditional model.")
+                self.cond_stage_model = None
+                # self.be_unconditional = True
+            else:
+                model = instantiate_from_config(config)
+                self.cond_stage_model = model.eval()
+                self.cond_stage_model.train = disabled_train
+                for param in self.cond_stage_model.parameters():
+                    param.requires_grad = False
+        else:
+            assert config != '__is_first_stage__'
+            assert config != '__is_unconditional__'
+            model = instantiate_from_config(config)
+            self.cond_stage_model = model
+
+    def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False):
+        denoise_row = []
+        for zd in tqdm(samples, desc=desc):
+            denoise_row.append(self.decode_first_stage(zd.to(self.device),
+                                                            force_not_quantize=force_no_decoder_quantization))
+        n_imgs_per_row = len(denoise_row)
+        denoise_row = torch.stack(denoise_row)  # n_log_step, n_row, C, H, W
+        denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
+        denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
+        denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
+        return denoise_grid
+
+    def get_first_stage_encoding(self, encoder_posterior):
+        if isinstance(encoder_posterior, DiagonalGaussianDistribution):
+            z = encoder_posterior.sample()
+        elif isinstance(encoder_posterior, torch.Tensor):
+            z = encoder_posterior
+        else:
+            raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
+        return self.scale_factor * z
+
+    def get_learned_conditioning(self, c):
+        if self.cond_stage_forward is None:
+            if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
+                c = self.cond_stage_model.encode(c)
+                if isinstance(c, DiagonalGaussianDistribution):
+                    c = c.mode()
+            else:
+                c = self.cond_stage_model(c)
+        else:
+            assert hasattr(self.cond_stage_model, self.cond_stage_forward)
+            c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
+        return c
+
+
+    @torch.no_grad()
+    def get_unconditional_conditioning(self, batch_size, null_label=None):
+        if null_label is not None:
+            xc = null_label
+            if isinstance(xc, ListConfig):
+                xc = list(xc)
+            if isinstance(xc, dict) or isinstance(xc, list):
+                c = self.get_learned_conditioning(xc)
+            else:
+                if hasattr(xc, "to"):
+                    xc = xc.to(self.device)
+                c = self.get_learned_conditioning(xc)
+        else:
+            if self.cond_stage_key in ["class_label", "cls"]:
+                xc = self.cond_stage_model.get_unconditional_conditioning(batch_size, device=self.device)
+                return self.get_learned_conditioning(xc)
+            else:
+                raise NotImplementedError("todo")
+        if isinstance(c, list):  # in case the encoder gives us a list
+            for i in range(len(c)):
+                c[i] = repeat(c[i], '1 ... -> b ...', b=batch_size).to(self.device)
+        else:
+            c = repeat(c, '1 ... -> b ...', b=batch_size).to(self.device)
+        return c
+
+    def meshgrid(self, h, w):
+        y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1)
+        x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1)
+
+        arr = torch.cat([y, x], dim=-1)
+        return arr
+
+    def delta_border(self, h, w):
+        """
+        :param h: height
+        :param w: width
+        :return: normalized distance to image border,
+         wtith min distance = 0 at border and max dist = 0.5 at image center
+        """
+        lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2)
+        arr = self.meshgrid(h, w) / lower_right_corner
+        dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0]
+        dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0]
+        edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0]
+        return edge_dist
+
+    def get_weighting(self, h, w, Ly, Lx, device):
+        weighting = self.delta_border(h, w)
+        weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"],
+                               self.split_input_params["clip_max_weight"], )
+        weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device)
+
+        if self.split_input_params["tie_braker"]:
+            L_weighting = self.delta_border(Ly, Lx)
+            L_weighting = torch.clip(L_weighting,
+                                     self.split_input_params["clip_min_tie_weight"],
+                                     self.split_input_params["clip_max_tie_weight"])
+
+            L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device)
+            weighting = weighting * L_weighting
+        return weighting
+
+    def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1):  # todo load once not every time, shorten code
+        """
+        :param x: img of size (bs, c, h, w)
+        :return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1])
+        """
+        bs, nc, h, w = x.shape
+
+        # number of crops in image
+        Ly = (h - kernel_size[0]) // stride[0] + 1
+        Lx = (w - kernel_size[1]) // stride[1] + 1
+
+        if uf == 1 and df == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params)
+
+            weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h, w)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx))
+
+        elif uf > 1 and df == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf),
+                                dilation=1, padding=0,
+                                stride=(stride[0] * uf, stride[1] * uf))
+            fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2)
+
+            weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h * uf, w * uf)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx))
+
+        elif df > 1 and uf == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df),
+                                dilation=1, padding=0,
+                                stride=(stride[0] // df, stride[1] // df))
+            fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2)
+
+            weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h // df, w // df)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx))
+
+        else:
+            raise NotImplementedError
+
+        return fold, unfold, normalization, weighting
+
+    @torch.no_grad()
+    def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False,
+                  cond_key=None, return_original_cond=False, bs=None):
+        x = super().get_input(batch, k)
+        if bs is not None:
+            x = x[:bs]
+        x = x.to(self.device)
+        encoder_posterior = self.encode_first_stage(x)
+        z = self.get_first_stage_encoding(encoder_posterior).detach()
+
+        if self.model.conditioning_key is not None:
+            if cond_key is None:
+                cond_key = self.cond_stage_key
+            if cond_key != self.first_stage_key:
+                if cond_key in ['caption', 'coordinates_bbox']:
+                    xc = batch[cond_key]
+                elif cond_key == 'class_label':
+                    xc = batch
+                else:
+                    xc = super().get_input(batch, cond_key).to(self.device)
+            else:
+                xc = x
+            if not self.cond_stage_trainable or force_c_encode:
+                if isinstance(xc, dict) or isinstance(xc, list):
+                    # import pudb; pudb.set_trace()
+                    c = self.get_learned_conditioning(xc)
+                else:
+                    c = self.get_learned_conditioning(xc.to(self.device))
+            else:
+                c = xc
+            if bs is not None:
+                c = c[:bs]
+            # Testing #
+            if cond_key == 'masked_image':
+                mask = super().get_input(batch, "mask")
+                cc = torch.nn.functional.interpolate(mask, size=c.shape[-2:]) # [B, 1, 10, 106]
+                c = torch.cat((c, cc), dim=1) # [B, 5, 10, 106]
+            # Testing #
+            if self.use_positional_encodings:
+                pos_x, pos_y = self.compute_latent_shifts(batch)
+                ckey = __conditioning_keys__[self.model.conditioning_key]
+                c = {ckey: c, 'pos_x': pos_x, 'pos_y': pos_y}
+
+        else:
+            c = None
+            xc = None
+            if self.use_positional_encodings:
+                pos_x, pos_y = self.compute_latent_shifts(batch)
+                c = {'pos_x': pos_x, 'pos_y': pos_y}
+        out = [z, c]
+        if return_first_stage_outputs:
+            xrec = self.decode_first_stage(z)
+            out.extend([x, xrec])
+        if return_original_cond:
+            out.append(xc)
+        return out
+
+    @torch.no_grad()
+    def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
+        if predict_cids:
+            if z.dim() == 4:
+                z = torch.argmax(z.exp(), dim=1).long()
+            z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
+            z = rearrange(z, 'b h w c -> b c h w').contiguous()
+
+        z = 1. / self.scale_factor * z
+
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                uf = self.split_input_params["vqf"]
+                bs, nc, h, w = z.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
+
+                z = unfold(z)  # (bn, nc * prod(**ks), L)
+                # 1. Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                # 2. apply model loop over last dim
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
+                                                                 force_not_quantize=predict_cids or force_not_quantize)
+                                   for i in range(z.shape[-1])]
+                else:
+
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
+                                   for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)  # # (bn, nc, ks[0], ks[1], L)
+                o = o * weighting
+                # Reverse 1. reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization  # norm is shape (1, 1, h, w)
+                return decoded
+            else:
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+                else:
+                    return self.first_stage_model.decode(z)
+
+        else:
+            if isinstance(self.first_stage_model, VQModelInterface):
+                return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+            else:
+                return self.first_stage_model.decode(z)
+
+    # same as above but without decorator
+    def differentiable_decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
+        if predict_cids:
+            if z.dim() == 4:
+                z = torch.argmax(z.exp(), dim=1).long()
+            z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
+            z = rearrange(z, 'b h w c -> b c h w').contiguous()
+
+        z = 1. / self.scale_factor * z
+
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                uf = self.split_input_params["vqf"]
+                bs, nc, h, w = z.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
+
+                z = unfold(z)  # (bn, nc * prod(**ks), L)
+                # 1. Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                # 2. apply model loop over last dim
+                if isinstance(self.first_stage_model, VQModelInterface):  
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
+                                                                 force_not_quantize=predict_cids or force_not_quantize)
+                                   for i in range(z.shape[-1])]
+                else:
+
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
+                                   for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)  # # (bn, nc, ks[0], ks[1], L)
+                o = o * weighting
+                # Reverse 1. reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization  # norm is shape (1, 1, h, w)
+                return decoded
+            else:
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+                else:
+                    return self.first_stage_model.decode(z)
+
+        else:
+            if isinstance(self.first_stage_model, VQModelInterface):
+                return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+            else:
+                return self.first_stage_model.decode(z)
+
+    @torch.no_grad()
+    def encode_first_stage(self, x):
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                df = self.split_input_params["vqf"]
+                self.split_input_params['original_image_size'] = x.shape[-2:]
+                bs, nc, h, w = x.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(x, ks, stride, df=df)
+                z = unfold(x)  # (bn, nc * prod(**ks), L)
+                # Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                output_list = [self.first_stage_model.encode(z[:, :, :, :, i])
+                               for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)
+                o = o * weighting
+
+                # Reverse reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization
+                return decoded
+
+            else:
+                return self.first_stage_model.encode(x)
+        else:
+            return self.first_stage_model.encode(x)
+
+    def shared_step(self, batch, **kwargs):
+        x, c = self.get_input(batch, self.first_stage_key)
+        loss = self(x, c)
+        return loss
+
+    def test_step(self,batch,batch_idx):
+        cond = batch[self.cond_stage_key] * self.test_repeat
+        cond = self.get_learned_conditioning(cond) # c: string -> [B, T, Context_dim]
+        batch_size = len(cond)
+        enc_emb = self.sample(cond,batch_size,timesteps=self.test_numsteps)# shape = [batch_size,self.channels,self.mel_dim,self.mel_length]
+        xrec = self.decode_first_stage(enc_emb)
+        reconstructions = (xrec + 1)/2 # to mel scale
+        test_ckpt_path = os.path.basename(self.trainer.tested_ckpt_path)
+        savedir = os.path.join(self.trainer.log_dir,f'output_imgs_{test_ckpt_path}','fake_class')
+        if not os.path.exists(savedir):
+            os.makedirs(savedir)
+
+        file_names = batch['f_name']
+        nfiles = len(file_names)
+        reconstructions = reconstructions.cpu().numpy().squeeze(1) # squuze channel dim
+        for k in range(reconstructions.shape[0]):
+            b,repeat = k % nfiles, k // nfiles
+            vname_num_split_index = file_names[b].rfind('_')# file_names[b]:video_name+'_'+num
+            v_n,num = file_names[b][:vname_num_split_index],file_names[b][vname_num_split_index+1:]
+            save_img_path = os.path.join(savedir,f'{v_n}_sample_{num}_{repeat}.npy')# the num_th caption, the repeat_th repitition
+            np.save(save_img_path,reconstructions[b])
+        
+        return None
+
+    def forward(self, x, c, *args, **kwargs):
+        t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
+        if self.model.conditioning_key is not None:
+            assert c is not None
+            if self.cond_stage_trainable:
+                c = self.get_learned_conditioning(c) # c: string -> [B, T, Context_dim]
+            if self.shorten_cond_schedule:  # TODO: drop this option
+                tc = self.cond_ids[t].to(self.device)
+                c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float()))
+        return self.p_losses(x, c, t, *args, **kwargs)
+
+    def _rescale_annotations(self, bboxes, crop_coordinates):  # TODO: move to dataset
+        def rescale_bbox(bbox):
+            x0 = clamp((bbox[0] - crop_coordinates[0]) / crop_coordinates[2])
+            y0 = clamp((bbox[1] - crop_coordinates[1]) / crop_coordinates[3])
+            w = min(bbox[2] / crop_coordinates[2], 1 - x0)
+            h = min(bbox[3] / crop_coordinates[3], 1 - y0)
+            return x0, y0, w, h
+
+        return [rescale_bbox(b) for b in bboxes]
+
+    def apply_model(self, x_noisy, t, cond, return_ids=False):
+
+        if isinstance(cond, dict):
+            # hybrid case, cond is exptected to be a dict
+            pass
+        else:
+            if not isinstance(cond, list):
+                cond = [cond]
+            key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn'
+            cond = {key: cond}
+
+        if hasattr(self, "split_input_params"):
+            assert len(cond) == 1  # todo can only deal with one conditioning atm
+            assert not return_ids  
+            ks = self.split_input_params["ks"]  # eg. (128, 128)
+            stride = self.split_input_params["stride"]  # eg. (64, 64)
+
+            h, w = x_noisy.shape[-2:]
+
+            fold, unfold, normalization, weighting = self.get_fold_unfold(x_noisy, ks, stride)
+
+            z = unfold(x_noisy)  # (bn, nc * prod(**ks), L)
+            # Reshape to img shape
+            z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+            z_list = [z[:, :, :, :, i] for i in range(z.shape[-1])]
+
+            if self.cond_stage_key in ["image", "LR_image", "segmentation",
+                                       'bbox_img'] and self.model.conditioning_key:  # todo check for completeness
+                c_key = next(iter(cond.keys()))  # get key
+                c = next(iter(cond.values()))  # get value
+                assert (len(c) == 1)  # todo extend to list with more than one elem
+                c = c[0]  # get element
+
+                c = unfold(c)
+                c = c.view((c.shape[0], -1, ks[0], ks[1], c.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                cond_list = [{c_key: [c[:, :, :, :, i]]} for i in range(c.shape[-1])]
+
+            elif self.cond_stage_key == 'coordinates_bbox':
+                assert 'original_image_size' in self.split_input_params, 'BoudingBoxRescaling is missing original_image_size'
+
+                # assuming padding of unfold is always 0 and its dilation is always 1
+                n_patches_per_row = int((w - ks[0]) / stride[0] + 1)
+                full_img_h, full_img_w = self.split_input_params['original_image_size']
+                # as we are operating on latents, we need the factor from the original image size to the
+                # spatial latent size to properly rescale the crops for regenerating the bbox annotations
+                num_downs = self.first_stage_model.encoder.num_resolutions - 1
+                rescale_latent = 2 ** (num_downs)
+
+                # get top left postions of patches as conforming for the bbbox tokenizer, therefore we
+                # need to rescale the tl patch coordinates to be in between (0,1)
+                tl_patch_coordinates = [(rescale_latent * stride[0] * (patch_nr % n_patches_per_row) / full_img_w,
+                                         rescale_latent * stride[1] * (patch_nr // n_patches_per_row) / full_img_h)
+                                        for patch_nr in range(z.shape[-1])]
+
+                # patch_limits are tl_coord, width and height coordinates as (x_tl, y_tl, h, w)
+                patch_limits = [(x_tl, y_tl,
+                                 rescale_latent * ks[0] / full_img_w,
+                                 rescale_latent * ks[1] / full_img_h) for x_tl, y_tl in tl_patch_coordinates]
+                # patch_values = [(np.arange(x_tl,min(x_tl+ks, 1.)),np.arange(y_tl,min(y_tl+ks, 1.))) for x_tl, y_tl in tl_patch_coordinates]
+
+                # tokenize crop coordinates for the bounding boxes of the respective patches
+                patch_limits_tknzd = [torch.LongTensor(self.bbox_tokenizer._crop_encoder(bbox))[None].to(self.device)
+                                      for bbox in patch_limits]  # list of length l with tensors of shape (1, 2)
+                print(patch_limits_tknzd[0].shape)
+                # cut tknzd crop position from conditioning
+                assert isinstance(cond, dict), 'cond must be dict to be fed into model'
+                cut_cond = cond['c_crossattn'][0][..., :-2].to(self.device)
+                print(cut_cond.shape)
+
+                adapted_cond = torch.stack([torch.cat([cut_cond, p], dim=1) for p in patch_limits_tknzd])
+                adapted_cond = rearrange(adapted_cond, 'l b n -> (l b) n')
+                print(adapted_cond.shape)
+                adapted_cond = self.get_learned_conditioning(adapted_cond)
+                print(adapted_cond.shape)
+                adapted_cond = rearrange(adapted_cond, '(l b) n d -> l b n d', l=z.shape[-1])
+                print(adapted_cond.shape)
+
+                cond_list = [{'c_crossattn': [e]} for e in adapted_cond]
+
+            else:
+                cond_list = [cond for i in range(z.shape[-1])]  # Todo make this more efficient
+
+            # apply model by loop over crops
+            output_list = [self.model(z_list[i], t, **cond_list[i]) for i in range(z.shape[-1])]
+            assert not isinstance(output_list[0],
+                                  tuple)  # todo cant deal with multiple model outputs check this never happens
+
+            o = torch.stack(output_list, axis=-1)
+            o = o * weighting
+            # Reverse reshape to img shape
+            o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+            # stitch crops together
+            x_recon = fold(o) / normalization
+
+        else:
+            x_recon = self.model(x_noisy, t, **cond)
+
+        if isinstance(x_recon, tuple) and not return_ids:
+            return x_recon[0]
+        else:
+            return x_recon
+
+    def _predict_eps_from_xstart(self, x_t, t, pred_xstart):
+        return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \
+               extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
+
+    def _prior_bpd(self, x_start):
+        """
+        Get the prior KL term for the variational lower-bound, measured in
+        bits-per-dim.
+        This term can't be optimized, as it only depends on the encoder.
+        :param x_start: the [N x C x ...] tensor of inputs.
+        :return: a batch of [N] KL values (in bits), one per batch element.
+        """
+        batch_size = x_start.shape[0]
+        t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device)
+        qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t)
+        kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0)
+        return mean_flat(kl_prior) / np.log(2.0)
+
+    def p_losses(self, x_start, cond, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+        model_output = self.apply_model(x_noisy, t, cond)
+
+        loss_dict = {}
+        prefix = 'train' if self.training else 'val'
+
+        if self.parameterization == "x0":
+            target = x_start
+        elif self.parameterization == "eps":
+            target = noise
+        else:
+            raise NotImplementedError()
+
+        loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3])
+        loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()})
+
+        logvar_t = self.logvar[t].to(self.device)
+        loss = loss_simple / torch.exp(logvar_t) + logvar_t
+        # loss = loss_simple / torch.exp(self.logvar) + self.logvar
+        if self.learn_logvar:
+            loss_dict.update({f'{prefix}/loss_gamma': loss.mean()})
+            loss_dict.update({'logvar': self.logvar.data.mean()})
+
+        loss = self.l_simple_weight * loss.mean()
+
+        loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3))
+        loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean()
+        loss_dict.update({f'{prefix}/loss_vlb': loss_vlb})
+        loss += (self.original_elbo_weight * loss_vlb)
+        loss_dict.update({f'{prefix}/loss': loss})
+
+        return loss, loss_dict
+
+    def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False,
+                        return_x0=False, score_corrector=None, corrector_kwargs=None):
+        t_in = t
+        model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids)
+
+        if score_corrector is not None:
+            assert self.parameterization == "eps"
+            model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs)
+
+        if return_codebook_ids:
+            model_out, logits = model_out
+
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        else:
+            raise NotImplementedError()
+
+        if clip_denoised:
+            x_recon.clamp_(-1., 1.)
+        if quantize_denoised:
+            x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon)
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+        if return_codebook_ids:
+            return model_mean, posterior_variance, posterior_log_variance, logits
+        elif return_x0:
+            return model_mean, posterior_variance, posterior_log_variance, x_recon
+        else:
+            return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False,
+                 return_codebook_ids=False, quantize_denoised=False, return_x0=False,
+                 temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None):
+        b, *_, device = *x.shape, x.device
+        outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised,
+                                       return_codebook_ids=return_codebook_ids,
+                                       quantize_denoised=quantize_denoised,
+                                       return_x0=return_x0,
+                                       score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
+        if return_codebook_ids:
+            raise DeprecationWarning("Support dropped.")
+            model_mean, _, model_log_variance, logits = outputs
+        elif return_x0:
+            model_mean, _, model_log_variance, x0 = outputs
+        else:
+            model_mean, _, model_log_variance = outputs
+
+        noise = noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        # no noise when t == 0
+        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
+
+        if return_codebook_ids:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1)
+        if return_x0:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0
+        else:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False,
+                              img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0.,
+                              score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None,
+                              log_every_t=None):
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        timesteps = self.num_timesteps
+        if batch_size is not None:
+            b = batch_size if batch_size is not None else shape[0]
+            shape = [batch_size] + list(shape)
+        else:
+            b = batch_size = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=self.device)
+        else:
+            img = x_T
+        intermediates = []
+        if cond is not None:
+            if isinstance(cond, dict):
+                cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
+                list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
+            else:
+                cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
+
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+        iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation',
+                        total=timesteps) if verbose else reversed(
+            range(0, timesteps))
+        if type(temperature) == float:
+            temperature = [temperature] * timesteps
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=self.device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != 'hybrid'
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img, x0_partial = self.p_sample(img, cond, ts,
+                                            clip_denoised=self.clip_denoised,
+                                            quantize_denoised=quantize_denoised, return_x0=True,
+                                            temperature=temperature[i], noise_dropout=noise_dropout,
+                                            score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1. - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(x0_partial)
+            if callback: callback(i)
+            if img_callback: img_callback(img, i)
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_loop(self, cond, shape, return_intermediates=False,
+                      x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, start_T=None,
+                      log_every_t=None):
+
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        device = self.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        intermediates = [img]
+        if timesteps is None:
+            timesteps = self.num_timesteps
+
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+        iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(
+            range(0, timesteps))
+
+        if mask is not None:
+            assert x0 is not None
+            assert x0.shape[2:3] == mask.shape[2:3]  # spatial size has to match
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != 'hybrid'
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img = self.p_sample(img, cond, ts,
+                                clip_denoised=self.clip_denoised,
+                                quantize_denoised=quantize_denoised)
+            if mask is not None:
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1. - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(img)
+            if callback: callback(i)
+            if img_callback: img_callback(img, i)
+
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+    @torch.no_grad()
+    def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None,
+               verbose=True, timesteps=None, quantize_denoised=False,
+               mask=None, x0=None, shape=None,**kwargs):
+        if shape is None:
+            shape = (batch_size, self.channels, self.mel_dim, self.mel_length)
+        if cond is not None:
+            if isinstance(cond, dict):
+                cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
+                list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
+            else:
+                cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
+        return self.p_sample_loop(cond,
+                                  shape,
+                                  return_intermediates=return_intermediates, x_T=x_T,
+                                  verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised,
+                                  mask=mask, x0=x0)
+
+    @torch.no_grad()
+    def sample_log(self,cond,batch_size,ddim, ddim_steps,**kwargs):
+
+        if ddim:
+            ddim_sampler = DDIMSampler(self)
+            shape = (self.channels, self.mel_dim, self.mel_length)
+            samples, intermediates =ddim_sampler.sample(ddim_steps,batch_size,
+                                                        shape,cond,verbose=False,**kwargs)
+
+        else:
+            samples, intermediates = self.sample(cond=cond, batch_size=batch_size,
+                                                 return_intermediates=True,**kwargs)
+
+        return samples, intermediates
+
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
+                   quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
+                   plot_diffusion_rows=True, **kwargs):
+
+        use_ddim = ddim_steps is not None
+
+        log = dict()
+        z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key,
+                                           return_first_stage_outputs=True,
+                                           force_c_encode=True,
+                                           return_original_cond=True,
+                                           bs=N)
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        log["inputs"] = x
+        log["reconstruction"] = xrec
+        if self.model.conditioning_key is not None:
+            if hasattr(self.cond_stage_model, "decode") and self.cond_stage_key != "masked_image":
+                xc = self.cond_stage_model.decode(c)
+                log["conditioning"] = xc
+            elif self.cond_stage_key == "masked_image":
+                log["mask"] = c[:, -1, :, :][:, None, :, :]
+                xc = self.cond_stage_model.decode(c[:, :self.cond_stage_model.embed_dim, :, :])
+                log["conditioning"] = xc
+            elif self.cond_stage_key in ["caption"]:
+                xc = log_txt_as_img((256, 256), batch["caption"])
+                log["conditioning"] = xc
+            elif self.cond_stage_key == 'class_label':
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
+                log['conditioning'] = xc
+            elif isimage(xc):
+                log["conditioning"] = xc
+            if ismap(xc):
+                log["original_conditioning"] = self.to_rgb(xc)
+
+        if plot_diffusion_rows:
+            # get diffusion row
+            diffusion_row = list()
+            z_start = z[:n_row]
+            for t in range(self.num_timesteps):
+                if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
+                    t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
+                    t = t.to(self.device).long()
+                    noise = torch.randn_like(z_start)
+                    z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
+                    diffusion_row.append(self.decode_first_stage(z_noisy))
+
+            diffusion_row = torch.stack(diffusion_row)  # n_log_step, n_row, C, H, W
+            diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
+            diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
+            diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
+            log["diffusion_row"] = diffusion_grid
+
+        if sample:
+            # get denoise row
+            with self.ema_scope("Plotting"):
+                samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
+                                                         ddim_steps=ddim_steps,eta=ddim_eta)
+                # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
+            x_samples = self.decode_first_stage(samples)
+            log["samples"] = x_samples
+            if plot_denoise_rows:
+                denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
+                log["denoise_row"] = denoise_grid
+
+            if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance(
+                    self.first_stage_model, IdentityFirstStage):
+                # also display when quantizing x0 while sampling
+                with self.ema_scope("Plotting Quantized Denoised"):
+                    samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
+                                                             ddim_steps=ddim_steps,eta=ddim_eta,
+                                                             quantize_denoised=True)
+                    # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True,
+                    #                                      quantize_denoised=True)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_x0_quantized"] = x_samples
+
+            if inpaint:
+                # make a simple center square
+                b, h, w = z.shape[0], z.shape[2], z.shape[3]
+                mask = torch.ones(N, h, w).to(self.device)
+                # zeros will be filled in
+                mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0.
+                mask = mask[:, None, ...]
+                with self.ema_scope("Plotting Inpaint"):
+
+                    samples, _ = self.sample_log(cond=c,batch_size=N,ddim=use_ddim, eta=ddim_eta,
+                                                ddim_steps=ddim_steps, x0=z[:N], mask=mask)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_inpainting"] = x_samples
+                log["mask_inpainting"] = mask
+
+                # outpaint
+                mask = 1 - mask
+                with self.ema_scope("Plotting Outpaint"):
+                    samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,eta=ddim_eta,
+                                                ddim_steps=ddim_steps, x0=z[:N], mask=mask)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_outpainting"] = x_samples
+                log["mask_outpainting"] = mask
+
+        if plot_progressive_rows:
+            with self.ema_scope("Plotting Progressives"):
+                img, progressives = self.progressive_denoising(c,
+                                                               shape=(self.channels, self.mel_dim, self.mel_length),
+                                                               batch_size=N)
+            prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation")
+            log["progressive_row"] = prog_row
+
+        if return_keys:
+            if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
+                return log
+            else:
+                return {key: log[key] for key in return_keys}
+        return log
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        params = list(self.model.parameters())
+        if self.cond_stage_trainable:
+            print(f"{self.__class__.__name__}: Also optimizing conditioner params!")
+            params = params + list(self.cond_stage_model.parameters())
+        if self.learn_logvar:
+            print('Diffusion model optimizing logvar')
+            params.append(self.logvar)
+        opt = torch.optim.AdamW(params, lr=lr)
+        if self.use_scheduler:
+            assert 'target' in self.scheduler_config
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                }]
+            return [opt], scheduler
+        return opt
+
+    @torch.no_grad()
+    def to_rgb(self, x):
+        x = x.float()
+        if not hasattr(self, "colorize"):
+            self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x)
+        x = nn.functional.conv2d(x, weight=self.colorize)
+        x = 2. * (x - x.min()) / (x.max() - x.min()) - 1.
+        return x
+
+
+class LatentFinetuneDiffusion(LatentDiffusion_audio):
+    """
+         Basis for different finetunas, such as inpainting or depth2image
+         To disable finetuning mode, set finetune_keys to None
+    """
+
+    def __init__(self,
+                 concat_keys: tuple,
+                 finetune_keys=("model.diffusion_model.input_blocks.0.0.weight",
+                                "model_ema.diffusion_modelinput_blocks00weight"
+                                ),
+                 keep_finetune_dims=4,
+                 # if model was trained without concat mode before and we would like to keep these channels
+                 c_concat_log_start=None,  # to log reconstruction of c_concat codes
+                 c_concat_log_end=None,
+                 *args, **kwargs
+                 ):
+        ckpt_path = kwargs.pop("ckpt_path", None)
+        ignore_keys = kwargs.pop("ignore_keys", list())
+        super().__init__(*args, **kwargs)
+        self.finetune_keys = finetune_keys
+        self.concat_keys = concat_keys
+        self.keep_dims = keep_finetune_dims
+        self.c_concat_log_start = c_concat_log_start
+        self.c_concat_log_end = c_concat_log_end
+
+        if exists(self.finetune_keys): assert exists(ckpt_path), 'can only finetune from a given checkpoint'
+        if exists(ckpt_path):
+            self.init_from_ckpt(ckpt_path, ignore_keys)
+
+    def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
+        sd = torch.load(path, map_location="cpu")
+        if "state_dict" in list(sd.keys()):
+            sd = sd["state_dict"]
+        keys = list(sd.keys())
+
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del sd[k]
+
+            # make it explicit, finetune by including extra input channels
+            if exists(self.finetune_keys) and k in self.finetune_keys:
+                new_entry = None
+                for name, param in self.named_parameters():
+                    if name in self.finetune_keys:
+                        print(
+                            f"modifying key '{name}' and keeping its original {self.keep_dims} (channels) dimensions only")
+                        new_entry = torch.zeros_like(param)  # zero init
+                assert exists(new_entry), 'did not find matching parameter to modify'
+                new_entry[:, :self.keep_dims, ...] = sd[k]
+                sd[k] = new_entry
+
+        missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(sd, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+        if len(unexpected) > 0:
+            print(f"Unexpected Keys: {unexpected}")
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
+                   quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
+                   plot_diffusion_rows=True, unconditional_guidance_scale=1., unconditional_guidance_label=None,
+                   use_ema_scope=True,
+                   **kwargs):
+        use_ddim = ddim_steps is not None
+
+        log = dict()
+        z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key, bs=N, return_first_stage_outputs=True)
+        c_cat, c = c["c_concat"][0], c["c_crossattn"][0]
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        log["inputs"] = x
+        log["reconstruction"] = xrec
+        if self.model.conditioning_key is not None:
+            if hasattr(self.cond_stage_model, "decode"):
+                xc = self.cond_stage_model.decode(c)
+                log["conditioning"] = xc
+            elif self.cond_stage_key in ["caption"]:
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["caption"])
+                log["conditioning"] = xc
+            elif self.cond_stage_key == 'class_label':
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
+                log['conditioning'] = xc
+            elif isimage(xc):
+                log["conditioning"] = xc
+            if ismap(xc):
+                log["original_conditioning"] = self.to_rgb(xc)
+
+        if not (self.c_concat_log_start is None and self.c_concat_log_end is None):
+            log["c_concat_decoded"] = self.decode_first_stage(c_cat[:, self.c_concat_log_start:self.c_concat_log_end])
+
+        if plot_diffusion_rows:
+            # get diffusion row
+            diffusion_row = list()
+            z_start = z[:n_row]
+            for t in range(self.num_timesteps):
+                if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
+                    t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
+                    t = t.to(self.device).long()
+                    noise = torch.randn_like(z_start)
+                    z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
+                    diffusion_row.append(self.decode_first_stage(z_noisy))
+
+            diffusion_row = torch.stack(diffusion_row)  # n_log_step, n_row, C, H, W
+            diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
+            diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
+            diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
+            log["diffusion_row"] = diffusion_grid
+
+        if sample:
+            # get denoise row
+            with self.ema_scope("Sampling"):
+                samples, z_denoise_row = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]},
+                                                         batch_size=N, ddim=use_ddim,
+                                                         ddim_steps=ddim_steps, eta=ddim_eta)
+                # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
+            x_samples = self.decode_first_stage(samples)
+            log["samples"] = x_samples
+            if plot_denoise_rows:
+                denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
+                log["denoise_row"] = denoise_grid
+
+        if unconditional_guidance_scale > 1.0:
+            uc_cross = self.get_unconditional_conditioning(N, unconditional_guidance_label)
+            uc_cat = c_cat
+            uc_full = {"c_concat": [uc_cat], "c_crossattn": [uc_cross]}
+            with self.ema_scope("Sampling with classifier-free guidance"):
+                samples_cfg, _ = self.sample_log(cond={"c_concat": [c_cat], "c_crossattn": [c]},
+                                                 batch_size=N, ddim=use_ddim,
+                                                 ddim_steps=ddim_steps, eta=ddim_eta,
+                                                 unconditional_guidance_scale=unconditional_guidance_scale,
+                                                 unconditional_conditioning=uc_full,
+                                                 )
+                x_samples_cfg = self.decode_first_stage(samples_cfg)
+                log[f"samples_cfg_scale_{unconditional_guidance_scale:.2f}"] = x_samples_cfg
+
+        return log
+
+
+class LatentInpaintDiffusion(LatentFinetuneDiffusion):
+    """
+    can either run as pure inpainting model (only concat mode) or with mixed conditionings,
+    e.g. mask as concat and text via cross-attn.
+    To disable finetuning mode, set finetune_keys to None
+     """
+
+    def __init__(self,
+                 concat_keys=("mask", "masked_image"),
+                 masked_image_key="masked_image",
+                 *args, **kwargs
+                 ):
+        super().__init__(concat_keys, *args, **kwargs)
+        self.masked_image_key = masked_image_key
+        assert self.masked_image_key in concat_keys
+
+    @torch.no_grad()
+    def get_input(self, batch, k, cond_key=None, bs=None, return_first_stage_outputs=False):
+        # note: restricted to non-trainable encoders currently
+        assert not self.cond_stage_trainable, 'trainable cond stages not yet supported for inpainting'
+        z, c, x, xrec, xc = super().get_input(batch, self.first_stage_key, return_first_stage_outputs=True,
+                                              force_c_encode=True, return_original_cond=True, bs=bs)
+
+        assert exists(self.concat_keys)
+        c_cat = list()
+        for ck in self.concat_keys:
+            if len(batch[ck].shape) == 3:
+                batch[ck] = batch[ck][..., None]
+            cc = rearrange(batch[ck], 'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float()
+            if bs is not None:
+                cc = cc[:bs]
+                cc = cc.to(self.device)
+            bchw = z.shape
+            if ck != self.masked_image_key:
+                cc = torch.nn.functional.interpolate(cc, size=bchw[-2:])
+            else:
+                cc = self.get_first_stage_encoding(self.encode_first_stage(cc))
+            c_cat.append(cc)
+        c_cat = torch.cat(c_cat, dim=1)
+        all_conds = {"c_concat": [c_cat], "c_crossattn": [c]}
+        if return_first_stage_outputs:
+            return z, all_conds, x, xrec, xc
+        return z, all_conds
+
+    @torch.no_grad()
+    def log_images(self, *args, **kwargs):
+        log = super(LatentInpaintDiffusion, self).log_images(*args, **kwargs)
+        log["masked_image"] = rearrange(args[0]["masked_image"],
+                                        'b h w c -> b c h w').to(memory_format=torch.contiguous_format).float()
+        return log

ldm/models/diffusion/ddpm_audio_inpaint.py

@@ -0,0 +1,1081 @@
+"""
+wild mixture of
+https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
+https://github.com/CompVis/taming-transformers
+-- merci
+"""
+import os
+import torch
+import torch.nn as nn
+import numpy as np
+import pytorch_lightning as pl
+from torch.optim.lr_scheduler import LambdaLR
+from einops import rearrange, repeat
+from contextlib import contextmanager
+from functools import partial
+from tqdm import tqdm
+from torchvision.utils import make_grid
+from pytorch_lightning.utilities.distributed import rank_zero_only
+
+from ldm.util import log_txt_as_img, exists, default, ismap, isimage, mean_flat, count_params, instantiate_from_config
+from ldm.modules.ema import LitEma
+from ldm.modules.distributions.distributions import normal_kl, DiagonalGaussianDistribution
+from ldm.models.autoencoder import VQModelInterface, IdentityFirstStage, AutoencoderKL
+from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor, noise_like
+from ldm.models.diffusion.ddim import DDIMSampler
+from ldm.models.diffusion.ddpm import DDPM, disabled_train
+
+__conditioning_keys__ = {'concat': 'c_concat',
+                         'crossattn': 'c_crossattn',
+                         'adm': 'y'}
+
+# add mel_dim and mel_length params to ensure correct shape
+class LatentDiffusion_audioinpaint(DDPM):
+    """main class"""
+    def __init__(self,
+                 first_stage_config,
+                 cond_stage_config,
+                 num_timesteps_cond=None,
+                 mel_dim=80,
+                 mel_length=848,
+                 cond_stage_key="image",
+                 cond_stage_trainable=False,
+                 concat_mode=True,
+                 cond_stage_forward=None,
+                 conditioning_key=None,
+                 scale_factor=1.0,
+                 scale_by_std=False,
+                 test_repeat=1,
+                 test_numsteps = None,
+                 *args, **kwargs):
+        self.num_timesteps_cond = default(num_timesteps_cond, 1)
+        self.scale_by_std = scale_by_std
+        assert self.num_timesteps_cond <= kwargs['timesteps']
+        # for backwards compatibility after implementation of DiffusionWrapper
+        if conditioning_key is None:
+            conditioning_key = 'concat' if concat_mode else 'crossattn'
+        if cond_stage_config == '__is_unconditional__':
+            conditioning_key = None
+        ckpt_path = kwargs.pop("ckpt_path", None)
+        ignore_keys = kwargs.pop("ignore_keys", [])
+        super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
+        self.test_repeat = test_repeat
+        if test_numsteps == None:
+            self.test_numsteps = self.num_timesteps
+        self.concat_mode = concat_mode
+        self.mel_dim = mel_dim
+        self.mel_length = mel_length
+        self.cond_stage_trainable = cond_stage_trainable
+        self.cond_stage_key = cond_stage_key
+        try:
+            self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
+        except:
+            self.num_downs = 0
+        if not scale_by_std:
+            self.scale_factor = scale_factor
+        else:
+            self.register_buffer('scale_factor', torch.tensor(scale_factor))
+        self.instantiate_first_stage(first_stage_config)
+        self.instantiate_cond_stage(cond_stage_config)
+        self.cond_stage_forward = cond_stage_forward
+        self.clip_denoised = False
+        self.bbox_tokenizer = None  
+
+        self.restarted_from_ckpt = False
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys)
+            self.restarted_from_ckpt = True
+
+    def make_cond_schedule(self, ):
+        self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
+        ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
+        self.cond_ids[:self.num_timesteps_cond] = ids
+
+    @rank_zero_only
+    @torch.no_grad()
+    def on_train_batch_start(self, batch, batch_idx, dataloader_idx):
+        # only for very first batch
+        if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 and batch_idx == 0 and not self.restarted_from_ckpt:
+            assert self.scale_factor == 1., 'rather not use custom rescaling and std-rescaling simultaneously'
+            # set rescale weight to 1./std of encodings
+            print("### USING STD-RESCALING ###")
+            x = super().get_input(batch, self.first_stage_key)
+            x = x.to(self.device)
+            encoder_posterior = self.encode_first_stage(x)
+            z = self.get_first_stage_encoding(encoder_posterior).detach()
+            del self.scale_factor
+            self.register_buffer('scale_factor', 1. / z.flatten().std())
+            print(f"setting self.scale_factor to {self.scale_factor}")
+            print("### USING STD-RESCALING ###")
+
+    def register_schedule(self,
+                          given_betas=None, beta_schedule="linear", timesteps=1000,
+                          linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+        super().register_schedule(given_betas, beta_schedule, timesteps, linear_start, linear_end, cosine_s)
+
+        self.shorten_cond_schedule = self.num_timesteps_cond > 1
+        if self.shorten_cond_schedule:
+            self.make_cond_schedule()
+
+    def instantiate_first_stage(self, config):
+        model = instantiate_from_config(config)
+        self.first_stage_model = model.eval()
+        self.first_stage_model.train = disabled_train
+        for param in self.first_stage_model.parameters():
+            param.requires_grad = False
+
+    def instantiate_cond_stage(self, config):
+        if not self.cond_stage_trainable:
+            if config == "__is_first_stage__":# for no_text inpainting task
+                print("Using first stage also as cond stage.")
+                self.cond_stage_model = self.first_stage_model
+            elif config == "__is_unconditional__":# for unconditional image generation such as human face、ImageNet
+                print(f"Training {self.__class__.__name__} as an unconditional model.")
+                self.cond_stage_model = None
+                # self.be_unconditional = True
+            else:
+                model = instantiate_from_config(config)
+                self.cond_stage_model = model.eval()
+                self.cond_stage_model.train = disabled_train
+                for param in self.cond_stage_model.parameters():
+                    param.requires_grad = False
+        else:
+            assert config != '__is_first_stage__'
+            assert config != '__is_unconditional__'
+            model = instantiate_from_config(config)
+            self.cond_stage_model = model
+
+    def _get_denoise_row_from_list(self, samples, desc='', force_no_decoder_quantization=False):
+        denoise_row = []
+        for zd in tqdm(samples, desc=desc):
+            denoise_row.append(self.decode_first_stage(zd.to(self.device),
+                                                            force_not_quantize=force_no_decoder_quantization))
+        n_imgs_per_row = len(denoise_row)
+        denoise_row = torch.stack(denoise_row)  # n_log_step, n_row, C, H, W
+        denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
+        denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
+        denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
+        return denoise_grid
+
+    def get_first_stage_encoding(self, encoder_posterior):# encode_emb from autoencoder
+        if isinstance(encoder_posterior, DiagonalGaussianDistribution):
+            z = encoder_posterior.sample()
+        elif isinstance(encoder_posterior, torch.Tensor):
+            z = encoder_posterior
+        else:
+            raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
+        return self.scale_factor * z
+
+    def get_learned_conditioning(self, c):
+        if self.cond_stage_forward is None:
+            if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
+                c = self.cond_stage_model.encode(c)
+                if isinstance(c, DiagonalGaussianDistribution):
+                    c = c.mode()
+            else:
+                c = self.cond_stage_model(c)
+        else:
+            assert hasattr(self.cond_stage_model, self.cond_stage_forward)
+            c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
+        return c
+
+    def meshgrid(self, h, w):
+        y = torch.arange(0, h).view(h, 1, 1).repeat(1, w, 1)
+        x = torch.arange(0, w).view(1, w, 1).repeat(h, 1, 1)
+
+        arr = torch.cat([y, x], dim=-1)
+        return arr
+
+    def delta_border(self, h, w):
+        """
+        :param h: height
+        :param w: width
+        :return: normalized distance to image border,
+         wtith min distance = 0 at border and max dist = 0.5 at image center
+        """
+        lower_right_corner = torch.tensor([h - 1, w - 1]).view(1, 1, 2)
+        arr = self.meshgrid(h, w) / lower_right_corner
+        dist_left_up = torch.min(arr, dim=-1, keepdims=True)[0]
+        dist_right_down = torch.min(1 - arr, dim=-1, keepdims=True)[0]
+        edge_dist = torch.min(torch.cat([dist_left_up, dist_right_down], dim=-1), dim=-1)[0]
+        return edge_dist
+
+    def get_weighting(self, h, w, Ly, Lx, device):
+        weighting = self.delta_border(h, w)
+        weighting = torch.clip(weighting, self.split_input_params["clip_min_weight"],
+                               self.split_input_params["clip_max_weight"], )
+        weighting = weighting.view(1, h * w, 1).repeat(1, 1, Ly * Lx).to(device)
+
+        if self.split_input_params["tie_braker"]:
+            L_weighting = self.delta_border(Ly, Lx)
+            L_weighting = torch.clip(L_weighting,
+                                     self.split_input_params["clip_min_tie_weight"],
+                                     self.split_input_params["clip_max_tie_weight"])
+
+            L_weighting = L_weighting.view(1, 1, Ly * Lx).to(device)
+            weighting = weighting * L_weighting
+        return weighting
+
+    def get_fold_unfold(self, x, kernel_size, stride, uf=1, df=1):  # todo load once not every time, shorten code
+        """
+        :param x: img of size (bs, c, h, w)
+        :return: n img crops of size (n, bs, c, kernel_size[0], kernel_size[1])
+        """
+        bs, nc, h, w = x.shape
+
+        # number of crops in image
+        Ly = (h - kernel_size[0]) // stride[0] + 1
+        Lx = (w - kernel_size[1]) // stride[1] + 1
+
+        if uf == 1 and df == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold = torch.nn.Fold(output_size=x.shape[2:], **fold_params)
+
+            weighting = self.get_weighting(kernel_size[0], kernel_size[1], Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h, w)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0], kernel_size[1], Ly * Lx))
+
+        elif uf > 1 and df == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold_params2 = dict(kernel_size=(kernel_size[0] * uf, kernel_size[0] * uf),
+                                dilation=1, padding=0,
+                                stride=(stride[0] * uf, stride[1] * uf))
+            fold = torch.nn.Fold(output_size=(x.shape[2] * uf, x.shape[3] * uf), **fold_params2)
+
+            weighting = self.get_weighting(kernel_size[0] * uf, kernel_size[1] * uf, Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h * uf, w * uf)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0] * uf, kernel_size[1] * uf, Ly * Lx))
+
+        elif df > 1 and uf == 1:
+            fold_params = dict(kernel_size=kernel_size, dilation=1, padding=0, stride=stride)
+            unfold = torch.nn.Unfold(**fold_params)
+
+            fold_params2 = dict(kernel_size=(kernel_size[0] // df, kernel_size[0] // df),
+                                dilation=1, padding=0,
+                                stride=(stride[0] // df, stride[1] // df))
+            fold = torch.nn.Fold(output_size=(x.shape[2] // df, x.shape[3] // df), **fold_params2)
+
+            weighting = self.get_weighting(kernel_size[0] // df, kernel_size[1] // df, Ly, Lx, x.device).to(x.dtype)
+            normalization = fold(weighting).view(1, 1, h // df, w // df)  # normalizes the overlap
+            weighting = weighting.view((1, 1, kernel_size[0] // df, kernel_size[1] // df, Ly * Lx))
+
+        else:
+            raise NotImplementedError
+
+        return fold, unfold, normalization, weighting
+
+    @torch.no_grad()
+    def get_input(self, batch, k, return_first_stage_outputs=False, force_c_encode=False,
+                  cond_key=None, return_original_cond=False, bs=None):
+        x = super().get_input(batch, k)
+        if bs is not None:
+            x = x[:bs]
+        x = x.to(self.device)
+        encoder_posterior = self.encode_first_stage(x)
+        z = self.get_first_stage_encoding(encoder_posterior).detach()
+
+        if self.model.conditioning_key is not None:# 'crossattn' for txt2image, 'hybird' for txt_inpaint
+            if cond_key is None:
+                cond_key = self.cond_stage_key # 'caption' for txt_inpaint
+            if self.model.conditioning_key == 'hybrid':
+                xc = {}
+                assert cond_key == 'caption' # only txt_inpaint is implemented now
+                assert 'masked_image' in batch.keys() 
+                assert 'mask' in batch.keys()
+                masked_image = super().get_input(batch,'masked_image')
+                mask = super().get_input(batch,'mask')
+                if bs is not None:
+                    masked_image,mask = masked_image[:bs],mask[:bs]
+                masked_image,mask = masked_image.to(self.device),mask.to(self.device)
+                masked_image = self.get_first_stage_encoding(self.encode_first_stage(masked_image)).detach()
+                resized_mask = torch.nn.functional.interpolate(mask,size=masked_image.shape[-2:])
+                xc['c_concat'] = torch.cat((masked_image,resized_mask),dim = 1)
+                xc[cond_key] = batch[cond_key]
+            else:
+                if cond_key != self.first_stage_key:
+                    if cond_key in ['caption', 'coordinates_bbox']:
+                        xc = batch[cond_key]
+                    elif cond_key == 'class_label':
+                        xc = batch
+                    else:
+                        xc = super().get_input(batch, cond_key).to(self.device)
+                else:# cond_key == 'image'
+                    xc = x
+            if not self.cond_stage_trainable or force_c_encode:# cond_stage_trainable is true for txt2img，force_c_encoder = True，when called in log_images
+                if isinstance(xc, list):
+                    # import pudb; pudb.set_trace()
+                    c = self.get_learned_conditioning(xc)# 因为log_images内接下来要调用sample_log,所以需要预先得到处理好的c
+                if isinstance(xc, dict):
+                    c = {}
+                    c['c_concat'] = xc['c_concat']
+                    c['c_crossattn'] = self.get_learned_conditioning(xc[cond_key])
+                else:
+                    c = self.get_learned_conditioning(xc.to(self.device))
+            else:
+                c = xc
+            if bs is not None:
+                if isinstance(c,dict):
+                    for k in c.keys():
+                        c[k] = c[k][:bs]
+                else:
+                    c = c[:bs]
+
+            if self.use_positional_encodings:
+                pos_x, pos_y = self.compute_latent_shifts(batch)
+                ckey = __conditioning_keys__[self.model.conditioning_key]
+                c = {ckey: c, 'pos_x': pos_x, 'pos_y': pos_y}
+
+        else:
+            c = None
+            xc = None
+            if self.use_positional_encodings:
+                pos_x, pos_y = self.compute_latent_shifts(batch)
+                c = {'pos_x': pos_x, 'pos_y': pos_y}
+        out = [z, c]
+        if return_first_stage_outputs:
+            xrec = self.decode_first_stage(z)
+            out.extend([x, xrec])
+        if return_original_cond:
+            out.append(xc)
+        return out
+
+    @torch.no_grad()
+    def decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
+        if predict_cids:
+            if z.dim() == 4:
+                z = torch.argmax(z.exp(), dim=1).long()
+            z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
+            z = rearrange(z, 'b h w c -> b c h w').contiguous()
+
+        z = 1. / self.scale_factor * z
+
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                uf = self.split_input_params["vqf"]
+                bs, nc, h, w = z.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
+
+                z = unfold(z)  # (bn, nc * prod(**ks), L)
+                # 1. Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                # 2. apply model loop over last dim
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
+                                                                 force_not_quantize=predict_cids or force_not_quantize)
+                                   for i in range(z.shape[-1])]
+                else:
+
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
+                                   for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)  # # (bn, nc, ks[0], ks[1], L)
+                o = o * weighting
+                # Reverse 1. reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization  # norm is shape (1, 1, h, w)
+                return decoded
+            else:
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+                else:
+                    return self.first_stage_model.decode(z)
+
+        else:
+            if isinstance(self.first_stage_model, VQModelInterface):
+                return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+            else:
+                return self.first_stage_model.decode(z)
+
+    # same as above but without decorator
+    def differentiable_decode_first_stage(self, z, predict_cids=False, force_not_quantize=False):
+        if predict_cids:
+            if z.dim() == 4:
+                z = torch.argmax(z.exp(), dim=1).long()
+            z = self.first_stage_model.quantize.get_codebook_entry(z, shape=None)
+            z = rearrange(z, 'b h w c -> b c h w').contiguous()
+
+        z = 1. / self.scale_factor * z
+
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                uf = self.split_input_params["vqf"]
+                bs, nc, h, w = z.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(z, ks, stride, uf=uf)
+
+                z = unfold(z)  # (bn, nc * prod(**ks), L)
+                # 1. Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                # 2. apply model loop over last dim
+                if isinstance(self.first_stage_model, VQModelInterface):  
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i],
+                                                                 force_not_quantize=predict_cids or force_not_quantize)
+                                   for i in range(z.shape[-1])]
+                else:
+
+                    output_list = [self.first_stage_model.decode(z[:, :, :, :, i])
+                                   for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)  # # (bn, nc, ks[0], ks[1], L)
+                o = o * weighting
+                # Reverse 1. reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization  # norm is shape (1, 1, h, w)
+                return decoded
+            else:
+                if isinstance(self.first_stage_model, VQModelInterface):
+                    return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+                else:
+                    return self.first_stage_model.decode(z)
+
+        else:
+            if isinstance(self.first_stage_model, VQModelInterface):
+                return self.first_stage_model.decode(z, force_not_quantize=predict_cids or force_not_quantize)
+            else:
+                return self.first_stage_model.decode(z)
+
+    @torch.no_grad()
+    def encode_first_stage(self, x):
+        if hasattr(self, "split_input_params"):
+            if self.split_input_params["patch_distributed_vq"]:
+                ks = self.split_input_params["ks"]  # eg. (128, 128)
+                stride = self.split_input_params["stride"]  # eg. (64, 64)
+                df = self.split_input_params["vqf"]
+                self.split_input_params['original_image_size'] = x.shape[-2:]
+                bs, nc, h, w = x.shape
+                if ks[0] > h or ks[1] > w:
+                    ks = (min(ks[0], h), min(ks[1], w))
+                    print("reducing Kernel")
+
+                if stride[0] > h or stride[1] > w:
+                    stride = (min(stride[0], h), min(stride[1], w))
+                    print("reducing stride")
+
+                fold, unfold, normalization, weighting = self.get_fold_unfold(x, ks, stride, df=df)
+                z = unfold(x)  # (bn, nc * prod(**ks), L)
+                # Reshape to img shape
+                z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                output_list = [self.first_stage_model.encode(z[:, :, :, :, i])
+                               for i in range(z.shape[-1])]
+
+                o = torch.stack(output_list, axis=-1)
+                o = o * weighting
+
+                # Reverse reshape to img shape
+                o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+                # stitch crops together
+                decoded = fold(o)
+                decoded = decoded / normalization
+                return decoded
+
+            else:
+                return self.first_stage_model.encode(x)
+        else:
+            return self.first_stage_model.encode(x)
+
+    def shared_step(self, batch, **kwargs):
+        x, c = self.get_input(batch, self.first_stage_key)# get latent and condition
+        loss = self(x, c)
+        return loss
+
+    def test_step(self,batch,batch_idx):
+        # TODO make self.test_repeat work
+        cond = {}
+        cond[self.cond_stage_key] = batch[self.cond_stage_key]
+        cond[self.cond_stage_key] = self.get_learned_conditioning(cond[self.cond_stage_key]) # c: string -> [B, T, Context_dim]
+        cond['c_crossattn'] = cond.pop(self.cond_stage_key)
+        masked_image = super().get_input(batch,'masked_image')
+        mask = super().get_input(batch,'mask')
+        masked_image,mask = masked_image.to(self.device),mask.to(self.device)
+        masked_image = self.get_first_stage_encoding(self.encode_first_stage(masked_image)).detach()
+        resized_mask = torch.nn.functional.interpolate(mask,size=masked_image.shape[-2:])
+        cond['c_concat'] = torch.cat((masked_image,resized_mask),dim = 1)
+        batch_size = len(batch[self.cond_stage_key])
+        # shape = [batch_size,self.channels,self.mel_dim,self.mel_length]
+        enc_emb = self.sample(cond,batch_size,timesteps=self.test_numsteps)
+        xrec = self.decode_first_stage(enc_emb)
+        reconstructions = (xrec + 1)/2 # to mel scale
+        test_ckpt_path = os.path.basename(self.trainer.tested_ckpt_path)
+        savedir = os.path.join(self.trainer.log_dir,f'output_imgs_{test_ckpt_path}','fake_class')
+        if not os.path.exists(savedir):
+            os.makedirs(savedir)
+
+        file_names = batch['f_name']
+        nfiles = len(file_names)
+        reconstructions = reconstructions.cpu().numpy().squeeze(1) # squuze channel dim
+        for k in range(reconstructions.shape[0]):
+            b,repeat = k % nfiles, k // nfiles
+            vname_num_split_index = file_names[b].rfind('_')# file_names[b]:video_name+'_'+num
+            v_n,num = file_names[b][:vname_num_split_index],file_names[b][vname_num_split_index+1:]
+            save_img_path = os.path.join(savedir,f'{v_n}_sample_{num}_{repeat}.npy')# the num_th caption, the repeat_th repitition
+            np.save(save_img_path,reconstructions[b])
+        
+        return None
+
+    def forward(self, x, c, *args, **kwargs):
+        t = torch.randint(0, self.num_timesteps, (x.shape[0],), device=self.device).long()
+        if self.model.conditioning_key is not None:
+            assert c is not None
+            if self.cond_stage_trainable:
+                if isinstance(c,dict):
+                    c[self.cond_stage_key] = self.get_learned_conditioning(c[self.cond_stage_key])
+                    c['c_crossattn'] = c.pop(self.cond_stage_key)
+                else:
+                    c = self.get_learned_conditioning(c) # c: string -> [B, T, Context_dim]
+            if self.shorten_cond_schedule:  # TODO: drop this option
+                tc = self.cond_ids[t].to(self.device)
+                c = self.q_sample(x_start=c, t=tc, noise=torch.randn_like(c.float()))
+        return self.p_losses(x, c, t, *args, **kwargs)
+
+    def _rescale_annotations(self, bboxes, crop_coordinates):  # TODO: move to dataset
+        def rescale_bbox(bbox):
+            x0 = torch.clamp((bbox[0] - crop_coordinates[0]) / crop_coordinates[2])
+            y0 = torch.clamp((bbox[1] - crop_coordinates[1]) / crop_coordinates[3])
+            w = min(bbox[2] / crop_coordinates[2], 1 - x0)
+            h = min(bbox[3] / crop_coordinates[3], 1 - y0)
+            return x0, y0, w, h
+
+        return [rescale_bbox(b) for b in bboxes]
+
+    def apply_model(self, x_noisy, t, cond, return_ids=False):
+        # make values to list to enable concat operation in 
+        if isinstance(cond, dict):
+            # hybrid case, cond is exptected to be a dict. (txt2inpaint)
+            cond_tmp = {}# use cond_tmp to avoid inplace edit
+            for k,v in cond.items():
+                if not isinstance(v, list):
+                    cond_tmp[k] = [cond[k]]
+                else:
+                    cond_tmp[k] = cond[k]
+            cond = cond_tmp
+        else:
+            if not isinstance(cond, list):
+                cond = [cond]
+            key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn'
+            cond = {key: cond}
+
+        if hasattr(self, "split_input_params"):
+            assert len(cond) == 1  # todo can only deal with one conditioning atm
+            assert not return_ids  
+            ks = self.split_input_params["ks"]  # eg. (128, 128)
+            stride = self.split_input_params["stride"]  # eg. (64, 64)
+
+            h, w = x_noisy.shape[-2:]
+
+            fold, unfold, normalization, weighting = self.get_fold_unfold(x_noisy, ks, stride)
+
+            z = unfold(x_noisy)  # (bn, nc * prod(**ks), L)
+            # Reshape to img shape
+            z = z.view((z.shape[0], -1, ks[0], ks[1], z.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+            z_list = [z[:, :, :, :, i] for i in range(z.shape[-1])]
+
+            if self.cond_stage_key in ["image", "LR_image", "segmentation",
+                                       'bbox_img'] and self.model.conditioning_key:  # todo check for completeness
+                c_key = next(iter(cond.keys()))  # get key
+                c = next(iter(cond.values()))  # get value
+                assert (len(c) == 1)  # todo extend to list with more than one elem
+                c = c[0]  # get element
+
+                c = unfold(c)
+                c = c.view((c.shape[0], -1, ks[0], ks[1], c.shape[-1]))  # (bn, nc, ks[0], ks[1], L )
+
+                cond_list = [{c_key: [c[:, :, :, :, i]]} for i in range(c.shape[-1])]
+
+            elif self.cond_stage_key == 'coordinates_bbox':
+                assert 'original_image_size' in self.split_input_params, 'BoudingBoxRescaling is missing original_image_size'
+
+                # assuming padding of unfold is always 0 and its dilation is always 1
+                n_patches_per_row = int((w - ks[0]) / stride[0] + 1)
+                full_img_h, full_img_w = self.split_input_params['original_image_size']
+                # as we are operating on latents, we need the factor from the original image size to the
+                # spatial latent size to properly rescale the crops for regenerating the bbox annotations
+                num_downs = self.first_stage_model.encoder.num_resolutions - 1
+                rescale_latent = 2 ** (num_downs)
+
+                # get top left postions of patches as conforming for the bbbox tokenizer, therefore we
+                # need to rescale the tl patch coordinates to be in between (0,1)
+                tl_patch_coordinates = [(rescale_latent * stride[0] * (patch_nr % n_patches_per_row) / full_img_w,
+                                         rescale_latent * stride[1] * (patch_nr // n_patches_per_row) / full_img_h)
+                                        for patch_nr in range(z.shape[-1])]
+
+                # patch_limits are tl_coord, width and height coordinates as (x_tl, y_tl, h, w)
+                patch_limits = [(x_tl, y_tl,
+                                 rescale_latent * ks[0] / full_img_w,
+                                 rescale_latent * ks[1] / full_img_h) for x_tl, y_tl in tl_patch_coordinates]
+                # patch_values = [(np.arange(x_tl,min(x_tl+ks, 1.)),np.arange(y_tl,min(y_tl+ks, 1.))) for x_tl, y_tl in tl_patch_coordinates]
+
+                # tokenize crop coordinates for the bounding boxes of the respective patches
+                patch_limits_tknzd = [torch.LongTensor(self.bbox_tokenizer._crop_encoder(bbox))[None].to(self.device)
+                                      for bbox in patch_limits]  # list of length l with tensors of shape (1, 2)
+                print(patch_limits_tknzd[0].shape)
+                # cut tknzd crop position from conditioning
+                assert isinstance(cond, dict), 'cond must be dict to be fed into model'
+                cut_cond = cond['c_crossattn'][0][..., :-2].to(self.device)
+                print(cut_cond.shape)
+
+                adapted_cond = torch.stack([torch.cat([cut_cond, p], dim=1) for p in patch_limits_tknzd])
+                adapted_cond = rearrange(adapted_cond, 'l b n -> (l b) n')
+                print(adapted_cond.shape)
+                adapted_cond = self.get_learned_conditioning(adapted_cond)
+                print(adapted_cond.shape)
+                adapted_cond = rearrange(adapted_cond, '(l b) n d -> l b n d', l=z.shape[-1])
+                print(adapted_cond.shape)
+
+                cond_list = [{'c_crossattn': [e]} for e in adapted_cond]
+
+            else:
+                cond_list = [cond for i in range(z.shape[-1])]  # Todo make this more efficient
+
+            # apply model by loop over crops
+            output_list = [self.model(z_list[i], t, **cond_list[i]) for i in range(z.shape[-1])]
+            assert not isinstance(output_list[0],
+                                  tuple)  # todo cant deal with multiple model outputs check this never happens
+
+            o = torch.stack(output_list, axis=-1)
+            o = o * weighting
+            # Reverse reshape to img shape
+            o = o.view((o.shape[0], -1, o.shape[-1]))  # (bn, nc * ks[0] * ks[1], L)
+            # stitch crops together
+            x_recon = fold(o) / normalization
+
+        else:
+            # x_noisy is tensor with shape [b,c,mel_len,T]
+            # if condition is caption ,cond['c_crossattn'] is a list, each item shape is [1, 77, 1280]
+            x_recon = self.model(x_noisy, t, **cond)# tensor with shape [b,c,mel_len,T]
+
+        if isinstance(x_recon, tuple) and not return_ids:
+            return x_recon[0]
+        else:
+            return x_recon
+
+    def _predict_eps_from_xstart(self, x_t, t, pred_xstart):
+        return (extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - pred_xstart) / \
+               extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
+
+    def _prior_bpd(self, x_start):
+        """
+        Get the prior KL term for the variational lower-bound, measured in
+        bits-per-dim.
+        This term can't be optimized, as it only depends on the encoder.
+        :param x_start: the [N x C x ...] tensor of inputs.
+        :return: a batch of [N] KL values (in bits), one per batch element.
+        """
+        batch_size = x_start.shape[0]
+        t = torch.tensor([self.num_timesteps - 1] * batch_size, device=x_start.device)
+        qt_mean, _, qt_log_variance = self.q_mean_variance(x_start, t)
+        kl_prior = normal_kl(mean1=qt_mean, logvar1=qt_log_variance, mean2=0.0, logvar2=0.0)
+        return mean_flat(kl_prior) / np.log(2.0)
+
+    def p_losses(self, x_start, cond, t, noise=None):
+        noise = default(noise, lambda: torch.randn_like(x_start))
+        x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
+        model_output = self.apply_model(x_noisy, t, cond)
+
+        loss_dict = {}
+        prefix = 'train' if self.training else 'val'
+
+        if self.parameterization == "x0":
+            target = x_start
+        elif self.parameterization == "eps":
+            target = noise
+        else:
+            raise NotImplementedError()
+
+        loss_simple = self.get_loss(model_output, target, mean=False).mean([1, 2, 3])
+        loss_dict.update({f'{prefix}/loss_simple': loss_simple.mean()})
+
+        logvar_t = self.logvar[t].to(self.device)
+        loss = loss_simple / torch.exp(logvar_t) + logvar_t
+        # loss = loss_simple / torch.exp(self.logvar) + self.logvar
+        if self.learn_logvar:
+            loss_dict.update({f'{prefix}/loss_gamma': loss.mean()})
+            loss_dict.update({'logvar': self.logvar.data.mean()})
+
+        loss = self.l_simple_weight * loss.mean()
+
+        loss_vlb = self.get_loss(model_output, target, mean=False).mean(dim=(1, 2, 3))
+        loss_vlb = (self.lvlb_weights[t] * loss_vlb).mean()
+        loss_dict.update({f'{prefix}/loss_vlb': loss_vlb})
+        loss += (self.original_elbo_weight * loss_vlb)
+        loss_dict.update({f'{prefix}/loss': loss})
+
+        return loss, loss_dict
+
+    def p_mean_variance(self, x, c, t, clip_denoised: bool, return_codebook_ids=False, quantize_denoised=False,
+                        return_x0=False, score_corrector=None, corrector_kwargs=None):
+        t_in = t
+        model_out = self.apply_model(x, t_in, c, return_ids=return_codebook_ids)
+
+        if score_corrector is not None:
+            assert self.parameterization == "eps"
+            model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs)
+
+        if return_codebook_ids:
+            model_out, logits = model_out
+
+        if self.parameterization == "eps":
+            x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
+        elif self.parameterization == "x0":
+            x_recon = model_out
+        else:
+            raise NotImplementedError()
+
+        if clip_denoised:
+            x_recon.clamp_(-1., 1.)
+        if quantize_denoised:
+            x_recon, _, [_, _, indices] = self.first_stage_model.quantize(x_recon)
+        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
+        if return_codebook_ids:
+            return model_mean, posterior_variance, posterior_log_variance, logits
+        elif return_x0:
+            return model_mean, posterior_variance, posterior_log_variance, x_recon
+        else:
+            return model_mean, posterior_variance, posterior_log_variance
+
+    @torch.no_grad()
+    def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False,
+                 return_codebook_ids=False, quantize_denoised=False, return_x0=False,
+                 temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None):
+        b, *_, device = *x.shape, x.device
+        outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised,
+                                       return_codebook_ids=return_codebook_ids,
+                                       quantize_denoised=quantize_denoised,
+                                       return_x0=return_x0,
+                                       score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
+        if return_codebook_ids:
+            raise DeprecationWarning("Support dropped.")
+            model_mean, _, model_log_variance, logits = outputs
+        elif return_x0:
+            model_mean, _, model_log_variance, x0 = outputs
+        else:
+            model_mean, _, model_log_variance = outputs
+
+        noise = noise_like(x.shape, device, repeat_noise) * temperature
+        if noise_dropout > 0.:
+            noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+        # no noise when t == 0
+        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
+
+        if return_codebook_ids:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, logits.argmax(dim=1)
+        if return_x0:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0
+        else:
+            return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
+
+    @torch.no_grad()
+    def progressive_denoising(self, cond, shape, verbose=True, callback=None, quantize_denoised=False,
+                              img_callback=None, mask=None, x0=None, temperature=1., noise_dropout=0.,
+                              score_corrector=None, corrector_kwargs=None, batch_size=None, x_T=None, start_T=None,
+                              log_every_t=None):
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        timesteps = self.num_timesteps
+        if batch_size is not None:
+            b = batch_size if batch_size is not None else shape[0]
+            shape = [batch_size] + list(shape)
+        else:
+            b = batch_size = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=self.device)
+        else:
+            img = x_T
+        intermediates = []
+        if cond is not None:
+            if isinstance(cond, dict):
+                cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
+                list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
+            else:
+                cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
+
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+        iterator = tqdm(reversed(range(0, timesteps)), desc='Progressive Generation',
+                        total=timesteps) if verbose else reversed(
+            range(0, timesteps))
+        if type(temperature) == float:
+            temperature = [temperature] * timesteps
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=self.device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != 'hybrid'
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img, x0_partial = self.p_sample(img, cond, ts,
+                                            clip_denoised=self.clip_denoised,
+                                            quantize_denoised=quantize_denoised, return_x0=True,
+                                            temperature=temperature[i], noise_dropout=noise_dropout,
+                                            score_corrector=score_corrector, corrector_kwargs=corrector_kwargs)
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1. - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(x0_partial)
+            if callback: callback(i)
+            if img_callback: img_callback(img, i)
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_loop(self, cond, shape, return_intermediates=False,
+                      x_T=None, verbose=True, callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, start_T=None,
+                      log_every_t=None):
+
+        if not log_every_t:
+            log_every_t = self.log_every_t
+        device = self.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        intermediates = [img]
+        if timesteps is None:
+            timesteps = self.num_timesteps
+
+        if start_T is not None:
+            timesteps = min(timesteps, start_T)
+        iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(
+            range(0, timesteps))
+
+        if mask is not None:
+            assert x0 is not None
+            assert x0.shape[2:3] == mask.shape[2:3]  # spatial size has to match
+
+        for i in iterator:
+            ts = torch.full((b,), i, device=device, dtype=torch.long)
+            if self.shorten_cond_schedule:
+                assert self.model.conditioning_key != 'hybrid'
+                tc = self.cond_ids[ts].to(cond.device)
+                cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
+
+            img = self.p_sample(img, cond, ts,
+                                clip_denoised=self.clip_denoised,
+                                quantize_denoised=quantize_denoised)
+            if mask is not None:
+                img_orig = self.q_sample(x0, ts)
+                img = img_orig * mask + (1. - mask) * img
+
+            if i % log_every_t == 0 or i == timesteps - 1:
+                intermediates.append(img)
+            if callback: callback(i)
+            if img_callback: img_callback(img, i)
+
+        if return_intermediates:
+            return img, intermediates
+        return img
+
+    @torch.no_grad()
+    def sample(self, cond, batch_size=16, return_intermediates=False, x_T=None,
+               verbose=True, timesteps=None, quantize_denoised=False,
+               mask=None, x0=None, shape=None,**kwargs):
+        if shape is None:
+            shape = (batch_size, self.channels, self.mel_dim, self.mel_length)
+        if cond is not None:
+            if isinstance(cond, dict):
+                cond = {key: cond[key][:batch_size] if not isinstance(cond[key], list) else
+                list(map(lambda x: x[:batch_size], cond[key])) for key in cond}
+            else:
+                cond = [c[:batch_size] for c in cond] if isinstance(cond, list) else cond[:batch_size]
+        return self.p_sample_loop(cond,
+                                  shape,
+                                  return_intermediates=return_intermediates, x_T=x_T,
+                                  verbose=verbose, timesteps=timesteps, quantize_denoised=quantize_denoised,
+                                  mask=mask, x0=x0)
+
+    @torch.no_grad()
+    def sample_log(self,cond,batch_size,ddim, ddim_steps,**kwargs):
+        if ddim:
+            ddim_sampler = DDIMSampler(self)
+            shape = (self.channels, self.mel_dim, self.mel_length)
+            samples, intermediates =ddim_sampler.sample(ddim_steps,batch_size,
+                                                        shape,cond,verbose=False,**kwargs)
+
+        else:
+            samples, intermediates = self.sample(cond=cond, batch_size=batch_size,
+                                                 return_intermediates=True,**kwargs)
+
+        return samples, intermediates
+
+    @torch.no_grad()
+    def log_images(self, batch, N=8, n_row=4, sample=True, ddim_steps=200, ddim_eta=1., return_keys=None,
+                   quantize_denoised=True, inpaint=True, plot_denoise_rows=False, plot_progressive_rows=True,
+                   plot_diffusion_rows=True, **kwargs):
+
+        use_ddim = ddim_steps is not None
+
+        log = dict()
+        z, c, x, xrec, xc = self.get_input(batch, self.first_stage_key,
+                                           return_first_stage_outputs=True,
+                                           force_c_encode=True,
+                                           return_original_cond=True,
+                                           bs=N)
+
+        N = min(x.shape[0], N)
+        n_row = min(x.shape[0], n_row)
+        log["inputs"] = x # 原始输入图像
+        log["reconstruction"] = xrec # 重建得到的图像
+        if self.model.conditioning_key is not None:
+            if hasattr(self.cond_stage_model, "decode"):# when cond_stage is first_stage. (bert embedder doesnot have decode)
+                xc = self.cond_stage_model.decode(c)# decoded masked image
+                log["conditioning"] = xc # 重建后的图像
+            elif self.cond_stage_key in ["caption"]:
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["caption"])
+                log["conditioning"] = xc # 含有文本的图像 
+                if self.model.conditioning_key == 'hybrid':
+                    log["decoded_maskedimg"] = self.first_stage_model.decode(c['c_concat'][:,:self.first_stage_model.embed_dim])# c_concat is the concat result of masked_img latent and resized mask. get latent here to decode
+            elif self.cond_stage_key == 'class_label':
+                xc = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"])
+                log['conditioning'] = xc # 文本为类标签的图像
+            elif isimage(xc):
+                log["conditioning"] = xc
+            if ismap(xc):
+                log["original_conditioning"] = self.to_rgb(xc)
+
+        if plot_diffusion_rows:# diffusion每一步的图像
+            # get diffusion row
+            diffusion_row = list()
+            z_start = z[:n_row]
+            for t in range(self.num_timesteps):
+                if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
+                    t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
+                    t = t.to(self.device).long()
+                    noise = torch.randn_like(z_start)
+                    z_noisy = self.q_sample(x_start=z_start, t=t, noise=noise)
+                    diffusion_row.append(self.decode_first_stage(z_noisy))
+
+            diffusion_row = torch.stack(diffusion_row)  # n_log_step, n_row, C, H, W
+            diffusion_grid = rearrange(diffusion_row, 'n b c h w -> b n c h w')
+            diffusion_grid = rearrange(diffusion_grid, 'b n c h w -> (b n) c h w')
+            diffusion_grid = make_grid(diffusion_grid, nrow=diffusion_row.shape[0])
+            log["diffusion_row"] = diffusion_grid
+
+        if sample:# 
+            # get denoise row
+            with self.ema_scope("Plotting"):
+                samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
+                                                         ddim_steps=ddim_steps,eta=ddim_eta)
+                # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True)
+            x_samples = self.decode_first_stage(samples)
+            log["samples"] = x_samples
+            if plot_denoise_rows:
+                denoise_grid = self._get_denoise_row_from_list(z_denoise_row)
+                log["denoise_row"] = denoise_grid
+
+            if quantize_denoised and not isinstance(self.first_stage_model, AutoencoderKL) and not isinstance(
+                    self.first_stage_model, IdentityFirstStage):
+                # also display when quantizing x0 while sampling
+                with self.ema_scope("Plotting Quantized Denoised"):
+                    samples, z_denoise_row = self.sample_log(cond=c,batch_size=N,ddim=use_ddim,
+                                                             ddim_steps=ddim_steps,eta=ddim_eta,
+                                                             quantize_denoised=True)
+                    # samples, z_denoise_row = self.sample(cond=c, batch_size=N, return_intermediates=True,
+                    #                                      quantize_denoised=True)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_x0_quantized"] = x_samples
+
+            if inpaint:
+                # make a simple center square
+                b, h, w = z.shape[0], z.shape[2], z.shape[3]
+                mask = torch.ones(N, h, w).to(self.device)
+                # zeros will be filled in
+                mask[:, h // 4:3 * h // 4, w // 4:3 * w // 4] = 0.
+                mask = mask[:, None, ...]# N,1,H,W
+                with self.ema_scope("Plotting Inpaint"):
+                    samples, _ = self.sample_log(cond=c,batch_size=N,ddim=use_ddim, eta=ddim_eta,
+                                                ddim_steps=ddim_steps, x0=z[:N], mask=mask)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_inpainting"] = x_samples
+                log["mask"] = mask
+
+                # outpaint
+                with self.ema_scope("Plotting Outpaint"):
+                    samples, _ = self.sample_log(cond=c, batch_size=N, ddim=use_ddim,eta=ddim_eta,
+                                                ddim_steps=ddim_steps, x0=z[:N], mask=mask)
+                x_samples = self.decode_first_stage(samples.to(self.device))
+                log["samples_outpainting"] = x_samples
+
+        if plot_progressive_rows:
+            with self.ema_scope("Plotting Progressives"):
+                img, progressives = self.progressive_denoising(c,
+                                                               shape=(self.channels, self.mel_dim, self.mel_length),
+                                                               batch_size=N)
+            prog_row = self._get_denoise_row_from_list(progressives, desc="Progressive Generation")
+            log["progressive_row"] = prog_row
+
+        if return_keys:
+            if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
+                return log
+            else:
+                return {key: log[key] for key in return_keys}
+        return log
+
+    def configure_optimizers(self):
+        lr = self.learning_rate
+        params = list(self.model.parameters())
+        if self.cond_stage_trainable:
+            print(f"{self.__class__.__name__}: Also optimizing conditioner params!")
+            params = params + list(self.cond_stage_model.parameters())
+        if self.learn_logvar:
+            print('Diffusion model optimizing logvar')
+            params.append(self.logvar)
+        opt = torch.optim.AdamW(params, lr=lr)
+        if self.use_scheduler:
+            assert 'target' in self.scheduler_config
+            scheduler = instantiate_from_config(self.scheduler_config)
+
+            print("Setting up LambdaLR scheduler...")
+            scheduler = [
+                {
+                    'scheduler': LambdaLR(opt, lr_lambda=scheduler.schedule),
+                    'interval': 'step',
+                    'frequency': 1
+                }]
+            return [opt], scheduler
+        return opt
+
+    @torch.no_grad()
+    def to_rgb(self, x):
+        x = x.float()
+        if not hasattr(self, "colorize"):
+            self.colorize = torch.randn(3, x.shape[1], 1, 1).to(x)
+        x = nn.functional.conv2d(x, weight=self.colorize)
+        x = 2. * (x - x.min()) / (x.max() - x.min()) - 1.
+        return x
+

ldm/models/diffusion/plms.py

@@ -0,0 +1,236 @@
+"""SAMPLING ONLY."""
+
+import torch
+import numpy as np
+from tqdm import tqdm
+from functools import partial
+
+from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like
+
+
+class PLMSSampler(object):
+    def __init__(self, model, schedule="linear", **kwargs):
+        super().__init__()
+        self.model = model
+        self.ddpm_num_timesteps = model.num_timesteps
+        self.schedule = schedule
+
+    def register_buffer(self, name, attr):
+        if type(attr) == torch.Tensor:
+            if attr.device != torch.device("cuda"):
+                attr = attr.to(torch.device("cuda"))
+        setattr(self, name, attr)
+
+    def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
+        if ddim_eta != 0:
+            raise ValueError('ddim_eta must be 0 for PLMS')
+        self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
+                                                  num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
+        alphas_cumprod = self.model.alphas_cumprod
+        assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
+        to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
+
+        self.register_buffer('betas', to_torch(self.model.betas))
+        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
+        self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
+
+        # calculations for diffusion q(x_t | x_{t-1}) and others
+        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
+        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
+        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
+
+        # ddim sampling parameters
+        ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
+                                                                                   ddim_timesteps=self.ddim_timesteps,
+                                                                                   eta=ddim_eta,verbose=verbose)
+        self.register_buffer('ddim_sigmas', ddim_sigmas)
+        self.register_buffer('ddim_alphas', ddim_alphas)
+        self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
+        self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
+        sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
+            (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
+                        1 - self.alphas_cumprod / self.alphas_cumprod_prev))
+        self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
+
+    @torch.no_grad()
+    def sample(self,
+               S,
+               batch_size,
+               shape,
+               conditioning=None,
+               callback=None,
+               normals_sequence=None,
+               img_callback=None,
+               quantize_x0=False,
+               eta=0.,
+               mask=None,
+               x0=None,
+               temperature=1.,
+               noise_dropout=0.,
+               score_corrector=None,
+               corrector_kwargs=None,
+               verbose=True,
+               x_T=None,
+               log_every_t=100,
+               unconditional_guidance_scale=1.,
+               unconditional_conditioning=None,
+               # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
+               **kwargs
+               ):
+        if conditioning is not None:
+            if isinstance(conditioning, dict):
+                cbs = conditioning[list(conditioning.keys())[0]].shape[0]
+                if cbs != batch_size:
+                    print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
+            else:
+                if conditioning.shape[0] != batch_size:
+                    print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
+
+        self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose)
+        # sampling
+        C, H, W = shape
+        size = (batch_size, C, H, W)
+        print(f'Data shape for PLMS sampling is {size}')
+
+        samples, intermediates = self.plms_sampling(conditioning, size,
+                                                    callback=callback,
+                                                    img_callback=img_callback,
+                                                    quantize_denoised=quantize_x0,
+                                                    mask=mask, x0=x0,
+                                                    ddim_use_original_steps=False,
+                                                    noise_dropout=noise_dropout,
+                                                    temperature=temperature,
+                                                    score_corrector=score_corrector,
+                                                    corrector_kwargs=corrector_kwargs,
+                                                    x_T=x_T,
+                                                    log_every_t=log_every_t,
+                                                    unconditional_guidance_scale=unconditional_guidance_scale,
+                                                    unconditional_conditioning=unconditional_conditioning,
+                                                    )
+        return samples, intermediates
+
+    @torch.no_grad()
+    def plms_sampling(self, cond, shape,
+                      x_T=None, ddim_use_original_steps=False,
+                      callback=None, timesteps=None, quantize_denoised=False,
+                      mask=None, x0=None, img_callback=None, log_every_t=100,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None,):
+        device = self.model.betas.device
+        b = shape[0]
+        if x_T is None:
+            img = torch.randn(shape, device=device)
+        else:
+            img = x_T
+
+        if timesteps is None:
+            timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
+        elif timesteps is not None and not ddim_use_original_steps:
+            subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
+            timesteps = self.ddim_timesteps[:subset_end]
+
+        intermediates = {'x_inter': [img], 'pred_x0': [img]}
+        time_range = list(reversed(range(0,timesteps))) if ddim_use_original_steps else np.flip(timesteps)
+        total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
+        print(f"Running PLMS Sampling with {total_steps} timesteps")
+
+        iterator = tqdm(time_range, desc='PLMS Sampler', total=total_steps)
+        old_eps = []
+
+        for i, step in enumerate(iterator):
+            index = total_steps - i - 1
+            ts = torch.full((b,), step, device=device, dtype=torch.long)
+            ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=device, dtype=torch.long)
+
+            if mask is not None:
+                assert x0 is not None
+                img_orig = self.model.q_sample(x0, ts)  # TODO: deterministic forward pass?
+                img = img_orig * mask + (1. - mask) * img
+
+            outs = self.p_sample_plms(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
+                                      quantize_denoised=quantize_denoised, temperature=temperature,
+                                      noise_dropout=noise_dropout, score_corrector=score_corrector,
+                                      corrector_kwargs=corrector_kwargs,
+                                      unconditional_guidance_scale=unconditional_guidance_scale,
+                                      unconditional_conditioning=unconditional_conditioning,
+                                      old_eps=old_eps, t_next=ts_next)
+            img, pred_x0, e_t = outs
+            old_eps.append(e_t)
+            if len(old_eps) >= 4:
+                old_eps.pop(0)
+            if callback: callback(i)
+            if img_callback: img_callback(pred_x0, i)
+
+            if index % log_every_t == 0 or index == total_steps - 1:
+                intermediates['x_inter'].append(img)
+                intermediates['pred_x0'].append(pred_x0)
+
+        return img, intermediates
+
+    @torch.no_grad()
+    def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
+                      temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
+                      unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None):
+        b, *_, device = *x.shape, x.device
+
+        def get_model_output(x, t):
+            if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
+                e_t = self.model.apply_model(x, t, c)
+            else:
+                x_in = torch.cat([x] * 2)
+                t_in = torch.cat([t] * 2)
+                c_in = torch.cat([unconditional_conditioning, c])
+                e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2)
+                e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
+
+            if score_corrector is not None:
+                assert self.model.parameterization == "eps"
+                e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
+
+            return e_t
+
+        alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
+        alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
+        sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
+        sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
+
+        def get_x_prev_and_pred_x0(e_t, index):
+            # select parameters corresponding to the currently considered timestep
+            a_t = torch.full((b, 1, 1, 1), alphas[index], device=device)
+            a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device)
+            sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device)
+            sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device)
+
+            # current prediction for x_0
+            pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
+            if quantize_denoised:
+                pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
+            # direction pointing to x_t
+            dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
+            noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
+            if noise_dropout > 0.:
+                noise = torch.nn.functional.dropout(noise, p=noise_dropout)
+            x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
+            return x_prev, pred_x0
+
+        e_t = get_model_output(x, t)
+        if len(old_eps) == 0:
+            # Pseudo Improved Euler (2nd order)
+            x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index)
+            e_t_next = get_model_output(x_prev, t_next)
+            e_t_prime = (e_t + e_t_next) / 2
+        elif len(old_eps) == 1:
+            # 2nd order Pseudo Linear Multistep (Adams-Bashforth)
+            e_t_prime = (3 * e_t - old_eps[-1]) / 2
+        elif len(old_eps) == 2:
+            # 3nd order Pseudo Linear Multistep (Adams-Bashforth)
+            e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12
+        elif len(old_eps) >= 3:
+            # 4nd order Pseudo Linear Multistep (Adams-Bashforth)
+            e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24
+
+        x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index)
+
+        return x_prev, pred_x0, e_t

ldm/modules/attention.py

@@ -0,0 +1,261 @@
+from inspect import isfunction
+import math
+import torch
+import torch.nn.functional as F
+from torch import nn, einsum
+from einops import rearrange, repeat
+
+from ldm.modules.diffusionmodules.util import checkpoint
+
+
+def exists(val):
+    return val is not None
+
+
+def uniq(arr):
+    return{el: True for el in arr}.keys()
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def max_neg_value(t):
+    return -torch.finfo(t.dtype).max
+
+
+def init_(tensor):
+    dim = tensor.shape[-1]
+    std = 1 / math.sqrt(dim)
+    tensor.uniform_(-std, std)
+    return tensor
+
+
+# feedforward
+class GEGLU(nn.Module):
+    def __init__(self, dim_in, dim_out):
+        super().__init__()
+        self.proj = nn.Linear(dim_in, dim_out * 2)
+
+    def forward(self, x):
+        x, gate = self.proj(x).chunk(2, dim=-1)
+        return x * F.gelu(gate)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
+        super().__init__()
+        inner_dim = int(dim * mult)
+        dim_out = default(dim_out, dim)
+        project_in = nn.Sequential(
+            nn.Linear(dim, inner_dim),
+            nn.GELU()
+        ) if not glu else GEGLU(dim, inner_dim)
+
+        self.net = nn.Sequential(
+            project_in,
+            nn.Dropout(dropout),
+            nn.Linear(inner_dim, dim_out)
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def Normalize(in_channels):
+    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class LinearAttention(nn.Module):
+    def __init__(self, dim, heads=4, dim_head=32):
+        super().__init__()
+        self.heads = heads
+        hidden_dim = dim_head * heads
+        self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
+        self.to_out = nn.Conv2d(hidden_dim, dim, 1)
+
+    def forward(self, x):
+        b, c, h, w = x.shape
+        qkv = self.to_qkv(x)
+        q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
+        k = k.softmax(dim=-1)  
+        context = torch.einsum('bhdn,bhen->bhde', k, v)
+        out = torch.einsum('bhde,bhdn->bhen', context, q)
+        out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
+        return self.to_out(out)
+
+
+class SpatialSelfAttention(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = rearrange(q, 'b c h w -> b (h w) c')
+        k = rearrange(k, 'b c h w -> b c (h w)')
+        w_ = torch.einsum('bij,bjk->bik', q, k)
+
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = rearrange(v, 'b c h w -> b c (h w)')
+        w_ = rearrange(w_, 'b i j -> b j i')
+        h_ = torch.einsum('bij,bjk->bik', v, w_)
+        h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.):# 如果设置了context_dim就不是自注意力了
+        super().__init__()
+        inner_dim = dim_head * heads # inner_dim == SpatialTransformer.model_channels
+        context_dim = default(context_dim, query_dim)
+
+        self.scale = dim_head ** -0.5
+        self.heads = heads
+
+        self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
+        self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
+        self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
+
+        self.to_out = nn.Sequential(
+            nn.Linear(inner_dim, query_dim),
+            nn.Dropout(dropout)
+        )
+
+    def forward(self, x, context=None, mask=None):# x:(b,h*w,c), context:(b,seq_len,context_dim)
+        h = self.heads
+
+        q = self.to_q(x)# q:(b,h*w,inner_dim)
+        context = default(context, x)
+        k = self.to_k(context)# (b,seq_len,inner_dim)
+        v = self.to_v(context)# (b,seq_len,inner_dim)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))# n is seq_len for k and v
+
+        sim = einsum('b i d, b j d -> b i j', q, k) * self.scale # (b*head,h*w,seq_len)
+
+        if exists(mask):# false
+            mask = rearrange(mask, 'b ... -> b (...)')
+            max_neg_value = -torch.finfo(sim.dtype).max
+            mask = repeat(mask, 'b j -> (b h) () j', h=h)
+            sim.masked_fill_(~mask, max_neg_value)
+
+        # attention, what we cannot get enough of
+        attn = sim.softmax(dim=-1)
+
+        out = einsum('b i j, b j d -> b i d', attn, v)# (b*head,h*w,inner_dim/head)
+        out = rearrange(out, '(b h) n d -> b n (h d)', h=h)# (b,h*w,inner_dim)
+        return self.to_out(out)
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True):
+        super().__init__()
+        self.attn1 = CrossAttention(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout)  # is a self-attention
+        self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
+        self.attn2 = CrossAttention(query_dim=dim, context_dim=context_dim,
+                                    heads=n_heads, dim_head=d_head, dropout=dropout)  # is self-attn if context is none
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+        self.norm3 = nn.LayerNorm(dim)
+        self.checkpoint = checkpoint
+
+    def forward(self, x, context=None):
+        return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint)
+
+    def _forward(self, x, context=None):
+        x = self.attn1(self.norm1(x)) + x
+        x = self.attn2(self.norm2(x), context=context) + x
+        x = self.ff(self.norm3(x)) + x
+        return x
+
+
+class SpatialTransformer(nn.Module):
+    """
+    Transformer block for image-like data.
+    First, project the input (aka embedding)
+    and reshape to b, t, d.
+    Then apply standard transformer action.
+    Finally, reshape to image
+    """
+    def __init__(self, in_channels, n_heads, d_head,
+                 depth=1, dropout=0., context_dim=None):
+        super().__init__()
+        self.in_channels = in_channels
+        inner_dim = n_heads * d_head 
+        self.norm = Normalize(in_channels)
+
+        self.proj_in = nn.Conv2d(in_channels,
+                                 inner_dim,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+
+        self.transformer_blocks = nn.ModuleList(
+            [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim)
+                for d in range(depth)]
+        )
+
+        self.proj_out = zero_module(nn.Conv2d(inner_dim,
+                                              in_channels,
+                                              kernel_size=1,
+                                              stride=1,
+                                              padding=0))
+
+    def forward(self, x, context=None):
+        # note: if no context is given, cross-attention defaults to self-attention
+        b, c, h, w = x.shape # such as [2,320,10,106]
+        x_in = x
+        x = self.norm(x)# group norm
+        x = self.proj_in(x)# no shape change
+        x = rearrange(x, 'b c h w -> b (h w) c')
+        for block in self.transformer_blocks:
+            x = block(x, context=context)# context shape [b,seq_len=77,context_dim]
+        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w)
+        x = self.proj_out(x)
+        return x + x_in

ldm/modules/diffusionmodules/custom_openaimodel.py

@@ -0,0 +1,368 @@
+from abc import abstractmethod
+from functools import partial
+import math
+from typing import Iterable
+
+import numpy as np
+import torch as th
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ldm.modules.diffusionmodules.util import (
+    checkpoint,
+    conv_nd,
+    linear,
+    avg_pool_nd,
+    zero_module,
+    normalization,
+    timestep_embedding,
+)
+from ldm.modules.attention import SpatialTransformer
+from ldm.modules.diffusionmodules.openaimodel import convert_module_to_f16, convert_module_to_f32, AttentionPool2d, \
+    TimestepBlock, TimestepEmbedSequential, Upsample, TransposedUpsample, Downsample, ResBlock, AttentionBlock, count_flops_attn, \
+    QKVAttentionLegacy, QKVAttention
+
+
+class UNetModel(nn.Module):
+    """
+    The full UNet model with attention and timestep embedding.
+    :param in_channels: channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param attention_resolutions: a collection of downsample rates at which
+        attention will take place. May be a set, list, or tuple.
+        For example, if this contains 4, then at 4x downsampling, attention
+        will be used.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param use_checkpoint: use gradient checkpointing to reduce memory usage.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    :param use_new_attention_order: use a different attention pattern for potentially
+                                    increased efficiency.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        num_classes=None,
+        use_checkpoint=False,
+        use_fp16=False,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        use_spatial_transformer=False,    # custom transformer support
+        transformer_depth=1,              # custom transformer support
+        context_dim=None,                 # custom transformer support
+        n_embed=None,                     # custom support for prediction of discrete ids into codebook of first stage vq model
+        legacy=True,
+        use_context_project=False,        # custom text to audio support
+        use_context_attn=True             # custom text to audio support
+    ):
+        super().__init__()
+        if use_spatial_transformer:
+            assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
+
+        if context_dim is not None and not use_context_project:
+            assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
+            from omegaconf.listconfig import ListConfig
+            if type(context_dim) == ListConfig:
+                context_dim = list(context_dim)
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        if num_heads == -1:
+            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
+
+        if num_head_channels == -1:
+            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
+
+        self.image_size = image_size
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        self.num_res_blocks = num_res_blocks
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.num_classes = num_classes
+        self.use_checkpoint = use_checkpoint
+        self.dtype = th.float16 if use_fp16 else th.float32
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.predict_codebook_ids = n_embed is not None
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+
+        if self.num_classes is not None:
+            self.label_emb = nn.Embedding(num_classes, time_embed_dim)
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    conv_nd(dims, in_channels, model_channels, 3, padding=1)
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for _ in range(num_res_blocks):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    layers.append(
+                        AttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads,
+                            num_head_channels=dim_head,
+                            use_new_attention_order=use_new_attention_order,
+                        ) if not use_spatial_transformer else SpatialTransformer(
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
+                        )
+                    )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        if legacy:
+            #num_heads = 1
+            dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            AttentionBlock(
+                ch,
+                use_checkpoint=use_checkpoint,
+                num_heads=num_heads,
+                num_head_channels=dim_head,
+                use_new_attention_order=use_new_attention_order,
+            ) if not use_spatial_transformer else SpatialTransformer(
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
+                        ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self._feature_size += ch
+
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(num_res_blocks + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    ResBlock(
+                        ch + ich,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=model_channels * mult,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = model_channels * mult
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    layers.append(
+                        AttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads_upsample,
+                            num_head_channels=dim_head,
+                            use_new_attention_order=use_new_attention_order,
+                        ) if not use_spatial_transformer else SpatialTransformer(
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
+                        )
+                    )
+                if level and i == num_res_blocks:
+                    out_ch = ch
+                    layers.append(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True,
+                        )
+                        if resblock_updown
+                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+
+        self.out = nn.Sequential(
+            normalization(ch),
+            nn.SiLU(),
+            zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
+        )
+        if self.predict_codebook_ids:
+            self.id_predictor = nn.Sequential(
+            normalization(ch),
+            conv_nd(dims, model_channels, n_embed, 1),
+            #nn.LogSoftmax(dim=1)  # change to cross_entropy and produce non-normalized logits
+        )
+
+        self.use_context_project = use_context_project
+        if use_context_project:
+            self.context_project = linear(context_dim, time_embed_dim)
+        self.use_context_attn = use_context_attn
+
+
+    def convert_to_fp16(self):
+        """
+        Convert the torso of the model to float16.
+        """
+        self.input_blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+        self.output_blocks.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self):
+        """
+        Convert the torso of the model to float32.
+        """
+        self.input_blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+        self.output_blocks.apply(convert_module_to_f32)
+
+    def forward(self, x, timesteps=None, context=None, y=None,**kwargs):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :param context: conditioning plugged in via crossattn
+        :param y: an [N] Tensor of labels, if class-conditional.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        assert (y is not None) == (
+            self.num_classes is not None
+        ), "must specify y if and only if the model is class-conditional"
+        hs = []
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
+        emb = self.time_embed(t_emb)
+
+        if self.num_classes is not None:
+            assert y.shape == (x.shape[0],)
+            emb = emb + self.label_emb(y)
+
+        # For text-to-audio using global CLIP
+        if self.use_context_project:
+            context = self.context_project(context)
+        emb = emb + context.squeeze(1)
+
+        h = x.type(self.dtype)
+        for module in self.input_blocks:
+            h = module(h, emb, context if self.use_context_attn else None)
+            hs.append(h)
+        h = self.middle_block(h, emb, context if self.use_context_attn else None)
+        for module in self.output_blocks:
+            h = th.cat([h, hs.pop()], dim=1)
+            h = module(h, emb, context if self.use_context_attn else None)
+        h = h.type(x.dtype)
+        if self.predict_codebook_ids:
+            return self.id_predictor(h)
+        else:
+            return self.out(h)

ldm/modules/diffusionmodules/model.py

@@ -0,0 +1,835 @@
+# pytorch_diffusion + derived encoder decoder
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import rearrange
+
+from ldm.util import instantiate_from_config
+from ldm.modules.attention import LinearAttention
+
+
+def get_timestep_embedding(timesteps, embedding_dim):
+    """
+    This matches the implementation in Denoising Diffusion Probabilistic Models:
+    From Fairseq.
+    Build sinusoidal embeddings.
+    This matches the implementation in tensor2tensor, but differs slightly
+    from the description in Section 3.5 of "Attention Is All You Need".
+    """
+    assert len(timesteps.shape) == 1
+
+    half_dim = embedding_dim // 2
+    emb = math.log(10000) / (half_dim - 1)
+    emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
+    emb = emb.to(device=timesteps.device)
+    emb = timesteps.float()[:, None] * emb[None, :]
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+    if embedding_dim % 2 == 1:  # zero pad
+        emb = torch.nn.functional.pad(emb, (0,1,0,0))
+    return emb
+
+
+def nonlinearity(x):
+    # swish
+    return x*torch.sigmoid(x)
+
+
+def Normalize(in_channels, num_groups=32):
+    return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=3,
+                                        stride=2,
+                                        padding=0)
+
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0,1,0,1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
+                 dropout, temb_channels=512):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = torch.nn.Conv2d(in_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if temb_channels > 0:
+            self.temb_proj = torch.nn.Linear(temb_channels,
+                                             out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = torch.nn.Conv2d(out_channels,
+                                     out_channels,
+                                     kernel_size=3,
+                                     stride=1,
+                                     padding=1)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = torch.nn.Conv2d(in_channels,
+                                                     out_channels,
+                                                     kernel_size=3,
+                                                     stride=1,
+                                                     padding=1)
+            else:
+                self.nin_shortcut = torch.nn.Conv2d(in_channels,
+                                                    out_channels,
+                                                    kernel_size=1,
+                                                    stride=1,
+                                                    padding=0)
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None]
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x+h
+
+
+class LinAttnBlock(LinearAttention):
+    """to match AttnBlock usage"""
+    def __init__(self, in_channels):
+        super().__init__(dim=in_channels, heads=1, dim_head=in_channels)
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.k = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.v = torch.nn.Conv2d(in_channels,
+                                 in_channels,
+                                 kernel_size=1,
+                                 stride=1,
+                                 padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=1,
+                                        stride=1,
+                                        padding=0)
+
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b,c,h,w = q.shape
+        q = q.reshape(b,c,h*w)
+        q = q.permute(0,2,1)   # b,hw,c
+        k = k.reshape(b,c,h*w) # b,c,hw
+        w_ = torch.bmm(q,k)     # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
+        w_ = w_ * (int(c)**(-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b,c,h*w)
+        w_ = w_.permute(0,2,1)   # b,hw,hw (first hw of k, second of q)
+        h_ = torch.bmm(v,w_)     # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
+        h_ = h_.reshape(b,c,h,w)
+
+        h_ = self.proj_out(h_)
+
+        return x+h_
+
+
+def make_attn(in_channels, attn_type="vanilla"):
+    assert attn_type in ["vanilla", "linear", "none"], f'attn_type {attn_type} unknown'
+    print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
+    if attn_type == "vanilla":
+        return AttnBlock(in_channels)
+    elif attn_type == "none":
+        return nn.Identity(in_channels)
+    else:
+        return LinAttnBlock(in_channels)
+
+
+class Model(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = self.ch*4
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        self.use_timestep = use_timestep
+        if self.use_timestep:
+            # timestep embedding
+            self.temb = nn.Module()
+            self.temb.dense = nn.ModuleList([
+                torch.nn.Linear(self.ch,
+                                self.temb_ch),
+                torch.nn.Linear(self.temb_ch,
+                                self.temb_ch),
+            ])
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            skip_in = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                if i_block == self.num_res_blocks:
+                    skip_in = ch*in_ch_mult[i_level]
+                block.append(ResnetBlock(in_channels=block_in+skip_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x, t=None, context=None):
+        #assert x.shape[2] == x.shape[3] == self.resolution
+        if context is not None:
+            # assume aligned context, cat along channel axis
+            x = torch.cat((x, context), dim=1)
+        if self.use_timestep:
+            # timestep embedding
+            assert t is not None
+            temb = get_timestep_embedding(t, self.ch)
+            temb = self.temb.dense[0](temb)
+            temb = nonlinearity(temb)
+            temb = self.temb.dense[1](temb)
+        else:
+            temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](
+                    torch.cat([h, hs.pop()], dim=1), temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+    def get_last_layer(self):
+        return self.conv_out.weight
+
+
+class Encoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla",
+                 **ignore_kwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels,
+                                       self.ch,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,)+tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch*in_ch_mult[i_level]
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))# vanilla attention
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions-1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # end
+        self.norm_out = Normalize(block_in)# GroupNorm
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        2*z_channels if double_z else z_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # timestep embedding
+        temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions-1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
+                 resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
+                 attn_type="vanilla", **ignorekwargs):
+        super().__init__()
+        if use_linear_attn: attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        in_ch_mult = (1,)+tuple(ch_mult)
+        block_in = ch*ch_mult[self.num_resolutions-1]
+        curr_res = resolution // 2**(self.num_resolutions-1)
+        self.z_shape = (1,z_channels,curr_res,curr_res)
+        print("Working with z of shape {} = {} dimensions.".format(
+            self.z_shape, np.prod(self.z_shape)))
+
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(z_channels,
+                                       block_in,
+                                       kernel_size=3,
+                                       stride=1,
+                                       padding=1)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in,
+                                       out_channels=block_in,
+                                       temb_channels=self.temb_ch,
+                                       dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch*ch_mult[i_level]
+            for i_block in range(self.num_res_blocks+1):
+                block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up) # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_ch,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, z):
+        #assert z.shape[1:] == self.z_shape[1:]
+        self.last_z_shape = z.shape
+
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks+1):
+                h = self.up[i_level].block[i_block](h, temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        if self.tanh_out:
+            h = torch.tanh(h)
+        return h
+
+
+class SimpleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, *args, **kwargs):
+        super().__init__()
+        self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1),
+                                     ResnetBlock(in_channels=in_channels,
+                                                 out_channels=2 * in_channels,
+                                                 temb_channels=0, dropout=0.0),
+                                     ResnetBlock(in_channels=2 * in_channels,
+                                                out_channels=4 * in_channels,
+                                                temb_channels=0, dropout=0.0),
+                                     ResnetBlock(in_channels=4 * in_channels,
+                                                out_channels=2 * in_channels,
+                                                temb_channels=0, dropout=0.0),
+                                     nn.Conv2d(2*in_channels, in_channels, 1),
+                                     Upsample(in_channels, with_conv=True)])
+        # end
+        self.norm_out = Normalize(in_channels)
+        self.conv_out = torch.nn.Conv2d(in_channels,
+                                        out_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        for i, layer in enumerate(self.model):
+            if i in [1,2,3]:
+                x = layer(x, None)
+            else:
+                x = layer(x)
+
+        h = self.norm_out(x)
+        h = nonlinearity(h)
+        x = self.conv_out(h)
+        return x
+
+
+class UpsampleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution,
+                 ch_mult=(2,2), dropout=0.0):
+        super().__init__()
+        # upsampling
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        block_in = in_channels
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.res_blocks = nn.ModuleList()
+        self.upsample_blocks = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            res_block = []
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                res_block.append(ResnetBlock(in_channels=block_in,
+                                         out_channels=block_out,
+                                         temb_channels=self.temb_ch,
+                                         dropout=dropout))
+                block_in = block_out
+            self.res_blocks.append(nn.ModuleList(res_block))
+            if i_level != self.num_resolutions - 1:
+                self.upsample_blocks.append(Upsample(block_in, True))
+                curr_res = curr_res * 2
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in,
+                                        out_channels,
+                                        kernel_size=3,
+                                        stride=1,
+                                        padding=1)
+
+    def forward(self, x):
+        # upsampling
+        h = x
+        for k, i_level in enumerate(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.res_blocks[i_level][i_block](h, None)
+            if i_level != self.num_resolutions - 1:
+                h = self.upsample_blocks[k](h)
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class LatentRescaler(nn.Module):
+    def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2):
+        super().__init__()
+        # residual block, interpolate, residual block
+        self.factor = factor
+        self.conv_in = nn.Conv2d(in_channels,
+                                 mid_channels,
+                                 kernel_size=3,
+                                 stride=1,
+                                 padding=1)
+        self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
+                                                     out_channels=mid_channels,
+                                                     temb_channels=0,
+                                                     dropout=0.0) for _ in range(depth)])
+        self.attn = AttnBlock(mid_channels)
+        self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
+                                                     out_channels=mid_channels,
+                                                     temb_channels=0,
+                                                     dropout=0.0) for _ in range(depth)])
+
+        self.conv_out = nn.Conv2d(mid_channels,
+                                  out_channels,
+                                  kernel_size=1,
+                                  )
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        for block in self.res_block1:
+            x = block(x, None)
+        x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor))))
+        x = self.attn(x)
+        for block in self.res_block2:
+            x = block(x, None)
+        x = self.conv_out(x)
+        return x
+
+
+class MergedRescaleEncoder(nn.Module):
+    def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks,
+                 attn_resolutions, dropout=0.0, resamp_with_conv=True,
+                 ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1):
+        super().__init__()
+        intermediate_chn = ch * ch_mult[-1]
+        self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult,
+                               z_channels=intermediate_chn, double_z=False, resolution=resolution,
+                               attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv,
+                               out_ch=None)
+        self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn,
+                                       mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth)
+
+    def forward(self, x):
+        x = self.encoder(x)
+        x = self.rescaler(x)
+        return x
+
+
+class MergedRescaleDecoder(nn.Module):
+    def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8),
+                 dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1):
+        super().__init__()
+        tmp_chn = z_channels*ch_mult[-1]
+        self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout,
+                               resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks,
+                               ch_mult=ch_mult, resolution=resolution, ch=ch)
+        self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn,
+                                       out_channels=tmp_chn, depth=rescale_module_depth)
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Upsampler(nn.Module):
+    def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2):
+        super().__init__()
+        assert out_size >= in_size
+        num_blocks = int(np.log2(out_size//in_size))+1
+        factor_up = 1.+ (out_size % in_size)
+        print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}")
+        self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels,
+                                       out_channels=in_channels)
+        self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2,
+                               attn_resolutions=[], in_channels=None, ch=in_channels,
+                               ch_mult=[ch_mult for _ in range(num_blocks)])
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Resize(nn.Module):
+    def __init__(self, in_channels=None, learned=False, mode="bilinear"):
+        super().__init__()
+        self.with_conv = learned
+        self.mode = mode
+        if self.with_conv:
+            print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode")
+            raise NotImplementedError()
+            assert in_channels is not None
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels,
+                                        in_channels,
+                                        kernel_size=4,
+                                        stride=2,
+                                        padding=1)
+
+    def forward(self, x, scale_factor=1.0):
+        if scale_factor==1.0:
+            return x
+        else:
+            x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor)
+        return x
+
+class FirstStagePostProcessor(nn.Module):
+
+    def __init__(self, ch_mult:list, in_channels,
+                 pretrained_model:nn.Module=None,
+                 reshape=False,
+                 n_channels=None,
+                 dropout=0.,
+                 pretrained_config=None):
+        super().__init__()
+        if pretrained_config is None:
+            assert pretrained_model is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.pretrained_model = pretrained_model
+        else:
+            assert pretrained_config is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
+            self.instantiate_pretrained(pretrained_config)
+
+        self.do_reshape = reshape
+
+        if n_channels is None:
+            n_channels = self.pretrained_model.encoder.ch
+
+        self.proj_norm = Normalize(in_channels,num_groups=in_channels//2)
+        self.proj = nn.Conv2d(in_channels,n_channels,kernel_size=3,
+                            stride=1,padding=1)
+
+        blocks = []
+        downs = []
+        ch_in = n_channels
+        for m in ch_mult:
+            blocks.append(ResnetBlock(in_channels=ch_in,out_channels=m*n_channels,dropout=dropout))
+            ch_in = m * n_channels
+            downs.append(Downsample(ch_in, with_conv=False))
+
+        self.model = nn.ModuleList(blocks)
+        self.downsampler = nn.ModuleList(downs)
+
+
+    def instantiate_pretrained(self, config):
+        model = instantiate_from_config(config)
+        self.pretrained_model = model.eval()
+        # self.pretrained_model.train = False
+        for param in self.pretrained_model.parameters():
+            param.requires_grad = False
+
+
+    @torch.no_grad()
+    def encode_with_pretrained(self,x):
+        c = self.pretrained_model.encode(x)
+        if isinstance(c, DiagonalGaussianDistribution):
+            c = c.mode()
+        return  c
+
+    def forward(self,x):
+        z_fs = self.encode_with_pretrained(x)
+        z = self.proj_norm(z_fs)
+        z = self.proj(z)
+        z = nonlinearity(z)
+
+        for submodel, downmodel in zip(self.model,self.downsampler):
+            z = submodel(z,temb=None)
+            z = downmodel(z)
+
+        if self.do_reshape:
+            z = rearrange(z,'b c h w -> b (h w) c')
+        return z
+

ldm/modules/diffusionmodules/openaimodel.py

@@ -0,0 +1,963 @@
+from abc import abstractmethod
+from functools import partial
+import math
+from typing import Iterable
+
+import numpy as np
+import torch as th
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ldm.modules.diffusionmodules.util import (
+    checkpoint,
+    conv_nd,
+    linear,
+    avg_pool_nd,
+    zero_module,
+    normalization,
+    timestep_embedding,
+)
+from ldm.modules.attention import SpatialTransformer
+
+
+# dummy replace
+def convert_module_to_f16(x):
+    pass
+
+def convert_module_to_f32(x):
+    pass
+
+
+## go
+class AttentionPool2d(nn.Module):
+    """
+    Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py
+    """
+
+    def __init__(
+        self,
+        spacial_dim: int,
+        embed_dim: int,
+        num_heads_channels: int,
+        output_dim: int = None,
+    ):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(th.randn(embed_dim, spacial_dim ** 2 + 1) / embed_dim ** 0.5)
+        self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1)
+        self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1)
+        self.num_heads = embed_dim // num_heads_channels
+        self.attention = QKVAttention(self.num_heads)
+
+    def forward(self, x):
+        b, c, *_spatial = x.shape
+        x = x.reshape(b, c, -1)  # NC(HW)
+        x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1)  # NC(HW+1)
+        x = x + self.positional_embedding[None, :, :].to(x.dtype)  # NC(HW+1)
+        x = self.qkv_proj(x)
+        x = self.attention(x)
+        x = self.c_proj(x)
+        return x[:, :, 0]
+
+
+class TimestepBlock(nn.Module):
+    """
+    Any module where forward() takes timestep embeddings as a second argument.
+    """
+
+    @abstractmethod
+    def forward(self, x, emb):
+        """
+        Apply the module to `x` given `emb` timestep embeddings.
+        """
+
+
+class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
+    """
+    A sequential module that passes timestep embeddings to the children that
+    support it as an extra input.
+    """
+
+    def forward(self, x, emb, context=None):
+        for layer in self:
+            if isinstance(layer, TimestepBlock):
+                x = layer(x, emb)
+            elif isinstance(layer, SpatialTransformer):
+                x = layer(x, context)
+            else:
+                x = layer(x)
+        return x
+
+
+class Upsample(nn.Module):
+    """
+    An upsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 upsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        if use_conv:
+            self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        if self.dims == 3:
+            x = F.interpolate(
+                x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest"
+            )
+        else:
+            x = F.interpolate(x, scale_factor=2, mode="nearest")
+        if self.use_conv:
+            x = self.conv(x)
+        return x
+
+class TransposedUpsample(nn.Module):
+    'Learned 2x upsampling without padding'
+    def __init__(self, channels, out_channels=None, ks=5):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+
+        self.up = nn.ConvTranspose2d(self.channels,self.out_channels,kernel_size=ks,stride=2)
+
+    def forward(self,x):
+        return self.up(x)
+
+
+class Downsample(nn.Module):
+    """
+    A downsampling layer with an optional convolution.
+    :param channels: channels in the inputs and outputs.
+    :param use_conv: a bool determining if a convolution is applied.
+    :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
+                 downsampling occurs in the inner-two dimensions.
+    """
+
+    def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.dims = dims
+        stride = 2 if dims != 3 else (1, 2, 2)
+        if use_conv:
+            self.op = conv_nd(
+                dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
+            )
+        else:
+            assert self.channels == self.out_channels
+            self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
+
+    def forward(self, x):
+        assert x.shape[1] == self.channels
+        return self.op(x)
+
+
+class ResBlock(TimestepBlock):
+    """
+    A residual block that can optionally change the number of channels.
+    :param channels: the number of input channels.
+    :param emb_channels: the number of timestep embedding channels.
+    :param dropout: the rate of dropout.
+    :param out_channels: if specified, the number of out channels.
+    :param use_conv: if True and out_channels is specified, use a spatial
+        convolution instead of a smaller 1x1 convolution to change the
+        channels in the skip connection.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param use_checkpoint: if True, use gradient checkpointing on this module.
+    :param up: if True, use this block for upsampling.
+    :param down: if True, use this block for downsampling.
+    """
+
+    def __init__(
+        self,
+        channels,
+        emb_channels,
+        dropout,
+        out_channels=None,
+        use_conv=False,
+        use_scale_shift_norm=False,
+        dims=2,
+        use_checkpoint=False,
+        up=False,
+        down=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.dropout = dropout
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_checkpoint = use_checkpoint
+        self.use_scale_shift_norm = use_scale_shift_norm
+
+        self.in_layers = nn.Sequential(
+            normalization(channels),
+            nn.SiLU(),
+            conv_nd(dims, channels, self.out_channels, 3, padding=1),
+        )
+
+        self.updown = up or down
+
+        if up:
+            self.h_upd = Upsample(channels, False, dims)
+            self.x_upd = Upsample(channels, False, dims)
+        elif down:
+            self.h_upd = Downsample(channels, False, dims)
+            self.x_upd = Downsample(channels, False, dims)
+        else:
+            self.h_upd = self.x_upd = nn.Identity()
+
+        self.emb_layers = nn.Sequential(
+            nn.SiLU(),
+            linear(
+                emb_channels,
+                2 * self.out_channels if use_scale_shift_norm else self.out_channels,
+            ),
+        )
+        self.out_layers = nn.Sequential(
+            normalization(self.out_channels),
+            nn.SiLU(),
+            nn.Dropout(p=dropout),
+            zero_module(
+                conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1)
+            ),
+        )
+
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        elif use_conv:
+            self.skip_connection = conv_nd(
+                dims, channels, self.out_channels, 3, padding=1
+            )
+        else:
+            self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
+
+    def forward(self, x, emb):
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+        :param x: an [N x C x ...] Tensor of features.
+        :param emb: an [N x emb_channels] Tensor of timestep embeddings.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        return checkpoint(
+            self._forward, (x, emb), self.parameters(), self.use_checkpoint
+        )
+
+
+    def _forward(self, x, emb):
+        if self.updown:
+            in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
+            h = in_rest(x)
+            h = self.h_upd(h)
+            x = self.x_upd(x)
+            h = in_conv(h)
+        else:
+            h = self.in_layers(x)
+        emb_out = self.emb_layers(emb).type(h.dtype)
+        while len(emb_out.shape) < len(h.shape):
+            emb_out = emb_out[..., None]
+        if self.use_scale_shift_norm:
+            out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
+            scale, shift = th.chunk(emb_out, 2, dim=1)
+            h = out_norm(h) * (1 + scale) + shift
+            h = out_rest(h)
+        else:
+            h = h + emb_out
+            h = self.out_layers(h)
+        return self.skip_connection(x) + h
+
+
+class AttentionBlock(nn.Module):
+    """
+    An attention block that allows spatial positions to attend to each other.
+    Originally ported from here, but adapted to the N-d case.
+    https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
+    """
+
+    def __init__(
+        self,
+        channels,
+        num_heads=1,
+        num_head_channels=-1,
+        use_checkpoint=False,
+        use_new_attention_order=False,
+    ):
+        super().__init__()
+        self.channels = channels
+        if num_head_channels == -1:
+            self.num_heads = num_heads
+        else:
+            assert (
+                channels % num_head_channels == 0
+            ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}"
+            self.num_heads = channels // num_head_channels
+        self.use_checkpoint = use_checkpoint
+        self.norm = normalization(channels)
+        self.qkv = conv_nd(1, channels, channels * 3, 1)
+        if use_new_attention_order:
+            # split qkv before split heads
+            self.attention = QKVAttention(self.num_heads)
+        else:
+            # split heads before split qkv
+            self.attention = QKVAttentionLegacy(self.num_heads)
+
+        self.proj_out = zero_module(conv_nd(1, channels, channels, 1))
+
+    def forward(self, x):
+        return checkpoint(self._forward, (x,), self.parameters(), True)   # TODO: check checkpoint usage, is True # TODO: fix the .half call!!!
+        #return pt_checkpoint(self._forward, x)  # pytorch
+
+    def _forward(self, x):
+        b, c, *spatial = x.shape
+        x = x.reshape(b, c, -1)
+        qkv = self.qkv(self.norm(x))
+        h = self.attention(qkv)
+        h = self.proj_out(h)
+        return (x + h).reshape(b, c, *spatial)
+
+
+def count_flops_attn(model, _x, y):
+    """
+    A counter for the `thop` package to count the operations in an
+    attention operation.
+    Meant to be used like:
+        macs, params = thop.profile(
+            model,
+            inputs=(inputs, timestamps),
+            custom_ops={QKVAttention: QKVAttention.count_flops},
+        )
+    """
+    b, c, *spatial = y[0].shape
+    num_spatial = int(np.prod(spatial))
+    # We perform two matmuls with the same number of ops.
+    # The first computes the weight matrix, the second computes
+    # the combination of the value vectors.
+    matmul_ops = 2 * b * (num_spatial ** 2) * c
+    model.total_ops += th.DoubleTensor([matmul_ops])
+
+
+class QKVAttentionLegacy(nn.Module):
+    """
+    A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping
+    """
+
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+        :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1)
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = th.einsum(
+            "bct,bcs->bts", q * scale, k * scale
+        )  # More stable with f16 than dividing afterwards
+        weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = th.einsum("bts,bcs->bct", weight, v)
+        return a.reshape(bs, -1, length)
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+
+
+class QKVAttention(nn.Module):
+    """
+    A module which performs QKV attention and splits in a different order.
+    """
+
+    def __init__(self, n_heads):
+        super().__init__()
+        self.n_heads = n_heads
+
+    def forward(self, qkv):
+        """
+        Apply QKV attention.
+        :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs.
+        :return: an [N x (H * C) x T] tensor after attention.
+        """
+        bs, width, length = qkv.shape
+        assert width % (3 * self.n_heads) == 0
+        ch = width // (3 * self.n_heads)
+        q, k, v = qkv.chunk(3, dim=1)
+        scale = 1 / math.sqrt(math.sqrt(ch))
+        weight = th.einsum(
+            "bct,bcs->bts",
+            (q * scale).view(bs * self.n_heads, ch, length),
+            (k * scale).view(bs * self.n_heads, ch, length),
+        )  # More stable with f16 than dividing afterwards
+        weight = th.softmax(weight.float(), dim=-1).type(weight.dtype)
+        a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length))
+        return a.reshape(bs, -1, length)
+
+    @staticmethod
+    def count_flops(model, _x, y):
+        return count_flops_attn(model, _x, y)
+
+
+class UNetModel(nn.Module):
+    """
+    The full UNet model with attention and timestep embedding.
+    :param in_channels: channels in the input Tensor.
+    :param model_channels: base channel count for the model.
+    :param out_channels: channels in the output Tensor.
+    :param num_res_blocks: number of residual blocks per downsample.
+    :param attention_resolutions: a collection of downsample rates at which
+        attention will take place. May be a set, list, or tuple.
+        For example, if this contains 4, then at 4x downsampling, attention
+        will be used.
+    :param dropout: the dropout probability.
+    :param channel_mult: channel multiplier for each level of the UNet.
+    :param conv_resample: if True, use learned convolutions for upsampling and
+        downsampling.
+    :param dims: determines if the signal is 1D, 2D, or 3D.
+    :param num_classes: if specified (as an int), then this model will be
+        class-conditional with `num_classes` classes.
+    :param use_checkpoint: use gradient checkpointing to reduce memory usage.
+    :param num_heads: the number of attention heads in each attention layer.
+    :param num_heads_channels: if specified, ignore num_heads and instead use
+                               a fixed channel width per attention head.
+    :param num_heads_upsample: works with num_heads to set a different number
+                               of heads for upsampling. Deprecated.
+    :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
+    :param resblock_updown: use residual blocks for up/downsampling.
+    :param use_new_attention_order: use a different attention pattern for potentially
+                                    increased efficiency.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        num_classes=None,
+        use_checkpoint=False,
+        use_fp16=False,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        use_spatial_transformer=False,    # custom transformer support
+        transformer_depth=1,              # custom transformer support
+        context_dim=None,                 # custom transformer support
+        n_embed=None,                     # custom support for prediction of discrete ids into codebook of first stage vq model
+        legacy=True,
+    ):
+        super().__init__()
+        if use_spatial_transformer:
+            assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
+
+        if context_dim is not None:
+            assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
+            from omegaconf.listconfig import ListConfig
+            if type(context_dim) == ListConfig:
+                context_dim = list(context_dim)
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        if num_heads == -1:
+            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
+
+        if num_head_channels == -1:
+            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
+
+        self.image_size = image_size
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        self.num_res_blocks = num_res_blocks
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.num_classes = num_classes
+        self.use_checkpoint = use_checkpoint
+        self.dtype = th.float16 if use_fp16 else th.float32
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.predict_codebook_ids = n_embed is not None
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+
+        if self.num_classes is not None:
+            self.label_emb = nn.Embedding(num_classes, time_embed_dim)
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    conv_nd(dims, in_channels, model_channels, 3, padding=1)# conv2d for txt2img/audio
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        # downsample blocks
+        for level, mult in enumerate(channel_mult):
+            for _ in range(num_res_blocks):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    layers.append(
+                        AttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads,
+                            num_head_channels=dim_head,
+                            use_new_attention_order=use_new_attention_order,
+                        ) if not use_spatial_transformer else SpatialTransformer(# transformer_depth is 1
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
+                        )
+                    )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        if legacy:
+            #num_heads = 1
+            dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            AttentionBlock(
+                ch,
+                use_checkpoint=use_checkpoint,
+                num_heads=num_heads,
+                num_head_channels=dim_head,
+                use_new_attention_order=use_new_attention_order,
+            ) if not use_spatial_transformer else SpatialTransformer(
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
+                        ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self._feature_size += ch
+        # upsample blocks
+        self.output_blocks = nn.ModuleList([])
+        for level, mult in list(enumerate(channel_mult))[::-1]:
+            for i in range(num_res_blocks + 1):
+                ich = input_block_chans.pop()
+                layers = [
+                    ResBlock(
+                        ch + ich,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=model_channels * mult,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = model_channels * mult
+                if ds in attention_resolutions:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    layers.append(
+                        AttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads_upsample,
+                            num_head_channels=dim_head,
+                            use_new_attention_order=use_new_attention_order,
+                        ) if not use_spatial_transformer else SpatialTransformer(
+                            ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim
+                        )
+                    )
+                if level and i == num_res_blocks:
+                    out_ch = ch
+                    layers.append(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            up=True,
+                        )
+                        if resblock_updown
+                        else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
+                    )
+                    ds //= 2
+                self.output_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+
+        self.out = nn.Sequential(
+            normalization(ch),
+            nn.SiLU(),
+            zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
+        )
+        if self.predict_codebook_ids:
+            self.id_predictor = nn.Sequential(
+            normalization(ch),
+            conv_nd(dims, model_channels, n_embed, 1),
+            #nn.LogSoftmax(dim=1)  # change to cross_entropy and produce non-normalized logits
+        )
+
+    def convert_to_fp16(self):
+        """
+        Convert the torso of the model to float16.
+        """
+        self.input_blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+        self.output_blocks.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self):
+        """
+        Convert the torso of the model to float32.
+        """
+        self.input_blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+        self.output_blocks.apply(convert_module_to_f32)
+
+    def forward(self, x, timesteps=None, context=None, y=None,**kwargs):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps,shape [N]
+        :param context: conditioning plugged in via crossattn. for txt2img shape is [N,77,context_dim]
+        :param y: an [N] Tensor of labels, if class-conditional.
+        :return: an [N x C x ...] Tensor of outputs.
+        """
+        # print(f"in unet {x.shape}")
+        assert (y is not None) == (
+            self.num_classes is not None
+        ), "must specify y if and only if the model is class-conditional"
+        hs = []
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)# shape [N,self.model_channels]
+        emb = self.time_embed(t_emb)# shape [N,context_dim]
+
+        if self.num_classes is not None:# only for class label
+            assert y.shape == (x.shape[0],)
+            emb = emb + self.label_emb(y)
+
+        h = x.type(self.dtype)# [N,C,10,106]
+        for module in self.input_blocks:
+            h = module(h, emb, context)# 0:[N,self.model_channels,10,106],1:[N,self.model_channels,10,106],2:[N,self.model_channels,10,106] 3:[N,self.model_channels,5,53] 4:[N,self.model_channels,5,53] 5:[N,self.model_channels*2,5,53]
+            hs.append(h)
+        h = self.middle_block(h, emb, context)# no shape change
+        for module in self.output_blocks:
+            h = th.cat([h, hs.pop()], dim=1)# 在这里c维度乘2或+self.model_channels，其余维度不变
+            h = module(h, emb, context)# 在这里c维度/2回到之前维度，h,w不变或*2
+        h = h.type(x.dtype)# 至此h维度和输入x维度回到相同状态
+        if self.predict_codebook_ids:
+            return self.id_predictor(h)
+        else:
+            return self.out(h)
+
+
+class EncoderUNetModel(nn.Module):
+    """
+    The half UNet model with attention and timestep embedding.
+    For usage, see UNet.
+    """
+
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        out_channels,
+        num_res_blocks,
+        attention_resolutions,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        use_checkpoint=False,
+        use_fp16=False,
+        num_heads=1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        pool="adaptive",
+        *args,
+        **kwargs
+    ):
+        super().__init__()
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.out_channels = out_channels
+        self.num_res_blocks = num_res_blocks
+        self.attention_resolutions = attention_resolutions
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.use_checkpoint = use_checkpoint
+        self.dtype = th.float16 if use_fp16 else th.float32
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            linear(model_channels, time_embed_dim),
+            nn.SiLU(),
+            linear(time_embed_dim, time_embed_dim),
+        )
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    conv_nd(dims, in_channels, model_channels, 3, padding=1)
+                )
+            ]
+        )
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for _ in range(num_res_blocks):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                    )
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    layers.append(
+                        AttentionBlock(
+                            ch,
+                            use_checkpoint=use_checkpoint,
+                            num_heads=num_heads,
+                            num_head_channels=num_head_channels,
+                            use_new_attention_order=use_new_attention_order,
+                        )
+                    )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                ds *= 2
+                self._feature_size += ch
+
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+            AttentionBlock(
+                ch,
+                use_checkpoint=use_checkpoint,
+                num_heads=num_heads,
+                num_head_channels=num_head_channels,
+                use_new_attention_order=use_new_attention_order,
+            ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+            ),
+        )
+        self._feature_size += ch
+        self.pool = pool
+        if pool == "adaptive":
+            self.out = nn.Sequential(
+                normalization(ch),
+                nn.SiLU(),
+                nn.AdaptiveAvgPool2d((1, 1)),
+                zero_module(conv_nd(dims, ch, out_channels, 1)),
+                nn.Flatten(),
+            )
+        elif pool == "attention":
+            assert num_head_channels != -1
+            self.out = nn.Sequential(
+                normalization(ch),
+                nn.SiLU(),
+                AttentionPool2d(
+                    (image_size // ds), ch, num_head_channels, out_channels
+                ),
+            )
+        elif pool == "spatial":
+            self.out = nn.Sequential(
+                nn.Linear(self._feature_size, 2048),
+                nn.ReLU(),
+                nn.Linear(2048, self.out_channels),
+            )
+        elif pool == "spatial_v2":
+            self.out = nn.Sequential(
+                nn.Linear(self._feature_size, 2048),
+                normalization(2048),
+                nn.SiLU(),
+                nn.Linear(2048, self.out_channels),
+            )
+        else:
+            raise NotImplementedError(f"Unexpected {pool} pooling")
+
+    def convert_to_fp16(self):
+        """
+        Convert the torso of the model to float16.
+        """
+        self.input_blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self):
+        """
+        Convert the torso of the model to float32.
+        """
+        self.input_blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+
+    def forward(self, x, timesteps):
+        """
+        Apply the model to an input batch.
+        :param x: an [N x C x ...] Tensor of inputs.
+        :param timesteps: a 1-D batch of timesteps.
+        :return: an [N x K] Tensor of outputs.
+        """
+        emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
+
+        results = []
+        h = x.type(self.dtype)
+        for module in self.input_blocks:
+            h = module(h, emb)
+            if self.pool.startswith("spatial"):
+                results.append(h.type(x.dtype).mean(dim=(2, 3)))
+        h = self.middle_block(h, emb)
+        if self.pool.startswith("spatial"):
+            results.append(h.type(x.dtype).mean(dim=(2, 3)))
+            h = th.cat(results, axis=-1)
+            return self.out(h)
+        else:
+            h = h.type(x.dtype)
+            return self.out(h)
+

ldm/modules/diffusionmodules/util.py

@@ -0,0 +1,267 @@
+# adopted from
+# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
+# and
+# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
+# and
+# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
+#
+# thanks!
+
+
+import os
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import repeat
+
+from ldm.util import instantiate_from_config
+
+
+def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
+    if schedule == "linear":
+        betas = (
+                torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
+        )
+
+    elif schedule == "cosine":
+        timesteps = (
+                torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
+        )
+        alphas = timesteps / (1 + cosine_s) * np.pi / 2
+        alphas = torch.cos(alphas).pow(2)
+        alphas = alphas / alphas[0]
+        betas = 1 - alphas[1:] / alphas[:-1]
+        betas = np.clip(betas, a_min=0, a_max=0.999)
+
+    elif schedule == "sqrt_linear":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
+    elif schedule == "sqrt":
+        betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5
+    else:
+        raise ValueError(f"schedule '{schedule}' unknown.")
+    return betas.numpy()
+
+
+def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True):
+    if ddim_discr_method == 'uniform':
+        c = num_ddpm_timesteps // num_ddim_timesteps
+        ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c)))
+    elif ddim_discr_method == 'quad':
+        ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int)
+    else:
+        raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"')
+
+    # assert ddim_timesteps.shape[0] == num_ddim_timesteps
+    # add one to get the final alpha values right (the ones from first scale to data during sampling)
+    steps_out = ddim_timesteps + 1
+    if verbose:
+        print(f'Selected timesteps for ddim sampler: {steps_out}')
+    return steps_out
+
+
+def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True):
+    # select alphas for computing the variance schedule
+    alphas = alphacums[ddim_timesteps]
+    alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
+
+    # according the the formula provided in https://arxiv.org/abs/2010.02502
+    sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev))
+    if verbose:
+        print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}')
+        print(f'For the chosen value of eta, which is {eta}, '
+              f'this results in the following sigma_t schedule for ddim sampler {sigmas}')
+    return sigmas, alphas, alphas_prev
+
+
+def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function,
+    which defines the cumulative product of (1-beta) over time from t = [0,1].
+    :param num_diffusion_timesteps: the number of betas to produce.
+    :param alpha_bar: a lambda that takes an argument t from 0 to 1 and
+                      produces the cumulative product of (1-beta) up to that
+                      part of the diffusion process.
+    :param max_beta: the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+    """
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
+    return np.array(betas)
+
+
+def extract_into_tensor(a, t, x_shape):
+    b, *_ = t.shape
+    out = a.gather(-1, t)
+    return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+
+def checkpoint(func, inputs, params, flag):
+    """
+    Evaluate a function without caching intermediate activations, allowing for
+    reduced memory at the expense of extra compute in the backward pass.
+    :param func: the function to evaluate.
+    :param inputs: the argument sequence to pass to `func`.
+    :param params: a sequence of parameters `func` depends on but does not
+                   explicitly take as arguments.
+    :param flag: if False, disable gradient checkpointing.
+    """
+    if flag:
+        args = tuple(inputs) + tuple(params)
+        return CheckpointFunction.apply(func, len(inputs), *args)
+    else:
+        return func(*inputs)
+
+
+class CheckpointFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, run_function, length, *args):
+        ctx.run_function = run_function
+        ctx.input_tensors = list(args[:length])
+        ctx.input_params = list(args[length:])
+
+        with torch.no_grad():
+            output_tensors = ctx.run_function(*ctx.input_tensors)
+        return output_tensors
+
+    @staticmethod
+    def backward(ctx, *output_grads):
+        ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
+        with torch.enable_grad():
+            # Fixes a bug where the first op in run_function modifies the
+            # Tensor storage in place, which is not allowed for detach()'d
+            # Tensors.
+            shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
+            output_tensors = ctx.run_function(*shallow_copies)
+        input_grads = torch.autograd.grad(
+            output_tensors,
+            ctx.input_tensors + ctx.input_params,
+            output_grads,
+            allow_unused=True,
+        )
+        del ctx.input_tensors
+        del ctx.input_params
+        del output_tensors
+        return (None, None) + input_grads
+
+
+def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    if not repeat_only:
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+        ).to(device=timesteps.device)
+        args = timesteps[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    else:
+        embedding = repeat(timesteps, 'b -> b d', d=dim)
+    return embedding
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def scale_module(module, scale):
+    """
+    Scale the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().mul_(scale)
+    return module
+
+
+def mean_flat(tensor):
+    """
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+
+def normalization(channels):
+    """
+    Make a standard normalization layer.
+    :param channels: number of input channels.
+    :return: an nn.Module for normalization.
+    """
+    return GroupNorm32(32, channels)
+
+
+# PyTorch 1.7 has SiLU, but we support PyTorch 1.5.
+class SiLU(nn.Module):
+    def forward(self, x):
+        return x * torch.sigmoid(x)
+
+
+class GroupNorm32(nn.GroupNorm):
+    def forward(self, x):
+        return super().forward(x.float()).type(x.dtype)
+
+def conv_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D convolution module.
+    """
+    if dims == 1:
+        return nn.Conv1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.Conv2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.Conv3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+def linear(*args, **kwargs):
+    """
+    Create a linear module.
+    """
+    return nn.Linear(*args, **kwargs)
+
+
+def avg_pool_nd(dims, *args, **kwargs):
+    """
+    Create a 1D, 2D, or 3D average pooling module.
+    """
+    if dims == 1:
+        return nn.AvgPool1d(*args, **kwargs)
+    elif dims == 2:
+        return nn.AvgPool2d(*args, **kwargs)
+    elif dims == 3:
+        return nn.AvgPool3d(*args, **kwargs)
+    raise ValueError(f"unsupported dimensions: {dims}")
+
+
+class HybridConditioner(nn.Module):
+
+    def __init__(self, c_concat_config, c_crossattn_config):
+        super().__init__()
+        self.concat_conditioner = instantiate_from_config(c_concat_config)
+        self.crossattn_conditioner = instantiate_from_config(c_crossattn_config)
+
+    def forward(self, c_concat, c_crossattn):
+        c_concat = self.concat_conditioner(c_concat)
+        c_crossattn = self.crossattn_conditioner(c_crossattn)
+        return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]}
+
+
+def noise_like(shape, device, repeat=False):
+    repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
+    noise = lambda: torch.randn(shape, device=device)
+    return repeat_noise() if repeat else noise()

ldm/modules/discriminator/model.py

@@ -0,0 +1,295 @@
+import functools
+import torch.nn as nn
+
+
+class ActNorm(nn.Module):
+    def __init__(self, num_features, logdet=False, affine=True,
+                 allow_reverse_init=False):
+        assert affine
+        super().__init__()
+        self.logdet = logdet
+        self.loc = nn.Parameter(torch.zeros(1, num_features, 1, 1))
+        self.scale = nn.Parameter(torch.ones(1, num_features, 1, 1))
+        self.allow_reverse_init = allow_reverse_init
+
+        self.register_buffer('initialized', torch.tensor(0, dtype=torch.uint8))
+
+    def initialize(self, input):
+        with torch.no_grad():
+            flatten = input.permute(1, 0, 2, 3).contiguous().view(input.shape[1], -1)
+            mean = (
+                flatten.mean(1)
+                .unsqueeze(1)
+                .unsqueeze(2)
+                .unsqueeze(3)
+                .permute(1, 0, 2, 3)
+            )
+            std = (
+                flatten.std(1)
+                .unsqueeze(1)
+                .unsqueeze(2)
+                .unsqueeze(3)
+                .permute(1, 0, 2, 3)
+            )
+
+            self.loc.data.copy_(-mean)
+            self.scale.data.copy_(1 / (std + 1e-6))
+
+    def forward(self, input, reverse=False):
+        if reverse:
+            return self.reverse(input)
+        if len(input.shape) == 2:
+            input = input[:, :, None, None]
+            squeeze = True
+        else:
+            squeeze = False
+
+        _, _, height, width = input.shape
+
+        if self.training and self.initialized.item() == 0:
+            self.initialize(input)
+            self.initialized.fill_(1)
+
+        h = self.scale * (input + self.loc)
+
+        if squeeze:
+            h = h.squeeze(-1).squeeze(-1)
+
+        if self.logdet:
+            log_abs = torch.log(torch.abs(self.scale))
+            logdet = height * width * torch.sum(log_abs)
+            logdet = logdet * torch.ones(input.shape[0]).to(input)
+            return h, logdet
+
+        return h
+
+    def reverse(self, output):
+        if self.training and self.initialized.item() == 0:
+            if not self.allow_reverse_init:
+                raise RuntimeError(
+                    "Initializing ActNorm in reverse direction is "
+                    "disabled by default. Use allow_reverse_init=True to enable."
+                )
+            else:
+                self.initialize(output)
+                self.initialized.fill_(1)
+
+        if len(output.shape) == 2:
+            output = output[:, :, None, None]
+            squeeze = True
+        else:
+            squeeze = False
+
+        h = output / self.scale - self.loc
+
+        if squeeze:
+            h = h.squeeze(-1).squeeze(-1)
+        return h
+
+def weights_init(m):
+    classname = m.__class__.__name__
+    if classname.find('Conv') != -1:
+        nn.init.normal_(m.weight.data, 0.0, 0.02)
+    elif classname.find('BatchNorm') != -1:
+        nn.init.normal_(m.weight.data, 1.0, 0.02)
+        nn.init.constant_(m.bias.data, 0)
+
+
+class NLayerDiscriminator(nn.Module):
+    """Defines a PatchGAN discriminator as in Pix2Pix
+        --> see https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/models/networks.py
+    """
+    def __init__(self, input_nc=3, ndf=64, n_layers=3, use_actnorm=False):
+        """Construct a PatchGAN discriminator
+        Parameters:
+            input_nc (int)  -- the number of channels in input images
+            ndf (int)       -- the number of filters in the last conv layer
+            n_layers (int)  -- the number of conv layers in the discriminator
+            norm_layer      -- normalization layer
+        """
+        super(NLayerDiscriminator, self).__init__()
+        if not use_actnorm:
+            norm_layer = nn.BatchNorm2d
+        else:
+            norm_layer = ActNorm
+        if type(norm_layer) == functools.partial:  # no need to use bias as BatchNorm2d has affine parameters
+            use_bias = norm_layer.func != nn.BatchNorm2d
+        else:
+            use_bias = norm_layer != nn.BatchNorm2d
+
+        kw = 4
+        padw = 1
+        sequence = [nn.Conv2d(input_nc, ndf, kernel_size=kw, stride=2, padding=padw), nn.LeakyReLU(0.2, True)]
+        nf_mult = 1
+        nf_mult_prev = 1
+        for n in range(1, n_layers):  # gradually increase the number of filters
+            nf_mult_prev = nf_mult
+            nf_mult = min(2 ** n, 8)
+            sequence += [
+                nn.Conv2d(ndf * nf_mult_prev, ndf * nf_mult, kernel_size=kw, stride=2, padding=padw, bias=use_bias),
+                norm_layer(ndf * nf_mult),
+                nn.LeakyReLU(0.2, True)
+            ]
+
+        nf_mult_prev = nf_mult
+        nf_mult = min(2 ** n_layers, 8)
+        sequence += [
+            nn.Conv2d(ndf * nf_mult_prev, ndf * nf_mult, kernel_size=kw, stride=1, padding=padw, bias=use_bias),
+            norm_layer(ndf * nf_mult),
+            nn.LeakyReLU(0.2, True)
+        ]
+        # output 1 channel prediction map
+        sequence += [nn.Conv2d(ndf * nf_mult, 1, kernel_size=kw, stride=1, padding=padw)]
+        self.main = nn.Sequential(*sequence)
+
+    def forward(self, input):
+        """Standard forward."""
+        return self.main(input)
+
+class NLayerDiscriminator1dFeats(NLayerDiscriminator):
+    """Defines a PatchGAN discriminator as in Pix2Pix
+        --> see https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/models/networks.py
+    """
+    def __init__(self, input_nc=3, ndf=64, n_layers=3, use_actnorm=False):
+        """Construct a PatchGAN discriminator
+        Parameters:
+            input_nc (int)  -- the number of channels in input feats
+            ndf (int)       -- the number of filters in the last conv layer
+            n_layers (int)  -- the number of conv layers in the discriminator
+            norm_layer      -- normalization layer
+        """
+        super().__init__(input_nc=input_nc, ndf=64, n_layers=n_layers, use_actnorm=use_actnorm)
+
+        if not use_actnorm:
+            norm_layer = nn.BatchNorm1d
+        else:
+            norm_layer = ActNorm
+        if type(norm_layer) == functools.partial:  # no need to use bias as BatchNorm has affine parameters
+            use_bias = norm_layer.func != nn.BatchNorm1d
+        else:
+            use_bias = norm_layer != nn.BatchNorm1d
+
+        kw = 4
+        padw = 1
+        sequence = [nn.Conv1d(input_nc, input_nc//2, kernel_size=kw, stride=2, padding=padw), nn.LeakyReLU(0.2, True)]
+        nf_mult = input_nc//2
+        nf_mult_prev = 1
+        for n in range(1, n_layers):  # gradually decrease the number of filters
+            nf_mult_prev = nf_mult
+            nf_mult = max(nf_mult_prev // (2 ** n), 8)
+            sequence += [
+                nn.Conv1d(nf_mult_prev, nf_mult, kernel_size=kw, stride=2, padding=padw, bias=use_bias),
+                norm_layer(nf_mult),
+                nn.LeakyReLU(0.2, True)
+            ]
+
+        nf_mult_prev = nf_mult
+        nf_mult = max(nf_mult_prev // (2 ** n), 8)
+        sequence += [
+            nn.Conv1d(nf_mult_prev, nf_mult, kernel_size=kw, stride=1, padding=padw, bias=use_bias),
+            norm_layer(nf_mult),
+            nn.LeakyReLU(0.2, True)
+        ]
+        nf_mult_prev = nf_mult
+        nf_mult = max(nf_mult_prev // (2 ** n), 8)
+        sequence += [
+            nn.Conv1d(nf_mult_prev, nf_mult, kernel_size=kw, stride=1, padding=padw, bias=use_bias),
+            norm_layer(nf_mult),
+            nn.LeakyReLU(0.2, True)
+        ]
+        # output 1 channel prediction map
+        sequence += [nn.Conv1d(nf_mult, 1, kernel_size=kw, stride=1, padding=padw)]
+        self.main = nn.Sequential(*sequence)
+
+
+class NLayerDiscriminator1dSpecs(NLayerDiscriminator):
+    """Defines a PatchGAN discriminator as in Pix2Pix
+        --> see https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/models/networks.py
+    """
+    def __init__(self, input_nc=80, ndf=64, n_layers=3, use_actnorm=False):
+        """Construct a PatchGAN discriminator
+        Parameters:
+            input_nc (int)  -- the number of channels in input specs
+            ndf (int)       -- the number of filters in the last conv layer
+            n_layers (int)  -- the number of conv layers in the discriminator
+            norm_layer      -- normalization layer
+        """
+        super().__init__(input_nc=input_nc, ndf=64, n_layers=n_layers, use_actnorm=use_actnorm)
+
+        if not use_actnorm:
+            norm_layer = nn.BatchNorm1d
+        else:
+            norm_layer = ActNorm
+        if type(norm_layer) == functools.partial:  # no need to use bias as BatchNorm has affine parameters
+            use_bias = norm_layer.func != nn.BatchNorm1d
+        else:
+            use_bias = norm_layer != nn.BatchNorm1d
+
+        kw = 4
+        padw = 1
+        sequence = [nn.Conv1d(input_nc, ndf, kernel_size=kw, stride=2, padding=padw), nn.LeakyReLU(0.2, True)]
+        nf_mult = 1
+        nf_mult_prev = 1
+        for n in range(1, n_layers):  # gradually decrease the number of filters
+            nf_mult_prev = nf_mult
+            nf_mult = min(2 ** n, 8)
+            sequence += [
+                nn.Conv1d(ndf * nf_mult_prev, ndf * nf_mult, kernel_size=kw, stride=2, padding=padw, bias=use_bias),
+                norm_layer(ndf * nf_mult),
+                nn.LeakyReLU(0.2, True)
+            ]
+
+        nf_mult_prev = nf_mult
+        nf_mult = min(2 ** n_layers, 8)
+        sequence += [
+            nn.Conv1d(ndf * nf_mult_prev, ndf * nf_mult, kernel_size=kw, stride=1, padding=padw, bias=use_bias),
+            norm_layer(ndf * nf_mult),
+            nn.LeakyReLU(0.2, True)
+        ]
+        # output 1 channel prediction map
+        sequence += [nn.Conv1d(ndf * nf_mult, 1, kernel_size=kw, stride=1, padding=padw)]
+        self.main = nn.Sequential(*sequence)
+
+    def forward(self, input):
+        """Standard forward."""
+        # (B, C, L)
+        input = input.squeeze(1)
+        input = self.main(input)
+        return input
+
+
+if __name__ == '__main__':
+    import torch
+
+    ## FEATURES
+    disc_in_channels = 2048
+    disc_num_layers = 2
+    use_actnorm = False
+    disc_ndf = 64
+    discriminator = NLayerDiscriminator1dFeats(input_nc=disc_in_channels, n_layers=disc_num_layers,
+                                            use_actnorm=use_actnorm, ndf=disc_ndf).apply(weights_init)
+    inputs = torch.rand((6, 2048, 212))
+    outputs = discriminator(inputs)
+    print(outputs.shape)
+
+    ## AUDIO
+    disc_in_channels = 1
+    disc_num_layers = 3
+    use_actnorm = False
+    disc_ndf = 64
+    discriminator = NLayerDiscriminator(input_nc=disc_in_channels, n_layers=disc_num_layers,
+                                        use_actnorm=use_actnorm, ndf=disc_ndf).apply(weights_init)
+    inputs = torch.rand((6, 1, 80, 848))
+    outputs = discriminator(inputs)
+    print(outputs.shape)
+
+    ## IMAGE
+    disc_in_channels = 3
+    disc_num_layers = 3
+    use_actnorm = False
+    disc_ndf = 64
+    discriminator = NLayerDiscriminator(input_nc=disc_in_channels, n_layers=disc_num_layers,
+                                        use_actnorm=use_actnorm, ndf=disc_ndf).apply(weights_init)
+    inputs = torch.rand((6, 3, 256, 256))
+    outputs = discriminator(inputs)
+    print(outputs.shape)

ldm/modules/discriminator/multi_window_disc.py

@@ -0,0 +1,196 @@
+import numpy as np
+import torch
+import torch.nn as nn
+
+
+class Discriminator2DFactory(nn.Module):
+    def __init__(self, time_length, freq_length=80, kernel=(3, 3), c_in=1, hidden_size=128,
+                 norm_type='bn', reduction='sum'):
+        super(Discriminator2DFactory, self).__init__()
+        padding = (kernel[0] // 2, kernel[1] // 2)
+
+        def discriminator_block(in_filters, out_filters, first=False):
+            """
+            Input: (B, in, 2H, 2W)
+            Output:(B, out, H,  W)
+            """
+            conv = nn.Conv2d(in_filters, out_filters, kernel, (2, 2), padding)
+            if norm_type == 'sn':
+                conv = nn.utils.spectral_norm(conv)
+            block = [
+                conv,  # padding = kernel//2
+                nn.LeakyReLU(0.2, inplace=True),
+                nn.Dropout2d(0.25)
+            ]
+            if norm_type == 'bn' and not first:
+                block.append(nn.BatchNorm2d(out_filters, 0.8))
+            if norm_type == 'in' and not first:
+                block.append(nn.InstanceNorm2d(out_filters, affine=True))
+            block = nn.Sequential(*block)
+            return block
+
+        self.model = nn.ModuleList([
+            discriminator_block(c_in, hidden_size, first=True),
+            discriminator_block(hidden_size, hidden_size),
+            discriminator_block(hidden_size, hidden_size),
+        ])
+
+        self.reduction = reduction
+        ds_size = (time_length // 2 ** 3, (freq_length + 7) // 2 ** 3)
+        if reduction != 'none':
+            # The height and width of downsampled image
+            self.adv_layer = nn.Linear(hidden_size * ds_size[0] * ds_size[1], 1)
+        else:
+            self.adv_layer = nn.Linear(hidden_size * ds_size[1], 1)
+
+    def forward(self, x):
+        """
+
+        :param x: [B, C, T, n_bins]
+        :return: validity: [B, 1], h: List of hiddens
+        """
+        h = []
+        for l in self.model:
+            x = l(x)
+            h.append(x)
+        if self.reduction != 'none':
+            x = x.view(x.shape[0], -1)
+            validity = self.adv_layer(x)  # [B, 1]
+        else:
+            B, _, T_, _ = x.shape
+            x = x.transpose(1, 2).reshape(B, T_, -1)
+            validity = self.adv_layer(x)[:, :, 0]  # [B, T]
+        return validity, h
+
+
+class MultiWindowDiscriminator(nn.Module):
+    def __init__(self, time_lengths, cond_size=0, freq_length=80, kernel=(3, 3),
+                 c_in=1, hidden_size=128, norm_type='bn', reduction='sum'):
+        super(MultiWindowDiscriminator, self).__init__()
+        self.win_lengths = time_lengths
+        self.reduction = reduction
+
+        self.conv_layers = nn.ModuleList()
+        if cond_size > 0:
+            self.cond_proj_layers = nn.ModuleList()
+            self.mel_proj_layers = nn.ModuleList()
+        for time_length in time_lengths:
+            conv_layer = [
+                Discriminator2DFactory(
+                    time_length, freq_length, kernel, c_in=c_in, hidden_size=hidden_size,
+                    norm_type=norm_type, reduction=reduction)
+            ]
+            self.conv_layers += conv_layer
+            if cond_size > 0:
+                self.cond_proj_layers.append(nn.Linear(cond_size, freq_length))
+                self.mel_proj_layers.append(nn.Linear(freq_length, freq_length))
+
+    def forward(self, x, x_len, cond=None, start_frames_wins=None):
+        '''
+        Args:
+            x (tensor): input mel, (B, c_in, T, n_bins).
+            x_length (tensor): len of per mel. (B,).
+
+        Returns:
+            tensor : (B).
+        '''
+        validity = []
+        if start_frames_wins is None:
+            start_frames_wins = [None] * len(self.conv_layers)
+        h = []
+        for i, start_frames in zip(range(len(self.conv_layers)), start_frames_wins):
+            x_clip, c_clip, start_frames = self.clip(
+                x, cond, x_len, self.win_lengths[i], start_frames)  # (B, win_length, C)
+            start_frames_wins[i] = start_frames
+            if x_clip is None:
+                continue
+            if cond is not None:
+                x_clip = self.mel_proj_layers[i](x_clip)  # (B, 1, win_length, C)
+                c_clip = self.cond_proj_layers[i](c_clip)[:, None]  # (B, 1, win_length, C)
+                x_clip = x_clip + c_clip
+            x_clip, h_ = self.conv_layers[i](x_clip)
+            h += h_
+            validity.append(x_clip)
+        if len(validity) != len(self.conv_layers):
+            return None, start_frames_wins, h
+        if self.reduction == 'sum':
+            validity = sum(validity)  # [B]
+        elif self.reduction == 'stack':
+            validity = torch.stack(validity, -1)  # [B, W_L]
+        elif self.reduction == 'none':
+            validity = torch.cat(validity, -1)  # [B, W_sum]
+        return validity, start_frames_wins, h
+
+    def clip(self, x, cond, x_len, win_length, start_frames=None):
+        '''Ramdom clip x to win_length.
+        Args:
+            x (tensor) : (B, c_in, T, n_bins).
+            cond (tensor) : (B, T, H).
+            x_len (tensor) : (B,).
+            win_length (int): target clip length
+
+        Returns:
+            (tensor) : (B, c_in, win_length, n_bins).
+
+        '''
+        T_start = 0
+        T_end = x_len.max() - win_length
+        if T_end < 0:
+            return None, None, start_frames
+        T_end = T_end.item()
+        if start_frames is None:
+            start_frame = np.random.randint(low=T_start, high=T_end + 1)
+            start_frames = [start_frame] * x.size(0)
+        else:
+            start_frame = start_frames[0]
+        x_batch = x[:, :, start_frame: start_frame + win_length]
+        c_batch = cond[:, start_frame: start_frame + win_length] if cond is not None else None
+        return x_batch, c_batch, start_frames
+
+
+class Discriminator(nn.Module):
+    def __init__(self, time_lengths=[32, 64, 128], freq_length=80, cond_size=0, kernel=(3, 3), c_in=1,
+                 hidden_size=128, norm_type='bn', reduction='sum', uncond_disc=True):
+        super(Discriminator, self).__init__()
+        self.time_lengths = time_lengths
+        self.cond_size = cond_size
+        self.reduction = reduction
+        self.uncond_disc = uncond_disc
+        if uncond_disc:
+            self.discriminator = MultiWindowDiscriminator(
+                freq_length=freq_length,
+                time_lengths=time_lengths,
+                kernel=kernel,
+                c_in=c_in, hidden_size=hidden_size, norm_type=norm_type,
+                reduction=reduction
+            )
+        if cond_size > 0:
+            self.cond_disc = MultiWindowDiscriminator(
+                freq_length=freq_length,
+                time_lengths=time_lengths,
+                cond_size=cond_size,
+                kernel=kernel,
+                c_in=c_in, hidden_size=hidden_size, norm_type=norm_type,
+                reduction=reduction
+            )
+
+    def forward(self, x, cond=None, start_frames_wins=None):
+        """
+
+        :param x: [B, T, 80]
+        :param cond: [B, T, cond_size]
+        :param return_y_only:
+        :return:
+        """
+        if len(x.shape) == 3:
+            x = x[:, None, :, :]
+        x_len = x.sum([1, -1]).ne(0).int().sum([-1])
+        ret = {'y_c': None, 'y': None}
+        if self.uncond_disc:
+            ret['y'], start_frames_wins, ret['h'] = self.discriminator(
+                x, x_len, start_frames_wins=start_frames_wins)
+        if self.cond_size > 0 and cond is not None:
+            ret['y_c'], start_frames_wins, ret['h_c'] = self.cond_disc(
+                x, x_len, cond, start_frames_wins=start_frames_wins)
+        ret['start_frames_wins'] = start_frames_wins
+        return ret

ldm/modules/distributions/distributions.py

@@ -0,0 +1,92 @@
+import torch
+import numpy as np
+
+
+class AbstractDistribution:
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters, deterministic=False):
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
+
+    def sample(self):
+        x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device)
+        return x
+
+    def kl(self, other=None):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        else:
+            if other is None:
+                return 0.5 * torch.sum(torch.pow(self.mean, 2)
+                                       + self.var - 1.0 - self.logvar,
+                                       dim=[1, 2, 3])
+            else:
+                return 0.5 * torch.sum(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var - 1.0 - self.logvar + other.logvar,
+                    dim=[1, 2, 3])
+
+    def nll(self, sample, dims=[1,2,3]):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims)
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, torch.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+
+    # Force variances to be Tensors. Broadcasting helps convert scalars to
+    # Tensors, but it does not work for torch.exp().
+    logvar1, logvar2 = [
+        x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
+        for x in (logvar1, logvar2)
+    ]
+
+    return 0.5 * (
+        -1.0
+        + logvar2
+        - logvar1
+        + torch.exp(logvar1 - logvar2)
+        + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
+    )

ldm/modules/ema.py

@@ -0,0 +1,76 @@
+import torch
+from torch import nn
+
+
+class LitEma(nn.Module):
+    def __init__(self, model, decay=0.9999, use_num_upates=True):
+        super().__init__()
+        if decay < 0.0 or decay > 1.0:
+            raise ValueError('Decay must be between 0 and 1')
+
+        self.m_name2s_name = {}
+        self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32))
+        self.register_buffer('num_updates', torch.tensor(0,dtype=torch.int) if use_num_upates
+                             else torch.tensor(-1,dtype=torch.int))
+
+        for name, p in model.named_parameters():
+            if p.requires_grad:
+                #remove as '.'-character is not allowed in buffers
+                s_name = name.replace('.','')
+                self.m_name2s_name.update({name:s_name})
+                self.register_buffer(s_name,p.clone().detach().data)
+
+        self.collected_params = []
+
+    def forward(self,model):
+        decay = self.decay
+
+        if self.num_updates >= 0:
+            self.num_updates += 1
+            decay = min(self.decay,(1 + self.num_updates) / (10 + self.num_updates))
+
+        one_minus_decay = 1.0 - decay
+
+        with torch.no_grad():
+            m_param = dict(model.named_parameters())
+            shadow_params = dict(self.named_buffers())
+
+            for key in m_param:
+                if m_param[key].requires_grad:
+                    sname = self.m_name2s_name[key]
+                    shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
+                    shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key]))
+                else:
+                    assert not key in self.m_name2s_name
+
+    def copy_to(self, model):
+        m_param = dict(model.named_parameters())
+        shadow_params = dict(self.named_buffers())
+        for key in m_param:
+            if m_param[key].requires_grad:
+                m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
+            else:
+                assert not key in self.m_name2s_name
+
+    def store(self, parameters):
+        """
+        Save the current parameters for restoring later.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            temporarily stored.
+        """
+        self.collected_params = [param.clone() for param in parameters]
+
+    def restore(self, parameters):
+        """
+        Restore the parameters stored with the `store` method.
+        Useful to validate the model with EMA parameters without affecting the
+        original optimization process. Store the parameters before the
+        `copy_to` method. After validation (or model saving), use this to
+        restore the former parameters.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            updated with the stored parameters.
+        """
+        for c_param, param in zip(self.collected_params, parameters):
+            param.data.copy_(c_param.data)

ldm/modules/encoders/CLAP/CLAPWrapper.py

@@ -0,0 +1,257 @@
+import random
+import torchaudio
+from torch._six import string_classes
+import collections
+import re
+import torch.nn.functional as F
+import numpy as np
+from transformers import AutoTokenizer
+from ldm.modules.encoders.CLAP.utils import read_config_as_args
+from ldm.modules.encoders.CLAP.clap import CLAP
+import math
+import torchaudio.transforms as T
+import os
+import torch
+from importlib_resources import files
+
+
+class CLAPWrapper():
+    """
+    A class for interfacing CLAP model.  
+    """
+
+    def __init__(self, model_fp, device):
+        self.np_str_obj_array_pattern = re.compile(r'[SaUO]')
+        self.file_path = os.path.realpath(__file__)
+        self.default_collate_err_msg_format = (
+            "default_collate: batch must contain tensors, numpy arrays, numbers, "
+            "dicts or lists; found {}")
+        self.config_as_str = files('ldm').joinpath('modules/encoders/CLAP/config.yml').read_text()
+        self.model_fp = model_fp
+        self.device = device
+        self.clap, self.tokenizer, self.args = self.load_clap()
+
+    def load_clap(self):
+        r"""Load CLAP model with args from config file"""
+
+        args = read_config_as_args(self.config_as_str, is_config_str=True)
+
+        if 'bert' in args.text_model:
+            self.token_keys = ['input_ids', 'token_type_ids', 'attention_mask']
+        else:
+            self.token_keys = ['input_ids', 'attention_mask']
+
+        clap = CLAP(
+            audioenc_name=args.audioenc_name,
+            sample_rate=args.sampling_rate,
+            window_size=args.window_size,
+            hop_size=args.hop_size,
+            mel_bins=args.mel_bins,
+            fmin=args.fmin,
+            fmax=args.fmax,
+            classes_num=args.num_classes,
+            out_emb=args.out_emb,
+            text_model=args.text_model,
+            transformer_embed_dim=args.transformer_embed_dim,
+            d_proj=args.d_proj
+        )
+
+        # Load pretrained weights for model
+        model_state_dict = torch.load(self.model_fp, map_location=torch.device('cpu'))['model']
+        clap.load_state_dict(model_state_dict)
+
+        clap.eval()  # set clap in eval mode
+        tokenizer = AutoTokenizer.from_pretrained(args.text_model)
+
+        clap = clap.to(self.device)
+        tokenizer = tokenizer.to(self.device)
+
+        return clap, tokenizer, args
+
+    def default_collate(self, batch):
+        r"""Puts each data field into a tensor with outer dimension batch size"""
+        elem = batch[0]
+        elem_type = type(elem)
+        if isinstance(elem, torch.Tensor):
+            out = None
+            if torch.utils.data.get_worker_info() is not None:
+                # If we're in a background process, concatenate directly into a
+                # shared memory tensor to avoid an extra copy
+                numel = sum([x.numel() for x in batch])
+                storage = elem.storage()._new_shared(numel)
+                out = elem.new(storage)
+            return torch.stack(batch, 0, out=out)
+        elif elem_type.__module__ == 'numpy' and elem_type.__name__ != 'str_' \
+                and elem_type.__name__ != 'string_':
+            if elem_type.__name__ == 'ndarray' or elem_type.__name__ == 'memmap':
+                # array of string classes and object
+                if self.np_str_obj_array_pattern.search(elem.dtype.str) is not None:
+                    raise TypeError(
+                        self.default_collate_err_msg_format.format(elem.dtype))
+
+                return self.default_collate([torch.as_tensor(b) for b in batch])
+            elif elem.shape == ():  # scalars
+                return torch.as_tensor(batch)
+        elif isinstance(elem, float):
+            return torch.tensor(batch, dtype=torch.float64)
+        elif isinstance(elem, int):
+            return torch.tensor(batch)
+        elif isinstance(elem, string_classes):
+            return batch
+        elif isinstance(elem, collections.abc.Mapping):
+            return {key: self.default_collate([d[key] for d in batch]) for key in elem}
+        elif isinstance(elem, tuple) and hasattr(elem, '_fields'):  # namedtuple
+            return elem_type(*(self.default_collate(samples) for samples in zip(*batch)))
+        elif isinstance(elem, collections.abc.Sequence):
+            # check to make sure that the elements in batch have consistent size
+            it = iter(batch)
+            elem_size = len(next(it))
+            if not all(len(elem) == elem_size for elem in it):
+                raise RuntimeError(
+                    'each element in list of batch should be of equal size')
+            transposed = zip(*batch)
+            return [self.default_collate(samples) for samples in transposed]
+
+        raise TypeError(self.default_collate_err_msg_format.format(elem_type))
+
+    def load_audio_into_tensor(self, audio_path, audio_duration, resample=False):
+        r"""Loads audio file and returns raw audio."""
+        # Randomly sample a segment of audio_duration from the clip or pad to match duration
+        audio_time_series, sample_rate = torchaudio.load(audio_path)
+        resample_rate = self.args.sampling_rate
+        if resample:
+            resampler = T.Resample(sample_rate, resample_rate)
+            audio_time_series = resampler(audio_time_series)
+        audio_time_series = audio_time_series.reshape(-1)
+
+        # audio_time_series is shorter than predefined audio duration,
+        # so audio_time_series is extended
+        if audio_duration*sample_rate >= audio_time_series.shape[0]:
+            repeat_factor = int(np.ceil((audio_duration*sample_rate) /
+                                        audio_time_series.shape[0]))
+            # Repeat audio_time_series by repeat_factor to match audio_duration
+            audio_time_series = audio_time_series.repeat(repeat_factor)
+            # remove excess part of audio_time_series
+            audio_time_series = audio_time_series[0:audio_duration*sample_rate]
+        else:
+            # audio_time_series is longer than predefined audio duration,
+            # so audio_time_series is trimmed
+            start_index = random.randrange(
+                audio_time_series.shape[0] - audio_duration*sample_rate)
+            audio_time_series = audio_time_series[start_index:start_index +
+                                                  audio_duration*sample_rate]
+        return torch.FloatTensor(audio_time_series)
+
+    def preprocess_audio(self, audio_files, resample):
+        r"""Load list of audio files and return raw audio"""
+        audio_tensors = []
+        for audio_file in audio_files:
+            audio_tensor = self.load_audio_into_tensor(
+                audio_file, self.args.duration, resample)
+            audio_tensor = audio_tensor.reshape(1, -1).to(self.device)
+            audio_tensors.append(audio_tensor)
+        return self.default_collate(audio_tensors)
+
+    def preprocess_text(self, text_queries, text_len=100):
+        r"""Load list of class labels and return tokenized text"""
+        device = next(self.clap.parameters()).device
+        tokenized_texts = []
+        for ttext in text_queries:
+            tok = self.tokenizer.encode_plus(
+                text=ttext, add_special_tokens=True, max_length=text_len, pad_to_max_length=True, return_tensors="pt")
+            for key in self.token_keys:
+                tok[key] = tok[key].reshape(-1).to(device)
+            tokenized_texts.append(tok)
+        return self.default_collate(tokenized_texts)
+
+    def get_text_embeddings(self, class_labels):
+        r"""Load list of class labels and return text embeddings"""
+        preprocessed_text = self.preprocess_text(class_labels)
+        text_embeddings = self._get_text_embeddings(preprocessed_text)
+        text_embeddings = text_embeddings/torch.norm(text_embeddings, dim=-1, keepdim=True)
+        return text_embeddings
+
+    def get_audio_embeddings(self, audio_files, resample):
+        r"""Load list of audio files and return a audio embeddings"""
+        preprocessed_audio = self.preprocess_audio(audio_files, resample)
+        audio_embeddings = self._get_audio_embeddings(preprocessed_audio)
+        audio_embeddings = audio_embeddings/torch.norm(audio_embeddings, dim=-1, keepdim=True)
+        return audio_embeddings
+
+    def _get_text_embeddings(self, preprocessed_text):
+        r"""Load preprocessed text and return text embeddings"""
+        with torch.no_grad():
+            text_embeddings = self.clap.caption_encoder(preprocessed_text)
+            text_embeddings = text_embeddings/torch.norm(text_embeddings, dim=-1, keepdim=True)
+            return text_embeddings
+
+    def _get_audio_embeddings(self, preprocessed_audio):
+        r"""Load preprocessed audio and return a audio embeddings"""
+        with torch.no_grad():
+            preprocessed_audio = preprocessed_audio.reshape(
+                preprocessed_audio.shape[0], preprocessed_audio.shape[2])
+            #Append [0] the audio emebdding, [1] has output class probabilities
+            audio_embeddings = self.clap.audio_encoder(preprocessed_audio)[0]
+            audio_embeddings = audio_embeddings/torch.norm(audio_embeddings, dim=-1, keepdim=True)
+            return audio_embeddings
+    
+    def compute_similarity(self, audio_embeddings, text_embeddings):
+        r"""Compute similarity between text and audio embeddings"""
+        logit_scale = self.clap.logit_scale.exp()
+        similarity = logit_scale*text_embeddings @ audio_embeddings.T
+        return similarity.T
+
+    def _generic_batch_inference(self, func, *args):
+        r"""Process audio and/or text per batch"""
+        input_tmp = args[0]
+        batch_size = args[-1]
+        # args[0] has audio_files, args[1] has class_labels
+        inputs = [args[0], args[1]] if len(args) == 3 else [args[0]]
+        args0_len = len(args[0])
+        # compute text_embeddings once for all the audio_files batches
+        if len(inputs) == 2:
+            text_embeddings = self.get_text_embeddings(args[1])
+            inputs = [args[0], args[1], text_embeddings]
+        dataset_idx = 0
+        for _ in range(math.ceil(args0_len/batch_size)):
+            next_batch_idx = dataset_idx + batch_size
+            # batch size is bigger than available audio/text items
+            if next_batch_idx >= args0_len:
+                inputs[0] = input_tmp[dataset_idx:]
+                return func(*tuple(inputs))
+            else:
+                inputs[0] = input_tmp[dataset_idx:next_batch_idx]
+                yield func(*tuple(inputs))
+            dataset_idx = next_batch_idx
+
+    def get_audio_embeddings_per_batch(self, audio_files, batch_size):
+        r"""Load preprocessed audio and return a audio embeddings per batch"""
+        return self._generic_batch_inference(self.get_audio_embeddings, audio_files, batch_size)
+
+    def get_text_embeddings_per_batch(self, class_labels, batch_size):
+        r"""Load preprocessed text and return text embeddings per batch"""
+        return self._generic_batch_inference(self.get_text_embeddings, class_labels, batch_size)
+
+    def classify_audio_files_per_batch(self, audio_files, class_labels, batch_size):
+        r"""Compute classification probabilities for each audio recording in a batch and each class label"""
+        return self._generic_batch_inference(self.classify_audio_files, audio_files, class_labels, batch_size)
+
+if __name__ == '__main__':
+
+    # Load and initialize CLAP
+    weights_path = "/home1/huangrongjie/Project/Diffusion/LatentDiffusion/CLAP/CLAP_weights_2022.pth"
+    clap_model = CLAPWrapper(weights_path, use_cuda=False)
+
+    y = ["A woman talks nearby as water pours", "Multiple clanging and clanking sounds"]
+    x = ['/home2/huangjiawei/data/audiocaps/train/Yr1nicOVtvkQ.wav', '/home2/huangjiawei/data/audiocaps/train/YUDGBjjwyaqE.wav']
+
+    # Computing text embeddings
+    text_embeddings = clap_model.get_text_embeddings(y)
+
+    import ipdb
+    ipdb.set_trace()
+
+    # Computing audio embeddings
+    audio_embeddings = clap_model.get_audio_embeddings(x, resample=True)
+    similarity = clap_model.compute_similarity(audio_embeddings, text_embeddings)
+

ldm/modules/encoders/CLAP/__init__.py

@@ -0,0 +1,3 @@
+from . import clap 
+from . import audio
+from . import utils

ldm/modules/encoders/CLAP/audio.py

@@ -0,0 +1,179 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchlibrosa.stft import Spectrogram, LogmelFilterBank
+
+def get_audio_encoder(name: str):
+    if name == "Cnn14":
+        return Cnn14
+    else:
+        raise Exception('The audio encoder name {} is incorrect or not supported'.format(name))
+
+
+class ConvBlock(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        
+        super(ConvBlock, self).__init__()
+        
+        self.conv1 = nn.Conv2d(in_channels=in_channels, 
+                              out_channels=out_channels,
+                              kernel_size=(3, 3), stride=(1, 1),
+                              padding=(1, 1), bias=False)
+                              
+        self.conv2 = nn.Conv2d(in_channels=out_channels, 
+                              out_channels=out_channels,
+                              kernel_size=(3, 3), stride=(1, 1),
+                              padding=(1, 1), bias=False)
+                              
+        self.bn1 = nn.BatchNorm2d(out_channels)
+        self.bn2 = nn.BatchNorm2d(out_channels)
+
+        
+    def forward(self, input, pool_size=(2, 2), pool_type='avg'):
+        
+        x = input
+        x = F.relu_(self.bn1(self.conv1(x)))
+        x = F.relu_(self.bn2(self.conv2(x)))
+        if pool_type == 'max':
+            x = F.max_pool2d(x, kernel_size=pool_size)
+        elif pool_type == 'avg':
+            x = F.avg_pool2d(x, kernel_size=pool_size)
+        elif pool_type == 'avg+max':
+            x1 = F.avg_pool2d(x, kernel_size=pool_size)
+            x2 = F.max_pool2d(x, kernel_size=pool_size)
+            x = x1 + x2
+        else:
+            raise Exception('Incorrect argument!')
+        
+        return x
+
+
+class ConvBlock5x5(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        
+        super(ConvBlock5x5, self).__init__()
+        
+        self.conv1 = nn.Conv2d(in_channels=in_channels, 
+                              out_channels=out_channels,
+                              kernel_size=(5, 5), stride=(1, 1),
+                              padding=(2, 2), bias=False)
+                              
+        self.bn1 = nn.BatchNorm2d(out_channels)
+
+        
+    def forward(self, input, pool_size=(2, 2), pool_type='avg'):
+        
+        x = input
+        x = F.relu_(self.bn1(self.conv1(x)))
+        if pool_type == 'max':
+            x = F.max_pool2d(x, kernel_size=pool_size)
+        elif pool_type == 'avg':
+            x = F.avg_pool2d(x, kernel_size=pool_size)
+        elif pool_type == 'avg+max':
+            x1 = F.avg_pool2d(x, kernel_size=pool_size)
+            x2 = F.max_pool2d(x, kernel_size=pool_size)
+            x = x1 + x2
+        else:
+            raise Exception('Incorrect argument!')
+        
+        return x
+
+
+class AttBlock(nn.Module):
+    def __init__(self, n_in, n_out, activation='linear', temperature=1.):
+        super(AttBlock, self).__init__()
+        
+        self.activation = activation
+        self.temperature = temperature
+        self.att = nn.Conv1d(in_channels=n_in, out_channels=n_out, kernel_size=1, stride=1, padding=0, bias=True)
+        self.cla = nn.Conv1d(in_channels=n_in, out_channels=n_out, kernel_size=1, stride=1, padding=0, bias=True)
+        
+        self.bn_att = nn.BatchNorm1d(n_out)
+         
+    def forward(self, x):
+        # x: (n_samples, n_in, n_time)
+        norm_att = torch.softmax(torch.clamp(self.att(x), -10, 10), dim=-1)
+        cla = self.nonlinear_transform(self.cla(x))
+        x = torch.sum(norm_att * cla, dim=2)
+        return x, norm_att, cla
+
+    def nonlinear_transform(self, x):
+        if self.activation == 'linear':
+            return x
+        elif self.activation == 'sigmoid':
+            return torch.sigmoid(x)
+
+
+class Cnn14(nn.Module):
+    def __init__(self, sample_rate, window_size, hop_size, mel_bins, fmin, 
+        fmax, classes_num, out_emb):
+        
+        super(Cnn14, self).__init__()
+
+        window = 'hann'
+        center = True
+        pad_mode = 'reflect'
+        ref = 1.0
+        amin = 1e-10
+        top_db = None
+
+        # Spectrogram extractor
+        self.spectrogram_extractor = Spectrogram(n_fft=window_size, hop_length=hop_size, 
+            win_length=window_size, window=window, center=center, pad_mode=pad_mode, 
+            freeze_parameters=True)
+
+        # Logmel feature extractor
+        self.logmel_extractor = LogmelFilterBank(sr=sample_rate, n_fft=window_size, 
+            n_mels=mel_bins, fmin=fmin, fmax=fmax, ref=ref, amin=amin, top_db=top_db, 
+            freeze_parameters=True)
+
+        self.bn0 = nn.BatchNorm2d(64)
+
+        self.conv_block1 = ConvBlock(in_channels=1, out_channels=64)
+        self.conv_block2 = ConvBlock(in_channels=64, out_channels=128)
+        self.conv_block3 = ConvBlock(in_channels=128, out_channels=256)
+        self.conv_block4 = ConvBlock(in_channels=256, out_channels=512)
+        self.conv_block5 = ConvBlock(in_channels=512, out_channels=1024)
+        self.conv_block6 = ConvBlock(in_channels=1024, out_channels=2048)
+        
+        # out_emb is 2048 for best Cnn14
+        self.fc1 = nn.Linear(2048, out_emb, bias=True)
+        self.fc_audioset = nn.Linear(out_emb, classes_num, bias=True)
+        
+    def forward(self, input, mixup_lambda=None):
+        """
+        Input: (batch_size, data_length)
+        """
+
+        x = self.spectrogram_extractor(input)   # (batch_size, 1, time_steps, freq_bins)
+        x = self.logmel_extractor(x)    # (batch_size, 1, time_steps, mel_bins)
+
+        x = x.transpose(1, 3)
+        x = self.bn0(x)
+        x = x.transpose(1, 3)
+
+        x = self.conv_block1(x, pool_size=(2, 2), pool_type='avg')
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block2(x, pool_size=(2, 2), pool_type='avg')
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block3(x, pool_size=(2, 2), pool_type='avg')
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block4(x, pool_size=(2, 2), pool_type='avg')
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block5(x, pool_size=(2, 2), pool_type='avg')
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = self.conv_block6(x, pool_size=(1, 1), pool_type='avg')
+        x = F.dropout(x, p=0.2, training=self.training)
+        x = torch.mean(x, dim=3)
+        
+        (x1, _) = torch.max(x, dim=2)
+        x2 = torch.mean(x, dim=2)
+        x = x1 + x2
+        x = F.dropout(x, p=0.5, training=self.training)
+        x = F.relu_(self.fc1(x))
+        embedding = F.dropout(x, p=0.5, training=self.training)
+        clipwise_output = torch.sigmoid(self.fc_audioset(x))
+        
+        output_dict = {'clipwise_output': clipwise_output, 'embedding': embedding}
+
+        return output_dict

ldm/modules/encoders/CLAP/clap.py

@@ -0,0 +1,89 @@
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+from transformers import AutoModel
+from .audio import get_audio_encoder
+
+class Projection(nn.Module):
+    def __init__(self, d_in: int, d_out: int, p: float=0.5) -> None:
+        super().__init__()
+        self.linear1 = nn.Linear(d_in, d_out, bias=False)
+        self.linear2 = nn.Linear(d_out, d_out, bias=False)
+        self.layer_norm = nn.LayerNorm(d_out)
+        self.drop = nn.Dropout(p)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        embed1 = self.linear1(x)
+        embed2 = self.drop(self.linear2(F.gelu(embed1)))
+        embeds = self.layer_norm(embed1 + embed2)
+        return embeds
+
+class AudioEncoder(nn.Module):
+    def __init__(self, audioenc_name:str, d_in: int, d_out: int, sample_rate: int, window_size: int,
+            hop_size: int, mel_bins: int, fmin: int, fmax: int, classes_num: int) -> None:
+        super().__init__()
+
+        audio_encoder = get_audio_encoder(audioenc_name)
+
+        self.base = audio_encoder(
+            sample_rate, window_size,
+            hop_size, mel_bins, fmin, fmax,
+            classes_num, d_in)
+
+        self.projection = Projection(d_in, d_out)
+
+    def forward(self, x):
+        out_dict = self.base(x)
+        audio_features, audio_classification_output = out_dict['embedding'], out_dict['clipwise_output']
+        projected_vec = self.projection(audio_features)
+        return projected_vec, audio_classification_output
+
+class TextEncoder(nn.Module):
+    def __init__(self, d_out: int, text_model: str, transformer_embed_dim: int) -> None:
+        super().__init__()
+        self.base = AutoModel.from_pretrained(text_model)
+        self.projection = Projection(transformer_embed_dim, d_out)
+
+    def forward(self, x):
+        out = self.base(**x)[0]
+        out = out[:, 0, :]  # get CLS token output
+        projected_vec = self.projection(out)
+        return projected_vec
+
+class CLAP(nn.Module):
+    def __init__(self,
+                # audio
+                audioenc_name: str,
+                sample_rate: int, 
+                window_size: int, 
+                hop_size: int, 
+                mel_bins: int, 
+                fmin: int, 
+                fmax: int, 
+                classes_num: int, 
+                out_emb: int,
+                # text
+                text_model: str,
+                transformer_embed_dim: int,
+                # common
+                d_proj: int,
+                ):
+        super().__init__()
+
+        
+        self.audio_encoder = AudioEncoder(
+            audioenc_name, out_emb, d_proj,
+            sample_rate, window_size, hop_size, mel_bins, fmin, fmax, classes_num)
+
+        self.caption_encoder = TextEncoder(
+            d_proj, text_model, transformer_embed_dim
+        )
+
+        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+
+    def forward(self, audio, text):
+        audio_embed, _ = self.audio_encoder(audio)
+        caption_embed = self.caption_encoder(text)
+
+        return caption_embed, audio_embed, self.logit_scale.exp()

ldm/modules/encoders/CLAP/config.yml

@@ -0,0 +1,26 @@
+# TEXT ENCODER CONFIG
+text_model: 'bert-base-uncased'
+text_len: 100
+transformer_embed_dim: 768
+freeze_text_encoder_weights: True
+
+# AUDIO ENCODER CONFIG
+audioenc_name: 'Cnn14'
+out_emb: 2048
+sampling_rate: 44100
+duration: 5
+fmin: 50
+fmax: 14000
+n_fft: 1028
+hop_size: 320
+mel_bins: 64
+window_size: 1024
+
+# PROJECTION SPACE CONFIG 
+d_proj: 1024
+temperature: 0.003
+
+# TRAINING AND EVALUATION CONFIG
+num_classes: 527
+batch_size: 1024
+demo: False

ldm/modules/encoders/CLAP/utils.py

@@ -0,0 +1,26 @@
+import argparse
+import yaml
+import sys
+
+def read_config_as_args(config_path,args=None,is_config_str=False):
+    return_dict = {}
+
+    if config_path is not None:
+        if is_config_str:
+            yml_config = yaml.load(config_path, Loader=yaml.FullLoader)
+        else:
+            with open(config_path, "r") as f:
+                yml_config = yaml.load(f, Loader=yaml.FullLoader)
+
+        if args != None:
+            for k, v in yml_config.items():
+                if k in args.__dict__:
+                    args.__dict__[k] = v
+                else:
+                    sys.stderr.write("Ignored unknown parameter {} in yaml.\n".format(k))
+        else:
+            for k, v in yml_config.items():
+                return_dict[k] = v
+
+    args = args if args != None else return_dict
+    return argparse.Namespace(**args)

ldm/modules/encoders/modules.py

@@ -0,0 +1,314 @@
+import torch
+import torch.nn as nn
+from functools import partial
+
+from ldm.modules.x_transformer import Encoder, TransformerWrapper  # TODO: can we directly rely on lucidrains code and simply add this as a reuirement? --> test
+from torch.utils.checkpoint import checkpoint
+from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel, AutoTokenizer
+from importlib_resources import files
+from ldm.modules.encoders.CLAP.utils import read_config_as_args
+from ldm.modules.encoders.CLAP.clap import TextEncoder
+from ldm.util import default, count_params
+
+
+class AbstractEncoder(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def encode(self, *args, **kwargs):
+        raise NotImplementedError
+
+
+class ClassEmbedder(nn.Module):
+    def __init__(self, embed_dim, n_classes=1000, key='class'):
+        super().__init__()
+        self.key = key
+        self.embedding = nn.Embedding(n_classes, embed_dim)
+
+    def forward(self, batch, key=None):
+        if key is None:
+            key = self.key
+        # this is for use in crossattn
+        c = batch[key][:, None]# (bsz,1)
+        c = self.embedding(c)
+        return c
+
+
+class TransformerEmbedder(AbstractEncoder):
+    """Some transformer encoder layers"""
+    def __init__(self, n_embed, n_layer, vocab_size, max_seq_len=77, device="cuda"):
+        super().__init__()
+        self.device = device
+        self.transformer = TransformerWrapper(num_tokens=vocab_size, max_seq_len=max_seq_len,
+                                              attn_layers=Encoder(dim=n_embed, depth=n_layer))
+
+    def forward(self, tokens):
+        tokens = tokens.to(self.device)  # meh
+        z = self.transformer(tokens, return_embeddings=True)
+        return z
+
+    def encode(self, x):
+        return self(x)
+
+
+class BERTTokenizer(AbstractEncoder):
+    """ Uses a pretrained BERT tokenizer by huggingface. Vocab size: 30522 (?)"""
+    def __init__(self, device="cuda", vq_interface=True, max_length=77):
+        super().__init__()
+        from transformers import BertTokenizerFast  # TODO: add to reuquirements
+        self.tokenizer = BertTokenizerFast.from_pretrained("bert-base-uncased")
+        self.device = device
+        self.vq_interface = vq_interface
+        self.max_length = max_length
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        return tokens
+
+    @torch.no_grad()
+    def encode(self, text):
+        tokens = self(text)
+        if not self.vq_interface:
+            return tokens
+        return None, None, [None, None, tokens]
+
+    def decode(self, text):
+        return text
+
+
+class BERTEmbedder(AbstractEncoder):# 这里不是用的pretrained bert,是用的transformers的BertTokenizer加自定义的TransformerWrapper
+    """Uses the BERT tokenizr model and add some transformer encoder layers"""
+    def __init__(self, n_embed, n_layer, vocab_size=30522, max_seq_len=77,
+                 device="cuda",use_tokenizer=True, embedding_dropout=0.0):
+        super().__init__()
+        self.use_tknz_fn = use_tokenizer
+        if self.use_tknz_fn:
+            self.tknz_fn = BERTTokenizer(vq_interface=False, max_length=max_seq_len)
+        self.device = device
+        self.transformer = TransformerWrapper(num_tokens=vocab_size, max_seq_len=max_seq_len,
+                                              attn_layers=Encoder(dim=n_embed, depth=n_layer),
+                                              emb_dropout=embedding_dropout)
+
+    def forward(self, text):
+        if self.use_tknz_fn:
+            tokens = self.tknz_fn(text)#.to(self.device)
+        else:
+            tokens = text
+        z = self.transformer(tokens, return_embeddings=True)
+        return z
+
+    def encode(self, text):
+        # output of length 77
+        return self(text)
+
+
+class SpatialRescaler(nn.Module):
+    def __init__(self,
+                 n_stages=1,
+                 method='bilinear',
+                 multiplier=0.5,
+                 in_channels=3,
+                 out_channels=None,
+                 bias=False):
+        super().__init__()
+        self.n_stages = n_stages
+        assert self.n_stages >= 0
+        assert method in ['nearest','linear','bilinear','trilinear','bicubic','area']
+        self.multiplier = multiplier
+        self.interpolator = partial(torch.nn.functional.interpolate, mode=method)
+        self.remap_output = out_channels is not None
+        if self.remap_output:
+            print(f'Spatial Rescaler mapping from {in_channels} to {out_channels} channels after resizing.')
+            self.channel_mapper = nn.Conv2d(in_channels,out_channels,1,bias=bias)
+
+    def forward(self,x):
+        for stage in range(self.n_stages):
+            x = self.interpolator(x, scale_factor=self.multiplier)
+
+
+        if self.remap_output:
+            x = self.channel_mapper(x)
+        return x
+
+    def encode(self, x):
+        return self(x)
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+class FrozenT5Embedder(AbstractEncoder):
+    """Uses the T5 transformer encoder for text"""
+    def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77, freeze=True):  # others are google/t5-v1_1-xl and google/t5-v1_1-xxl
+        super().__init__()
+        self.tokenizer = T5Tokenizer.from_pretrained(version)
+        self.transformer = T5EncoderModel.from_pretrained(version)
+        self.device = device
+        self.max_length = max_length   # TODO: typical value?
+        if freeze:
+            self.freeze()
+
+    def freeze(self):
+        self.transformer = self.transformer.eval()
+        #self.train = disabled_train
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.transformer(input_ids=tokens)
+
+        z = outputs.last_hidden_state
+        return z
+
+    def encode(self, text):
+        return self(text)
+
+
+class FrozenCLAPEmbedder(AbstractEncoder):
+    """Uses the CLAP transformer encoder for text (from huggingface)"""
+    def __init__(self, weights_path, freeze=True, device="cuda", max_length=77):  # clip-vit-base-patch32
+        super().__init__()
+
+        model_state_dict = torch.load(weights_path, map_location=torch.device('cpu'))['model']
+        match_params = dict()
+        for key in list(model_state_dict.keys()):
+            if 'caption_encoder' in key:
+                match_params[key.replace('caption_encoder.', '')] = model_state_dict[key]
+
+        config_as_str = files('ldm').joinpath('modules/encoders/CLAP/config.yml').read_text()
+        args = read_config_as_args(config_as_str, is_config_str=True)
+
+        # To device
+        self.tokenizer = AutoTokenizer.from_pretrained(args.text_model) # args.text_model
+        self.caption_encoder = TextEncoder(
+            args.d_proj, args.text_model, args.transformer_embed_dim
+        )
+
+        self.max_length = max_length
+        self.device = device
+        if freeze: self.freeze()
+
+        print(f"{self.caption_encoder.__class__.__name__} comes with {count_params(self.caption_encoder) * 1.e-6:.2f} M params.")
+
+    def freeze(self):
+        self.caption_encoder.base = self.caption_encoder.base.eval()
+        for param in self.caption_encoder.base.parameters():
+            param.requires_grad = False
+
+
+    def encode(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+
+        outputs = self.caption_encoder.base(input_ids=tokens)
+        z = self.caption_encoder.projection(outputs.last_hidden_state)
+        return z
+
+class FrozenCLAPEmbedderNoLoad(AbstractEncoder):
+    def __init__(self, config, freeze=True, device="cpu", max_length=77):
+        super().__init__()
+        args = config
+
+        # To device
+        self.tokenizer = AutoTokenizer.from_pretrained(args.text_model) # args.text_model
+        self.caption_encoder = TextEncoder(
+            args.d_proj, args.text_model, args.transformer_embed_dim
+        )
+
+        self.max_length = max_length
+        self.device = device
+        if freeze: self.freeze()
+
+        print(f"{self.caption_encoder.__class__.__name__} comes with {count_params(self.caption_encoder) * 1.e-6:.2f} M params.")        
+
+    def freeze(self):
+        self.caption_encoder.base = self.caption_encoder.base.eval()
+        for param in self.caption_encoder.base.parameters():
+            param.requires_grad = False
+
+
+    def encode(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+
+        outputs = self.caption_encoder.base(input_ids=tokens)
+        z = self.caption_encoder.projection(outputs.last_hidden_state)
+        return z
+
+
+class NewFrozenCLAPEmbedder(AbstractEncoder):
+    """Uses the CLAP transformer encoder for text (from huggingface)"""
+    def __init__(self, weights_path, freeze=True, device="cuda", max_length=77):  # clip-vit-base-patch32
+        super().__init__()
+        # To device
+        from transformers import RobertaTokenizer
+        from ldm.modules.encoders.open_clap import create_model
+
+
+        model, model_cfg = create_model(
+            'HTSAT-tiny',
+            'roberta',
+            weights_path,
+            enable_fusion=True,
+            fusion_type='aff_2d'
+        )
+
+        del model.audio_branch, model.audio_transform, model.audio_projection
+        self.tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
+        self.model = model
+
+        self.max_length = max_length
+        self.device = device
+        if freeze: self.freeze()
+
+        param_num = sum(p.numel() for p in model.parameters() if p.requires_grad)
+        print(f'{self.model.__class__.__name__} comes with: {param_num / 1e+6:.3f} M params.')
+
+    def freeze(self):
+        self.model = self.model.eval()
+        for param in self.model.parameters():
+            param.requires_grad = False
+
+    def encode(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        outputs = self.model.text_branch(input_ids=batch_encoding["input_ids"].to(self.device), attention_mask=batch_encoding["attention_mask"].to(self.device))
+        z = self.model.text_projection(outputs.last_hidden_state)
+        return z
+
+class FrozenFLANEmbedder(AbstractEncoder):
+    """Uses the T5 transformer encoder for text"""
+    def __init__(self, version="google/flan-t5-large", device="cuda", max_length=77, freeze=True):  # others are google/t5-v1_1-xl and google/t5-v1_1-xxl
+        super().__init__()
+        self.tokenizer = T5Tokenizer.from_pretrained(version)
+        self.transformer = T5EncoderModel.from_pretrained(version)
+        self.device = device
+        self.max_length = max_length   # TODO: typical value?
+        if freeze:
+            self.freeze()
+
+    def freeze(self):
+        self.transformer = self.transformer.eval()
+        #self.train = disabled_train
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, text):
+        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
+                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.transformer(input_ids=tokens)
+
+        z = outputs.last_hidden_state
+        return z
+
+    def encode(self, text):
+        return self(text)

ldm/modules/encoders/open_clap/__init__.py

@@ -0,0 +1,8 @@
+from .factory import list_models, create_model, create_model_and_transforms, add_model_config
+from .loss import ClipLoss, gather_features, LPLoss, lp_gather_features, LPMetrics
+from .model import CLAP, CLAPTextCfg, CLAPVisionCfg, CLAPAudioCfp, convert_weights_to_fp16, trace_model
+from .openai import load_openai_model, list_openai_models
+from .pretrained import list_pretrained, list_pretrained_tag_models, list_pretrained_model_tags,\
+    get_pretrained_url, download_pretrained
+from .tokenizer import SimpleTokenizer, tokenize
+from .transform import image_transform

ldm/modules/encoders/open_clap/bert.py

@@ -0,0 +1,32 @@
+from transformers import BertTokenizer, BertModel
+tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+model = BertModel.from_pretrained("bert-base-uncased")
+text = "Replace me by any text you'd like."
+
+def bert_embeddings(text):
+    # text = "Replace me by any text you'd like."
+    encoded_input = tokenizer(text, return_tensors='pt')
+    output = model(**encoded_input)
+    return output
+    
+from transformers import RobertaTokenizer, RobertaModel
+
+tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
+model = RobertaModel.from_pretrained('roberta-base')
+text = "Replace me by any text you'd like."
+def Roberta_embeddings(text):
+    # text = "Replace me by any text you'd like."
+    encoded_input = tokenizer(text, return_tensors='pt')
+    output = model(**encoded_input)
+    return output
+
+from transformers import BartTokenizer, BartModel
+
+tokenizer = BartTokenizer.from_pretrained('facebook/bart-base')
+model = BartModel.from_pretrained('facebook/bart-base')
+text = "Replace me by any text you'd like."
+def bart_embeddings(text):
+    # text = "Replace me by any text you'd like."
+    encoded_input = tokenizer(text, return_tensors='pt')
+    output = model(**encoded_input)
+    return output

ldm/modules/encoders/open_clap/bpe_simple_vocab_16e6.txt.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a
+size 1356917

ldm/modules/encoders/open_clap/factory.py

@@ -0,0 +1,257 @@
+import json
+import logging
+import os
+import pathlib
+import re
+from copy import deepcopy
+from pathlib import Path
+
+import torch
+
+from .model import CLAP, convert_weights_to_fp16
+from .openai import load_openai_model
+from .pretrained import get_pretrained_url, download_pretrained
+from .transform import image_transform
+
+_MODEL_CONFIG_PATHS = [Path(__file__).parent / f"model_configs/"]
+_MODEL_CONFIGS = {}  # directory (model_name: config) of model architecture configs
+
+
+def _natural_key(string_):
+    return [int(s) if s.isdigit() else s for s in re.split(r"(\d+)", string_.lower())]
+
+
+def _rescan_model_configs():
+    global _MODEL_CONFIGS
+
+    config_ext = (".json",)
+    config_files = []
+    for config_path in _MODEL_CONFIG_PATHS:
+        if config_path.is_file() and config_path.suffix in config_ext:
+            config_files.append(config_path)
+        elif config_path.is_dir():
+            for ext in config_ext:
+                config_files.extend(config_path.glob(f"*{ext}"))
+
+    for cf in config_files:
+        with open(cf, "r") as f:
+            model_cfg = json.load(f)
+            if all(a in model_cfg for a in ("embed_dim", "audio_cfg", "text_cfg")):
+                _MODEL_CONFIGS[cf.stem] = model_cfg
+
+    _MODEL_CONFIGS = {
+        k: v
+        for k, v in sorted(_MODEL_CONFIGS.items(), key=lambda x: _natural_key(x[0]))
+    }
+
+
+_rescan_model_configs()  # initial populate of model config registry
+
+
+def load_state_dict(checkpoint_path: str, map_location="cpu", skip_params=True):
+    checkpoint = torch.load(checkpoint_path, map_location=map_location)
+    if isinstance(checkpoint, dict) and "state_dict" in checkpoint:
+        state_dict = checkpoint["state_dict"]
+    else:
+        state_dict = checkpoint
+    if skip_params:
+        if next(iter(state_dict.items()))[0].startswith("module"):
+            state_dict = {k[7:]: v for k, v in state_dict.items()}
+    # for k in state_dict:
+    #     if k.startswith('transformer'):
+    #         v = state_dict.pop(k)
+    #         state_dict['text_branch.' + k[12:]] = v
+    return state_dict
+
+
+def create_model(
+    amodel_name: str,
+    tmodel_name: str,
+    pretrained: str = "",
+    precision: str = "fp32",
+    device: torch.device = torch.device("cpu"),
+    jit: bool = False,
+    force_quick_gelu: bool = False,
+    openai_model_cache_dir: str = os.path.expanduser("~/.cache/clip"),
+    skip_params=True,
+    pretrained_audio: str = "",
+    pretrained_text: str = "",
+    enable_fusion: bool = False,
+    fusion_type: str = 'None'
+    # pretrained_image: bool = False,
+):
+    amodel_name = amodel_name.replace(
+        "/", "-"
+    )  # for callers using old naming with / in ViT names
+    pretrained_orig = pretrained
+    pretrained = pretrained.lower()
+    if pretrained == "openai":
+        if amodel_name in _MODEL_CONFIGS:
+            logging.info(f"Loading {amodel_name} model config.")
+            model_cfg = deepcopy(_MODEL_CONFIGS[amodel_name])
+        else:
+            logging.error(
+                f"Model config for {amodel_name} not found; available models {list_models()}."
+            )
+            raise RuntimeError(f"Model config for {amodel_name} not found.")
+
+        logging.info(f"Loading pretrained ViT-B-16 text encoder from OpenAI.")
+        # Hard Code in model name
+        model_cfg["text_cfg"]["model_type"] = tmodel_name
+        model = load_openai_model(
+            "ViT-B-16",
+            model_cfg,
+            device=device,
+            jit=jit,
+            cache_dir=openai_model_cache_dir,
+            enable_fusion=enable_fusion,
+            fusion_type=fusion_type
+        )
+        # See https://discuss.pytorch.org/t/valueerror-attemting-to-unscale-fp16-gradients/81372
+        if precision == "amp" or precision == "fp32":
+            model = model.float()
+    else:
+        if amodel_name in _MODEL_CONFIGS:
+            logging.info(f"Loading {amodel_name} model config.")
+            model_cfg = deepcopy(_MODEL_CONFIGS[amodel_name])
+        else:
+            logging.error(
+                f"Model config for {amodel_name} not found; available models {list_models()}."
+            )
+            raise RuntimeError(f"Model config for {amodel_name} not found.")
+
+        if force_quick_gelu:
+            # override for use of QuickGELU on non-OpenAI transformer models
+            model_cfg["quick_gelu"] = True
+
+        # if pretrained_image:
+        #     if 'timm_amodel_name' in model_cfg.get('vision_cfg', {}):
+        #         # pretrained weight loading for timm models set via vision_cfg
+        #         model_cfg['vision_cfg']['timm_model_pretrained'] = True
+        #     else:
+        #         assert False, 'pretrained image towers currently only supported for timm models'
+        model_cfg["text_cfg"]["model_type"] = tmodel_name
+        model_cfg["enable_fusion"] = enable_fusion
+        model_cfg["fusion_type"] = fusion_type
+        model = CLAP(**model_cfg)
+
+        if pretrained:
+            checkpoint_path = ""
+            url = get_pretrained_url(amodel_name, pretrained)
+            if url:
+                checkpoint_path = download_pretrained(url, root=openai_model_cache_dir)
+            elif os.path.exists(pretrained_orig):
+                checkpoint_path = pretrained_orig
+            if checkpoint_path:
+                logging.info(f"Loading pretrained {amodel_name}-{tmodel_name} weights ({pretrained}).")
+                ckpt = load_state_dict(checkpoint_path, skip_params=True)
+                model.load_state_dict(ckpt)
+                param_names = [n for n, p in model.named_parameters()]
+                for n in param_names:
+                    print(n, "\t", "Loaded" if n in ckpt else "Unloaded")
+            else:
+                logging.warning(
+                    f"Pretrained weights ({pretrained}) not found for model {amodel_name}."
+                )
+                raise RuntimeError(
+                    f"Pretrained weights ({pretrained}) not found for model {amodel_name}."
+                )
+
+        if pretrained_audio:
+            if amodel_name.startswith('PANN'):
+                if 'Cnn14_mAP' in pretrained_audio:  # official checkpoint
+                    audio_ckpt = torch.load(pretrained_audio, map_location='cpu')
+                    audio_ckpt = audio_ckpt['model']
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if 'spectrogram_extractor' not in key and 'logmel_extractor' not in key:
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt['audio_branch.' + key] = v
+                elif os.path.basename(pretrained_audio).startswith('PANN'):  # checkpoint trained via HTSAT codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location='cpu')
+                    audio_ckpt = audio_ckpt['state_dict']
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if key.startswith('sed_model'):
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt['audio_branch.' + key[10:]] = v
+                elif os.path.basename(pretrained_audio).startswith('finetuned'):  # checkpoint trained via linear probe codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location='cpu')
+                else:
+                    raise ValueError('Unknown audio checkpoint')
+            elif amodel_name.startswith('HTSAT'):
+                if 'HTSAT_AudioSet_Saved' in pretrained_audio:  # official checkpoint
+                    audio_ckpt = torch.load(pretrained_audio, map_location='cpu')
+                    audio_ckpt = audio_ckpt['state_dict']
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if key.startswith('sed_model') and ('spectrogram_extractor' not in key
+                                                            and 'logmel_extractor' not in key):
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt['audio_branch.' + key[10:]] = v
+                elif os.path.basename(pretrained_audio).startswith('HTSAT'):  # checkpoint trained via HTSAT codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location='cpu')
+                    audio_ckpt = audio_ckpt['state_dict']
+                    keys = list(audio_ckpt.keys())
+                    for key in keys:
+                        if key.startswith('sed_model'):
+                            v = audio_ckpt.pop(key)
+                            audio_ckpt['audio_branch.' + key[10:]] = v
+                elif os.path.basename(pretrained_audio).startswith('finetuned'):  # checkpoint trained via linear probe codebase
+                    audio_ckpt = torch.load(pretrained_audio, map_location='cpu')
+                else:
+                    raise ValueError('Unknown audio checkpoint')
+            else:
+                raise f'this audio encoder pretrained checkpoint is not support'
+
+            model.load_state_dict(audio_ckpt, strict=False)
+            logging.info(f"Loading pretrained {amodel_name} weights ({pretrained_audio}).")
+            param_names = [n for n, p in model.named_parameters()]
+            for n in param_names:
+                print(n, "\t", "Loaded" if n in audio_ckpt else "Unloaded")
+            
+        model.to(device=device)
+        if precision == "fp16":
+            assert device.type != "cpu"
+            convert_weights_to_fp16(model)
+
+        if jit:
+            model = torch.jit.script(model)
+
+    return model, model_cfg
+
+
+def create_model_and_transforms(
+    model_name: str,
+    pretrained: str = "",
+    precision: str = "fp32",
+    device: torch.device = torch.device("cpu"),
+    jit: bool = False,
+    force_quick_gelu: bool = False,
+    # pretrained_image: bool = False,
+):
+    model = create_model(
+        model_name,
+        pretrained,
+        precision,
+        device,
+        jit,
+        force_quick_gelu=force_quick_gelu,
+        # pretrained_image=pretrained_image
+    )
+    preprocess_train = image_transform(model.visual.image_size, is_train=True)
+    preprocess_val = image_transform(model.visual.image_size, is_train=False)
+    return model, preprocess_train, preprocess_val
+
+
+def list_models():
+    """enumerate available model architectures based on config files"""
+    return list(_MODEL_CONFIGS.keys())
+
+
+def add_model_config(path):
+    """add model config path or file and update registry"""
+    if not isinstance(path, Path):
+        path = Path(path)
+    _MODEL_CONFIG_PATHS.append(path)
+    _rescan_model_configs()

ldm/modules/encoders/open_clap/feature_fusion.py

@@ -0,0 +1,193 @@
+'''
+Feature Fusion for Varible-Length Data Processing
+AFF/iAFF is referred and modified from https://github.com/YimianDai/open-aff/blob/master/aff_pytorch/aff_net/fusion.py
+According to the paper: Yimian Dai et al, Attentional Feature Fusion, IEEE Winter Conference on Applications of Computer Vision, WACV 2021
+'''
+
+import torch
+import torch.nn as nn
+
+
+class DAF(nn.Module):
+    '''
+    直接相加 DirectAddFuse
+    '''
+
+    def __init__(self):
+        super(DAF, self).__init__()
+
+    def forward(self, x, residual):
+        return x + residual
+
+
+class iAFF(nn.Module):
+    '''
+    多特征融合 iAFF
+    '''
+
+    def __init__(self, channels=64, r=4, type='2D'):
+        super(iAFF, self).__init__()
+        inter_channels = int(channels // r)
+
+        if type == '1D':
+            # 本地注意力
+            self.local_att = nn.Sequential(
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+
+            # 全局注意力
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool1d(1),
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+
+            # 第二次本地注意力
+            self.local_att2 = nn.Sequential(
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+            # 第二次全局注意力
+            self.global_att2 = nn.Sequential(
+                nn.AdaptiveAvgPool1d(1),
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+        elif type == '2D':
+            # 本地注意力
+            self.local_att = nn.Sequential(
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+
+            # 全局注意力
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+
+            # 第二次本地注意力
+            self.local_att2 = nn.Sequential(
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+            # 第二次全局注意力
+            self.global_att2 = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+        else:
+            raise f'the type is not supported'
+
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x, residual):
+        flag = False
+        xa = x + residual
+        if xa.size(0) == 1:
+            xa = torch.cat([xa,xa],dim=0)
+            flag = True
+        xl = self.local_att(xa)
+        xg = self.global_att(xa)
+        xlg = xl + xg
+        wei = self.sigmoid(xlg)
+        xi = x * wei + residual * (1 - wei)
+
+        xl2 = self.local_att2(xi)
+        xg2 = self.global_att(xi)
+        xlg2 = xl2 + xg2
+        wei2 = self.sigmoid(xlg2)
+        xo = x * wei2 + residual * (1 - wei2)
+        if flag:
+            xo = xo[0].unsqueeze(0)
+        return xo
+
+
+class AFF(nn.Module):
+    '''
+    多特征融合 AFF
+    '''
+
+    def __init__(self, channels=64, r=4, type='2D'):
+        super(AFF, self).__init__()
+        inter_channels = int(channels // r)
+
+        if type == '1D':
+            self.local_att = nn.Sequential(
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool1d(1),
+                nn.Conv1d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv1d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm1d(channels),
+            )
+        elif type == '2D':
+            self.local_att = nn.Sequential(
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+            self.global_att = nn.Sequential(
+                nn.AdaptiveAvgPool2d(1),
+                nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(inter_channels),
+                nn.ReLU(inplace=True),
+                nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
+                nn.BatchNorm2d(channels),
+            )
+        else:
+            raise f'the type is not supported.'
+        
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x, residual):
+        flag = False
+        xa = x + residual
+        if xa.size(0) == 1:
+            xa = torch.cat([xa,xa],dim=0)
+            flag = True
+        xl = self.local_att(xa)
+        xg = self.global_att(xa)
+        xlg = xl + xg
+        wei = self.sigmoid(xlg)
+        xo = 2 * x * wei + 2 * residual * (1 - wei)
+        if flag:
+            xo = xo[0].unsqueeze(0)
+        return xo
+

ldm/modules/encoders/open_clap/htsat.py

@@ -0,0 +1,1022 @@
+# Ke Chen
+# [email protected]
+# HTS-AT: A HIERARCHICAL TOKEN-SEMANTIC AUDIO TRANSFORMER FOR SOUND CLASSIFICATION AND DETECTION
+# Some layers designed on the model
+# below codes are based and referred from https://github.com/microsoft/Swin-Transformer
+# Swin Transformer for Computer Vision: https://arxiv.org/pdf/2103.14030.pdf
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from itertools import repeat
+import collections.abc
+import math
+import warnings
+
+from torch.nn.init import _calculate_fan_in_and_fan_out
+import torch.utils.checkpoint as checkpoint
+
+import random
+
+from torchlibrosa.stft import Spectrogram, LogmelFilterBank
+from torchlibrosa.augmentation import SpecAugmentation
+
+from itertools import repeat
+from .utils import do_mixup, interpolate
+
+from .feature_fusion import iAFF, AFF, DAF
+
+# from PyTorch internals
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.abc.Iterable):
+            return x
+        return tuple(repeat(x, n))
+    return parse
+
+to_1tuple = _ntuple(1)
+to_2tuple = _ntuple(2)
+to_3tuple = _ntuple(3)
+to_4tuple = _ntuple(4)
+to_ntuple = _ntuple
+
+def drop_path(x, drop_prob: float = 0., training: bool = False):
+    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+    This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
+    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
+    changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
+    'survival rate' as the argument.
+    """
+    if drop_prob == 0. or not training:
+        return x
+    keep_prob = 1 - drop_prob
+    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
+    random_tensor.floor_()  # binarize
+    output = x.div(keep_prob) * random_tensor
+    return output
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+
+class PatchEmbed(nn.Module):
+    """ 2D Image to Patch Embedding
+    """
+    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768, norm_layer=None, flatten=True, patch_stride = 16,
+        enable_fusion=False, fusion_type='None'):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        patch_stride = to_2tuple(patch_stride)
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patch_stride = patch_stride
+        self.grid_size = (img_size[0] // patch_stride[0], img_size[1] // patch_stride[1])
+        self.num_patches = self.grid_size[0] * self.grid_size[1]
+        self.flatten = flatten
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+        
+        padding = ((patch_size[0] - patch_stride[0]) // 2, (patch_size[1] - patch_stride[1]) // 2)
+
+        if (self.enable_fusion) and (self.fusion_type == 'channel_map'):
+            self.proj = nn.Conv2d(in_chans*4, embed_dim, kernel_size=patch_size, stride=patch_stride, padding=padding)
+        else:
+            self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_stride, padding=padding)
+        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
+
+        if (self.enable_fusion) and (self.fusion_type in ['daf_2d','aff_2d','iaff_2d']):
+            self.mel_conv2d = nn.Conv2d(in_chans, embed_dim, kernel_size=(patch_size[0], patch_size[1]*3), stride=(patch_stride[0], patch_stride[1] * 3), padding=padding)
+            if self.fusion_type == 'daf_2d':
+                self.fusion_model = DAF()
+            elif self.fusion_type == 'aff_2d':
+                self.fusion_model = AFF(channels=embed_dim, type='2D')
+            elif self.fusion_type == 'iaff_2d':
+                self.fusion_model = iAFF(channels=embed_dim, type='2D')    
+    def forward(self, x, longer_idx = None):
+        if (self.enable_fusion) and (self.fusion_type in ['daf_2d','aff_2d','iaff_2d']):
+            global_x = x[:,0:1,:,:]
+            
+
+            # global processing
+            B, C, H, W = global_x.shape
+            assert H == self.img_size[0] and W == self.img_size[1], \
+                f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+            global_x = self.proj(global_x)
+            TW = global_x.size(-1)
+            if len(longer_idx) > 0:
+                # local processing
+                local_x = x[longer_idx,1:,:,:].contiguous()
+                B, C, H, W = local_x.shape
+                local_x = local_x.view(B*C,1,H,W)
+                local_x = self.mel_conv2d(local_x)
+                local_x = local_x.view(B,C,local_x.size(1),local_x.size(2),local_x.size(3))
+                local_x = local_x.permute((0,2,3,1,4)).contiguous().flatten(3)
+                TB,TC,TH,_ = local_x.size()
+                if local_x.size(-1) < TW:
+                    local_x = torch.cat([local_x, torch.zeros((TB,TC,TH,TW-local_x.size(-1)), device=global_x.device)], dim=-1)
+                else:
+                    local_x = local_x[:,:,:,:TW]
+                
+                global_x[longer_idx] = self.fusion_model(global_x[longer_idx],local_x)
+            x = global_x
+        else:
+            B, C, H, W = x.shape
+            assert H == self.img_size[0] and W == self.img_size[1], \
+                f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+            x = self.proj(x)
+        
+        if self.flatten:
+            x = x.flatten(2).transpose(1, 2)  # BCHW -> BNC
+        x = self.norm(x)
+        return x
+
+class Mlp(nn.Module):
+    """ MLP as used in Vision Transformer, MLP-Mixer and related networks
+    """
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+def _no_grad_trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1. + math.erf(x / math.sqrt(2.))) / 2.
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn("mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+                      "The distribution of values may be incorrect.",
+                      stacklevel=2)
+
+    with torch.no_grad():
+        # Values are generated by using a truncated uniform distribution and
+        # then using the inverse CDF for the normal distribution.
+        # Get upper and lower cdf values
+        l = norm_cdf((a - mean) / std)
+        u = norm_cdf((b - mean) / std)
+
+        # Uniformly fill tensor with values from [l, u], then translate to
+        # [2l-1, 2u-1].
+        tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+        # Use inverse cdf transform for normal distribution to get truncated
+        # standard normal
+        tensor.erfinv_()
+
+        # Transform to proper mean, std
+        tensor.mul_(std * math.sqrt(2.))
+        tensor.add_(mean)
+
+        # Clamp to ensure it's in the proper range
+        tensor.clamp_(min=a, max=b)
+        return tensor
+
+
+def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
+    # type: (Tensor, float, float, float, float) -> Tensor
+    r"""Fills the input Tensor with values drawn from a truncated
+    normal distribution. The values are effectively drawn from the
+    normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \leq \text{mean} \leq b`.
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    Examples:
+        >>> w = torch.empty(3, 5)
+        >>> nn.init.trunc_normal_(w)
+    """
+    return _no_grad_trunc_normal_(tensor, mean, std, a, b)
+
+
+def variance_scaling_(tensor, scale=1.0, mode='fan_in', distribution='normal'):
+    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
+    if mode == 'fan_in':
+        denom = fan_in
+    elif mode == 'fan_out':
+        denom = fan_out
+    elif mode == 'fan_avg':
+        denom = (fan_in + fan_out) / 2
+
+    variance = scale / denom
+
+    if distribution == "truncated_normal":
+        # constant is stddev of standard normal truncated to (-2, 2)
+        trunc_normal_(tensor, std=math.sqrt(variance) / .87962566103423978)
+    elif distribution == "normal":
+        tensor.normal_(std=math.sqrt(variance))
+    elif distribution == "uniform":
+        bound = math.sqrt(3 * variance)
+        tensor.uniform_(-bound, bound)
+    else:
+        raise ValueError(f"invalid distribution {distribution}")
+
+
+def lecun_normal_(tensor):
+    variance_scaling_(tensor, mode='fan_in', distribution='truncated_normal')
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    r""" Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x, attn
+
+    def extra_repr(self):
+        return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}'
+
+
+# We use the model based on Swintransformer Block, therefore we can use the swin-transformer pretrained model
+class SwinTransformerBlock(nn.Module):
+    r""" Swin Transformer Block.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resulotion.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm, norm_before_mlp='ln'):
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        self.norm_before_mlp = norm_before_mlp
+        if min(self.input_resolution) <= self.window_size:
+            # if window size is larger than input resolution, we don't partition windows
+            self.shift_size = 0
+            self.window_size = min(self.input_resolution)
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        if self.norm_before_mlp == 'ln':
+            self.norm2 = nn.LayerNorm(dim)
+        elif self.norm_before_mlp == 'bn':
+            self.norm2 = lambda x: nn.BatchNorm1d(dim)(x.transpose(1, 2)).transpose(1, 2)
+        else:
+            raise NotImplementedError
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        if self.shift_size > 0:
+            # calculate attention mask for SW-MSA
+            H, W = self.input_resolution
+            img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
+            h_slices = (slice(0, -self.window_size),
+                        slice(-self.window_size, -self.shift_size),
+                        slice(-self.shift_size, None))
+            w_slices = (slice(0, -self.window_size),
+                        slice(-self.window_size, -self.shift_size),
+                        slice(-self.shift_size, None))
+            cnt = 0
+            for h in h_slices:
+                for w in w_slices:
+                    img_mask[:, h, w, :] = cnt
+                    cnt += 1
+
+            mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+            mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+            attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+            attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+        else:
+            attn_mask = None
+
+        self.register_buffer("attn_mask", attn_mask)
+
+    def forward(self, x):
+        # pdb.set_trace()
+        H, W = self.input_resolution
+        # print("H: ", H)
+        # print("W: ", W)
+        # pdb.set_trace()
+        B, L, C = x.shape
+        # assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+        else:
+            shifted_x = x
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows, attn = self.attn(x_windows, mask=self.attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+        x = x.view(B, H * W, C)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x, attn
+
+    def extra_repr(self):
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
+               f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
+
+
+
+class PatchMerging(nn.Module):
+    r""" Patch Merging Layer.
+    Args:
+        input_resolution (tuple[int]): Resolution of input feature.
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x):
+        """
+        x: B, H*W, C
+        """
+        H, W = self.input_resolution
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+        assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
+
+        x = x.view(B, H, W, C)
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+    def extra_repr(self):
+        return f"input_resolution={self.input_resolution}, dim={self.dim}"
+
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self, dim, input_resolution, depth, num_heads, window_size,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
+                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False,
+                 norm_before_mlp='ln'):
+
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
+                                 num_heads=num_heads, window_size=window_size,
+                                 shift_size=0 if (i % 2 == 0) else window_size // 2,
+                                 mlp_ratio=mlp_ratio,
+                                 qkv_bias=qkv_bias, qk_scale=qk_scale,
+                                 drop=drop, attn_drop=attn_drop,
+                                 drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                                 norm_layer=norm_layer, norm_before_mlp=norm_before_mlp)
+            for i in range(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x):
+        attns = []
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x)
+            else:
+                x, attn = blk(x)
+                if not self.training:
+                    attns.append(attn.unsqueeze(0))
+        if self.downsample is not None:
+            x = self.downsample(x)
+        if not self.training:
+            attn = torch.cat(attns, dim = 0)
+            attn = torch.mean(attn, dim = 0)
+        return x, attn
+
+    def extra_repr(self):
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
+
+
+# The Core of HTSAT
+class HTSAT_Swin_Transformer(nn.Module):
+    r"""HTSAT based on the Swin Transformer
+    Args:
+        spec_size (int | tuple(int)): Input Spectrogram size. Default 256
+        patch_size (int | tuple(int)): Patch size. Default: 4
+        path_stride (iot | tuple(int)): Patch Stride for Frequency and Time Axis. Default: 4
+        in_chans (int): Number of input image channels. Default: 1 (mono)
+        num_classes (int): Number of classes for classification head. Default: 527
+        embed_dim (int): Patch embedding dimension. Default: 96
+        depths (tuple(int)): Depth of each HTSAT-Swin Transformer layer.
+        num_heads (tuple(int)): Number of attention heads in different layers.
+        window_size (int): Window size. Default: 8
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None
+        drop_rate (float): Dropout rate. Default: 0
+        attn_drop_rate (float): Attention dropout rate. Default: 0
+        drop_path_rate (float): Stochastic depth rate. Default: 0.1
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
+        config (module): The configuration Module from config.py
+    """
+
+    def __init__(self, spec_size=256, patch_size=4, patch_stride=(4,4), 
+                in_chans=1, num_classes=527,
+                 embed_dim=96, depths=[2, 2, 6, 2], num_heads=[4, 8, 16, 32],
+                 window_size=8, mlp_ratio=4., qkv_bias=True, qk_scale=None,
+                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
+                 norm_layer=nn.LayerNorm, 
+                 ape=False, patch_norm=True,
+                 use_checkpoint=False, norm_before_mlp='ln', config = None, 
+                 enable_fusion = False, fusion_type = 'None', **kwargs):
+        super(HTSAT_Swin_Transformer, self).__init__()
+
+        self.config = config
+        self.spec_size = spec_size 
+        self.patch_stride = patch_stride
+        self.patch_size = patch_size
+        self.window_size = window_size
+        self.embed_dim = embed_dim
+        self.depths = depths
+        self.ape = ape
+        self.in_chans = in_chans
+        self.num_classes = num_classes
+        self.num_heads = num_heads
+        self.num_layers = len(self.depths)
+        self.num_features = int(self.embed_dim * 2 ** (self.num_layers - 1))
+        
+        self.drop_rate = drop_rate
+        self.attn_drop_rate = attn_drop_rate
+        self.drop_path_rate = drop_path_rate
+
+        self.qkv_bias = qkv_bias
+        self.qk_scale = None
+
+        self.patch_norm = patch_norm
+        self.norm_layer = norm_layer if self.patch_norm else None
+        self.norm_before_mlp = norm_before_mlp
+        self.mlp_ratio = mlp_ratio
+
+        self.use_checkpoint = use_checkpoint
+
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+
+        #  process mel-spec ; used only once
+        self.freq_ratio = self.spec_size // self.config.mel_bins
+        window = 'hann'
+        center = True
+        pad_mode = 'reflect'
+        ref = 1.0
+        amin = 1e-10
+        top_db = None
+        self.interpolate_ratio = 32     # Downsampled ratio
+        # Spectrogram extractor
+        self.spectrogram_extractor = Spectrogram(n_fft=config.window_size, hop_length=config.hop_size, 
+            win_length=config.window_size, window=window, center=center, pad_mode=pad_mode, 
+            freeze_parameters=True)
+        # Logmel feature extractor
+        self.logmel_extractor = LogmelFilterBank(sr=config.sample_rate, n_fft=config.window_size, 
+            n_mels=config.mel_bins, fmin=config.fmin, fmax=config.fmax, ref=ref, amin=amin, top_db=top_db, 
+            freeze_parameters=True)
+        # Spec augmenter
+        self.spec_augmenter = SpecAugmentation(time_drop_width=64, time_stripes_num=2, 
+            freq_drop_width=8, freq_stripes_num=2) # 2 2
+        self.bn0 = nn.BatchNorm2d(self.config.mel_bins)
+
+
+        # split spctrogram into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            img_size=self.spec_size, patch_size=self.patch_size, in_chans=self.in_chans, 
+            embed_dim=self.embed_dim, norm_layer=self.norm_layer, patch_stride = patch_stride,
+            enable_fusion=self.enable_fusion, fusion_type=self.fusion_type
+            )
+
+        num_patches = self.patch_embed.num_patches
+        patches_resolution = self.patch_embed.grid_size
+        self.patches_resolution = patches_resolution
+
+        # absolute position embedding
+        if self.ape:
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, self.embed_dim))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=self.drop_rate)
+
+        # stochastic depth
+        dpr = [x.item() for x in torch.linspace(0, self.drop_path_rate, sum(self.depths))]  # stochastic depth decay rule
+
+        # build layers
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = BasicLayer(dim=int(self.embed_dim * 2 ** i_layer),
+                input_resolution=(patches_resolution[0] // (2 ** i_layer),
+                                    patches_resolution[1] // (2 ** i_layer)),
+                depth=self.depths[i_layer],
+                num_heads=self.num_heads[i_layer],
+                window_size=self.window_size,
+                mlp_ratio=self.mlp_ratio,
+                qkv_bias=self.qkv_bias, qk_scale=self.qk_scale,
+                drop=self.drop_rate, attn_drop=self.attn_drop_rate,
+                drop_path=dpr[sum(self.depths[:i_layer]):sum(self.depths[:i_layer + 1])],
+                norm_layer=self.norm_layer,
+                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                use_checkpoint=use_checkpoint,
+                norm_before_mlp=self.norm_before_mlp)
+            self.layers.append(layer)
+
+        self.norm = self.norm_layer(self.num_features)
+        self.avgpool = nn.AdaptiveAvgPool1d(1)
+        self.maxpool = nn.AdaptiveMaxPool1d(1)
+        
+        SF = self.spec_size // (2 ** (len(self.depths) - 1)) // self.patch_stride[0] // self.freq_ratio
+        self.tscam_conv = nn.Conv2d(
+            in_channels = self.num_features,
+            out_channels = self.num_classes,
+            kernel_size = (SF,3),
+            padding = (0,1)
+        )
+        self.head = nn.Linear(num_classes, num_classes)
+
+        if (self.enable_fusion) and (self.fusion_type in ['daf_1d','aff_1d','iaff_1d']):
+            self.mel_conv1d = nn.Sequential(
+                nn.Conv1d(64, 64, kernel_size=5, stride=3, padding=2),
+                nn.BatchNorm1d(64)
+            )
+            if self.fusion_type == 'daf_1d':
+                self.fusion_model = DAF()
+            elif self.fusion_type == 'aff_1d':
+                self.fusion_model = AFF(channels=64, type='1D')
+            elif self.fusion_type == 'iaff_1d':
+                self.fusion_model = iAFF(channels=64, type='1D')
+                
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'absolute_pos_embed'}
+
+    @torch.jit.ignore
+    def no_weight_decay_keywords(self):
+        return {'relative_position_bias_table'}
+
+
+    def forward_features(self, x, longer_idx = None):
+        # A deprecated optimization for using a hierarchical output from different blocks
+
+        frames_num = x.shape[2]        
+        x = self.patch_embed(x, longer_idx = longer_idx)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+        for i, layer in enumerate(self.layers):
+            x, attn = layer(x)
+        # for x
+        x = self.norm(x)
+        B, N, C = x.shape
+        SF = frames_num // (2 ** (len(self.depths) - 1)) // self.patch_stride[0]
+        ST = frames_num // (2 ** (len(self.depths) - 1)) // self.patch_stride[1]
+        x = x.permute(0,2,1).contiguous().reshape(B, C, SF, ST)
+        B, C, F, T = x.shape
+        # group 2D CNN
+        c_freq_bin = F // self.freq_ratio
+        x = x.reshape(B, C, F // c_freq_bin, c_freq_bin, T)
+        x = x.permute(0,1,3,2,4).contiguous().reshape(B, C, c_freq_bin, -1)
+        # get latent_output
+        fine_grained_latent_output = torch.mean(x, dim = 2)
+        fine_grained_latent_output = interpolate(fine_grained_latent_output.permute(0,2,1).contiguous(), 8 * self.patch_stride[1]) 
+        
+        latent_output = self.avgpool(torch.flatten(x,2))
+        latent_output = torch.flatten(latent_output, 1)
+
+        # display the attention map, if needed
+
+        x = self.tscam_conv(x)
+        x = torch.flatten(x, 2) # B, C, T
+ 
+        fpx = interpolate(torch.sigmoid(x).permute(0,2,1).contiguous(), 8 * self.patch_stride[1]) 
+            
+        x = self.avgpool(x)
+        x = torch.flatten(x, 1)
+
+        output_dict = {
+            'framewise_output': fpx, # already sigmoided
+            'clipwise_output': torch.sigmoid(x),
+            'fine_grained_embedding': fine_grained_latent_output,
+            'embedding': latent_output
+        }
+
+        return output_dict
+
+    def crop_wav(self, x, crop_size, spe_pos = None):
+        time_steps = x.shape[2]
+        tx = torch.zeros(x.shape[0], x.shape[1], crop_size, x.shape[3]).to(x.device)
+        for i in range(len(x)):
+            if spe_pos is None:
+                crop_pos = random.randint(0, time_steps - crop_size - 1)
+            else:
+                crop_pos = spe_pos
+            tx[i][0] = x[i, 0, crop_pos:crop_pos + crop_size,:]
+        return tx
+
+    # Reshape the wavform to a img size, if you want to use the pretrained swin transformer model
+    def reshape_wav2img(self, x):
+        B, C, T, F = x.shape
+        target_T = int(self.spec_size * self.freq_ratio)
+        target_F = self.spec_size // self.freq_ratio
+        assert T <= target_T and F <= target_F, "the wav size should less than or equal to the swin input size"
+        # to avoid bicubic zero error
+        if T < target_T:
+            x = nn.functional.interpolate(x, (target_T, x.shape[3]), mode="bicubic", align_corners=True)
+        if F < target_F:
+            x = nn.functional.interpolate(x, (x.shape[2], target_F), mode="bicubic", align_corners=True)
+        x = x.permute(0,1,3,2).contiguous()
+        x = x.reshape(x.shape[0], x.shape[1], x.shape[2], self.freq_ratio, x.shape[3] // self.freq_ratio)
+        # print(x.shape)
+        x = x.permute(0,1,3,2,4).contiguous()
+        x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3], x.shape[4])
+        return x
+    
+    # Repeat the wavform to a img size, if you want to use the pretrained swin transformer model
+    def repeat_wat2img(self, x, cur_pos):
+        B, C, T, F = x.shape
+        target_T = int(self.spec_size * self.freq_ratio)
+        target_F = self.spec_size // self.freq_ratio
+        assert T <= target_T and F <= target_F, "the wav size should less than or equal to the swin input size"
+        # to avoid bicubic zero error
+        if T < target_T:
+            x = nn.functional.interpolate(x, (target_T, x.shape[3]), mode="bicubic", align_corners=True)
+        if F < target_F:
+            x = nn.functional.interpolate(x, (x.shape[2], target_F), mode="bicubic", align_corners=True)  
+        x = x.permute(0,1,3,2).contiguous() # B C F T
+        x = x[:,:,:,cur_pos:cur_pos + self.spec_size]
+        x = x.repeat(repeats = (1,1,4,1))
+        return x
+
+    def forward(self, x: torch.Tensor, mixup_lambda = None, infer_mode = False, device=None):# out_feat_keys: List[str] = None):
+
+        if self.enable_fusion and x["longer"].sum() == 0:
+            # if no audio is longer than 10s, then randomly select one audio to be longer
+            x["longer"][torch.randint(0, x["longer"].shape[0], (1,))] = True
+
+        if not self.enable_fusion:
+            x = x["waveform"].to(device=device, non_blocking=True)
+            x = self.spectrogram_extractor(x)   # (batch_size, 1, time_steps, freq_bins)
+            x = self.logmel_extractor(x)    # (batch_size, 1, time_steps, mel_bins)
+            x = x.transpose(1, 3)
+            x = self.bn0(x)
+            x = x.transpose(1, 3)
+            if self.training:
+                x = self.spec_augmenter(x)
+
+            if self.training and mixup_lambda is not None:
+                x = do_mixup(x, mixup_lambda)
+                
+            x = self.reshape_wav2img(x)
+            output_dict = self.forward_features(x)
+        else:
+            longer_list = x["longer"].to(device=device, non_blocking=True)
+            x = x["mel_fusion"].to(device=device, non_blocking=True)
+            x = x.transpose(1, 3)
+            x = self.bn0(x)
+            x = x.transpose(1, 3)
+            longer_list_idx = torch.where(longer_list)[0]
+            if self.fusion_type in ['daf_1d','aff_1d','iaff_1d']:
+                new_x = x[:,0:1,:,:].clone().contiguous()
+                if len(longer_list_idx) > 0:
+                # local processing
+                    fusion_x_local = x[longer_list_idx,1:,:,:].clone().contiguous()
+                    FB,FC,FT,FF = fusion_x_local.size()
+                    fusion_x_local = fusion_x_local.view(FB * FC, FT, FF)
+                    fusion_x_local = torch.permute(fusion_x_local, (0,2,1)).contiguous()
+                    fusion_x_local = self.mel_conv1d(fusion_x_local)
+                    fusion_x_local = fusion_x_local.view(FB,FC,FF,fusion_x_local.size(-1))
+                    fusion_x_local = torch.permute(fusion_x_local, (0,2,1,3)).contiguous().flatten(2)
+                    if fusion_x_local.size(-1) < FT:
+                        fusion_x_local = torch.cat([fusion_x_local, torch.zeros((FB,FF,FT- fusion_x_local.size(-1)), device=device)], dim=-1)
+                    else:
+                        fusion_x_local = fusion_x_local[:,:,:FT]
+                    # 1D fusion
+                    new_x = new_x.squeeze(1).permute((0,2,1)).contiguous()
+                    new_x[longer_list_idx] = self.fusion_model(new_x[longer_list_idx], fusion_x_local)
+                    x = new_x.permute((0,2,1)).contiguous()[:,None,:,:]
+                else:
+                    x = new_x
+
+            elif self.fusion_type in ['daf_2d','aff_2d','iaff_2d','channel_map']:
+                x = x # no change
+
+            if self.training:
+                x = self.spec_augmenter(x)
+            if self.training and mixup_lambda is not None:
+                x = do_mixup(x, mixup_lambda)
+
+            x = self.reshape_wav2img(x)
+            output_dict = self.forward_features(x, longer_idx = longer_list_idx)
+       
+        # if infer_mode:
+        #     # in infer mode. we need to handle different length audio input
+        #     frame_num = x.shape[2]
+        #     target_T = int(self.spec_size * self.freq_ratio)
+        #     repeat_ratio = math.floor(target_T / frame_num)
+        #     x = x.repeat(repeats=(1,1,repeat_ratio,1))
+        #     x = self.reshape_wav2img(x)
+        #     output_dict = self.forward_features(x)
+        # else:
+        #     if x.shape[2] > self.freq_ratio * self.spec_size:
+        #         if self.training:
+        #             x = self.crop_wav(x, crop_size=self.freq_ratio * self.spec_size)
+        #             x = self.reshape_wav2img(x)
+        #             output_dict = self.forward_features(x)
+        #         else:
+        #             # Change: Hard code here
+        #             overlap_size = (x.shape[2] - 1) // 4
+        #             output_dicts = []
+        #             crop_size = (x.shape[2] - 1) // 2
+        #             for cur_pos in range(0, x.shape[2] - crop_size - 1, overlap_size):
+        #                 tx = self.crop_wav(x, crop_size = crop_size, spe_pos = cur_pos)
+        #                 tx = self.reshape_wav2img(tx)
+        #                 output_dicts.append(self.forward_features(tx))
+        #             clipwise_output = torch.zeros_like(output_dicts[0]["clipwise_output"]).float().to(x.device)
+        #             framewise_output = torch.zeros_like(output_dicts[0]["framewise_output"]).float().to(x.device)
+        #             for d in output_dicts:
+        #                 clipwise_output += d["clipwise_output"]
+        #                 framewise_output += d["framewise_output"]
+        #             clipwise_output  = clipwise_output / len(output_dicts)
+        #             framewise_output = framewise_output / len(output_dicts)
+        #             output_dict = {
+        #                 'framewise_output': framewise_output, 
+        #                 'clipwise_output': clipwise_output
+        #             }
+        #     else: # this part is typically used, and most easy one
+        #         x = self.reshape_wav2img(x)
+        #         output_dict = self.forward_features(x)
+        # x = self.head(x)
+
+        # We process the data in the dataloader part, in that here we only consider the input_T < fixed_T
+        
+        
+
+        return output_dict
+
+def create_htsat_model(audio_cfg, enable_fusion=False, fusion_type='None'):
+    try:
+
+        assert audio_cfg.model_name in ["tiny", "base", "large"], "model name for HTS-AT is wrong!"
+        if audio_cfg.model_name == "tiny":
+            model = HTSAT_Swin_Transformer(
+                spec_size=256,
+                patch_size=4,
+                patch_stride=(4,4),
+                num_classes=audio_cfg.class_num,
+                embed_dim=96,
+                depths=[2,2,6,2],
+                num_heads=[4,8,16,32],
+                window_size=8,
+                config = audio_cfg,
+                enable_fusion = enable_fusion,
+                fusion_type = fusion_type
+            )
+        elif audio_cfg.model_name == "base":
+            model = HTSAT_Swin_Transformer(
+                spec_size=256,
+                patch_size=4,
+                patch_stride=(4,4),
+                num_classes=audio_cfg.class_num,
+                embed_dim=128,
+                depths=[2,2,12,2],
+                num_heads=[4,8,16,32],
+                window_size=8,
+                config = audio_cfg,
+                enable_fusion = enable_fusion,
+                fusion_type = fusion_type
+            )
+        elif audio_cfg.model_name == "large":
+            model = HTSAT_Swin_Transformer(
+                spec_size=256,
+                patch_size=4,
+                patch_stride=(4,4),
+                num_classes=audio_cfg.class_num,
+                embed_dim=256,
+                depths=[2,2,12,2],
+                num_heads=[4,8,16,32],
+                window_size=8,
+                config = audio_cfg,
+                enable_fusion = enable_fusion,
+                fusion_type = fusion_type
+            )
+        
+        return model
+    except:
+        raise RuntimeError(f'Import Model for {audio_cfg.model_name} not found, or the audio cfg parameters are not enough.')
+        

ldm/modules/encoders/open_clap/linear_probe.py

@@ -0,0 +1,63 @@
+import numpy as np
+import torch.nn.functional as F
+from torch import nn
+from .model import MLPLayers
+
+
+class LinearProbe(nn.Module):
+    def __init__(self, model, mlp, freeze, in_ch, out_ch, act=None):
+        """
+        Args:
+            model: nn.Module
+            mlp: bool, if True, then use the MLP layer as the linear probe module
+            freeze: bool, if Ture, then freeze all the CLAP model's layers when training the linear probe
+            in_ch: int, the output channel from CLAP model
+            out_ch: int, the output channel from linear probe (class_num)
+            act: torch.nn.functional, the activation function before the loss function
+        """
+        super().__init__()
+        in_ch = 512
+        self.clap_model = model
+        self.clap_model.text_branch = None  # to save memory
+        self.freeze = freeze
+        if mlp:
+            self.lp_layer = MLPLayers(units=[in_ch, in_ch * 2, out_ch])
+        else:
+            self.lp_layer = nn.Linear(in_ch, out_ch)
+
+        if self.freeze:
+            for param in self.clap_model.parameters():
+                param.requires_grad = False
+
+        if act == 'None':
+            self.act = None
+        elif act == 'relu':
+            self.act = nn.ReLU()
+        elif act == 'elu':
+            self.act = nn.ELU()
+        elif act == 'prelu':
+            self.act = nn.PReLU(num_parameters=in_ch)
+        elif act == 'softmax':
+            self.act = nn.Softmax(dim=-1)
+        elif act == 'sigmoid':
+            self.act = nn.Sigmoid()
+
+    def forward(self, x, mix_lambda=None, device=None):
+        """
+        Args:
+            x: waveform, torch.tensor [batch, t_samples] / batch of mel_spec and longer list
+            mix_lambda: torch.tensor [batch], the mixup lambda
+        Returns:
+            class_prob: torch.tensor [batch, class_num]
+
+        """
+        # batchnorm cancel grandient
+        if self.freeze:
+            self.clap_model.eval()
+
+        x = self.clap_model.audio_projection(
+            self.clap_model.audio_branch(x, mixup_lambda=mix_lambda, device=device)["embedding"])
+        out = self.lp_layer(x)
+        if self.act is not None:
+            out = self.act(out)
+        return out

ldm/modules/encoders/open_clap/loss.py

@@ -0,0 +1,307 @@
+from multiprocessing.sharedctypes import Value
+import torch
+import torch.distributed.nn
+from torch import distributed as dist, nn as nn
+from torch.nn import functional as F
+import numpy as np
+from sklearn.metrics import average_precision_score, roc_auc_score, accuracy_score 
+
+try:
+    import horovod.torch as hvd
+except ImportError:
+    hvd = None
+
+
+def gather_features(
+        audio_features,
+        text_features,
+        audio_features_mlp=None, 
+        text_features_mlp=None,
+        local_loss=False,
+        gather_with_grad=False,
+        rank=0,
+        world_size=1,
+        use_horovod=False,
+        mlp_loss=False
+):
+    if use_horovod:
+        assert hvd is not None, 'Please install horovod'
+        if gather_with_grad:
+            all_audio_features = hvd.allgather(audio_features)
+            all_text_features = hvd.allgather(text_features)
+            if mlp_loss:
+                all_audio_features_mlp = hvd.allgather(audio_features_mlp)
+                all_text_features_mlp = hvd.allgather(text_features_mlp)
+        else:
+            with torch.no_grad():
+                all_audio_features = hvd.allgather(audio_features)
+                all_text_features = hvd.allgather(text_features)
+                if mlp_loss:
+                    all_audio_features_mlp = hvd.allgather(audio_features_mlp)
+                    all_text_features_mlp = hvd.allgather(text_features_mlp)
+            if not local_loss:
+                # ensure grads for local rank when all_* features don't have a gradient
+                gathered_audio_features = list(all_audio_features.chunk(world_size, dim=0))
+                gathered_text_features = list(all_text_features.chunk(world_size, dim=0))
+                gathered_audio_features[rank] = audio_features
+                gathered_text_features[rank] = text_features
+                all_audio_features = torch.cat(gathered_audio_features, dim=0)
+                all_text_features = torch.cat(gathered_text_features, dim=0)
+                if mlp_loss:
+                    gathered_audio_features_mlp = list(all_audio_features_mlp.chunk(world_size, dim=0))
+                    gathered_text_features_mlp = list(all_text_features_mlp.chunk(world_size, dim=0))
+                    gathered_audio_features_mlp[rank] = audio_features_mlp
+                    gathered_text_features_mlp[rank] = text_features_mlp
+                    all_audio_features_mlp = torch.cat(gathered_audio_features_mlp, dim=0)
+                    all_text_features_mlp = torch.cat(gathered_text_features_mlp, dim=0)
+    else:
+        # We gather tensors from all gpus
+        if gather_with_grad:
+            all_audio_features = torch.cat(torch.distributed.nn.all_gather(audio_features), dim=0)
+            all_text_features = torch.cat(torch.distributed.nn.all_gather(text_features), dim=0)
+            if mlp_loss:
+                all_audio_features_mlp = torch.cat(torch.distributed.nn.all_gather(audio_features_mlp), dim=0)
+                all_text_features_mlp = torch.cat(torch.distributed.nn.all_gather(text_features_mlp), dim=0)
+        else:
+            gathered_audio_features = [torch.zeros_like(audio_features) for _ in range(world_size)]
+            gathered_text_features = [torch.zeros_like(text_features) for _ in range(world_size)]
+            dist.all_gather(gathered_audio_features, audio_features)
+            dist.all_gather(gathered_text_features, text_features)
+            if mlp_loss:
+                gathered_audio_features_mlp = [torch.zeros_like(audio_features_mlp) for _ in range(world_size)]
+                gathered_text_features_mlp = [torch.zeros_like(text_features_mlp) for _ in range(world_size)]
+                dist.all_gather(gathered_audio_features_mlp, audio_features_mlp)
+                dist.all_gather(gathered_text_features_mlp, text_features_mlp)
+            if not local_loss:
+                # ensure grads for local rank when all_* features don't have a gradient
+                gathered_audio_features[rank] = audio_features
+                gathered_text_features[rank] = text_features
+                if mlp_loss:
+                    gathered_audio_features_mlp[rank] = audio_features_mlp
+                    gathered_text_features_mlp[rank] = text_features_mlp
+
+            all_audio_features = torch.cat(gathered_audio_features, dim=0)
+            all_text_features = torch.cat(gathered_text_features, dim=0)
+            if mlp_loss:
+                all_audio_features_mlp = torch.cat(gathered_audio_features_mlp, dim=0)
+                all_text_features_mlp = torch.cat(gathered_text_features_mlp, dim=0)
+    if mlp_loss:
+        return all_audio_features, all_text_features, all_audio_features_mlp, all_text_features_mlp
+    else:
+        return all_audio_features, all_text_features
+
+class ClipLoss(nn.Module):
+
+    def __init__(
+            self,
+            local_loss=False,
+            gather_with_grad=False,
+            cache_labels=False,
+            rank=0,
+            world_size=1,
+            use_horovod=False,
+            mlp_loss=False,
+            weight_loss_kappa=0,
+    ):
+        super().__init__()
+        self.local_loss = local_loss
+        self.gather_with_grad = gather_with_grad
+        self.cache_labels = cache_labels
+        self.rank = rank
+        self.world_size = world_size
+        self.use_horovod = use_horovod
+        self.mlp_loss = mlp_loss
+        self.weighted_loss = bool(weight_loss_kappa!=0)
+        self.weight_loss_kappa = weight_loss_kappa
+        # cache state
+        self.prev_num_logits = 0
+        self.labels = {}
+
+    def forward(self, audio_features, text_features, logit_scale_a, logit_scale_t=None, audio_features_mlp=None, text_features_mlp=None):
+        device = audio_features.device
+        if self.mlp_loss:
+            if self.world_size > 1:
+                all_audio_features, all_text_features, all_audio_features_mlp, all_text_features_mlp = gather_features(
+                    audio_features=audio_features,text_features=text_features,
+                    audio_features_mlp=audio_features_mlp,text_features_mlp=text_features_mlp,
+                    local_loss=self.local_loss,gather_with_grad=self.gather_with_grad,
+                    rank=self.rank,world_size=self.world_size,use_horovod=self.use_horovod,
+                    mlp_loss=self.mlp_loss
+                )
+                if self.local_loss:
+                    a_logits_per_audio = logit_scale_a * audio_features @ all_text_features_mlp.T
+                    a_logits_per_text = logit_scale_a * text_features_mlp @ all_audio_features.T
+                    t_logits_per_audio = logit_scale_t * audio_features_mlp @ all_text_features.T
+                    t_logits_per_text = logit_scale_t * text_features @ all_audio_features_mlp.T
+                else:
+                    a_logits_per_audio = logit_scale_a * all_audio_features @ all_text_features_mlp.T
+                    a_logits_per_text = a_logits_per_audio.T
+                    t_logits_per_audio = logit_scale_t * all_audio_features_mlp @ all_text_features.T
+                    t_logits_per_text = t_logits_per_audio.T
+            else:
+                a_logits_per_audio = logit_scale_a * audio_features @ text_features_mlp.T
+                a_logits_per_text = logit_scale_a * text_features_mlp @ audio_features.T
+                t_logits_per_audio = logit_scale_t * audio_features_mlp @ text_features.T
+                t_logits_per_text = logit_scale_t * text_features @ audio_features_mlp.T
+
+            # calculated ground-truth and cache if enabled
+            num_logits = a_logits_per_audio.shape[0]
+            if self.prev_num_logits != num_logits or device not in self.labels:
+                labels = torch.arange(num_logits, device=device, dtype=torch.long)
+                if self.world_size > 1 and self.local_loss:
+                    labels = labels + num_logits * self.rank
+                if self.cache_labels:
+                    self.labels[device] = labels
+                    self.prev_num_logits = num_logits
+            else:
+                labels = self.labels[device]
+
+            if not self.weighted_loss:
+                total_loss = (
+                    F.cross_entropy(a_logits_per_audio, labels) +
+                    F.cross_entropy(a_logits_per_text, labels) + 
+                    F.cross_entropy(t_logits_per_audio, labels) +
+                    F.cross_entropy(t_logits_per_text, labels) 
+                    ) / 4
+            else:
+                audio_weight = (audio_features@audio_features.T).detach()
+                audio_weight = (torch.exp(torch.sum(audio_weight, axis=1)/(self.weight_loss_kappa*len(audio_weight)))).detach()
+                text_weight = (text_features@text_features.T).detach()
+                text_weight = (torch.exp(torch.sum(text_weight, axis=1)/(self.weight_loss_kappa*len(text_features)))).detach()
+                total_loss = (
+                    F.cross_entropy(a_logits_per_audio, labels, weight=audio_weight) +
+                    F.cross_entropy(a_logits_per_text, labels, weight=audio_weight) + 
+                    F.cross_entropy(t_logits_per_audio, labels, weight=text_weight) +
+                    F.cross_entropy(t_logits_per_text, labels, weight=text_weight) 
+                    ) / 4
+        else:
+            if self.world_size > 1:
+                all_audio_features, all_text_features = gather_features(
+                    audio_features=audio_features,text_features=text_features,
+                    local_loss=self.local_loss,gather_with_grad=self.gather_with_grad,
+                    rank=self.rank,world_size=self.world_size,use_horovod=self.use_horovod,
+                    mlp_loss=self.mlp_loss
+                )
+
+                if self.local_loss:
+                    logits_per_audio = logit_scale_a * audio_features @ all_text_features.T
+                    logits_per_text = logit_scale_a * text_features @ all_audio_features.T
+                else:
+                    logits_per_audio = logit_scale_a * all_audio_features @ all_text_features.T
+                    logits_per_text = logits_per_audio.T
+            else:
+                logits_per_audio = logit_scale_a * audio_features @ text_features.T
+                logits_per_text = logit_scale_a * text_features @ audio_features.T
+
+            # calculated ground-truth and cache if enabled
+            num_logits = logits_per_audio.shape[0]
+            if self.prev_num_logits != num_logits or device not in self.labels:
+                labels = torch.arange(num_logits, device=device, dtype=torch.long)
+                if self.world_size > 1 and self.local_loss:
+                    labels = labels + num_logits * self.rank
+                if self.cache_labels:
+                    self.labels[device] = labels
+                    self.prev_num_logits = num_logits
+            else:
+                labels = self.labels[device]
+            if not self.weighted_loss:
+                total_loss = (
+                    F.cross_entropy(logits_per_audio, labels) +
+                    F.cross_entropy(logits_per_text, labels)
+                    ) / 2
+            else:
+                audio_weight = (all_audio_features@all_audio_features.T).detach()
+                audio_weight = (torch.exp(torch.sum(audio_weight, axis=1)/(self.weight_loss_kappa*len(all_audio_features)))).detach()
+                text_weight = (all_text_features@all_text_features.T).detach()
+                text_weight = (torch.exp(torch.sum(text_weight, axis=1)/(self.weight_loss_kappa*len(all_text_features)))).detach()
+                total_loss = (
+                    F.cross_entropy(logits_per_audio, labels, weight=text_weight) +
+                    F.cross_entropy(logits_per_text, labels, weight=audio_weight)
+                    ) / 2
+        return total_loss
+
+def lp_gather_features(
+        pred,
+        target,
+        world_size=1,
+        use_horovod=False
+):
+    if use_horovod:
+        assert hvd is not None, 'Please install horovod'
+        with torch.no_grad():
+            all_preds = hvd.allgather(pred)
+            all_targets = hvd.allgath(target)
+    else:
+        gathered_preds = [torch.zeros_like(pred) for _ in range(world_size)]
+        gathered_targets = [torch.zeros_like(target) for _ in range(world_size)]
+
+        dist.all_gather(gathered_preds, pred)
+        dist.all_gather(gathered_targets, target)
+        all_preds = torch.cat(gathered_preds, dim=0)
+        all_targets = torch.cat(gathered_targets, dim=0)
+
+    return all_preds, all_targets
+
+
+def get_map(pred, target):
+    pred = torch.sigmoid(pred).numpy()
+    target = target.numpy()
+    return np.mean(average_precision_score(target, pred, average=None))
+
+def get_acc(pred, target):
+    pred = torch.argmax(pred,1).numpy()
+    target = torch.argmax(target,1).numpy()
+    return accuracy_score(target, pred)
+
+def get_mauc(pred, target):
+    pred = torch.sigmoid(pred).numpy()
+    target = target.numpy()
+    return np.mean(roc_auc_score(target, pred, average=None))
+
+
+class LPMetrics(object):
+    def __init__(self, metric_names = ['map','acc','mauc']):
+        self.metrics = []
+        for name in metric_names:
+            self.metrics.append(self.get_metric(name))
+        self.metric_names = metric_names
+
+    def get_metric(self,name):
+        if name == 'map':
+            return get_map
+        elif name == 'acc':
+            return get_acc
+        elif name == 'mauc':
+            return get_mauc
+        else:
+            raise ValueError(f'the metric should be at least one of [map, acc, mauc]')
+
+    def evaluate_mertics(self, pred, target):
+        metric_dict = {}
+        for i in range(len(self.metric_names)):
+            metric_dict[self.metric_names[i]] = self.metrics[i](pred, target)
+        return metric_dict
+
+
+def calc_celoss(pred, target):
+    target = torch.argmax(target, 1).long()
+    return nn.CrossEntropyLoss()(pred, target)
+
+
+class LPLoss(nn.Module):
+
+    def __init__(self, loss_name):
+        super().__init__()
+        if loss_name == 'bce':
+            self.loss_func = nn.BCEWithLogitsLoss()
+        elif loss_name == 'ce':
+            self.loss_func = calc_celoss
+        elif loss_name == 'mse':
+            self.loss_func = nn.MSELoss()
+        else:
+            raise ValueError(f'the loss func should be at least one of [bce, ce, mse]')
+
+    def forward(self, pred, target):
+        loss = self.loss_func(pred, target)
+        return loss
+        

ldm/modules/encoders/open_clap/model.py

@@ -0,0 +1,913 @@
+""" CLAP Model
+
+Adapted from CLIP: https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
+Adapted to the Audio Task.
+"""
+
+from collections import OrderedDict
+from dataclasses import dataclass
+from email.mime import audio
+from typing import Tuple, Union, Callable, Optional
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from .timm_model import TimmModel
+import logging
+from .utils import freeze_batch_norm_2d
+
+from .pann_model import create_pann_model
+from .htsat import create_htsat_model
+from transformers import BertModel, RobertaModel, BartModel
+from transformers.tokenization_utils_base import BatchEncoding
+
+
+class MLPLayers(nn.Module):
+    def __init__(self, units=[512, 512, 512], nonlin=nn.ReLU(), dropout=0.1):
+        super(MLPLayers, self).__init__()
+        self.nonlin = nonlin
+        self.dropout = dropout
+
+        sequence = []
+        for u0, u1 in zip(units[:-1], units[1:]):
+            sequence.append(nn.Linear(u0, u1))
+            sequence.append(self.nonlin)
+            sequence.append(nn.Dropout(self.dropout))
+        sequence = sequence[:-2]
+
+        self.sequential = nn.Sequential(*sequence)
+
+    def forward(self, X):
+        X = self.sequential(X)
+        return X
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1):
+        super().__init__()
+
+        # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1
+        self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+
+        self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+
+        self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity()
+
+        self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)
+        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
+
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = None
+        self.stride = stride
+
+        if stride > 1 or inplanes != planes * Bottleneck.expansion:
+            # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1
+            self.downsample = nn.Sequential(
+                OrderedDict(
+                    [
+                        ("-1", nn.AvgPool2d(stride)),
+                        (
+                            "0",
+                            nn.Conv2d(
+                                inplanes,
+                                planes * self.expansion,
+                                1,
+                                stride=1,
+                                bias=False,
+                            ),
+                        ),
+                        ("1", nn.BatchNorm2d(planes * self.expansion)),
+                    ]
+                )
+            )
+
+    def forward(self, x: torch.Tensor):
+        identity = x
+
+        out = self.relu(self.bn1(self.conv1(x)))
+        out = self.relu(self.bn2(self.conv2(out)))
+        out = self.avgpool(out)
+        out = self.bn3(self.conv3(out))
+
+        if self.downsample is not None:
+            identity = self.downsample(x)
+
+        out += identity
+        out = self.relu(out)
+        return out
+
+
+class AttentionPool2d(nn.Module):
+    def __init__(
+        self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None
+    ):
+        super().__init__()
+        self.positional_embedding = nn.Parameter(
+            torch.randn(spacial_dim**2 + 1, embed_dim) / embed_dim**0.5
+        )
+        self.k_proj = nn.Linear(embed_dim, embed_dim)
+        self.q_proj = nn.Linear(embed_dim, embed_dim)
+        self.v_proj = nn.Linear(embed_dim, embed_dim)
+        self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
+        self.num_heads = num_heads
+
+    def forward(self, x):
+        x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3]).permute(
+            2, 0, 1
+        )  # NCHW -> (HW)NC
+        x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0)  # (HW+1)NC
+        x = x + self.positional_embedding[:, None, :].to(x.dtype)  # (HW+1)NC
+        x, _ = F.multi_head_attention_forward(
+            query=x,
+            key=x,
+            value=x,
+            embed_dim_to_check=x.shape[-1],
+            num_heads=self.num_heads,
+            q_proj_weight=self.q_proj.weight,
+            k_proj_weight=self.k_proj.weight,
+            v_proj_weight=self.v_proj.weight,
+            in_proj_weight=None,
+            in_proj_bias=torch.cat(
+                [self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]
+            ),
+            bias_k=None,
+            bias_v=None,
+            add_zero_attn=False,
+            dropout_p=0,
+            out_proj_weight=self.c_proj.weight,
+            out_proj_bias=self.c_proj.bias,
+            use_separate_proj_weight=True,
+            training=self.training,
+            need_weights=False,
+        )
+
+        return x[0]
+
+
+class ModifiedResNet(nn.Module):
+    """
+    A ResNet class that is similar to torchvision's but contains the following changes:
+    - There are now 3 "stem" convolutions as opposed to 1, with an average pool instead of a max pool.
+    - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1
+    - The final pooling layer is a QKV attention instead of an average pool
+    """
+
+    def __init__(self, layers, output_dim, heads, image_size=224, width=64):
+        super().__init__()
+        self.output_dim = output_dim
+        self.image_size = image_size
+
+        # the 3-layer stem
+        self.conv1 = nn.Conv2d(
+            3, width // 2, kernel_size=3, stride=2, padding=1, bias=False
+        )
+        self.bn1 = nn.BatchNorm2d(width // 2)
+        self.conv2 = nn.Conv2d(
+            width // 2, width // 2, kernel_size=3, padding=1, bias=False
+        )
+        self.bn2 = nn.BatchNorm2d(width // 2)
+        self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(width)
+        self.avgpool = nn.AvgPool2d(2)
+        self.relu = nn.ReLU(inplace=True)
+
+        # residual layers
+        self._inplanes = width  # this is a *mutable* variable used during construction
+        self.layer1 = self._make_layer(width, layers[0])
+        self.layer2 = self._make_layer(width * 2, layers[1], stride=2)
+        self.layer3 = self._make_layer(width * 4, layers[2], stride=2)
+        self.layer4 = self._make_layer(width * 8, layers[3], stride=2)
+
+        embed_dim = width * 32  # the ResNet feature dimension
+        self.attnpool = AttentionPool2d(image_size // 32, embed_dim, heads, output_dim)
+
+        self.init_parameters()
+
+    def _make_layer(self, planes, blocks, stride=1):
+        layers = [Bottleneck(self._inplanes, planes, stride)]
+
+        self._inplanes = planes * Bottleneck.expansion
+        for _ in range(1, blocks):
+            layers.append(Bottleneck(self._inplanes, planes))
+
+        return nn.Sequential(*layers)
+
+    def init_parameters(self):
+        if self.attnpool is not None:
+            std = self.attnpool.c_proj.in_features**-0.5
+            nn.init.normal_(self.attnpool.q_proj.weight, std=std)
+            nn.init.normal_(self.attnpool.k_proj.weight, std=std)
+            nn.init.normal_(self.attnpool.v_proj.weight, std=std)
+            nn.init.normal_(self.attnpool.c_proj.weight, std=std)
+
+        for resnet_block in [self.layer1, self.layer2, self.layer3, self.layer4]:
+            for name, param in resnet_block.named_parameters():
+                if name.endswith("bn3.weight"):
+                    nn.init.zeros_(param)
+
+    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
+        assert (
+            unlocked_groups == 0
+        ), "partial locking not currently supported for this model"
+        for param in self.parameters():
+            param.requires_grad = False
+        if freeze_bn_stats:
+            freeze_batch_norm_2d(self)
+
+    def stem(self, x):
+        for conv, bn in [
+            (self.conv1, self.bn1),
+            (self.conv2, self.bn2),
+            (self.conv3, self.bn3),
+        ]:
+            x = self.relu(bn(conv(x)))
+        x = self.avgpool(x)
+        return x
+
+    def forward(self, x):
+        x = self.stem(x)
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.attnpool(x)
+
+        return x
+
+
+class LayerNorm(nn.LayerNorm):
+    """Subclass torch's LayerNorm to handle fp16."""
+
+    def forward(self, x: torch.Tensor):
+        orig_type = x.dtype
+        x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
+        return x.to(orig_type)
+
+
+class QuickGELU(nn.Module):
+    # NOTE This is slower than nn.GELU or nn.SiLU and uses more GPU memory
+    def forward(self, x: torch.Tensor):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(self, d_model: int, n_head: int, act_layer: Callable = nn.GELU):
+        super().__init__()
+
+        self.attn = nn.MultiheadAttention(d_model, n_head)
+        self.ln_1 = LayerNorm(d_model)
+        self.mlp = nn.Sequential(
+            OrderedDict(
+                [
+                    ("c_fc", nn.Linear(d_model, d_model * 4)),
+                    ("gelu", act_layer()),
+                    ("c_proj", nn.Linear(d_model * 4, d_model)),
+                ]
+            )
+        )
+        self.ln_2 = LayerNorm(d_model)
+
+    def attention(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        return self.attn(x, x, x, need_weights=False, attn_mask=attn_mask)[0]
+
+    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        x = x + self.attention(self.ln_1(x), attn_mask=attn_mask)
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self, width: int, layers: int, heads: int, act_layer: Callable = nn.GELU
+    ):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(width, heads, act_layer=act_layer)
+                for _ in range(layers)
+            ]
+        )
+
+    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
+        for r in self.resblocks:
+            x = r(x, attn_mask=attn_mask)
+        return x
+
+
+class VisualTransformer(nn.Module):
+    def __init__(
+        self,
+        image_size: int,
+        patch_size: int,
+        width: int,
+        layers: int,
+        heads: int,
+        output_dim: int,
+        act_layer: Callable = nn.GELU,
+    ):
+        super().__init__()
+        self.image_size = image_size
+        self.output_dim = output_dim
+        self.conv1 = nn.Conv2d(
+            in_channels=3,
+            out_channels=width,
+            kernel_size=patch_size,
+            stride=patch_size,
+            bias=False,
+        )
+
+        scale = width**-0.5
+        self.class_embedding = nn.Parameter(scale * torch.randn(width))
+        self.positional_embedding = nn.Parameter(
+            scale * torch.randn((image_size // patch_size) ** 2 + 1, width)
+        )
+        self.ln_pre = LayerNorm(width)
+
+        self.text_branch = Transformer(width, layers, heads, act_layer=act_layer)
+
+        self.ln_post = LayerNorm(width)
+        self.proj = nn.Parameter(scale * torch.randn(width, output_dim))
+
+    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
+        assert (
+            unlocked_groups == 0
+        ), "partial locking not currently supported for this model"
+        for param in self.parameters():
+            param.requires_grad = False
+
+    def forward(self, x: torch.Tensor):
+        x = self.conv1(x)  # shape = [*, width, grid, grid]
+        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
+        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]
+        x = torch.cat(
+            [
+                self.class_embedding.to(x.dtype)
+                + torch.zeros(
+                    x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device
+                ),
+                x,
+            ],
+            dim=1,
+        )  # shape = [*, grid ** 2 + 1, width]
+        x = x + self.positional_embedding.to(x.dtype)
+        x = self.ln_pre(x)
+
+        x = x.permute(1, 0, 2)  # NLD -> LND
+        x = self.text_branch(x)
+        x = x.permute(1, 0, 2)  # LND -> NLD
+
+        x = self.ln_post(x[:, 0, :])
+
+        if self.proj is not None:
+            x = x @ self.proj
+
+        return x
+
+
+@dataclass
+class CLAPVisionCfg:
+    layers: Union[Tuple[int, int, int, int], int] = 12
+    width: int = 768
+    patch_size: int = 16
+    image_size: Union[Tuple[int, int], int] = 224
+    timm_model_name: str = (
+        None  # a valid model name overrides layers, width, patch_size
+    )
+    timm_model_pretrained: bool = (
+        False  # use (imagenet) pretrained weights for named model
+    )
+    timm_pool: str = (
+        "avg"  # feature pooling for timm model ('abs_attn', 'rot_attn', 'avg', '')
+    )
+    timm_proj: str = (
+        "linear"  # linear projection for timm model output ('linear', 'mlp', '')
+    )
+
+
+# Audio Config Class
+@dataclass
+class CLAPAudioCfp:
+    model_type: str = "PANN"
+    model_name: str = "Cnn14"
+    sample_rate: int = 48000
+    # Param
+    audio_length: int = 1024
+    window_size: int = 1024
+    hop_size: int = 1024
+    fmin: int = 50
+    fmax: int = 14000
+    class_num: int = 527
+    mel_bins: int = 64
+    clip_samples: int = 480000
+
+
+@dataclass
+class CLAPTextCfg:
+    context_length: int
+    vocab_size: int
+    width: int
+    heads: int
+    layers: int
+    model_type: str
+
+
+class CLAP(nn.Module):
+    def __init__(
+        self,
+        embed_dim: int,
+        audio_cfg: CLAPAudioCfp,
+        text_cfg: CLAPTextCfg,
+        quick_gelu: bool = False,
+        enable_fusion: bool = False,
+        fusion_type: str = 'None',
+        joint_embed_shape: int = 512,
+        mlp_act: str = 'relu',
+    ):
+        super().__init__()
+        if isinstance(audio_cfg, dict):
+            audio_cfg = CLAPAudioCfp(**audio_cfg)
+        if isinstance(text_cfg, dict):
+            text_cfg = CLAPTextCfg(**text_cfg)
+
+        self.audio_cfg = audio_cfg
+        self.text_cfg = text_cfg
+        self.enable_fusion = enable_fusion
+        self.fusion_type = fusion_type
+        self.joint_embed_shape = joint_embed_shape
+        self.mlp_act = mlp_act
+
+
+        self.context_length = text_cfg.context_length
+
+        # OpenAI models are pretrained w/ QuickGELU but native nn.GELU is both faster and more
+        # memory efficient in recent PyTorch releases (>= 1.10).
+        # NOTE: timm models always use native GELU regardless of quick_gelu flag.
+        act_layer = QuickGELU if quick_gelu else nn.GELU
+
+        if mlp_act == 'relu':
+            mlp_act_layer = nn.ReLU()
+        elif mlp_act == 'gelu':
+            mlp_act_layer = nn.GELU()
+        else:
+            raise NotImplementedError
+
+        # audio branch
+        # audio branch parameters
+        if audio_cfg.model_type == "PANN":
+            self.audio_branch = create_pann_model(audio_cfg, enable_fusion, fusion_type)
+        elif audio_cfg.model_type == "HTSAT":
+            self.audio_branch = create_htsat_model(audio_cfg, enable_fusion, fusion_type)
+        else:
+            logging.error(f"Model config for {audio_cfg.model_type} not found")
+            raise RuntimeError(f"Model config for {audio_cfg.model_type} not found.")
+
+
+        # text branch
+        # text branch parameters
+        if text_cfg.model_type == "transformer":
+            self.text_branch = Transformer(
+                width=text_cfg.width,
+                layers=text_cfg.layers,
+                heads=text_cfg.heads,
+                act_layer=act_layer,
+            )
+            self.vocab_size = text_cfg.vocab_size
+            self.token_embedding = nn.Embedding(text_cfg.vocab_size, text_cfg.width)
+            self.positional_embedding = nn.Parameter(
+                torch.empty(self.context_length, text_cfg.width)
+            )
+            self.ln_final = LayerNorm(text_cfg.width)
+            self.text_transform = MLPLayers(units=[self.joint_embed_shape,
+                                                   self.joint_embed_shape,
+                                                   self.joint_embed_shape], dropout=0.1)
+            self.text_projection = nn.Sequential(
+                nn.Linear(text_cfg.width, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape)
+            )
+        elif text_cfg.model_type == "bert":
+            self.text_branch = BertModel.from_pretrained("bert-base-uncased")
+            self.text_transform = MLPLayers(units=[self.joint_embed_shape,
+                                                   self.joint_embed_shape,
+                                                   self.joint_embed_shape], dropout=0.1)
+            self.text_projection = nn.Sequential(
+                nn.Linear(768, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape)
+            )
+        elif text_cfg.model_type == "roberta":
+            self.text_branch = RobertaModel.from_pretrained('roberta-base')
+            self.text_transform = MLPLayers(units=[self.joint_embed_shape,
+                                                   self.joint_embed_shape,
+                                                   self.joint_embed_shape], dropout=0.1)
+            self.text_projection = nn.Sequential(
+                nn.Linear(768, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape)
+            )
+        elif text_cfg.model_type == "bart":
+            self.text_branch = BartModel.from_pretrained('facebook/bart-base')
+            self.text_transform = MLPLayers(units=[self.joint_embed_shape,
+                                                   self.joint_embed_shape,
+                                                   self.joint_embed_shape], dropout=0.1)
+            self.text_projection = nn.Sequential(
+                nn.Linear(768, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape)
+            )
+        else:
+            logging.error(f"Model config for {text_cfg.model_type} not found")
+            raise RuntimeError(f"Model config for {text_cfg.model_type} not found.")
+        self.text_branch_type = text_cfg.model_type
+        # text branch parameters
+
+        # audio branch parameters
+        self.audio_transform = MLPLayers(units=[self.joint_embed_shape,
+                                                self.joint_embed_shape,
+                                                self.joint_embed_shape], dropout=0.1)
+
+        # below here is text branch parameters
+
+        # ============================================================================================================
+        self.audio_projection = nn.Sequential(
+                nn.Linear(embed_dim, self.joint_embed_shape),
+                mlp_act_layer,
+                nn.Linear(self.joint_embed_shape, self.joint_embed_shape)
+            )
+
+        self.logit_scale_a = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+        self.logit_scale_t = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
+        self.register_buffer("attn_mask", self.build_attention_mask(), persistent=False)
+
+        self.init_text_branch_parameters()
+
+    def init_text_branch_parameters(self):
+        if self.text_branch_type == "transformer":
+            nn.init.normal_(self.token_embedding.weight, std=0.02)
+            nn.init.normal_(self.positional_embedding, std=0.01)
+            proj_std = (self.text_branch.width**-0.5) * (
+                (2 * self.text_branch.layers) ** -0.5
+            )
+            attn_std = self.text_branch.width**-0.5
+            fc_std = (2 * self.text_branch.width) ** -0.5
+            for block in self.text_branch.resblocks:
+                nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
+                nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
+                nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
+                nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
+        if self.text_branch_type == "bert" or self.text_branch_type == "roberta":
+            width = self.text_branch.embeddings.word_embeddings.weight.shape[-1]
+        elif self.text_branch_type == "bart":
+            width = self.text_branch.shared.weight.shape[-1]
+        else:
+            width = self.text_branch.width
+        nn.init.constant_(self.logit_scale_a, np.log(1 / 0.07))
+        nn.init.constant_(self.logit_scale_t, np.log(1 / 0.07))
+
+        # deprecated
+        # if hasattr(self.visual, 'init_parameters'):
+        # self.visual.init_parameters()
+
+        # if self.text_projection is not None:
+        #     nn.init.normal_(self.text_projection, std=width**-0.5)
+
+    def build_attention_mask(self):
+        # lazily create causal attention mask, with full attention between the vision tokens
+        # pytorch uses additive attention mask; fill with -inf
+        mask = torch.empty(self.context_length, self.context_length)
+        mask.fill_(float("-inf"))
+        mask.triu_(1)  # zero out the lower diagonal
+        return mask
+
+    def encode_audio(self, audio, device):
+        return self.audio_branch(audio, mixup_lambda=None, device=device)  # mix lambda needs to add
+
+    # def list_of_dict_of_tensor2dict_of_tensor(self, x, device):
+    #     tmp = {}
+    #     for k in x[0].keys():
+    #         tmp[k] = []
+    #         for i in range(len(x)):
+    #             tmp[k].append(x[i][k][:77])
+    #     for k in x[0].keys():
+    #         tmp[k] = torch.tensor(tmp[k]).to(device=device, non_blocking=True)
+    #     return tmp
+
+    def encode_text(self, text, device):
+        if self.text_branch_type == "transformer":
+            text = text.to(device=device, non_blocking=True)
+            x = self.token_embedding(text)  # [batch_size, n_ctx, d_model]
+
+            x = x + self.positional_embedding
+            x = x.permute(1, 0, 2)  # NLD -> LND
+            x = self.text_branch(x, attn_mask=self.attn_mask)
+            x = x.permute(1, 0, 2)  # LND -> NLD
+            x = self.ln_final(x)
+
+            # x.shape = [batch_size, n_ctx, transformer.width]
+            # take features from the eot embedding (eot_token is the highest number in each sequence)
+            x = self.text_projection(x[torch.arange(x.shape[0]), text.argmax(dim=-1)])
+        elif self.text_branch_type == "bert":
+            # text = self.list_of_dict_of_tensor2dict_of_tensor(text, device)
+            # text = BatchEncoding(text)
+            x = self.text_branch(
+                input_ids=text["input_ids"].to(device=device, non_blocking=True),
+                attention_mask=text["attention_mask"].to(
+                    device=device, non_blocking=True
+                ),
+                token_type_ids=text["token_type_ids"].to(
+                    device=device, non_blocking=True
+                ),
+            )["pooler_output"]
+            x = self.text_projection(x)
+        elif self.text_branch_type == "roberta":
+            x = self.text_branch(
+                input_ids=text["input_ids"].to(device=device, non_blocking=True),
+                attention_mask=text["attention_mask"].to(
+                    device=device, non_blocking=True
+                ),
+            )["pooler_output"]
+
+            x = self.text_projection(x)
+        elif self.text_branch_type == "bart":
+            x = torch.mean(self.text_branch(
+                input_ids=text["input_ids"].to(device=device, non_blocking=True),
+                attention_mask=text["attention_mask"].to(
+                    device=device, non_blocking=True
+                ),
+            )["encoder_last_hidden_state"],axis=1)
+            x = self.text_projection(x)
+        else:
+            logging.error(f"Model type {self.text_branch_type} not found")
+            raise RuntimeError(f"Model type {self.text_branch_type} not found.")
+        return x
+
+    def forward(self, audio, text, device=None):
+        """Forward audio and text into the CLAP
+
+        Parameters
+        ----------
+        audio: torch.Tensor (batch_size, audio_length)
+            the time-domain audio input / the batch of mel_spec and longer list.
+        text: torch.Tensor () // need to add
+            the text token input
+        """
+        if device is None:
+            if audio is not None:
+                device = audio.device
+            elif text is not None:
+                device = text.device
+        if audio is None and text is None:
+            # a hack to get the logit scale
+            return self.logit_scale_a.exp(), self.logit_scale_t.exp()
+        elif audio is None:
+            return self.encode_text(text, device=device)
+        elif text is None:
+            return self.audio_projection(self.encode_audio(audio, device=device)["embedding"])
+        audio_features = self.audio_projection(self.encode_audio(audio, device=device)["embedding"])
+        audio_features = F.normalize(audio_features, dim=-1)
+
+        text_features = self.encode_text(
+            text, device=device
+        )
+        # print("text_features", text_features)
+        # print("text_features.shape", text_features.shape)
+        # print("text_features.type", type(text_features))
+        text_features = F.normalize(text_features, dim=-1)
+
+        audio_features_mlp = self.audio_transform(audio_features)
+        text_features_mlp = self.text_transform(text_features)
+        # Four outputs: audio features (basic & MLP), text features (basic & MLP)
+        return (
+            audio_features,
+            text_features,
+            audio_features_mlp,
+            text_features_mlp,
+            self.logit_scale_a.exp(),
+            self.logit_scale_t.exp(),
+        )
+
+    def get_logit_scale(self):
+        return self.logit_scale_a.exp(), self.logit_scale_t.exp()
+
+    def get_textual_embedding(self, data):
+
+        device = next(self.parameters()).device
+        for k in data:
+            data[k] = data[k].to(device)
+
+        # if self.text_branch_type == "roberta":
+        text_embeds = self.text_branch(
+                input_ids=data["input_ids"].to(device=device, non_blocking=True),
+                attention_mask=data["attention_mask"].to(device=device, non_blocking=True),
+            )["last_hidden_state"]
+
+        text_embeds = self.text_projection(text_embeds)
+
+        text_embeds = F.normalize(text_embeds, dim=-1)
+
+        return text_embeds
+
+    def get_text_embedding(self, data):
+        """Get the text embedding from the model
+
+        Parameters
+        ----------
+        data: torch.Tensor 
+            a tensor of text embedding
+
+        Returns
+        ----------
+        text_embed: torch.Tensor
+            a tensor of text_embeds (N, D)
+
+        """
+        device = next(self.parameters()).device
+        for k in data:
+            data[k] = data[k].to(device)
+        text_embeds = self.encode_text(data, device=device)
+        text_embeds = F.normalize(text_embeds, dim=-1)
+        
+        return text_embeds
+
+    def get_audio_embedding(self, data):
+        """Get the audio embedding from the model
+
+        Parameters
+        ----------
+        data: a list of dict
+            the audio input dict list from 'get_audio_feature' method
+
+        Returns
+        ----------
+        audio_embed: torch.Tensor
+            a tensor of audio_embeds (N, D)
+
+        """
+        device = next(self.parameters()).device
+        input_dict = {}
+        keys = data[0].keys()
+        for k in keys:
+            input_dict[k] = torch.cat([d[k].unsqueeze(0) for d in data], dim=0).to(device)
+        
+        audio_embeds = self.audio_projection(self.encode_audio(input_dict, device=device)["embedding"])
+        audio_embeds = F.normalize(audio_embeds, dim=-1)
+        
+        return audio_embeds
+
+            
+
+    def audio_infer(self, audio, hopsize=None, device=None):
+        """Forward one audio and produce the audio embedding
+
+        Parameters
+        ----------
+        audio:  (audio_length)
+            the time-domain audio input, notice that it must be only one input
+        hopsize: int
+            the overlap hopsize as the sliding window
+
+        Returns
+        ----------
+        output_dict: {
+            key: [n, (embedding_shape)] if "HTS-AT"
+            or
+            key: [(embedding_shape)] if "PANN"
+        }
+            the list of key values of the audio branch
+
+        """
+
+        assert not self.training, "the inference mode must be run at eval stage"
+        output_dict = {}
+        # PANN
+        if self.audio_cfg.model_type == "PANN":
+            audio_input = audio.unsqueeze(dim=0)
+            output_dict[key] = self.encode_audio(audio_input, device=device)[key].squeeze(dim=0)
+        elif self.audio_cfg.model_type == "HTSAT":
+            # repeat
+            audio_len = len(audio)
+            k = self.audio_cfg.clip_samples // audio_len
+            if k > 1:
+                audio = audio.repeat(k)
+                audio_len = len(audio)
+
+            if hopsize is None:
+                hopsize = min(hopsize, audio_len)
+
+            if audio_len > self.audio_cfg.clip_samples:
+                audio_input = [
+                    audio[pos : pos + self.audio_cfg.clip_samples].clone()
+                    for pos in range(
+                        0, audio_len - self.audio_cfg.clip_samples, hopsize
+                    )
+                ]
+                audio_input.append(audio[-self.audio_cfg.clip_samples :].clone())
+                audio_input = torch.stack(audio_input)
+                output_dict[key] = self.encode_audio(audio_input, device=device)[key]
+            else:
+                audio_input = audio.unsqueeze(dim=0)
+                output_dict[key] = self.encode_audio(audio_input, device=device)[key].squeeze(dim=0)
+
+        return output_dict
+
+
+def convert_weights_to_fp16(model: nn.Module):
+    """Convert applicable model parameters to fp16"""
+
+    def _convert_weights_to_fp16(l):
+        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
+            l.weight.data = l.weight.data.half()
+            if l.bias is not None:
+                l.bias.data = l.bias.data.half()
+
+        if isinstance(l, nn.MultiheadAttention):
+            for attr in [
+                *[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]],
+                "in_proj_bias",
+                "bias_k",
+                "bias_v",
+            ]:
+                tensor = getattr(l, attr)
+                if tensor is not None:
+                    tensor.data = tensor.data.half()
+
+        for name in ["text_projection", "proj"]:
+            if hasattr(l, name):
+                attr = getattr(l, name)
+                if attr is not None:
+                    attr.data = attr.data.half()
+
+    model.apply(_convert_weights_to_fp16)
+
+
+# Ignore the state dict of the vision part
+def build_model_from_openai_state_dict(state_dict: dict, model_cfg, enable_fusion: bool = False, fusion_type: str = 'None'):
+
+    embed_dim = model_cfg["embed_dim"]
+    audio_cfg = model_cfg["audio_cfg"]
+    text_cfg = model_cfg["text_cfg"]
+    context_length = state_dict["positional_embedding"].shape[0]
+    vocab_size = state_dict["token_embedding.weight"].shape[0]
+    transformer_width = state_dict["ln_final.weight"].shape[0]
+    transformer_heads = transformer_width // 64
+    transformer_layers = len(
+        set(
+            k.split(".")[2]
+            for k in state_dict
+            if k.startswith(f"transformer.resblocks")
+        )
+    )
+
+    audio_cfg = CLAPAudioCfp(**audio_cfg)
+    text_cfg = CLAPTextCfg(**text_cfg)
+
+    model = CLAP(
+        embed_dim,
+        audio_cfg=audio_cfg,
+        text_cfg=text_cfg,
+        quick_gelu=True,  # OpenAI models were trained with QuickGELU
+        enable_fusion=enable_fusion,
+        fusion_type=fusion_type
+    )
+    state_dict["logit_scale_a"] = state_dict["logit_scale"]
+    state_dict["logit_scale_t"] = state_dict["logit_scale"]
+    pop_keys = list(state_dict.keys())[::]
+    # pop the visual branch saved weights
+    for key in pop_keys:
+        if key.startswith("visual."):
+            state_dict.pop(key, None)
+
+    for key in ["logit_scale", "input_resolution", "context_length", "vocab_size"]:
+        state_dict.pop(key, None)
+
+    # not use fp16
+    # convert_weights_to_fp16(model)
+    model.load_state_dict(state_dict, strict=False)
+    return model.eval()
+
+
+def trace_model(model, batch_size=256, device=torch.device("cpu")):
+    model.eval()
+    audio_length = model.audio_cfg.audio_length
+    example_audio = torch.ones((batch_size, audio_length), device=device)
+    example_text = torch.zeros(
+        (batch_size, model.context_length), dtype=torch.int, device=device
+    )
+    model = torch.jit.trace_module(
+        model,
+        inputs=dict(
+            forward=(example_audio, example_text),
+            encode_text=(example_text,),
+            encode_image=(example_audio,),
+        ),
+    )
+    model.audio_cfg.audio_length = audio_length  # Question: what does this do?
+    return model

ldm/modules/encoders/open_clap/model_configs/HTSAT-base.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 1024,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "base"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

ldm/modules/encoders/open_clap/model_configs/HTSAT-large.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "large"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

ldm/modules/encoders/open_clap/model_configs/HTSAT-tiny-win-1536.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 768,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1536,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "tiny"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

ldm/modules/encoders/open_clap/model_configs/HTSAT-tiny.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 768,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "HTSAT",
+        "model_name": "tiny"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

ldm/modules/encoders/open_clap/model_configs/PANN-10.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 1024,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn10"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

ldm/modules/encoders/open_clap/model_configs/PANN-14-fmax-18k.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 18000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

ldm/modules/encoders/open_clap/model_configs/PANN-14-fmax-8k-20s.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 960000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 360,
+        "fmin": 50,
+        "fmax": 8000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 12
+    }
+}

ldm/modules/encoders/open_clap/model_configs/PANN-14-tiny-transformer.json

@@ -0,0 +1,23 @@
+{
+    "embed_dim": 2048,
+    "audio_cfg": {
+        "audio_length": 1024,
+        "clip_samples": 480000,
+        "mel_bins": 64,
+        "sample_rate": 48000,
+        "window_size": 1024,
+        "hop_size": 480,
+        "fmin": 50,
+        "fmax": 14000,
+        "class_num": 527,
+        "model_type": "PANN",
+        "model_name": "Cnn14"
+    },
+    "text_cfg": {
+        "context_length": 77,
+        "vocab_size": 49408,
+        "width": 512,
+        "heads": 8,
+        "layers": 4
+    }
+}