Anonymous commited on
Commit
e9a044b
1 Parent(s): 6c71197

update app

Browse files
Files changed (39) hide show
  1. app.py +247 -0
  2. configs/inference_t2v_1024_v1.0.yaml +77 -0
  3. configs/inference_t2v_1024_v1.0_freenoise.yaml +77 -0
  4. configs/inference_t2v_tconv256_v1.0.yaml +77 -0
  5. configs/inference_t2v_tconv256_v1.0_freenoise.yaml +77 -0
  6. configs/inference_t2v_tconv512_v1.0.yaml +77 -0
  7. configs/inference_t2v_tconv512_v1.0_freenoise.yaml +77 -0
  8. lvdm/basics.py +100 -0
  9. lvdm/common.py +95 -0
  10. lvdm/distributions.py +95 -0
  11. lvdm/ema.py +76 -0
  12. lvdm/models/autoencoder.py +219 -0
  13. lvdm/models/ddpm3d.py +763 -0
  14. lvdm/models/samplers/ddim.py +336 -0
  15. lvdm/models/samplers/ddim_mp.py +340 -0
  16. lvdm/models/utils_diffusion.py +104 -0
  17. lvdm/modules/attention.py +475 -0
  18. lvdm/modules/attention_freenoise.py +565 -0
  19. lvdm/modules/encoders/condition.py +392 -0
  20. lvdm/modules/encoders/ip_resampler.py +136 -0
  21. lvdm/modules/networks/ae_modules.py +845 -0
  22. lvdm/modules/networks/openaimodel3d.py +581 -0
  23. lvdm/modules/networks/openaimodel3d_freenoise.py +581 -0
  24. lvdm/modules/x_transformer.py +640 -0
  25. prompts/mp_prompts.txt +2 -0
  26. prompts/single_prompts.txt +2 -0
  27. requirements.txt +23 -0
  28. scripts/evaluation/ddp_wrapper.py +46 -0
  29. scripts/evaluation/funcs.py +353 -0
  30. scripts/evaluation/inference.py +137 -0
  31. scripts/evaluation/inference_freenoise.py +147 -0
  32. scripts/evaluation/inference_freenoise_mp.py +153 -0
  33. scripts/run_text2video.sh +22 -0
  34. scripts/run_text2video_freenoise_1024.sh +24 -0
  35. scripts/run_text2video_freenoise_256.sh +25 -0
  36. scripts/run_text2video_freenoise_512.sh +24 -0
  37. scripts/run_text2video_freenoise_mp_256.sh +25 -0
  38. scripts/run_text2video_freenoise_mp_512.sh +24 -0
  39. utils/utils.py +77 -0
app.py ADDED
@@ -0,0 +1,247 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import gradio as gr
2
+
3
+ import os
4
+ import sys
5
+ import argparse
6
+ import random
7
+ from omegaconf import OmegaConf
8
+ import torch
9
+ import torchvision
10
+ from pytorch_lightning import seed_everything
11
+ from huggingface_hub import hf_hub_download
12
+
13
+ sys.path.insert(0, "scripts/evaluation")
14
+ from funcs import (
15
+ batch_ddim_sampling_freenoise,
16
+ load_model_checkpoint,
17
+ )
18
+ from utils.utils import instantiate_from_config
19
+
20
+ def infer(prompt):
21
+ output_size = "256x256"
22
+ num_frames = 32
23
+ ddim_steps = 50
24
+ unconditional_guidance_scale = 12.0
25
+ seed = 123
26
+ save_fps = 10
27
+ window_size = 16
28
+ window_stride = 4
29
+
30
+ if output_size == "576x1024":
31
+ width = 1024
32
+ height = 576
33
+ ckpt_path_1024 = "checkpoints/base_1024_v1/model.ckpt"
34
+ hf_hub_download(repo_id="VideoCrafter/Text2Video-1024", filename="model.ckpt", local_dir=ckpt_path_1024)
35
+ config_1024 = "configs/inference_t2v_1024_v1.0_freenoise.yaml"
36
+ config_1024 = OmegaConf.load(config_1024)
37
+ model_config_1024 = config_1024.pop("model", OmegaConf.create())
38
+ model_1024 = instantiate_from_config(model_config_1024)
39
+ # model_1024 = model_1024.cuda()
40
+ model_1024 = load_model_checkpoint(model_1024, ckpt_path_1024)
41
+ model_1024.eval()
42
+ model = model_1024
43
+ fps = 24
44
+ elif output_size == "256x256":
45
+ width = 256
46
+ height = 256
47
+ ckpt_path_256 = "checkpoints/base_256_v1/model.pth"
48
+ config_256 = "configs/inference_t2v_tconv256_v1.0_freenoise.yaml"
49
+ hf_hub_download(repo_id="MoonQiu/LongerCrafter", filename="model.pth", local_dir=model_config_256)
50
+ config_256 = OmegaConf.load(config_256)
51
+ model_config_256 = config_256.pop("model", OmegaConf.create())
52
+ model_256 = instantiate_from_config(model_config_256)
53
+ # model_256 = model_256.cuda()
54
+ model_256 = load_model_checkpoint(model_256, ckpt_path_256)
55
+ model_256.eval()
56
+ model = model_256
57
+ fps = 8
58
+
59
+ if seed is None:
60
+ seed = int.from_bytes(os.urandom(2), "big")
61
+ print(f"Using seed: {seed}")
62
+ seed_everything(seed)
63
+
64
+ args = argparse.Namespace(
65
+ mode="base",
66
+ savefps=save_fps,
67
+ n_samples=1,
68
+ ddim_steps=ddim_steps,
69
+ ddim_eta=0.0,
70
+ bs=1,
71
+ height=height,
72
+ width=width,
73
+ frames=num_frames,
74
+ fps=fps,
75
+ unconditional_guidance_scale=unconditional_guidance_scale,
76
+ unconditional_guidance_scale_temporal=None,
77
+ cond_input=None,
78
+ window_size=window_size,
79
+ window_stride=window_stride,
80
+ )
81
+
82
+ ## latent noise shape
83
+ h, w = args.height // 8, args.width // 8
84
+ frames = model.temporal_length if args.frames < 0 else args.frames
85
+ channels = model.channels
86
+
87
+ x_T_total = torch.randn(
88
+ [args.n_samples, 1, channels, frames, h, w], device=model.device
89
+ ).repeat(1, args.bs, 1, 1, 1, 1)
90
+ for frame_index in range(args.window_size, args.frames, args.window_stride):
91
+ list_index = list(
92
+ range(
93
+ frame_index - args.window_size,
94
+ frame_index + args.window_stride - args.window_size,
95
+ )
96
+ )
97
+ random.shuffle(list_index)
98
+ x_T_total[
99
+ :, :, :, frame_index : frame_index + args.window_stride
100
+ ] = x_T_total[:, :, :, list_index]
101
+
102
+ batch_size = 1
103
+ noise_shape = [batch_size, channels, frames, h, w]
104
+ fps = torch.tensor([args.fps] * batch_size).to(model.device).long()
105
+ prompts = [prompt]
106
+ text_emb = model.get_learned_conditioning(prompts)
107
+
108
+ cond = {"c_crossattn": [text_emb], "fps": fps}
109
+
110
+ ## inference
111
+ batch_samples = batch_ddim_sampling_freenoise(
112
+ model,
113
+ cond,
114
+ noise_shape,
115
+ args.n_samples,
116
+ args.ddim_steps,
117
+ args.ddim_eta,
118
+ args.unconditional_guidance_scale,
119
+ args=args,
120
+ x_T_total=x_T_total,
121
+ )
122
+
123
+ video_path = "/tmp/output.mp4"
124
+ vid_tensor = batch_samples[0]
125
+ video = vid_tensor.detach().cpu()
126
+ video = torch.clamp(video.float(), -1.0, 1.0)
127
+ video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w
128
+
129
+ frame_grids = [
130
+ torchvision.utils.make_grid(framesheet, nrow=int(args.n_samples))
131
+ for framesheet in video
132
+ ] # [3, 1*h, n*w]
133
+ grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [t, 3, n*h, w]
134
+ grid = (grid + 1.0) / 2.0
135
+ grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1)
136
+ torchvision.io.write_video(
137
+ video_path,
138
+ grid,
139
+ fps=args.savefps,
140
+ video_codec="h264",
141
+ options={"crf": "10"},
142
+ )
143
+
144
+ print(video_path)
145
+ return video_path, gr.Group.update(visible=True)
146
+
147
+ css = """
148
+ #col-container {max-width: 510px; margin-left: auto; margin-right: auto;}
149
+ a {text-decoration-line: underline; font-weight: 600;}
150
+ .animate-spin {
151
+ animation: spin 1s linear infinite;
152
+ }
153
+ @keyframes spin {
154
+ from {
155
+ transform: rotate(0deg);
156
+ }
157
+ to {
158
+ transform: rotate(360deg);
159
+ }
160
+ }
161
+ #share-btn-container {
162
+ display: flex;
163
+ padding-left: 0.5rem !important;
164
+ padding-right: 0.5rem !important;
165
+ background-color: #000000;
166
+ justify-content: center;
167
+ align-items: center;
168
+ border-radius: 9999px !important;
169
+ max-width: 15rem;
170
+ height: 36px;
171
+ }
172
+ div#share-btn-container > div {
173
+ flex-direction: row;
174
+ background: black;
175
+ align-items: center;
176
+ }
177
+ #share-btn-container:hover {
178
+ background-color: #060606;
179
+ }
180
+ #share-btn {
181
+ all: initial;
182
+ color: #ffffff;
183
+ font-weight: 600;
184
+ cursor:pointer;
185
+ font-family: 'IBM Plex Sans', sans-serif;
186
+ margin-left: 0.5rem !important;
187
+ padding-top: 0.5rem !important;
188
+ padding-bottom: 0.5rem !important;
189
+ right:0;
190
+ }
191
+ #share-btn * {
192
+ all: unset;
193
+ }
194
+ #share-btn-container div:nth-child(-n+2){
195
+ width: auto !important;
196
+ min-height: 0px !important;
197
+ }
198
+ #share-btn-container .wrap {
199
+ display: none !important;
200
+ }
201
+ #share-btn-container.hidden {
202
+ display: none!important;
203
+ }
204
+ img[src*='#center'] {
205
+ display: inline-block;
206
+ margin: unset;
207
+ }
208
+ .footer {
209
+ margin-bottom: 45px;
210
+ margin-top: 10px;
211
+ text-align: center;
212
+ border-bottom: 1px solid #e5e5e5;
213
+ }
214
+ .footer>p {
215
+ font-size: .8rem;
216
+ display: inline-block;
217
+ padding: 0 10px;
218
+ transform: translateY(10px);
219
+ background: white;
220
+ }
221
+ .dark .footer {
222
+ border-color: #303030;
223
+ }
224
+ .dark .footer>p {
225
+ background: #0b0f19;
226
+ }
227
+ """
228
+
229
+ with gr.Blocks(css=css) as demo:
230
+ with gr.Column(elem_id="col-container"):
231
+ gr.Markdown(
232
+ """
233
+ <h1 style="text-align: center;">LongerCrafter(FreeNoise) Text-to-Video</h1>
234
+ <p style="text-align: center;">
235
+ Tuning-Free Longer Video Diffusion via Noise Rescheduling <br />
236
+ </p>
237
+
238
+ """
239
+ )
240
+
241
+ prompt_in = gr.Textbox(label="Prompt", placeholder="Darth Vader is surfing on waves", elem_id="prompt-in")
242
+ #neg_prompt = gr.Textbox(label="Negative prompt", value="text, watermark, copyright, blurry, nsfw", elem_id="neg-prompt-in")
243
+ #inference_steps = gr.Slider(label="Inference Steps", minimum=10, maximum=100, step=1, value=40, interactive=False)
244
+ submit_btn = gr.Button("Submit")
245
+ video_result = gr.Video(label="Video Output", elem_id="video-output")
246
+
247
+ demo.queue(max_size=12).launch(show_api=True)
configs/inference_t2v_1024_v1.0.yaml ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model:
2
+ target: lvdm.models.ddpm3d.LatentDiffusion
3
+ params:
4
+ linear_start: 0.00085
5
+ linear_end: 0.012
6
+ num_timesteps_cond: 1
7
+ timesteps: 1000
8
+ first_stage_key: video
9
+ cond_stage_key: caption
10
+ cond_stage_trainable: false
11
+ conditioning_key: crossattn
12
+ image_size:
13
+ - 72
14
+ - 128
15
+ channels: 4
16
+ scale_by_std: false
17
+ scale_factor: 0.18215
18
+ use_ema: false
19
+ uncond_type: empty_seq
20
+ use_scale: true
21
+ fix_scale_bug: true
22
+ unet_config:
23
+ target: lvdm.modules.networks.openaimodel3d.UNetModel
24
+ params:
25
+ in_channels: 4
26
+ out_channels: 4
27
+ model_channels: 320
28
+ attention_resolutions:
29
+ - 4
30
+ - 2
31
+ - 1
32
+ num_res_blocks: 2
33
+ channel_mult:
34
+ - 1
35
+ - 2
36
+ - 4
37
+ - 4
38
+ num_head_channels: 64
39
+ transformer_depth: 1
40
+ context_dim: 1024
41
+ use_linear: true
42
+ use_checkpoint: true
43
+ temporal_conv: false
44
+ temporal_attention: true
45
+ temporal_selfatt_only: true
46
+ use_relative_position: true
47
+ use_causal_attention: false
48
+ temporal_length: 16
49
+ addition_attention: true
50
+ fps_cond: true
51
+ first_stage_config:
52
+ target: lvdm.models.autoencoder.AutoencoderKL
53
+ params:
54
+ embed_dim: 4
55
+ monitor: val/rec_loss
56
+ ddconfig:
57
+ double_z: true
58
+ z_channels: 4
59
+ resolution: 512
60
+ in_channels: 3
61
+ out_ch: 3
62
+ ch: 128
63
+ ch_mult:
64
+ - 1
65
+ - 2
66
+ - 4
67
+ - 4
68
+ num_res_blocks: 2
69
+ attn_resolutions: []
70
+ dropout: 0.0
71
+ lossconfig:
72
+ target: torch.nn.Identity
73
+ cond_stage_config:
74
+ target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder
75
+ params:
76
+ freeze: true
77
+ layer: penultimate
configs/inference_t2v_1024_v1.0_freenoise.yaml ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model:
2
+ target: lvdm.models.ddpm3d.LatentDiffusion
3
+ params:
4
+ linear_start: 0.00085
5
+ linear_end: 0.012
6
+ num_timesteps_cond: 1
7
+ timesteps: 1000
8
+ first_stage_key: video
9
+ cond_stage_key: caption
10
+ cond_stage_trainable: false
11
+ conditioning_key: crossattn
12
+ image_size:
13
+ - 72
14
+ - 128
15
+ channels: 4
16
+ scale_by_std: false
17
+ scale_factor: 0.18215
18
+ use_ema: false
19
+ uncond_type: empty_seq
20
+ use_scale: true
21
+ fix_scale_bug: true
22
+ unet_config:
23
+ target: lvdm.modules.networks.openaimodel3d_freenoise.UNetModel
24
+ params:
25
+ in_channels: 4
26
+ out_channels: 4
27
+ model_channels: 320
28
+ attention_resolutions:
29
+ - 4
30
+ - 2
31
+ - 1
32
+ num_res_blocks: 2
33
+ channel_mult:
34
+ - 1
35
+ - 2
36
+ - 4
37
+ - 4
38
+ num_head_channels: 64
39
+ transformer_depth: 1
40
+ context_dim: 1024
41
+ use_linear: true
42
+ use_checkpoint: true
43
+ temporal_conv: false
44
+ temporal_attention: true
45
+ temporal_selfatt_only: true
46
+ use_relative_position: true
47
+ use_causal_attention: false
48
+ temporal_length: 16
49
+ addition_attention: true
50
+ fps_cond: true
51
+ first_stage_config:
52
+ target: lvdm.models.autoencoder.AutoencoderKL
53
+ params:
54
+ embed_dim: 4
55
+ monitor: val/rec_loss
56
+ ddconfig:
57
+ double_z: true
58
+ z_channels: 4
59
+ resolution: 512
60
+ in_channels: 3
61
+ out_ch: 3
62
+ ch: 128
63
+ ch_mult:
64
+ - 1
65
+ - 2
66
+ - 4
67
+ - 4
68
+ num_res_blocks: 2
69
+ attn_resolutions: []
70
+ dropout: 0.0
71
+ lossconfig:
72
+ target: torch.nn.Identity
73
+ cond_stage_config:
74
+ target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder
75
+ params:
76
+ freeze: true
77
+ layer: penultimate
configs/inference_t2v_tconv256_v1.0.yaml ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model:
2
+ target: lvdm.models.ddpm3d.LatentDiffusion
3
+ params:
4
+ linear_start: 0.00085
5
+ linear_end: 0.012
6
+ num_timesteps_cond: 1
7
+ timesteps: 1000
8
+ first_stage_key: video
9
+ cond_stage_key: caption
10
+ cond_stage_trainable: false
11
+ conditioning_key: crossattn
12
+ image_size:
13
+ - 32
14
+ - 32
15
+ channels: 4
16
+ scale_by_std: false
17
+ scale_factor: 0.18215
18
+ use_ema: false
19
+ uncond_type: empty_seq
20
+ use_scale: false
21
+ fix_scale_bug: true
22
+ unet_config:
23
+ target: lvdm.modules.networks.openaimodel3d_freenoise.UNetModel
24
+ params:
25
+ in_channels: 4
26
+ out_channels: 4
27
+ model_channels: 320
28
+ attention_resolutions:
29
+ - 4
30
+ - 2
31
+ - 1
32
+ num_res_blocks: 2
33
+ channel_mult:
34
+ - 1
35
+ - 2
36
+ - 4
37
+ - 4
38
+ num_head_channels: 64
39
+ transformer_depth: 1
40
+ context_dim: 1024
41
+ use_linear: true
42
+ use_checkpoint: true
43
+ temporal_conv: true
44
+ temporal_attention: true
45
+ temporal_selfatt_only: true
46
+ use_relative_position: false
47
+ use_causal_attention: false
48
+ temporal_length: 16
49
+ addition_attention: true
50
+ fps_cond: true
51
+ first_stage_config:
52
+ target: lvdm.models.autoencoder.AutoencoderKL
53
+ params:
54
+ embed_dim: 4
55
+ monitor: val/rec_loss
56
+ ddconfig:
57
+ double_z: true
58
+ z_channels: 4
59
+ resolution: 512
60
+ in_channels: 3
61
+ out_ch: 3
62
+ ch: 128
63
+ ch_mult:
64
+ - 1
65
+ - 2
66
+ - 4
67
+ - 4
68
+ num_res_blocks: 2
69
+ attn_resolutions: []
70
+ dropout: 0.0
71
+ lossconfig:
72
+ target: torch.nn.Identity
73
+ cond_stage_config:
74
+ target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder
75
+ params:
76
+ freeze: true
77
+ layer: penultimate
configs/inference_t2v_tconv256_v1.0_freenoise.yaml ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model:
2
+ target: lvdm.models.ddpm3d.LatentDiffusion
3
+ params:
4
+ linear_start: 0.00085
5
+ linear_end: 0.012
6
+ num_timesteps_cond: 1
7
+ timesteps: 1000
8
+ first_stage_key: video
9
+ cond_stage_key: caption
10
+ cond_stage_trainable: false
11
+ conditioning_key: crossattn
12
+ image_size:
13
+ - 32
14
+ - 32
15
+ channels: 4
16
+ scale_by_std: false
17
+ scale_factor: 0.18215
18
+ use_ema: false
19
+ uncond_type: empty_seq
20
+ use_scale: false
21
+ fix_scale_bug: true
22
+ unet_config:
23
+ target: lvdm.modules.networks.openaimodel3d_freenoise.UNetModel
24
+ params:
25
+ in_channels: 4
26
+ out_channels: 4
27
+ model_channels: 320
28
+ attention_resolutions:
29
+ - 4
30
+ - 2
31
+ - 1
32
+ num_res_blocks: 2
33
+ channel_mult:
34
+ - 1
35
+ - 2
36
+ - 4
37
+ - 4
38
+ num_head_channels: 64
39
+ transformer_depth: 1
40
+ context_dim: 1024
41
+ use_linear: true
42
+ use_checkpoint: true
43
+ temporal_conv: true
44
+ temporal_attention: true
45
+ temporal_selfatt_only: true
46
+ use_relative_position: false
47
+ use_causal_attention: false
48
+ temporal_length: 16
49
+ addition_attention: true
50
+ fps_cond: true
51
+ first_stage_config:
52
+ target: lvdm.models.autoencoder.AutoencoderKL
53
+ params:
54
+ embed_dim: 4
55
+ monitor: val/rec_loss
56
+ ddconfig:
57
+ double_z: true
58
+ z_channels: 4
59
+ resolution: 512
60
+ in_channels: 3
61
+ out_ch: 3
62
+ ch: 128
63
+ ch_mult:
64
+ - 1
65
+ - 2
66
+ - 4
67
+ - 4
68
+ num_res_blocks: 2
69
+ attn_resolutions: []
70
+ dropout: 0.0
71
+ lossconfig:
72
+ target: torch.nn.Identity
73
+ cond_stage_config:
74
+ target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder
75
+ params:
76
+ freeze: true
77
+ layer: penultimate
configs/inference_t2v_tconv512_v1.0.yaml ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model:
2
+ target: lvdm.models.ddpm3d.LatentDiffusion
3
+ params:
4
+ linear_start: 0.00085
5
+ linear_end: 0.012
6
+ num_timesteps_cond: 1
7
+ timesteps: 1000
8
+ first_stage_key: video
9
+ cond_stage_key: caption
10
+ cond_stage_trainable: false
11
+ conditioning_key: crossattn
12
+ image_size:
13
+ - 40
14
+ - 64
15
+ channels: 4
16
+ scale_by_std: false
17
+ scale_factor: 0.18215
18
+ use_ema: false
19
+ uncond_type: empty_seq
20
+ use_scale: true
21
+ scale_b: 0.7
22
+ unet_config:
23
+ target: lvdm.modules.networks.openaimodel3d.UNetModel
24
+ params:
25
+ in_channels: 4
26
+ out_channels: 4
27
+ model_channels: 320
28
+ attention_resolutions:
29
+ - 4
30
+ - 2
31
+ - 1
32
+ num_res_blocks: 2
33
+ channel_mult:
34
+ - 1
35
+ - 2
36
+ - 4
37
+ - 4
38
+ num_head_channels: 64
39
+ transformer_depth: 1
40
+ context_dim: 1024
41
+ use_linear: true
42
+ use_checkpoint: true
43
+ temporal_conv: true
44
+ temporal_attention: true
45
+ temporal_selfatt_only: true
46
+ use_relative_position: false
47
+ use_causal_attention: false
48
+ temporal_length: 16
49
+ addition_attention: true
50
+ fps_cond: true
51
+ first_stage_config:
52
+ target: lvdm.models.autoencoder.AutoencoderKL
53
+ params:
54
+ embed_dim: 4
55
+ monitor: val/rec_loss
56
+ ddconfig:
57
+ double_z: true
58
+ z_channels: 4
59
+ resolution: 512
60
+ in_channels: 3
61
+ out_ch: 3
62
+ ch: 128
63
+ ch_mult:
64
+ - 1
65
+ - 2
66
+ - 4
67
+ - 4
68
+ num_res_blocks: 2
69
+ attn_resolutions: []
70
+ dropout: 0.0
71
+ lossconfig:
72
+ target: torch.nn.Identity
73
+ cond_stage_config:
74
+ target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder
75
+ params:
76
+ freeze: true
77
+ layer: penultimate
configs/inference_t2v_tconv512_v1.0_freenoise.yaml ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model:
2
+ target: lvdm.models.ddpm3d.LatentDiffusion
3
+ params:
4
+ linear_start: 0.00085
5
+ linear_end: 0.012
6
+ num_timesteps_cond: 1
7
+ timesteps: 1000
8
+ first_stage_key: video
9
+ cond_stage_key: caption
10
+ cond_stage_trainable: false
11
+ conditioning_key: crossattn
12
+ image_size:
13
+ - 40
14
+ - 64
15
+ channels: 4
16
+ scale_by_std: false
17
+ scale_factor: 0.18215
18
+ use_ema: false
19
+ uncond_type: empty_seq
20
+ use_scale: true
21
+ scale_b: 0.7
22
+ unet_config:
23
+ target: lvdm.modules.networks.openaimodel3d_freenoise.UNetModel
24
+ params:
25
+ in_channels: 4
26
+ out_channels: 4
27
+ model_channels: 320
28
+ attention_resolutions:
29
+ - 4
30
+ - 2
31
+ - 1
32
+ num_res_blocks: 2
33
+ channel_mult:
34
+ - 1
35
+ - 2
36
+ - 4
37
+ - 4
38
+ num_head_channels: 64
39
+ transformer_depth: 1
40
+ context_dim: 1024
41
+ use_linear: true
42
+ use_checkpoint: true
43
+ temporal_conv: true
44
+ temporal_attention: true
45
+ temporal_selfatt_only: true
46
+ use_relative_position: false
47
+ use_causal_attention: false
48
+ temporal_length: 16
49
+ addition_attention: true
50
+ fps_cond: true
51
+ first_stage_config:
52
+ target: lvdm.models.autoencoder.AutoencoderKL
53
+ params:
54
+ embed_dim: 4
55
+ monitor: val/rec_loss
56
+ ddconfig:
57
+ double_z: true
58
+ z_channels: 4
59
+ resolution: 512
60
+ in_channels: 3
61
+ out_ch: 3
62
+ ch: 128
63
+ ch_mult:
64
+ - 1
65
+ - 2
66
+ - 4
67
+ - 4
68
+ num_res_blocks: 2
69
+ attn_resolutions: []
70
+ dropout: 0.0
71
+ lossconfig:
72
+ target: torch.nn.Identity
73
+ cond_stage_config:
74
+ target: lvdm.modules.encoders.condition.FrozenOpenCLIPEmbedder
75
+ params:
76
+ freeze: true
77
+ layer: penultimate
lvdm/basics.py ADDED
@@ -0,0 +1,100 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # adopted from
2
+ # https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
3
+ # and
4
+ # https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
5
+ # and
6
+ # https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py
7
+ #
8
+ # thanks!
9
+
10
+ import torch.nn as nn
11
+ from utils.utils import instantiate_from_config
12
+
13
+
14
+ def disabled_train(self, mode=True):
15
+ """Overwrite model.train with this function to make sure train/eval mode
16
+ does not change anymore."""
17
+ return self
18
+
19
+ def zero_module(module):
20
+ """
21
+ Zero out the parameters of a module and return it.
22
+ """
23
+ for p in module.parameters():
24
+ p.detach().zero_()
25
+ return module
26
+
27
+ def scale_module(module, scale):
28
+ """
29
+ Scale the parameters of a module and return it.
30
+ """
31
+ for p in module.parameters():
32
+ p.detach().mul_(scale)
33
+ return module
34
+
35
+
36
+ def conv_nd(dims, *args, **kwargs):
37
+ """
38
+ Create a 1D, 2D, or 3D convolution module.
39
+ """
40
+ if dims == 1:
41
+ return nn.Conv1d(*args, **kwargs)
42
+ elif dims == 2:
43
+ return nn.Conv2d(*args, **kwargs)
44
+ elif dims == 3:
45
+ return nn.Conv3d(*args, **kwargs)
46
+ raise ValueError(f"unsupported dimensions: {dims}")
47
+
48
+
49
+ def linear(*args, **kwargs):
50
+ """
51
+ Create a linear module.
52
+ """
53
+ return nn.Linear(*args, **kwargs)
54
+
55
+
56
+ def avg_pool_nd(dims, *args, **kwargs):
57
+ """
58
+ Create a 1D, 2D, or 3D average pooling module.
59
+ """
60
+ if dims == 1:
61
+ return nn.AvgPool1d(*args, **kwargs)
62
+ elif dims == 2:
63
+ return nn.AvgPool2d(*args, **kwargs)
64
+ elif dims == 3:
65
+ return nn.AvgPool3d(*args, **kwargs)
66
+ raise ValueError(f"unsupported dimensions: {dims}")
67
+
68
+
69
+ def nonlinearity(type='silu'):
70
+ if type == 'silu':
71
+ return nn.SiLU()
72
+ elif type == 'leaky_relu':
73
+ return nn.LeakyReLU()
74
+
75
+
76
+ class GroupNormSpecific(nn.GroupNorm):
77
+ def forward(self, x):
78
+ return super().forward(x.float()).type(x.dtype)
79
+
80
+
81
+ def normalization(channels, num_groups=32):
82
+ """
83
+ Make a standard normalization layer.
84
+ :param channels: number of input channels.
85
+ :return: an nn.Module for normalization.
86
+ """
87
+ return GroupNormSpecific(num_groups, channels)
88
+
89
+
90
+ class HybridConditioner(nn.Module):
91
+
92
+ def __init__(self, c_concat_config, c_crossattn_config):
93
+ super().__init__()
94
+ self.concat_conditioner = instantiate_from_config(c_concat_config)
95
+ self.crossattn_conditioner = instantiate_from_config(c_crossattn_config)
96
+
97
+ def forward(self, c_concat, c_crossattn):
98
+ c_concat = self.concat_conditioner(c_concat)
99
+ c_crossattn = self.crossattn_conditioner(c_crossattn)
100
+ return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]}
lvdm/common.py ADDED
@@ -0,0 +1,95 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ from inspect import isfunction
3
+ import torch
4
+ from torch import nn
5
+ import torch.distributed as dist
6
+
7
+
8
+ def gather_data(data, return_np=True):
9
+ ''' gather data from multiple processes to one list '''
10
+ data_list = [torch.zeros_like(data) for _ in range(dist.get_world_size())]
11
+ dist.all_gather(data_list, data) # gather not supported with NCCL
12
+ if return_np:
13
+ data_list = [data.cpu().numpy() for data in data_list]
14
+ return data_list
15
+
16
+ def autocast(f):
17
+ def do_autocast(*args, **kwargs):
18
+ with torch.cuda.amp.autocast(enabled=True,
19
+ dtype=torch.get_autocast_gpu_dtype(),
20
+ cache_enabled=torch.is_autocast_cache_enabled()):
21
+ return f(*args, **kwargs)
22
+ return do_autocast
23
+
24
+
25
+ def extract_into_tensor(a, t, x_shape):
26
+ b, *_ = t.shape
27
+ out = a.gather(-1, t)
28
+ return out.reshape(b, *((1,) * (len(x_shape) - 1)))
29
+
30
+
31
+ def noise_like(shape, device, repeat=False):
32
+ repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
33
+ noise = lambda: torch.randn(shape, device=device)
34
+ return repeat_noise() if repeat else noise()
35
+
36
+
37
+ def default(val, d):
38
+ if exists(val):
39
+ return val
40
+ return d() if isfunction(d) else d
41
+
42
+ def exists(val):
43
+ return val is not None
44
+
45
+ def identity(*args, **kwargs):
46
+ return nn.Identity()
47
+
48
+ def uniq(arr):
49
+ return{el: True for el in arr}.keys()
50
+
51
+ def mean_flat(tensor):
52
+ """
53
+ Take the mean over all non-batch dimensions.
54
+ """
55
+ return tensor.mean(dim=list(range(1, len(tensor.shape))))
56
+
57
+ def ismap(x):
58
+ if not isinstance(x, torch.Tensor):
59
+ return False
60
+ return (len(x.shape) == 4) and (x.shape[1] > 3)
61
+
62
+ def isimage(x):
63
+ if not isinstance(x,torch.Tensor):
64
+ return False
65
+ return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1)
66
+
67
+ def max_neg_value(t):
68
+ return -torch.finfo(t.dtype).max
69
+
70
+ def shape_to_str(x):
71
+ shape_str = "x".join([str(x) for x in x.shape])
72
+ return shape_str
73
+
74
+ def init_(tensor):
75
+ dim = tensor.shape[-1]
76
+ std = 1 / math.sqrt(dim)
77
+ tensor.uniform_(-std, std)
78
+ return tensor
79
+
80
+ ckpt = torch.utils.checkpoint.checkpoint
81
+ def checkpoint(func, inputs, params, flag):
82
+ """
83
+ Evaluate a function without caching intermediate activations, allowing for
84
+ reduced memory at the expense of extra compute in the backward pass.
85
+ :param func: the function to evaluate.
86
+ :param inputs: the argument sequence to pass to `func`.
87
+ :param params: a sequence of parameters `func` depends on but does not
88
+ explicitly take as arguments.
89
+ :param flag: if False, disable gradient checkpointing.
90
+ """
91
+ if flag:
92
+ return ckpt(func, *inputs)
93
+ else:
94
+ return func(*inputs)
95
+
lvdm/distributions.py ADDED
@@ -0,0 +1,95 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import numpy as np
3
+
4
+
5
+ class AbstractDistribution:
6
+ def sample(self):
7
+ raise NotImplementedError()
8
+
9
+ def mode(self):
10
+ raise NotImplementedError()
11
+
12
+
13
+ class DiracDistribution(AbstractDistribution):
14
+ def __init__(self, value):
15
+ self.value = value
16
+
17
+ def sample(self):
18
+ return self.value
19
+
20
+ def mode(self):
21
+ return self.value
22
+
23
+
24
+ class DiagonalGaussianDistribution(object):
25
+ def __init__(self, parameters, deterministic=False):
26
+ self.parameters = parameters
27
+ self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
28
+ self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
29
+ self.deterministic = deterministic
30
+ self.std = torch.exp(0.5 * self.logvar)
31
+ self.var = torch.exp(self.logvar)
32
+ if self.deterministic:
33
+ self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device)
34
+
35
+ def sample(self, noise=None):
36
+ if noise is None:
37
+ noise = torch.randn(self.mean.shape)
38
+
39
+ x = self.mean + self.std * noise.to(device=self.parameters.device)
40
+ return x
41
+
42
+ def kl(self, other=None):
43
+ if self.deterministic:
44
+ return torch.Tensor([0.])
45
+ else:
46
+ if other is None:
47
+ return 0.5 * torch.sum(torch.pow(self.mean, 2)
48
+ + self.var - 1.0 - self.logvar,
49
+ dim=[1, 2, 3])
50
+ else:
51
+ return 0.5 * torch.sum(
52
+ torch.pow(self.mean - other.mean, 2) / other.var
53
+ + self.var / other.var - 1.0 - self.logvar + other.logvar,
54
+ dim=[1, 2, 3])
55
+
56
+ def nll(self, sample, dims=[1,2,3]):
57
+ if self.deterministic:
58
+ return torch.Tensor([0.])
59
+ logtwopi = np.log(2.0 * np.pi)
60
+ return 0.5 * torch.sum(
61
+ logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
62
+ dim=dims)
63
+
64
+ def mode(self):
65
+ return self.mean
66
+
67
+
68
+ def normal_kl(mean1, logvar1, mean2, logvar2):
69
+ """
70
+ source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
71
+ Compute the KL divergence between two gaussians.
72
+ Shapes are automatically broadcasted, so batches can be compared to
73
+ scalars, among other use cases.
74
+ """
75
+ tensor = None
76
+ for obj in (mean1, logvar1, mean2, logvar2):
77
+ if isinstance(obj, torch.Tensor):
78
+ tensor = obj
79
+ break
80
+ assert tensor is not None, "at least one argument must be a Tensor"
81
+
82
+ # Force variances to be Tensors. Broadcasting helps convert scalars to
83
+ # Tensors, but it does not work for torch.exp().
84
+ logvar1, logvar2 = [
85
+ x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
86
+ for x in (logvar1, logvar2)
87
+ ]
88
+
89
+ return 0.5 * (
90
+ -1.0
91
+ + logvar2
92
+ - logvar1
93
+ + torch.exp(logvar1 - logvar2)
94
+ + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
95
+ )
lvdm/ema.py ADDED
@@ -0,0 +1,76 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch import nn
3
+
4
+
5
+ class LitEma(nn.Module):
6
+ def __init__(self, model, decay=0.9999, use_num_upates=True):
7
+ super().__init__()
8
+ if decay < 0.0 or decay > 1.0:
9
+ raise ValueError('Decay must be between 0 and 1')
10
+
11
+ self.m_name2s_name = {}
12
+ self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32))
13
+ self.register_buffer('num_updates', torch.tensor(0,dtype=torch.int) if use_num_upates
14
+ else torch.tensor(-1,dtype=torch.int))
15
+
16
+ for name, p in model.named_parameters():
17
+ if p.requires_grad:
18
+ #remove as '.'-character is not allowed in buffers
19
+ s_name = name.replace('.','')
20
+ self.m_name2s_name.update({name:s_name})
21
+ self.register_buffer(s_name,p.clone().detach().data)
22
+
23
+ self.collected_params = []
24
+
25
+ def forward(self,model):
26
+ decay = self.decay
27
+
28
+ if self.num_updates >= 0:
29
+ self.num_updates += 1
30
+ decay = min(self.decay,(1 + self.num_updates) / (10 + self.num_updates))
31
+
32
+ one_minus_decay = 1.0 - decay
33
+
34
+ with torch.no_grad():
35
+ m_param = dict(model.named_parameters())
36
+ shadow_params = dict(self.named_buffers())
37
+
38
+ for key in m_param:
39
+ if m_param[key].requires_grad:
40
+ sname = self.m_name2s_name[key]
41
+ shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
42
+ shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key]))
43
+ else:
44
+ assert not key in self.m_name2s_name
45
+
46
+ def copy_to(self, model):
47
+ m_param = dict(model.named_parameters())
48
+ shadow_params = dict(self.named_buffers())
49
+ for key in m_param:
50
+ if m_param[key].requires_grad:
51
+ m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
52
+ else:
53
+ assert not key in self.m_name2s_name
54
+
55
+ def store(self, parameters):
56
+ """
57
+ Save the current parameters for restoring later.
58
+ Args:
59
+ parameters: Iterable of `torch.nn.Parameter`; the parameters to be
60
+ temporarily stored.
61
+ """
62
+ self.collected_params = [param.clone() for param in parameters]
63
+
64
+ def restore(self, parameters):
65
+ """
66
+ Restore the parameters stored with the `store` method.
67
+ Useful to validate the model with EMA parameters without affecting the
68
+ original optimization process. Store the parameters before the
69
+ `copy_to` method. After validation (or model saving), use this to
70
+ restore the former parameters.
71
+ Args:
72
+ parameters: Iterable of `torch.nn.Parameter`; the parameters to be
73
+ updated with the stored parameters.
74
+ """
75
+ for c_param, param in zip(self.collected_params, parameters):
76
+ param.data.copy_(c_param.data)
lvdm/models/autoencoder.py ADDED
@@ -0,0 +1,219 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ from contextlib import contextmanager
3
+ import torch
4
+ import numpy as np
5
+ from einops import rearrange
6
+ import torch.nn.functional as F
7
+ import pytorch_lightning as pl
8
+ from lvdm.modules.networks.ae_modules import Encoder, Decoder
9
+ from lvdm.distributions import DiagonalGaussianDistribution
10
+ from utils.utils import instantiate_from_config
11
+
12
+
13
+ class AutoencoderKL(pl.LightningModule):
14
+ def __init__(self,
15
+ ddconfig,
16
+ lossconfig,
17
+ embed_dim,
18
+ ckpt_path=None,
19
+ ignore_keys=[],
20
+ image_key="image",
21
+ colorize_nlabels=None,
22
+ monitor=None,
23
+ test=False,
24
+ logdir=None,
25
+ input_dim=4,
26
+ test_args=None,
27
+ ):
28
+ super().__init__()
29
+ self.image_key = image_key
30
+ self.encoder = Encoder(**ddconfig)
31
+ self.decoder = Decoder(**ddconfig)
32
+ self.loss = instantiate_from_config(lossconfig)
33
+ assert ddconfig["double_z"]
34
+ self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1)
35
+ self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1)
36
+ self.embed_dim = embed_dim
37
+ self.input_dim = input_dim
38
+ self.test = test
39
+ self.test_args = test_args
40
+ self.logdir = logdir
41
+ if colorize_nlabels is not None:
42
+ assert type(colorize_nlabels)==int
43
+ self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1))
44
+ if monitor is not None:
45
+ self.monitor = monitor
46
+ if ckpt_path is not None:
47
+ self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
48
+ if self.test:
49
+ self.init_test()
50
+
51
+ def init_test(self,):
52
+ self.test = True
53
+ save_dir = os.path.join(self.logdir, "test")
54
+ if 'ckpt' in self.test_args:
55
+ ckpt_name = os.path.basename(self.test_args.ckpt).split('.ckpt')[0] + f'_epoch{self._cur_epoch}'
56
+ self.root = os.path.join(save_dir, ckpt_name)
57
+ else:
58
+ self.root = save_dir
59
+ if 'test_subdir' in self.test_args:
60
+ self.root = os.path.join(save_dir, self.test_args.test_subdir)
61
+
62
+ self.root_zs = os.path.join(self.root, "zs")
63
+ self.root_dec = os.path.join(self.root, "reconstructions")
64
+ self.root_inputs = os.path.join(self.root, "inputs")
65
+ os.makedirs(self.root, exist_ok=True)
66
+
67
+ if self.test_args.save_z:
68
+ os.makedirs(self.root_zs, exist_ok=True)
69
+ if self.test_args.save_reconstruction:
70
+ os.makedirs(self.root_dec, exist_ok=True)
71
+ if self.test_args.save_input:
72
+ os.makedirs(self.root_inputs, exist_ok=True)
73
+ assert(self.test_args is not None)
74
+ self.test_maximum = getattr(self.test_args, 'test_maximum', None)
75
+ self.count = 0
76
+ self.eval_metrics = {}
77
+ self.decodes = []
78
+ self.save_decode_samples = 2048
79
+
80
+ def init_from_ckpt(self, path, ignore_keys=list()):
81
+ sd = torch.load(path, map_location="cpu")
82
+ try:
83
+ self._cur_epoch = sd['epoch']
84
+ sd = sd["state_dict"]
85
+ except:
86
+ self._cur_epoch = 'null'
87
+ keys = list(sd.keys())
88
+ for k in keys:
89
+ for ik in ignore_keys:
90
+ if k.startswith(ik):
91
+ print("Deleting key {} from state_dict.".format(k))
92
+ del sd[k]
93
+ self.load_state_dict(sd, strict=False)
94
+ # self.load_state_dict(sd, strict=True)
95
+ print(f"Restored from {path}")
96
+
97
+ def encode(self, x, **kwargs):
98
+
99
+ h = self.encoder(x)
100
+ moments = self.quant_conv(h)
101
+ posterior = DiagonalGaussianDistribution(moments)
102
+ return posterior
103
+
104
+ def decode(self, z, **kwargs):
105
+ z = self.post_quant_conv(z)
106
+ dec = self.decoder(z)
107
+ return dec
108
+
109
+ def forward(self, input, sample_posterior=True):
110
+ posterior = self.encode(input)
111
+ if sample_posterior:
112
+ z = posterior.sample()
113
+ else:
114
+ z = posterior.mode()
115
+ dec = self.decode(z)
116
+ return dec, posterior
117
+
118
+ def get_input(self, batch, k):
119
+ x = batch[k]
120
+ if x.dim() == 5 and self.input_dim == 4:
121
+ b,c,t,h,w = x.shape
122
+ self.b = b
123
+ self.t = t
124
+ x = rearrange(x, 'b c t h w -> (b t) c h w')
125
+
126
+ return x
127
+
128
+ def training_step(self, batch, batch_idx, optimizer_idx):
129
+ inputs = self.get_input(batch, self.image_key)
130
+ reconstructions, posterior = self(inputs)
131
+
132
+ if optimizer_idx == 0:
133
+ # train encoder+decoder+logvar
134
+ aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
135
+ last_layer=self.get_last_layer(), split="train")
136
+ self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
137
+ self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False)
138
+ return aeloss
139
+
140
+ if optimizer_idx == 1:
141
+ # train the discriminator
142
+ discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step,
143
+ last_layer=self.get_last_layer(), split="train")
144
+
145
+ self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True)
146
+ self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False)
147
+ return discloss
148
+
149
+ def validation_step(self, batch, batch_idx):
150
+ inputs = self.get_input(batch, self.image_key)
151
+ reconstructions, posterior = self(inputs)
152
+ aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step,
153
+ last_layer=self.get_last_layer(), split="val")
154
+
155
+ discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step,
156
+ last_layer=self.get_last_layer(), split="val")
157
+
158
+ self.log("val/rec_loss", log_dict_ae["val/rec_loss"])
159
+ self.log_dict(log_dict_ae)
160
+ self.log_dict(log_dict_disc)
161
+ return self.log_dict
162
+
163
+ def configure_optimizers(self):
164
+ lr = self.learning_rate
165
+ opt_ae = torch.optim.Adam(list(self.encoder.parameters())+
166
+ list(self.decoder.parameters())+
167
+ list(self.quant_conv.parameters())+
168
+ list(self.post_quant_conv.parameters()),
169
+ lr=lr, betas=(0.5, 0.9))
170
+ opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(),
171
+ lr=lr, betas=(0.5, 0.9))
172
+ return [opt_ae, opt_disc], []
173
+
174
+ def get_last_layer(self):
175
+ return self.decoder.conv_out.weight
176
+
177
+ @torch.no_grad()
178
+ def log_images(self, batch, only_inputs=False, **kwargs):
179
+ log = dict()
180
+ x = self.get_input(batch, self.image_key)
181
+ x = x.to(self.device)
182
+ if not only_inputs:
183
+ xrec, posterior = self(x)
184
+ if x.shape[1] > 3:
185
+ # colorize with random projection
186
+ assert xrec.shape[1] > 3
187
+ x = self.to_rgb(x)
188
+ xrec = self.to_rgb(xrec)
189
+ log["samples"] = self.decode(torch.randn_like(posterior.sample()))
190
+ log["reconstructions"] = xrec
191
+ log["inputs"] = x
192
+ return log
193
+
194
+ def to_rgb(self, x):
195
+ assert self.image_key == "segmentation"
196
+ if not hasattr(self, "colorize"):
197
+ self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x))
198
+ x = F.conv2d(x, weight=self.colorize)
199
+ x = 2.*(x-x.min())/(x.max()-x.min()) - 1.
200
+ return x
201
+
202
+ class IdentityFirstStage(torch.nn.Module):
203
+ def __init__(self, *args, vq_interface=False, **kwargs):
204
+ self.vq_interface = vq_interface # TODO: Should be true by default but check to not break older stuff
205
+ super().__init__()
206
+
207
+ def encode(self, x, *args, **kwargs):
208
+ return x
209
+
210
+ def decode(self, x, *args, **kwargs):
211
+ return x
212
+
213
+ def quantize(self, x, *args, **kwargs):
214
+ if self.vq_interface:
215
+ return x, None, [None, None, None]
216
+ return x
217
+
218
+ def forward(self, x, *args, **kwargs):
219
+ return x
lvdm/models/ddpm3d.py ADDED
@@ -0,0 +1,763 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ wild mixture of
3
+ https://github.com/openai/improved-diffusion/blob/e94489283bb876ac1477d5dd7709bbbd2d9902ce/improved_diffusion/gaussian_diffusion.py
4
+ https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py
5
+ https://github.com/CompVis/taming-transformers
6
+ -- merci
7
+ """
8
+
9
+ from functools import partial
10
+ from contextlib import contextmanager
11
+ import numpy as np
12
+ from tqdm import tqdm
13
+ from einops import rearrange, repeat
14
+ import logging
15
+ mainlogger = logging.getLogger('mainlogger')
16
+ import torch
17
+ import torch.nn as nn
18
+ from torchvision.utils import make_grid
19
+ import pytorch_lightning as pl
20
+ from utils.utils import instantiate_from_config
21
+ from lvdm.ema import LitEma
22
+ from lvdm.distributions import DiagonalGaussianDistribution
23
+ from lvdm.models.utils_diffusion import make_beta_schedule
24
+ from lvdm.modules.encoders.ip_resampler import ImageProjModel, Resampler
25
+ from lvdm.basics import disabled_train
26
+ from lvdm.common import (
27
+ extract_into_tensor,
28
+ noise_like,
29
+ exists,
30
+ default
31
+ )
32
+
33
+
34
+ __conditioning_keys__ = {'concat': 'c_concat',
35
+ 'crossattn': 'c_crossattn',
36
+ 'adm': 'y'}
37
+
38
+ class DDPM(pl.LightningModule):
39
+ # classic DDPM with Gaussian diffusion, in image space
40
+ def __init__(self,
41
+ unet_config,
42
+ timesteps=1000,
43
+ beta_schedule="linear",
44
+ loss_type="l2",
45
+ ckpt_path=None,
46
+ ignore_keys=[],
47
+ load_only_unet=False,
48
+ monitor=None,
49
+ use_ema=True,
50
+ first_stage_key="image",
51
+ image_size=256,
52
+ channels=3,
53
+ log_every_t=100,
54
+ clip_denoised=True,
55
+ linear_start=1e-4,
56
+ linear_end=2e-2,
57
+ cosine_s=8e-3,
58
+ given_betas=None,
59
+ original_elbo_weight=0.,
60
+ v_posterior=0., # weight for choosing posterior variance as sigma = (1-v) * beta_tilde + v * beta
61
+ l_simple_weight=1.,
62
+ conditioning_key=None,
63
+ parameterization="eps", # all assuming fixed variance schedules
64
+ scheduler_config=None,
65
+ use_positional_encodings=False,
66
+ learn_logvar=False,
67
+ logvar_init=0.
68
+ ):
69
+ super().__init__()
70
+ assert parameterization in ["eps", "x0"], 'currently only supporting "eps" and "x0"'
71
+ self.parameterization = parameterization
72
+ mainlogger.info(f"{self.__class__.__name__}: Running in {self.parameterization}-prediction mode")
73
+ self.cond_stage_model = None
74
+ self.clip_denoised = clip_denoised
75
+ self.log_every_t = log_every_t
76
+ self.first_stage_key = first_stage_key
77
+ self.channels = channels
78
+ self.temporal_length = unet_config.params.temporal_length
79
+ self.image_size = image_size
80
+ if isinstance(self.image_size, int):
81
+ self.image_size = [self.image_size, self.image_size]
82
+ self.use_positional_encodings = use_positional_encodings
83
+ self.model = DiffusionWrapper(unet_config, conditioning_key)
84
+ self.use_ema = use_ema
85
+ if self.use_ema:
86
+ self.model_ema = LitEma(self.model)
87
+ mainlogger.info(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
88
+
89
+ self.use_scheduler = scheduler_config is not None
90
+ if self.use_scheduler:
91
+ self.scheduler_config = scheduler_config
92
+
93
+ self.v_posterior = v_posterior
94
+ self.original_elbo_weight = original_elbo_weight
95
+ self.l_simple_weight = l_simple_weight
96
+
97
+ if monitor is not None:
98
+ self.monitor = monitor
99
+ if ckpt_path is not None:
100
+ self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys, only_model=load_only_unet)
101
+
102
+ self.register_schedule(given_betas=given_betas, beta_schedule=beta_schedule, timesteps=timesteps,
103
+ linear_start=linear_start, linear_end=linear_end, cosine_s=cosine_s)
104
+
105
+ self.loss_type = loss_type
106
+
107
+ self.learn_logvar = learn_logvar
108
+ self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,))
109
+ if self.learn_logvar:
110
+ self.logvar = nn.Parameter(self.logvar, requires_grad=True)
111
+
112
+
113
+ def register_schedule(self, given_betas=None, beta_schedule="linear", timesteps=1000,
114
+ linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
115
+ if exists(given_betas):
116
+ betas = given_betas
117
+ else:
118
+ betas = make_beta_schedule(beta_schedule, timesteps, linear_start=linear_start, linear_end=linear_end,
119
+ cosine_s=cosine_s)
120
+ alphas = 1. - betas
121
+ alphas_cumprod = np.cumprod(alphas, axis=0)
122
+ alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
123
+
124
+ timesteps, = betas.shape
125
+ self.num_timesteps = int(timesteps)
126
+ self.linear_start = linear_start
127
+ self.linear_end = linear_end
128
+ assert alphas_cumprod.shape[0] == self.num_timesteps, 'alphas have to be defined for each timestep'
129
+
130
+ to_torch = partial(torch.tensor, dtype=torch.float32)
131
+
132
+ self.register_buffer('betas', to_torch(betas))
133
+ self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
134
+ self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
135
+
136
+ # calculations for diffusion q(x_t | x_{t-1}) and others
137
+ self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
138
+ self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
139
+ self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
140
+ self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
141
+ self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
142
+
143
+ # calculations for posterior q(x_{t-1} | x_t, x_0)
144
+ posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / (
145
+ 1. - alphas_cumprod) + self.v_posterior * betas
146
+ # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
147
+ self.register_buffer('posterior_variance', to_torch(posterior_variance))
148
+ # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
149
+ self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
150
+ self.register_buffer('posterior_mean_coef1', to_torch(
151
+ betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
152
+ self.register_buffer('posterior_mean_coef2', to_torch(
153
+ (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
154
+
155
+ if self.parameterization == "eps":
156
+ lvlb_weights = self.betas ** 2 / (
157
+ 2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod))
158
+ elif self.parameterization == "x0":
159
+ lvlb_weights = 0.5 * np.sqrt(torch.Tensor(alphas_cumprod)) / (2. * 1 - torch.Tensor(alphas_cumprod))
160
+ else:
161
+ raise NotImplementedError("mu not supported")
162
+ # TODO how to choose this term
163
+ lvlb_weights[0] = lvlb_weights[1]
164
+ self.register_buffer('lvlb_weights', lvlb_weights, persistent=False)
165
+ assert not torch.isnan(self.lvlb_weights).all()
166
+
167
+ @contextmanager
168
+ def ema_scope(self, context=None):
169
+ if self.use_ema:
170
+ self.model_ema.store(self.model.parameters())
171
+ self.model_ema.copy_to(self.model)
172
+ if context is not None:
173
+ mainlogger.info(f"{context}: Switched to EMA weights")
174
+ try:
175
+ yield None
176
+ finally:
177
+ if self.use_ema:
178
+ self.model_ema.restore(self.model.parameters())
179
+ if context is not None:
180
+ mainlogger.info(f"{context}: Restored training weights")
181
+
182
+ def init_from_ckpt(self, path, ignore_keys=list(), only_model=False):
183
+ sd = torch.load(path, map_location="cpu")
184
+ if "state_dict" in list(sd.keys()):
185
+ sd = sd["state_dict"]
186
+ keys = list(sd.keys())
187
+ for k in keys:
188
+ for ik in ignore_keys:
189
+ if k.startswith(ik):
190
+ mainlogger.info("Deleting key {} from state_dict.".format(k))
191
+ del sd[k]
192
+ missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict(
193
+ sd, strict=False)
194
+ mainlogger.info(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
195
+ if len(missing) > 0:
196
+ mainlogger.info(f"Missing Keys: {missing}")
197
+ if len(unexpected) > 0:
198
+ mainlogger.info(f"Unexpected Keys: {unexpected}")
199
+
200
+ def q_mean_variance(self, x_start, t):
201
+ """
202
+ Get the distribution q(x_t | x_0).
203
+ :param x_start: the [N x C x ...] tensor of noiseless inputs.
204
+ :param t: the number of diffusion steps (minus 1). Here, 0 means one step.
205
+ :return: A tuple (mean, variance, log_variance), all of x_start's shape.
206
+ """
207
+ mean = (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start)
208
+ variance = extract_into_tensor(1.0 - self.alphas_cumprod, t, x_start.shape)
209
+ log_variance = extract_into_tensor(self.log_one_minus_alphas_cumprod, t, x_start.shape)
210
+ return mean, variance, log_variance
211
+
212
+ def predict_start_from_noise(self, x_t, t, noise):
213
+ return (
214
+ extract_into_tensor(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
215
+ extract_into_tensor(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
216
+ )
217
+
218
+ def q_posterior(self, x_start, x_t, t):
219
+ posterior_mean = (
220
+ extract_into_tensor(self.posterior_mean_coef1, t, x_t.shape) * x_start +
221
+ extract_into_tensor(self.posterior_mean_coef2, t, x_t.shape) * x_t
222
+ )
223
+ posterior_variance = extract_into_tensor(self.posterior_variance, t, x_t.shape)
224
+ posterior_log_variance_clipped = extract_into_tensor(self.posterior_log_variance_clipped, t, x_t.shape)
225
+ return posterior_mean, posterior_variance, posterior_log_variance_clipped
226
+
227
+ def p_mean_variance(self, x, t, clip_denoised: bool):
228
+ model_out = self.model(x, t)
229
+ if self.parameterization == "eps":
230
+ x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
231
+ elif self.parameterization == "x0":
232
+ x_recon = model_out
233
+ if clip_denoised:
234
+ x_recon.clamp_(-1., 1.)
235
+
236
+ model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
237
+ return model_mean, posterior_variance, posterior_log_variance
238
+
239
+ @torch.no_grad()
240
+ def p_sample(self, x, t, clip_denoised=True, repeat_noise=False):
241
+ b, *_, device = *x.shape, x.device
242
+ model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, clip_denoised=clip_denoised)
243
+ noise = noise_like(x.shape, device, repeat_noise)
244
+ # no noise when t == 0
245
+ nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
246
+ return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
247
+
248
+ @torch.no_grad()
249
+ def p_sample_loop(self, shape, return_intermediates=False):
250
+ device = self.betas.device
251
+ b = shape[0]
252
+ img = torch.randn(shape, device=device)
253
+ intermediates = [img]
254
+ for i in tqdm(reversed(range(0, self.num_timesteps)), desc='Sampling t', total=self.num_timesteps):
255
+ img = self.p_sample(img, torch.full((b,), i, device=device, dtype=torch.long),
256
+ clip_denoised=self.clip_denoised)
257
+ if i % self.log_every_t == 0 or i == self.num_timesteps - 1:
258
+ intermediates.append(img)
259
+ if return_intermediates:
260
+ return img, intermediates
261
+ return img
262
+
263
+ @torch.no_grad()
264
+ def sample(self, batch_size=16, return_intermediates=False):
265
+ image_size = self.image_size
266
+ channels = self.channels
267
+ return self.p_sample_loop((batch_size, channels, image_size, image_size),
268
+ return_intermediates=return_intermediates)
269
+
270
+ def q_sample(self, x_start, t, noise=None):
271
+ noise = default(noise, lambda: torch.randn_like(x_start))
272
+ return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start *
273
+ extract_into_tensor(self.scale_arr, t, x_start.shape) +
274
+ extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
275
+
276
+ def get_input(self, batch, k):
277
+ x = batch[k]
278
+ x = x.to(memory_format=torch.contiguous_format).float()
279
+ return x
280
+
281
+ def _get_rows_from_list(self, samples):
282
+ n_imgs_per_row = len(samples)
283
+ denoise_grid = rearrange(samples, 'n b c h w -> b n c h w')
284
+ denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
285
+ denoise_grid = make_grid(denoise_grid, nrow=n_imgs_per_row)
286
+ return denoise_grid
287
+
288
+ @torch.no_grad()
289
+ def log_images(self, batch, N=8, n_row=2, sample=True, return_keys=None, **kwargs):
290
+ log = dict()
291
+ x = self.get_input(batch, self.first_stage_key)
292
+ N = min(x.shape[0], N)
293
+ n_row = min(x.shape[0], n_row)
294
+ x = x.to(self.device)[:N]
295
+ log["inputs"] = x
296
+
297
+ # get diffusion row
298
+ diffusion_row = list()
299
+ x_start = x[:n_row]
300
+
301
+ for t in range(self.num_timesteps):
302
+ if t % self.log_every_t == 0 or t == self.num_timesteps - 1:
303
+ t = repeat(torch.tensor([t]), '1 -> b', b=n_row)
304
+ t = t.to(self.device).long()
305
+ noise = torch.randn_like(x_start)
306
+ x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
307
+ diffusion_row.append(x_noisy)
308
+
309
+ log["diffusion_row"] = self._get_rows_from_list(diffusion_row)
310
+
311
+ if sample:
312
+ # get denoise row
313
+ with self.ema_scope("Plotting"):
314
+ samples, denoise_row = self.sample(batch_size=N, return_intermediates=True)
315
+
316
+ log["samples"] = samples
317
+ log["denoise_row"] = self._get_rows_from_list(denoise_row)
318
+
319
+ if return_keys:
320
+ if np.intersect1d(list(log.keys()), return_keys).shape[0] == 0:
321
+ return log
322
+ else:
323
+ return {key: log[key] for key in return_keys}
324
+ return log
325
+
326
+
327
+ class LatentDiffusion(DDPM):
328
+ """main class"""
329
+ def __init__(self,
330
+ first_stage_config,
331
+ cond_stage_config,
332
+ num_timesteps_cond=None,
333
+ cond_stage_key="caption",
334
+ cond_stage_trainable=False,
335
+ cond_stage_forward=None,
336
+ conditioning_key=None,
337
+ uncond_prob=0.2,
338
+ uncond_type="empty_seq",
339
+ scale_factor=1.0,
340
+ scale_by_std=False,
341
+ encoder_type="2d",
342
+ only_model=False,
343
+ use_scale=False,
344
+ scale_a=1,
345
+ scale_b=0.3,
346
+ mid_step=400,
347
+ fix_scale_bug=False,
348
+ *args, **kwargs):
349
+ self.num_timesteps_cond = default(num_timesteps_cond, 1)
350
+ self.scale_by_std = scale_by_std
351
+ assert self.num_timesteps_cond <= kwargs['timesteps']
352
+ # for backwards compatibility after implementation of DiffusionWrapper
353
+ ckpt_path = kwargs.pop("ckpt_path", None)
354
+ ignore_keys = kwargs.pop("ignore_keys", [])
355
+ conditioning_key = default(conditioning_key, 'crossattn')
356
+ super().__init__(conditioning_key=conditioning_key, *args, **kwargs)
357
+
358
+ self.cond_stage_trainable = cond_stage_trainable
359
+ self.cond_stage_key = cond_stage_key
360
+
361
+ # scale factor
362
+ self.use_scale=use_scale
363
+ if self.use_scale:
364
+ self.scale_a=scale_a
365
+ self.scale_b=scale_b
366
+ if fix_scale_bug:
367
+ scale_step=self.num_timesteps-mid_step
368
+ else: #bug
369
+ scale_step = self.num_timesteps
370
+
371
+ scale_arr1 = np.linspace(scale_a, scale_b, mid_step)
372
+ scale_arr2 = np.full(scale_step, scale_b)
373
+ scale_arr = np.concatenate((scale_arr1, scale_arr2))
374
+ scale_arr_prev = np.append(scale_a, scale_arr[:-1])
375
+ to_torch = partial(torch.tensor, dtype=torch.float32)
376
+ self.register_buffer('scale_arr', to_torch(scale_arr))
377
+
378
+ try:
379
+ self.num_downs = len(first_stage_config.params.ddconfig.ch_mult) - 1
380
+ except:
381
+ self.num_downs = 0
382
+ if not scale_by_std:
383
+ self.scale_factor = scale_factor
384
+ else:
385
+ self.register_buffer('scale_factor', torch.tensor(scale_factor))
386
+ self.instantiate_first_stage(first_stage_config)
387
+ self.instantiate_cond_stage(cond_stage_config)
388
+ self.first_stage_config = first_stage_config
389
+ self.cond_stage_config = cond_stage_config
390
+ self.clip_denoised = False
391
+
392
+ self.cond_stage_forward = cond_stage_forward
393
+ self.encoder_type = encoder_type
394
+ assert(encoder_type in ["2d", "3d"])
395
+ self.uncond_prob = uncond_prob
396
+ self.classifier_free_guidance = True if uncond_prob > 0 else False
397
+ assert(uncond_type in ["zero_embed", "empty_seq"])
398
+ self.uncond_type = uncond_type
399
+
400
+
401
+ self.restarted_from_ckpt = False
402
+ if ckpt_path is not None:
403
+ self.init_from_ckpt(ckpt_path, ignore_keys, only_model=only_model)
404
+ self.restarted_from_ckpt = True
405
+
406
+
407
+ def make_cond_schedule(self, ):
408
+ self.cond_ids = torch.full(size=(self.num_timesteps,), fill_value=self.num_timesteps - 1, dtype=torch.long)
409
+ ids = torch.round(torch.linspace(0, self.num_timesteps - 1, self.num_timesteps_cond)).long()
410
+ self.cond_ids[:self.num_timesteps_cond] = ids
411
+
412
+ def q_sample(self, x_start, t, noise=None):
413
+ noise = default(noise, lambda: torch.randn_like(x_start))
414
+ if self.use_scale:
415
+ return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start *
416
+ extract_into_tensor(self.scale_arr, t, x_start.shape) +
417
+ extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
418
+ else:
419
+ return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
420
+ extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise)
421
+
422
+
423
+ def _freeze_model(self):
424
+ for name, para in self.model.diffusion_model.named_parameters():
425
+ para.requires_grad = False
426
+
427
+ def instantiate_first_stage(self, config):
428
+ model = instantiate_from_config(config)
429
+ self.first_stage_model = model.eval()
430
+ self.first_stage_model.train = disabled_train
431
+ for param in self.first_stage_model.parameters():
432
+ param.requires_grad = False
433
+
434
+ def instantiate_cond_stage(self, config):
435
+ if not self.cond_stage_trainable:
436
+ model = instantiate_from_config(config)
437
+ self.cond_stage_model = model.eval()
438
+ self.cond_stage_model.train = disabled_train
439
+ for param in self.cond_stage_model.parameters():
440
+ param.requires_grad = False
441
+ else:
442
+ model = instantiate_from_config(config)
443
+ self.cond_stage_model = model
444
+
445
+ def get_learned_conditioning(self, c):
446
+ if self.cond_stage_forward is None:
447
+ if hasattr(self.cond_stage_model, 'encode') and callable(self.cond_stage_model.encode):
448
+ c = self.cond_stage_model.encode(c)
449
+ if isinstance(c, DiagonalGaussianDistribution):
450
+ c = c.mode()
451
+ else:
452
+ c = self.cond_stage_model(c)
453
+ else:
454
+ assert hasattr(self.cond_stage_model, self.cond_stage_forward)
455
+ c = getattr(self.cond_stage_model, self.cond_stage_forward)(c)
456
+ return c
457
+
458
+ def get_first_stage_encoding(self, encoder_posterior, noise=None):
459
+ if isinstance(encoder_posterior, DiagonalGaussianDistribution):
460
+ z = encoder_posterior.sample(noise=noise)
461
+ elif isinstance(encoder_posterior, torch.Tensor):
462
+ z = encoder_posterior
463
+ else:
464
+ raise NotImplementedError(f"encoder_posterior of type '{type(encoder_posterior)}' not yet implemented")
465
+ return self.scale_factor * z
466
+
467
+ @torch.no_grad()
468
+ def encode_first_stage(self, x):
469
+ if self.encoder_type == "2d" and x.dim() == 5:
470
+ b, _, t, _, _ = x.shape
471
+ x = rearrange(x, 'b c t h w -> (b t) c h w')
472
+ reshape_back = True
473
+ else:
474
+ reshape_back = False
475
+
476
+ encoder_posterior = self.first_stage_model.encode(x)
477
+ results = self.get_first_stage_encoding(encoder_posterior).detach()
478
+
479
+ if reshape_back:
480
+ results = rearrange(results, '(b t) c h w -> b c t h w', b=b,t=t)
481
+
482
+ return results
483
+
484
+ @torch.no_grad()
485
+ def encode_first_stage_2DAE(self, x):
486
+
487
+ b, _, t, _, _ = x.shape
488
+ results = torch.cat([self.get_first_stage_encoding(self.first_stage_model.encode(x[:,:,i])).detach().unsqueeze(2) for i in range(t)], dim=2)
489
+
490
+ return results
491
+
492
+ def decode_core(self, z, **kwargs):
493
+ if self.encoder_type == "2d" and z.dim() == 5:
494
+ b, _, t, _, _ = z.shape
495
+ z = rearrange(z, 'b c t h w -> (b t) c h w')
496
+ reshape_back = True
497
+ else:
498
+ reshape_back = False
499
+
500
+ z = 1. / self.scale_factor * z
501
+
502
+ results = self.first_stage_model.decode(z, **kwargs)
503
+
504
+ if reshape_back:
505
+ results = rearrange(results, '(b t) c h w -> b c t h w', b=b,t=t)
506
+ return results
507
+
508
+ @torch.no_grad()
509
+ def decode_first_stage(self, z, **kwargs):
510
+ return self.decode_core(z, **kwargs)
511
+
512
+ def apply_model(self, x_noisy, t, cond, **kwargs):
513
+ if isinstance(cond, dict):
514
+ # hybrid case, cond is exptected to be a dict
515
+ pass
516
+ else:
517
+ if not isinstance(cond, list):
518
+ cond = [cond]
519
+ key = 'c_concat' if self.model.conditioning_key == 'concat' else 'c_crossattn'
520
+ cond = {key: cond}
521
+
522
+ x_recon = self.model(x_noisy, t, **cond, **kwargs)
523
+
524
+ if isinstance(x_recon, tuple):
525
+ return x_recon[0]
526
+ else:
527
+ return x_recon
528
+
529
+ def _get_denoise_row_from_list(self, samples, desc=''):
530
+ denoise_row = []
531
+ for zd in tqdm(samples, desc=desc):
532
+ denoise_row.append(self.decode_first_stage(zd.to(self.device)))
533
+ n_log_timesteps = len(denoise_row)
534
+
535
+ denoise_row = torch.stack(denoise_row) # n_log_timesteps, b, C, H, W
536
+
537
+ if denoise_row.dim() == 5:
538
+ # img, num_imgs= n_log_timesteps * bs, grid_size=[bs,n_log_timesteps]
539
+ denoise_grid = rearrange(denoise_row, 'n b c h w -> b n c h w')
540
+ denoise_grid = rearrange(denoise_grid, 'b n c h w -> (b n) c h w')
541
+ denoise_grid = make_grid(denoise_grid, nrow=n_log_timesteps)
542
+ elif denoise_row.dim() == 6:
543
+ # video, grid_size=[n_log_timesteps*bs, t]
544
+ video_length = denoise_row.shape[3]
545
+ denoise_grid = rearrange(denoise_row, 'n b c t h w -> b n c t h w')
546
+ denoise_grid = rearrange(denoise_grid, 'b n c t h w -> (b n) c t h w')
547
+ denoise_grid = rearrange(denoise_grid, 'n c t h w -> (n t) c h w')
548
+ denoise_grid = make_grid(denoise_grid, nrow=video_length)
549
+ else:
550
+ raise ValueError
551
+
552
+ return denoise_grid
553
+
554
+
555
+ @torch.no_grad()
556
+ def decode_first_stage_2DAE(self, z, **kwargs):
557
+
558
+ b, _, t, _, _ = z.shape
559
+ z = 1. / self.scale_factor * z
560
+ results = torch.cat([self.first_stage_model.decode(z[:,:,i], **kwargs).unsqueeze(2) for i in range(t)], dim=2)
561
+
562
+ return results
563
+
564
+
565
+ def p_mean_variance(self, x, c, t, clip_denoised: bool, return_x0=False, score_corrector=None, corrector_kwargs=None, **kwargs):
566
+ t_in = t
567
+ model_out = self.apply_model(x, t_in, c, **kwargs)
568
+
569
+ if score_corrector is not None:
570
+ assert self.parameterization == "eps"
571
+ model_out = score_corrector.modify_score(self, model_out, x, t, c, **corrector_kwargs)
572
+
573
+ if self.parameterization == "eps":
574
+ x_recon = self.predict_start_from_noise(x, t=t, noise=model_out)
575
+ elif self.parameterization == "x0":
576
+ x_recon = model_out
577
+ else:
578
+ raise NotImplementedError()
579
+
580
+ if clip_denoised:
581
+ x_recon.clamp_(-1., 1.)
582
+
583
+ model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
584
+
585
+ if return_x0:
586
+ return model_mean, posterior_variance, posterior_log_variance, x_recon
587
+ else:
588
+ return model_mean, posterior_variance, posterior_log_variance
589
+
590
+ @torch.no_grad()
591
+ def p_sample(self, x, c, t, clip_denoised=False, repeat_noise=False, return_x0=False, \
592
+ temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, **kwargs):
593
+ b, *_, device = *x.shape, x.device
594
+ outputs = self.p_mean_variance(x=x, c=c, t=t, clip_denoised=clip_denoised, return_x0=return_x0, \
595
+ score_corrector=score_corrector, corrector_kwargs=corrector_kwargs, **kwargs)
596
+ if return_x0:
597
+ model_mean, _, model_log_variance, x0 = outputs
598
+ else:
599
+ model_mean, _, model_log_variance = outputs
600
+
601
+ noise = noise_like(x.shape, device, repeat_noise) * temperature
602
+ if noise_dropout > 0.:
603
+ noise = torch.nn.functional.dropout(noise, p=noise_dropout)
604
+ # no noise when t == 0
605
+ nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
606
+
607
+ if return_x0:
608
+ return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise, x0
609
+ else:
610
+ return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
611
+
612
+ @torch.no_grad()
613
+ def p_sample_loop(self, cond, shape, return_intermediates=False, x_T=None, verbose=True, callback=None, \
614
+ timesteps=None, mask=None, x0=None, img_callback=None, start_T=None, log_every_t=None, **kwargs):
615
+
616
+ if not log_every_t:
617
+ log_every_t = self.log_every_t
618
+ device = self.betas.device
619
+ b = shape[0]
620
+ # sample an initial noise
621
+ if x_T is None:
622
+ img = torch.randn(shape, device=device)
623
+ else:
624
+ img = x_T
625
+
626
+ intermediates = [img]
627
+ if timesteps is None:
628
+ timesteps = self.num_timesteps
629
+ if start_T is not None:
630
+ timesteps = min(timesteps, start_T)
631
+
632
+ iterator = tqdm(reversed(range(0, timesteps)), desc='Sampling t', total=timesteps) if verbose else reversed(range(0, timesteps))
633
+
634
+ if mask is not None:
635
+ assert x0 is not None
636
+ assert x0.shape[2:3] == mask.shape[2:3] # spatial size has to match
637
+
638
+ for i in iterator:
639
+ ts = torch.full((b,), i, device=device, dtype=torch.long)
640
+ if self.shorten_cond_schedule:
641
+ assert self.model.conditioning_key != 'hybrid'
642
+ tc = self.cond_ids[ts].to(cond.device)
643
+ cond = self.q_sample(x_start=cond, t=tc, noise=torch.randn_like(cond))
644
+
645
+ img = self.p_sample(img, cond, ts, clip_denoised=self.clip_denoised, **kwargs)
646
+ if mask is not None:
647
+ img_orig = self.q_sample(x0, ts)
648
+ img = img_orig * mask + (1. - mask) * img
649
+
650
+ if i % log_every_t == 0 or i == timesteps - 1:
651
+ intermediates.append(img)
652
+ if callback: callback(i)
653
+ if img_callback: img_callback(img, i)
654
+
655
+ if return_intermediates:
656
+ return img, intermediates
657
+ return img
658
+
659
+
660
+ class LatentVisualDiffusion(LatentDiffusion):
661
+ def __init__(self, cond_img_config, finegrained=False, random_cond=False, *args, **kwargs):
662
+ super().__init__(*args, **kwargs)
663
+ self.random_cond = random_cond
664
+ self.instantiate_img_embedder(cond_img_config, freeze=True)
665
+ num_tokens = 16 if finegrained else 4
666
+ self.image_proj_model = self.init_projector(use_finegrained=finegrained, num_tokens=num_tokens, input_dim=1024,\
667
+ cross_attention_dim=1024, dim=1280)
668
+
669
+ def instantiate_img_embedder(self, config, freeze=True):
670
+ embedder = instantiate_from_config(config)
671
+ if freeze:
672
+ self.embedder = embedder.eval()
673
+ self.embedder.train = disabled_train
674
+ for param in self.embedder.parameters():
675
+ param.requires_grad = False
676
+
677
+ def init_projector(self, use_finegrained, num_tokens, input_dim, cross_attention_dim, dim):
678
+ if not use_finegrained:
679
+ image_proj_model = ImageProjModel(clip_extra_context_tokens=num_tokens, cross_attention_dim=cross_attention_dim,
680
+ clip_embeddings_dim=input_dim
681
+ )
682
+ else:
683
+ image_proj_model = Resampler(dim=input_dim, depth=4, dim_head=64, heads=12, num_queries=num_tokens,
684
+ embedding_dim=dim, output_dim=cross_attention_dim, ff_mult=4
685
+ )
686
+ return image_proj_model
687
+
688
+ ## Never delete this func: it is used in log_images() and inference stage
689
+ def get_image_embeds(self, batch_imgs):
690
+ ## img: b c h w
691
+ img_token = self.embedder(batch_imgs)
692
+ img_emb = self.image_proj_model(img_token)
693
+ return img_emb
694
+
695
+
696
+ class DiffusionWrapper(pl.LightningModule):
697
+ def __init__(self, diff_model_config, conditioning_key):
698
+ super().__init__()
699
+ self.diffusion_model = instantiate_from_config(diff_model_config)
700
+ self.conditioning_key = conditioning_key
701
+
702
+ def forward(self, x, t, c_concat: list = None, c_crossattn: list = None,
703
+ c_adm=None, s=None, mask=None, **kwargs):
704
+ # temporal_context = fps is foNone
705
+ if self.conditioning_key is None:
706
+ out = self.diffusion_model(x, t)
707
+ elif self.conditioning_key == 'concat':
708
+ xc = torch.cat([x] + c_concat, dim=1)
709
+ out = self.diffusion_model(xc, t, **kwargs)
710
+ elif self.conditioning_key == 'crossattn':
711
+ cc = torch.cat(c_crossattn, 1)
712
+ out = self.diffusion_model(x, t, context=cc, **kwargs)
713
+ elif self.conditioning_key == 'hybrid':
714
+ ## it is just right [b,c,t,h,w]: concatenate in channel dim
715
+ xc = torch.cat([x] + c_concat, dim=1)
716
+ cc = torch.cat(c_crossattn, 1)
717
+ out = self.diffusion_model(xc, t, context=cc)
718
+ elif self.conditioning_key == 'resblockcond':
719
+ cc = c_crossattn[0]
720
+ out = self.diffusion_model(x, t, context=cc)
721
+ elif self.conditioning_key == 'adm':
722
+ cc = c_crossattn[0]
723
+ out = self.diffusion_model(x, t, y=cc)
724
+ elif self.conditioning_key == 'hybrid-adm':
725
+ assert c_adm is not None
726
+ xc = torch.cat([x] + c_concat, dim=1)
727
+ cc = torch.cat(c_crossattn, 1)
728
+ out = self.diffusion_model(xc, t, context=cc, y=c_adm)
729
+ elif self.conditioning_key == 'hybrid-time':
730
+ assert s is not None
731
+ xc = torch.cat([x] + c_concat, dim=1)
732
+ cc = torch.cat(c_crossattn, 1)
733
+ out = self.diffusion_model(xc, t, context=cc, s=s)
734
+ elif self.conditioning_key == 'concat-time-mask':
735
+ # assert s is not None
736
+ # mainlogger.info('x & mask:',x.shape,c_concat[0].shape)
737
+ xc = torch.cat([x] + c_concat, dim=1)
738
+ out = self.diffusion_model(xc, t, context=None, s=s, mask=mask)
739
+ elif self.conditioning_key == 'concat-adm-mask':
740
+ # assert s is not None
741
+ # mainlogger.info('x & mask:',x.shape,c_concat[0].shape)
742
+ if c_concat is not None:
743
+ xc = torch.cat([x] + c_concat, dim=1)
744
+ else:
745
+ xc = x
746
+ out = self.diffusion_model(xc, t, context=None, y=s, mask=mask)
747
+ elif self.conditioning_key == 'hybrid-adm-mask':
748
+ cc = torch.cat(c_crossattn, 1)
749
+ if c_concat is not None:
750
+ xc = torch.cat([x] + c_concat, dim=1)
751
+ else:
752
+ xc = x
753
+ out = self.diffusion_model(xc, t, context=cc, y=s, mask=mask)
754
+ elif self.conditioning_key == 'hybrid-time-adm': # adm means y, e.g., class index
755
+ # assert s is not None
756
+ assert c_adm is not None
757
+ xc = torch.cat([x] + c_concat, dim=1)
758
+ cc = torch.cat(c_crossattn, 1)
759
+ out = self.diffusion_model(xc, t, context=cc, s=s, y=c_adm)
760
+ else:
761
+ raise NotImplementedError()
762
+
763
+ return out
lvdm/models/samplers/ddim.py ADDED
@@ -0,0 +1,336 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ from tqdm import tqdm
3
+ import torch
4
+ from lvdm.models.utils_diffusion import make_ddim_sampling_parameters, make_ddim_timesteps
5
+ from lvdm.common import noise_like
6
+
7
+
8
+ class DDIMSampler(object):
9
+ def __init__(self, model, schedule="linear", **kwargs):
10
+ super().__init__()
11
+ self.model = model
12
+ self.ddpm_num_timesteps = model.num_timesteps
13
+ self.schedule = schedule
14
+ self.counter = 0
15
+
16
+ def register_buffer(self, name, attr):
17
+ if type(attr) == torch.Tensor:
18
+ if attr.device != torch.device("cuda"):
19
+ attr = attr.to(torch.device("cuda"))
20
+ setattr(self, name, attr)
21
+
22
+ def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
23
+ self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
24
+ num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
25
+ alphas_cumprod = self.model.alphas_cumprod
26
+ assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
27
+ to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
28
+
29
+ self.register_buffer('betas', to_torch(self.model.betas))
30
+ self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
31
+ self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
32
+ self.use_scale = self.model.use_scale
33
+ print('DDIM scale', self.use_scale)
34
+
35
+ if self.use_scale:
36
+ self.register_buffer('scale_arr', to_torch(self.model.scale_arr))
37
+ ddim_scale_arr = self.scale_arr.cpu()[self.ddim_timesteps]
38
+ self.register_buffer('ddim_scale_arr', ddim_scale_arr)
39
+ ddim_scale_arr = np.asarray([self.scale_arr.cpu()[0]] + self.scale_arr.cpu()[self.ddim_timesteps[:-1]].tolist())
40
+ self.register_buffer('ddim_scale_arr_prev', ddim_scale_arr)
41
+
42
+ # calculations for diffusion q(x_t | x_{t-1}) and others
43
+ self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
44
+ self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
45
+ self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
46
+ self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
47
+ self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
48
+
49
+ # ddim sampling parameters
50
+ ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
51
+ ddim_timesteps=self.ddim_timesteps,
52
+ eta=ddim_eta,verbose=verbose)
53
+ self.register_buffer('ddim_sigmas', ddim_sigmas)
54
+ self.register_buffer('ddim_alphas', ddim_alphas)
55
+ self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
56
+ self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
57
+ sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
58
+ (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
59
+ 1 - self.alphas_cumprod / self.alphas_cumprod_prev))
60
+ self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
61
+
62
+ @torch.no_grad()
63
+ def sample(self,
64
+ S,
65
+ batch_size,
66
+ shape,
67
+ conditioning=None,
68
+ callback=None,
69
+ normals_sequence=None,
70
+ img_callback=None,
71
+ quantize_x0=False,
72
+ eta=0.,
73
+ mask=None,
74
+ x0=None,
75
+ temperature=1.,
76
+ noise_dropout=0.,
77
+ score_corrector=None,
78
+ corrector_kwargs=None,
79
+ verbose=True,
80
+ schedule_verbose=False,
81
+ x_T=None,
82
+ log_every_t=100,
83
+ unconditional_guidance_scale=1.,
84
+ unconditional_conditioning=None,
85
+ # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
86
+ **kwargs
87
+ ):
88
+
89
+ # check condition bs
90
+ if conditioning is not None:
91
+ if isinstance(conditioning, dict):
92
+ try:
93
+ cbs = conditioning[list(conditioning.keys())[0]].shape[0]
94
+ except:
95
+ cbs = conditioning[list(conditioning.keys())[0]][0].shape[0]
96
+
97
+ if cbs != batch_size:
98
+ print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
99
+ else:
100
+ if conditioning.shape[0] != batch_size:
101
+ print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
102
+
103
+ self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=schedule_verbose)
104
+
105
+ # make shape
106
+ if len(shape) == 3:
107
+ C, H, W = shape
108
+ size = (batch_size, C, H, W)
109
+ elif len(shape) == 4:
110
+ C, T, H, W = shape
111
+ size = (batch_size, C, T, H, W)
112
+ # print(f'Data shape for DDIM sampling is {size}, eta {eta}')
113
+
114
+ samples, intermediates = self.ddim_sampling(conditioning, size,
115
+ callback=callback,
116
+ img_callback=img_callback,
117
+ quantize_denoised=quantize_x0,
118
+ mask=mask, x0=x0,
119
+ ddim_use_original_steps=False,
120
+ noise_dropout=noise_dropout,
121
+ temperature=temperature,
122
+ score_corrector=score_corrector,
123
+ corrector_kwargs=corrector_kwargs,
124
+ x_T=x_T,
125
+ log_every_t=log_every_t,
126
+ unconditional_guidance_scale=unconditional_guidance_scale,
127
+ unconditional_conditioning=unconditional_conditioning,
128
+ verbose=verbose,
129
+ **kwargs)
130
+ return samples, intermediates
131
+
132
+ @torch.no_grad()
133
+ def ddim_sampling(self, cond, shape,
134
+ x_T=None, ddim_use_original_steps=False,
135
+ callback=None, timesteps=None, quantize_denoised=False,
136
+ mask=None, x0=None, img_callback=None, log_every_t=100,
137
+ temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
138
+ unconditional_guidance_scale=1., unconditional_conditioning=None, verbose=True,
139
+ cond_tau=1., target_size=None, start_timesteps=None,
140
+ **kwargs):
141
+ device = self.model.betas.device
142
+ print('ddim device', device)
143
+ b = shape[0]
144
+ if x_T is None:
145
+ img = torch.randn(shape, device=device)
146
+ else:
147
+ img = x_T
148
+
149
+ if timesteps is None:
150
+ timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
151
+ elif timesteps is not None and not ddim_use_original_steps:
152
+ subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
153
+ timesteps = self.ddim_timesteps[:subset_end]
154
+
155
+ intermediates = {'x_inter': [img], 'pred_x0': [img]}
156
+ time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
157
+ total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
158
+ if verbose:
159
+ iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
160
+ else:
161
+ iterator = time_range
162
+
163
+ init_x0 = False
164
+ clean_cond = kwargs.pop("clean_cond", False)
165
+ for i, step in enumerate(iterator):
166
+ index = total_steps - i - 1
167
+ ts = torch.full((b,), step, device=device, dtype=torch.long)
168
+ if start_timesteps is not None:
169
+ assert x0 is not None
170
+ if step > start_timesteps*time_range[0]:
171
+ continue
172
+ elif not init_x0:
173
+ img = self.model.q_sample(x0, ts)
174
+ init_x0 = True
175
+
176
+ # use mask to blend noised original latent (img_orig) & new sampled latent (img)
177
+ if mask is not None:
178
+ assert x0 is not None
179
+ if clean_cond:
180
+ img_orig = x0
181
+ else:
182
+ img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass? <ddim inversion>
183
+ img = img_orig * mask + (1. - mask) * img # keep original & modify use img
184
+
185
+ index_clip = int((1 - cond_tau) * total_steps)
186
+ if index <= index_clip and target_size is not None:
187
+ target_size_ = [target_size[0], target_size[1]//8, target_size[2]//8]
188
+ img = torch.nn.functional.interpolate(
189
+ img,
190
+ size=target_size_,
191
+ mode="nearest",
192
+ )
193
+ outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
194
+ quantize_denoised=quantize_denoised, temperature=temperature,
195
+ noise_dropout=noise_dropout, score_corrector=score_corrector,
196
+ corrector_kwargs=corrector_kwargs,
197
+ unconditional_guidance_scale=unconditional_guidance_scale,
198
+ unconditional_conditioning=unconditional_conditioning,
199
+ x0=x0,
200
+ **kwargs)
201
+
202
+ img, pred_x0 = outs
203
+ if callback: callback(i)
204
+ if img_callback: img_callback(pred_x0, i)
205
+
206
+ if index % log_every_t == 0 or index == total_steps - 1:
207
+ intermediates['x_inter'].append(img)
208
+ intermediates['pred_x0'].append(pred_x0)
209
+
210
+ return img, intermediates
211
+
212
+ @torch.no_grad()
213
+ def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
214
+ temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
215
+ unconditional_guidance_scale=1., unconditional_conditioning=None,
216
+ uc_type=None, conditional_guidance_scale_temporal=None, **kwargs):
217
+ b, *_, device = *x.shape, x.device
218
+ if x.dim() == 5:
219
+ is_video = True
220
+ else:
221
+ is_video = False
222
+ if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
223
+ e_t = self.model.apply_model(x, t, c, **kwargs) # unet denoiser
224
+ else:
225
+ # with unconditional condition
226
+ if isinstance(c, torch.Tensor):
227
+ e_t = self.model.apply_model(x, t, c, **kwargs)
228
+ e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **kwargs)
229
+ elif isinstance(c, dict):
230
+ e_t = self.model.apply_model(x, t, c, **kwargs)
231
+ e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **kwargs)
232
+ else:
233
+ raise NotImplementedError
234
+ # text cfg
235
+ if uc_type is None:
236
+ e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
237
+ else:
238
+ if uc_type == 'cfg_original':
239
+ e_t = e_t + unconditional_guidance_scale * (e_t - e_t_uncond)
240
+ elif uc_type == 'cfg_ours':
241
+ e_t = e_t + unconditional_guidance_scale * (e_t_uncond - e_t)
242
+ else:
243
+ raise NotImplementedError
244
+ # temporal guidance
245
+ if conditional_guidance_scale_temporal is not None:
246
+ e_t_temporal = self.model.apply_model(x, t, c, **kwargs)
247
+ e_t_image = self.model.apply_model(x, t, c, no_temporal_attn=True, **kwargs)
248
+ e_t = e_t + conditional_guidance_scale_temporal * (e_t_temporal - e_t_image)
249
+
250
+ if score_corrector is not None:
251
+ assert self.model.parameterization == "eps"
252
+ e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
253
+
254
+ alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
255
+ alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
256
+ sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
257
+ sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
258
+ # select parameters corresponding to the currently considered timestep
259
+
260
+ if is_video:
261
+ size = (b, 1, 1, 1, 1)
262
+ else:
263
+ size = (b, 1, 1, 1)
264
+ a_t = torch.full(size, alphas[index], device=device)
265
+ a_prev = torch.full(size, alphas_prev[index], device=device)
266
+ sigma_t = torch.full(size, sigmas[index], device=device)
267
+ sqrt_one_minus_at = torch.full(size, sqrt_one_minus_alphas[index],device=device)
268
+
269
+ # current prediction for x_0
270
+ pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
271
+ if quantize_denoised:
272
+ pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
273
+ # direction pointing to x_t
274
+ dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
275
+
276
+ noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
277
+ if noise_dropout > 0.:
278
+ noise = torch.nn.functional.dropout(noise, p=noise_dropout)
279
+
280
+ alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
281
+ if self.use_scale:
282
+ scale_arr = self.model.scale_arr if use_original_steps else self.ddim_scale_arr
283
+ scale_t = torch.full(size, scale_arr[index], device=device)
284
+ scale_arr_prev = self.model.scale_arr_prev if use_original_steps else self.ddim_scale_arr_prev
285
+ scale_t_prev = torch.full(size, scale_arr_prev[index], device=device)
286
+ pred_x0 /= scale_t
287
+ x_prev = a_prev.sqrt() * scale_t_prev * pred_x0 + dir_xt + noise
288
+ else:
289
+ x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
290
+
291
+ return x_prev, pred_x0
292
+
293
+
294
+ @torch.no_grad()
295
+ def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
296
+ # fast, but does not allow for exact reconstruction
297
+ # t serves as an index to gather the correct alphas
298
+ if use_original_steps:
299
+ sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
300
+ sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
301
+ else:
302
+ sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
303
+ sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
304
+
305
+ if noise is None:
306
+ noise = torch.randn_like(x0)
307
+
308
+ def extract_into_tensor(a, t, x_shape):
309
+ b, *_ = t.shape
310
+ out = a.gather(-1, t)
311
+ return out.reshape(b, *((1,) * (len(x_shape) - 1)))
312
+
313
+ return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
314
+ extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise)
315
+
316
+ @torch.no_grad()
317
+ def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None,
318
+ use_original_steps=False):
319
+
320
+ timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps
321
+ timesteps = timesteps[:t_start]
322
+
323
+ time_range = np.flip(timesteps)
324
+ total_steps = timesteps.shape[0]
325
+ print(f"Running DDIM Sampling with {total_steps} timesteps")
326
+
327
+ iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
328
+ x_dec = x_latent
329
+ for i, step in enumerate(iterator):
330
+ index = total_steps - i - 1
331
+ ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long)
332
+ x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps,
333
+ unconditional_guidance_scale=unconditional_guidance_scale,
334
+ unconditional_conditioning=unconditional_conditioning)
335
+ return x_dec
336
+
lvdm/models/samplers/ddim_mp.py ADDED
@@ -0,0 +1,340 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ from tqdm import tqdm
3
+ import torch
4
+ from lvdm.models.utils_diffusion import make_ddim_sampling_parameters, make_ddim_timesteps
5
+ from lvdm.common import noise_like
6
+
7
+
8
+ class DDIMSampler(object):
9
+ def __init__(self, model, schedule="linear", **kwargs):
10
+ super().__init__()
11
+ self.model = model
12
+ self.ddpm_num_timesteps = model.num_timesteps
13
+ self.schedule = schedule
14
+ self.counter = 0
15
+
16
+ def register_buffer(self, name, attr):
17
+ if type(attr) == torch.Tensor:
18
+ if attr.device != torch.device("cuda"):
19
+ attr = attr.to(torch.device("cuda"))
20
+ setattr(self, name, attr)
21
+
22
+ def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True):
23
+ self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps,
24
+ num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose)
25
+ alphas_cumprod = self.model.alphas_cumprod
26
+ assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep'
27
+ to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device)
28
+
29
+ self.register_buffer('betas', to_torch(self.model.betas))
30
+ self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
31
+ self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev))
32
+ self.use_scale = self.model.use_scale
33
+ print('DDIM scale', self.use_scale)
34
+
35
+ if self.use_scale:
36
+ self.register_buffer('scale_arr', to_torch(self.model.scale_arr))
37
+ ddim_scale_arr = self.scale_arr.cpu()[self.ddim_timesteps]
38
+ self.register_buffer('ddim_scale_arr', ddim_scale_arr)
39
+ ddim_scale_arr = np.asarray([self.scale_arr.cpu()[0]] + self.scale_arr.cpu()[self.ddim_timesteps[:-1]].tolist())
40
+ self.register_buffer('ddim_scale_arr_prev', ddim_scale_arr)
41
+
42
+ # calculations for diffusion q(x_t | x_{t-1}) and others
43
+ self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu())))
44
+ self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu())))
45
+ self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu())))
46
+ self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu())))
47
+ self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1)))
48
+
49
+ # ddim sampling parameters
50
+ ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(),
51
+ ddim_timesteps=self.ddim_timesteps,
52
+ eta=ddim_eta,verbose=verbose)
53
+ self.register_buffer('ddim_sigmas', ddim_sigmas)
54
+ self.register_buffer('ddim_alphas', ddim_alphas)
55
+ self.register_buffer('ddim_alphas_prev', ddim_alphas_prev)
56
+ self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas))
57
+ sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt(
58
+ (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * (
59
+ 1 - self.alphas_cumprod / self.alphas_cumprod_prev))
60
+ self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps)
61
+
62
+ @torch.no_grad()
63
+ def sample(self,
64
+ S,
65
+ batch_size,
66
+ shape,
67
+ conditioning=None,
68
+ callback=None,
69
+ normals_sequence=None,
70
+ img_callback=None,
71
+ quantize_x0=False,
72
+ eta=0.,
73
+ mask=None,
74
+ x0=None,
75
+ temperature=1.,
76
+ noise_dropout=0.,
77
+ score_corrector=None,
78
+ corrector_kwargs=None,
79
+ verbose=True,
80
+ schedule_verbose=False,
81
+ x_T=None,
82
+ log_every_t=100,
83
+ unconditional_guidance_scale=1.,
84
+ unconditional_conditioning=None,
85
+ # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ...
86
+ **kwargs
87
+ ):
88
+
89
+ # check condition bs
90
+ # if conditioning is not None:
91
+ # if isinstance(conditioning, dict):
92
+ # try:
93
+ # cbs = conditioning[list(conditioning.keys())[0]].shape[0]
94
+ # except:
95
+ # cbs = conditioning[list(conditioning.keys())[0]][0].shape[0]
96
+
97
+ # if cbs != batch_size:
98
+ # print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}")
99
+ # else:
100
+ # if conditioning.shape[0] != batch_size:
101
+ # print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}")
102
+
103
+ self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=schedule_verbose)
104
+
105
+ # make shape
106
+ if len(shape) == 3:
107
+ C, H, W = shape
108
+ size = (batch_size, C, H, W)
109
+ elif len(shape) == 4:
110
+ C, T, H, W = shape
111
+ size = (batch_size, C, T, H, W)
112
+ # print(f'Data shape for DDIM sampling is {size}, eta {eta}')
113
+
114
+ samples, intermediates = self.ddim_sampling(conditioning, size,
115
+ callback=callback,
116
+ img_callback=img_callback,
117
+ quantize_denoised=quantize_x0,
118
+ mask=mask, x0=x0,
119
+ ddim_use_original_steps=False,
120
+ noise_dropout=noise_dropout,
121
+ temperature=temperature,
122
+ score_corrector=score_corrector,
123
+ corrector_kwargs=corrector_kwargs,
124
+ x_T=x_T,
125
+ log_every_t=log_every_t,
126
+ unconditional_guidance_scale=unconditional_guidance_scale,
127
+ unconditional_conditioning=unconditional_conditioning,
128
+ verbose=verbose,
129
+ **kwargs)
130
+ return samples, intermediates
131
+
132
+ @torch.no_grad()
133
+ def ddim_sampling(self, cond, shape,
134
+ x_T=None, ddim_use_original_steps=False,
135
+ callback=None, timesteps=None, quantize_denoised=False,
136
+ mask=None, x0=None, img_callback=None, log_every_t=100,
137
+ temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
138
+ unconditional_guidance_scale=1., unconditional_conditioning=None, verbose=True,
139
+ cond_tau=1., target_size=None, start_timesteps=None,
140
+ **kwargs):
141
+ device = self.model.betas.device
142
+ print('ddim device', device)
143
+ b = shape[0]
144
+ if x_T is None:
145
+ img = torch.randn(shape, device=device)
146
+ else:
147
+ img = x_T
148
+
149
+ if timesteps is None:
150
+ timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps
151
+ elif timesteps is not None and not ddim_use_original_steps:
152
+ subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1
153
+ timesteps = self.ddim_timesteps[:subset_end]
154
+
155
+ intermediates = {'x_inter': [img], 'pred_x0': [img]}
156
+ time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps)
157
+ total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0]
158
+ if verbose:
159
+ iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps)
160
+ else:
161
+ iterator = time_range
162
+
163
+ init_x0 = False
164
+ clean_cond = kwargs.pop("clean_cond", False)
165
+ for i, step in enumerate(iterator):
166
+ index = total_steps - i - 1
167
+ ts = torch.full((b,), step, device=device, dtype=torch.long)
168
+ if start_timesteps is not None:
169
+ assert x0 is not None
170
+ if step > start_timesteps*time_range[0]:
171
+ continue
172
+ elif not init_x0:
173
+ img = self.model.q_sample(x0, ts)
174
+ init_x0 = True
175
+
176
+ # use mask to blend noised original latent (img_orig) & new sampled latent (img)
177
+ if mask is not None:
178
+ assert x0 is not None
179
+ if clean_cond:
180
+ img_orig = x0
181
+ else:
182
+ img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass? <ddim inversion>
183
+ img = img_orig * mask + (1. - mask) * img # keep original & modify use img
184
+
185
+ index_clip = int((1 - cond_tau) * total_steps)
186
+ if index <= index_clip and target_size is not None:
187
+ target_size_ = [target_size[0], target_size[1]//8, target_size[2]//8]
188
+ img = torch.nn.functional.interpolate(
189
+ img,
190
+ size=target_size_,
191
+ mode="nearest",
192
+ )
193
+ outs = self.p_sample_ddim(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps,
194
+ quantize_denoised=quantize_denoised, temperature=temperature,
195
+ noise_dropout=noise_dropout, score_corrector=score_corrector,
196
+ corrector_kwargs=corrector_kwargs,
197
+ unconditional_guidance_scale=unconditional_guidance_scale,
198
+ unconditional_conditioning=unconditional_conditioning,
199
+ x0=x0,
200
+ step=i,
201
+ **kwargs)
202
+
203
+ img, pred_x0 = outs
204
+ if callback: callback(i)
205
+ if img_callback: img_callback(pred_x0, i)
206
+
207
+ if index % log_every_t == 0 or index == total_steps - 1:
208
+ intermediates['x_inter'].append(img)
209
+ intermediates['pred_x0'].append(pred_x0)
210
+
211
+ return img, intermediates
212
+
213
+ @torch.no_grad()
214
+ def p_sample_ddim(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False,
215
+ temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None,
216
+ unconditional_guidance_scale=1., unconditional_conditioning=None,
217
+ uc_type=None, conditional_guidance_scale_temporal=None, step=0, **kwargs):
218
+ b, *_, device = *x.shape, x.device
219
+ if x.dim() == 5:
220
+ is_video = True
221
+ else:
222
+ is_video = False
223
+ if unconditional_conditioning is None or unconditional_guidance_scale == 1.:
224
+ e_t = self.model.apply_model(x, t, c, **kwargs) # unet denoiser
225
+ else:
226
+ # with unconditional condition
227
+ if step < 5 or step > 15:
228
+ e_t = self.model.apply_model(x, t, c, use_injection=True, **kwargs)
229
+ e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **kwargs)
230
+ elif isinstance(c, torch.Tensor):
231
+ e_t = self.model.apply_model(x, t, c, **kwargs)
232
+ e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **kwargs)
233
+ elif isinstance(c, dict):
234
+ e_t = self.model.apply_model(x, t, c, **kwargs)
235
+ e_t_uncond = self.model.apply_model(x, t, unconditional_conditioning, **kwargs)
236
+ else:
237
+ raise NotImplementedError
238
+ # text cfg
239
+ if uc_type is None:
240
+ e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond)
241
+ else:
242
+ if uc_type == 'cfg_original':
243
+ e_t = e_t + unconditional_guidance_scale * (e_t - e_t_uncond)
244
+ elif uc_type == 'cfg_ours':
245
+ e_t = e_t + unconditional_guidance_scale * (e_t_uncond - e_t)
246
+ else:
247
+ raise NotImplementedError
248
+ # temporal guidance
249
+ if conditional_guidance_scale_temporal is not None:
250
+ e_t_temporal = self.model.apply_model(x, t, c, **kwargs)
251
+ e_t_image = self.model.apply_model(x, t, c, no_temporal_attn=True, **kwargs)
252
+ e_t = e_t + conditional_guidance_scale_temporal * (e_t_temporal - e_t_image)
253
+
254
+ if score_corrector is not None:
255
+ assert self.model.parameterization == "eps"
256
+ e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs)
257
+
258
+ alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
259
+ alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev
260
+ sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas
261
+ sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas
262
+ # select parameters corresponding to the currently considered timestep
263
+
264
+ if is_video:
265
+ size = (b, 1, 1, 1, 1)
266
+ else:
267
+ size = (b, 1, 1, 1)
268
+ a_t = torch.full(size, alphas[index], device=device)
269
+ a_prev = torch.full(size, alphas_prev[index], device=device)
270
+ sigma_t = torch.full(size, sigmas[index], device=device)
271
+ sqrt_one_minus_at = torch.full(size, sqrt_one_minus_alphas[index],device=device)
272
+
273
+ # current prediction for x_0
274
+ pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt()
275
+ if quantize_denoised:
276
+ pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0)
277
+ # direction pointing to x_t
278
+ dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t
279
+
280
+ noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature
281
+ if noise_dropout > 0.:
282
+ noise = torch.nn.functional.dropout(noise, p=noise_dropout)
283
+
284
+ alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas
285
+ if self.use_scale:
286
+ scale_arr = self.model.scale_arr if use_original_steps else self.ddim_scale_arr
287
+ scale_t = torch.full(size, scale_arr[index], device=device)
288
+ scale_arr_prev = self.model.scale_arr_prev if use_original_steps else self.ddim_scale_arr_prev
289
+ scale_t_prev = torch.full(size, scale_arr_prev[index], device=device)
290
+ pred_x0 /= scale_t
291
+ x_prev = a_prev.sqrt() * scale_t_prev * pred_x0 + dir_xt + noise
292
+ else:
293
+ x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise
294
+
295
+ return x_prev, pred_x0
296
+
297
+
298
+ @torch.no_grad()
299
+ def stochastic_encode(self, x0, t, use_original_steps=False, noise=None):
300
+ # fast, but does not allow for exact reconstruction
301
+ # t serves as an index to gather the correct alphas
302
+ if use_original_steps:
303
+ sqrt_alphas_cumprod = self.sqrt_alphas_cumprod
304
+ sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod
305
+ else:
306
+ sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas)
307
+ sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas
308
+
309
+ if noise is None:
310
+ noise = torch.randn_like(x0)
311
+
312
+ def extract_into_tensor(a, t, x_shape):
313
+ b, *_ = t.shape
314
+ out = a.gather(-1, t)
315
+ return out.reshape(b, *((1,) * (len(x_shape) - 1)))
316
+
317
+ return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 +
318
+ extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise)
319
+
320
+ @torch.no_grad()
321
+ def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None,
322
+ use_original_steps=False):
323
+
324
+ timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps
325
+ timesteps = timesteps[:t_start]
326
+
327
+ time_range = np.flip(timesteps)
328
+ total_steps = timesteps.shape[0]
329
+ print(f"Running DDIM Sampling with {total_steps} timesteps")
330
+
331
+ iterator = tqdm(time_range, desc='Decoding image', total=total_steps)
332
+ x_dec = x_latent
333
+ for i, step in enumerate(iterator):
334
+ index = total_steps - i - 1
335
+ ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long)
336
+ x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps,
337
+ unconditional_guidance_scale=unconditional_guidance_scale,
338
+ unconditional_conditioning=unconditional_conditioning)
339
+ return x_dec
340
+
lvdm/models/utils_diffusion.py ADDED
@@ -0,0 +1,104 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import numpy as np
3
+ from einops import repeat
4
+ import torch
5
+ import torch.nn.functional as F
6
+
7
+
8
+ def timestep_embedding(timesteps, dim, max_period=10000, repeat_only=False):
9
+ """
10
+ Create sinusoidal timestep embeddings.
11
+ :param timesteps: a 1-D Tensor of N indices, one per batch element.
12
+ These may be fractional.
13
+ :param dim: the dimension of the output.
14
+ :param max_period: controls the minimum frequency of the embeddings.
15
+ :return: an [N x dim] Tensor of positional embeddings.
16
+ """
17
+ if not repeat_only:
18
+ half = dim // 2
19
+ freqs = torch.exp(
20
+ -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
21
+ ).to(device=timesteps.device)
22
+ args = timesteps[:, None].float() * freqs[None]
23
+ embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
24
+ if dim % 2:
25
+ embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
26
+ else:
27
+ embedding = repeat(timesteps, 'b -> b d', d=dim)
28
+ return embedding
29
+
30
+
31
+ def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3):
32
+ if schedule == "linear":
33
+ betas = (
34
+ torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2
35
+ )
36
+
37
+ elif schedule == "cosine":
38
+ timesteps = (
39
+ torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s
40
+ )
41
+ alphas = timesteps / (1 + cosine_s) * np.pi / 2
42
+ alphas = torch.cos(alphas).pow(2)
43
+ alphas = alphas / alphas[0]
44
+ betas = 1 - alphas[1:] / alphas[:-1]
45
+ betas = np.clip(betas, a_min=0, a_max=0.999)
46
+
47
+ elif schedule == "sqrt_linear":
48
+ betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64)
49
+ elif schedule == "sqrt":
50
+ betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5
51
+ else:
52
+ raise ValueError(f"schedule '{schedule}' unknown.")
53
+ return betas.numpy()
54
+
55
+
56
+ def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True):
57
+ if ddim_discr_method == 'uniform':
58
+ c = num_ddpm_timesteps // num_ddim_timesteps
59
+ ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c)))
60
+ elif ddim_discr_method == 'quad':
61
+ ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int)
62
+ else:
63
+ raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"')
64
+
65
+ # assert ddim_timesteps.shape[0] == num_ddim_timesteps
66
+ # add one to get the final alpha values right (the ones from first scale to data during sampling)
67
+ steps_out = ddim_timesteps + 1
68
+ if verbose:
69
+ print(f'Selected timesteps for ddim sampler: {steps_out}')
70
+ return steps_out
71
+
72
+
73
+ def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True):
74
+ # select alphas for computing the variance schedule
75
+ # print(f'ddim_timesteps={ddim_timesteps}, len_alphacums={len(alphacums)}')
76
+ alphas = alphacums[ddim_timesteps]
77
+ alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
78
+
79
+ # according the the formula provided in https://arxiv.org/abs/2010.02502
80
+ sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev))
81
+ if verbose:
82
+ print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}')
83
+ print(f'For the chosen value of eta, which is {eta}, '
84
+ f'this results in the following sigma_t schedule for ddim sampler {sigmas}')
85
+ return sigmas, alphas, alphas_prev
86
+
87
+
88
+ def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999):
89
+ """
90
+ Create a beta schedule that discretizes the given alpha_t_bar function,
91
+ which defines the cumulative product of (1-beta) over time from t = [0,1].
92
+ :param num_diffusion_timesteps: the number of betas to produce.
93
+ :param alpha_bar: a lambda that takes an argument t from 0 to 1 and
94
+ produces the cumulative product of (1-beta) up to that
95
+ part of the diffusion process.
96
+ :param max_beta: the maximum beta to use; use values lower than 1 to
97
+ prevent singularities.
98
+ """
99
+ betas = []
100
+ for i in range(num_diffusion_timesteps):
101
+ t1 = i / num_diffusion_timesteps
102
+ t2 = (i + 1) / num_diffusion_timesteps
103
+ betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
104
+ return np.array(betas)
lvdm/modules/attention.py ADDED
@@ -0,0 +1,475 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from functools import partial
2
+ import torch
3
+ from torch import nn, einsum
4
+ import torch.nn.functional as F
5
+ from einops import rearrange, repeat
6
+ try:
7
+ import xformers
8
+ import xformers.ops
9
+ XFORMERS_IS_AVAILBLE = True
10
+ except:
11
+ XFORMERS_IS_AVAILBLE = False
12
+ from lvdm.common import (
13
+ checkpoint,
14
+ exists,
15
+ default,
16
+ )
17
+ from lvdm.basics import (
18
+ zero_module,
19
+ )
20
+
21
+ class RelativePosition(nn.Module):
22
+ """ https://github.com/evelinehong/Transformer_Relative_Position_PyTorch/blob/master/relative_position.py """
23
+
24
+ def __init__(self, num_units, max_relative_position):
25
+ super().__init__()
26
+ self.num_units = num_units
27
+ self.max_relative_position = max_relative_position
28
+ self.embeddings_table = nn.Parameter(torch.Tensor(max_relative_position * 2 + 1, num_units))
29
+ nn.init.xavier_uniform_(self.embeddings_table)
30
+
31
+ def forward(self, length_q, length_k):
32
+ device = self.embeddings_table.device
33
+ range_vec_q = torch.arange(length_q, device=device)
34
+ range_vec_k = torch.arange(length_k, device=device)
35
+ distance_mat = range_vec_k[None, :] - range_vec_q[:, None]
36
+ distance_mat_clipped = torch.clamp(distance_mat, -self.max_relative_position, self.max_relative_position)
37
+ final_mat = distance_mat_clipped + self.max_relative_position
38
+ final_mat = final_mat.long()
39
+ embeddings = self.embeddings_table[final_mat]
40
+ return embeddings
41
+
42
+
43
+ class CrossAttention(nn.Module):
44
+
45
+ def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.,
46
+ relative_position=False, temporal_length=None, img_cross_attention=False):
47
+ super().__init__()
48
+ inner_dim = dim_head * heads
49
+ context_dim = default(context_dim, query_dim)
50
+
51
+ self.scale = dim_head**-0.5
52
+ self.heads = heads
53
+ self.dim_head = dim_head
54
+ self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
55
+ self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
56
+ self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
57
+ self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout))
58
+
59
+ self.image_cross_attention_scale = 1.0
60
+ self.text_context_len = 77
61
+ self.img_cross_attention = img_cross_attention
62
+ if self.img_cross_attention:
63
+ self.to_k_ip = nn.Linear(context_dim, inner_dim, bias=False)
64
+ self.to_v_ip = nn.Linear(context_dim, inner_dim, bias=False)
65
+
66
+ self.relative_position = relative_position
67
+ if self.relative_position:
68
+ assert(temporal_length is not None)
69
+ self.relative_position_k = RelativePosition(num_units=dim_head, max_relative_position=temporal_length)
70
+ self.relative_position_v = RelativePosition(num_units=dim_head, max_relative_position=temporal_length)
71
+ else:
72
+ ## only used for spatial attention, while NOT for temporal attention
73
+ if XFORMERS_IS_AVAILBLE and temporal_length is None:
74
+ self.forward = self.efficient_forward
75
+
76
+ def forward(self, x, context=None, mask=None):
77
+ h = self.heads
78
+
79
+ q = self.to_q(x)
80
+ context = default(context, x)
81
+ ## considering image token additionally
82
+ if context is not None and self.img_cross_attention:
83
+ context, context_img = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:]
84
+ k = self.to_k(context)
85
+ v = self.to_v(context)
86
+ k_ip = self.to_k_ip(context_img)
87
+ v_ip = self.to_v_ip(context_img)
88
+ else:
89
+ k = self.to_k(context)
90
+ v = self.to_v(context)
91
+
92
+ q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
93
+ sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale
94
+ if self.relative_position:
95
+ len_q, len_k, len_v = q.shape[1], k.shape[1], v.shape[1]
96
+ k2 = self.relative_position_k(len_q, len_k)
97
+ sim2 = einsum('b t d, t s d -> b t s', q, k2) * self.scale # TODO check
98
+ sim += sim2
99
+ del k
100
+
101
+ if exists(mask):
102
+ ## feasible for causal attention mask only
103
+ max_neg_value = -torch.finfo(sim.dtype).max
104
+ mask = repeat(mask, 'b i j -> (b h) i j', h=h)
105
+ sim.masked_fill_(~(mask>0.5), max_neg_value)
106
+
107
+ # attention, what we cannot get enough of
108
+ sim = sim.softmax(dim=-1)
109
+ out = torch.einsum('b i j, b j d -> b i d', sim, v)
110
+ if self.relative_position:
111
+ v2 = self.relative_position_v(len_q, len_v)
112
+ out2 = einsum('b t s, t s d -> b t d', sim, v2) # TODO check
113
+ out += out2
114
+ out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
115
+
116
+ ## considering image token additionally
117
+ if context is not None and self.img_cross_attention:
118
+ k_ip, v_ip = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (k_ip, v_ip))
119
+ sim_ip = torch.einsum('b i d, b j d -> b i j', q, k_ip) * self.scale
120
+ del k_ip
121
+ sim_ip = sim_ip.softmax(dim=-1)
122
+ out_ip = torch.einsum('b i j, b j d -> b i d', sim_ip, v_ip)
123
+ out_ip = rearrange(out_ip, '(b h) n d -> b n (h d)', h=h)
124
+ out = out + self.image_cross_attention_scale * out_ip
125
+ del q
126
+
127
+ return self.to_out(out)
128
+
129
+ def efficient_forward(self, x, context=None, mask=None):
130
+ q = self.to_q(x)
131
+ context = default(context, x)
132
+
133
+ ## considering image token additionally
134
+ if context is not None and self.img_cross_attention:
135
+ context, context_img = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:]
136
+ k = self.to_k(context)
137
+ v = self.to_v(context)
138
+ k_ip = self.to_k_ip(context_img)
139
+ v_ip = self.to_v_ip(context_img)
140
+ else:
141
+ k = self.to_k(context)
142
+ v = self.to_v(context)
143
+
144
+ b, _, _ = q.shape
145
+ q, k, v = map(
146
+ lambda t: t.unsqueeze(3)
147
+ .reshape(b, t.shape[1], self.heads, self.dim_head)
148
+ .permute(0, 2, 1, 3)
149
+ .reshape(b * self.heads, t.shape[1], self.dim_head)
150
+ .contiguous(),
151
+ (q, k, v),
152
+ )
153
+ # actually compute the attention, what we cannot get enough of
154
+ out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=None)
155
+
156
+ ## considering image token additionally
157
+ if context is not None and self.img_cross_attention:
158
+ k_ip, v_ip = map(
159
+ lambda t: t.unsqueeze(3)
160
+ .reshape(b, t.shape[1], self.heads, self.dim_head)
161
+ .permute(0, 2, 1, 3)
162
+ .reshape(b * self.heads, t.shape[1], self.dim_head)
163
+ .contiguous(),
164
+ (k_ip, v_ip),
165
+ )
166
+ out_ip = xformers.ops.memory_efficient_attention(q, k_ip, v_ip, attn_bias=None, op=None)
167
+ out_ip = (
168
+ out_ip.unsqueeze(0)
169
+ .reshape(b, self.heads, out.shape[1], self.dim_head)
170
+ .permute(0, 2, 1, 3)
171
+ .reshape(b, out.shape[1], self.heads * self.dim_head)
172
+ )
173
+
174
+ if exists(mask):
175
+ raise NotImplementedError
176
+ out = (
177
+ out.unsqueeze(0)
178
+ .reshape(b, self.heads, out.shape[1], self.dim_head)
179
+ .permute(0, 2, 1, 3)
180
+ .reshape(b, out.shape[1], self.heads * self.dim_head)
181
+ )
182
+ if context is not None and self.img_cross_attention:
183
+ out = out + self.image_cross_attention_scale * out_ip
184
+ return self.to_out(out)
185
+
186
+
187
+ class BasicTransformerBlock(nn.Module):
188
+
189
+ def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True,
190
+ disable_self_attn=False, attention_cls=None, img_cross_attention=False):
191
+ super().__init__()
192
+ attn_cls = CrossAttention if attention_cls is None else attention_cls
193
+ self.disable_self_attn = disable_self_attn
194
+ self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout,
195
+ context_dim=context_dim if self.disable_self_attn else None)
196
+ self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
197
+ self.attn2 = attn_cls(query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, dropout=dropout,
198
+ img_cross_attention=img_cross_attention)
199
+ self.norm1 = nn.LayerNorm(dim)
200
+ self.norm2 = nn.LayerNorm(dim)
201
+ self.norm3 = nn.LayerNorm(dim)
202
+ self.checkpoint = checkpoint
203
+
204
+ def forward(self, x, context=None, mask=None):
205
+ ## implementation tricks: because checkpointing doesn't support non-tensor (e.g. None or scalar) arguments
206
+ input_tuple = (x,) ## should not be (x), otherwise *input_tuple will decouple x into multiple arguments
207
+ if context is not None:
208
+ input_tuple = (x, context)
209
+ if mask is not None:
210
+ forward_mask = partial(self._forward, mask=mask)
211
+ return checkpoint(forward_mask, (x,), self.parameters(), self.checkpoint)
212
+ if context is not None and mask is not None:
213
+ input_tuple = (x, context, mask)
214
+ return checkpoint(self._forward, input_tuple, self.parameters(), self.checkpoint)
215
+
216
+ def _forward(self, x, context=None, mask=None):
217
+ x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None, mask=mask) + x
218
+ x = self.attn2(self.norm2(x), context=context, mask=mask) + x
219
+ x = self.ff(self.norm3(x)) + x
220
+ return x
221
+
222
+
223
+ class SpatialTransformer(nn.Module):
224
+ """
225
+ Transformer block for image-like data in spatial axis.
226
+ First, project the input (aka embedding)
227
+ and reshape to b, t, d.
228
+ Then apply standard transformer action.
229
+ Finally, reshape to image
230
+ NEW: use_linear for more efficiency instead of the 1x1 convs
231
+ """
232
+
233
+ def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None,
234
+ use_checkpoint=True, disable_self_attn=False, use_linear=False, img_cross_attention=False):
235
+ super().__init__()
236
+ self.in_channels = in_channels
237
+ inner_dim = n_heads * d_head
238
+ self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
239
+ if not use_linear:
240
+ self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
241
+ else:
242
+ self.proj_in = nn.Linear(in_channels, inner_dim)
243
+
244
+ self.transformer_blocks = nn.ModuleList([
245
+ BasicTransformerBlock(
246
+ inner_dim,
247
+ n_heads,
248
+ d_head,
249
+ dropout=dropout,
250
+ context_dim=context_dim,
251
+ img_cross_attention=img_cross_attention,
252
+ disable_self_attn=disable_self_attn,
253
+ checkpoint=use_checkpoint) for d in range(depth)
254
+ ])
255
+ if not use_linear:
256
+ self.proj_out = zero_module(nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0))
257
+ else:
258
+ self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
259
+ self.use_linear = use_linear
260
+
261
+
262
+ def forward(self, x, context=None):
263
+ b, c, h, w = x.shape
264
+ x_in = x
265
+ x = self.norm(x)
266
+ if not self.use_linear:
267
+ x = self.proj_in(x)
268
+ x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
269
+ if self.use_linear:
270
+ x = self.proj_in(x)
271
+ for i, block in enumerate(self.transformer_blocks):
272
+ x = block(x, context=context)
273
+ if self.use_linear:
274
+ x = self.proj_out(x)
275
+ x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
276
+ if not self.use_linear:
277
+ x = self.proj_out(x)
278
+ return x + x_in
279
+
280
+
281
+ class TemporalTransformer(nn.Module):
282
+ """
283
+ Transformer block for image-like data in temporal axis.
284
+ First, reshape to b, t, d.
285
+ Then apply standard transformer action.
286
+ Finally, reshape to image
287
+ """
288
+ def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None,
289
+ use_checkpoint=True, use_linear=False, only_self_att=True, causal_attention=False,
290
+ relative_position=False, temporal_length=None):
291
+ super().__init__()
292
+ self.only_self_att = only_self_att
293
+ self.relative_position = relative_position
294
+ self.causal_attention = causal_attention
295
+ self.in_channels = in_channels
296
+ inner_dim = n_heads * d_head
297
+ self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
298
+ self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
299
+ if not use_linear:
300
+ self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
301
+ else:
302
+ self.proj_in = nn.Linear(in_channels, inner_dim)
303
+
304
+ if relative_position:
305
+ assert(temporal_length is not None)
306
+ attention_cls = partial(CrossAttention, relative_position=True, temporal_length=temporal_length)
307
+ else:
308
+ attention_cls = None
309
+ if self.causal_attention:
310
+ assert(temporal_length is not None)
311
+ self.mask = torch.tril(torch.ones([1, temporal_length, temporal_length]))
312
+
313
+ if self.only_self_att:
314
+ context_dim = None
315
+ self.transformer_blocks = nn.ModuleList([
316
+ BasicTransformerBlock(
317
+ inner_dim,
318
+ n_heads,
319
+ d_head,
320
+ dropout=dropout,
321
+ context_dim=context_dim,
322
+ attention_cls=attention_cls,
323
+ checkpoint=use_checkpoint) for d in range(depth)
324
+ ])
325
+ if not use_linear:
326
+ self.proj_out = zero_module(nn.Conv1d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0))
327
+ else:
328
+ self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
329
+ self.use_linear = use_linear
330
+
331
+ def forward(self, x, context=None):
332
+ b, c, t, h, w = x.shape
333
+ x_in = x
334
+ x = self.norm(x)
335
+ x = rearrange(x, 'b c t h w -> (b h w) c t').contiguous()
336
+ if not self.use_linear:
337
+ x = self.proj_in(x)
338
+ x = rearrange(x, 'bhw c t -> bhw t c').contiguous()
339
+ if self.use_linear:
340
+ x = self.proj_in(x)
341
+
342
+ if self.causal_attention:
343
+ mask = self.mask.to(x.device)
344
+ mask = repeat(mask, 'l i j -> (l bhw) i j', bhw=b*h*w)
345
+ else:
346
+ mask = None
347
+
348
+ if self.only_self_att:
349
+ ## note: if no context is given, cross-attention defaults to self-attention
350
+ for i, block in enumerate(self.transformer_blocks):
351
+ x = block(x, mask=mask)
352
+ x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous()
353
+ else:
354
+ x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous()
355
+ context = rearrange(context, '(b t) l con -> b t l con', t=t).contiguous()
356
+ for i, block in enumerate(self.transformer_blocks):
357
+ # calculate each batch one by one (since number in shape could not greater then 65,535 for some package)
358
+ for j in range(b):
359
+ context_j = repeat(
360
+ context[j],
361
+ 't l con -> (t r) l con', r=(h * w) // t, t=t).contiguous()
362
+ ## note: causal mask will not applied in cross-attention case
363
+ x[j] = block(x[j], context=context_j)
364
+
365
+ if self.use_linear:
366
+ x = self.proj_out(x)
367
+ x = rearrange(x, 'b (h w) t c -> b c t h w', h=h, w=w).contiguous()
368
+ if not self.use_linear:
369
+ x = rearrange(x, 'b hw t c -> (b hw) c t').contiguous()
370
+ x = self.proj_out(x)
371
+ x = rearrange(x, '(b h w) c t -> b c t h w', b=b, h=h, w=w).contiguous()
372
+
373
+ return x + x_in
374
+
375
+
376
+ class GEGLU(nn.Module):
377
+ def __init__(self, dim_in, dim_out):
378
+ super().__init__()
379
+ self.proj = nn.Linear(dim_in, dim_out * 2)
380
+
381
+ def forward(self, x):
382
+ x, gate = self.proj(x).chunk(2, dim=-1)
383
+ return x * F.gelu(gate)
384
+
385
+
386
+ class FeedForward(nn.Module):
387
+ def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
388
+ super().__init__()
389
+ inner_dim = int(dim * mult)
390
+ dim_out = default(dim_out, dim)
391
+ project_in = nn.Sequential(
392
+ nn.Linear(dim, inner_dim),
393
+ nn.GELU()
394
+ ) if not glu else GEGLU(dim, inner_dim)
395
+
396
+ self.net = nn.Sequential(
397
+ project_in,
398
+ nn.Dropout(dropout),
399
+ nn.Linear(inner_dim, dim_out)
400
+ )
401
+
402
+ def forward(self, x):
403
+ return self.net(x)
404
+
405
+
406
+ class LinearAttention(nn.Module):
407
+ def __init__(self, dim, heads=4, dim_head=32):
408
+ super().__init__()
409
+ self.heads = heads
410
+ hidden_dim = dim_head * heads
411
+ self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
412
+ self.to_out = nn.Conv2d(hidden_dim, dim, 1)
413
+
414
+ def forward(self, x):
415
+ b, c, h, w = x.shape
416
+ qkv = self.to_qkv(x)
417
+ q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
418
+ k = k.softmax(dim=-1)
419
+ context = torch.einsum('bhdn,bhen->bhde', k, v)
420
+ out = torch.einsum('bhde,bhdn->bhen', context, q)
421
+ out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
422
+ return self.to_out(out)
423
+
424
+
425
+ class SpatialSelfAttention(nn.Module):
426
+ def __init__(self, in_channels):
427
+ super().__init__()
428
+ self.in_channels = in_channels
429
+
430
+ self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
431
+ self.q = torch.nn.Conv2d(in_channels,
432
+ in_channels,
433
+ kernel_size=1,
434
+ stride=1,
435
+ padding=0)
436
+ self.k = torch.nn.Conv2d(in_channels,
437
+ in_channels,
438
+ kernel_size=1,
439
+ stride=1,
440
+ padding=0)
441
+ self.v = torch.nn.Conv2d(in_channels,
442
+ in_channels,
443
+ kernel_size=1,
444
+ stride=1,
445
+ padding=0)
446
+ self.proj_out = torch.nn.Conv2d(in_channels,
447
+ in_channels,
448
+ kernel_size=1,
449
+ stride=1,
450
+ padding=0)
451
+
452
+ def forward(self, x):
453
+ h_ = x
454
+ h_ = self.norm(h_)
455
+ q = self.q(h_)
456
+ k = self.k(h_)
457
+ v = self.v(h_)
458
+
459
+ # compute attention
460
+ b,c,h,w = q.shape
461
+ q = rearrange(q, 'b c h w -> b (h w) c')
462
+ k = rearrange(k, 'b c h w -> b c (h w)')
463
+ w_ = torch.einsum('bij,bjk->bik', q, k)
464
+
465
+ w_ = w_ * (int(c)**(-0.5))
466
+ w_ = torch.nn.functional.softmax(w_, dim=2)
467
+
468
+ # attend to values
469
+ v = rearrange(v, 'b c h w -> b c (h w)')
470
+ w_ = rearrange(w_, 'b i j -> b j i')
471
+ h_ = torch.einsum('bij,bjk->bik', v, w_)
472
+ h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
473
+ h_ = self.proj_out(h_)
474
+
475
+ return x+h_
lvdm/modules/attention_freenoise.py ADDED
@@ -0,0 +1,565 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from functools import partial
2
+ import torch
3
+ from torch import nn, einsum
4
+ import torch.nn.functional as F
5
+ from einops import rearrange, repeat
6
+ try:
7
+ import xformers
8
+ import xformers.ops
9
+ XFORMERS_IS_AVAILBLE = True
10
+ except:
11
+ XFORMERS_IS_AVAILBLE = False
12
+ from lvdm.common import (
13
+ checkpoint,
14
+ exists,
15
+ default,
16
+ )
17
+ from lvdm.basics import (
18
+ zero_module,
19
+ )
20
+
21
+ def generate_weight_sequence(n):
22
+ if n % 2 == 0:
23
+ max_weight = n // 2
24
+ weight_sequence = list(range(1, max_weight + 1, 1)) + list(range(max_weight, 0, -1))
25
+ else:
26
+ max_weight = (n + 1) // 2
27
+ weight_sequence = list(range(1, max_weight, 1)) + [max_weight] + list(range(max_weight - 1, 0, -1))
28
+ return weight_sequence
29
+
30
+ class RelativePosition(nn.Module):
31
+ """ https://github.com/evelinehong/Transformer_Relative_Position_PyTorch/blob/master/relative_position.py """
32
+
33
+ def __init__(self, num_units, max_relative_position):
34
+ super().__init__()
35
+ self.num_units = num_units
36
+ self.max_relative_position = max_relative_position
37
+ self.embeddings_table = nn.Parameter(torch.Tensor(max_relative_position * 2 + 1, num_units))
38
+ nn.init.xavier_uniform_(self.embeddings_table)
39
+
40
+ def forward(self, length_q, length_k):
41
+ device = self.embeddings_table.device
42
+ range_vec_q = torch.arange(length_q, device=device)
43
+ range_vec_k = torch.arange(length_k, device=device)
44
+ distance_mat = range_vec_k[None, :] - range_vec_q[:, None]
45
+ distance_mat_clipped = torch.clamp(distance_mat, -self.max_relative_position, self.max_relative_position)
46
+ final_mat = distance_mat_clipped + self.max_relative_position
47
+ final_mat = final_mat.long()
48
+ embeddings = self.embeddings_table[final_mat]
49
+ return embeddings
50
+
51
+
52
+ class CrossAttention(nn.Module):
53
+
54
+ def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.,
55
+ relative_position=False, temporal_length=None, img_cross_attention=False, injection=False):
56
+ super().__init__()
57
+ inner_dim = dim_head * heads
58
+ context_dim = default(context_dim, query_dim)
59
+
60
+ self.scale = dim_head**-0.5
61
+ self.heads = heads
62
+ self.dim_head = dim_head
63
+ self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
64
+ self.to_k = nn.Linear(context_dim, inner_dim, bias=False)
65
+ self.to_v = nn.Linear(context_dim, inner_dim, bias=False)
66
+ self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout))
67
+
68
+ self.image_cross_attention_scale = 1.0
69
+ self.text_context_len = 77
70
+ self.img_cross_attention = img_cross_attention
71
+ if self.img_cross_attention:
72
+ self.to_k_ip = nn.Linear(context_dim, inner_dim, bias=False)
73
+ self.to_v_ip = nn.Linear(context_dim, inner_dim, bias=False)
74
+
75
+ self.relative_position = relative_position
76
+ if self.relative_position:
77
+ assert(temporal_length is not None)
78
+ self.relative_position_k = RelativePosition(num_units=dim_head, max_relative_position=temporal_length)
79
+ self.relative_position_v = RelativePosition(num_units=dim_head, max_relative_position=temporal_length)
80
+ else:
81
+ ## only used for spatial attention, while NOT for temporal attention
82
+ if XFORMERS_IS_AVAILBLE and temporal_length is None:
83
+ self.forward = self.efficient_forward
84
+
85
+ self.injection = injection
86
+
87
+ def forward(self, x, context=None, mask=None, context_next=None, use_injection=False):
88
+
89
+ sa_flag = False
90
+ if context is None:
91
+ sa_flag = True
92
+
93
+ h = self.heads
94
+
95
+ all_q = self.to_q(x)
96
+ context = default(context, x)
97
+ ## considering image token additionally
98
+ if context is not None and self.img_cross_attention:
99
+ context, context_img = context[:,:self.text_context_len,:], context[:,self.text_context_len:,:]
100
+ all_k = self.to_k(context)
101
+ all_v = self.to_v(context)
102
+ all_k_ip = self.to_k_ip(context_img)
103
+ all_v_ip = self.to_v_ip(context_img)
104
+ else:
105
+ all_k = self.to_k(context)
106
+ all_v = self.to_v(context)
107
+
108
+ count = torch.zeros_like(all_k)
109
+ value = torch.zeros_like(all_k)
110
+
111
+ if (sa_flag) and (context_next is not None):
112
+ all_q, all_k, all_v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (all_q, all_k, all_v))
113
+ if context is not None and self.img_cross_attention:
114
+ all_k_ip, all_v_ip = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (all_k_ip, all_v_ip))
115
+ for t_start, t_end in context_next:
116
+ weight_sequence = generate_weight_sequence(t_end - t_start)
117
+ weight_tensor = torch.ones_like(count[:, t_start:t_end])
118
+ weight_tensor = weight_tensor * torch.Tensor(weight_sequence).to(x.device).unsqueeze(0).unsqueeze(-1)
119
+
120
+ q = all_q[:, t_start:t_end]
121
+ k = all_k[:, t_start:t_end]
122
+ v = all_v[:, t_start:t_end]
123
+
124
+ sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale
125
+ if self.relative_position:
126
+ len_q, len_k, len_v = q.shape[1], k.shape[1], v.shape[1]
127
+ k2 = self.relative_position_k(len_q, len_k)
128
+ sim2 = einsum('b t d, t s d -> b t s', q, k2) * self.scale # TODO check
129
+ sim += sim2
130
+ del k
131
+
132
+ if exists(mask):
133
+ ## feasible for causal attention mask only
134
+ max_neg_value = -torch.finfo(sim.dtype).max
135
+ mask = repeat(mask, 'b i j -> (b h) i j', h=h)
136
+ sim.masked_fill_(~(mask>0.5), max_neg_value)
137
+
138
+ # attention, what we cannot get enough of
139
+ sim = sim.softmax(dim=-1)
140
+ out = torch.einsum('b i j, b j d -> b i d', sim, v)
141
+ if self.relative_position:
142
+ v2 = self.relative_position_v(len_q, len_v)
143
+ out2 = einsum('b t s, t s d -> b t d', sim, v2) # TODO check
144
+ out += out2
145
+ out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
146
+
147
+ ## considering image token additionally
148
+ if context is not None and self.img_cross_attention:
149
+ k_ip = all_k_ip[:, t_start:t_end]
150
+ v_ip = all_v_ip[:, t_start:t_end]
151
+ sim_ip = torch.einsum('b i d, b j d -> b i j', q, k_ip) * self.scale
152
+ del k_ip
153
+ sim_ip = sim_ip.softmax(dim=-1)
154
+ out_ip = torch.einsum('b i j, b j d -> b i d', sim_ip, v_ip)
155
+ out_ip = rearrange(out_ip, '(b h) n d -> b n (h d)', h=h)
156
+ out = out + self.image_cross_attention_scale * out_ip
157
+ del q
158
+
159
+ value[:,t_start:t_end] += out * weight_tensor
160
+ count[:,t_start:t_end] += weight_tensor
161
+
162
+ final_out = torch.where(count>0, value/count, value)
163
+
164
+ else:
165
+ q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (all_q, all_k, all_v))
166
+ sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale
167
+ if self.relative_position:
168
+ len_q, len_k, len_v = q.shape[1], k.shape[1], v.shape[1]
169
+ k2 = self.relative_position_k(len_q, len_k)
170
+ sim2 = einsum('b t d, t s d -> b t s', q, k2) * self.scale # TODO check
171
+ sim += sim2
172
+ del k
173
+
174
+ if exists(mask):
175
+ ## feasible for causal attention mask only
176
+ max_neg_value = -torch.finfo(sim.dtype).max
177
+ mask = repeat(mask, 'b i j -> (b h) i j', h=h)
178
+ sim.masked_fill_(~(mask>0.5), max_neg_value)
179
+
180
+ # attention, what we cannot get enough of
181
+ sim = sim.softmax(dim=-1)
182
+ out = torch.einsum('b i j, b j d -> b i d', sim, v)
183
+ if self.relative_position:
184
+ v2 = self.relative_position_v(len_q, len_v)
185
+ out2 = einsum('b t s, t s d -> b t d', sim, v2) # TODO check
186
+ out += out2
187
+ final_out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
188
+
189
+ ## considering image token additionally
190
+ if context is not None and self.img_cross_attention:
191
+ k_ip, v_ip = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (all_k_ip, all_v_ip))
192
+ sim_ip = torch.einsum('b i d, b j d -> b i j', q, k_ip) * self.scale
193
+ del k_ip
194
+ sim_ip = sim_ip.softmax(dim=-1)
195
+ out_ip = torch.einsum('b i j, b j d -> b i d', sim_ip, v_ip)
196
+ out_ip = rearrange(out_ip, '(b h) n d -> b n (h d)', h=h)
197
+ final_out = final_out + self.image_cross_attention_scale * out_ip
198
+ del q
199
+
200
+ return self.to_out(final_out)
201
+
202
+ def efficient_forward(self, x, context=None, mask=None, context_next=None, use_injection=False):
203
+
204
+ sa_flag = False
205
+ if context is None:
206
+ sa_flag = True
207
+
208
+ q = self.to_q(x)
209
+ context = default(context, x)
210
+
211
+ if not sa_flag:
212
+ sq_size = x.shape[0]
213
+ if self.injection and use_injection:
214
+ context_new = context[-sq_size:]
215
+ else:
216
+ context_new = context[:sq_size]
217
+ else:
218
+ context_new = context.clone()
219
+
220
+ ## considering image token additionally
221
+ if context is not None and self.img_cross_attention:
222
+ context, context_img = context_new[:,:self.text_context_len,:], context_new[:,self.text_context_len:,:]
223
+ k = self.to_k(context)
224
+ v = self.to_v(context)
225
+ k_ip = self.to_k_ip(context_img)
226
+ v_ip = self.to_v_ip(context_img)
227
+ else:
228
+ k = self.to_k(context_new)
229
+ v = self.to_v(context_new)
230
+
231
+ b, _, _ = q.shape
232
+ q, k, v = map(
233
+ lambda t: t.unsqueeze(3)
234
+ .reshape(b, t.shape[1], self.heads, self.dim_head)
235
+ .permute(0, 2, 1, 3)
236
+ .reshape(b * self.heads, t.shape[1], self.dim_head)
237
+ .contiguous(),
238
+ (q, k, v),
239
+ )
240
+ # actually compute the attention, what we cannot get enough of
241
+ out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=None)
242
+
243
+ ## considering image token additionally
244
+ if context is not None and self.img_cross_attention:
245
+ k_ip, v_ip = map(
246
+ lambda t: t.unsqueeze(3)
247
+ .reshape(b, t.shape[1], self.heads, self.dim_head)
248
+ .permute(0, 2, 1, 3)
249
+ .reshape(b * self.heads, t.shape[1], self.dim_head)
250
+ .contiguous(),
251
+ (k_ip, v_ip),
252
+ )
253
+ out_ip = xformers.ops.memory_efficient_attention(q, k_ip, v_ip, attn_bias=None, op=None)
254
+ out_ip = (
255
+ out_ip.unsqueeze(0)
256
+ .reshape(b, self.heads, out.shape[1], self.dim_head)
257
+ .permute(0, 2, 1, 3)
258
+ .reshape(b, out.shape[1], self.heads * self.dim_head)
259
+ )
260
+
261
+ if exists(mask):
262
+ raise NotImplementedError
263
+ out = (
264
+ out.unsqueeze(0)
265
+ .reshape(b, self.heads, out.shape[1], self.dim_head)
266
+ .permute(0, 2, 1, 3)
267
+ .reshape(b, out.shape[1], self.heads * self.dim_head)
268
+ )
269
+ if context is not None and self.img_cross_attention:
270
+ out = out + self.image_cross_attention_scale * out_ip
271
+ return self.to_out(out)
272
+
273
+
274
+ class BasicTransformerBlock(nn.Module):
275
+
276
+ def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True,
277
+ disable_self_attn=False, attention_cls=None, img_cross_attention=False, injection=False):
278
+ super().__init__()
279
+ attn_cls = CrossAttention if attention_cls is None else attention_cls
280
+ self.disable_self_attn = disable_self_attn
281
+ self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout,
282
+ context_dim=context_dim if self.disable_self_attn else None, injection=injection)
283
+ self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff)
284
+ self.attn2 = attn_cls(query_dim=dim, context_dim=context_dim, heads=n_heads, dim_head=d_head, dropout=dropout,
285
+ img_cross_attention=img_cross_attention, injection=injection)
286
+ self.norm1 = nn.LayerNorm(dim)
287
+ self.norm2 = nn.LayerNorm(dim)
288
+ self.norm3 = nn.LayerNorm(dim)
289
+ self.checkpoint = checkpoint
290
+
291
+ def forward(self, x, context=None, mask=None, context_next=None, use_injection=False, **kwargs):
292
+ ## implementation tricks: because checkpointing doesn't support non-tensor (e.g. None or scalar) arguments
293
+ input_tuple = (x,) ## should not be (x), otherwise *input_tuple will decouple x into multiple arguments
294
+ if context is not None:
295
+ input_tuple = (x, context)
296
+ if mask is not None:
297
+ forward_mask = partial(self._forward, mask=mask)
298
+ return checkpoint(forward_mask, (x,), self.parameters(), self.checkpoint)
299
+ if context is not None and mask is not None:
300
+ input_tuple = (x, context, mask)
301
+ input_tuple = (x, context, mask, context_next, use_injection)
302
+ return checkpoint(self._forward, input_tuple, self.parameters(), self.checkpoint)
303
+
304
+ def _forward(self, x, context=None, mask=None, context_next=None, use_injection=False):
305
+ x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None, mask=mask, context_next=context_next, use_injection=False) + x
306
+ x = self.attn2(self.norm2(x), context=context, mask=mask, context_next=context_next, use_injection=use_injection) + x
307
+ x = self.ff(self.norm3(x)) + x
308
+ return x
309
+
310
+
311
+ class SpatialTransformer(nn.Module):
312
+ """
313
+ Transformer block for image-like data in spatial axis.
314
+ First, project the input (aka embedding)
315
+ and reshape to b, t, d.
316
+ Then apply standard transformer action.
317
+ Finally, reshape to image
318
+ NEW: use_linear for more efficiency instead of the 1x1 convs
319
+ """
320
+
321
+ def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None,
322
+ use_checkpoint=True, disable_self_attn=False, use_linear=False, img_cross_attention=False, injection=False):
323
+ super().__init__()
324
+ self.in_channels = in_channels
325
+ inner_dim = n_heads * d_head
326
+ self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
327
+ if not use_linear:
328
+ self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
329
+ else:
330
+ self.proj_in = nn.Linear(in_channels, inner_dim)
331
+
332
+ self.transformer_blocks = nn.ModuleList([
333
+ BasicTransformerBlock(
334
+ inner_dim,
335
+ n_heads,
336
+ d_head,
337
+ dropout=dropout,
338
+ context_dim=context_dim,
339
+ img_cross_attention=img_cross_attention,
340
+ disable_self_attn=disable_self_attn,
341
+ checkpoint=use_checkpoint,
342
+ injection=injection) for d in range(depth)
343
+ ])
344
+ if not use_linear:
345
+ self.proj_out = zero_module(nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0))
346
+ else:
347
+ self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
348
+ self.use_linear = use_linear
349
+
350
+
351
+ def forward(self, x, context=None, **kwargs):
352
+ b, c, h, w = x.shape
353
+ x_in = x
354
+ x = self.norm(x)
355
+ if not self.use_linear:
356
+ x = self.proj_in(x)
357
+ x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
358
+ if self.use_linear:
359
+ x = self.proj_in(x)
360
+ for i, block in enumerate(self.transformer_blocks):
361
+ x = block(x, context=context, **kwargs)
362
+ if self.use_linear:
363
+ x = self.proj_out(x)
364
+ x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
365
+ if not self.use_linear:
366
+ x = self.proj_out(x)
367
+ return x + x_in
368
+
369
+
370
+ class TemporalTransformer(nn.Module):
371
+ """
372
+ Transformer block for image-like data in temporal axis.
373
+ First, reshape to b, t, d.
374
+ Then apply standard transformer action.
375
+ Finally, reshape to image
376
+ """
377
+ def __init__(self, in_channels, n_heads, d_head, depth=1, dropout=0., context_dim=None,
378
+ use_checkpoint=True, use_linear=False, only_self_att=True, causal_attention=False,
379
+ relative_position=False, temporal_length=None, injection=False):
380
+ super().__init__()
381
+ self.only_self_att = only_self_att
382
+ self.relative_position = relative_position
383
+ self.causal_attention = causal_attention
384
+ self.in_channels = in_channels
385
+ inner_dim = n_heads * d_head
386
+ self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
387
+ self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
388
+ if not use_linear:
389
+ self.proj_in = nn.Conv1d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
390
+ else:
391
+ self.proj_in = nn.Linear(in_channels, inner_dim)
392
+
393
+ if relative_position:
394
+ assert(temporal_length is not None)
395
+ attention_cls = partial(CrossAttention, relative_position=True, temporal_length=temporal_length)
396
+ else:
397
+ attention_cls = partial(CrossAttention, temporal_length=temporal_length)
398
+ if self.causal_attention:
399
+ assert(temporal_length is not None)
400
+ self.mask = torch.tril(torch.ones([1, temporal_length, temporal_length]))
401
+
402
+ if self.only_self_att:
403
+ context_dim = None
404
+ self.transformer_blocks = nn.ModuleList([
405
+ BasicTransformerBlock(
406
+ inner_dim,
407
+ n_heads,
408
+ d_head,
409
+ dropout=dropout,
410
+ context_dim=context_dim,
411
+ attention_cls=attention_cls,
412
+ checkpoint=use_checkpoint,
413
+ injection=injection) for d in range(depth)
414
+ ])
415
+ if not use_linear:
416
+ self.proj_out = zero_module(nn.Conv1d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0))
417
+ else:
418
+ self.proj_out = zero_module(nn.Linear(inner_dim, in_channels))
419
+ self.use_linear = use_linear
420
+
421
+ def forward(self, x, context=None, **kwargs):
422
+ b, c, t, h, w = x.shape
423
+ x_in = x
424
+ x = self.norm(x)
425
+ x = rearrange(x, 'b c t h w -> (b h w) c t').contiguous()
426
+ if not self.use_linear:
427
+ x = self.proj_in(x)
428
+ x = rearrange(x, 'bhw c t -> bhw t c').contiguous()
429
+ if self.use_linear:
430
+ x = self.proj_in(x)
431
+
432
+ if self.causal_attention:
433
+ mask = self.mask.to(x.device)
434
+ mask = repeat(mask, 'l i j -> (l bhw) i j', bhw=b*h*w)
435
+ else:
436
+ mask = None
437
+
438
+ if self.only_self_att:
439
+ ## note: if no context is given, cross-attention defaults to self-attention
440
+ for i, block in enumerate(self.transformer_blocks):
441
+ x = block(x, mask=mask, **kwargs)
442
+ x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous()
443
+ else:
444
+ x = rearrange(x, '(b hw) t c -> b hw t c', b=b).contiguous()
445
+ context = rearrange(context, '(b t) l con -> b t l con', t=t).contiguous()
446
+ for i, block in enumerate(self.transformer_blocks):
447
+ # calculate each batch one by one (since number in shape could not greater then 65,535 for some package)
448
+ for j in range(b):
449
+ context_j = repeat(
450
+ context[j],
451
+ 't l con -> (t r) l con', r=(h * w) // t, t=t).contiguous()
452
+ ## note: causal mask will not applied in cross-attention case
453
+ x[j] = block(x[j], context=context_j, **kwargs)
454
+
455
+ if self.use_linear:
456
+ x = self.proj_out(x)
457
+ x = rearrange(x, 'b (h w) t c -> b c t h w', h=h, w=w).contiguous()
458
+ if not self.use_linear:
459
+ x = rearrange(x, 'b hw t c -> (b hw) c t').contiguous()
460
+ x = self.proj_out(x)
461
+ x = rearrange(x, '(b h w) c t -> b c t h w', b=b, h=h, w=w).contiguous()
462
+
463
+ return x + x_in
464
+
465
+
466
+ class GEGLU(nn.Module):
467
+ def __init__(self, dim_in, dim_out):
468
+ super().__init__()
469
+ self.proj = nn.Linear(dim_in, dim_out * 2)
470
+
471
+ def forward(self, x):
472
+ x, gate = self.proj(x).chunk(2, dim=-1)
473
+ return x * F.gelu(gate)
474
+
475
+
476
+ class FeedForward(nn.Module):
477
+ def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
478
+ super().__init__()
479
+ inner_dim = int(dim * mult)
480
+ dim_out = default(dim_out, dim)
481
+ project_in = nn.Sequential(
482
+ nn.Linear(dim, inner_dim),
483
+ nn.GELU()
484
+ ) if not glu else GEGLU(dim, inner_dim)
485
+
486
+ self.net = nn.Sequential(
487
+ project_in,
488
+ nn.Dropout(dropout),
489
+ nn.Linear(inner_dim, dim_out)
490
+ )
491
+
492
+ def forward(self, x):
493
+ return self.net(x)
494
+
495
+
496
+ class LinearAttention(nn.Module):
497
+ def __init__(self, dim, heads=4, dim_head=32):
498
+ super().__init__()
499
+ self.heads = heads
500
+ hidden_dim = dim_head * heads
501
+ self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False)
502
+ self.to_out = nn.Conv2d(hidden_dim, dim, 1)
503
+
504
+ def forward(self, x):
505
+ b, c, h, w = x.shape
506
+ qkv = self.to_qkv(x)
507
+ q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3)
508
+ k = k.softmax(dim=-1)
509
+ context = torch.einsum('bhdn,bhen->bhde', k, v)
510
+ out = torch.einsum('bhde,bhdn->bhen', context, q)
511
+ out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w)
512
+ return self.to_out(out)
513
+
514
+
515
+ class SpatialSelfAttention(nn.Module):
516
+ def __init__(self, in_channels):
517
+ super().__init__()
518
+ self.in_channels = in_channels
519
+
520
+ self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
521
+ self.q = torch.nn.Conv2d(in_channels,
522
+ in_channels,
523
+ kernel_size=1,
524
+ stride=1,
525
+ padding=0)
526
+ self.k = torch.nn.Conv2d(in_channels,
527
+ in_channels,
528
+ kernel_size=1,
529
+ stride=1,
530
+ padding=0)
531
+ self.v = torch.nn.Conv2d(in_channels,
532
+ in_channels,
533
+ kernel_size=1,
534
+ stride=1,
535
+ padding=0)
536
+ self.proj_out = torch.nn.Conv2d(in_channels,
537
+ in_channels,
538
+ kernel_size=1,
539
+ stride=1,
540
+ padding=0)
541
+
542
+ def forward(self, x, **kwargs):
543
+ h_ = x
544
+ h_ = self.norm(h_)
545
+ q = self.q(h_)
546
+ k = self.k(h_)
547
+ v = self.v(h_)
548
+
549
+ # compute attention
550
+ b,c,h,w = q.shape
551
+ q = rearrange(q, 'b c h w -> b (h w) c')
552
+ k = rearrange(k, 'b c h w -> b c (h w)')
553
+ w_ = torch.einsum('bij,bjk->bik', q, k)
554
+
555
+ w_ = w_ * (int(c)**(-0.5))
556
+ w_ = torch.nn.functional.softmax(w_, dim=2)
557
+
558
+ # attend to values
559
+ v = rearrange(v, 'b c h w -> b c (h w)')
560
+ w_ = rearrange(w_, 'b i j -> b j i')
561
+ h_ = torch.einsum('bij,bjk->bik', v, w_)
562
+ h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h)
563
+ h_ = self.proj_out(h_)
564
+
565
+ return x+h_
lvdm/modules/encoders/condition.py ADDED
@@ -0,0 +1,392 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn as nn
3
+ from torch.utils.checkpoint import checkpoint
4
+ import kornia
5
+ import open_clip
6
+ from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel
7
+ from lvdm.common import autocast
8
+ from utils.utils import count_params
9
+
10
+ class AbstractEncoder(nn.Module):
11
+ def __init__(self):
12
+ super().__init__()
13
+
14
+ def encode(self, *args, **kwargs):
15
+ raise NotImplementedError
16
+
17
+
18
+ class IdentityEncoder(AbstractEncoder):
19
+
20
+ def encode(self, x):
21
+ return x
22
+
23
+
24
+ class ClassEmbedder(nn.Module):
25
+ def __init__(self, embed_dim, n_classes=1000, key='class', ucg_rate=0.1):
26
+ super().__init__()
27
+ self.key = key
28
+ self.embedding = nn.Embedding(n_classes, embed_dim)
29
+ self.n_classes = n_classes
30
+ self.ucg_rate = ucg_rate
31
+
32
+ def forward(self, batch, key=None, disable_dropout=False):
33
+ if key is None:
34
+ key = self.key
35
+ # this is for use in crossattn
36
+ c = batch[key][:, None]
37
+ if self.ucg_rate > 0. and not disable_dropout:
38
+ mask = 1. - torch.bernoulli(torch.ones_like(c) * self.ucg_rate)
39
+ c = mask * c + (1 - mask) * torch.ones_like(c) * (self.n_classes - 1)
40
+ c = c.long()
41
+ c = self.embedding(c)
42
+ return c
43
+
44
+ def get_unconditional_conditioning(self, bs, device="cuda"):
45
+ uc_class = self.n_classes - 1 # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000)
46
+ uc = torch.ones((bs,), device=device) * uc_class
47
+ uc = {self.key: uc}
48
+ return uc
49
+
50
+
51
+ def disabled_train(self, mode=True):
52
+ """Overwrite model.train with this function to make sure train/eval mode
53
+ does not change anymore."""
54
+ return self
55
+
56
+
57
+ class FrozenT5Embedder(AbstractEncoder):
58
+ """Uses the T5 transformer encoder for text"""
59
+
60
+ def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77,
61
+ freeze=True): # others are google/t5-v1_1-xl and google/t5-v1_1-xxl
62
+ super().__init__()
63
+ self.tokenizer = T5Tokenizer.from_pretrained(version)
64
+ self.transformer = T5EncoderModel.from_pretrained(version)
65
+ self.device = device
66
+ self.max_length = max_length # TODO: typical value?
67
+ if freeze:
68
+ self.freeze()
69
+
70
+ def freeze(self):
71
+ self.transformer = self.transformer.eval()
72
+ # self.train = disabled_train
73
+ for param in self.parameters():
74
+ param.requires_grad = False
75
+
76
+ def forward(self, text):
77
+ batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
78
+ return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
79
+ tokens = batch_encoding["input_ids"].to(self.device)
80
+ outputs = self.transformer(input_ids=tokens)
81
+
82
+ z = outputs.last_hidden_state
83
+ return z
84
+
85
+ def encode(self, text):
86
+ return self(text)
87
+
88
+
89
+ class FrozenCLIPEmbedder(AbstractEncoder):
90
+ """Uses the CLIP transformer encoder for text (from huggingface)"""
91
+ LAYERS = [
92
+ "last",
93
+ "pooled",
94
+ "hidden"
95
+ ]
96
+
97
+ def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77,
98
+ freeze=True, layer="last", layer_idx=None): # clip-vit-base-patch32
99
+ super().__init__()
100
+ assert layer in self.LAYERS
101
+ self.tokenizer = CLIPTokenizer.from_pretrained(version)
102
+ self.transformer = CLIPTextModel.from_pretrained(version)
103
+ self.device = device
104
+ self.max_length = max_length
105
+ if freeze:
106
+ self.freeze()
107
+ self.layer = layer
108
+ self.layer_idx = layer_idx
109
+ if layer == "hidden":
110
+ assert layer_idx is not None
111
+ assert 0 <= abs(layer_idx) <= 12
112
+
113
+ def freeze(self):
114
+ self.transformer = self.transformer.eval()
115
+ # self.train = disabled_train
116
+ for param in self.parameters():
117
+ param.requires_grad = False
118
+
119
+ def forward(self, text):
120
+ batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
121
+ return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
122
+ tokens = batch_encoding["input_ids"].to(self.device)
123
+ outputs = self.transformer(input_ids=tokens, output_hidden_states=self.layer == "hidden")
124
+ if self.layer == "last":
125
+ z = outputs.last_hidden_state
126
+ elif self.layer == "pooled":
127
+ z = outputs.pooler_output[:, None, :]
128
+ else:
129
+ z = outputs.hidden_states[self.layer_idx]
130
+ return z
131
+
132
+ def encode(self, text):
133
+ return self(text)
134
+
135
+
136
+ class ClipImageEmbedder(nn.Module):
137
+ def __init__(
138
+ self,
139
+ model,
140
+ jit=False,
141
+ device='cuda' if torch.cuda.is_available() else 'cpu',
142
+ antialias=True,
143
+ ucg_rate=0.
144
+ ):
145
+ super().__init__()
146
+ from clip import load as load_clip
147
+ self.model, _ = load_clip(name=model, device=device, jit=jit)
148
+
149
+ self.antialias = antialias
150
+
151
+ self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
152
+ self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
153
+ self.ucg_rate = ucg_rate
154
+
155
+ def preprocess(self, x):
156
+ # normalize to [0,1]
157
+ x = kornia.geometry.resize(x, (224, 224),
158
+ interpolation='bicubic', align_corners=True,
159
+ antialias=self.antialias)
160
+ x = (x + 1.) / 2.
161
+ # re-normalize according to clip
162
+ x = kornia.enhance.normalize(x, self.mean, self.std)
163
+ return x
164
+
165
+ def forward(self, x, no_dropout=False):
166
+ # x is assumed to be in range [-1,1]
167
+ out = self.model.encode_image(self.preprocess(x))
168
+ out = out.to(x.dtype)
169
+ if self.ucg_rate > 0. and not no_dropout:
170
+ out = torch.bernoulli((1. - self.ucg_rate) * torch.ones(out.shape[0], device=out.device))[:, None] * out
171
+ return out
172
+
173
+
174
+ class FrozenOpenCLIPEmbedder(AbstractEncoder):
175
+ """
176
+ Uses the OpenCLIP transformer encoder for text
177
+ """
178
+ LAYERS = [
179
+ # "pooled",
180
+ "last",
181
+ "penultimate"
182
+ ]
183
+
184
+ def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77,
185
+ freeze=True, layer="last"):
186
+ super().__init__()
187
+ assert layer in self.LAYERS
188
+ model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'))
189
+ del model.visual
190
+ self.model = model
191
+
192
+ self.device = device
193
+ self.max_length = max_length
194
+ if freeze:
195
+ self.freeze()
196
+ self.layer = layer
197
+ if self.layer == "last":
198
+ self.layer_idx = 0
199
+ elif self.layer == "penultimate":
200
+ self.layer_idx = 1
201
+ else:
202
+ raise NotImplementedError()
203
+
204
+ def freeze(self):
205
+ self.model = self.model.eval()
206
+ for param in self.parameters():
207
+ param.requires_grad = False
208
+
209
+ def forward(self, text):
210
+ self.device = self.model.positional_embedding.device
211
+ tokens = open_clip.tokenize(text)
212
+ z = self.encode_with_transformer(tokens.to(self.device))
213
+ return z
214
+
215
+ def encode_with_transformer(self, text):
216
+ x = self.model.token_embedding(text) # [batch_size, n_ctx, d_model]
217
+ x = x + self.model.positional_embedding
218
+ x = x.permute(1, 0, 2) # NLD -> LND
219
+ x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask)
220
+ x = x.permute(1, 0, 2) # LND -> NLD
221
+ x = self.model.ln_final(x)
222
+ return x
223
+
224
+ def text_transformer_forward(self, x: torch.Tensor, attn_mask=None):
225
+ for i, r in enumerate(self.model.transformer.resblocks):
226
+ if i == len(self.model.transformer.resblocks) - self.layer_idx:
227
+ break
228
+ if self.model.transformer.grad_checkpointing and not torch.jit.is_scripting():
229
+ x = checkpoint(r, x, attn_mask)
230
+ else:
231
+ x = r(x, attn_mask=attn_mask)
232
+ return x
233
+
234
+ def encode(self, text):
235
+ return self(text)
236
+
237
+
238
+ class FrozenOpenCLIPImageEmbedder(AbstractEncoder):
239
+ """
240
+ Uses the OpenCLIP vision transformer encoder for images
241
+ """
242
+
243
+ def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77,
244
+ freeze=True, layer="pooled", antialias=True, ucg_rate=0.):
245
+ super().__init__()
246
+ model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'),
247
+ pretrained=version, )
248
+ del model.transformer
249
+ self.model = model
250
+
251
+ self.device = device
252
+ self.max_length = max_length
253
+ if freeze:
254
+ self.freeze()
255
+ self.layer = layer
256
+ if self.layer == "penultimate":
257
+ raise NotImplementedError()
258
+ self.layer_idx = 1
259
+
260
+ self.antialias = antialias
261
+
262
+ self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
263
+ self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
264
+ self.ucg_rate = ucg_rate
265
+
266
+ def preprocess(self, x):
267
+ # normalize to [0,1]
268
+ x = kornia.geometry.resize(x, (224, 224),
269
+ interpolation='bicubic', align_corners=True,
270
+ antialias=self.antialias)
271
+ x = (x + 1.) / 2.
272
+ # renormalize according to clip
273
+ x = kornia.enhance.normalize(x, self.mean, self.std)
274
+ return x
275
+
276
+ def freeze(self):
277
+ self.model = self.model.eval()
278
+ for param in self.parameters():
279
+ param.requires_grad = False
280
+
281
+ @autocast
282
+ def forward(self, image, no_dropout=False):
283
+ z = self.encode_with_vision_transformer(image)
284
+ if self.ucg_rate > 0. and not no_dropout:
285
+ z = torch.bernoulli((1. - self.ucg_rate) * torch.ones(z.shape[0], device=z.device))[:, None] * z
286
+ return z
287
+
288
+ def encode_with_vision_transformer(self, img):
289
+ img = self.preprocess(img)
290
+ x = self.model.visual(img)
291
+ return x
292
+
293
+ def encode(self, text):
294
+ return self(text)
295
+
296
+
297
+
298
+ class FrozenOpenCLIPImageEmbedderV2(AbstractEncoder):
299
+ """
300
+ Uses the OpenCLIP vision transformer encoder for images
301
+ """
302
+
303
+ def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda",
304
+ freeze=True, layer="pooled", antialias=True):
305
+ super().__init__()
306
+ model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'),
307
+ pretrained=version, )
308
+ del model.transformer
309
+ self.model = model
310
+ self.device = device
311
+
312
+ if freeze:
313
+ self.freeze()
314
+ self.layer = layer
315
+ if self.layer == "penultimate":
316
+ raise NotImplementedError()
317
+ self.layer_idx = 1
318
+
319
+ self.antialias = antialias
320
+ self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
321
+ self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
322
+
323
+
324
+ def preprocess(self, x):
325
+ # normalize to [0,1]
326
+ x = kornia.geometry.resize(x, (224, 224),
327
+ interpolation='bicubic', align_corners=True,
328
+ antialias=self.antialias)
329
+ x = (x + 1.) / 2.
330
+ # renormalize according to clip
331
+ x = kornia.enhance.normalize(x, self.mean, self.std)
332
+ return x
333
+
334
+ def freeze(self):
335
+ self.model = self.model.eval()
336
+ for param in self.model.parameters():
337
+ param.requires_grad = False
338
+
339
+ def forward(self, image, no_dropout=False):
340
+ ## image: b c h w
341
+ z = self.encode_with_vision_transformer(image)
342
+ return z
343
+
344
+ def encode_with_vision_transformer(self, x):
345
+ x = self.preprocess(x)
346
+
347
+ # to patches - whether to use dual patchnorm - https://arxiv.org/abs/2302.01327v1
348
+ if self.model.visual.input_patchnorm:
349
+ # einops - rearrange(x, 'b c (h p1) (w p2) -> b (h w) (c p1 p2)')
350
+ x = x.reshape(x.shape[0], x.shape[1], self.model.visual.grid_size[0], self.model.visual.patch_size[0], self.model.visual.grid_size[1], self.model.visual.patch_size[1])
351
+ x = x.permute(0, 2, 4, 1, 3, 5)
352
+ x = x.reshape(x.shape[0], self.model.visual.grid_size[0] * self.model.visual.grid_size[1], -1)
353
+ x = self.model.visual.patchnorm_pre_ln(x)
354
+ x = self.model.visual.conv1(x)
355
+ else:
356
+ x = self.model.visual.conv1(x) # shape = [*, width, grid, grid]
357
+ x = x.reshape(x.shape[0], x.shape[1], -1) # shape = [*, width, grid ** 2]
358
+ x = x.permute(0, 2, 1) # shape = [*, grid ** 2, width]
359
+
360
+ # class embeddings and positional embeddings
361
+ x = torch.cat(
362
+ [self.model.visual.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device),
363
+ x], dim=1) # shape = [*, grid ** 2 + 1, width]
364
+ x = x + self.model.visual.positional_embedding.to(x.dtype)
365
+
366
+ # a patch_dropout of 0. would mean it is disabled and this function would do nothing but return what was passed in
367
+ x = self.model.visual.patch_dropout(x)
368
+ x = self.model.visual.ln_pre(x)
369
+
370
+ x = x.permute(1, 0, 2) # NLD -> LND
371
+ x = self.model.visual.transformer(x)
372
+ x = x.permute(1, 0, 2) # LND -> NLD
373
+
374
+ return x
375
+
376
+
377
+ class FrozenCLIPT5Encoder(AbstractEncoder):
378
+ def __init__(self, clip_version="openai/clip-vit-large-patch14", t5_version="google/t5-v1_1-xl", device="cuda",
379
+ clip_max_length=77, t5_max_length=77):
380
+ super().__init__()
381
+ self.clip_encoder = FrozenCLIPEmbedder(clip_version, device, max_length=clip_max_length)
382
+ self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length)
383
+ print(f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder) * 1.e-6:.2f} M parameters, "
384
+ f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder) * 1.e-6:.2f} M params.")
385
+
386
+ def encode(self, text):
387
+ return self(text)
388
+
389
+ def forward(self, text):
390
+ clip_z = self.clip_encoder.encode(text)
391
+ t5_z = self.t5_encoder.encode(text)
392
+ return [clip_z, t5_z]
lvdm/modules/encoders/ip_resampler.py ADDED
@@ -0,0 +1,136 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # modified from https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py
2
+ import math
3
+ import torch
4
+ import torch.nn as nn
5
+
6
+
7
+ class ImageProjModel(nn.Module):
8
+ """Projection Model"""
9
+ def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4):
10
+ super().__init__()
11
+ self.cross_attention_dim = cross_attention_dim
12
+ self.clip_extra_context_tokens = clip_extra_context_tokens
13
+ self.proj = nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
14
+ self.norm = nn.LayerNorm(cross_attention_dim)
15
+
16
+ def forward(self, image_embeds):
17
+ #embeds = image_embeds
18
+ embeds = image_embeds.type(list(self.proj.parameters())[0].dtype)
19
+ clip_extra_context_tokens = self.proj(embeds).reshape(-1, self.clip_extra_context_tokens, self.cross_attention_dim)
20
+ clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
21
+ return clip_extra_context_tokens
22
+
23
+ # FFN
24
+ def FeedForward(dim, mult=4):
25
+ inner_dim = int(dim * mult)
26
+ return nn.Sequential(
27
+ nn.LayerNorm(dim),
28
+ nn.Linear(dim, inner_dim, bias=False),
29
+ nn.GELU(),
30
+ nn.Linear(inner_dim, dim, bias=False),
31
+ )
32
+
33
+
34
+ def reshape_tensor(x, heads):
35
+ bs, length, width = x.shape
36
+ #(bs, length, width) --> (bs, length, n_heads, dim_per_head)
37
+ x = x.view(bs, length, heads, -1)
38
+ # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
39
+ x = x.transpose(1, 2)
40
+ # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
41
+ x = x.reshape(bs, heads, length, -1)
42
+ return x
43
+
44
+
45
+ class PerceiverAttention(nn.Module):
46
+ def __init__(self, *, dim, dim_head=64, heads=8):
47
+ super().__init__()
48
+ self.scale = dim_head**-0.5
49
+ self.dim_head = dim_head
50
+ self.heads = heads
51
+ inner_dim = dim_head * heads
52
+
53
+ self.norm1 = nn.LayerNorm(dim)
54
+ self.norm2 = nn.LayerNorm(dim)
55
+
56
+ self.to_q = nn.Linear(dim, inner_dim, bias=False)
57
+ self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
58
+ self.to_out = nn.Linear(inner_dim, dim, bias=False)
59
+
60
+
61
+ def forward(self, x, latents):
62
+ """
63
+ Args:
64
+ x (torch.Tensor): image features
65
+ shape (b, n1, D)
66
+ latent (torch.Tensor): latent features
67
+ shape (b, n2, D)
68
+ """
69
+ x = self.norm1(x)
70
+ latents = self.norm2(latents)
71
+
72
+ b, l, _ = latents.shape
73
+
74
+ q = self.to_q(latents)
75
+ kv_input = torch.cat((x, latents), dim=-2)
76
+ k, v = self.to_kv(kv_input).chunk(2, dim=-1)
77
+
78
+ q = reshape_tensor(q, self.heads)
79
+ k = reshape_tensor(k, self.heads)
80
+ v = reshape_tensor(v, self.heads)
81
+
82
+ # attention
83
+ scale = 1 / math.sqrt(math.sqrt(self.dim_head))
84
+ weight = (q * scale) @ (k * scale).transpose(-2, -1) # More stable with f16 than dividing afterwards
85
+ weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
86
+ out = weight @ v
87
+
88
+ out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
89
+
90
+ return self.to_out(out)
91
+
92
+
93
+ class Resampler(nn.Module):
94
+ def __init__(
95
+ self,
96
+ dim=1024,
97
+ depth=8,
98
+ dim_head=64,
99
+ heads=16,
100
+ num_queries=8,
101
+ embedding_dim=768,
102
+ output_dim=1024,
103
+ ff_mult=4,
104
+ ):
105
+ super().__init__()
106
+
107
+ self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
108
+
109
+ self.proj_in = nn.Linear(embedding_dim, dim)
110
+
111
+ self.proj_out = nn.Linear(dim, output_dim)
112
+ self.norm_out = nn.LayerNorm(output_dim)
113
+
114
+ self.layers = nn.ModuleList([])
115
+ for _ in range(depth):
116
+ self.layers.append(
117
+ nn.ModuleList(
118
+ [
119
+ PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
120
+ FeedForward(dim=dim, mult=ff_mult),
121
+ ]
122
+ )
123
+ )
124
+
125
+ def forward(self, x):
126
+
127
+ latents = self.latents.repeat(x.size(0), 1, 1)
128
+
129
+ x = self.proj_in(x)
130
+
131
+ for attn, ff in self.layers:
132
+ latents = attn(x, latents) + latents
133
+ latents = ff(latents) + latents
134
+
135
+ latents = self.proj_out(latents)
136
+ return self.norm_out(latents)
lvdm/modules/networks/ae_modules.py ADDED
@@ -0,0 +1,845 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pytorch_diffusion + derived encoder decoder
2
+ import math
3
+ import torch
4
+ import numpy as np
5
+ import torch.nn as nn
6
+ from einops import rearrange
7
+ from utils.utils import instantiate_from_config
8
+ from lvdm.modules.attention import LinearAttention
9
+
10
+ def nonlinearity(x):
11
+ # swish
12
+ return x*torch.sigmoid(x)
13
+
14
+
15
+ def Normalize(in_channels, num_groups=32):
16
+ return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True)
17
+
18
+
19
+
20
+ class LinAttnBlock(LinearAttention):
21
+ """to match AttnBlock usage"""
22
+ def __init__(self, in_channels):
23
+ super().__init__(dim=in_channels, heads=1, dim_head=in_channels)
24
+
25
+
26
+ class AttnBlock(nn.Module):
27
+ def __init__(self, in_channels):
28
+ super().__init__()
29
+ self.in_channels = in_channels
30
+
31
+ self.norm = Normalize(in_channels)
32
+ self.q = torch.nn.Conv2d(in_channels,
33
+ in_channels,
34
+ kernel_size=1,
35
+ stride=1,
36
+ padding=0)
37
+ self.k = torch.nn.Conv2d(in_channels,
38
+ in_channels,
39
+ kernel_size=1,
40
+ stride=1,
41
+ padding=0)
42
+ self.v = torch.nn.Conv2d(in_channels,
43
+ in_channels,
44
+ kernel_size=1,
45
+ stride=1,
46
+ padding=0)
47
+ self.proj_out = torch.nn.Conv2d(in_channels,
48
+ in_channels,
49
+ kernel_size=1,
50
+ stride=1,
51
+ padding=0)
52
+
53
+ def forward(self, x):
54
+ h_ = x
55
+ h_ = self.norm(h_)
56
+ q = self.q(h_)
57
+ k = self.k(h_)
58
+ v = self.v(h_)
59
+
60
+ # compute attention
61
+ b,c,h,w = q.shape
62
+ q = q.reshape(b,c,h*w) # bcl
63
+ q = q.permute(0,2,1) # bcl -> blc l=hw
64
+ k = k.reshape(b,c,h*w) # bcl
65
+
66
+ w_ = torch.bmm(q,k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
67
+ w_ = w_ * (int(c)**(-0.5))
68
+ w_ = torch.nn.functional.softmax(w_, dim=2)
69
+
70
+ # attend to values
71
+ v = v.reshape(b,c,h*w)
72
+ w_ = w_.permute(0,2,1) # b,hw,hw (first hw of k, second of q)
73
+ 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]
74
+ h_ = h_.reshape(b,c,h,w)
75
+
76
+ h_ = self.proj_out(h_)
77
+
78
+ return x+h_
79
+
80
+ def make_attn(in_channels, attn_type="vanilla"):
81
+ assert attn_type in ["vanilla", "linear", "none"], f'attn_type {attn_type} unknown'
82
+ #print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
83
+ if attn_type == "vanilla":
84
+ return AttnBlock(in_channels)
85
+ elif attn_type == "none":
86
+ return nn.Identity(in_channels)
87
+ else:
88
+ return LinAttnBlock(in_channels)
89
+
90
+ class Downsample(nn.Module):
91
+ def __init__(self, in_channels, with_conv):
92
+ super().__init__()
93
+ self.with_conv = with_conv
94
+ self.in_channels = in_channels
95
+ if self.with_conv:
96
+ # no asymmetric padding in torch conv, must do it ourselves
97
+ self.conv = torch.nn.Conv2d(in_channels,
98
+ in_channels,
99
+ kernel_size=3,
100
+ stride=2,
101
+ padding=0)
102
+ def forward(self, x):
103
+ if self.with_conv:
104
+ pad = (0,1,0,1)
105
+ x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
106
+ x = self.conv(x)
107
+ else:
108
+ x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
109
+ return x
110
+
111
+ class Upsample(nn.Module):
112
+ def __init__(self, in_channels, with_conv):
113
+ super().__init__()
114
+ self.with_conv = with_conv
115
+ self.in_channels = in_channels
116
+ if self.with_conv:
117
+ self.conv = torch.nn.Conv2d(in_channels,
118
+ in_channels,
119
+ kernel_size=3,
120
+ stride=1,
121
+ padding=1)
122
+
123
+ def forward(self, x):
124
+ x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
125
+ if self.with_conv:
126
+ x = self.conv(x)
127
+ return x
128
+
129
+ def get_timestep_embedding(timesteps, embedding_dim):
130
+ """
131
+ This matches the implementation in Denoising Diffusion Probabilistic Models:
132
+ From Fairseq.
133
+ Build sinusoidal embeddings.
134
+ This matches the implementation in tensor2tensor, but differs slightly
135
+ from the description in Section 3.5 of "Attention Is All You Need".
136
+ """
137
+ assert len(timesteps.shape) == 1
138
+
139
+ half_dim = embedding_dim // 2
140
+ emb = math.log(10000) / (half_dim - 1)
141
+ emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
142
+ emb = emb.to(device=timesteps.device)
143
+ emb = timesteps.float()[:, None] * emb[None, :]
144
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
145
+ if embedding_dim % 2 == 1: # zero pad
146
+ emb = torch.nn.functional.pad(emb, (0,1,0,0))
147
+ return emb
148
+
149
+
150
+
151
+ class ResnetBlock(nn.Module):
152
+ def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False,
153
+ dropout, temb_channels=512):
154
+ super().__init__()
155
+ self.in_channels = in_channels
156
+ out_channels = in_channels if out_channels is None else out_channels
157
+ self.out_channels = out_channels
158
+ self.use_conv_shortcut = conv_shortcut
159
+
160
+ self.norm1 = Normalize(in_channels)
161
+ self.conv1 = torch.nn.Conv2d(in_channels,
162
+ out_channels,
163
+ kernel_size=3,
164
+ stride=1,
165
+ padding=1)
166
+ if temb_channels > 0:
167
+ self.temb_proj = torch.nn.Linear(temb_channels,
168
+ out_channels)
169
+ self.norm2 = Normalize(out_channels)
170
+ self.dropout = torch.nn.Dropout(dropout)
171
+ self.conv2 = torch.nn.Conv2d(out_channels,
172
+ out_channels,
173
+ kernel_size=3,
174
+ stride=1,
175
+ padding=1)
176
+ if self.in_channels != self.out_channels:
177
+ if self.use_conv_shortcut:
178
+ self.conv_shortcut = torch.nn.Conv2d(in_channels,
179
+ out_channels,
180
+ kernel_size=3,
181
+ stride=1,
182
+ padding=1)
183
+ else:
184
+ self.nin_shortcut = torch.nn.Conv2d(in_channels,
185
+ out_channels,
186
+ kernel_size=1,
187
+ stride=1,
188
+ padding=0)
189
+
190
+ def forward(self, x, temb):
191
+ h = x
192
+ h = self.norm1(h)
193
+ h = nonlinearity(h)
194
+ h = self.conv1(h)
195
+
196
+ if temb is not None:
197
+ h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None]
198
+
199
+ h = self.norm2(h)
200
+ h = nonlinearity(h)
201
+ h = self.dropout(h)
202
+ h = self.conv2(h)
203
+
204
+ if self.in_channels != self.out_channels:
205
+ if self.use_conv_shortcut:
206
+ x = self.conv_shortcut(x)
207
+ else:
208
+ x = self.nin_shortcut(x)
209
+
210
+ return x+h
211
+
212
+ class Model(nn.Module):
213
+ def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
214
+ attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
215
+ resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"):
216
+ super().__init__()
217
+ if use_linear_attn: attn_type = "linear"
218
+ self.ch = ch
219
+ self.temb_ch = self.ch*4
220
+ self.num_resolutions = len(ch_mult)
221
+ self.num_res_blocks = num_res_blocks
222
+ self.resolution = resolution
223
+ self.in_channels = in_channels
224
+
225
+ self.use_timestep = use_timestep
226
+ if self.use_timestep:
227
+ # timestep embedding
228
+ self.temb = nn.Module()
229
+ self.temb.dense = nn.ModuleList([
230
+ torch.nn.Linear(self.ch,
231
+ self.temb_ch),
232
+ torch.nn.Linear(self.temb_ch,
233
+ self.temb_ch),
234
+ ])
235
+
236
+ # downsampling
237
+ self.conv_in = torch.nn.Conv2d(in_channels,
238
+ self.ch,
239
+ kernel_size=3,
240
+ stride=1,
241
+ padding=1)
242
+
243
+ curr_res = resolution
244
+ in_ch_mult = (1,)+tuple(ch_mult)
245
+ self.down = nn.ModuleList()
246
+ for i_level in range(self.num_resolutions):
247
+ block = nn.ModuleList()
248
+ attn = nn.ModuleList()
249
+ block_in = ch*in_ch_mult[i_level]
250
+ block_out = ch*ch_mult[i_level]
251
+ for i_block in range(self.num_res_blocks):
252
+ block.append(ResnetBlock(in_channels=block_in,
253
+ out_channels=block_out,
254
+ temb_channels=self.temb_ch,
255
+ dropout=dropout))
256
+ block_in = block_out
257
+ if curr_res in attn_resolutions:
258
+ attn.append(make_attn(block_in, attn_type=attn_type))
259
+ down = nn.Module()
260
+ down.block = block
261
+ down.attn = attn
262
+ if i_level != self.num_resolutions-1:
263
+ down.downsample = Downsample(block_in, resamp_with_conv)
264
+ curr_res = curr_res // 2
265
+ self.down.append(down)
266
+
267
+ # middle
268
+ self.mid = nn.Module()
269
+ self.mid.block_1 = ResnetBlock(in_channels=block_in,
270
+ out_channels=block_in,
271
+ temb_channels=self.temb_ch,
272
+ dropout=dropout)
273
+ self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
274
+ self.mid.block_2 = ResnetBlock(in_channels=block_in,
275
+ out_channels=block_in,
276
+ temb_channels=self.temb_ch,
277
+ dropout=dropout)
278
+
279
+ # upsampling
280
+ self.up = nn.ModuleList()
281
+ for i_level in reversed(range(self.num_resolutions)):
282
+ block = nn.ModuleList()
283
+ attn = nn.ModuleList()
284
+ block_out = ch*ch_mult[i_level]
285
+ skip_in = ch*ch_mult[i_level]
286
+ for i_block in range(self.num_res_blocks+1):
287
+ if i_block == self.num_res_blocks:
288
+ skip_in = ch*in_ch_mult[i_level]
289
+ block.append(ResnetBlock(in_channels=block_in+skip_in,
290
+ out_channels=block_out,
291
+ temb_channels=self.temb_ch,
292
+ dropout=dropout))
293
+ block_in = block_out
294
+ if curr_res in attn_resolutions:
295
+ attn.append(make_attn(block_in, attn_type=attn_type))
296
+ up = nn.Module()
297
+ up.block = block
298
+ up.attn = attn
299
+ if i_level != 0:
300
+ up.upsample = Upsample(block_in, resamp_with_conv)
301
+ curr_res = curr_res * 2
302
+ self.up.insert(0, up) # prepend to get consistent order
303
+
304
+ # end
305
+ self.norm_out = Normalize(block_in)
306
+ self.conv_out = torch.nn.Conv2d(block_in,
307
+ out_ch,
308
+ kernel_size=3,
309
+ stride=1,
310
+ padding=1)
311
+
312
+ def forward(self, x, t=None, context=None):
313
+ #assert x.shape[2] == x.shape[3] == self.resolution
314
+ if context is not None:
315
+ # assume aligned context, cat along channel axis
316
+ x = torch.cat((x, context), dim=1)
317
+ if self.use_timestep:
318
+ # timestep embedding
319
+ assert t is not None
320
+ temb = get_timestep_embedding(t, self.ch)
321
+ temb = self.temb.dense[0](temb)
322
+ temb = nonlinearity(temb)
323
+ temb = self.temb.dense[1](temb)
324
+ else:
325
+ temb = None
326
+
327
+ # downsampling
328
+ hs = [self.conv_in(x)]
329
+ for i_level in range(self.num_resolutions):
330
+ for i_block in range(self.num_res_blocks):
331
+ h = self.down[i_level].block[i_block](hs[-1], temb)
332
+ if len(self.down[i_level].attn) > 0:
333
+ h = self.down[i_level].attn[i_block](h)
334
+ hs.append(h)
335
+ if i_level != self.num_resolutions-1:
336
+ hs.append(self.down[i_level].downsample(hs[-1]))
337
+
338
+ # middle
339
+ h = hs[-1]
340
+ h = self.mid.block_1(h, temb)
341
+ h = self.mid.attn_1(h)
342
+ h = self.mid.block_2(h, temb)
343
+
344
+ # upsampling
345
+ for i_level in reversed(range(self.num_resolutions)):
346
+ for i_block in range(self.num_res_blocks+1):
347
+ h = self.up[i_level].block[i_block](
348
+ torch.cat([h, hs.pop()], dim=1), temb)
349
+ if len(self.up[i_level].attn) > 0:
350
+ h = self.up[i_level].attn[i_block](h)
351
+ if i_level != 0:
352
+ h = self.up[i_level].upsample(h)
353
+
354
+ # end
355
+ h = self.norm_out(h)
356
+ h = nonlinearity(h)
357
+ h = self.conv_out(h)
358
+ return h
359
+
360
+ def get_last_layer(self):
361
+ return self.conv_out.weight
362
+
363
+
364
+ class Encoder(nn.Module):
365
+ def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
366
+ attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
367
+ resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla",
368
+ **ignore_kwargs):
369
+ super().__init__()
370
+ if use_linear_attn: attn_type = "linear"
371
+ self.ch = ch
372
+ self.temb_ch = 0
373
+ self.num_resolutions = len(ch_mult)
374
+ self.num_res_blocks = num_res_blocks
375
+ self.resolution = resolution
376
+ self.in_channels = in_channels
377
+
378
+ # downsampling
379
+ self.conv_in = torch.nn.Conv2d(in_channels,
380
+ self.ch,
381
+ kernel_size=3,
382
+ stride=1,
383
+ padding=1)
384
+
385
+ curr_res = resolution
386
+ in_ch_mult = (1,)+tuple(ch_mult)
387
+ self.in_ch_mult = in_ch_mult
388
+ self.down = nn.ModuleList()
389
+ for i_level in range(self.num_resolutions):
390
+ block = nn.ModuleList()
391
+ attn = nn.ModuleList()
392
+ block_in = ch*in_ch_mult[i_level]
393
+ block_out = ch*ch_mult[i_level]
394
+ for i_block in range(self.num_res_blocks):
395
+ block.append(ResnetBlock(in_channels=block_in,
396
+ out_channels=block_out,
397
+ temb_channels=self.temb_ch,
398
+ dropout=dropout))
399
+ block_in = block_out
400
+ if curr_res in attn_resolutions:
401
+ attn.append(make_attn(block_in, attn_type=attn_type))
402
+ down = nn.Module()
403
+ down.block = block
404
+ down.attn = attn
405
+ if i_level != self.num_resolutions-1:
406
+ down.downsample = Downsample(block_in, resamp_with_conv)
407
+ curr_res = curr_res // 2
408
+ self.down.append(down)
409
+
410
+ # middle
411
+ self.mid = nn.Module()
412
+ self.mid.block_1 = ResnetBlock(in_channels=block_in,
413
+ out_channels=block_in,
414
+ temb_channels=self.temb_ch,
415
+ dropout=dropout)
416
+ self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
417
+ self.mid.block_2 = ResnetBlock(in_channels=block_in,
418
+ out_channels=block_in,
419
+ temb_channels=self.temb_ch,
420
+ dropout=dropout)
421
+
422
+ # end
423
+ self.norm_out = Normalize(block_in)
424
+ self.conv_out = torch.nn.Conv2d(block_in,
425
+ 2*z_channels if double_z else z_channels,
426
+ kernel_size=3,
427
+ stride=1,
428
+ padding=1)
429
+
430
+ def forward(self, x):
431
+ # timestep embedding
432
+ temb = None
433
+
434
+ # print(f'encoder-input={x.shape}')
435
+ # downsampling
436
+ hs = [self.conv_in(x)]
437
+ # print(f'encoder-conv in feat={hs[0].shape}')
438
+ for i_level in range(self.num_resolutions):
439
+ for i_block in range(self.num_res_blocks):
440
+ h = self.down[i_level].block[i_block](hs[-1], temb)
441
+ # print(f'encoder-down feat={h.shape}')
442
+ if len(self.down[i_level].attn) > 0:
443
+ h = self.down[i_level].attn[i_block](h)
444
+ hs.append(h)
445
+ if i_level != self.num_resolutions-1:
446
+ # print(f'encoder-downsample (input)={hs[-1].shape}')
447
+ hs.append(self.down[i_level].downsample(hs[-1]))
448
+ # print(f'encoder-downsample (output)={hs[-1].shape}')
449
+
450
+ # middle
451
+ h = hs[-1]
452
+ h = self.mid.block_1(h, temb)
453
+ # print(f'encoder-mid1 feat={h.shape}')
454
+ h = self.mid.attn_1(h)
455
+ h = self.mid.block_2(h, temb)
456
+ # print(f'encoder-mid2 feat={h.shape}')
457
+
458
+ # end
459
+ h = self.norm_out(h)
460
+ h = nonlinearity(h)
461
+ h = self.conv_out(h)
462
+ # print(f'end feat={h.shape}')
463
+ return h
464
+
465
+
466
+ class Decoder(nn.Module):
467
+ def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks,
468
+ attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels,
469
+ resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
470
+ attn_type="vanilla", **ignorekwargs):
471
+ super().__init__()
472
+ if use_linear_attn: attn_type = "linear"
473
+ self.ch = ch
474
+ self.temb_ch = 0
475
+ self.num_resolutions = len(ch_mult)
476
+ self.num_res_blocks = num_res_blocks
477
+ self.resolution = resolution
478
+ self.in_channels = in_channels
479
+ self.give_pre_end = give_pre_end
480
+ self.tanh_out = tanh_out
481
+
482
+ # compute in_ch_mult, block_in and curr_res at lowest res
483
+ in_ch_mult = (1,)+tuple(ch_mult)
484
+ block_in = ch*ch_mult[self.num_resolutions-1]
485
+ curr_res = resolution // 2**(self.num_resolutions-1)
486
+ self.z_shape = (1,z_channels,curr_res,curr_res)
487
+ print("AE working on z of shape {} = {} dimensions.".format(
488
+ self.z_shape, np.prod(self.z_shape)))
489
+
490
+ # z to block_in
491
+ self.conv_in = torch.nn.Conv2d(z_channels,
492
+ block_in,
493
+ kernel_size=3,
494
+ stride=1,
495
+ padding=1)
496
+
497
+ # middle
498
+ self.mid = nn.Module()
499
+ self.mid.block_1 = ResnetBlock(in_channels=block_in,
500
+ out_channels=block_in,
501
+ temb_channels=self.temb_ch,
502
+ dropout=dropout)
503
+ self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
504
+ self.mid.block_2 = ResnetBlock(in_channels=block_in,
505
+ out_channels=block_in,
506
+ temb_channels=self.temb_ch,
507
+ dropout=dropout)
508
+
509
+ # upsampling
510
+ self.up = nn.ModuleList()
511
+ for i_level in reversed(range(self.num_resolutions)):
512
+ block = nn.ModuleList()
513
+ attn = nn.ModuleList()
514
+ block_out = ch*ch_mult[i_level]
515
+ for i_block in range(self.num_res_blocks+1):
516
+ block.append(ResnetBlock(in_channels=block_in,
517
+ out_channels=block_out,
518
+ temb_channels=self.temb_ch,
519
+ dropout=dropout))
520
+ block_in = block_out
521
+ if curr_res in attn_resolutions:
522
+ attn.append(make_attn(block_in, attn_type=attn_type))
523
+ up = nn.Module()
524
+ up.block = block
525
+ up.attn = attn
526
+ if i_level != 0:
527
+ up.upsample = Upsample(block_in, resamp_with_conv)
528
+ curr_res = curr_res * 2
529
+ self.up.insert(0, up) # prepend to get consistent order
530
+
531
+ # end
532
+ self.norm_out = Normalize(block_in)
533
+ self.conv_out = torch.nn.Conv2d(block_in,
534
+ out_ch,
535
+ kernel_size=3,
536
+ stride=1,
537
+ padding=1)
538
+
539
+ def forward(self, z):
540
+ #assert z.shape[1:] == self.z_shape[1:]
541
+ self.last_z_shape = z.shape
542
+
543
+ # print(f'decoder-input={z.shape}')
544
+ # timestep embedding
545
+ temb = None
546
+
547
+ # z to block_in
548
+ h = self.conv_in(z)
549
+ # print(f'decoder-conv in feat={h.shape}')
550
+
551
+ # middle
552
+ h = self.mid.block_1(h, temb)
553
+ h = self.mid.attn_1(h)
554
+ h = self.mid.block_2(h, temb)
555
+ # print(f'decoder-mid feat={h.shape}')
556
+
557
+ # upsampling
558
+ for i_level in reversed(range(self.num_resolutions)):
559
+ for i_block in range(self.num_res_blocks+1):
560
+ h = self.up[i_level].block[i_block](h, temb)
561
+ if len(self.up[i_level].attn) > 0:
562
+ h = self.up[i_level].attn[i_block](h)
563
+ # print(f'decoder-up feat={h.shape}')
564
+ if i_level != 0:
565
+ h = self.up[i_level].upsample(h)
566
+ # print(f'decoder-upsample feat={h.shape}')
567
+
568
+ # end
569
+ if self.give_pre_end:
570
+ return h
571
+
572
+ h = self.norm_out(h)
573
+ h = nonlinearity(h)
574
+ h = self.conv_out(h)
575
+ # print(f'decoder-conv_out feat={h.shape}')
576
+ if self.tanh_out:
577
+ h = torch.tanh(h)
578
+ return h
579
+
580
+
581
+ class SimpleDecoder(nn.Module):
582
+ def __init__(self, in_channels, out_channels, *args, **kwargs):
583
+ super().__init__()
584
+ self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1),
585
+ ResnetBlock(in_channels=in_channels,
586
+ out_channels=2 * in_channels,
587
+ temb_channels=0, dropout=0.0),
588
+ ResnetBlock(in_channels=2 * in_channels,
589
+ out_channels=4 * in_channels,
590
+ temb_channels=0, dropout=0.0),
591
+ ResnetBlock(in_channels=4 * in_channels,
592
+ out_channels=2 * in_channels,
593
+ temb_channels=0, dropout=0.0),
594
+ nn.Conv2d(2*in_channels, in_channels, 1),
595
+ Upsample(in_channels, with_conv=True)])
596
+ # end
597
+ self.norm_out = Normalize(in_channels)
598
+ self.conv_out = torch.nn.Conv2d(in_channels,
599
+ out_channels,
600
+ kernel_size=3,
601
+ stride=1,
602
+ padding=1)
603
+
604
+ def forward(self, x):
605
+ for i, layer in enumerate(self.model):
606
+ if i in [1,2,3]:
607
+ x = layer(x, None)
608
+ else:
609
+ x = layer(x)
610
+
611
+ h = self.norm_out(x)
612
+ h = nonlinearity(h)
613
+ x = self.conv_out(h)
614
+ return x
615
+
616
+
617
+ class UpsampleDecoder(nn.Module):
618
+ def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution,
619
+ ch_mult=(2,2), dropout=0.0):
620
+ super().__init__()
621
+ # upsampling
622
+ self.temb_ch = 0
623
+ self.num_resolutions = len(ch_mult)
624
+ self.num_res_blocks = num_res_blocks
625
+ block_in = in_channels
626
+ curr_res = resolution // 2 ** (self.num_resolutions - 1)
627
+ self.res_blocks = nn.ModuleList()
628
+ self.upsample_blocks = nn.ModuleList()
629
+ for i_level in range(self.num_resolutions):
630
+ res_block = []
631
+ block_out = ch * ch_mult[i_level]
632
+ for i_block in range(self.num_res_blocks + 1):
633
+ res_block.append(ResnetBlock(in_channels=block_in,
634
+ out_channels=block_out,
635
+ temb_channels=self.temb_ch,
636
+ dropout=dropout))
637
+ block_in = block_out
638
+ self.res_blocks.append(nn.ModuleList(res_block))
639
+ if i_level != self.num_resolutions - 1:
640
+ self.upsample_blocks.append(Upsample(block_in, True))
641
+ curr_res = curr_res * 2
642
+
643
+ # end
644
+ self.norm_out = Normalize(block_in)
645
+ self.conv_out = torch.nn.Conv2d(block_in,
646
+ out_channels,
647
+ kernel_size=3,
648
+ stride=1,
649
+ padding=1)
650
+
651
+ def forward(self, x):
652
+ # upsampling
653
+ h = x
654
+ for k, i_level in enumerate(range(self.num_resolutions)):
655
+ for i_block in range(self.num_res_blocks + 1):
656
+ h = self.res_blocks[i_level][i_block](h, None)
657
+ if i_level != self.num_resolutions - 1:
658
+ h = self.upsample_blocks[k](h)
659
+ h = self.norm_out(h)
660
+ h = nonlinearity(h)
661
+ h = self.conv_out(h)
662
+ return h
663
+
664
+
665
+ class LatentRescaler(nn.Module):
666
+ def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2):
667
+ super().__init__()
668
+ # residual block, interpolate, residual block
669
+ self.factor = factor
670
+ self.conv_in = nn.Conv2d(in_channels,
671
+ mid_channels,
672
+ kernel_size=3,
673
+ stride=1,
674
+ padding=1)
675
+ self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
676
+ out_channels=mid_channels,
677
+ temb_channels=0,
678
+ dropout=0.0) for _ in range(depth)])
679
+ self.attn = AttnBlock(mid_channels)
680
+ self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels,
681
+ out_channels=mid_channels,
682
+ temb_channels=0,
683
+ dropout=0.0) for _ in range(depth)])
684
+
685
+ self.conv_out = nn.Conv2d(mid_channels,
686
+ out_channels,
687
+ kernel_size=1,
688
+ )
689
+
690
+ def forward(self, x):
691
+ x = self.conv_in(x)
692
+ for block in self.res_block1:
693
+ x = block(x, None)
694
+ x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor))))
695
+ x = self.attn(x)
696
+ for block in self.res_block2:
697
+ x = block(x, None)
698
+ x = self.conv_out(x)
699
+ return x
700
+
701
+
702
+ class MergedRescaleEncoder(nn.Module):
703
+ def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks,
704
+ attn_resolutions, dropout=0.0, resamp_with_conv=True,
705
+ ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1):
706
+ super().__init__()
707
+ intermediate_chn = ch * ch_mult[-1]
708
+ self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult,
709
+ z_channels=intermediate_chn, double_z=False, resolution=resolution,
710
+ attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv,
711
+ out_ch=None)
712
+ self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn,
713
+ mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth)
714
+
715
+ def forward(self, x):
716
+ x = self.encoder(x)
717
+ x = self.rescaler(x)
718
+ return x
719
+
720
+
721
+ class MergedRescaleDecoder(nn.Module):
722
+ def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8),
723
+ dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1):
724
+ super().__init__()
725
+ tmp_chn = z_channels*ch_mult[-1]
726
+ self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout,
727
+ resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks,
728
+ ch_mult=ch_mult, resolution=resolution, ch=ch)
729
+ self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn,
730
+ out_channels=tmp_chn, depth=rescale_module_depth)
731
+
732
+ def forward(self, x):
733
+ x = self.rescaler(x)
734
+ x = self.decoder(x)
735
+ return x
736
+
737
+
738
+ class Upsampler(nn.Module):
739
+ def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2):
740
+ super().__init__()
741
+ assert out_size >= in_size
742
+ num_blocks = int(np.log2(out_size//in_size))+1
743
+ factor_up = 1.+ (out_size % in_size)
744
+ print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}")
745
+ self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels,
746
+ out_channels=in_channels)
747
+ self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2,
748
+ attn_resolutions=[], in_channels=None, ch=in_channels,
749
+ ch_mult=[ch_mult for _ in range(num_blocks)])
750
+
751
+ def forward(self, x):
752
+ x = self.rescaler(x)
753
+ x = self.decoder(x)
754
+ return x
755
+
756
+
757
+ class Resize(nn.Module):
758
+ def __init__(self, in_channels=None, learned=False, mode="bilinear"):
759
+ super().__init__()
760
+ self.with_conv = learned
761
+ self.mode = mode
762
+ if self.with_conv:
763
+ print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode")
764
+ raise NotImplementedError()
765
+ assert in_channels is not None
766
+ # no asymmetric padding in torch conv, must do it ourselves
767
+ self.conv = torch.nn.Conv2d(in_channels,
768
+ in_channels,
769
+ kernel_size=4,
770
+ stride=2,
771
+ padding=1)
772
+
773
+ def forward(self, x, scale_factor=1.0):
774
+ if scale_factor==1.0:
775
+ return x
776
+ else:
777
+ x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor)
778
+ return x
779
+
780
+ class FirstStagePostProcessor(nn.Module):
781
+
782
+ def __init__(self, ch_mult:list, in_channels,
783
+ pretrained_model:nn.Module=None,
784
+ reshape=False,
785
+ n_channels=None,
786
+ dropout=0.,
787
+ pretrained_config=None):
788
+ super().__init__()
789
+ if pretrained_config is None:
790
+ assert pretrained_model is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
791
+ self.pretrained_model = pretrained_model
792
+ else:
793
+ assert pretrained_config is not None, 'Either "pretrained_model" or "pretrained_config" must not be None'
794
+ self.instantiate_pretrained(pretrained_config)
795
+
796
+ self.do_reshape = reshape
797
+
798
+ if n_channels is None:
799
+ n_channels = self.pretrained_model.encoder.ch
800
+
801
+ self.proj_norm = Normalize(in_channels,num_groups=in_channels//2)
802
+ self.proj = nn.Conv2d(in_channels,n_channels,kernel_size=3,
803
+ stride=1,padding=1)
804
+
805
+ blocks = []
806
+ downs = []
807
+ ch_in = n_channels
808
+ for m in ch_mult:
809
+ blocks.append(ResnetBlock(in_channels=ch_in,out_channels=m*n_channels,dropout=dropout))
810
+ ch_in = m * n_channels
811
+ downs.append(Downsample(ch_in, with_conv=False))
812
+
813
+ self.model = nn.ModuleList(blocks)
814
+ self.downsampler = nn.ModuleList(downs)
815
+
816
+
817
+ def instantiate_pretrained(self, config):
818
+ model = instantiate_from_config(config)
819
+ self.pretrained_model = model.eval()
820
+ # self.pretrained_model.train = False
821
+ for param in self.pretrained_model.parameters():
822
+ param.requires_grad = False
823
+
824
+
825
+ @torch.no_grad()
826
+ def encode_with_pretrained(self,x):
827
+ c = self.pretrained_model.encode(x)
828
+ if isinstance(c, DiagonalGaussianDistribution):
829
+ c = c.mode()
830
+ return c
831
+
832
+ def forward(self,x):
833
+ z_fs = self.encode_with_pretrained(x)
834
+ z = self.proj_norm(z_fs)
835
+ z = self.proj(z)
836
+ z = nonlinearity(z)
837
+
838
+ for submodel, downmodel in zip(self.model,self.downsampler):
839
+ z = submodel(z,temb=None)
840
+ z = downmodel(z)
841
+
842
+ if self.do_reshape:
843
+ z = rearrange(z,'b c h w -> b (h w) c')
844
+ return z
845
+
lvdm/modules/networks/openaimodel3d.py ADDED
@@ -0,0 +1,581 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from functools import partial
2
+ from abc import abstractmethod
3
+ import torch
4
+ import torch.nn as nn
5
+ from einops import rearrange
6
+ import torch.nn.functional as F
7
+ from lvdm.models.utils_diffusion import timestep_embedding
8
+ from lvdm.common import checkpoint
9
+ from lvdm.basics import (
10
+ zero_module,
11
+ conv_nd,
12
+ linear,
13
+ avg_pool_nd,
14
+ normalization
15
+ )
16
+ from lvdm.modules.attention import SpatialTransformer, TemporalTransformer
17
+
18
+
19
+ class TimestepBlock(nn.Module):
20
+ """
21
+ Any module where forward() takes timestep embeddings as a second argument.
22
+ """
23
+ @abstractmethod
24
+ def forward(self, x, emb):
25
+ """
26
+ Apply the module to `x` given `emb` timestep embeddings.
27
+ """
28
+
29
+
30
+ class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
31
+ """
32
+ A sequential module that passes timestep embeddings to the children that
33
+ support it as an extra input.
34
+ """
35
+
36
+ def forward(self, x, emb, context=None, batch_size=None):
37
+ for layer in self:
38
+ if isinstance(layer, TimestepBlock):
39
+ x = layer(x, emb, batch_size)
40
+ elif isinstance(layer, SpatialTransformer):
41
+ x = layer(x, context)
42
+ elif isinstance(layer, TemporalTransformer):
43
+ x = rearrange(x, '(b f) c h w -> b c f h w', b=batch_size)
44
+ x = layer(x, context)
45
+ x = rearrange(x, 'b c f h w -> (b f) c h w')
46
+ else:
47
+ x = layer(x,)
48
+ return x
49
+
50
+
51
+ class Downsample(nn.Module):
52
+ """
53
+ A downsampling layer with an optional convolution.
54
+ :param channels: channels in the inputs and outputs.
55
+ :param use_conv: a bool determining if a convolution is applied.
56
+ :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
57
+ downsampling occurs in the inner-two dimensions.
58
+ """
59
+
60
+ def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
61
+ super().__init__()
62
+ self.channels = channels
63
+ self.out_channels = out_channels or channels
64
+ self.use_conv = use_conv
65
+ self.dims = dims
66
+ stride = 2 if dims != 3 else (1, 2, 2)
67
+ if use_conv:
68
+ self.op = conv_nd(
69
+ dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
70
+ )
71
+ else:
72
+ assert self.channels == self.out_channels
73
+ self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
74
+
75
+ def forward(self, x):
76
+ assert x.shape[1] == self.channels
77
+ return self.op(x)
78
+
79
+
80
+ class Upsample(nn.Module):
81
+ """
82
+ An upsampling layer with an optional convolution.
83
+ :param channels: channels in the inputs and outputs.
84
+ :param use_conv: a bool determining if a convolution is applied.
85
+ :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
86
+ upsampling occurs in the inner-two dimensions.
87
+ """
88
+
89
+ def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
90
+ super().__init__()
91
+ self.channels = channels
92
+ self.out_channels = out_channels or channels
93
+ self.use_conv = use_conv
94
+ self.dims = dims
95
+ if use_conv:
96
+ self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding)
97
+
98
+ def forward(self, x):
99
+ assert x.shape[1] == self.channels
100
+ if self.dims == 3:
101
+ x = F.interpolate(x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode='nearest')
102
+ else:
103
+ x = F.interpolate(x, scale_factor=2, mode='nearest')
104
+ if self.use_conv:
105
+ x = self.conv(x)
106
+ return x
107
+
108
+
109
+ class ResBlock(TimestepBlock):
110
+ """
111
+ A residual block that can optionally change the number of channels.
112
+ :param channels: the number of input channels.
113
+ :param emb_channels: the number of timestep embedding channels.
114
+ :param dropout: the rate of dropout.
115
+ :param out_channels: if specified, the number of out channels.
116
+ :param use_conv: if True and out_channels is specified, use a spatial
117
+ convolution instead of a smaller 1x1 convolution to change the
118
+ channels in the skip connection.
119
+ :param dims: determines if the signal is 1D, 2D, or 3D.
120
+ :param up: if True, use this block for upsampling.
121
+ :param down: if True, use this block for downsampling.
122
+ """
123
+
124
+ def __init__(
125
+ self,
126
+ channels,
127
+ emb_channels,
128
+ dropout,
129
+ out_channels=None,
130
+ use_scale_shift_norm=False,
131
+ dims=2,
132
+ use_checkpoint=False,
133
+ use_conv=False,
134
+ up=False,
135
+ down=False,
136
+ use_temporal_conv=False,
137
+ tempspatial_aware=False
138
+ ):
139
+ super().__init__()
140
+ self.channels = channels
141
+ self.emb_channels = emb_channels
142
+ self.dropout = dropout
143
+ self.out_channels = out_channels or channels
144
+ self.use_conv = use_conv
145
+ self.use_checkpoint = use_checkpoint
146
+ self.use_scale_shift_norm = use_scale_shift_norm
147
+ self.use_temporal_conv = use_temporal_conv
148
+
149
+ self.in_layers = nn.Sequential(
150
+ normalization(channels),
151
+ nn.SiLU(),
152
+ conv_nd(dims, channels, self.out_channels, 3, padding=1),
153
+ )
154
+
155
+ self.updown = up or down
156
+
157
+ if up:
158
+ self.h_upd = Upsample(channels, False, dims)
159
+ self.x_upd = Upsample(channels, False, dims)
160
+ elif down:
161
+ self.h_upd = Downsample(channels, False, dims)
162
+ self.x_upd = Downsample(channels, False, dims)
163
+ else:
164
+ self.h_upd = self.x_upd = nn.Identity()
165
+
166
+ self.emb_layers = nn.Sequential(
167
+ nn.SiLU(),
168
+ nn.Linear(
169
+ emb_channels,
170
+ 2 * self.out_channels if use_scale_shift_norm else self.out_channels,
171
+ ),
172
+ )
173
+ self.out_layers = nn.Sequential(
174
+ normalization(self.out_channels),
175
+ nn.SiLU(),
176
+ nn.Dropout(p=dropout),
177
+ zero_module(nn.Conv2d(self.out_channels, self.out_channels, 3, padding=1)),
178
+ )
179
+
180
+ if self.out_channels == channels:
181
+ self.skip_connection = nn.Identity()
182
+ elif use_conv:
183
+ self.skip_connection = conv_nd(dims, channels, self.out_channels, 3, padding=1)
184
+ else:
185
+ self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
186
+
187
+ if self.use_temporal_conv:
188
+ self.temopral_conv = TemporalConvBlock(
189
+ self.out_channels,
190
+ self.out_channels,
191
+ dropout=0.1,
192
+ spatial_aware=tempspatial_aware
193
+ )
194
+
195
+ def forward(self, x, emb, batch_size=None):
196
+ """
197
+ Apply the block to a Tensor, conditioned on a timestep embedding.
198
+ :param x: an [N x C x ...] Tensor of features.
199
+ :param emb: an [N x emb_channels] Tensor of timestep embeddings.
200
+ :return: an [N x C x ...] Tensor of outputs.
201
+ """
202
+ input_tuple = (x, emb,)
203
+ if batch_size:
204
+ forward_batchsize = partial(self._forward, batch_size=batch_size)
205
+ return checkpoint(forward_batchsize, input_tuple, self.parameters(), self.use_checkpoint)
206
+ return checkpoint(self._forward, input_tuple, self.parameters(), self.use_checkpoint)
207
+
208
+ def _forward(self, x, emb, batch_size=None,):
209
+ if self.updown:
210
+ in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
211
+ h = in_rest(x)
212
+ h = self.h_upd(h)
213
+ x = self.x_upd(x)
214
+ h = in_conv(h)
215
+ else:
216
+ h = self.in_layers(x)
217
+ emb_out = self.emb_layers(emb).type(h.dtype)
218
+ while len(emb_out.shape) < len(h.shape):
219
+ emb_out = emb_out[..., None]
220
+ if self.use_scale_shift_norm:
221
+ out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
222
+ scale, shift = torch.chunk(emb_out, 2, dim=1)
223
+ h = out_norm(h) * (1 + scale) + shift
224
+ h = out_rest(h)
225
+ else:
226
+ h = h + emb_out
227
+ h = self.out_layers(h)
228
+ h = self.skip_connection(x) + h
229
+
230
+ if self.use_temporal_conv and batch_size:
231
+ h = rearrange(h, '(b t) c h w -> b c t h w', b=batch_size)
232
+ h = self.temopral_conv(h)
233
+ h = rearrange(h, 'b c t h w -> (b t) c h w')
234
+ return h
235
+
236
+
237
+ class TemporalConvBlock(nn.Module):
238
+ """
239
+ Adapted from modelscope: https://github.com/modelscope/modelscope/blob/master/modelscope/models/multi_modal/video_synthesis/unet_sd.py
240
+ """
241
+
242
+ def __init__(self, in_channels, out_channels=None, dropout=0.0, spatial_aware=False):
243
+ super(TemporalConvBlock, self).__init__()
244
+ if out_channels is None:
245
+ out_channels = in_channels
246
+ self.in_channels = in_channels
247
+ self.out_channels = out_channels
248
+ kernel_shape = (3, 1, 1) if not spatial_aware else (3, 3, 3)
249
+ padding_shape = (1, 0, 0) if not spatial_aware else (1, 1, 1)
250
+
251
+ # conv layers
252
+ self.conv1 = nn.Sequential(
253
+ nn.GroupNorm(32, in_channels), nn.SiLU(),
254
+ nn.Conv3d(in_channels, out_channels, kernel_shape, padding=padding_shape))
255
+ self.conv2 = nn.Sequential(
256
+ nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout),
257
+ nn.Conv3d(out_channels, in_channels, kernel_shape, padding=padding_shape))
258
+ self.conv3 = nn.Sequential(
259
+ nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout),
260
+ nn.Conv3d(out_channels, in_channels, (3, 1, 1), padding=(1, 0, 0)))
261
+ self.conv4 = nn.Sequential(
262
+ nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout),
263
+ nn.Conv3d(out_channels, in_channels, (3, 1, 1), padding=(1, 0, 0)))
264
+
265
+ # zero out the last layer params,so the conv block is identity
266
+ nn.init.zeros_(self.conv4[-1].weight)
267
+ nn.init.zeros_(self.conv4[-1].bias)
268
+
269
+ def forward(self, x):
270
+ identity = x
271
+ x = self.conv1(x)
272
+ x = self.conv2(x)
273
+ x = self.conv3(x)
274
+ x = self.conv4(x)
275
+
276
+ return x + identity
277
+
278
+
279
+ class UNetModel(nn.Module):
280
+ """
281
+ The full UNet model with attention and timestep embedding.
282
+ :param in_channels: in_channels in the input Tensor.
283
+ :param model_channels: base channel count for the model.
284
+ :param out_channels: channels in the output Tensor.
285
+ :param num_res_blocks: number of residual blocks per downsample.
286
+ :param attention_resolutions: a collection of downsample rates at which
287
+ attention will take place. May be a set, list, or tuple.
288
+ For example, if this contains 4, then at 4x downsampling, attention
289
+ will be used.
290
+ :param dropout: the dropout probability.
291
+ :param channel_mult: channel multiplier for each level of the UNet.
292
+ :param conv_resample: if True, use learned convolutions for upsampling and
293
+ downsampling.
294
+ :param dims: determines if the signal is 1D, 2D, or 3D.
295
+ :param num_classes: if specified (as an int), then this model will be
296
+ class-conditional with `num_classes` classes.
297
+ :param use_checkpoint: use gradient checkpointing to reduce memory usage.
298
+ :param num_heads: the number of attention heads in each attention layer.
299
+ :param num_heads_channels: if specified, ignore num_heads and instead use
300
+ a fixed channel width per attention head.
301
+ :param num_heads_upsample: works with num_heads to set a different number
302
+ of heads for upsampling. Deprecated.
303
+ :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
304
+ :param resblock_updown: use residual blocks for up/downsampling.
305
+ """
306
+
307
+ def __init__(self,
308
+ in_channels,
309
+ model_channels,
310
+ out_channels,
311
+ num_res_blocks,
312
+ attention_resolutions,
313
+ dropout=0.0,
314
+ channel_mult=(1, 2, 4, 8),
315
+ conv_resample=True,
316
+ dims=2,
317
+ context_dim=None,
318
+ use_scale_shift_norm=False,
319
+ resblock_updown=False,
320
+ num_heads=-1,
321
+ num_head_channels=-1,
322
+ transformer_depth=1,
323
+ use_linear=False,
324
+ use_checkpoint=False,
325
+ temporal_conv=False,
326
+ tempspatial_aware=False,
327
+ temporal_attention=True,
328
+ temporal_selfatt_only=True,
329
+ use_relative_position=True,
330
+ use_causal_attention=False,
331
+ temporal_length=None,
332
+ use_fp16=False,
333
+ addition_attention=False,
334
+ use_image_attention=False,
335
+ temporal_transformer_depth=1,
336
+ fps_cond=False,
337
+ ):
338
+ super(UNetModel, self).__init__()
339
+ if num_heads == -1:
340
+ assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
341
+ if num_head_channels == -1:
342
+ assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
343
+
344
+ self.in_channels = in_channels
345
+ self.model_channels = model_channels
346
+ self.out_channels = out_channels
347
+ self.num_res_blocks = num_res_blocks
348
+ self.attention_resolutions = attention_resolutions
349
+ self.dropout = dropout
350
+ self.channel_mult = channel_mult
351
+ self.conv_resample = conv_resample
352
+ self.temporal_attention = temporal_attention
353
+ time_embed_dim = model_channels * 4
354
+ self.use_checkpoint = use_checkpoint
355
+ self.dtype = torch.float16 if use_fp16 else torch.float32
356
+ self.addition_attention=addition_attention
357
+ self.use_image_attention = use_image_attention
358
+ self.fps_cond=fps_cond
359
+
360
+
361
+
362
+ self.time_embed = nn.Sequential(
363
+ linear(model_channels, time_embed_dim),
364
+ nn.SiLU(),
365
+ linear(time_embed_dim, time_embed_dim),
366
+ )
367
+ if self.fps_cond:
368
+ self.fps_embedding = nn.Sequential(
369
+ linear(model_channels, time_embed_dim),
370
+ nn.SiLU(),
371
+ linear(time_embed_dim, time_embed_dim),
372
+ )
373
+
374
+ self.input_blocks = nn.ModuleList(
375
+ [
376
+ TimestepEmbedSequential(conv_nd(dims, in_channels, model_channels, 3, padding=1))
377
+ ]
378
+ )
379
+ if self.addition_attention:
380
+ self.init_attn=TimestepEmbedSequential(
381
+ TemporalTransformer(
382
+ model_channels,
383
+ n_heads=8,
384
+ d_head=num_head_channels,
385
+ depth=transformer_depth,
386
+ context_dim=context_dim,
387
+ use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only,
388
+ causal_attention=use_causal_attention, relative_position=use_relative_position,
389
+ temporal_length=temporal_length))
390
+
391
+ input_block_chans = [model_channels]
392
+ ch = model_channels
393
+ ds = 1
394
+ for level, mult in enumerate(channel_mult):
395
+ for _ in range(num_res_blocks):
396
+ layers = [
397
+ ResBlock(ch, time_embed_dim, dropout,
398
+ out_channels=mult * model_channels, dims=dims, use_checkpoint=use_checkpoint,
399
+ use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
400
+ use_temporal_conv=temporal_conv
401
+ )
402
+ ]
403
+ ch = mult * model_channels
404
+ if ds in attention_resolutions:
405
+ if num_head_channels == -1:
406
+ dim_head = ch // num_heads
407
+ else:
408
+ num_heads = ch // num_head_channels
409
+ dim_head = num_head_channels
410
+ layers.append(
411
+ SpatialTransformer(ch, num_heads, dim_head,
412
+ depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
413
+ use_checkpoint=use_checkpoint, disable_self_attn=False,
414
+ img_cross_attention=self.use_image_attention
415
+ )
416
+ )
417
+ if self.temporal_attention:
418
+ layers.append(
419
+ TemporalTransformer(ch, num_heads, dim_head,
420
+ depth=temporal_transformer_depth, context_dim=context_dim, use_linear=use_linear,
421
+ use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only,
422
+ causal_attention=use_causal_attention, relative_position=use_relative_position,
423
+ temporal_length=temporal_length
424
+ )
425
+ )
426
+ self.input_blocks.append(TimestepEmbedSequential(*layers))
427
+ input_block_chans.append(ch)
428
+ if level != len(channel_mult) - 1:
429
+ out_ch = ch
430
+ self.input_blocks.append(
431
+ TimestepEmbedSequential(
432
+ ResBlock(ch, time_embed_dim, dropout,
433
+ out_channels=out_ch, dims=dims, use_checkpoint=use_checkpoint,
434
+ use_scale_shift_norm=use_scale_shift_norm,
435
+ down=True
436
+ )
437
+ if resblock_updown
438
+ else Downsample(ch, conv_resample, dims=dims, out_channels=out_ch)
439
+ )
440
+ )
441
+ ch = out_ch
442
+ input_block_chans.append(ch)
443
+ ds *= 2
444
+
445
+ if num_head_channels == -1:
446
+ dim_head = ch // num_heads
447
+ else:
448
+ num_heads = ch // num_head_channels
449
+ dim_head = num_head_channels
450
+ layers = [
451
+ ResBlock(ch, time_embed_dim, dropout,
452
+ dims=dims, use_checkpoint=use_checkpoint,
453
+ use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
454
+ use_temporal_conv=temporal_conv
455
+ ),
456
+ SpatialTransformer(ch, num_heads, dim_head,
457
+ depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
458
+ use_checkpoint=use_checkpoint, disable_self_attn=False,
459
+ img_cross_attention=self.use_image_attention
460
+ )
461
+ ]
462
+ if self.temporal_attention:
463
+ layers.append(
464
+ TemporalTransformer(ch, num_heads, dim_head,
465
+ depth=temporal_transformer_depth, context_dim=context_dim, use_linear=use_linear,
466
+ use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only,
467
+ causal_attention=use_causal_attention, relative_position=use_relative_position,
468
+ temporal_length=temporal_length
469
+ )
470
+ )
471
+ layers.append(
472
+ ResBlock(ch, time_embed_dim, dropout,
473
+ dims=dims, use_checkpoint=use_checkpoint,
474
+ use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
475
+ use_temporal_conv=temporal_conv
476
+ )
477
+ )
478
+ self.middle_block = TimestepEmbedSequential(*layers)
479
+
480
+ self.output_blocks = nn.ModuleList([])
481
+ for level, mult in list(enumerate(channel_mult))[::-1]:
482
+ for i in range(num_res_blocks + 1):
483
+ ich = input_block_chans.pop()
484
+ layers = [
485
+ ResBlock(ch + ich, time_embed_dim, dropout,
486
+ out_channels=mult * model_channels, dims=dims, use_checkpoint=use_checkpoint,
487
+ use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
488
+ use_temporal_conv=temporal_conv
489
+ )
490
+ ]
491
+ ch = model_channels * mult
492
+ if ds in attention_resolutions:
493
+ if num_head_channels == -1:
494
+ dim_head = ch // num_heads
495
+ else:
496
+ num_heads = ch // num_head_channels
497
+ dim_head = num_head_channels
498
+ layers.append(
499
+ SpatialTransformer(ch, num_heads, dim_head,
500
+ depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
501
+ use_checkpoint=use_checkpoint, disable_self_attn=False,
502
+ img_cross_attention=self.use_image_attention
503
+ )
504
+ )
505
+ if self.temporal_attention:
506
+ layers.append(
507
+ TemporalTransformer(ch, num_heads, dim_head,
508
+ depth=temporal_transformer_depth, context_dim=context_dim, use_linear=use_linear,
509
+ use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only,
510
+ causal_attention=use_causal_attention, relative_position=use_relative_position,
511
+ temporal_length=temporal_length
512
+ )
513
+ )
514
+ if level and i == num_res_blocks:
515
+ out_ch = ch
516
+ layers.append(
517
+ ResBlock(ch, time_embed_dim, dropout,
518
+ out_channels=out_ch, dims=dims, use_checkpoint=use_checkpoint,
519
+ use_scale_shift_norm=use_scale_shift_norm,
520
+ up=True
521
+ )
522
+ if resblock_updown
523
+ else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
524
+ )
525
+ ds //= 2
526
+ self.output_blocks.append(TimestepEmbedSequential(*layers))
527
+
528
+ self.out = nn.Sequential(
529
+ normalization(ch),
530
+ nn.SiLU(),
531
+ zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
532
+ )
533
+
534
+ def forward(self, x, timesteps, context=None, features_adapter=None, fps=16, **kwargs):
535
+ t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
536
+ emb = self.time_embed(t_emb)
537
+
538
+ if self.fps_cond:
539
+ if type(fps) == int:
540
+ fps = torch.full_like(timesteps, fps)
541
+ fps_emb = timestep_embedding(fps,self.model_channels, repeat_only=False)
542
+ emb += self.fps_embedding(fps_emb)
543
+
544
+ b,_,t,_,_ = x.shape
545
+ ## repeat t times for context [(b t) 77 768] & time embedding
546
+ if len(context.shape) < 4:
547
+ context = context.repeat_interleave(repeats=t, dim=0)
548
+ else:
549
+ context = context.view(-1, context.shape[2], context.shape[3])
550
+ # context = context.repeat_interleave(repeats=t, dim=0)
551
+ emb = emb.repeat_interleave(repeats=t, dim=0)
552
+
553
+ ## always in shape (b t) c h w, except for temporal layer
554
+ x = rearrange(x, 'b c t h w -> (b t) c h w')
555
+
556
+ h = x.type(self.dtype)
557
+ adapter_idx = 0
558
+ hs = []
559
+ for id, module in enumerate(self.input_blocks):
560
+ h = module(h, emb, context=context, batch_size=b)
561
+ if id ==0 and self.addition_attention:
562
+ h = self.init_attn(h, emb, context=context, batch_size=b)
563
+ ## plug-in adapter features
564
+ if ((id+1)%3 == 0) and features_adapter is not None:
565
+ h = h + features_adapter[adapter_idx]
566
+ adapter_idx += 1
567
+ hs.append(h)
568
+ if features_adapter is not None:
569
+ assert len(features_adapter)==adapter_idx, 'Wrong features_adapter'
570
+
571
+ h = self.middle_block(h, emb, context=context, batch_size=b)
572
+ for module in self.output_blocks:
573
+ h = torch.cat([h, hs.pop()], dim=1)
574
+ h = module(h, emb, context=context, batch_size=b)
575
+ h = h.type(x.dtype)
576
+ y = self.out(h)
577
+
578
+ # reshape back to (b c t h w)
579
+ y = rearrange(y, '(b t) c h w -> b c t h w', b=b)
580
+ return y
581
+
lvdm/modules/networks/openaimodel3d_freenoise.py ADDED
@@ -0,0 +1,581 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from functools import partial
2
+ from abc import abstractmethod
3
+ import torch
4
+ import torch.nn as nn
5
+ from einops import rearrange
6
+ import torch.nn.functional as F
7
+ from lvdm.models.utils_diffusion import timestep_embedding
8
+ from lvdm.common import checkpoint
9
+ from lvdm.basics import (
10
+ zero_module,
11
+ conv_nd,
12
+ linear,
13
+ avg_pool_nd,
14
+ normalization
15
+ )
16
+ from lvdm.modules.attention_freenoise import SpatialTransformer, TemporalTransformer
17
+
18
+
19
+ class TimestepBlock(nn.Module):
20
+ """
21
+ Any module where forward() takes timestep embeddings as a second argument.
22
+ """
23
+ @abstractmethod
24
+ def forward(self, x, emb):
25
+ """
26
+ Apply the module to `x` given `emb` timestep embeddings.
27
+ """
28
+
29
+
30
+ class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
31
+ """
32
+ A sequential module that passes timestep embeddings to the children that
33
+ support it as an extra input.
34
+ """
35
+
36
+ def forward(self, x, emb, context=None, batch_size=None, use_injection=False, **kwargs):
37
+ for layer in self:
38
+ if isinstance(layer, TimestepBlock):
39
+ x = layer(x, emb, batch_size)
40
+ elif isinstance(layer, SpatialTransformer):
41
+ x = layer(x, context, use_injection=use_injection)
42
+ elif isinstance(layer, TemporalTransformer):
43
+ x = rearrange(x, '(b f) c h w -> b c f h w', b=batch_size)
44
+ x = layer(x, context, **kwargs)
45
+ x = rearrange(x, 'b c f h w -> (b f) c h w')
46
+ else:
47
+ x = layer(x,)
48
+ return x
49
+
50
+
51
+ class Downsample(nn.Module):
52
+ """
53
+ A downsampling layer with an optional convolution.
54
+ :param channels: channels in the inputs and outputs.
55
+ :param use_conv: a bool determining if a convolution is applied.
56
+ :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
57
+ downsampling occurs in the inner-two dimensions.
58
+ """
59
+
60
+ def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
61
+ super().__init__()
62
+ self.channels = channels
63
+ self.out_channels = out_channels or channels
64
+ self.use_conv = use_conv
65
+ self.dims = dims
66
+ stride = 2 if dims != 3 else (1, 2, 2)
67
+ if use_conv:
68
+ self.op = conv_nd(
69
+ dims, self.channels, self.out_channels, 3, stride=stride, padding=padding
70
+ )
71
+ else:
72
+ assert self.channels == self.out_channels
73
+ self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride)
74
+
75
+ def forward(self, x):
76
+ assert x.shape[1] == self.channels
77
+ return self.op(x)
78
+
79
+
80
+ class Upsample(nn.Module):
81
+ """
82
+ An upsampling layer with an optional convolution.
83
+ :param channels: channels in the inputs and outputs.
84
+ :param use_conv: a bool determining if a convolution is applied.
85
+ :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then
86
+ upsampling occurs in the inner-two dimensions.
87
+ """
88
+
89
+ def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1):
90
+ super().__init__()
91
+ self.channels = channels
92
+ self.out_channels = out_channels or channels
93
+ self.use_conv = use_conv
94
+ self.dims = dims
95
+ if use_conv:
96
+ self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding)
97
+
98
+ def forward(self, x):
99
+ assert x.shape[1] == self.channels
100
+ if self.dims == 3:
101
+ x = F.interpolate(x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode='nearest')
102
+ else:
103
+ x = F.interpolate(x, scale_factor=2, mode='nearest')
104
+ if self.use_conv:
105
+ x = self.conv(x)
106
+ return x
107
+
108
+
109
+ class ResBlock(TimestepBlock):
110
+ """
111
+ A residual block that can optionally change the number of channels.
112
+ :param channels: the number of input channels.
113
+ :param emb_channels: the number of timestep embedding channels.
114
+ :param dropout: the rate of dropout.
115
+ :param out_channels: if specified, the number of out channels.
116
+ :param use_conv: if True and out_channels is specified, use a spatial
117
+ convolution instead of a smaller 1x1 convolution to change the
118
+ channels in the skip connection.
119
+ :param dims: determines if the signal is 1D, 2D, or 3D.
120
+ :param up: if True, use this block for upsampling.
121
+ :param down: if True, use this block for downsampling.
122
+ """
123
+
124
+ def __init__(
125
+ self,
126
+ channels,
127
+ emb_channels,
128
+ dropout,
129
+ out_channels=None,
130
+ use_scale_shift_norm=False,
131
+ dims=2,
132
+ use_checkpoint=False,
133
+ use_conv=False,
134
+ up=False,
135
+ down=False,
136
+ use_temporal_conv=False,
137
+ tempspatial_aware=False
138
+ ):
139
+ super().__init__()
140
+ self.channels = channels
141
+ self.emb_channels = emb_channels
142
+ self.dropout = dropout
143
+ self.out_channels = out_channels or channels
144
+ self.use_conv = use_conv
145
+ self.use_checkpoint = use_checkpoint
146
+ self.use_scale_shift_norm = use_scale_shift_norm
147
+ self.use_temporal_conv = use_temporal_conv
148
+
149
+ self.in_layers = nn.Sequential(
150
+ normalization(channels),
151
+ nn.SiLU(),
152
+ conv_nd(dims, channels, self.out_channels, 3, padding=1),
153
+ )
154
+
155
+ self.updown = up or down
156
+
157
+ if up:
158
+ self.h_upd = Upsample(channels, False, dims)
159
+ self.x_upd = Upsample(channels, False, dims)
160
+ elif down:
161
+ self.h_upd = Downsample(channels, False, dims)
162
+ self.x_upd = Downsample(channels, False, dims)
163
+ else:
164
+ self.h_upd = self.x_upd = nn.Identity()
165
+
166
+ self.emb_layers = nn.Sequential(
167
+ nn.SiLU(),
168
+ nn.Linear(
169
+ emb_channels,
170
+ 2 * self.out_channels if use_scale_shift_norm else self.out_channels,
171
+ ),
172
+ )
173
+ self.out_layers = nn.Sequential(
174
+ normalization(self.out_channels),
175
+ nn.SiLU(),
176
+ nn.Dropout(p=dropout),
177
+ zero_module(nn.Conv2d(self.out_channels, self.out_channels, 3, padding=1)),
178
+ )
179
+
180
+ if self.out_channels == channels:
181
+ self.skip_connection = nn.Identity()
182
+ elif use_conv:
183
+ self.skip_connection = conv_nd(dims, channels, self.out_channels, 3, padding=1)
184
+ else:
185
+ self.skip_connection = conv_nd(dims, channels, self.out_channels, 1)
186
+
187
+ if self.use_temporal_conv:
188
+ self.temopral_conv = TemporalConvBlock(
189
+ self.out_channels,
190
+ self.out_channels,
191
+ dropout=0.1,
192
+ spatial_aware=tempspatial_aware
193
+ )
194
+
195
+ def forward(self, x, emb, batch_size=None):
196
+ """
197
+ Apply the block to a Tensor, conditioned on a timestep embedding.
198
+ :param x: an [N x C x ...] Tensor of features.
199
+ :param emb: an [N x emb_channels] Tensor of timestep embeddings.
200
+ :return: an [N x C x ...] Tensor of outputs.
201
+ """
202
+ input_tuple = (x, emb,)
203
+ if batch_size:
204
+ forward_batchsize = partial(self._forward, batch_size=batch_size)
205
+ return checkpoint(forward_batchsize, input_tuple, self.parameters(), self.use_checkpoint)
206
+ return checkpoint(self._forward, input_tuple, self.parameters(), self.use_checkpoint)
207
+
208
+ def _forward(self, x, emb, batch_size=None,):
209
+ if self.updown:
210
+ in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
211
+ h = in_rest(x)
212
+ h = self.h_upd(h)
213
+ x = self.x_upd(x)
214
+ h = in_conv(h)
215
+ else:
216
+ h = self.in_layers(x)
217
+ emb_out = self.emb_layers(emb).type(h.dtype)
218
+ while len(emb_out.shape) < len(h.shape):
219
+ emb_out = emb_out[..., None]
220
+ if self.use_scale_shift_norm:
221
+ out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
222
+ scale, shift = torch.chunk(emb_out, 2, dim=1)
223
+ h = out_norm(h) * (1 + scale) + shift
224
+ h = out_rest(h)
225
+ else:
226
+ h = h + emb_out
227
+ h = self.out_layers(h)
228
+ h = self.skip_connection(x) + h
229
+
230
+ if self.use_temporal_conv and batch_size:
231
+ h = rearrange(h, '(b t) c h w -> b c t h w', b=batch_size)
232
+ h = self.temopral_conv(h)
233
+ h = rearrange(h, 'b c t h w -> (b t) c h w')
234
+ return h
235
+
236
+
237
+ class TemporalConvBlock(nn.Module):
238
+ """
239
+ Adapted from modelscope: https://github.com/modelscope/modelscope/blob/master/modelscope/models/multi_modal/video_synthesis/unet_sd.py
240
+ """
241
+
242
+ def __init__(self, in_channels, out_channels=None, dropout=0.0, spatial_aware=False):
243
+ super(TemporalConvBlock, self).__init__()
244
+ if out_channels is None:
245
+ out_channels = in_channels
246
+ self.in_channels = in_channels
247
+ self.out_channels = out_channels
248
+ kernel_shape = (3, 1, 1) if not spatial_aware else (3, 3, 3)
249
+ padding_shape = (1, 0, 0) if not spatial_aware else (1, 1, 1)
250
+
251
+ # conv layers
252
+ self.conv1 = nn.Sequential(
253
+ nn.GroupNorm(32, in_channels), nn.SiLU(),
254
+ nn.Conv3d(in_channels, out_channels, kernel_shape, padding=padding_shape))
255
+ self.conv2 = nn.Sequential(
256
+ nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout),
257
+ nn.Conv3d(out_channels, in_channels, kernel_shape, padding=padding_shape))
258
+ self.conv3 = nn.Sequential(
259
+ nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout),
260
+ nn.Conv3d(out_channels, in_channels, (3, 1, 1), padding=(1, 0, 0)))
261
+ self.conv4 = nn.Sequential(
262
+ nn.GroupNorm(32, out_channels), nn.SiLU(), nn.Dropout(dropout),
263
+ nn.Conv3d(out_channels, in_channels, (3, 1, 1), padding=(1, 0, 0)))
264
+
265
+ # zero out the last layer params,so the conv block is identity
266
+ nn.init.zeros_(self.conv4[-1].weight)
267
+ nn.init.zeros_(self.conv4[-1].bias)
268
+
269
+ def forward(self, x):
270
+ identity = x
271
+ x = self.conv1(x)
272
+ x = self.conv2(x)
273
+ x = self.conv3(x)
274
+ x = self.conv4(x)
275
+
276
+ return x + identity
277
+
278
+
279
+ class UNetModel(nn.Module):
280
+ """
281
+ The full UNet model with attention and timestep embedding.
282
+ :param in_channels: in_channels in the input Tensor.
283
+ :param model_channels: base channel count for the model.
284
+ :param out_channels: channels in the output Tensor.
285
+ :param num_res_blocks: number of residual blocks per downsample.
286
+ :param attention_resolutions: a collection of downsample rates at which
287
+ attention will take place. May be a set, list, or tuple.
288
+ For example, if this contains 4, then at 4x downsampling, attention
289
+ will be used.
290
+ :param dropout: the dropout probability.
291
+ :param channel_mult: channel multiplier for each level of the UNet.
292
+ :param conv_resample: if True, use learned convolutions for upsampling and
293
+ downsampling.
294
+ :param dims: determines if the signal is 1D, 2D, or 3D.
295
+ :param num_classes: if specified (as an int), then this model will be
296
+ class-conditional with `num_classes` classes.
297
+ :param use_checkpoint: use gradient checkpointing to reduce memory usage.
298
+ :param num_heads: the number of attention heads in each attention layer.
299
+ :param num_heads_channels: if specified, ignore num_heads and instead use
300
+ a fixed channel width per attention head.
301
+ :param num_heads_upsample: works with num_heads to set a different number
302
+ of heads for upsampling. Deprecated.
303
+ :param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
304
+ :param resblock_updown: use residual blocks for up/downsampling.
305
+ """
306
+
307
+ def __init__(self,
308
+ in_channels,
309
+ model_channels,
310
+ out_channels,
311
+ num_res_blocks,
312
+ attention_resolutions,
313
+ dropout=0.0,
314
+ channel_mult=(1, 2, 4, 8),
315
+ conv_resample=True,
316
+ dims=2,
317
+ context_dim=None,
318
+ use_scale_shift_norm=False,
319
+ resblock_updown=False,
320
+ num_heads=-1,
321
+ num_head_channels=-1,
322
+ transformer_depth=1,
323
+ use_linear=False,
324
+ use_checkpoint=False,
325
+ temporal_conv=False,
326
+ tempspatial_aware=False,
327
+ temporal_attention=True,
328
+ temporal_selfatt_only=True,
329
+ use_relative_position=True,
330
+ use_causal_attention=False,
331
+ temporal_length=None,
332
+ use_fp16=False,
333
+ addition_attention=False,
334
+ use_image_attention=False,
335
+ temporal_transformer_depth=1,
336
+ fps_cond=False,
337
+ ):
338
+ super(UNetModel, self).__init__()
339
+ if num_heads == -1:
340
+ assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
341
+ if num_head_channels == -1:
342
+ assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
343
+
344
+ self.in_channels = in_channels
345
+ self.model_channels = model_channels
346
+ self.out_channels = out_channels
347
+ self.num_res_blocks = num_res_blocks
348
+ self.attention_resolutions = attention_resolutions
349
+ self.dropout = dropout
350
+ self.channel_mult = channel_mult
351
+ self.conv_resample = conv_resample
352
+ self.temporal_attention = temporal_attention
353
+ time_embed_dim = model_channels * 4
354
+ self.use_checkpoint = use_checkpoint
355
+ self.dtype = torch.float16 if use_fp16 else torch.float32
356
+ self.addition_attention=addition_attention
357
+ self.use_image_attention = use_image_attention
358
+ self.fps_cond=fps_cond
359
+
360
+
361
+
362
+ self.time_embed = nn.Sequential(
363
+ linear(model_channels, time_embed_dim),
364
+ nn.SiLU(),
365
+ linear(time_embed_dim, time_embed_dim),
366
+ )
367
+ if self.fps_cond:
368
+ self.fps_embedding = nn.Sequential(
369
+ linear(model_channels, time_embed_dim),
370
+ nn.SiLU(),
371
+ linear(time_embed_dim, time_embed_dim),
372
+ )
373
+
374
+ self.input_blocks = nn.ModuleList(
375
+ [
376
+ TimestepEmbedSequential(conv_nd(dims, in_channels, model_channels, 3, padding=1))
377
+ ]
378
+ )
379
+ if self.addition_attention:
380
+ self.init_attn=TimestepEmbedSequential(
381
+ TemporalTransformer(
382
+ model_channels,
383
+ n_heads=8,
384
+ d_head=num_head_channels,
385
+ depth=transformer_depth,
386
+ context_dim=context_dim,
387
+ use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only,
388
+ causal_attention=use_causal_attention, relative_position=use_relative_position,
389
+ temporal_length=temporal_length))
390
+
391
+ input_block_chans = [model_channels]
392
+ ch = model_channels
393
+ ds = 1
394
+ for level, mult in enumerate(channel_mult):
395
+ for _ in range(num_res_blocks):
396
+ layers = [
397
+ ResBlock(ch, time_embed_dim, dropout,
398
+ out_channels=mult * model_channels, dims=dims, use_checkpoint=use_checkpoint,
399
+ use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
400
+ use_temporal_conv=temporal_conv
401
+ )
402
+ ]
403
+ ch = mult * model_channels
404
+ if ds in attention_resolutions:
405
+ if num_head_channels == -1:
406
+ dim_head = ch // num_heads
407
+ else:
408
+ num_heads = ch // num_head_channels
409
+ dim_head = num_head_channels
410
+ layers.append(
411
+ SpatialTransformer(ch, num_heads, dim_head,
412
+ depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
413
+ use_checkpoint=use_checkpoint, disable_self_attn=False,
414
+ img_cross_attention=self.use_image_attention, injection=True
415
+ )
416
+ )
417
+ if self.temporal_attention:
418
+ layers.append(
419
+ TemporalTransformer(ch, num_heads, dim_head,
420
+ depth=temporal_transformer_depth, context_dim=context_dim, use_linear=use_linear,
421
+ use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only,
422
+ causal_attention=use_causal_attention, relative_position=use_relative_position,
423
+ temporal_length=temporal_length
424
+ )
425
+ )
426
+ self.input_blocks.append(TimestepEmbedSequential(*layers))
427
+ input_block_chans.append(ch)
428
+ if level != len(channel_mult) - 1:
429
+ out_ch = ch
430
+ self.input_blocks.append(
431
+ TimestepEmbedSequential(
432
+ ResBlock(ch, time_embed_dim, dropout,
433
+ out_channels=out_ch, dims=dims, use_checkpoint=use_checkpoint,
434
+ use_scale_shift_norm=use_scale_shift_norm,
435
+ down=True
436
+ )
437
+ if resblock_updown
438
+ else Downsample(ch, conv_resample, dims=dims, out_channels=out_ch)
439
+ )
440
+ )
441
+ ch = out_ch
442
+ input_block_chans.append(ch)
443
+ ds *= 2
444
+
445
+ if num_head_channels == -1:
446
+ dim_head = ch // num_heads
447
+ else:
448
+ num_heads = ch // num_head_channels
449
+ dim_head = num_head_channels
450
+ layers = [
451
+ ResBlock(ch, time_embed_dim, dropout,
452
+ dims=dims, use_checkpoint=use_checkpoint,
453
+ use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
454
+ use_temporal_conv=temporal_conv
455
+ ),
456
+ SpatialTransformer(ch, num_heads, dim_head,
457
+ depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
458
+ use_checkpoint=use_checkpoint, disable_self_attn=False,
459
+ img_cross_attention=self.use_image_attention, injection=True
460
+ )
461
+ ]
462
+ if self.temporal_attention:
463
+ layers.append(
464
+ TemporalTransformer(ch, num_heads, dim_head,
465
+ depth=temporal_transformer_depth, context_dim=context_dim, use_linear=use_linear,
466
+ use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only,
467
+ causal_attention=use_causal_attention, relative_position=use_relative_position,
468
+ temporal_length=temporal_length
469
+ )
470
+ )
471
+ layers.append(
472
+ ResBlock(ch, time_embed_dim, dropout,
473
+ dims=dims, use_checkpoint=use_checkpoint,
474
+ use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
475
+ use_temporal_conv=temporal_conv
476
+ )
477
+ )
478
+ self.middle_block = TimestepEmbedSequential(*layers)
479
+
480
+ self.output_blocks = nn.ModuleList([])
481
+ for level, mult in list(enumerate(channel_mult))[::-1]:
482
+ for i in range(num_res_blocks + 1):
483
+ ich = input_block_chans.pop()
484
+ layers = [
485
+ ResBlock(ch + ich, time_embed_dim, dropout,
486
+ out_channels=mult * model_channels, dims=dims, use_checkpoint=use_checkpoint,
487
+ use_scale_shift_norm=use_scale_shift_norm, tempspatial_aware=tempspatial_aware,
488
+ use_temporal_conv=temporal_conv
489
+ )
490
+ ]
491
+ ch = model_channels * mult
492
+ if ds in attention_resolutions:
493
+ if num_head_channels == -1:
494
+ dim_head = ch // num_heads
495
+ else:
496
+ num_heads = ch // num_head_channels
497
+ dim_head = num_head_channels
498
+ layers.append(
499
+ SpatialTransformer(ch, num_heads, dim_head,
500
+ depth=transformer_depth, context_dim=context_dim, use_linear=use_linear,
501
+ use_checkpoint=use_checkpoint, disable_self_attn=False,
502
+ img_cross_attention=self.use_image_attention, injection=False
503
+ )
504
+ )
505
+ if self.temporal_attention:
506
+ layers.append(
507
+ TemporalTransformer(ch, num_heads, dim_head,
508
+ depth=temporal_transformer_depth, context_dim=context_dim, use_linear=use_linear,
509
+ use_checkpoint=use_checkpoint, only_self_att=temporal_selfatt_only,
510
+ causal_attention=use_causal_attention, relative_position=use_relative_position,
511
+ temporal_length=temporal_length
512
+ )
513
+ )
514
+ if level and i == num_res_blocks:
515
+ out_ch = ch
516
+ layers.append(
517
+ ResBlock(ch, time_embed_dim, dropout,
518
+ out_channels=out_ch, dims=dims, use_checkpoint=use_checkpoint,
519
+ use_scale_shift_norm=use_scale_shift_norm,
520
+ up=True
521
+ )
522
+ if resblock_updown
523
+ else Upsample(ch, conv_resample, dims=dims, out_channels=out_ch)
524
+ )
525
+ ds //= 2
526
+ self.output_blocks.append(TimestepEmbedSequential(*layers))
527
+
528
+ self.out = nn.Sequential(
529
+ normalization(ch),
530
+ nn.SiLU(),
531
+ zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)),
532
+ )
533
+
534
+ def forward(self, x, timesteps, context=None, features_adapter=None, fps=16, **kwargs):
535
+ t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
536
+ emb = self.time_embed(t_emb)
537
+
538
+ if self.fps_cond:
539
+ if type(fps) == int:
540
+ fps = torch.full_like(timesteps, fps)
541
+ fps_emb = timestep_embedding(fps,self.model_channels, repeat_only=False)
542
+ emb += self.fps_embedding(fps_emb)
543
+
544
+ b,_,t,_,_ = x.shape
545
+ ## repeat t times for context [(b t) 77 768] & time embedding
546
+ if len(context.shape) < 4:
547
+ context = context.repeat_interleave(repeats=t, dim=0)
548
+ else:
549
+ context = context.view(-1, context.shape[2], context.shape[3])
550
+ # context = context.repeat_interleave(repeats=t, dim=0)
551
+ emb = emb.repeat_interleave(repeats=t, dim=0)
552
+
553
+ ## always in shape (b t) c h w, except for temporal layer
554
+ x = rearrange(x, 'b c t h w -> (b t) c h w')
555
+
556
+ h = x.type(self.dtype)
557
+ adapter_idx = 0
558
+ hs = []
559
+ for id, module in enumerate(self.input_blocks):
560
+ h = module(h, emb, context=context, batch_size=b, **kwargs)
561
+ if id ==0 and self.addition_attention:
562
+ h = self.init_attn(h, emb, context=context, batch_size=b, **kwargs)
563
+ ## plug-in adapter features
564
+ if ((id+1)%3 == 0) and features_adapter is not None:
565
+ h = h + features_adapter[adapter_idx]
566
+ adapter_idx += 1
567
+ hs.append(h)
568
+ if features_adapter is not None:
569
+ assert len(features_adapter)==adapter_idx, 'Wrong features_adapter'
570
+
571
+ h = self.middle_block(h, emb, context=context, batch_size=b, **kwargs)
572
+ for module in self.output_blocks:
573
+ h = torch.cat([h, hs.pop()], dim=1)
574
+ h = module(h, emb, context=context, batch_size=b, **kwargs)
575
+ h = h.type(x.dtype)
576
+ y = self.out(h)
577
+
578
+ # reshape back to (b c t h w)
579
+ y = rearrange(y, '(b t) c h w -> b c t h w', b=b)
580
+ return y
581
+
lvdm/modules/x_transformer.py ADDED
@@ -0,0 +1,640 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """shout-out to https://github.com/lucidrains/x-transformers/tree/main/x_transformers"""
2
+ from functools import partial
3
+ from inspect import isfunction
4
+ from collections import namedtuple
5
+ from einops import rearrange, repeat
6
+ import torch
7
+ from torch import nn, einsum
8
+ import torch.nn.functional as F
9
+
10
+ # constants
11
+ DEFAULT_DIM_HEAD = 64
12
+
13
+ Intermediates = namedtuple('Intermediates', [
14
+ 'pre_softmax_attn',
15
+ 'post_softmax_attn'
16
+ ])
17
+
18
+ LayerIntermediates = namedtuple('Intermediates', [
19
+ 'hiddens',
20
+ 'attn_intermediates'
21
+ ])
22
+
23
+
24
+ class AbsolutePositionalEmbedding(nn.Module):
25
+ def __init__(self, dim, max_seq_len):
26
+ super().__init__()
27
+ self.emb = nn.Embedding(max_seq_len, dim)
28
+ self.init_()
29
+
30
+ def init_(self):
31
+ nn.init.normal_(self.emb.weight, std=0.02)
32
+
33
+ def forward(self, x):
34
+ n = torch.arange(x.shape[1], device=x.device)
35
+ return self.emb(n)[None, :, :]
36
+
37
+
38
+ class FixedPositionalEmbedding(nn.Module):
39
+ def __init__(self, dim):
40
+ super().__init__()
41
+ inv_freq = 1. / (10000 ** (torch.arange(0, dim, 2).float() / dim))
42
+ self.register_buffer('inv_freq', inv_freq)
43
+
44
+ def forward(self, x, seq_dim=1, offset=0):
45
+ t = torch.arange(x.shape[seq_dim], device=x.device).type_as(self.inv_freq) + offset
46
+ sinusoid_inp = torch.einsum('i , j -> i j', t, self.inv_freq)
47
+ emb = torch.cat((sinusoid_inp.sin(), sinusoid_inp.cos()), dim=-1)
48
+ return emb[None, :, :]
49
+
50
+
51
+ # helpers
52
+
53
+ def exists(val):
54
+ return val is not None
55
+
56
+
57
+ def default(val, d):
58
+ if exists(val):
59
+ return val
60
+ return d() if isfunction(d) else d
61
+
62
+
63
+ def always(val):
64
+ def inner(*args, **kwargs):
65
+ return val
66
+ return inner
67
+
68
+
69
+ def not_equals(val):
70
+ def inner(x):
71
+ return x != val
72
+ return inner
73
+
74
+
75
+ def equals(val):
76
+ def inner(x):
77
+ return x == val
78
+ return inner
79
+
80
+
81
+ def max_neg_value(tensor):
82
+ return -torch.finfo(tensor.dtype).max
83
+
84
+
85
+ # keyword argument helpers
86
+
87
+ def pick_and_pop(keys, d):
88
+ values = list(map(lambda key: d.pop(key), keys))
89
+ return dict(zip(keys, values))
90
+
91
+
92
+ def group_dict_by_key(cond, d):
93
+ return_val = [dict(), dict()]
94
+ for key in d.keys():
95
+ match = bool(cond(key))
96
+ ind = int(not match)
97
+ return_val[ind][key] = d[key]
98
+ return (*return_val,)
99
+
100
+
101
+ def string_begins_with(prefix, str):
102
+ return str.startswith(prefix)
103
+
104
+
105
+ def group_by_key_prefix(prefix, d):
106
+ return group_dict_by_key(partial(string_begins_with, prefix), d)
107
+
108
+
109
+ def groupby_prefix_and_trim(prefix, d):
110
+ kwargs_with_prefix, kwargs = group_dict_by_key(partial(string_begins_with, prefix), d)
111
+ kwargs_without_prefix = dict(map(lambda x: (x[0][len(prefix):], x[1]), tuple(kwargs_with_prefix.items())))
112
+ return kwargs_without_prefix, kwargs
113
+
114
+
115
+ # classes
116
+ class Scale(nn.Module):
117
+ def __init__(self, value, fn):
118
+ super().__init__()
119
+ self.value = value
120
+ self.fn = fn
121
+
122
+ def forward(self, x, **kwargs):
123
+ x, *rest = self.fn(x, **kwargs)
124
+ return (x * self.value, *rest)
125
+
126
+
127
+ class Rezero(nn.Module):
128
+ def __init__(self, fn):
129
+ super().__init__()
130
+ self.fn = fn
131
+ self.g = nn.Parameter(torch.zeros(1))
132
+
133
+ def forward(self, x, **kwargs):
134
+ x, *rest = self.fn(x, **kwargs)
135
+ return (x * self.g, *rest)
136
+
137
+
138
+ class ScaleNorm(nn.Module):
139
+ def __init__(self, dim, eps=1e-5):
140
+ super().__init__()
141
+ self.scale = dim ** -0.5
142
+ self.eps = eps
143
+ self.g = nn.Parameter(torch.ones(1))
144
+
145
+ def forward(self, x):
146
+ norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
147
+ return x / norm.clamp(min=self.eps) * self.g
148
+
149
+
150
+ class RMSNorm(nn.Module):
151
+ def __init__(self, dim, eps=1e-8):
152
+ super().__init__()
153
+ self.scale = dim ** -0.5
154
+ self.eps = eps
155
+ self.g = nn.Parameter(torch.ones(dim))
156
+
157
+ def forward(self, x):
158
+ norm = torch.norm(x, dim=-1, keepdim=True) * self.scale
159
+ return x / norm.clamp(min=self.eps) * self.g
160
+
161
+
162
+ class Residual(nn.Module):
163
+ def forward(self, x, residual):
164
+ return x + residual
165
+
166
+
167
+ class GRUGating(nn.Module):
168
+ def __init__(self, dim):
169
+ super().__init__()
170
+ self.gru = nn.GRUCell(dim, dim)
171
+
172
+ def forward(self, x, residual):
173
+ gated_output = self.gru(
174
+ rearrange(x, 'b n d -> (b n) d'),
175
+ rearrange(residual, 'b n d -> (b n) d')
176
+ )
177
+
178
+ return gated_output.reshape_as(x)
179
+
180
+
181
+ # feedforward
182
+
183
+ class GEGLU(nn.Module):
184
+ def __init__(self, dim_in, dim_out):
185
+ super().__init__()
186
+ self.proj = nn.Linear(dim_in, dim_out * 2)
187
+
188
+ def forward(self, x):
189
+ x, gate = self.proj(x).chunk(2, dim=-1)
190
+ return x * F.gelu(gate)
191
+
192
+
193
+ class FeedForward(nn.Module):
194
+ def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
195
+ super().__init__()
196
+ inner_dim = int(dim * mult)
197
+ dim_out = default(dim_out, dim)
198
+ project_in = nn.Sequential(
199
+ nn.Linear(dim, inner_dim),
200
+ nn.GELU()
201
+ ) if not glu else GEGLU(dim, inner_dim)
202
+
203
+ self.net = nn.Sequential(
204
+ project_in,
205
+ nn.Dropout(dropout),
206
+ nn.Linear(inner_dim, dim_out)
207
+ )
208
+
209
+ def forward(self, x):
210
+ return self.net(x)
211
+
212
+
213
+ # attention.
214
+ class Attention(nn.Module):
215
+ def __init__(
216
+ self,
217
+ dim,
218
+ dim_head=DEFAULT_DIM_HEAD,
219
+ heads=8,
220
+ causal=False,
221
+ mask=None,
222
+ talking_heads=False,
223
+ sparse_topk=None,
224
+ use_entmax15=False,
225
+ num_mem_kv=0,
226
+ dropout=0.,
227
+ on_attn=False
228
+ ):
229
+ super().__init__()
230
+ if use_entmax15:
231
+ raise NotImplementedError("Check out entmax activation instead of softmax activation!")
232
+ self.scale = dim_head ** -0.5
233
+ self.heads = heads
234
+ self.causal = causal
235
+ self.mask = mask
236
+
237
+ inner_dim = dim_head * heads
238
+
239
+ self.to_q = nn.Linear(dim, inner_dim, bias=False)
240
+ self.to_k = nn.Linear(dim, inner_dim, bias=False)
241
+ self.to_v = nn.Linear(dim, inner_dim, bias=False)
242
+ self.dropout = nn.Dropout(dropout)
243
+
244
+ # talking heads
245
+ self.talking_heads = talking_heads
246
+ if talking_heads:
247
+ self.pre_softmax_proj = nn.Parameter(torch.randn(heads, heads))
248
+ self.post_softmax_proj = nn.Parameter(torch.randn(heads, heads))
249
+
250
+ # explicit topk sparse attention
251
+ self.sparse_topk = sparse_topk
252
+
253
+ # entmax
254
+ #self.attn_fn = entmax15 if use_entmax15 else F.softmax
255
+ self.attn_fn = F.softmax
256
+
257
+ # add memory key / values
258
+ self.num_mem_kv = num_mem_kv
259
+ if num_mem_kv > 0:
260
+ self.mem_k = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))
261
+ self.mem_v = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head))
262
+
263
+ # attention on attention
264
+ self.attn_on_attn = on_attn
265
+ self.to_out = nn.Sequential(nn.Linear(inner_dim, dim * 2), nn.GLU()) if on_attn else nn.Linear(inner_dim, dim)
266
+
267
+ def forward(
268
+ self,
269
+ x,
270
+ context=None,
271
+ mask=None,
272
+ context_mask=None,
273
+ rel_pos=None,
274
+ sinusoidal_emb=None,
275
+ prev_attn=None,
276
+ mem=None
277
+ ):
278
+ b, n, _, h, talking_heads, device = *x.shape, self.heads, self.talking_heads, x.device
279
+ kv_input = default(context, x)
280
+
281
+ q_input = x
282
+ k_input = kv_input
283
+ v_input = kv_input
284
+
285
+ if exists(mem):
286
+ k_input = torch.cat((mem, k_input), dim=-2)
287
+ v_input = torch.cat((mem, v_input), dim=-2)
288
+
289
+ if exists(sinusoidal_emb):
290
+ # in shortformer, the query would start at a position offset depending on the past cached memory
291
+ offset = k_input.shape[-2] - q_input.shape[-2]
292
+ q_input = q_input + sinusoidal_emb(q_input, offset=offset)
293
+ k_input = k_input + sinusoidal_emb(k_input)
294
+
295
+ q = self.to_q(q_input)
296
+ k = self.to_k(k_input)
297
+ v = self.to_v(v_input)
298
+
299
+ q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h=h), (q, k, v))
300
+
301
+ input_mask = None
302
+ if any(map(exists, (mask, context_mask))):
303
+ q_mask = default(mask, lambda: torch.ones((b, n), device=device).bool())
304
+ k_mask = q_mask if not exists(context) else context_mask
305
+ k_mask = default(k_mask, lambda: torch.ones((b, k.shape[-2]), device=device).bool())
306
+ q_mask = rearrange(q_mask, 'b i -> b () i ()')
307
+ k_mask = rearrange(k_mask, 'b j -> b () () j')
308
+ input_mask = q_mask * k_mask
309
+
310
+ if self.num_mem_kv > 0:
311
+ mem_k, mem_v = map(lambda t: repeat(t, 'h n d -> b h n d', b=b), (self.mem_k, self.mem_v))
312
+ k = torch.cat((mem_k, k), dim=-2)
313
+ v = torch.cat((mem_v, v), dim=-2)
314
+ if exists(input_mask):
315
+ input_mask = F.pad(input_mask, (self.num_mem_kv, 0), value=True)
316
+
317
+ dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale
318
+ mask_value = max_neg_value(dots)
319
+
320
+ if exists(prev_attn):
321
+ dots = dots + prev_attn
322
+
323
+ pre_softmax_attn = dots
324
+
325
+ if talking_heads:
326
+ dots = einsum('b h i j, h k -> b k i j', dots, self.pre_softmax_proj).contiguous()
327
+
328
+ if exists(rel_pos):
329
+ dots = rel_pos(dots)
330
+
331
+ if exists(input_mask):
332
+ dots.masked_fill_(~input_mask, mask_value)
333
+ del input_mask
334
+
335
+ if self.causal:
336
+ i, j = dots.shape[-2:]
337
+ r = torch.arange(i, device=device)
338
+ mask = rearrange(r, 'i -> () () i ()') < rearrange(r, 'j -> () () () j')
339
+ mask = F.pad(mask, (j - i, 0), value=False)
340
+ dots.masked_fill_(mask, mask_value)
341
+ del mask
342
+
343
+ if exists(self.sparse_topk) and self.sparse_topk < dots.shape[-1]:
344
+ top, _ = dots.topk(self.sparse_topk, dim=-1)
345
+ vk = top[..., -1].unsqueeze(-1).expand_as(dots)
346
+ mask = dots < vk
347
+ dots.masked_fill_(mask, mask_value)
348
+ del mask
349
+
350
+ attn = self.attn_fn(dots, dim=-1)
351
+ post_softmax_attn = attn
352
+
353
+ attn = self.dropout(attn)
354
+
355
+ if talking_heads:
356
+ attn = einsum('b h i j, h k -> b k i j', attn, self.post_softmax_proj).contiguous()
357
+
358
+ out = einsum('b h i j, b h j d -> b h i d', attn, v)
359
+ out = rearrange(out, 'b h n d -> b n (h d)')
360
+
361
+ intermediates = Intermediates(
362
+ pre_softmax_attn=pre_softmax_attn,
363
+ post_softmax_attn=post_softmax_attn
364
+ )
365
+
366
+ return self.to_out(out), intermediates
367
+
368
+
369
+ class AttentionLayers(nn.Module):
370
+ def __init__(
371
+ self,
372
+ dim,
373
+ depth,
374
+ heads=8,
375
+ causal=False,
376
+ cross_attend=False,
377
+ only_cross=False,
378
+ use_scalenorm=False,
379
+ use_rmsnorm=False,
380
+ use_rezero=False,
381
+ rel_pos_num_buckets=32,
382
+ rel_pos_max_distance=128,
383
+ position_infused_attn=False,
384
+ custom_layers=None,
385
+ sandwich_coef=None,
386
+ par_ratio=None,
387
+ residual_attn=False,
388
+ cross_residual_attn=False,
389
+ macaron=False,
390
+ pre_norm=True,
391
+ gate_residual=False,
392
+ **kwargs
393
+ ):
394
+ super().__init__()
395
+ ff_kwargs, kwargs = groupby_prefix_and_trim('ff_', kwargs)
396
+ attn_kwargs, _ = groupby_prefix_and_trim('attn_', kwargs)
397
+
398
+ dim_head = attn_kwargs.get('dim_head', DEFAULT_DIM_HEAD)
399
+
400
+ self.dim = dim
401
+ self.depth = depth
402
+ self.layers = nn.ModuleList([])
403
+
404
+ self.has_pos_emb = position_infused_attn
405
+ self.pia_pos_emb = FixedPositionalEmbedding(dim) if position_infused_attn else None
406
+ self.rotary_pos_emb = always(None)
407
+
408
+ assert rel_pos_num_buckets <= rel_pos_max_distance, 'number of relative position buckets must be less than the relative position max distance'
409
+ self.rel_pos = None
410
+
411
+ self.pre_norm = pre_norm
412
+
413
+ self.residual_attn = residual_attn
414
+ self.cross_residual_attn = cross_residual_attn
415
+
416
+ norm_class = ScaleNorm if use_scalenorm else nn.LayerNorm
417
+ norm_class = RMSNorm if use_rmsnorm else norm_class
418
+ norm_fn = partial(norm_class, dim)
419
+
420
+ norm_fn = nn.Identity if use_rezero else norm_fn
421
+ branch_fn = Rezero if use_rezero else None
422
+
423
+ if cross_attend and not only_cross:
424
+ default_block = ('a', 'c', 'f')
425
+ elif cross_attend and only_cross:
426
+ default_block = ('c', 'f')
427
+ else:
428
+ default_block = ('a', 'f')
429
+
430
+ if macaron:
431
+ default_block = ('f',) + default_block
432
+
433
+ if exists(custom_layers):
434
+ layer_types = custom_layers
435
+ elif exists(par_ratio):
436
+ par_depth = depth * len(default_block)
437
+ assert 1 < par_ratio <= par_depth, 'par ratio out of range'
438
+ default_block = tuple(filter(not_equals('f'), default_block))
439
+ par_attn = par_depth // par_ratio
440
+ depth_cut = par_depth * 2 // 3 # 2 / 3 attention layer cutoff suggested by PAR paper
441
+ par_width = (depth_cut + depth_cut // par_attn) // par_attn
442
+ assert len(default_block) <= par_width, 'default block is too large for par_ratio'
443
+ par_block = default_block + ('f',) * (par_width - len(default_block))
444
+ par_head = par_block * par_attn
445
+ layer_types = par_head + ('f',) * (par_depth - len(par_head))
446
+ elif exists(sandwich_coef):
447
+ assert sandwich_coef > 0 and sandwich_coef <= depth, 'sandwich coefficient should be less than the depth'
448
+ layer_types = ('a',) * sandwich_coef + default_block * (depth - sandwich_coef) + ('f',) * sandwich_coef
449
+ else:
450
+ layer_types = default_block * depth
451
+
452
+ self.layer_types = layer_types
453
+ self.num_attn_layers = len(list(filter(equals('a'), layer_types)))
454
+
455
+ for layer_type in self.layer_types:
456
+ if layer_type == 'a':
457
+ layer = Attention(dim, heads=heads, causal=causal, **attn_kwargs)
458
+ elif layer_type == 'c':
459
+ layer = Attention(dim, heads=heads, **attn_kwargs)
460
+ elif layer_type == 'f':
461
+ layer = FeedForward(dim, **ff_kwargs)
462
+ layer = layer if not macaron else Scale(0.5, layer)
463
+ else:
464
+ raise Exception(f'invalid layer type {layer_type}')
465
+
466
+ if isinstance(layer, Attention) and exists(branch_fn):
467
+ layer = branch_fn(layer)
468
+
469
+ if gate_residual:
470
+ residual_fn = GRUGating(dim)
471
+ else:
472
+ residual_fn = Residual()
473
+
474
+ self.layers.append(nn.ModuleList([
475
+ norm_fn(),
476
+ layer,
477
+ residual_fn
478
+ ]))
479
+
480
+ def forward(
481
+ self,
482
+ x,
483
+ context=None,
484
+ mask=None,
485
+ context_mask=None,
486
+ mems=None,
487
+ return_hiddens=False
488
+ ):
489
+ hiddens = []
490
+ intermediates = []
491
+ prev_attn = None
492
+ prev_cross_attn = None
493
+
494
+ mems = mems.copy() if exists(mems) else [None] * self.num_attn_layers
495
+
496
+ for ind, (layer_type, (norm, block, residual_fn)) in enumerate(zip(self.layer_types, self.layers)):
497
+ is_last = ind == (len(self.layers) - 1)
498
+
499
+ if layer_type == 'a':
500
+ hiddens.append(x)
501
+ layer_mem = mems.pop(0)
502
+
503
+ residual = x
504
+
505
+ if self.pre_norm:
506
+ x = norm(x)
507
+
508
+ if layer_type == 'a':
509
+ out, inter = block(x, mask=mask, sinusoidal_emb=self.pia_pos_emb, rel_pos=self.rel_pos,
510
+ prev_attn=prev_attn, mem=layer_mem)
511
+ elif layer_type == 'c':
512
+ out, inter = block(x, context=context, mask=mask, context_mask=context_mask, prev_attn=prev_cross_attn)
513
+ elif layer_type == 'f':
514
+ out = block(x)
515
+
516
+ x = residual_fn(out, residual)
517
+
518
+ if layer_type in ('a', 'c'):
519
+ intermediates.append(inter)
520
+
521
+ if layer_type == 'a' and self.residual_attn:
522
+ prev_attn = inter.pre_softmax_attn
523
+ elif layer_type == 'c' and self.cross_residual_attn:
524
+ prev_cross_attn = inter.pre_softmax_attn
525
+
526
+ if not self.pre_norm and not is_last:
527
+ x = norm(x)
528
+
529
+ if return_hiddens:
530
+ intermediates = LayerIntermediates(
531
+ hiddens=hiddens,
532
+ attn_intermediates=intermediates
533
+ )
534
+
535
+ return x, intermediates
536
+
537
+ return x
538
+
539
+
540
+ class Encoder(AttentionLayers):
541
+ def __init__(self, **kwargs):
542
+ assert 'causal' not in kwargs, 'cannot set causality on encoder'
543
+ super().__init__(causal=False, **kwargs)
544
+
545
+
546
+
547
+ class TransformerWrapper(nn.Module):
548
+ def __init__(
549
+ self,
550
+ *,
551
+ num_tokens,
552
+ max_seq_len,
553
+ attn_layers,
554
+ emb_dim=None,
555
+ max_mem_len=0.,
556
+ emb_dropout=0.,
557
+ num_memory_tokens=None,
558
+ tie_embedding=False,
559
+ use_pos_emb=True
560
+ ):
561
+ super().__init__()
562
+ assert isinstance(attn_layers, AttentionLayers), 'attention layers must be one of Encoder or Decoder'
563
+
564
+ dim = attn_layers.dim
565
+ emb_dim = default(emb_dim, dim)
566
+
567
+ self.max_seq_len = max_seq_len
568
+ self.max_mem_len = max_mem_len
569
+ self.num_tokens = num_tokens
570
+
571
+ self.token_emb = nn.Embedding(num_tokens, emb_dim)
572
+ self.pos_emb = AbsolutePositionalEmbedding(emb_dim, max_seq_len) if (
573
+ use_pos_emb and not attn_layers.has_pos_emb) else always(0)
574
+ self.emb_dropout = nn.Dropout(emb_dropout)
575
+
576
+ self.project_emb = nn.Linear(emb_dim, dim) if emb_dim != dim else nn.Identity()
577
+ self.attn_layers = attn_layers
578
+ self.norm = nn.LayerNorm(dim)
579
+
580
+ self.init_()
581
+
582
+ self.to_logits = nn.Linear(dim, num_tokens) if not tie_embedding else lambda t: t @ self.token_emb.weight.t()
583
+
584
+ # memory tokens (like [cls]) from Memory Transformers paper
585
+ num_memory_tokens = default(num_memory_tokens, 0)
586
+ self.num_memory_tokens = num_memory_tokens
587
+ if num_memory_tokens > 0:
588
+ self.memory_tokens = nn.Parameter(torch.randn(num_memory_tokens, dim))
589
+
590
+ # let funnel encoder know number of memory tokens, if specified
591
+ if hasattr(attn_layers, 'num_memory_tokens'):
592
+ attn_layers.num_memory_tokens = num_memory_tokens
593
+
594
+ def init_(self):
595
+ nn.init.normal_(self.token_emb.weight, std=0.02)
596
+
597
+ def forward(
598
+ self,
599
+ x,
600
+ return_embeddings=False,
601
+ mask=None,
602
+ return_mems=False,
603
+ return_attn=False,
604
+ mems=None,
605
+ **kwargs
606
+ ):
607
+ b, n, device, num_mem = *x.shape, x.device, self.num_memory_tokens
608
+ x = self.token_emb(x)
609
+ x += self.pos_emb(x)
610
+ x = self.emb_dropout(x)
611
+
612
+ x = self.project_emb(x)
613
+
614
+ if num_mem > 0:
615
+ mem = repeat(self.memory_tokens, 'n d -> b n d', b=b)
616
+ x = torch.cat((mem, x), dim=1)
617
+
618
+ # auto-handle masking after appending memory tokens
619
+ if exists(mask):
620
+ mask = F.pad(mask, (num_mem, 0), value=True)
621
+
622
+ x, intermediates = self.attn_layers(x, mask=mask, mems=mems, return_hiddens=True, **kwargs)
623
+ x = self.norm(x)
624
+
625
+ mem, x = x[:, :num_mem], x[:, num_mem:]
626
+
627
+ out = self.to_logits(x) if not return_embeddings else x
628
+
629
+ if return_mems:
630
+ hiddens = intermediates.hiddens
631
+ new_mems = list(map(lambda pair: torch.cat(pair, dim=-2), zip(mems, hiddens))) if exists(mems) else hiddens
632
+ new_mems = list(map(lambda t: t[..., -self.max_mem_len:, :].detach(), new_mems))
633
+ return out, new_mems
634
+
635
+ if return_attn:
636
+ attn_maps = list(map(lambda t: t.post_softmax_attn, intermediates.attn_intermediates))
637
+ return out, attn_maps
638
+
639
+ return out
640
+
prompts/mp_prompts.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ A bigfoot giving a thumbs up in the snow, towards the camera;A bigfoot waving hands in the snow, towards the camera
2
+ A woman with red dress waving hands on the beach in sunset;A woman with red dress dancing on the beach in sunset
prompts/single_prompts.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ A chihuahua in astronaut suit floating in space, cinematic lighting, glow effect
2
+ A corgi is swimming
requirements.txt ADDED
@@ -0,0 +1,23 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ decord==0.6.0
2
+ einops==0.3.0
3
+ imageio==2.9.0
4
+ numpy==1.24.2
5
+ omegaconf==2.1.1
6
+ opencv_python
7
+ pandas==2.0.0
8
+ Pillow==9.5.0
9
+ pytorch_lightning==1.8.3
10
+ PyYAML==6.0
11
+ setuptools==65.6.3
12
+ torch==2.0.0
13
+ torchvision
14
+ tqdm==4.65.0
15
+ transformers==4.25.1
16
+ moviepy
17
+ av
18
+ xformers
19
+ gradio
20
+ timm
21
+ scikit-learn
22
+ open_clip_torch
23
+ kornia
scripts/evaluation/ddp_wrapper.py ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import datetime
2
+ import argparse, importlib
3
+ from pytorch_lightning import seed_everything
4
+
5
+ import torch
6
+ import torch.distributed as dist
7
+
8
+ def setup_dist(local_rank):
9
+ if dist.is_initialized():
10
+ return
11
+ torch.cuda.set_device(local_rank)
12
+ torch.distributed.init_process_group('nccl', init_method='env://')
13
+
14
+
15
+ def get_dist_info():
16
+ if dist.is_available():
17
+ initialized = dist.is_initialized()
18
+ else:
19
+ initialized = False
20
+ if initialized:
21
+ rank = dist.get_rank()
22
+ world_size = dist.get_world_size()
23
+ else:
24
+ rank = 0
25
+ world_size = 1
26
+ return rank, world_size
27
+
28
+
29
+ if __name__ == '__main__':
30
+ now = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S")
31
+ parser = argparse.ArgumentParser()
32
+ parser.add_argument("--module", type=str, help="module name", default="inference")
33
+ parser.add_argument("--local_rank", type=int, nargs="?", help="for ddp", default=0)
34
+ args, unknown = parser.parse_known_args()
35
+ inference_api = importlib.import_module(args.module, package=None)
36
+
37
+ inference_parser = inference_api.get_parser()
38
+ inference_args, unknown = inference_parser.parse_known_args()
39
+
40
+ seed_everything(inference_args.seed)
41
+ setup_dist(args.local_rank)
42
+ torch.backends.cudnn.benchmark = True
43
+ rank, gpu_num = get_dist_info()
44
+
45
+ print("@CoLVDM Inference [rank%d]: %s"%(rank, now))
46
+ inference_api.run_inference(inference_args, gpu_num, rank)
scripts/evaluation/funcs.py ADDED
@@ -0,0 +1,353 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os, sys, glob
2
+ import numpy as np
3
+ from collections import OrderedDict
4
+ from decord import VideoReader, cpu
5
+ import cv2
6
+
7
+ import torch
8
+ import torchvision
9
+ sys.path.insert(1, os.path.join(sys.path[0], '..', '..'))
10
+ from lvdm.models.samplers.ddim import DDIMSampler
11
+ from lvdm.models.samplers.ddim_mp import DDIMSampler as DDIMSampler_mp
12
+
13
+ def get_views(video_length, window_size=16, stride=4):
14
+ num_blocks_time = (video_length - window_size) // stride + 1
15
+ views = []
16
+ for i in range(num_blocks_time):
17
+ t_start = int(i * stride)
18
+ t_end = t_start + window_size
19
+ views.append((t_start,t_end))
20
+ return views
21
+
22
+ def batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=50, ddim_eta=1.0,\
23
+ cfg_scale=1.0, temporal_cfg_scale=None, **kwargs):
24
+ ddim_sampler = DDIMSampler(model)
25
+ uncond_type = model.uncond_type
26
+ batch_size = noise_shape[0]
27
+
28
+ ## construct unconditional guidance
29
+ if cfg_scale != 1.0:
30
+ if uncond_type == "empty_seq":
31
+ prompts = batch_size * [""]
32
+ #prompts = N * T * [""] ## if is_imgbatch=True
33
+ uc_emb = model.get_learned_conditioning(prompts)
34
+ elif uncond_type == "zero_embed":
35
+ c_emb = cond["c_crossattn"][0] if isinstance(cond, dict) else cond
36
+ uc_emb = torch.zeros_like(c_emb)
37
+
38
+ ## process image embedding token
39
+ if hasattr(model, 'embedder'):
40
+ uc_img = torch.zeros(noise_shape[0],3,224,224).to(model.device)
41
+ ## img: b c h w >> b l c
42
+ uc_img = model.get_image_embeds(uc_img)
43
+ uc_emb = torch.cat([uc_emb, uc_img], dim=1)
44
+
45
+ if isinstance(cond, dict):
46
+ uc = {key:cond[key] for key in cond.keys()}
47
+ uc.update({'c_crossattn': [uc_emb]})
48
+ else:
49
+ uc = uc_emb
50
+ else:
51
+ uc = None
52
+
53
+ x_T = None
54
+ batch_variants = []
55
+ #batch_variants1, batch_variants2 = [], []
56
+ for _ in range(n_samples):
57
+ if ddim_sampler is not None:
58
+ kwargs.update({"clean_cond": True})
59
+ samples, _ = ddim_sampler.sample(S=ddim_steps,
60
+ conditioning=cond,
61
+ batch_size=noise_shape[0],
62
+ shape=noise_shape[1:],
63
+ verbose=False,
64
+ unconditional_guidance_scale=cfg_scale,
65
+ unconditional_conditioning=uc,
66
+ eta=ddim_eta,
67
+ temporal_length=noise_shape[2],
68
+ conditional_guidance_scale_temporal=temporal_cfg_scale,
69
+ x_T=x_T,
70
+ **kwargs
71
+ )
72
+ ## reconstruct from latent to pixel space
73
+ batch_images = model.decode_first_stage_2DAE(samples)
74
+ batch_variants.append(batch_images)
75
+ ## batch, <samples>, c, t, h, w
76
+ batch_variants = torch.stack(batch_variants, dim=1)
77
+ return batch_variants
78
+
79
+ def batch_ddim_sampling_freenoise(model, cond, noise_shape, n_samples=1, ddim_steps=50, ddim_eta=1.0,\
80
+ cfg_scale=1.0, temporal_cfg_scale=None, args=None, x_T_total=None, **kwargs):
81
+ ddim_sampler = DDIMSampler(model)
82
+ uncond_type = model.uncond_type
83
+ batch_size = noise_shape[0]
84
+
85
+ ## construct unconditional guidance
86
+ if cfg_scale != 1.0:
87
+ if uncond_type == "empty_seq":
88
+ prompts = batch_size * [""]
89
+ #prompts = N * T * [""] ## if is_imgbatch=True
90
+ uc_emb = model.get_learned_conditioning(prompts)
91
+ elif uncond_type == "zero_embed":
92
+ c_emb = cond["c_crossattn"][0] if isinstance(cond, dict) else cond
93
+ uc_emb = torch.zeros_like(c_emb)
94
+
95
+ ## process image embedding token
96
+ if hasattr(model, 'embedder'):
97
+ uc_img = torch.zeros(noise_shape[0],3,224,224).to(model.device)
98
+ ## img: b c h w >> b l c
99
+ uc_img = model.get_image_embeds(uc_img)
100
+ uc_emb = torch.cat([uc_emb, uc_img], dim=1)
101
+
102
+ if isinstance(cond, dict):
103
+ uc = {key:cond[key] for key in cond.keys()}
104
+ uc.update({'c_crossattn': [uc_emb]})
105
+ else:
106
+ uc = uc_emb
107
+ else:
108
+ uc = None
109
+
110
+ views = get_views(args.frames, args.window_size, args.window_stride)
111
+
112
+ batch_variants = []
113
+ #batch_variants1, batch_variants2 = [], []
114
+ for _ in range(n_samples):
115
+ x_T = x_T_total[_]
116
+ if ddim_sampler is not None:
117
+ kwargs.update({"clean_cond": True})
118
+ samples, _ = ddim_sampler.sample(S=ddim_steps,
119
+ conditioning=cond,
120
+ batch_size=noise_shape[0],
121
+ shape=noise_shape[1:],
122
+ verbose=False,
123
+ unconditional_guidance_scale=cfg_scale,
124
+ unconditional_conditioning=uc,
125
+ eta=ddim_eta,
126
+ temporal_length=noise_shape[2],
127
+ conditional_guidance_scale_temporal=temporal_cfg_scale,
128
+ x_T=x_T,
129
+ context_next=views,
130
+ **kwargs
131
+ )
132
+ ## reconstruct from latent to pixel space
133
+ batch_images = model.decode_first_stage_2DAE(samples)
134
+ batch_variants.append(batch_images)
135
+ ## batch, <samples>, c, t, h, w
136
+ batch_variants = torch.stack(batch_variants, dim=1)
137
+ return batch_variants
138
+
139
+ def batch_ddim_sampling_freenoise_mp(model, cond, noise_shape, n_samples=1, ddim_steps=50, ddim_eta=1.0,\
140
+ cfg_scale=1.0, temporal_cfg_scale=None, args=None, x_T_total=None, **kwargs):
141
+ ddim_sampler = DDIMSampler_mp(model)
142
+ uncond_type = model.uncond_type
143
+ batch_size = noise_shape[0]
144
+
145
+ ## construct unconditional guidance
146
+ if cfg_scale != 1.0:
147
+ if uncond_type == "empty_seq":
148
+ prompts = batch_size * [""]
149
+ #prompts = N * T * [""] ## if is_imgbatch=True
150
+ uc_emb = model.get_learned_conditioning(prompts)
151
+ elif uncond_type == "zero_embed":
152
+ c_emb = cond["c_crossattn"][0] if isinstance(cond, dict) else cond
153
+ uc_emb = torch.zeros_like(c_emb)
154
+
155
+ ## process image embedding token
156
+ if hasattr(model, 'embedder'):
157
+ uc_img = torch.zeros(noise_shape[0],3,224,224).to(model.device)
158
+ ## img: b c h w >> b l c
159
+ uc_img = model.get_image_embeds(uc_img)
160
+ uc_emb = torch.cat([uc_emb, uc_img], dim=1)
161
+
162
+ if isinstance(cond, dict):
163
+ uc = {key:cond[key] for key in cond.keys()}
164
+ uc.update({'c_crossattn': [uc_emb]})
165
+ else:
166
+ uc = uc_emb
167
+ else:
168
+ uc = None
169
+
170
+ views = get_views(args.frames, args.window_size, args.window_stride)
171
+
172
+ conditioning = cond['c_crossattn'][0]
173
+ len1 = int(args.frames * 3 // 8)
174
+ len2 = args.frames - len1 * 2
175
+ cond_diff1 = (conditioning[[1]] - conditioning[[0]]) / (len2 - 1)
176
+ cond_list1 = []
177
+ for i in range(len2):
178
+ cond_list1.append((conditioning[[0]] + cond_diff1 * i).unsqueeze(0))
179
+
180
+ cond1 = torch.cat([conditioning[[0]].unsqueeze(0).repeat(1, len1, 1, 1), torch.cat(cond_list1, 1), conditioning[[1]].unsqueeze(0).repeat(1, len1, 1, 1)], 1)
181
+ cond2 = torch.cat([conditioning[[1]].unsqueeze(0).repeat(1, args.frames, 1, 1)], 1)
182
+
183
+ cond_all = torch.cat([cond1, cond2], 0)
184
+
185
+ cond['c_crossattn'] = [cond_all]
186
+
187
+ batch_variants = []
188
+ #batch_variants1, batch_variants2 = [], []
189
+ for _ in range(n_samples):
190
+ x_T = x_T_total[_]
191
+ if ddim_sampler is not None:
192
+ kwargs.update({"clean_cond": True})
193
+ samples, _ = ddim_sampler.sample(S=ddim_steps,
194
+ conditioning=cond,
195
+ batch_size=noise_shape[0],
196
+ shape=noise_shape[1:],
197
+ verbose=False,
198
+ unconditional_guidance_scale=cfg_scale,
199
+ unconditional_conditioning=uc,
200
+ eta=ddim_eta,
201
+ temporal_length=noise_shape[2],
202
+ conditional_guidance_scale_temporal=temporal_cfg_scale,
203
+ x_T=x_T,
204
+ context_next=views,
205
+ **kwargs
206
+ )
207
+ ## reconstruct from latent to pixel space
208
+ batch_images = model.decode_first_stage_2DAE(samples)
209
+ batch_variants.append(batch_images)
210
+ ## batch, <samples>, c, t, h, w
211
+ batch_variants = torch.stack(batch_variants, dim=1)
212
+ return batch_variants
213
+
214
+ def get_filelist(data_dir, ext='*'):
215
+ file_list = glob.glob(os.path.join(data_dir, '*.%s'%ext))
216
+ file_list.sort()
217
+ return file_list
218
+
219
+ def get_dirlist(path):
220
+ list = []
221
+ if (os.path.exists(path)):
222
+ files = os.listdir(path)
223
+ for file in files:
224
+ m = os.path.join(path,file)
225
+ if (os.path.isdir(m)):
226
+ list.append(m)
227
+ list.sort()
228
+ return list
229
+
230
+
231
+ def load_model_checkpoint(model, ckpt):
232
+ def load_checkpoint(model, ckpt, full_strict):
233
+ state_dict = torch.load(ckpt, map_location="cpu")
234
+ try:
235
+ ## deepspeed
236
+ new_pl_sd = OrderedDict()
237
+ for key in state_dict['module'].keys():
238
+ new_pl_sd[key[16:]]=state_dict['module'][key]
239
+ model.load_state_dict(new_pl_sd, strict=full_strict)
240
+ except:
241
+ if "state_dict" in list(state_dict.keys()):
242
+ state_dict = state_dict["state_dict"]
243
+ model.load_state_dict(state_dict, strict=full_strict)
244
+ return model
245
+ load_checkpoint(model, ckpt, full_strict=True)
246
+ print('>>> model checkpoint loaded.')
247
+ return model
248
+
249
+
250
+ def load_prompts(prompt_file):
251
+ f = open(prompt_file, 'r')
252
+ prompt_list = []
253
+ for idx, line in enumerate(f.readlines()):
254
+ l = line.strip()
255
+ if len(l) != 0:
256
+ prompt_list.append(l)
257
+ f.close()
258
+ return prompt_list
259
+
260
+ def load_prompts_mp(prompt_file):
261
+ f = open(prompt_file, 'r')
262
+ prompt_list = []
263
+ for idx, line in enumerate(f.readlines()):
264
+ l = []
265
+ line = line.strip()
266
+ prompts = line.split(';')
267
+ for prompt in prompts:
268
+ prompt = prompt.strip()
269
+ if len(prompt) != 0:
270
+ l.append(prompt)
271
+ if len(l) != 0:
272
+ prompt_list.append(l)
273
+ f.close()
274
+ print(prompt_list)
275
+ return prompt_list
276
+
277
+ def load_video_batch(filepath_list, frame_stride, video_size=(256,256), video_frames=16):
278
+ '''
279
+ Notice about some special cases:
280
+ 1. video_frames=-1 means to take all the frames (with fs=1)
281
+ 2. when the total video frames is less than required, padding strategy will be used (repreated last frame)
282
+ '''
283
+ fps_list = []
284
+ batch_tensor = []
285
+ assert frame_stride > 0, "valid frame stride should be a positive integer!"
286
+ for filepath in filepath_list:
287
+ padding_num = 0
288
+ vidreader = VideoReader(filepath, ctx=cpu(0), width=video_size[1], height=video_size[0])
289
+ fps = vidreader.get_avg_fps()
290
+ total_frames = len(vidreader)
291
+ max_valid_frames = (total_frames-1) // frame_stride + 1
292
+ if video_frames < 0:
293
+ ## all frames are collected: fs=1 is a must
294
+ required_frames = total_frames
295
+ frame_stride = 1
296
+ else:
297
+ required_frames = video_frames
298
+ query_frames = min(required_frames, max_valid_frames)
299
+ frame_indices = [frame_stride*i for i in range(query_frames)]
300
+
301
+ ## [t,h,w,c] -> [c,t,h,w]
302
+ frames = vidreader.get_batch(frame_indices)
303
+ frame_tensor = torch.tensor(frames.asnumpy()).permute(3, 0, 1, 2).float()
304
+ frame_tensor = (frame_tensor / 255. - 0.5) * 2
305
+ if max_valid_frames < required_frames:
306
+ padding_num = required_frames - max_valid_frames
307
+ frame_tensor = torch.cat([frame_tensor, *([frame_tensor[:,-1:,:,:]]*padding_num)], dim=1)
308
+ print(f'{os.path.split(filepath)[1]} is not long enough: {padding_num} frames padded.')
309
+ batch_tensor.append(frame_tensor)
310
+ sample_fps = int(fps/frame_stride)
311
+ fps_list.append(sample_fps)
312
+
313
+ return torch.stack(batch_tensor, dim=0)
314
+
315
+ from PIL import Image
316
+ def load_image_batch(filepath_list, image_size=(256,256)):
317
+ batch_tensor = []
318
+ for filepath in filepath_list:
319
+ _, filename = os.path.split(filepath)
320
+ _, ext = os.path.splitext(filename)
321
+ if ext == '.mp4':
322
+ vidreader = VideoReader(filepath, ctx=cpu(0), width=image_size[1], height=image_size[0])
323
+ frame = vidreader.get_batch([0])
324
+ img_tensor = torch.tensor(frame.asnumpy()).squeeze(0).permute(2, 0, 1).float()
325
+ elif ext == '.png' or ext == '.jpg':
326
+ img = Image.open(filepath).convert("RGB")
327
+ rgb_img = np.array(img, np.float32)
328
+ #bgr_img = cv2.imread(filepath, cv2.IMREAD_COLOR)
329
+ #bgr_img = cv2.cvtColor(bgr_img, cv2.COLOR_BGR2RGB)
330
+ rgb_img = cv2.resize(rgb_img, (image_size[1],image_size[0]), interpolation=cv2.INTER_LINEAR)
331
+ img_tensor = torch.from_numpy(rgb_img).permute(2, 0, 1).float()
332
+ else:
333
+ print(f'ERROR: <{ext}> image loading only support format: [mp4], [png], [jpg]')
334
+ raise NotImplementedError
335
+ img_tensor = (img_tensor / 255. - 0.5) * 2
336
+ batch_tensor.append(img_tensor)
337
+ return torch.stack(batch_tensor, dim=0)
338
+
339
+
340
+ def save_videos(batch_tensors, savedir, filenames, fps=10):
341
+ # b,samples,c,t,h,w
342
+ n_samples = batch_tensors.shape[1]
343
+ for idx, vid_tensor in enumerate(batch_tensors):
344
+ video = vid_tensor.detach().cpu()
345
+ video = torch.clamp(video.float(), -1., 1.)
346
+ video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w
347
+ frame_grids = [torchvision.utils.make_grid(framesheet, nrow=int(n_samples)) for framesheet in video] #[3, 1*h, n*w]
348
+ grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [t, 3, n*h, w]
349
+ grid = (grid + 1.0) / 2.0
350
+ grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1)
351
+ savepath = os.path.join(savedir, f"{filenames[idx]}.mp4")
352
+ torchvision.io.write_video(savepath, grid, fps=fps, video_codec='h264', options={'crf': '10'})
353
+
scripts/evaluation/inference.py ADDED
@@ -0,0 +1,137 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse, os, sys, glob, yaml, math, random
2
+ import datetime, time
3
+ import numpy as np
4
+ from omegaconf import OmegaConf
5
+ from collections import OrderedDict
6
+ from tqdm import trange, tqdm
7
+ from einops import repeat
8
+ from einops import rearrange, repeat
9
+ from functools import partial
10
+ import torch
11
+ from pytorch_lightning import seed_everything
12
+
13
+ from funcs import load_model_checkpoint, load_prompts, load_image_batch, get_filelist, save_videos
14
+ from funcs import batch_ddim_sampling
15
+ from utils.utils import instantiate_from_config
16
+
17
+
18
+ def get_parser():
19
+ parser = argparse.ArgumentParser()
20
+ parser.add_argument("--seed", type=int, default=20230211, help="seed for seed_everything")
21
+ parser.add_argument("--mode", default="base", type=str, help="which kind of inference mode: {'base', 'i2v'}")
22
+ parser.add_argument("--ckpt_path", type=str, default=None, help="checkpoint path")
23
+ parser.add_argument("--config", type=str, help="config (yaml) path")
24
+ parser.add_argument("--prompt_file", type=str, default=None, help="a text file containing many prompts")
25
+ parser.add_argument("--savedir", type=str, default=None, help="results saving path")
26
+ parser.add_argument("--savefps", type=str, default=10, help="video fps to generate")
27
+ parser.add_argument("--n_samples", type=int, default=1, help="num of samples per prompt",)
28
+ parser.add_argument("--ddim_steps", type=int, default=50, help="steps of ddim if positive, otherwise use DDPM",)
29
+ parser.add_argument("--ddim_eta", type=float, default=1.0, help="eta for ddim sampling (0.0 yields deterministic sampling)",)
30
+ parser.add_argument("--bs", type=int, default=1, help="batch size for inference")
31
+ parser.add_argument("--height", type=int, default=512, help="image height, in pixel space")
32
+ parser.add_argument("--width", type=int, default=512, help="image width, in pixel space")
33
+ parser.add_argument("--frames", type=int, default=-1, help="frames num to inference")
34
+ parser.add_argument("--fps", type=int, default=24)
35
+ parser.add_argument("--unconditional_guidance_scale", type=float, default=1.0, help="prompt classifier-free guidance")
36
+ parser.add_argument("--unconditional_guidance_scale_temporal", type=float, default=None, help="temporal consistency guidance")
37
+ ## for conditional i2v only
38
+ parser.add_argument("--cond_input", type=str, default=None, help="data dir of conditional input")
39
+ return parser
40
+
41
+
42
+ def run_inference(args, gpu_num, gpu_no, **kwargs):
43
+ ## step 1: model config
44
+ ## -----------------------------------------------------------------
45
+ config = OmegaConf.load(args.config)
46
+ #data_config = config.pop("data", OmegaConf.create())
47
+ model_config = config.pop("model", OmegaConf.create())
48
+ model = instantiate_from_config(model_config)
49
+ model = model.cuda(gpu_no)
50
+ assert os.path.exists(args.ckpt_path), f"Error: checkpoint [{args.ckpt_path}] Not Found!"
51
+ model = load_model_checkpoint(model, args.ckpt_path)
52
+ model.eval()
53
+
54
+ ## sample shape
55
+ assert (args.height % 16 == 0) and (args.width % 16 == 0), "Error: image size [h,w] should be multiples of 16!"
56
+ ## latent noise shape
57
+ h, w = args.height // 8, args.width // 8
58
+ frames = model.temporal_length if args.frames < 0 else args.frames
59
+ channels = model.channels
60
+
61
+ ## saving folders
62
+ os.makedirs(args.savedir, exist_ok=True)
63
+
64
+ ## step 2: load data
65
+ ## -----------------------------------------------------------------
66
+ assert os.path.exists(args.prompt_file), "Error: prompt file NOT Found!"
67
+ prompt_list = load_prompts(args.prompt_file)
68
+ num_samples = len(prompt_list)
69
+ filename_list = [f"{id+1:04d}" for id in range(num_samples)]
70
+
71
+ samples_split = num_samples // gpu_num
72
+ residual_tail = num_samples % gpu_num
73
+ print(f'[rank:{gpu_no}] {samples_split}/{num_samples} samples loaded.')
74
+ indices = list(range(samples_split*gpu_no, samples_split*(gpu_no+1)))
75
+ if gpu_no == 0 and residual_tail != 0:
76
+ indices = indices + list(range(num_samples-residual_tail, num_samples))
77
+ prompt_list_rank = [prompt_list[i] for i in indices]
78
+
79
+ ## conditional input
80
+ if args.mode == "i2v":
81
+ ## each video or frames dir per prompt
82
+ cond_inputs = get_filelist(args.cond_input, ext='[mpj][pn][4gj]') # '[mpj][pn][4gj]'
83
+ assert len(cond_inputs) == num_samples, f"Error: conditional input ({len(cond_inputs)}) NOT match prompt ({num_samples})!"
84
+ filename_list = [f"{os.path.split(cond_inputs[id])[-1][:-4]}" for id in range(num_samples)]
85
+ cond_inputs_rank = [cond_inputs[i] for i in indices]
86
+
87
+ filename_list_rank = [filename_list[i] for i in indices]
88
+
89
+ ## step 3: run over samples
90
+ ## -----------------------------------------------------------------
91
+ start = time.time()
92
+ n_rounds = len(prompt_list_rank) // args.bs
93
+ n_rounds = n_rounds+1 if len(prompt_list_rank) % args.bs != 0 else n_rounds
94
+ for idx in range(0, n_rounds):
95
+ print(f'[rank:{gpu_no}] batch-{idx+1} ({args.bs})x{args.n_samples} ...')
96
+ idx_s = idx*args.bs
97
+ idx_e = min(idx_s+args.bs, len(prompt_list_rank))
98
+ batch_size = idx_e - idx_s
99
+ filenames = filename_list_rank[idx_s:idx_e]
100
+ noise_shape = [batch_size, channels, frames, h, w]
101
+ fps = torch.tensor([args.fps]*batch_size).to(model.device).long()
102
+
103
+ prompts = prompt_list_rank[idx_s:idx_e]
104
+ if isinstance(prompts, str):
105
+ prompts = [prompts]
106
+ #prompts = batch_size * [""]
107
+ text_emb = model.get_learned_conditioning(prompts)
108
+
109
+ if args.mode == 'base':
110
+ cond = {"c_crossattn": [text_emb], "fps": fps}
111
+ elif args.mode == 'i2v':
112
+ #cond_images = torch.zeros(noise_shape[0],3,224,224).to(model.device)
113
+ cond_images = load_image_batch(cond_inputs_rank[idx_s:idx_e], (args.height, args.width))
114
+ cond_images = cond_images.to(model.device)
115
+ img_emb = model.get_image_embeds(cond_images)
116
+ imtext_cond = torch.cat([text_emb, img_emb], dim=1)
117
+ cond = {"c_crossattn": [imtext_cond], "fps": fps}
118
+ else:
119
+ raise NotImplementedError
120
+
121
+ ## inference
122
+ batch_samples = batch_ddim_sampling(model, cond, noise_shape, args.n_samples, \
123
+ args.ddim_steps, args.ddim_eta, args.unconditional_guidance_scale, **kwargs)
124
+ ## b,samples,c,t,h,w
125
+ save_videos(batch_samples, args.savedir, filenames, fps=args.savefps)
126
+
127
+ print(f"Saved in {args.savedir}. Time used: {(time.time() - start):.2f} seconds")
128
+
129
+
130
+ if __name__ == '__main__':
131
+ now = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S")
132
+ print("@CoLVDM Inference: %s"%now)
133
+ parser = get_parser()
134
+ args = parser.parse_args()
135
+ seed_everything(args.seed)
136
+ rank, gpu_num = 0, 1
137
+ run_inference(args, gpu_num, rank)
scripts/evaluation/inference_freenoise.py ADDED
@@ -0,0 +1,147 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse, os, sys, glob, yaml, math, random
2
+ import datetime, time
3
+ import numpy as np
4
+ from omegaconf import OmegaConf
5
+ from collections import OrderedDict
6
+ from tqdm import trange, tqdm
7
+ from einops import repeat
8
+ from einops import rearrange, repeat
9
+ from functools import partial
10
+ import torch
11
+ from pytorch_lightning import seed_everything
12
+
13
+ from funcs import load_model_checkpoint, load_prompts, load_image_batch, get_filelist, save_videos
14
+ from funcs import batch_ddim_sampling_freenoise
15
+ from utils.utils import instantiate_from_config
16
+
17
+
18
+ def get_parser():
19
+ parser = argparse.ArgumentParser()
20
+ parser.add_argument("--seed", type=int, default=20230211, help="seed for seed_everything")
21
+ parser.add_argument("--mode", default="base", type=str, help="which kind of inference mode: {'base', 'i2v'}")
22
+ parser.add_argument("--ckpt_path", type=str, default=None, help="checkpoint path")
23
+ parser.add_argument("--config", type=str, help="config (yaml) path")
24
+ parser.add_argument("--prompt_file", type=str, default=None, help="a text file containing many prompts")
25
+ parser.add_argument("--savedir", type=str, default=None, help="results saving path")
26
+ parser.add_argument("--savefps", type=str, default=10, help="video fps to generate")
27
+ parser.add_argument("--n_samples", type=int, default=1, help="num of samples per prompt",)
28
+ parser.add_argument("--ddim_steps", type=int, default=50, help="steps of ddim if positive, otherwise use DDPM",)
29
+ parser.add_argument("--ddim_eta", type=float, default=1.0, help="eta for ddim sampling (0.0 yields deterministic sampling)",)
30
+ parser.add_argument("--bs", type=int, default=1, help="batch size for inference")
31
+ parser.add_argument("--height", type=int, default=512, help="image height, in pixel space")
32
+ parser.add_argument("--width", type=int, default=512, help="image width, in pixel space")
33
+ parser.add_argument("--frames", type=int, default=-1, help="frames num to inference")
34
+ parser.add_argument("--fps", type=int, default=24)
35
+ parser.add_argument("--unconditional_guidance_scale", type=float, default=1.0, help="prompt classifier-free guidance")
36
+ parser.add_argument("--unconditional_guidance_scale_temporal", type=float, default=None, help="temporal consistency guidance")
37
+ ## for conditional i2v only
38
+ parser.add_argument("--cond_input", type=str, default=None, help="data dir of conditional input")
39
+ ## for freenoise only
40
+ parser.add_argument("--window_size", type=int, default=16, help="window_size")
41
+ parser.add_argument("--window_stride", type=int, default=4, help="window_stride")
42
+ return parser
43
+
44
+
45
+ def run_inference(args, gpu_num, gpu_no, **kwargs):
46
+ ## step 1: model config
47
+ ## -----------------------------------------------------------------
48
+ config = OmegaConf.load(args.config)
49
+ #data_config = config.pop("data", OmegaConf.create())
50
+ model_config = config.pop("model", OmegaConf.create())
51
+ model = instantiate_from_config(model_config)
52
+ model = model.cuda(gpu_no)
53
+ assert os.path.exists(args.ckpt_path), f"Error: checkpoint [{args.ckpt_path}] Not Found!"
54
+ model = load_model_checkpoint(model, args.ckpt_path)
55
+ model.eval()
56
+
57
+ ## sample shape
58
+ assert (args.height % 16 == 0) and (args.width % 16 == 0), "Error: image size [h,w] should be multiples of 16!"
59
+ ## latent noise shape
60
+ h, w = args.height // 8, args.width // 8
61
+ frames = model.temporal_length if args.frames < 0 else args.frames
62
+ channels = model.channels
63
+
64
+ ## saving folders
65
+ os.makedirs(args.savedir, exist_ok=True)
66
+
67
+ ## step 2: load data
68
+ ## -----------------------------------------------------------------
69
+ assert os.path.exists(args.prompt_file), "Error: prompt file NOT Found!"
70
+ prompt_list = load_prompts(args.prompt_file)
71
+ num_samples = len(prompt_list)
72
+ filename_list = [f"{id+1:04d}" for id in range(num_samples)]
73
+
74
+ samples_split = num_samples // gpu_num
75
+ residual_tail = num_samples % gpu_num
76
+ print(f'[rank:{gpu_no}] {samples_split}/{num_samples} samples loaded.')
77
+ indices = list(range(samples_split*gpu_no, samples_split*(gpu_no+1)))
78
+ if gpu_no == 0 and residual_tail != 0:
79
+ indices = indices + list(range(num_samples-residual_tail, num_samples))
80
+ prompt_list_rank = [prompt_list[i] for i in indices]
81
+
82
+ ## conditional input
83
+ if args.mode == "i2v":
84
+ ## each video or frames dir per prompt
85
+ cond_inputs = get_filelist(args.cond_input, ext='[mpj][pn][4gj]') # '[mpj][pn][4gj]'
86
+ assert len(cond_inputs) == num_samples, f"Error: conditional input ({len(cond_inputs)}) NOT match prompt ({num_samples})!"
87
+ filename_list = [f"{os.path.split(cond_inputs[id])[-1][:-4]}" for id in range(num_samples)]
88
+ cond_inputs_rank = [cond_inputs[i] for i in indices]
89
+
90
+ filename_list_rank = [filename_list[i] for i in indices]
91
+
92
+ ## step 3: run over samples
93
+ ## -----------------------------------------------------------------
94
+ start = time.time()
95
+ n_rounds = len(prompt_list_rank) // args.bs
96
+ n_rounds = n_rounds+1 if len(prompt_list_rank) % args.bs != 0 else n_rounds
97
+
98
+ x_T_total = torch.randn([args.n_samples, 1, channels, frames, h, w], device=model.device).repeat(1, args.bs, 1, 1, 1, 1)
99
+ for frame_index in range(args.window_size, args.frames, args.window_stride):
100
+ list_index = list(range(frame_index-args.window_size, frame_index+args.window_stride-args.window_size))
101
+ random.shuffle(list_index)
102
+ x_T_total[:, :, :, frame_index:frame_index+args.window_stride] = x_T_total[:, :, :, list_index]
103
+
104
+ for idx in range(0, n_rounds):
105
+ print(f'[rank:{gpu_no}] batch-{idx+1} ({args.bs})x{args.n_samples} ...')
106
+ idx_s = idx*args.bs
107
+ idx_e = min(idx_s+args.bs, len(prompt_list_rank))
108
+ batch_size = idx_e - idx_s
109
+ filenames = filename_list_rank[idx_s:idx_e]
110
+ noise_shape = [batch_size, channels, frames, h, w]
111
+ fps = torch.tensor([args.fps]*batch_size).to(model.device).long()
112
+
113
+ prompts = prompt_list_rank[idx_s:idx_e]
114
+ if isinstance(prompts, str):
115
+ prompts = [prompts]
116
+ #prompts = batch_size * [""]
117
+ text_emb = model.get_learned_conditioning(prompts)
118
+
119
+ if args.mode == 'base':
120
+ cond = {"c_crossattn": [text_emb], "fps": fps}
121
+ elif args.mode == 'i2v':
122
+ #cond_images = torch.zeros(noise_shape[0],3,224,224).to(model.device)
123
+ cond_images = load_image_batch(cond_inputs_rank[idx_s:idx_e], (args.height, args.width))
124
+ cond_images = cond_images.to(model.device)
125
+ img_emb = model.get_image_embeds(cond_images)
126
+ imtext_cond = torch.cat([text_emb, img_emb], dim=1)
127
+ cond = {"c_crossattn": [imtext_cond], "fps": fps}
128
+ else:
129
+ raise NotImplementedError
130
+
131
+ ## inference
132
+ batch_samples = batch_ddim_sampling_freenoise(model, cond, noise_shape, args.n_samples, \
133
+ args.ddim_steps, args.ddim_eta, args.unconditional_guidance_scale, args=args, x_T_total=x_T_total, **kwargs)
134
+ ## b,samples,c,t,h,w
135
+ save_videos(batch_samples, args.savedir, filenames, fps=args.savefps)
136
+
137
+ print(f"Saved in {args.savedir}. Time used: {(time.time() - start):.2f} seconds")
138
+
139
+
140
+ if __name__ == '__main__':
141
+ now = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S")
142
+ print("@CoLVDM Inference: %s"%now)
143
+ parser = get_parser()
144
+ args = parser.parse_args()
145
+ seed_everything(args.seed)
146
+ rank, gpu_num = 0, 1
147
+ run_inference(args, gpu_num, rank)
scripts/evaluation/inference_freenoise_mp.py ADDED
@@ -0,0 +1,153 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse, os, sys, glob, yaml, math, random
2
+ import datetime, time
3
+ import numpy as np
4
+ from omegaconf import OmegaConf
5
+ from collections import OrderedDict
6
+ from tqdm import trange, tqdm
7
+ from einops import repeat
8
+ from einops import rearrange, repeat
9
+ from functools import partial
10
+ import torch
11
+ from pytorch_lightning import seed_everything
12
+
13
+ from funcs import load_model_checkpoint, load_prompts_mp, load_image_batch, get_filelist, save_videos
14
+ from funcs import batch_ddim_sampling_freenoise_mp
15
+ from utils.utils import instantiate_from_config
16
+
17
+
18
+ def get_parser():
19
+ parser = argparse.ArgumentParser()
20
+ parser.add_argument("--seed", type=int, default=20230211, help="seed for seed_everything")
21
+ parser.add_argument("--mode", default="base", type=str, help="which kind of inference mode: {'base', 'i2v'}")
22
+ parser.add_argument("--ckpt_path", type=str, default=None, help="checkpoint path")
23
+ parser.add_argument("--config", type=str, help="config (yaml) path")
24
+ parser.add_argument("--prompt_file", type=str, default=None, help="a text file containing many prompts")
25
+ parser.add_argument("--savedir", type=str, default=None, help="results saving path")
26
+ parser.add_argument("--savefps", type=str, default=10, help="video fps to generate")
27
+ parser.add_argument("--n_samples", type=int, default=1, help="num of samples per prompt",)
28
+ parser.add_argument("--ddim_steps", type=int, default=50, help="steps of ddim if positive, otherwise use DDPM",)
29
+ parser.add_argument("--ddim_eta", type=float, default=1.0, help="eta for ddim sampling (0.0 yields deterministic sampling)",)
30
+ parser.add_argument("--bs", type=int, default=1, help="batch size for inference")
31
+ parser.add_argument("--height", type=int, default=512, help="image height, in pixel space")
32
+ parser.add_argument("--width", type=int, default=512, help="image width, in pixel space")
33
+ parser.add_argument("--frames", type=int, default=-1, help="frames num to inference")
34
+ parser.add_argument("--fps", type=int, default=24)
35
+ parser.add_argument("--unconditional_guidance_scale", type=float, default=1.0, help="prompt classifier-free guidance")
36
+ parser.add_argument("--unconditional_guidance_scale_temporal", type=float, default=None, help="temporal consistency guidance")
37
+ ## for conditional i2v only
38
+ parser.add_argument("--cond_input", type=str, default=None, help="data dir of conditional input")
39
+ ## for freenoise only
40
+ parser.add_argument("--window_size", type=int, default=16, help="window_size")
41
+ parser.add_argument("--window_stride", type=int, default=4, help="window_stride")
42
+ return parser
43
+
44
+
45
+ def run_inference(args, gpu_num, gpu_no, **kwargs):
46
+ ## step 1: model config
47
+ ## -----------------------------------------------------------------
48
+ config = OmegaConf.load(args.config)
49
+ #data_config = config.pop("data", OmegaConf.create())
50
+ model_config = config.pop("model", OmegaConf.create())
51
+ model = instantiate_from_config(model_config)
52
+ model = model.cuda(gpu_no)
53
+ assert os.path.exists(args.ckpt_path), f"Error: checkpoint [{args.ckpt_path}] Not Found!"
54
+ model = load_model_checkpoint(model, args.ckpt_path)
55
+ model.eval()
56
+
57
+ ## sample shape
58
+ assert (args.height % 16 == 0) and (args.width % 16 == 0), "Error: image size [h,w] should be multiples of 16!"
59
+ ## latent noise shape
60
+ h, w = args.height // 8, args.width // 8
61
+ frames = model.temporal_length if args.frames < 0 else args.frames
62
+ channels = model.channels
63
+
64
+ ## saving folders
65
+ os.makedirs(args.savedir, exist_ok=True)
66
+
67
+ ## step 2: load data
68
+ ## -----------------------------------------------------------------
69
+ assert os.path.exists(args.prompt_file), "Error: prompt file NOT Found!"
70
+ prompt_list = load_prompts_mp(args.prompt_file)
71
+ num_samples = len(prompt_list)
72
+ filename_list = [f"{id+1:04d}" for id in range(num_samples)]
73
+
74
+ samples_split = num_samples // gpu_num
75
+ residual_tail = num_samples % gpu_num
76
+ print(f'[rank:{gpu_no}] {samples_split}/{num_samples} samples loaded.')
77
+ indices = list(range(samples_split*gpu_no, samples_split*(gpu_no+1)))
78
+ if gpu_no == 0 and residual_tail != 0:
79
+ indices = indices + list(range(num_samples-residual_tail, num_samples))
80
+ prompt_list_rank = [prompt_list[i] for i in indices]
81
+
82
+ ## conditional input
83
+ if args.mode == "i2v":
84
+ ## each video or frames dir per prompt
85
+ cond_inputs = get_filelist(args.cond_input, ext='[mpj][pn][4gj]') # '[mpj][pn][4gj]'
86
+ assert len(cond_inputs) == num_samples, f"Error: conditional input ({len(cond_inputs)}) NOT match prompt ({num_samples})!"
87
+ filename_list = [f"{os.path.split(cond_inputs[id])[-1][:-4]}" for id in range(num_samples)]
88
+ cond_inputs_rank = [cond_inputs[i] for i in indices]
89
+
90
+ filename_list_rank = [filename_list[i] for i in indices]
91
+
92
+ ## step 3: run over samples
93
+ ## -----------------------------------------------------------------
94
+ start = time.time()
95
+ n_rounds = len(prompt_list_rank) // args.bs
96
+ n_rounds = n_rounds+1 if len(prompt_list_rank) % args.bs != 0 else n_rounds
97
+
98
+ x_T_total = torch.randn([args.n_samples, 1, channels, frames, h, w], device=model.device).repeat(1, args.bs, 1, 1, 1, 1)
99
+ for frame_index in range(args.window_size, args.frames, args.window_stride):
100
+ list_index = list(range(frame_index-args.window_size, frame_index+args.window_stride-args.window_size))
101
+ random.shuffle(list_index)
102
+ x_T_total[:, :, :, frame_index:frame_index+args.window_stride] = x_T_total[:, :, :, list_index]
103
+
104
+ for idx in range(0, n_rounds):
105
+ print(f'[rank:{gpu_no}] batch-{idx+1} ({args.bs})x{args.n_samples} ...')
106
+ idx_s = idx*args.bs
107
+ idx_e = min(idx_s+args.bs, len(prompt_list_rank))
108
+ batch_size = idx_e - idx_s
109
+ filenames = filename_list_rank[idx_s:idx_e]
110
+ noise_shape = [batch_size, channels, frames, h, w]
111
+ fps = torch.tensor([args.fps]*batch_size).to(model.device).long()
112
+
113
+ prompts_list = prompt_list_rank[idx_s:idx_e]
114
+ if isinstance(prompts_list, str):
115
+ prompts_list = [prompts_list]
116
+ #prompts = batch_size * [""]
117
+ text_emb_list = []
118
+ for prompts in prompts_list:
119
+ # text_emb = model.get_learned_conditioning(prompts)
120
+ # text_emb_list.append(text_emb)
121
+ text_emb = [model.get_learned_conditioning(prompt) for prompt in prompts]
122
+ text_emb = torch.cat(text_emb, 0)
123
+ text_emb_list.append(text_emb)
124
+
125
+ if args.mode == 'base':
126
+ cond = {"c_crossattn": text_emb_list, "fps": fps}
127
+ elif args.mode == 'i2v':
128
+ #cond_images = torch.zeros(noise_shape[0],3,224,224).to(model.device)
129
+ cond_images = load_image_batch(cond_inputs_rank[idx_s:idx_e], (args.height, args.width))
130
+ cond_images = cond_images.to(model.device)
131
+ img_emb = model.get_image_embeds(cond_images)
132
+ imtext_cond = torch.cat([text_emb, img_emb], dim=1)
133
+ cond = {"c_crossattn": [imtext_cond], "fps": fps}
134
+ else:
135
+ raise NotImplementedError
136
+
137
+ ## inference
138
+ batch_samples = batch_ddim_sampling_freenoise_mp(model, cond, noise_shape, args.n_samples, \
139
+ args.ddim_steps, args.ddim_eta, args.unconditional_guidance_scale, args=args, x_T_total=x_T_total, **kwargs)
140
+ ## b,samples,c,t,h,w
141
+ save_videos(batch_samples, args.savedir, filenames, fps=args.savefps)
142
+
143
+ print(f"Saved in {args.savedir}. Time used: {(time.time() - start):.2f} seconds")
144
+
145
+
146
+ if __name__ == '__main__':
147
+ now = datetime.datetime.now().strftime("%Y-%m-%d-%H-%M-%S")
148
+ print("@CoLVDM Inference: %s"%now)
149
+ parser = get_parser()
150
+ args = parser.parse_args()
151
+ seed_everything(args.seed)
152
+ rank, gpu_num = 0, 1
153
+ run_inference(args, gpu_num, rank)
scripts/run_text2video.sh ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name="base_512_test"
2
+
3
+ ckpt='checkpoints/base_512_v1/model.ckpt'
4
+ config='configs/inference_t2v_tconv512_v1.0.yaml'
5
+
6
+ prompt_file="prompts/single_prompts.txt"
7
+ res_dir="results_single_512"
8
+
9
+ python3 scripts/evaluation/inference.py \
10
+ --seed 123 \
11
+ --mode 'base' \
12
+ --ckpt_path $ckpt \
13
+ --config $config \
14
+ --savedir $res_dir/$name \
15
+ --n_samples 3 \
16
+ --bs 1 --height 320 --width 512 \
17
+ --unconditional_guidance_scale 12.0 \
18
+ --ddim_steps 50 \
19
+ --ddim_eta 0.0 \
20
+ --prompt_file $prompt_file \
21
+ --fps 8 \
22
+ --frames 16
scripts/run_text2video_freenoise_1024.sh ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name="base_1024_test"
2
+
3
+ ckpt='checkpoints/base_1024_v1/model.ckpt'
4
+ config='configs/inference_t2v_1024_v1.0_freenoise.yaml'
5
+
6
+ prompt_file="prompts/single_prompts.txt"
7
+ res_dir="results_freenoise_single_1024"
8
+
9
+ python3 scripts/evaluation/inference_freenoise.py \
10
+ --seed 123 \
11
+ --mode 'base' \
12
+ --ckpt_path $ckpt \
13
+ --config $config \
14
+ --savedir $res_dir/$name \
15
+ --n_samples 3 \
16
+ --bs 1 --height 576 --width 1024 \
17
+ --unconditional_guidance_scale 12.0 \
18
+ --ddim_steps 50 \
19
+ --ddim_eta 0.0 \
20
+ --prompt_file $prompt_file \
21
+ --fps 28 \
22
+ --frames 64 \
23
+ --window_size 16 \
24
+ --window_stride 4
scripts/run_text2video_freenoise_256.sh ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name="base_256_test"
2
+
3
+ ckpt='checkpoints/base_256_v1/model.pth'
4
+ config='configs/inference_t2v_tconv256_v1.0_freenoise.yaml'
5
+
6
+ prompt_file="prompts/single_prompts.txt"
7
+ res_dir="results_freenoise_single_256"
8
+
9
+ python3 scripts/evaluation/inference_freenoise.py \
10
+ --seed 123 \
11
+ --mode 'base' \
12
+ --ckpt_path $ckpt \
13
+ --config $config \
14
+ --savedir $res_dir/$name \
15
+ --n_samples 3 \
16
+ --bs 1 --height 256 --width 256 \
17
+ --unconditional_guidance_scale 15.0 \
18
+ --ddim_steps 50 \
19
+ --ddim_eta 0.0 \
20
+ --prompt_file $prompt_file \
21
+ --fps 8 \
22
+ --frames 128 \
23
+ --window_size 16 \
24
+ --window_stride 4
25
+
scripts/run_text2video_freenoise_512.sh ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name="base_512_test"
2
+
3
+ ckpt='checkpoints/base_512_v1/model.ckpt'
4
+ config='configs/inference_t2v_tconv512_v1.0_freenoise.yaml'
5
+
6
+ prompt_file="prompts/single_prompts.txt"
7
+ res_dir="results_freenoise_single_512"
8
+
9
+ python3 scripts/evaluation/inference_freenoise.py \
10
+ --seed 123 \
11
+ --mode 'base' \
12
+ --ckpt_path $ckpt \
13
+ --config $config \
14
+ --savedir $res_dir/$name \
15
+ --n_samples 3 \
16
+ --bs 1 --height 320 --width 512 \
17
+ --unconditional_guidance_scale 12.0 \
18
+ --ddim_steps 50 \
19
+ --ddim_eta 0.0 \
20
+ --prompt_file $prompt_file \
21
+ --fps 8 \
22
+ --frames 64 \
23
+ --window_size 16 \
24
+ --window_stride 4
scripts/run_text2video_freenoise_mp_256.sh ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name="base_256_test"
2
+
3
+ ckpt='checkpoints/base_256_v1/model.pth'
4
+ config='configs/inference_t2v_tconv256_v1.0_freenoise.yaml'
5
+
6
+ prompt_file="prompts/mp_prompts.txt"
7
+ res_dir="results_freenoise_mp_256"
8
+
9
+ python3 scripts/evaluation/inference_freenoise_mp.py \
10
+ --seed 123 \
11
+ --mode 'base' \
12
+ --ckpt_path $ckpt \
13
+ --config $config \
14
+ --savedir $res_dir/$name \
15
+ --n_samples 3 \
16
+ --bs 1 --height 256 --width 256 \
17
+ --unconditional_guidance_scale 15.0 \
18
+ --ddim_steps 50 \
19
+ --ddim_eta 0.0 \
20
+ --prompt_file $prompt_file \
21
+ --fps 8 \
22
+ --frames 64 \
23
+ --window_size 16 \
24
+ --window_stride 4
25
+
scripts/run_text2video_freenoise_mp_512.sh ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name="base_512_test"
2
+
3
+ ckpt='checkpoints/base_512_v1/model.ckpt'
4
+ config='configs/inference_t2v_tconv512_v1.0_freenoise.yaml'
5
+
6
+ prompt_file="prompts/mp_prompts.txt"
7
+ res_dir="results_freenoise_mp_512"
8
+
9
+ python3 scripts/evaluation/inference_freenoise_mp.py \
10
+ --seed 123 \
11
+ --mode 'base' \
12
+ --ckpt_path $ckpt \
13
+ --config $config \
14
+ --savedir $res_dir/$name \
15
+ --n_samples 3 \
16
+ --bs 1 --height 320 --width 512 \
17
+ --unconditional_guidance_scale 12.0 \
18
+ --ddim_steps 50 \
19
+ --ddim_eta 0.0 \
20
+ --prompt_file $prompt_file \
21
+ --fps 8 \
22
+ --frames 64 \
23
+ --window_size 16 \
24
+ --window_stride 4
utils/utils.py ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import importlib
2
+ import numpy as np
3
+ import cv2
4
+ import torch
5
+ import torch.distributed as dist
6
+
7
+
8
+ def count_params(model, verbose=False):
9
+ total_params = sum(p.numel() for p in model.parameters())
10
+ if verbose:
11
+ print(f"{model.__class__.__name__} has {total_params*1.e-6:.2f} M params.")
12
+ return total_params
13
+
14
+
15
+ def check_istarget(name, para_list):
16
+ """
17
+ name: full name of source para
18
+ para_list: partial name of target para
19
+ """
20
+ istarget=False
21
+ for para in para_list:
22
+ if para in name:
23
+ return True
24
+ return istarget
25
+
26
+
27
+ def instantiate_from_config(config):
28
+ if not "target" in config:
29
+ if config == '__is_first_stage__':
30
+ return None
31
+ elif config == "__is_unconditional__":
32
+ return None
33
+ raise KeyError("Expected key `target` to instantiate.")
34
+ return get_obj_from_str(config["target"])(**config.get("params", dict()))
35
+
36
+
37
+ def get_obj_from_str(string, reload=False):
38
+ module, cls = string.rsplit(".", 1)
39
+ if reload:
40
+ module_imp = importlib.import_module(module)
41
+ importlib.reload(module_imp)
42
+ return getattr(importlib.import_module(module, package=None), cls)
43
+
44
+
45
+ def load_npz_from_dir(data_dir):
46
+ data = [np.load(os.path.join(data_dir, data_name))['arr_0'] for data_name in os.listdir(data_dir)]
47
+ data = np.concatenate(data, axis=0)
48
+ return data
49
+
50
+
51
+ def load_npz_from_paths(data_paths):
52
+ data = [np.load(data_path)['arr_0'] for data_path in data_paths]
53
+ data = np.concatenate(data, axis=0)
54
+ return data
55
+
56
+
57
+ def resize_numpy_image(image, max_resolution=512 * 512, resize_short_edge=None):
58
+ h, w = image.shape[:2]
59
+ if resize_short_edge is not None:
60
+ k = resize_short_edge / min(h, w)
61
+ else:
62
+ k = max_resolution / (h * w)
63
+ k = k**0.5
64
+ h = int(np.round(h * k / 64)) * 64
65
+ w = int(np.round(w * k / 64)) * 64
66
+ image = cv2.resize(image, (w, h), interpolation=cv2.INTER_LANCZOS4)
67
+ return image
68
+
69
+
70
+ def setup_dist(args):
71
+ if dist.is_initialized():
72
+ return
73
+ torch.cuda.set_device(args.local_rank)
74
+ torch.distributed.init_process_group(
75
+ 'nccl',
76
+ init_method='env://'
77
+ )