Upload 84 files
Browse filesThis view is limited to 50 files because it contains too many changes.
See raw diff
- api.py +117 -0
- src/.idea/.gitignore +8 -0
- src/.idea/inspectionProfiles/Project_Default.xml +34 -0
- src/.idea/inspectionProfiles/profiles_settings.xml +6 -0
- src/.idea/misc.xml +7 -0
- src/.idea/modules.xml +8 -0
- src/.idea/src.iml +12 -0
- src/.idea/workspace.xml +128 -0
- src/inference.py +169 -0
- src/models/blocks.py +325 -0
- src/models/conditioners.py +180 -0
- src/models/udit.py +356 -0
- src/models/utils/.ipynb_checkpoints/__init__-checkpoint.py +0 -0
- src/models/utils/.ipynb_checkpoints/attention-checkpoint.py +290 -0
- src/models/utils/.ipynb_checkpoints/modules-checkpoint.py +374 -0
- src/models/utils/.ipynb_checkpoints/rotary-checkpoint.py +91 -0
- src/models/utils/.ipynb_checkpoints/span_mask-checkpoint.py +146 -0
- src/models/utils/.ipynb_checkpoints/timm-checkpoint.py +114 -0
- src/models/utils/__init__.py +0 -0
- src/models/utils/__pycache__/__init__.cpython-310.pyc +0 -0
- src/models/utils/__pycache__/__init__.cpython-311.pyc +0 -0
- src/models/utils/__pycache__/attention.cpython-310.pyc +0 -0
- src/models/utils/__pycache__/attention.cpython-311.pyc +0 -0
- src/models/utils/__pycache__/modules.cpython-310.pyc +0 -0
- src/models/utils/__pycache__/modules.cpython-311.pyc +0 -0
- src/models/utils/__pycache__/rotary.cpython-310.pyc +0 -0
- src/models/utils/__pycache__/rotary.cpython-311.pyc +0 -0
- src/models/utils/__pycache__/span_mask.cpython-310.pyc +0 -0
- src/models/utils/__pycache__/span_mask.cpython-311.pyc +0 -0
- src/models/utils/__pycache__/timm.cpython-310.pyc +0 -0
- src/models/utils/__pycache__/timm.cpython-311.pyc +0 -0
- src/models/utils/attention.py +290 -0
- src/models/utils/bk/.ipynb_checkpoints/attention-checkpoint.py +99 -0
- src/models/utils/bk/.ipynb_checkpoints/llama_rotary-checkpoint.py +74 -0
- src/models/utils/bk/__pycache__/rotary.cpython-311.pyc +0 -0
- src/models/utils/bk/attention.py +99 -0
- src/models/utils/bk/llama_rotary.py +74 -0
- src/models/utils/modules.py +374 -0
- src/models/utils/rotary.py +91 -0
- src/models/utils/span_mask.py +146 -0
- src/models/utils/timm.py +114 -0
- src/modules/autoencoder_wrapper.py +83 -0
- src/modules/clap_wrapper.py +0 -0
- src/modules/dac/__init__.py +16 -0
- src/modules/dac/__main__.py +36 -0
- src/modules/dac/compare/__init__.py +0 -0
- src/modules/dac/compare/encodec.py +54 -0
- src/modules/dac/model/__init__.py +4 -0
- src/modules/dac/model/base.py +294 -0
- src/modules/dac/model/dac.py +364 -0
api.py
ADDED
@@ -0,0 +1,117 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import torch
|
3 |
+
import random
|
4 |
+
import numpy as np
|
5 |
+
import gradio as gr
|
6 |
+
import soundfile as sf
|
7 |
+
from transformers import T5Tokenizer, T5EncoderModel
|
8 |
+
from diffusers import DDIMScheduler
|
9 |
+
from src.models.conditioners import MaskDiT
|
10 |
+
from src.modules.autoencoder_wrapper import Autoencoder
|
11 |
+
from src.inference import inference
|
12 |
+
from src.utils import load_yaml_with_includes
|
13 |
+
|
14 |
+
|
15 |
+
# Load model and configs
|
16 |
+
def load_models(config_name, ckpt_path, vae_path, device):
|
17 |
+
params = load_yaml_with_includes(config_name)
|
18 |
+
|
19 |
+
# Load codec model
|
20 |
+
autoencoder = Autoencoder(ckpt_path=vae_path,
|
21 |
+
model_type=params['autoencoder']['name'],
|
22 |
+
quantization_first=params['autoencoder']['q_first']).to(device)
|
23 |
+
autoencoder.eval()
|
24 |
+
|
25 |
+
# Load text encoder
|
26 |
+
tokenizer = T5Tokenizer.from_pretrained(params['text_encoder']['model'])
|
27 |
+
text_encoder = T5EncoderModel.from_pretrained(params['text_encoder']['model']).to(device)
|
28 |
+
text_encoder.eval()
|
29 |
+
|
30 |
+
# Load main U-Net model
|
31 |
+
unet = MaskDiT(**params['model']).to(device)
|
32 |
+
unet.load_state_dict(torch.load(ckpt_path)['model'])
|
33 |
+
unet.eval()
|
34 |
+
|
35 |
+
# Load noise scheduler
|
36 |
+
noise_scheduler = DDIMScheduler(**params['diff'])
|
37 |
+
|
38 |
+
return autoencoder, unet, tokenizer, text_encoder, noise_scheduler, params
|
39 |
+
|
40 |
+
MAX_SEED = np.iinfo(np.int32).max
|
41 |
+
|
42 |
+
# Model and config paths
|
43 |
+
config_name = 'ckpts/ezaudio-xl.yml'
|
44 |
+
ckpt_path = 'ckpts/s3/ezaudio_s3_xl.pt'
|
45 |
+
vae_path = 'ckpts/vae/1m.pt'
|
46 |
+
save_path = 'output/'
|
47 |
+
os.makedirs(save_path, exist_ok=True)
|
48 |
+
|
49 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
50 |
+
|
51 |
+
autoencoder, unet, tokenizer, text_encoder, noise_scheduler, params = load_models(config_name, ckpt_path, vae_path,
|
52 |
+
device)
|
53 |
+
|
54 |
+
latents = torch.randn((1, 128, 128), device=device)
|
55 |
+
noise = torch.randn_like(latents)
|
56 |
+
timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (1,), device=device)
|
57 |
+
_ = noise_scheduler.add_noise(latents, noise, timesteps)
|
58 |
+
|
59 |
+
|
60 |
+
# Inference function
|
61 |
+
def generate_audio(text, length,
|
62 |
+
guidance_scale, guidance_rescale, ddim_steps, eta,
|
63 |
+
random_seed, randomize_seed):
|
64 |
+
neg_text = None
|
65 |
+
length = length * params['autoencoder']['latent_sr']
|
66 |
+
|
67 |
+
if randomize_seed:
|
68 |
+
random_seed = random.randint(0, MAX_SEED)
|
69 |
+
|
70 |
+
pred = inference(autoencoder, unet, None, None,
|
71 |
+
tokenizer, text_encoder,
|
72 |
+
params, noise_scheduler,
|
73 |
+
text, neg_text,
|
74 |
+
length,
|
75 |
+
guidance_scale, guidance_rescale,
|
76 |
+
ddim_steps, eta, random_seed,
|
77 |
+
device)
|
78 |
+
|
79 |
+
pred = pred.cpu().numpy().squeeze(0).squeeze(0)
|
80 |
+
# output_file = f"{save_path}/{text}.wav"
|
81 |
+
# sf.write(output_file, pred, samplerate=params['autoencoder']['sr'])
|
82 |
+
|
83 |
+
return params['autoencoder']['sr'], pred
|
84 |
+
|
85 |
+
|
86 |
+
# Gradio Interface
|
87 |
+
def gradio_interface():
|
88 |
+
# Input components
|
89 |
+
text_input = gr.Textbox(label="Text Prompt", value="the sound of dog barking")
|
90 |
+
length_input = gr.Slider(minimum=1, maximum=10, step=1, value=10, label="Audio Length (in seconds)")
|
91 |
+
|
92 |
+
# Advanced settings
|
93 |
+
guidance_scale_input = gr.Slider(minimum=1.0, maximum=10, step=0.1, value=5, label="Guidance Scale")
|
94 |
+
guidance_rescale_input = gr.Slider(minimum=0.0, maximum=1, step=0.05, value=0.75, label="Guidance Rescale")
|
95 |
+
ddim_steps_input = gr.Slider(minimum=25, maximum=200, step=5, value=100, label="DDIM Steps")
|
96 |
+
eta_input = gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=1, label="Eta")
|
97 |
+
random_seed_input = gr.Slider(minimum=0, maximum=MAX_SEED, step=1, value=0,)
|
98 |
+
|
99 |
+
randomize_seed = gr.Checkbox(label="Randomize seed", value=False)
|
100 |
+
|
101 |
+
# Output component
|
102 |
+
output_audio = gr.Audio(label="Converted Audio", type="numpy")
|
103 |
+
|
104 |
+
# Interface
|
105 |
+
gr.Interface(
|
106 |
+
fn=generate_audio,
|
107 |
+
inputs=[text_input, length_input, guidance_scale_input, guidance_rescale_input, ddim_steps_input, eta_input,
|
108 |
+
random_seed_input, randomize_seed],
|
109 |
+
outputs=output_audio,
|
110 |
+
title="EzAudio Text-to-Audio Generator",
|
111 |
+
description="Generate audio from text using a diffusion model. Adjust advanced settings for more control.",
|
112 |
+
allow_flagging="never"
|
113 |
+
).launch()
|
114 |
+
|
115 |
+
|
116 |
+
if __name__ == "__main__":
|
117 |
+
gradio_interface()
|
src/.idea/.gitignore
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Default ignored files
|
2 |
+
/shelf/
|
3 |
+
/workspace.xml
|
4 |
+
# Editor-based HTTP Client requests
|
5 |
+
/httpRequests/
|
6 |
+
# Datasource local storage ignored files
|
7 |
+
/dataSources/
|
8 |
+
/dataSources.local.xml
|
src/.idea/inspectionProfiles/Project_Default.xml
ADDED
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
<component name="InspectionProjectProfileManager">
|
2 |
+
<profile version="1.0">
|
3 |
+
<option name="myName" value="Project Default" />
|
4 |
+
<inspection_tool class="PyPackageRequirementsInspection" enabled="true" level="WARNING" enabled_by_default="true">
|
5 |
+
<option name="ignoredPackages">
|
6 |
+
<value>
|
7 |
+
<list size="21">
|
8 |
+
<item index="0" class="java.lang.String" itemvalue="numba" />
|
9 |
+
<item index="1" class="java.lang.String" itemvalue="scipy" />
|
10 |
+
<item index="2" class="java.lang.String" itemvalue="decorator" />
|
11 |
+
<item index="3" class="java.lang.String" itemvalue="six" />
|
12 |
+
<item index="4" class="java.lang.String" itemvalue="joblib" />
|
13 |
+
<item index="5" class="java.lang.String" itemvalue="threadpoolctl" />
|
14 |
+
<item index="6" class="java.lang.String" itemvalue="scikit-learn" />
|
15 |
+
<item index="7" class="java.lang.String" itemvalue="python-dateutil" />
|
16 |
+
<item index="8" class="java.lang.String" itemvalue="cffi" />
|
17 |
+
<item index="9" class="java.lang.String" itemvalue="SoundFile" />
|
18 |
+
<item index="10" class="java.lang.String" itemvalue="audioread" />
|
19 |
+
<item index="11" class="java.lang.String" itemvalue="kiwisolver" />
|
20 |
+
<item index="12" class="java.lang.String" itemvalue="cycler" />
|
21 |
+
<item index="13" class="java.lang.String" itemvalue="llvmlite" />
|
22 |
+
<item index="14" class="java.lang.String" itemvalue="mido" />
|
23 |
+
<item index="15" class="java.lang.String" itemvalue="matplotlib" />
|
24 |
+
<item index="16" class="java.lang.String" itemvalue="resampy" />
|
25 |
+
<item index="17" class="java.lang.String" itemvalue="librosa" />
|
26 |
+
<item index="18" class="java.lang.String" itemvalue="pyparsing" />
|
27 |
+
<item index="19" class="java.lang.String" itemvalue="pretty-midi" />
|
28 |
+
<item index="20" class="java.lang.String" itemvalue="Pillow" />
|
29 |
+
</list>
|
30 |
+
</value>
|
31 |
+
</option>
|
32 |
+
</inspection_tool>
|
33 |
+
</profile>
|
34 |
+
</component>
|
src/.idea/inspectionProfiles/profiles_settings.xml
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
<component name="InspectionProjectProfileManager">
|
2 |
+
<settings>
|
3 |
+
<option name="USE_PROJECT_PROFILE" value="false" />
|
4 |
+
<version value="1.0" />
|
5 |
+
</settings>
|
6 |
+
</component>
|
src/.idea/misc.xml
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
<?xml version="1.0" encoding="UTF-8"?>
|
2 |
+
<project version="4">
|
3 |
+
<component name="Black">
|
4 |
+
<option name="sdkName" value="Python 3.10" />
|
5 |
+
</component>
|
6 |
+
<component name="ProjectRootManager" version="2" project-jdk-name="Python 3.10" project-jdk-type="Python SDK" />
|
7 |
+
</project>
|
src/.idea/modules.xml
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
<?xml version="1.0" encoding="UTF-8"?>
|
2 |
+
<project version="4">
|
3 |
+
<component name="ProjectModuleManager">
|
4 |
+
<modules>
|
5 |
+
<module fileurl="file://$PROJECT_DIR$/.idea/src.iml" filepath="$PROJECT_DIR$/.idea/src.iml" />
|
6 |
+
</modules>
|
7 |
+
</component>
|
8 |
+
</project>
|
src/.idea/src.iml
ADDED
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
<?xml version="1.0" encoding="UTF-8"?>
|
2 |
+
<module type="PYTHON_MODULE" version="4">
|
3 |
+
<component name="NewModuleRootManager">
|
4 |
+
<content url="file://$MODULE_DIR$" />
|
5 |
+
<orderEntry type="inheritedJdk" />
|
6 |
+
<orderEntry type="sourceFolder" forTests="false" />
|
7 |
+
</component>
|
8 |
+
<component name="PyDocumentationSettings">
|
9 |
+
<option name="format" value="PLAIN" />
|
10 |
+
<option name="myDocStringFormat" value="Plain" />
|
11 |
+
</component>
|
12 |
+
</module>
|
src/.idea/workspace.xml
ADDED
@@ -0,0 +1,128 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
<?xml version="1.0" encoding="UTF-8"?>
|
2 |
+
<project version="4">
|
3 |
+
<component name="AutoImportSettings">
|
4 |
+
<option name="autoReloadType" value="SELECTIVE" />
|
5 |
+
</component>
|
6 |
+
<component name="ChangeListManager">
|
7 |
+
<list default="true" id="cb82860d-7ce6-451e-932b-96d3a6e7b20d" name="Changes" comment="" />
|
8 |
+
<option name="SHOW_DIALOG" value="false" />
|
9 |
+
<option name="HIGHLIGHT_CONFLICTS" value="true" />
|
10 |
+
<option name="HIGHLIGHT_NON_ACTIVE_CHANGELIST" value="false" />
|
11 |
+
<option name="LAST_RESOLUTION" value="IGNORE" />
|
12 |
+
</component>
|
13 |
+
<component name="ProjectColorInfo"><![CDATA[{
|
14 |
+
"associatedIndex": 4
|
15 |
+
}]]></component>
|
16 |
+
<component name="ProjectId" id="2m8UaG5ZprDRDpwT0ASOoKWJVNg" />
|
17 |
+
<component name="ProjectViewState">
|
18 |
+
<option name="hideEmptyMiddlePackages" value="true" />
|
19 |
+
<option name="showLibraryContents" value="true" />
|
20 |
+
</component>
|
21 |
+
<component name="PropertiesComponent"><![CDATA[{
|
22 |
+
"keyToString": {
|
23 |
+
"Python.api.executor": "Run",
|
24 |
+
"Python.clean.executor": "Run",
|
25 |
+
"Python.gradio.executor": "Run",
|
26 |
+
"RunOnceActivity.ShowReadmeOnStart": "true",
|
27 |
+
"node.js.detected.package.eslint": "true",
|
28 |
+
"node.js.detected.package.tslint": "true",
|
29 |
+
"node.js.selected.package.eslint": "(autodetect)",
|
30 |
+
"node.js.selected.package.tslint": "(autodetect)",
|
31 |
+
"nodejs_package_manager_path": "npm",
|
32 |
+
"settings.editor.selected.configurable": "com.jetbrains.python.configuration.PyActiveSdkModuleConfigurable",
|
33 |
+
"vue.rearranger.settings.migration": "true"
|
34 |
+
}
|
35 |
+
}]]></component>
|
36 |
+
<component name="RunManager" selected="Python.api">
|
37 |
+
<configuration name="api" type="PythonConfigurationType" factoryName="Python" temporary="true" nameIsGenerated="true">
|
38 |
+
<module name="src" />
|
39 |
+
<option name="ENV_FILES" value="" />
|
40 |
+
<option name="INTERPRETER_OPTIONS" value="" />
|
41 |
+
<option name="PARENT_ENVS" value="true" />
|
42 |
+
<option name="SDK_HOME" value="" />
|
43 |
+
<option name="WORKING_DIRECTORY" value="$PROJECT_DIR$/.." />
|
44 |
+
<option name="IS_MODULE_SDK" value="false" />
|
45 |
+
<option name="ADD_CONTENT_ROOTS" value="true" />
|
46 |
+
<option name="ADD_SOURCE_ROOTS" value="true" />
|
47 |
+
<EXTENSION ID="PythonCoverageRunConfigurationExtension" runner="coverage.py" />
|
48 |
+
<option name="SCRIPT_NAME" value="$PROJECT_DIR$/../api.py" />
|
49 |
+
<option name="PARAMETERS" value="" />
|
50 |
+
<option name="SHOW_COMMAND_LINE" value="false" />
|
51 |
+
<option name="EMULATE_TERMINAL" value="false" />
|
52 |
+
<option name="MODULE_MODE" value="false" />
|
53 |
+
<option name="REDIRECT_INPUT" value="false" />
|
54 |
+
<option name="INPUT_FILE" value="" />
|
55 |
+
<method v="2" />
|
56 |
+
</configuration>
|
57 |
+
<configuration name="clean" type="PythonConfigurationType" factoryName="Python" temporary="true" nameIsGenerated="true">
|
58 |
+
<module name="src" />
|
59 |
+
<option name="ENV_FILES" value="" />
|
60 |
+
<option name="INTERPRETER_OPTIONS" value="" />
|
61 |
+
<option name="PARENT_ENVS" value="true" />
|
62 |
+
<option name="SDK_HOME" value="" />
|
63 |
+
<option name="WORKING_DIRECTORY" value="$PROJECT_DIR$/../ckpts/vae" />
|
64 |
+
<option name="IS_MODULE_SDK" value="false" />
|
65 |
+
<option name="ADD_CONTENT_ROOTS" value="true" />
|
66 |
+
<option name="ADD_SOURCE_ROOTS" value="true" />
|
67 |
+
<EXTENSION ID="PythonCoverageRunConfigurationExtension" runner="coverage.py" />
|
68 |
+
<option name="SCRIPT_NAME" value="$PROJECT_DIR$/../ckpts/vae/clean.py" />
|
69 |
+
<option name="PARAMETERS" value="" />
|
70 |
+
<option name="SHOW_COMMAND_LINE" value="false" />
|
71 |
+
<option name="EMULATE_TERMINAL" value="false" />
|
72 |
+
<option name="MODULE_MODE" value="false" />
|
73 |
+
<option name="REDIRECT_INPUT" value="false" />
|
74 |
+
<option name="INPUT_FILE" value="" />
|
75 |
+
<method v="2" />
|
76 |
+
</configuration>
|
77 |
+
<configuration name="gradio" type="PythonConfigurationType" factoryName="Python" temporary="true" nameIsGenerated="true">
|
78 |
+
<module name="src" />
|
79 |
+
<option name="ENV_FILES" value="" />
|
80 |
+
<option name="INTERPRETER_OPTIONS" value="" />
|
81 |
+
<option name="PARENT_ENVS" value="true" />
|
82 |
+
<option name="SDK_HOME" value="" />
|
83 |
+
<option name="WORKING_DIRECTORY" value="$PROJECT_DIR$/.." />
|
84 |
+
<option name="IS_MODULE_SDK" value="false" />
|
85 |
+
<option name="ADD_CONTENT_ROOTS" value="true" />
|
86 |
+
<option name="ADD_SOURCE_ROOTS" value="true" />
|
87 |
+
<EXTENSION ID="PythonCoverageRunConfigurationExtension" runner="coverage.py" />
|
88 |
+
<option name="SCRIPT_NAME" value="$PROJECT_DIR$/../gradio.py" />
|
89 |
+
<option name="PARAMETERS" value="" />
|
90 |
+
<option name="SHOW_COMMAND_LINE" value="false" />
|
91 |
+
<option name="EMULATE_TERMINAL" value="false" />
|
92 |
+
<option name="MODULE_MODE" value="false" />
|
93 |
+
<option name="REDIRECT_INPUT" value="false" />
|
94 |
+
<option name="INPUT_FILE" value="" />
|
95 |
+
<method v="2" />
|
96 |
+
</configuration>
|
97 |
+
<recent_temporary>
|
98 |
+
<list>
|
99 |
+
<item itemvalue="Python.api" />
|
100 |
+
<item itemvalue="Python.gradio" />
|
101 |
+
<item itemvalue="Python.clean" />
|
102 |
+
</list>
|
103 |
+
</recent_temporary>
|
104 |
+
</component>
|
105 |
+
<component name="SharedIndexes">
|
106 |
+
<attachedChunks>
|
107 |
+
<set>
|
108 |
+
<option value="bundled-js-predefined-1d06a55b98c1-74d2a5396914-JavaScript-PY-241.14494.241" />
|
109 |
+
<option value="bundled-python-sdk-0509580d9d50-28c9f5db9ffe-com.jetbrains.pycharm.pro.sharedIndexes.bundled-PY-241.14494.241" />
|
110 |
+
</set>
|
111 |
+
</attachedChunks>
|
112 |
+
</component>
|
113 |
+
<component name="SpellCheckerSettings" RuntimeDictionaries="0" Folders="0" CustomDictionaries="0" DefaultDictionary="application-level" UseSingleDictionary="true" transferred="true" />
|
114 |
+
<component name="TaskManager">
|
115 |
+
<task active="true" id="Default" summary="Default task">
|
116 |
+
<changelist id="cb82860d-7ce6-451e-932b-96d3a6e7b20d" name="Changes" comment="" />
|
117 |
+
<created>1726457759523</created>
|
118 |
+
<option name="number" value="Default" />
|
119 |
+
<option name="presentableId" value="Default" />
|
120 |
+
<updated>1726457759523</updated>
|
121 |
+
<workItem from="1726457760668" duration="3668000" />
|
122 |
+
</task>
|
123 |
+
<servers />
|
124 |
+
</component>
|
125 |
+
<component name="TypeScriptGeneratedFilesManager">
|
126 |
+
<option name="version" value="3" />
|
127 |
+
</component>
|
128 |
+
</project>
|
src/inference.py
ADDED
@@ -0,0 +1,169 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import random
|
3 |
+
import pandas as pd
|
4 |
+
import torch
|
5 |
+
import librosa
|
6 |
+
import numpy as np
|
7 |
+
import soundfile as sf
|
8 |
+
from tqdm import tqdm
|
9 |
+
from utils import scale_shift_re
|
10 |
+
|
11 |
+
|
12 |
+
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
|
13 |
+
"""
|
14 |
+
Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
|
15 |
+
Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
|
16 |
+
"""
|
17 |
+
std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
|
18 |
+
std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
|
19 |
+
# rescale the results from guidance (fixes overexposure)
|
20 |
+
noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
|
21 |
+
# mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
|
22 |
+
noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
|
23 |
+
return noise_cfg
|
24 |
+
|
25 |
+
|
26 |
+
@torch.no_grad()
|
27 |
+
def inference(autoencoder, unet, gt, gt_mask,
|
28 |
+
tokenizer, text_encoder,
|
29 |
+
params, noise_scheduler,
|
30 |
+
text_raw, neg_text=None,
|
31 |
+
audio_frames=500,
|
32 |
+
guidance_scale=3, guidance_rescale=0.0,
|
33 |
+
ddim_steps=50, eta=1, random_seed=2024,
|
34 |
+
device='cuda',
|
35 |
+
):
|
36 |
+
if neg_text is None:
|
37 |
+
neg_text = [""]
|
38 |
+
if tokenizer is not None:
|
39 |
+
text_batch = tokenizer(text_raw,
|
40 |
+
max_length=params['text_encoder']['max_length'],
|
41 |
+
padding="max_length", truncation=True, return_tensors="pt")
|
42 |
+
text, text_mask = text_batch.input_ids.to(device), text_batch.attention_mask.to(device).bool()
|
43 |
+
text = text_encoder(input_ids=text, attention_mask=text_mask).last_hidden_state
|
44 |
+
|
45 |
+
uncond_text_batch = tokenizer(neg_text,
|
46 |
+
max_length=params['text_encoder']['max_length'],
|
47 |
+
padding="max_length", truncation=True, return_tensors="pt")
|
48 |
+
uncond_text, uncond_text_mask = uncond_text_batch.input_ids.to(device), uncond_text_batch.attention_mask.to(device).bool()
|
49 |
+
uncond_text = text_encoder(input_ids=uncond_text,
|
50 |
+
attention_mask=uncond_text_mask).last_hidden_state
|
51 |
+
else:
|
52 |
+
text, text_mask = None, None
|
53 |
+
guidance_scale = None
|
54 |
+
|
55 |
+
codec_dim = params['model']['out_chans']
|
56 |
+
unet.eval()
|
57 |
+
|
58 |
+
if random_seed is not None:
|
59 |
+
generator = torch.Generator(device=device).manual_seed(random_seed)
|
60 |
+
else:
|
61 |
+
generator = torch.Generator(device=device)
|
62 |
+
generator.seed()
|
63 |
+
|
64 |
+
noise_scheduler.set_timesteps(ddim_steps)
|
65 |
+
|
66 |
+
# init noise
|
67 |
+
noise = torch.randn((1, codec_dim, audio_frames), generator=generator, device=device)
|
68 |
+
latents = noise
|
69 |
+
|
70 |
+
for t in noise_scheduler.timesteps:
|
71 |
+
latents = noise_scheduler.scale_model_input(latents, t)
|
72 |
+
|
73 |
+
if guidance_scale:
|
74 |
+
|
75 |
+
latents_combined = torch.cat([latents, latents], dim=0)
|
76 |
+
text_combined = torch.cat([text, uncond_text], dim=0)
|
77 |
+
text_mask_combined = torch.cat([text_mask, uncond_text_mask], dim=0)
|
78 |
+
|
79 |
+
if gt is not None:
|
80 |
+
gt_combined = torch.cat([gt, gt], dim=0)
|
81 |
+
gt_mask_combined = torch.cat([gt_mask, gt_mask], dim=0)
|
82 |
+
else:
|
83 |
+
gt_combined = None
|
84 |
+
gt_mask_combined = None
|
85 |
+
|
86 |
+
output_combined, _ = unet(latents_combined, t, text_combined, context_mask=text_mask_combined,
|
87 |
+
cls_token=None, gt=gt_combined, mae_mask_infer=gt_mask_combined)
|
88 |
+
output_text, output_uncond = torch.chunk(output_combined, 2, dim=0)
|
89 |
+
|
90 |
+
output_pred = output_uncond + guidance_scale * (output_text - output_uncond)
|
91 |
+
if guidance_rescale > 0.0:
|
92 |
+
output_pred = rescale_noise_cfg(output_pred, output_text,
|
93 |
+
guidance_rescale=guidance_rescale)
|
94 |
+
else:
|
95 |
+
output_pred, mae_mask = unet(latents, t, text, context_mask=text_mask,
|
96 |
+
cls_token=None, gt=gt, mae_mask_infer=gt_mask)
|
97 |
+
|
98 |
+
latents = noise_scheduler.step(model_output=output_pred, timestep=t,
|
99 |
+
sample=latents,
|
100 |
+
eta=eta, generator=generator).prev_sample
|
101 |
+
|
102 |
+
pred = scale_shift_re(latents, params['autoencoder']['scale'],
|
103 |
+
params['autoencoder']['shift'])
|
104 |
+
if gt is not None:
|
105 |
+
pred[~gt_mask] = gt[~gt_mask]
|
106 |
+
pred_wav = autoencoder(embedding=pred)
|
107 |
+
return pred_wav
|
108 |
+
|
109 |
+
|
110 |
+
@torch.no_grad()
|
111 |
+
def eval_udit(autoencoder, unet,
|
112 |
+
tokenizer, text_encoder,
|
113 |
+
params, noise_scheduler,
|
114 |
+
val_df, subset,
|
115 |
+
audio_frames, mae=False,
|
116 |
+
guidance_scale=3, guidance_rescale=0.0,
|
117 |
+
ddim_steps=50, eta=1, random_seed=2023,
|
118 |
+
device='cuda',
|
119 |
+
epoch=0, save_path='logs/eval/', val_num=5):
|
120 |
+
val_df = pd.read_csv(val_df)
|
121 |
+
val_df = val_df[val_df['split'] == subset]
|
122 |
+
if mae:
|
123 |
+
val_df = val_df[val_df['audio_length'] != 0]
|
124 |
+
|
125 |
+
save_path = save_path + str(epoch) + '/'
|
126 |
+
os.makedirs(save_path, exist_ok=True)
|
127 |
+
|
128 |
+
for i in tqdm(range(len(val_df))):
|
129 |
+
row = val_df.iloc[i]
|
130 |
+
text = [row['caption']]
|
131 |
+
if mae:
|
132 |
+
audio_path = params['data']['val_dir'] + str(row['audio_path'])
|
133 |
+
gt, sr = librosa.load(audio_path, sr=params['data']['sr'])
|
134 |
+
gt = gt / (np.max(np.abs(gt)) + 1e-9)
|
135 |
+
sf.write(save_path + text[0] + '_gt.wav', gt, samplerate=params['data']['sr'])
|
136 |
+
num_samples = 10 * sr
|
137 |
+
if len(gt) < num_samples:
|
138 |
+
padding = num_samples - len(gt)
|
139 |
+
gt = np.pad(gt, (0, padding), 'constant')
|
140 |
+
else:
|
141 |
+
gt = gt[:num_samples]
|
142 |
+
gt = torch.tensor(gt).unsqueeze(0).unsqueeze(1).to(device)
|
143 |
+
gt = autoencoder(audio=gt)
|
144 |
+
B, D, L = gt.shape
|
145 |
+
mask_len = int(L * 0.2)
|
146 |
+
gt_mask = torch.zeros(B, D, L).to(device)
|
147 |
+
for _ in range(2):
|
148 |
+
start = random.randint(0, L - mask_len)
|
149 |
+
gt_mask[:, :, start:start + mask_len] = 1
|
150 |
+
gt_mask = gt_mask.bool()
|
151 |
+
else:
|
152 |
+
gt = None
|
153 |
+
gt_mask = None
|
154 |
+
|
155 |
+
pred = inference(autoencoder, unet, gt, gt_mask,
|
156 |
+
tokenizer, text_encoder,
|
157 |
+
params, noise_scheduler,
|
158 |
+
text, neg_text=None,
|
159 |
+
audio_frames=audio_frames,
|
160 |
+
guidance_scale=guidance_scale, guidance_rescale=guidance_rescale,
|
161 |
+
ddim_steps=ddim_steps, eta=eta, random_seed=random_seed,
|
162 |
+
device=device)
|
163 |
+
|
164 |
+
pred = pred.cpu().numpy().squeeze(0).squeeze(0)
|
165 |
+
|
166 |
+
sf.write(save_path + text[0] + '.wav', pred, samplerate=params['data']['sr'])
|
167 |
+
|
168 |
+
if i + 1 >= val_num:
|
169 |
+
break
|
src/models/blocks.py
ADDED
@@ -0,0 +1,325 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
from torch.utils.checkpoint import checkpoint
|
4 |
+
from .utils.attention import Attention, JointAttention
|
5 |
+
from .utils.modules import unpatchify, FeedForward
|
6 |
+
from .utils.modules import film_modulate
|
7 |
+
|
8 |
+
|
9 |
+
class AdaLN(nn.Module):
|
10 |
+
def __init__(self, dim, ada_mode='ada', r=None, alpha=None):
|
11 |
+
super().__init__()
|
12 |
+
self.ada_mode = ada_mode
|
13 |
+
self.scale_shift_table = None
|
14 |
+
if ada_mode == 'ada':
|
15 |
+
# move nn.silu outside
|
16 |
+
self.time_ada = nn.Linear(dim, 6 * dim, bias=True)
|
17 |
+
elif ada_mode == 'ada_single':
|
18 |
+
# adaln used in pixel-art alpha
|
19 |
+
self.scale_shift_table = nn.Parameter(torch.zeros(6, dim))
|
20 |
+
elif ada_mode in ['ada_lora', 'ada_lora_bias']:
|
21 |
+
self.lora_a = nn.Linear(dim, r * 6, bias=False)
|
22 |
+
self.lora_b = nn.Linear(r * 6, dim * 6, bias=False)
|
23 |
+
self.scaling = alpha / r
|
24 |
+
if ada_mode == 'ada_lora_bias':
|
25 |
+
# take bias out for consistency
|
26 |
+
self.scale_shift_table = nn.Parameter(torch.zeros(6, dim))
|
27 |
+
else:
|
28 |
+
raise NotImplementedError
|
29 |
+
|
30 |
+
def forward(self, time_token=None, time_ada=None):
|
31 |
+
if self.ada_mode == 'ada':
|
32 |
+
assert time_ada is None
|
33 |
+
B = time_token.shape[0]
|
34 |
+
time_ada = self.time_ada(time_token).reshape(B, 6, -1)
|
35 |
+
elif self.ada_mode == 'ada_single':
|
36 |
+
B = time_ada.shape[0]
|
37 |
+
time_ada = time_ada.reshape(B, 6, -1)
|
38 |
+
time_ada = self.scale_shift_table[None] + time_ada
|
39 |
+
elif self.ada_mode in ['ada_lora', 'ada_lora_bias']:
|
40 |
+
B = time_ada.shape[0]
|
41 |
+
time_ada_lora = self.lora_b(self.lora_a(time_token)) * self.scaling
|
42 |
+
time_ada = time_ada + time_ada_lora
|
43 |
+
time_ada = time_ada.reshape(B, 6, -1)
|
44 |
+
if self.scale_shift_table is not None:
|
45 |
+
time_ada = self.scale_shift_table[None] + time_ada
|
46 |
+
else:
|
47 |
+
raise NotImplementedError
|
48 |
+
return time_ada
|
49 |
+
|
50 |
+
|
51 |
+
class DiTBlock(nn.Module):
|
52 |
+
"""
|
53 |
+
A modified PixArt block with adaptive layer norm (adaLN-single) conditioning.
|
54 |
+
"""
|
55 |
+
|
56 |
+
def __init__(self, dim, context_dim=None,
|
57 |
+
num_heads=8, mlp_ratio=4.,
|
58 |
+
qkv_bias=False, qk_scale=None, qk_norm=None,
|
59 |
+
act_layer='gelu', norm_layer=nn.LayerNorm,
|
60 |
+
time_fusion='none',
|
61 |
+
ada_lora_rank=None, ada_lora_alpha=None,
|
62 |
+
skip=False, skip_norm=False,
|
63 |
+
rope_mode='none',
|
64 |
+
context_norm=False,
|
65 |
+
use_checkpoint=False):
|
66 |
+
|
67 |
+
super().__init__()
|
68 |
+
self.norm1 = norm_layer(dim)
|
69 |
+
self.attn = Attention(dim=dim,
|
70 |
+
num_heads=num_heads,
|
71 |
+
qkv_bias=qkv_bias, qk_scale=qk_scale,
|
72 |
+
qk_norm=qk_norm,
|
73 |
+
rope_mode=rope_mode)
|
74 |
+
|
75 |
+
if context_dim is not None:
|
76 |
+
self.use_context = True
|
77 |
+
self.cross_attn = Attention(dim=dim,
|
78 |
+
num_heads=num_heads,
|
79 |
+
context_dim=context_dim,
|
80 |
+
qkv_bias=qkv_bias, qk_scale=qk_scale,
|
81 |
+
qk_norm=qk_norm,
|
82 |
+
rope_mode='none')
|
83 |
+
self.norm2 = norm_layer(dim)
|
84 |
+
if context_norm:
|
85 |
+
self.norm_context = norm_layer(context_dim)
|
86 |
+
else:
|
87 |
+
self.norm_context = nn.Identity()
|
88 |
+
else:
|
89 |
+
self.use_context = False
|
90 |
+
|
91 |
+
self.norm3 = norm_layer(dim)
|
92 |
+
self.mlp = FeedForward(dim=dim, mult=mlp_ratio,
|
93 |
+
activation_fn=act_layer, dropout=0)
|
94 |
+
|
95 |
+
self.use_adanorm = True if time_fusion != 'token' else False
|
96 |
+
if self.use_adanorm:
|
97 |
+
self.adaln = AdaLN(dim, ada_mode=time_fusion,
|
98 |
+
r=ada_lora_rank, alpha=ada_lora_alpha)
|
99 |
+
if skip:
|
100 |
+
self.skip_norm = norm_layer(2 * dim) if skip_norm else nn.Identity()
|
101 |
+
self.skip_linear = nn.Linear(2 * dim, dim)
|
102 |
+
else:
|
103 |
+
self.skip_linear = None
|
104 |
+
|
105 |
+
self.use_checkpoint = use_checkpoint
|
106 |
+
|
107 |
+
def forward(self, x, time_token=None, time_ada=None,
|
108 |
+
skip=None, context=None,
|
109 |
+
x_mask=None, context_mask=None, extras=None):
|
110 |
+
if self.use_checkpoint:
|
111 |
+
return checkpoint(self._forward, x,
|
112 |
+
time_token, time_ada, skip, context,
|
113 |
+
x_mask, context_mask, extras,
|
114 |
+
use_reentrant=False)
|
115 |
+
else:
|
116 |
+
return self._forward(x,
|
117 |
+
time_token, time_ada, skip, context,
|
118 |
+
x_mask, context_mask, extras)
|
119 |
+
|
120 |
+
def _forward(self, x, time_token=None, time_ada=None,
|
121 |
+
skip=None, context=None,
|
122 |
+
x_mask=None, context_mask=None, extras=None):
|
123 |
+
B, T, C = x.shape
|
124 |
+
if self.skip_linear is not None:
|
125 |
+
assert skip is not None
|
126 |
+
cat = torch.cat([x, skip], dim=-1)
|
127 |
+
cat = self.skip_norm(cat)
|
128 |
+
x = self.skip_linear(cat)
|
129 |
+
|
130 |
+
if self.use_adanorm:
|
131 |
+
time_ada = self.adaln(time_token, time_ada)
|
132 |
+
(shift_msa, scale_msa, gate_msa,
|
133 |
+
shift_mlp, scale_mlp, gate_mlp) = time_ada.chunk(6, dim=1)
|
134 |
+
|
135 |
+
# self attention
|
136 |
+
if self.use_adanorm:
|
137 |
+
x_norm = film_modulate(self.norm1(x), shift=shift_msa,
|
138 |
+
scale=scale_msa)
|
139 |
+
x = x + (1 - gate_msa) * self.attn(x_norm, context=None,
|
140 |
+
context_mask=x_mask,
|
141 |
+
extras=extras)
|
142 |
+
else:
|
143 |
+
x = x + self.attn(self.norm1(x), context=None, context_mask=x_mask,
|
144 |
+
extras=extras)
|
145 |
+
|
146 |
+
# cross attention
|
147 |
+
if self.use_context:
|
148 |
+
assert context is not None
|
149 |
+
x = x + self.cross_attn(x=self.norm2(x),
|
150 |
+
context=self.norm_context(context),
|
151 |
+
context_mask=context_mask, extras=extras)
|
152 |
+
|
153 |
+
# mlp
|
154 |
+
if self.use_adanorm:
|
155 |
+
x_norm = film_modulate(self.norm3(x), shift=shift_mlp, scale=scale_mlp)
|
156 |
+
x = x + (1 - gate_mlp) * self.mlp(x_norm)
|
157 |
+
else:
|
158 |
+
x = x + self.mlp(self.norm3(x))
|
159 |
+
|
160 |
+
return x
|
161 |
+
|
162 |
+
|
163 |
+
class JointDiTBlock(nn.Module):
|
164 |
+
"""
|
165 |
+
A modified PixArt block with adaptive layer norm (adaLN-single) conditioning.
|
166 |
+
"""
|
167 |
+
|
168 |
+
def __init__(self, dim, context_dim=None,
|
169 |
+
num_heads=8, mlp_ratio=4.,
|
170 |
+
qkv_bias=False, qk_scale=None, qk_norm=None,
|
171 |
+
act_layer='gelu', norm_layer=nn.LayerNorm,
|
172 |
+
time_fusion='none',
|
173 |
+
ada_lora_rank=None, ada_lora_alpha=None,
|
174 |
+
skip=(False, False),
|
175 |
+
rope_mode=False,
|
176 |
+
context_norm=False,
|
177 |
+
use_checkpoint=False,):
|
178 |
+
|
179 |
+
super().__init__()
|
180 |
+
# no cross attention
|
181 |
+
assert context_dim is None
|
182 |
+
self.attn_norm_x = norm_layer(dim)
|
183 |
+
self.attn_norm_c = norm_layer(dim)
|
184 |
+
self.attn = JointAttention(dim=dim,
|
185 |
+
num_heads=num_heads,
|
186 |
+
qkv_bias=qkv_bias, qk_scale=qk_scale,
|
187 |
+
qk_norm=qk_norm,
|
188 |
+
rope_mode=rope_mode)
|
189 |
+
self.ffn_norm_x = norm_layer(dim)
|
190 |
+
self.ffn_norm_c = norm_layer(dim)
|
191 |
+
self.mlp_x = FeedForward(dim=dim, mult=mlp_ratio,
|
192 |
+
activation_fn=act_layer, dropout=0)
|
193 |
+
self.mlp_c = FeedForward(dim=dim, mult=mlp_ratio,
|
194 |
+
activation_fn=act_layer, dropout=0)
|
195 |
+
|
196 |
+
# Zero-out the shift table
|
197 |
+
self.use_adanorm = True if time_fusion != 'token' else False
|
198 |
+
if self.use_adanorm:
|
199 |
+
self.adaln = AdaLN(dim, ada_mode=time_fusion,
|
200 |
+
r=ada_lora_rank, alpha=ada_lora_alpha)
|
201 |
+
|
202 |
+
if skip is False:
|
203 |
+
skip_x, skip_c = False, False
|
204 |
+
else:
|
205 |
+
skip_x, skip_c = skip
|
206 |
+
|
207 |
+
self.skip_linear_x = nn.Linear(2 * dim, dim) if skip_x else None
|
208 |
+
self.skip_linear_c = nn.Linear(2 * dim, dim) if skip_c else None
|
209 |
+
|
210 |
+
self.use_checkpoint = use_checkpoint
|
211 |
+
|
212 |
+
def forward(self, x, time_token=None, time_ada=None,
|
213 |
+
skip=None, context=None,
|
214 |
+
x_mask=None, context_mask=None, extras=None):
|
215 |
+
if self.use_checkpoint:
|
216 |
+
return checkpoint(self._forward, x,
|
217 |
+
time_token, time_ada, skip,
|
218 |
+
context, x_mask, context_mask, extras,
|
219 |
+
use_reentrant=False)
|
220 |
+
else:
|
221 |
+
return self._forward(x,
|
222 |
+
time_token, time_ada, skip,
|
223 |
+
context, x_mask, context_mask, extras)
|
224 |
+
|
225 |
+
def _forward(self, x, time_token=None, time_ada=None,
|
226 |
+
skip=None, context=None,
|
227 |
+
x_mask=None, context_mask=None, extras=None):
|
228 |
+
|
229 |
+
assert context is None and context_mask is None
|
230 |
+
|
231 |
+
context, x = x[:, :extras, :], x[:, extras:, :]
|
232 |
+
context_mask, x_mask = x_mask[:, :extras], x_mask[:, extras:]
|
233 |
+
|
234 |
+
if skip is not None:
|
235 |
+
skip_c, skip_x = skip[:, :extras, :], skip[:, extras:, :]
|
236 |
+
|
237 |
+
B, T, C = x.shape
|
238 |
+
if self.skip_linear_x is not None:
|
239 |
+
x = self.skip_linear_x(torch.cat([x, skip_x], dim=-1))
|
240 |
+
|
241 |
+
if self.skip_linear_c is not None:
|
242 |
+
context = self.skip_linear_c(torch.cat([context, skip_c], dim=-1))
|
243 |
+
|
244 |
+
if self.use_adanorm:
|
245 |
+
time_ada = self.adaln(time_token, time_ada)
|
246 |
+
(shift_msa, scale_msa, gate_msa,
|
247 |
+
shift_mlp, scale_mlp, gate_mlp) = time_ada.chunk(6, dim=1)
|
248 |
+
|
249 |
+
# self attention
|
250 |
+
x_norm = self.attn_norm_x(x)
|
251 |
+
c_norm = self.attn_norm_c(context)
|
252 |
+
if self.use_adanorm:
|
253 |
+
x_norm = film_modulate(x_norm, shift=shift_msa, scale=scale_msa)
|
254 |
+
x_out, c_out = self.attn(x_norm, context=c_norm,
|
255 |
+
x_mask=x_mask, context_mask=context_mask,
|
256 |
+
extras=extras)
|
257 |
+
if self.use_adanorm:
|
258 |
+
x = x + (1 - gate_msa) * x_out
|
259 |
+
else:
|
260 |
+
x = x + x_out
|
261 |
+
context = context + c_out
|
262 |
+
|
263 |
+
# mlp
|
264 |
+
if self.use_adanorm:
|
265 |
+
x_norm = film_modulate(self.ffn_norm_x(x),
|
266 |
+
shift=shift_mlp, scale=scale_mlp)
|
267 |
+
x = x + (1 - gate_mlp) * self.mlp_x(x_norm)
|
268 |
+
else:
|
269 |
+
x = x + self.mlp_x(self.ffn_norm_x(x))
|
270 |
+
|
271 |
+
c_norm = self.ffn_norm_c(context)
|
272 |
+
context = context + self.mlp_c(c_norm)
|
273 |
+
|
274 |
+
return torch.cat((context, x), dim=1)
|
275 |
+
|
276 |
+
|
277 |
+
class FinalBlock(nn.Module):
|
278 |
+
def __init__(self, embed_dim, patch_size, in_chans,
|
279 |
+
img_size,
|
280 |
+
input_type='2d',
|
281 |
+
norm_layer=nn.LayerNorm,
|
282 |
+
use_conv=True,
|
283 |
+
use_adanorm=True):
|
284 |
+
super().__init__()
|
285 |
+
self.in_chans = in_chans
|
286 |
+
self.img_size = img_size
|
287 |
+
self.input_type = input_type
|
288 |
+
|
289 |
+
self.norm = norm_layer(embed_dim)
|
290 |
+
if use_adanorm:
|
291 |
+
self.use_adanorm = True
|
292 |
+
else:
|
293 |
+
self.use_adanorm = False
|
294 |
+
|
295 |
+
if input_type == '2d':
|
296 |
+
self.patch_dim = patch_size ** 2 * in_chans
|
297 |
+
self.linear = nn.Linear(embed_dim, self.patch_dim, bias=True)
|
298 |
+
if use_conv:
|
299 |
+
self.final_layer = nn.Conv2d(self.in_chans, self.in_chans,
|
300 |
+
3, padding=1)
|
301 |
+
else:
|
302 |
+
self.final_layer = nn.Identity()
|
303 |
+
|
304 |
+
elif input_type == '1d':
|
305 |
+
self.patch_dim = patch_size * in_chans
|
306 |
+
self.linear = nn.Linear(embed_dim, self.patch_dim, bias=True)
|
307 |
+
if use_conv:
|
308 |
+
self.final_layer = nn.Conv1d(self.in_chans, self.in_chans,
|
309 |
+
3, padding=1)
|
310 |
+
else:
|
311 |
+
self.final_layer = nn.Identity()
|
312 |
+
|
313 |
+
def forward(self, x, time_ada=None, extras=0):
|
314 |
+
B, T, C = x.shape
|
315 |
+
x = x[:, extras:, :]
|
316 |
+
# only handle generation target
|
317 |
+
if self.use_adanorm:
|
318 |
+
shift, scale = time_ada.reshape(B, 2, -1).chunk(2, dim=1)
|
319 |
+
x = film_modulate(self.norm(x), shift, scale)
|
320 |
+
else:
|
321 |
+
x = self.norm(x)
|
322 |
+
x = self.linear(x)
|
323 |
+
x = unpatchify(x, self.in_chans, self.input_type, self.img_size)
|
324 |
+
x = self.final_layer(x)
|
325 |
+
return x
|
src/models/conditioners.py
ADDED
@@ -0,0 +1,180 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from einops import repeat
|
5 |
+
import math
|
6 |
+
from .udit import UDiT
|
7 |
+
from .utils.span_mask import compute_mask_indices
|
8 |
+
|
9 |
+
|
10 |
+
class EmbeddingCFG(nn.Module):
|
11 |
+
"""
|
12 |
+
Handles label dropout for classifier-free guidance.
|
13 |
+
"""
|
14 |
+
# todo: support 2D input
|
15 |
+
|
16 |
+
def __init__(self, in_channels):
|
17 |
+
super().__init__()
|
18 |
+
self.cfg_embedding = nn.Parameter(
|
19 |
+
torch.randn(in_channels) / in_channels ** 0.5)
|
20 |
+
|
21 |
+
def token_drop(self, condition, condition_mask, cfg_prob):
|
22 |
+
"""
|
23 |
+
Drops labels to enable classifier-free guidance.
|
24 |
+
"""
|
25 |
+
b, t, device = condition.shape[0], condition.shape[1], condition.device
|
26 |
+
drop_ids = torch.rand(b, device=device) < cfg_prob
|
27 |
+
uncond = repeat(self.cfg_embedding, "c -> b t c", b=b, t=t)
|
28 |
+
condition = torch.where(drop_ids[:, None, None], uncond, condition)
|
29 |
+
if condition_mask is not None:
|
30 |
+
condition_mask[drop_ids] = False
|
31 |
+
condition_mask[drop_ids, 0] = True
|
32 |
+
|
33 |
+
return condition, condition_mask
|
34 |
+
|
35 |
+
def forward(self, condition, condition_mask, cfg_prob=0.0):
|
36 |
+
if condition_mask is not None:
|
37 |
+
condition_mask = condition_mask.clone()
|
38 |
+
if cfg_prob > 0:
|
39 |
+
condition, condition_mask = self.token_drop(condition,
|
40 |
+
condition_mask,
|
41 |
+
cfg_prob)
|
42 |
+
return condition, condition_mask
|
43 |
+
|
44 |
+
|
45 |
+
class DiscreteCFG(nn.Module):
|
46 |
+
def __init__(self, replace_id=2):
|
47 |
+
super(DiscreteCFG, self).__init__()
|
48 |
+
self.replace_id = replace_id
|
49 |
+
|
50 |
+
def forward(self, context, context_mask, cfg_prob):
|
51 |
+
context = context.clone()
|
52 |
+
if context_mask is not None:
|
53 |
+
context_mask = context_mask.clone()
|
54 |
+
if cfg_prob > 0:
|
55 |
+
cfg_mask = torch.rand(len(context)) < cfg_prob
|
56 |
+
if torch.any(cfg_mask):
|
57 |
+
context[cfg_mask] = 0
|
58 |
+
context[cfg_mask, 0] = self.replace_id
|
59 |
+
if context_mask is not None:
|
60 |
+
context_mask[cfg_mask] = False
|
61 |
+
context_mask[cfg_mask, 0] = True
|
62 |
+
return context, context_mask
|
63 |
+
|
64 |
+
|
65 |
+
class CFGModel(nn.Module):
|
66 |
+
def __init__(self, context_dim, backbone):
|
67 |
+
super().__init__()
|
68 |
+
self.model = backbone
|
69 |
+
self.context_cfg = EmbeddingCFG(context_dim)
|
70 |
+
|
71 |
+
def forward(self, x, timesteps,
|
72 |
+
context, x_mask=None, context_mask=None,
|
73 |
+
cfg_prob=0.0):
|
74 |
+
context = self.context_cfg(context, cfg_prob)
|
75 |
+
x = self.model(x=x, timesteps=timesteps,
|
76 |
+
context=context,
|
77 |
+
x_mask=x_mask, context_mask=context_mask)
|
78 |
+
return x
|
79 |
+
|
80 |
+
|
81 |
+
class ConcatModel(nn.Module):
|
82 |
+
def __init__(self, backbone, in_dim, stride=[]):
|
83 |
+
super().__init__()
|
84 |
+
self.model = backbone
|
85 |
+
|
86 |
+
self.downsample_layers = nn.ModuleList()
|
87 |
+
for i, s in enumerate(stride):
|
88 |
+
downsample_layer = nn.Conv1d(
|
89 |
+
in_dim,
|
90 |
+
in_dim * 2,
|
91 |
+
kernel_size=2 * s,
|
92 |
+
stride=s,
|
93 |
+
padding=math.ceil(s / 2),
|
94 |
+
)
|
95 |
+
self.downsample_layers.append(downsample_layer)
|
96 |
+
in_dim = in_dim * 2
|
97 |
+
|
98 |
+
self.context_cfg = EmbeddingCFG(in_dim)
|
99 |
+
|
100 |
+
def forward(self, x, timesteps,
|
101 |
+
context, x_mask=None,
|
102 |
+
cfg=False, cfg_prob=0.0):
|
103 |
+
|
104 |
+
# todo: support 2D input
|
105 |
+
# x: B, C, L
|
106 |
+
# context: B, C, L
|
107 |
+
|
108 |
+
for downsample_layer in self.downsample_layers:
|
109 |
+
context = downsample_layer(context)
|
110 |
+
|
111 |
+
context = context.transpose(1, 2)
|
112 |
+
context = self.context_cfg(caption=context,
|
113 |
+
cfg=cfg, cfg_prob=cfg_prob)
|
114 |
+
context = context.transpose(1, 2)
|
115 |
+
|
116 |
+
assert context.shape[-1] == x.shape[-1]
|
117 |
+
x = torch.cat([context, x], dim=1)
|
118 |
+
x = self.model(x=x, timesteps=timesteps,
|
119 |
+
context=None, x_mask=x_mask, context_mask=None)
|
120 |
+
return x
|
121 |
+
|
122 |
+
|
123 |
+
class MaskDiT(nn.Module):
|
124 |
+
def __init__(self, mae=False, mae_prob=0.5, mask_ratio=[0.25, 1.0], mask_span=10, **kwargs):
|
125 |
+
super().__init__()
|
126 |
+
self.model = UDiT(**kwargs)
|
127 |
+
self.mae = mae
|
128 |
+
if self.mae:
|
129 |
+
out_channel = kwargs.pop('out_chans', None)
|
130 |
+
self.mask_embed = nn.Parameter(torch.zeros((out_channel)))
|
131 |
+
self.mae_prob = mae_prob
|
132 |
+
self.mask_ratio = mask_ratio
|
133 |
+
self.mask_span = mask_span
|
134 |
+
|
135 |
+
def random_masking(self, gt, mask_ratios, mae_mask_infer=None):
|
136 |
+
B, D, L = gt.shape
|
137 |
+
if mae_mask_infer is None:
|
138 |
+
# mask = torch.rand(B, L).to(gt.device) < mask_ratios.unsqueeze(1)
|
139 |
+
mask_ratios = mask_ratios.cpu().numpy()
|
140 |
+
mask = compute_mask_indices(shape=[B, L],
|
141 |
+
padding_mask=None,
|
142 |
+
mask_prob=mask_ratios,
|
143 |
+
mask_length=self.mask_span,
|
144 |
+
mask_type="static",
|
145 |
+
mask_other=0.0,
|
146 |
+
min_masks=1,
|
147 |
+
no_overlap=False,
|
148 |
+
min_space=0,)
|
149 |
+
mask = mask.unsqueeze(1).expand_as(gt)
|
150 |
+
else:
|
151 |
+
mask = mae_mask_infer
|
152 |
+
mask = mask.expand_as(gt)
|
153 |
+
gt[mask] = self.mask_embed.view(1, D, 1).expand_as(gt)[mask]
|
154 |
+
return gt, mask.type_as(gt)
|
155 |
+
|
156 |
+
def forward(self, x, timesteps, context,
|
157 |
+
x_mask=None, context_mask=None, cls_token=None,
|
158 |
+
gt=None, mae_mask_infer=None):
|
159 |
+
mae_mask = torch.ones_like(x)
|
160 |
+
if self.mae:
|
161 |
+
if gt is not None:
|
162 |
+
B, D, L = gt.shape
|
163 |
+
mask_ratios = torch.FloatTensor(B).uniform_(*self.mask_ratio).to(gt.device)
|
164 |
+
gt, mae_mask = self.random_masking(gt, mask_ratios, mae_mask_infer)
|
165 |
+
# apply mae only to the selected batches
|
166 |
+
if mae_mask_infer is None:
|
167 |
+
# determine mae batch
|
168 |
+
mae_batch = torch.rand(B) < self.mae_prob
|
169 |
+
gt[~mae_batch] = self.mask_embed.view(1, D, 1).expand_as(gt)[~mae_batch]
|
170 |
+
mae_mask[~mae_batch] = 1.0
|
171 |
+
else:
|
172 |
+
B, D, L = x.shape
|
173 |
+
gt = self.mask_embed.view(1, D, 1).expand_as(x)
|
174 |
+
x = torch.cat([x, gt, mae_mask[:, 0:1, :]], dim=1)
|
175 |
+
|
176 |
+
x = self.model(x=x, timesteps=timesteps, context=context,
|
177 |
+
x_mask=x_mask, context_mask=context_mask,
|
178 |
+
cls_token=cls_token)
|
179 |
+
# print(mae_mask[:, 0, :].sum(dim=-1))
|
180 |
+
return x, mae_mask
|
src/models/udit.py
ADDED
@@ -0,0 +1,356 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.utils.checkpoint
|
4 |
+
import math
|
5 |
+
from .utils.modules import PatchEmbed, TimestepEmbedder
|
6 |
+
from .utils.modules import PE_wrapper, RMSNorm
|
7 |
+
from .blocks import DiTBlock, JointDiTBlock, FinalBlock
|
8 |
+
|
9 |
+
|
10 |
+
class UDiT(nn.Module):
|
11 |
+
def __init__(self,
|
12 |
+
img_size=224, patch_size=16, in_chans=3,
|
13 |
+
input_type='2d', out_chans=None,
|
14 |
+
embed_dim=768, depth=12, num_heads=12, mlp_ratio=4.,
|
15 |
+
qkv_bias=False, qk_scale=None, qk_norm=None,
|
16 |
+
act_layer='gelu', norm_layer='layernorm',
|
17 |
+
context_norm=False,
|
18 |
+
use_checkpoint=False,
|
19 |
+
# time fusion ada or token
|
20 |
+
time_fusion='token',
|
21 |
+
ada_lora_rank=None, ada_lora_alpha=None,
|
22 |
+
cls_dim=None,
|
23 |
+
# max length is only used for concat
|
24 |
+
context_dim=768, context_fusion='concat',
|
25 |
+
context_max_length=128, context_pe_method='sinu',
|
26 |
+
pe_method='abs', rope_mode='none',
|
27 |
+
use_conv=True,
|
28 |
+
skip=True, skip_norm=True):
|
29 |
+
super().__init__()
|
30 |
+
self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models
|
31 |
+
|
32 |
+
# input
|
33 |
+
self.in_chans = in_chans
|
34 |
+
self.input_type = input_type
|
35 |
+
if self.input_type == '2d':
|
36 |
+
num_patches = (img_size[0] // patch_size) * (img_size[1] // patch_size)
|
37 |
+
elif self.input_type == '1d':
|
38 |
+
num_patches = img_size // patch_size
|
39 |
+
self.patch_embed = PatchEmbed(patch_size=patch_size, in_chans=in_chans,
|
40 |
+
embed_dim=embed_dim, input_type=input_type)
|
41 |
+
out_chans = in_chans if out_chans is None else out_chans
|
42 |
+
self.out_chans = out_chans
|
43 |
+
|
44 |
+
# position embedding
|
45 |
+
self.rope = rope_mode
|
46 |
+
self.x_pe = PE_wrapper(dim=embed_dim, method=pe_method,
|
47 |
+
length=num_patches)
|
48 |
+
|
49 |
+
print(f'x position embedding: {pe_method}')
|
50 |
+
print(f'rope mode: {self.rope}')
|
51 |
+
|
52 |
+
# time embed
|
53 |
+
self.time_embed = TimestepEmbedder(embed_dim)
|
54 |
+
self.time_fusion = time_fusion
|
55 |
+
self.use_adanorm = False
|
56 |
+
|
57 |
+
# cls embed
|
58 |
+
if cls_dim is not None:
|
59 |
+
self.cls_embed = nn.Sequential(
|
60 |
+
nn.Linear(cls_dim, embed_dim, bias=True),
|
61 |
+
nn.SiLU(),
|
62 |
+
nn.Linear(embed_dim, embed_dim, bias=True),)
|
63 |
+
else:
|
64 |
+
self.cls_embed = None
|
65 |
+
|
66 |
+
# time fusion
|
67 |
+
if time_fusion == 'token':
|
68 |
+
# put token at the beginning of sequence
|
69 |
+
self.extras = 2 if self.cls_embed else 1
|
70 |
+
self.time_pe = PE_wrapper(dim=embed_dim, method='abs', length=self.extras)
|
71 |
+
elif time_fusion in ['ada', 'ada_single', 'ada_lora', 'ada_lora_bias']:
|
72 |
+
self.use_adanorm = True
|
73 |
+
# aviod repetitive silu for each adaln block
|
74 |
+
self.time_act = nn.SiLU()
|
75 |
+
self.extras = 0
|
76 |
+
self.time_ada_final = nn.Linear(embed_dim, 2 * embed_dim, bias=True)
|
77 |
+
if time_fusion in ['ada_single', 'ada_lora', 'ada_lora_bias']:
|
78 |
+
# shared adaln
|
79 |
+
self.time_ada = nn.Linear(embed_dim, 6 * embed_dim, bias=True)
|
80 |
+
else:
|
81 |
+
self.time_ada = None
|
82 |
+
else:
|
83 |
+
raise NotImplementedError
|
84 |
+
print(f'time fusion mode: {self.time_fusion}')
|
85 |
+
|
86 |
+
# context
|
87 |
+
# use a simple projection
|
88 |
+
self.use_context = False
|
89 |
+
self.context_cross = False
|
90 |
+
self.context_max_length = context_max_length
|
91 |
+
self.context_fusion = 'none'
|
92 |
+
if context_dim is not None:
|
93 |
+
self.use_context = True
|
94 |
+
self.context_embed = nn.Sequential(
|
95 |
+
nn.Linear(context_dim, embed_dim, bias=True),
|
96 |
+
nn.SiLU(),
|
97 |
+
nn.Linear(embed_dim, embed_dim, bias=True),)
|
98 |
+
self.context_fusion = context_fusion
|
99 |
+
if context_fusion == 'concat' or context_fusion == 'joint':
|
100 |
+
self.extras += context_max_length
|
101 |
+
self.context_pe = PE_wrapper(dim=embed_dim,
|
102 |
+
method=context_pe_method,
|
103 |
+
length=context_max_length)
|
104 |
+
# no cross attention layers
|
105 |
+
context_dim = None
|
106 |
+
elif context_fusion == 'cross':
|
107 |
+
self.context_pe = PE_wrapper(dim=embed_dim,
|
108 |
+
method=context_pe_method,
|
109 |
+
length=context_max_length)
|
110 |
+
self.context_cross = True
|
111 |
+
context_dim = embed_dim
|
112 |
+
else:
|
113 |
+
raise NotImplementedError
|
114 |
+
print(f'context fusion mode: {context_fusion}')
|
115 |
+
print(f'context position embedding: {context_pe_method}')
|
116 |
+
|
117 |
+
if self.context_fusion == 'joint':
|
118 |
+
Block = JointDiTBlock
|
119 |
+
self.use_skip = skip[0]
|
120 |
+
else:
|
121 |
+
Block = DiTBlock
|
122 |
+
self.use_skip = skip
|
123 |
+
|
124 |
+
# norm layers
|
125 |
+
if norm_layer == 'layernorm':
|
126 |
+
norm_layer = nn.LayerNorm
|
127 |
+
elif norm_layer == 'rmsnorm':
|
128 |
+
norm_layer = RMSNorm
|
129 |
+
else:
|
130 |
+
raise NotImplementedError
|
131 |
+
|
132 |
+
print(f'use long skip connection: {skip}')
|
133 |
+
self.in_blocks = nn.ModuleList([
|
134 |
+
Block(
|
135 |
+
dim=embed_dim, context_dim=context_dim, num_heads=num_heads,
|
136 |
+
mlp_ratio=mlp_ratio,
|
137 |
+
qkv_bias=qkv_bias, qk_scale=qk_scale, qk_norm=qk_norm,
|
138 |
+
act_layer=act_layer, norm_layer=norm_layer,
|
139 |
+
time_fusion=time_fusion,
|
140 |
+
ada_lora_rank=ada_lora_rank, ada_lora_alpha=ada_lora_alpha,
|
141 |
+
skip=False, skip_norm=False,
|
142 |
+
rope_mode=self.rope,
|
143 |
+
context_norm=context_norm,
|
144 |
+
use_checkpoint=use_checkpoint)
|
145 |
+
for _ in range(depth // 2)])
|
146 |
+
|
147 |
+
self.mid_block = Block(
|
148 |
+
dim=embed_dim, context_dim=context_dim, num_heads=num_heads,
|
149 |
+
mlp_ratio=mlp_ratio,
|
150 |
+
qkv_bias=qkv_bias, qk_scale=qk_scale, qk_norm=qk_norm,
|
151 |
+
act_layer=act_layer, norm_layer=norm_layer,
|
152 |
+
time_fusion=time_fusion,
|
153 |
+
ada_lora_rank=ada_lora_rank, ada_lora_alpha=ada_lora_alpha,
|
154 |
+
skip=False, skip_norm=False,
|
155 |
+
rope_mode=self.rope,
|
156 |
+
context_norm=context_norm,
|
157 |
+
use_checkpoint=use_checkpoint)
|
158 |
+
|
159 |
+
self.out_blocks = nn.ModuleList([
|
160 |
+
Block(
|
161 |
+
dim=embed_dim, context_dim=context_dim, num_heads=num_heads,
|
162 |
+
mlp_ratio=mlp_ratio,
|
163 |
+
qkv_bias=qkv_bias, qk_scale=qk_scale, qk_norm=qk_norm,
|
164 |
+
act_layer=act_layer, norm_layer=norm_layer,
|
165 |
+
time_fusion=time_fusion,
|
166 |
+
ada_lora_rank=ada_lora_rank, ada_lora_alpha=ada_lora_alpha,
|
167 |
+
skip=skip, skip_norm=skip_norm,
|
168 |
+
rope_mode=self.rope,
|
169 |
+
context_norm=context_norm,
|
170 |
+
use_checkpoint=use_checkpoint)
|
171 |
+
for _ in range(depth // 2)])
|
172 |
+
|
173 |
+
# FinalLayer block
|
174 |
+
self.use_conv = use_conv
|
175 |
+
self.final_block = FinalBlock(embed_dim=embed_dim,
|
176 |
+
patch_size=patch_size,
|
177 |
+
img_size=img_size,
|
178 |
+
in_chans=out_chans,
|
179 |
+
input_type=input_type,
|
180 |
+
norm_layer=norm_layer,
|
181 |
+
use_conv=use_conv,
|
182 |
+
use_adanorm=self.use_adanorm)
|
183 |
+
self.initialize_weights()
|
184 |
+
|
185 |
+
def _init_ada(self):
|
186 |
+
if self.time_fusion == 'ada':
|
187 |
+
nn.init.constant_(self.time_ada_final.weight, 0)
|
188 |
+
nn.init.constant_(self.time_ada_final.bias, 0)
|
189 |
+
for block in self.in_blocks:
|
190 |
+
nn.init.constant_(block.adaln.time_ada.weight, 0)
|
191 |
+
nn.init.constant_(block.adaln.time_ada.bias, 0)
|
192 |
+
nn.init.constant_(self.mid_block.adaln.time_ada.weight, 0)
|
193 |
+
nn.init.constant_(self.mid_block.adaln.time_ada.bias, 0)
|
194 |
+
for block in self.out_blocks:
|
195 |
+
nn.init.constant_(block.adaln.time_ada.weight, 0)
|
196 |
+
nn.init.constant_(block.adaln.time_ada.bias, 0)
|
197 |
+
elif self.time_fusion == 'ada_single':
|
198 |
+
nn.init.constant_(self.time_ada.weight, 0)
|
199 |
+
nn.init.constant_(self.time_ada.bias, 0)
|
200 |
+
nn.init.constant_(self.time_ada_final.weight, 0)
|
201 |
+
nn.init.constant_(self.time_ada_final.bias, 0)
|
202 |
+
elif self.time_fusion in ['ada_lora', 'ada_lora_bias']:
|
203 |
+
nn.init.constant_(self.time_ada.weight, 0)
|
204 |
+
nn.init.constant_(self.time_ada.bias, 0)
|
205 |
+
nn.init.constant_(self.time_ada_final.weight, 0)
|
206 |
+
nn.init.constant_(self.time_ada_final.bias, 0)
|
207 |
+
for block in self.in_blocks:
|
208 |
+
nn.init.kaiming_uniform_(block.adaln.lora_a.weight,
|
209 |
+
a=math.sqrt(5))
|
210 |
+
nn.init.constant_(block.adaln.lora_b.weight, 0)
|
211 |
+
nn.init.kaiming_uniform_(self.mid_block.adaln.lora_a.weight,
|
212 |
+
a=math.sqrt(5))
|
213 |
+
nn.init.constant_(self.mid_block.adaln.lora_b.weight, 0)
|
214 |
+
for block in self.out_blocks:
|
215 |
+
nn.init.kaiming_uniform_(block.adaln.lora_a.weight,
|
216 |
+
a=math.sqrt(5))
|
217 |
+
nn.init.constant_(block.adaln.lora_b.weight, 0)
|
218 |
+
|
219 |
+
def initialize_weights(self):
|
220 |
+
# Basic init for all layers
|
221 |
+
def _basic_init(module):
|
222 |
+
if isinstance(module, nn.Linear):
|
223 |
+
torch.nn.init.xavier_uniform_(module.weight)
|
224 |
+
if module.bias is not None:
|
225 |
+
nn.init.constant_(module.bias, 0)
|
226 |
+
self.apply(_basic_init)
|
227 |
+
|
228 |
+
# init patch Conv like Linear
|
229 |
+
w = self.patch_embed.proj.weight.data
|
230 |
+
nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
|
231 |
+
nn.init.constant_(self.patch_embed.proj.bias, 0)
|
232 |
+
|
233 |
+
# Zero-out AdaLN
|
234 |
+
if self.use_adanorm:
|
235 |
+
self._init_ada()
|
236 |
+
|
237 |
+
# Zero-out Cross Attention
|
238 |
+
if self.context_cross:
|
239 |
+
for block in self.in_blocks:
|
240 |
+
nn.init.constant_(block.cross_attn.proj.weight, 0)
|
241 |
+
nn.init.constant_(block.cross_attn.proj.bias, 0)
|
242 |
+
nn.init.constant_(self.mid_block.cross_attn.proj.weight, 0)
|
243 |
+
nn.init.constant_(self.mid_block.cross_attn.proj.bias, 0)
|
244 |
+
for block in self.out_blocks:
|
245 |
+
nn.init.constant_(block.cross_attn.proj.weight, 0)
|
246 |
+
nn.init.constant_(block.cross_attn.proj.bias, 0)
|
247 |
+
|
248 |
+
# Zero-out cls embedding
|
249 |
+
if self.cls_embed:
|
250 |
+
if self.use_adanorm:
|
251 |
+
nn.init.constant_(self.cls_embed[-1].weight, 0)
|
252 |
+
nn.init.constant_(self.cls_embed[-1].bias, 0)
|
253 |
+
|
254 |
+
# Zero-out Output
|
255 |
+
# might not zero-out this when using v-prediction
|
256 |
+
# it could be good when using noise-prediction
|
257 |
+
# nn.init.constant_(self.final_block.linear.weight, 0)
|
258 |
+
# nn.init.constant_(self.final_block.linear.bias, 0)
|
259 |
+
# if self.use_conv:
|
260 |
+
# nn.init.constant_(self.final_block.final_layer.weight.data, 0)
|
261 |
+
# nn.init.constant_(self.final_block.final_layer.bias, 0)
|
262 |
+
|
263 |
+
# init out Conv
|
264 |
+
if self.use_conv:
|
265 |
+
nn.init.xavier_uniform_(self.final_block.final_layer.weight)
|
266 |
+
nn.init.constant_(self.final_block.final_layer.bias, 0)
|
267 |
+
|
268 |
+
def _concat_x_context(self, x, context, x_mask=None, context_mask=None):
|
269 |
+
assert context.shape[-2] == self.context_max_length
|
270 |
+
# Check if either x_mask or context_mask is provided
|
271 |
+
B = x.shape[0]
|
272 |
+
# Create default masks if they are not provided
|
273 |
+
if x_mask is None:
|
274 |
+
x_mask = torch.ones(B, x.shape[-2], device=x.device).bool()
|
275 |
+
if context_mask is None:
|
276 |
+
context_mask = torch.ones(B, context.shape[-2],
|
277 |
+
device=context.device).bool()
|
278 |
+
# Concatenate the masks along the second dimension (dim=1)
|
279 |
+
x_mask = torch.cat([context_mask, x_mask], dim=1)
|
280 |
+
# Concatenate context and x along the second dimension (dim=1)
|
281 |
+
x = torch.cat((context, x), dim=1)
|
282 |
+
return x, x_mask
|
283 |
+
|
284 |
+
def forward(self, x, timesteps, context,
|
285 |
+
x_mask=None, context_mask=None,
|
286 |
+
cls_token=None
|
287 |
+
):
|
288 |
+
# make it compatible with int time step during inference
|
289 |
+
if timesteps.dim() == 0:
|
290 |
+
timesteps = timesteps.expand(x.shape[0]).to(x.device, dtype=torch.long)
|
291 |
+
|
292 |
+
x = self.patch_embed(x)
|
293 |
+
x = self.x_pe(x)
|
294 |
+
|
295 |
+
B, L, D = x.shape
|
296 |
+
|
297 |
+
if self.use_context:
|
298 |
+
context_token = self.context_embed(context)
|
299 |
+
context_token = self.context_pe(context_token)
|
300 |
+
if self.context_fusion == 'concat' or self.context_fusion == 'joint':
|
301 |
+
x, x_mask = self._concat_x_context(x=x, context=context_token,
|
302 |
+
x_mask=x_mask,
|
303 |
+
context_mask=context_mask)
|
304 |
+
context_token, context_mask = None, None
|
305 |
+
else:
|
306 |
+
context_token, context_mask = None, None
|
307 |
+
|
308 |
+
time_token = self.time_embed(timesteps)
|
309 |
+
if self.cls_embed:
|
310 |
+
cls_token = self.cls_embed(cls_token)
|
311 |
+
time_ada = None
|
312 |
+
time_ada_final = None
|
313 |
+
if self.use_adanorm:
|
314 |
+
if self.cls_embed:
|
315 |
+
time_token = time_token + cls_token
|
316 |
+
time_token = self.time_act(time_token)
|
317 |
+
time_ada_final = self.time_ada_final(time_token)
|
318 |
+
if self.time_ada is not None:
|
319 |
+
time_ada = self.time_ada(time_token)
|
320 |
+
else:
|
321 |
+
time_token = time_token.unsqueeze(dim=1)
|
322 |
+
if self.cls_embed:
|
323 |
+
cls_token = cls_token.unsqueeze(dim=1)
|
324 |
+
time_token = torch.cat([time_token, cls_token], dim=1)
|
325 |
+
time_token = self.time_pe(time_token)
|
326 |
+
x = torch.cat((time_token, x), dim=1)
|
327 |
+
if x_mask is not None:
|
328 |
+
x_mask = torch.cat(
|
329 |
+
[torch.ones(B, time_token.shape[1], device=x_mask.device).bool(),
|
330 |
+
x_mask], dim=1)
|
331 |
+
time_token = None
|
332 |
+
|
333 |
+
skips = []
|
334 |
+
for blk in self.in_blocks:
|
335 |
+
x = blk(x=x, time_token=time_token, time_ada=time_ada,
|
336 |
+
skip=None, context=context_token,
|
337 |
+
x_mask=x_mask, context_mask=context_mask,
|
338 |
+
extras=self.extras)
|
339 |
+
if self.use_skip:
|
340 |
+
skips.append(x)
|
341 |
+
|
342 |
+
x = self.mid_block(x=x, time_token=time_token, time_ada=time_ada,
|
343 |
+
skip=None, context=context_token,
|
344 |
+
x_mask=x_mask, context_mask=context_mask,
|
345 |
+
extras=self.extras)
|
346 |
+
|
347 |
+
for blk in self.out_blocks:
|
348 |
+
skip = skips.pop() if self.use_skip else None
|
349 |
+
x = blk(x=x, time_token=time_token, time_ada=time_ada,
|
350 |
+
skip=skip, context=context_token,
|
351 |
+
x_mask=x_mask, context_mask=context_mask,
|
352 |
+
extras=self.extras)
|
353 |
+
|
354 |
+
x = self.final_block(x, time_ada=time_ada_final, extras=self.extras)
|
355 |
+
|
356 |
+
return x
|
src/models/utils/.ipynb_checkpoints/__init__-checkpoint.py
ADDED
File without changes
|
src/models/utils/.ipynb_checkpoints/attention-checkpoint.py
ADDED
@@ -0,0 +1,290 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import torch.utils.checkpoint
|
5 |
+
import einops
|
6 |
+
from einops import rearrange, repeat
|
7 |
+
from inspect import isfunction
|
8 |
+
from .rotary import RotaryEmbedding
|
9 |
+
from .modules import RMSNorm
|
10 |
+
|
11 |
+
|
12 |
+
if hasattr(nn.functional, 'scaled_dot_product_attention'):
|
13 |
+
ATTENTION_MODE = 'flash'
|
14 |
+
else:
|
15 |
+
ATTENTION_MODE = 'math'
|
16 |
+
print(f'attention mode is {ATTENTION_MODE}')
|
17 |
+
|
18 |
+
|
19 |
+
def add_mask(sim, mask):
|
20 |
+
b, ndim = sim.shape[0], mask.ndim
|
21 |
+
if ndim == 3:
|
22 |
+
mask = rearrange(mask, "b n m -> b 1 n m")
|
23 |
+
if ndim == 2:
|
24 |
+
mask = repeat(mask, "n m -> b 1 n m", b=b)
|
25 |
+
max_neg_value = -torch.finfo(sim.dtype).max
|
26 |
+
sim = sim.masked_fill(~mask, max_neg_value)
|
27 |
+
return sim
|
28 |
+
|
29 |
+
|
30 |
+
def create_mask(q_shape, k_shape, device, q_mask=None, k_mask=None):
|
31 |
+
def default(val, d):
|
32 |
+
return val if val is not None else (d() if isfunction(d) else d)
|
33 |
+
b, i, j, device = q_shape[0], q_shape[-2], k_shape[-2], device
|
34 |
+
q_mask = default(q_mask, torch.ones((b, i), device=device, dtype=torch.bool))
|
35 |
+
k_mask = default(k_mask, torch.ones((b, j), device=device, dtype=torch.bool))
|
36 |
+
attn_mask = rearrange(q_mask, 'b i -> b 1 i 1') * rearrange(k_mask, 'b j -> b 1 1 j')
|
37 |
+
return attn_mask
|
38 |
+
|
39 |
+
|
40 |
+
class Attention(nn.Module):
|
41 |
+
def __init__(self, dim, context_dim=None, num_heads=8,
|
42 |
+
qkv_bias=False, qk_scale=None, qk_norm=None,
|
43 |
+
attn_drop=0., proj_drop=0., rope_mode='none'):
|
44 |
+
super().__init__()
|
45 |
+
self.num_heads = num_heads
|
46 |
+
head_dim = dim // num_heads
|
47 |
+
self.scale = qk_scale or head_dim ** -0.5
|
48 |
+
|
49 |
+
if context_dim is None:
|
50 |
+
self.cross_attn = False
|
51 |
+
else:
|
52 |
+
self.cross_attn = True
|
53 |
+
|
54 |
+
context_dim = dim if context_dim is None else context_dim
|
55 |
+
|
56 |
+
self.to_q = nn.Linear(dim, dim, bias=qkv_bias)
|
57 |
+
self.to_k = nn.Linear(context_dim, dim, bias=qkv_bias)
|
58 |
+
self.to_v = nn.Linear(context_dim, dim, bias=qkv_bias)
|
59 |
+
|
60 |
+
if qk_norm is None:
|
61 |
+
self.norm_q = nn.Identity()
|
62 |
+
self.norm_k = nn.Identity()
|
63 |
+
elif qk_norm == 'layernorm':
|
64 |
+
self.norm_q = nn.LayerNorm(head_dim)
|
65 |
+
self.norm_k = nn.LayerNorm(head_dim)
|
66 |
+
elif qk_norm == 'rmsnorm':
|
67 |
+
self.norm_q = RMSNorm(head_dim)
|
68 |
+
self.norm_k = RMSNorm(head_dim)
|
69 |
+
else:
|
70 |
+
raise NotImplementedError
|
71 |
+
|
72 |
+
self.attn_drop_p = attn_drop
|
73 |
+
self.attn_drop = nn.Dropout(attn_drop)
|
74 |
+
self.proj = nn.Linear(dim, dim)
|
75 |
+
self.proj_drop = nn.Dropout(proj_drop)
|
76 |
+
|
77 |
+
if self.cross_attn:
|
78 |
+
assert rope_mode == 'none'
|
79 |
+
self.rope_mode = rope_mode
|
80 |
+
if self.rope_mode == 'shared' or self.rope_mode == 'x_only':
|
81 |
+
self.rotary = RotaryEmbedding(dim=head_dim)
|
82 |
+
elif self.rope_mode == 'dual':
|
83 |
+
self.rotary_x = RotaryEmbedding(dim=head_dim)
|
84 |
+
self.rotary_c = RotaryEmbedding(dim=head_dim)
|
85 |
+
|
86 |
+
def _rotary(self, q, k, extras):
|
87 |
+
if self.rope_mode == 'shared':
|
88 |
+
q, k = self.rotary(q=q, k=k)
|
89 |
+
elif self.rope_mode == 'x_only':
|
90 |
+
q_x, k_x = self.rotary(q=q[:, :, extras:, :], k=k[:, :, extras:, :])
|
91 |
+
q_c, k_c = q[:, :, :extras, :], k[:, :, :extras, :]
|
92 |
+
q = torch.cat((q_c, q_x), dim=2)
|
93 |
+
k = torch.cat((k_c, k_x), dim=2)
|
94 |
+
elif self.rope_mode == 'dual':
|
95 |
+
q_x, k_x = self.rotary_x(q=q[:, :, extras:, :], k=k[:, :, extras:, :])
|
96 |
+
q_c, k_c = self.rotary_c(q=q[:, :, :extras, :], k=k[:, :, :extras, :])
|
97 |
+
q = torch.cat((q_c, q_x), dim=2)
|
98 |
+
k = torch.cat((k_c, k_x), dim=2)
|
99 |
+
elif self.rope_mode == 'none':
|
100 |
+
pass
|
101 |
+
else:
|
102 |
+
raise NotImplementedError
|
103 |
+
return q, k
|
104 |
+
|
105 |
+
def _attn(self, q, k, v, mask_binary):
|
106 |
+
if ATTENTION_MODE == 'flash':
|
107 |
+
x = F.scaled_dot_product_attention(q, k, v,
|
108 |
+
dropout_p=self.attn_drop_p,
|
109 |
+
attn_mask=mask_binary)
|
110 |
+
x = einops.rearrange(x, 'B H L D -> B L (H D)')
|
111 |
+
elif ATTENTION_MODE == 'math':
|
112 |
+
attn = (q @ k.transpose(-2, -1)) * self.scale
|
113 |
+
attn = add_mask(attn, mask_binary) if mask_binary is not None else attn
|
114 |
+
attn = attn.softmax(dim=-1)
|
115 |
+
attn = self.attn_drop(attn)
|
116 |
+
x = (attn @ v).transpose(1, 2)
|
117 |
+
x = einops.rearrange(x, 'B H L D -> B L (H D)')
|
118 |
+
else:
|
119 |
+
raise NotImplementedError
|
120 |
+
return x
|
121 |
+
|
122 |
+
def forward(self, x, context=None, context_mask=None, extras=0):
|
123 |
+
B, L, C = x.shape
|
124 |
+
if context is None:
|
125 |
+
context = x
|
126 |
+
|
127 |
+
q = self.to_q(x)
|
128 |
+
k = self.to_k(context)
|
129 |
+
v = self.to_v(context)
|
130 |
+
|
131 |
+
if context_mask is not None:
|
132 |
+
mask_binary = create_mask(x.shape, context.shape,
|
133 |
+
x.device, None, context_mask)
|
134 |
+
else:
|
135 |
+
mask_binary = None
|
136 |
+
|
137 |
+
q = einops.rearrange(q, 'B L (H D) -> B H L D', H=self.num_heads)
|
138 |
+
k = einops.rearrange(k, 'B L (H D) -> B H L D', H=self.num_heads)
|
139 |
+
v = einops.rearrange(v, 'B L (H D) -> B H L D', H=self.num_heads)
|
140 |
+
|
141 |
+
q = self.norm_q(q)
|
142 |
+
k = self.norm_k(k)
|
143 |
+
|
144 |
+
q, k = self._rotary(q, k, extras)
|
145 |
+
|
146 |
+
x = self._attn(q, k, v, mask_binary)
|
147 |
+
|
148 |
+
x = self.proj(x)
|
149 |
+
x = self.proj_drop(x)
|
150 |
+
return x
|
151 |
+
|
152 |
+
|
153 |
+
class JointAttention(nn.Module):
|
154 |
+
def __init__(self, dim, num_heads=8,
|
155 |
+
qkv_bias=False, qk_scale=None, qk_norm=None,
|
156 |
+
attn_drop=0., proj_drop=0.,
|
157 |
+
rope_mode='none'):
|
158 |
+
super().__init__()
|
159 |
+
self.num_heads = num_heads
|
160 |
+
head_dim = dim // num_heads
|
161 |
+
self.scale = qk_scale or head_dim ** -0.5
|
162 |
+
|
163 |
+
self.to_qx, self.to_kx, self.to_vx = self._make_qkv_layers(dim, qkv_bias)
|
164 |
+
self.to_qc, self.to_kc, self.to_vc = self._make_qkv_layers(dim, qkv_bias)
|
165 |
+
|
166 |
+
self.norm_qx, self.norm_kx = self._make_norm_layers(qk_norm, head_dim)
|
167 |
+
self.norm_qc, self.norm_kc = self._make_norm_layers(qk_norm, head_dim)
|
168 |
+
|
169 |
+
self.attn_drop_p = attn_drop
|
170 |
+
self.attn_drop = nn.Dropout(attn_drop)
|
171 |
+
|
172 |
+
self.proj_x = nn.Linear(dim, dim)
|
173 |
+
self.proj_drop_x = nn.Dropout(proj_drop)
|
174 |
+
|
175 |
+
self.proj_c = nn.Linear(dim, dim)
|
176 |
+
self.proj_drop_c = nn.Dropout(proj_drop)
|
177 |
+
|
178 |
+
self.rope_mode = rope_mode
|
179 |
+
if self.rope_mode == 'shared' or self.rope_mode == 'x_only':
|
180 |
+
self.rotary = RotaryEmbedding(dim=head_dim)
|
181 |
+
elif self.rope_mode == 'dual':
|
182 |
+
self.rotary_x = RotaryEmbedding(dim=head_dim)
|
183 |
+
self.rotary_c = RotaryEmbedding(dim=head_dim)
|
184 |
+
|
185 |
+
def _make_qkv_layers(self, dim, qkv_bias):
|
186 |
+
return (nn.Linear(dim, dim, bias=qkv_bias),
|
187 |
+
nn.Linear(dim, dim, bias=qkv_bias),
|
188 |
+
nn.Linear(dim, dim, bias=qkv_bias))
|
189 |
+
|
190 |
+
def _make_norm_layers(self, qk_norm, head_dim):
|
191 |
+
if qk_norm is None:
|
192 |
+
norm_q = nn.Identity()
|
193 |
+
norm_k = nn.Identity()
|
194 |
+
elif qk_norm == 'layernorm':
|
195 |
+
norm_q = nn.LayerNorm(head_dim)
|
196 |
+
norm_k = nn.LayerNorm(head_dim)
|
197 |
+
elif qk_norm == 'rmsnorm':
|
198 |
+
norm_q = RMSNorm(head_dim)
|
199 |
+
norm_k = RMSNorm(head_dim)
|
200 |
+
else:
|
201 |
+
raise NotImplementedError
|
202 |
+
return norm_q, norm_k
|
203 |
+
|
204 |
+
def _rotary(self, q, k, extras):
|
205 |
+
if self.rope_mode == 'shared':
|
206 |
+
q, k = self.rotary(q=q, k=k)
|
207 |
+
elif self.rope_mode == 'x_only':
|
208 |
+
q_x, k_x = self.rotary(q=q[:, :, extras:, :], k=k[:, :, extras:, :])
|
209 |
+
q_c, k_c = q[:, :, :extras, :], k[:, :, :extras, :]
|
210 |
+
q = torch.cat((q_c, q_x), dim=2)
|
211 |
+
k = torch.cat((k_c, k_x), dim=2)
|
212 |
+
elif self.rope_mode == 'dual':
|
213 |
+
q_x, k_x = self.rotary_x(q=q[:, :, extras:, :], k=k[:, :, extras:, :])
|
214 |
+
q_c, k_c = self.rotary_c(q=q[:, :, :extras, :], k=k[:, :, :extras, :])
|
215 |
+
q = torch.cat((q_c, q_x), dim=2)
|
216 |
+
k = torch.cat((k_c, k_x), dim=2)
|
217 |
+
elif self.rope_mode == 'none':
|
218 |
+
pass
|
219 |
+
else:
|
220 |
+
raise NotImplementedError
|
221 |
+
return q, k
|
222 |
+
|
223 |
+
def _attn(self, q, k, v, mask_binary):
|
224 |
+
if ATTENTION_MODE == 'flash':
|
225 |
+
x = F.scaled_dot_product_attention(q, k, v,
|
226 |
+
dropout_p=self.attn_drop_p,
|
227 |
+
attn_mask=mask_binary)
|
228 |
+
x = einops.rearrange(x, 'B H L D -> B L (H D)')
|
229 |
+
elif ATTENTION_MODE == 'math':
|
230 |
+
attn = (q @ k.transpose(-2, -1)) * self.scale
|
231 |
+
attn = add_mask(attn, mask_binary) if mask_binary is not None else attn
|
232 |
+
attn = attn.softmax(dim=-1)
|
233 |
+
attn = self.attn_drop(attn)
|
234 |
+
x = (attn @ v).transpose(1, 2)
|
235 |
+
x = einops.rearrange(x, 'B H L D -> B L (H D)')
|
236 |
+
else:
|
237 |
+
raise NotImplementedError
|
238 |
+
return x
|
239 |
+
|
240 |
+
def _cat_mask(self, x, context, x_mask=None, context_mask=None):
|
241 |
+
B = x.shape[0]
|
242 |
+
if x_mask is None:
|
243 |
+
x_mask = torch.ones(B, x.shape[-2], device=x.device).bool()
|
244 |
+
if context_mask is None:
|
245 |
+
context_mask = torch.ones(B, context.shape[-2], device=context.device).bool()
|
246 |
+
mask = torch.cat([context_mask, x_mask], dim=1)
|
247 |
+
return mask
|
248 |
+
|
249 |
+
def forward(self, x, context, x_mask=None, context_mask=None, extras=0):
|
250 |
+
B, Lx, C = x.shape
|
251 |
+
_, Lc, _ = context.shape
|
252 |
+
if x_mask is not None or context_mask is not None:
|
253 |
+
mask = self._cat_mask(x, context,
|
254 |
+
x_mask=x_mask,
|
255 |
+
context_mask=context_mask)
|
256 |
+
shape = [B, Lx+Lc, C]
|
257 |
+
mask_binary = create_mask(q_shape=shape, k_shape=shape,
|
258 |
+
device=x.device,
|
259 |
+
q_mask=None, k_mask=mask)
|
260 |
+
else:
|
261 |
+
mask_binary = None
|
262 |
+
|
263 |
+
qx, kx, vx = self.to_qx(x), self.to_kx(x), self.to_vx(x)
|
264 |
+
qc, kc, vc = self.to_qc(context), self.to_kc(context), self.to_vc(context)
|
265 |
+
|
266 |
+
qx, kx, vx = map(lambda t: einops.rearrange(t, 'B L (H D) -> B H L D',
|
267 |
+
H=self.num_heads), [qx, kx, vx])
|
268 |
+
qc, kc, vc = map(lambda t: einops.rearrange(t, 'B L (H D) -> B H L D',
|
269 |
+
H=self.num_heads), [qc, kc, vc])
|
270 |
+
|
271 |
+
qx, kx = self.norm_qx(qx), self.norm_kx(kx)
|
272 |
+
qc, kc = self.norm_qc(qc), self.norm_kc(kc)
|
273 |
+
|
274 |
+
q, k, v = (torch.cat([qc, qx], dim=2),
|
275 |
+
torch.cat([kc, kx], dim=2),
|
276 |
+
torch.cat([vc, vx], dim=2))
|
277 |
+
|
278 |
+
q, k = self._rotary(q, k, extras)
|
279 |
+
|
280 |
+
x = self._attn(q, k, v, mask_binary)
|
281 |
+
|
282 |
+
context, x = x[:, :Lc, :], x[:, Lc:, :]
|
283 |
+
|
284 |
+
x = self.proj_x(x)
|
285 |
+
x = self.proj_drop_x(x)
|
286 |
+
|
287 |
+
context = self.proj_c(context)
|
288 |
+
context = self.proj_drop_c(context)
|
289 |
+
|
290 |
+
return x, context
|
src/models/utils/.ipynb_checkpoints/modules-checkpoint.py
ADDED
@@ -0,0 +1,374 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import torch.utils.checkpoint
|
5 |
+
from torch.cuda.amp import autocast
|
6 |
+
import math
|
7 |
+
import einops
|
8 |
+
from einops import rearrange, repeat
|
9 |
+
from inspect import isfunction
|
10 |
+
from .timm import trunc_normal_
|
11 |
+
|
12 |
+
|
13 |
+
# disable in checkpoint mode
|
14 |
+
# @torch.jit.script
|
15 |
+
def film_modulate(x, shift, scale):
|
16 |
+
return x * (1 + scale) + shift
|
17 |
+
|
18 |
+
|
19 |
+
def timestep_embedding(timesteps, dim, max_period=10000):
|
20 |
+
"""
|
21 |
+
Create sinusoidal timestep embeddings.
|
22 |
+
|
23 |
+
:param timesteps: a 1-D Tensor of N indices, one per batch element.
|
24 |
+
These may be fractional.
|
25 |
+
:param dim: the dimension of the output.
|
26 |
+
:param max_period: controls the minimum frequency of the embeddings.
|
27 |
+
:return: an [N x dim] Tensor of positional embeddings.
|
28 |
+
"""
|
29 |
+
half = dim // 2
|
30 |
+
freqs = torch.exp(
|
31 |
+
-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
|
32 |
+
).to(device=timesteps.device)
|
33 |
+
args = timesteps[:, None].float() * freqs[None]
|
34 |
+
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
35 |
+
if dim % 2:
|
36 |
+
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
|
37 |
+
return embedding
|
38 |
+
|
39 |
+
|
40 |
+
class TimestepEmbedder(nn.Module):
|
41 |
+
"""
|
42 |
+
Embeds scalar timesteps into vector representations.
|
43 |
+
"""
|
44 |
+
|
45 |
+
def __init__(self, hidden_size, frequency_embedding_size=256,
|
46 |
+
out_size=None):
|
47 |
+
super().__init__()
|
48 |
+
if out_size is None:
|
49 |
+
out_size = hidden_size
|
50 |
+
self.mlp = nn.Sequential(
|
51 |
+
nn.Linear(frequency_embedding_size, hidden_size, bias=True),
|
52 |
+
nn.SiLU(),
|
53 |
+
nn.Linear(hidden_size, out_size, bias=True),
|
54 |
+
)
|
55 |
+
self.frequency_embedding_size = frequency_embedding_size
|
56 |
+
|
57 |
+
def forward(self, t):
|
58 |
+
t_freq = timestep_embedding(t, self.frequency_embedding_size).type(
|
59 |
+
self.mlp[0].weight.dtype)
|
60 |
+
t_emb = self.mlp(t_freq)
|
61 |
+
return t_emb
|
62 |
+
|
63 |
+
|
64 |
+
def patchify(imgs, patch_size, input_type='2d'):
|
65 |
+
if input_type == '2d':
|
66 |
+
x = einops.rearrange(imgs, 'B C (h p1) (w p2) -> B (h w) (p1 p2 C)', p1=patch_size, p2=patch_size)
|
67 |
+
elif input_type == '1d':
|
68 |
+
x = einops.rearrange(imgs, 'B C (h p1) -> B h (p1 C)', p1=patch_size)
|
69 |
+
return x
|
70 |
+
|
71 |
+
|
72 |
+
def unpatchify(x, channels=3, input_type='2d', img_size=None):
|
73 |
+
if input_type == '2d':
|
74 |
+
patch_size = int((x.shape[2] // channels) ** 0.5)
|
75 |
+
# h = w = int(x.shape[1] ** .5)
|
76 |
+
h, w = img_size[0] // patch_size, img_size[1] // patch_size
|
77 |
+
assert h * w == x.shape[1] and patch_size ** 2 * channels == x.shape[2]
|
78 |
+
x = einops.rearrange(x, 'B (h w) (p1 p2 C) -> B C (h p1) (w p2)', h=h,
|
79 |
+
p1=patch_size, p2=patch_size)
|
80 |
+
elif input_type == '1d':
|
81 |
+
patch_size = int((x.shape[2] // channels))
|
82 |
+
h = x.shape[1]
|
83 |
+
assert patch_size * channels == x.shape[2]
|
84 |
+
x = einops.rearrange(x, 'B h (p1 C) -> B C (h p1)', h=h, p1=patch_size)
|
85 |
+
return x
|
86 |
+
|
87 |
+
|
88 |
+
class PatchEmbed(nn.Module):
|
89 |
+
"""
|
90 |
+
Image to Patch Embedding
|
91 |
+
"""
|
92 |
+
|
93 |
+
def __init__(self, patch_size, in_chans=3, embed_dim=768, input_type='2d'):
|
94 |
+
super().__init__()
|
95 |
+
self.patch_size = patch_size
|
96 |
+
self.input_type = input_type
|
97 |
+
if input_type == '2d':
|
98 |
+
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=True)
|
99 |
+
elif input_type == '1d':
|
100 |
+
self.proj = nn.Conv1d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=True)
|
101 |
+
|
102 |
+
def forward(self, x):
|
103 |
+
if self.input_type == '2d':
|
104 |
+
B, C, H, W = x.shape
|
105 |
+
assert H % self.patch_size == 0 and W % self.patch_size == 0
|
106 |
+
elif self.input_type == '1d':
|
107 |
+
B, C, H = x.shape
|
108 |
+
assert H % self.patch_size == 0
|
109 |
+
|
110 |
+
x = self.proj(x).flatten(2).transpose(1, 2)
|
111 |
+
return x
|
112 |
+
|
113 |
+
|
114 |
+
class PositionalConvEmbedding(nn.Module):
|
115 |
+
"""
|
116 |
+
Relative positional embedding used in HuBERT
|
117 |
+
"""
|
118 |
+
|
119 |
+
def __init__(self, dim=768, kernel_size=128, groups=16):
|
120 |
+
super().__init__()
|
121 |
+
self.conv = nn.Conv1d(
|
122 |
+
dim,
|
123 |
+
dim,
|
124 |
+
kernel_size=kernel_size,
|
125 |
+
padding=kernel_size // 2,
|
126 |
+
groups=groups,
|
127 |
+
bias=True
|
128 |
+
)
|
129 |
+
self.conv = nn.utils.parametrizations.weight_norm(self.conv, name="weight", dim=2)
|
130 |
+
|
131 |
+
def forward(self, x):
|
132 |
+
# B C T
|
133 |
+
x = self.conv(x)
|
134 |
+
x = F.gelu(x[:, :, :-1])
|
135 |
+
return x
|
136 |
+
|
137 |
+
|
138 |
+
class SinusoidalPositionalEncoding(nn.Module):
|
139 |
+
def __init__(self, dim, length):
|
140 |
+
super(SinusoidalPositionalEncoding, self).__init__()
|
141 |
+
self.length = length
|
142 |
+
self.dim = dim
|
143 |
+
self.register_buffer('pe', self._generate_positional_encoding(length, dim))
|
144 |
+
|
145 |
+
def _generate_positional_encoding(self, length, dim):
|
146 |
+
pe = torch.zeros(length, dim)
|
147 |
+
position = torch.arange(0, length, dtype=torch.float).unsqueeze(1)
|
148 |
+
div_term = torch.exp(torch.arange(0, dim, 2).float() * (-math.log(10000.0) / dim))
|
149 |
+
|
150 |
+
pe[:, 0::2] = torch.sin(position * div_term)
|
151 |
+
pe[:, 1::2] = torch.cos(position * div_term)
|
152 |
+
|
153 |
+
pe = pe.unsqueeze(0)
|
154 |
+
return pe
|
155 |
+
|
156 |
+
def forward(self, x):
|
157 |
+
x = x + self.pe[:, :x.size(1)]
|
158 |
+
return x
|
159 |
+
|
160 |
+
|
161 |
+
class PE_wrapper(nn.Module):
|
162 |
+
def __init__(self, dim=768, method='abs', length=None, **kwargs):
|
163 |
+
super().__init__()
|
164 |
+
self.method = method
|
165 |
+
if method == 'abs':
|
166 |
+
# init absolute pe like UViT
|
167 |
+
self.length = length
|
168 |
+
self.abs_pe = nn.Parameter(torch.zeros(1, length, dim))
|
169 |
+
trunc_normal_(self.abs_pe, std=.02)
|
170 |
+
elif method == 'conv':
|
171 |
+
self.conv_pe = PositionalConvEmbedding(dim=dim, **kwargs)
|
172 |
+
elif method == 'sinu':
|
173 |
+
self.sinu_pe = SinusoidalPositionalEncoding(dim=dim, length=length)
|
174 |
+
elif method == 'none':
|
175 |
+
# skip pe
|
176 |
+
self.id = nn.Identity()
|
177 |
+
else:
|
178 |
+
raise NotImplementedError
|
179 |
+
|
180 |
+
def forward(self, x):
|
181 |
+
if self.method == 'abs':
|
182 |
+
_, L, _ = x.shape
|
183 |
+
assert L <= self.length
|
184 |
+
x = x + self.abs_pe[:, :L, :]
|
185 |
+
elif self.method == 'conv':
|
186 |
+
x = x + self.conv_pe(x)
|
187 |
+
elif self.method == 'sinu':
|
188 |
+
x = self.sinu_pe(x)
|
189 |
+
elif self.method == 'none':
|
190 |
+
x = self.id(x)
|
191 |
+
else:
|
192 |
+
raise NotImplementedError
|
193 |
+
return x
|
194 |
+
|
195 |
+
|
196 |
+
class RMSNorm(torch.nn.Module):
|
197 |
+
def __init__(self, dim: int, eps: float = 1e-6):
|
198 |
+
"""
|
199 |
+
Initialize the RMSNorm normalization layer.
|
200 |
+
|
201 |
+
Args:
|
202 |
+
dim (int): The dimension of the input tensor.
|
203 |
+
eps (float, optional): A small value added to the denominator for numerical stability. Default is 1e-6.
|
204 |
+
|
205 |
+
Attributes:
|
206 |
+
eps (float): A small value added to the denominator for numerical stability.
|
207 |
+
weight (nn.Parameter): Learnable scaling parameter.
|
208 |
+
|
209 |
+
"""
|
210 |
+
super().__init__()
|
211 |
+
self.eps = eps
|
212 |
+
self.weight = nn.Parameter(torch.ones(dim))
|
213 |
+
|
214 |
+
def _norm(self, x):
|
215 |
+
"""
|
216 |
+
Apply the RMSNorm normalization to the input tensor.
|
217 |
+
|
218 |
+
Args:
|
219 |
+
x (torch.Tensor): The input tensor.
|
220 |
+
|
221 |
+
Returns:
|
222 |
+
torch.Tensor: The normalized tensor.
|
223 |
+
|
224 |
+
"""
|
225 |
+
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
|
226 |
+
|
227 |
+
def forward(self, x):
|
228 |
+
"""
|
229 |
+
Forward pass through the RMSNorm layer.
|
230 |
+
|
231 |
+
Args:
|
232 |
+
x (torch.Tensor): The input tensor.
|
233 |
+
|
234 |
+
Returns:
|
235 |
+
torch.Tensor: The output tensor after applying RMSNorm.
|
236 |
+
|
237 |
+
"""
|
238 |
+
output = self._norm(x.float()).type_as(x)
|
239 |
+
return output * self.weight
|
240 |
+
|
241 |
+
|
242 |
+
class GELU(nn.Module):
|
243 |
+
|
244 |
+
def __init__(self, dim_in: int, dim_out: int, approximate: str = "none",
|
245 |
+
bias: bool = True):
|
246 |
+
super().__init__()
|
247 |
+
self.proj = nn.Linear(dim_in, dim_out, bias=bias)
|
248 |
+
self.approximate = approximate
|
249 |
+
|
250 |
+
def gelu(self, gate: torch.Tensor) -> torch.Tensor:
|
251 |
+
if gate.device.type != "mps":
|
252 |
+
return F.gelu(gate, approximate=self.approximate)
|
253 |
+
# mps: gelu is not implemented for float16
|
254 |
+
return F.gelu(gate.to(dtype=torch.float32),
|
255 |
+
approximate=self.approximate).to(dtype=gate.dtype)
|
256 |
+
|
257 |
+
def forward(self, hidden_states):
|
258 |
+
hidden_states = self.proj(hidden_states)
|
259 |
+
hidden_states = self.gelu(hidden_states)
|
260 |
+
return hidden_states
|
261 |
+
|
262 |
+
|
263 |
+
class GEGLU(nn.Module):
|
264 |
+
def __init__(self, dim_in: int, dim_out: int, bias: bool = True):
|
265 |
+
super().__init__()
|
266 |
+
self.proj = nn.Linear(dim_in, dim_out * 2, bias=bias)
|
267 |
+
|
268 |
+
def gelu(self, gate: torch.Tensor) -> torch.Tensor:
|
269 |
+
if gate.device.type != "mps":
|
270 |
+
return F.gelu(gate)
|
271 |
+
# mps: gelu is not implemented for float16
|
272 |
+
return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)
|
273 |
+
|
274 |
+
def forward(self, hidden_states):
|
275 |
+
hidden_states = self.proj(hidden_states)
|
276 |
+
hidden_states, gate = hidden_states.chunk(2, dim=-1)
|
277 |
+
return hidden_states * self.gelu(gate)
|
278 |
+
|
279 |
+
|
280 |
+
class ApproximateGELU(nn.Module):
|
281 |
+
def __init__(self, dim_in: int, dim_out: int, bias: bool = True):
|
282 |
+
super().__init__()
|
283 |
+
self.proj = nn.Linear(dim_in, dim_out, bias=bias)
|
284 |
+
|
285 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
286 |
+
x = self.proj(x)
|
287 |
+
return x * torch.sigmoid(1.702 * x)
|
288 |
+
|
289 |
+
|
290 |
+
# disable in checkpoint mode
|
291 |
+
# @torch.jit.script
|
292 |
+
def snake_beta(x, alpha, beta):
|
293 |
+
return x + beta * torch.sin(x * alpha).pow(2)
|
294 |
+
|
295 |
+
|
296 |
+
class Snake(nn.Module):
|
297 |
+
def __init__(self, dim_in, dim_out, bias,
|
298 |
+
alpha_trainable=True):
|
299 |
+
super().__init__()
|
300 |
+
self.proj = nn.Linear(dim_in, dim_out, bias=bias)
|
301 |
+
self.alpha = nn.Parameter(torch.ones(1, 1, dim_out))
|
302 |
+
self.beta = nn.Parameter(torch.ones(1, 1, dim_out))
|
303 |
+
self.alpha.requires_grad = alpha_trainable
|
304 |
+
self.beta.requires_grad = alpha_trainable
|
305 |
+
|
306 |
+
def forward(self, x):
|
307 |
+
x = self.proj(x)
|
308 |
+
x = snake_beta(x, self.alpha, self.beta)
|
309 |
+
return x
|
310 |
+
|
311 |
+
|
312 |
+
class GESnake(nn.Module):
|
313 |
+
def __init__(self, dim_in, dim_out, bias,
|
314 |
+
alpha_trainable=True):
|
315 |
+
super().__init__()
|
316 |
+
self.proj = nn.Linear(dim_in, dim_out * 2, bias=bias)
|
317 |
+
self.alpha = nn.Parameter(torch.ones(1, 1, dim_out))
|
318 |
+
self.beta = nn.Parameter(torch.ones(1, 1, dim_out))
|
319 |
+
self.alpha.requires_grad = alpha_trainable
|
320 |
+
self.beta.requires_grad = alpha_trainable
|
321 |
+
|
322 |
+
def forward(self, x):
|
323 |
+
x = self.proj(x)
|
324 |
+
x, gate = x.chunk(2, dim=-1)
|
325 |
+
return x * snake_beta(gate, self.alpha, self.beta)
|
326 |
+
|
327 |
+
|
328 |
+
class FeedForward(nn.Module):
|
329 |
+
def __init__(
|
330 |
+
self,
|
331 |
+
dim,
|
332 |
+
dim_out=None,
|
333 |
+
mult=4,
|
334 |
+
dropout=0.0,
|
335 |
+
activation_fn="geglu",
|
336 |
+
final_dropout=False,
|
337 |
+
inner_dim=None,
|
338 |
+
bias=True,
|
339 |
+
):
|
340 |
+
super().__init__()
|
341 |
+
if inner_dim is None:
|
342 |
+
inner_dim = int(dim * mult)
|
343 |
+
dim_out = dim_out if dim_out is not None else dim
|
344 |
+
|
345 |
+
if activation_fn == "gelu":
|
346 |
+
act_fn = GELU(dim, inner_dim, bias=bias)
|
347 |
+
elif activation_fn == "gelu-approximate":
|
348 |
+
act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias)
|
349 |
+
elif activation_fn == "geglu":
|
350 |
+
act_fn = GEGLU(dim, inner_dim, bias=bias)
|
351 |
+
elif activation_fn == "geglu-approximate":
|
352 |
+
act_fn = ApproximateGELU(dim, inner_dim, bias=bias)
|
353 |
+
elif activation_fn == "snake":
|
354 |
+
act_fn = Snake(dim, inner_dim, bias=bias)
|
355 |
+
elif activation_fn == "gesnake":
|
356 |
+
act_fn = GESnake(dim, inner_dim, bias=bias)
|
357 |
+
else:
|
358 |
+
raise NotImplementedError
|
359 |
+
|
360 |
+
self.net = nn.ModuleList([])
|
361 |
+
# project in
|
362 |
+
self.net.append(act_fn)
|
363 |
+
# project dropout
|
364 |
+
self.net.append(nn.Dropout(dropout))
|
365 |
+
# project out
|
366 |
+
self.net.append(nn.Linear(inner_dim, dim_out, bias=bias))
|
367 |
+
# FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
|
368 |
+
if final_dropout:
|
369 |
+
self.net.append(nn.Dropout(dropout))
|
370 |
+
|
371 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
372 |
+
for module in self.net:
|
373 |
+
hidden_states = module(hidden_states)
|
374 |
+
return hidden_states
|
src/models/utils/.ipynb_checkpoints/rotary-checkpoint.py
ADDED
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
"this rope is faster than llama rope with jit script"
|
4 |
+
|
5 |
+
|
6 |
+
def rotate_half(x):
|
7 |
+
x1, x2 = x.chunk(2, dim=-1)
|
8 |
+
return torch.cat((-x2, x1), dim=-1)
|
9 |
+
|
10 |
+
|
11 |
+
# disable in checkpoint mode
|
12 |
+
# @torch.jit.script
|
13 |
+
def apply_rotary_pos_emb(x, cos, sin):
|
14 |
+
# NOTE: This could probably be moved to Triton
|
15 |
+
# Handle a possible sequence length mismatch in between q and k
|
16 |
+
cos = cos[:, :, : x.shape[-2], :]
|
17 |
+
sin = sin[:, :, : x.shape[-2], :]
|
18 |
+
return (x * cos) + (rotate_half(x) * sin)
|
19 |
+
|
20 |
+
|
21 |
+
class RotaryEmbedding(torch.nn.Module):
|
22 |
+
"""
|
23 |
+
The rotary position embeddings from RoFormer_ (Su et. al).
|
24 |
+
A crucial insight from the method is that the query and keys are
|
25 |
+
transformed by rotation matrices which depend on the relative positions.
|
26 |
+
|
27 |
+
Other implementations are available in the Rotary Transformer repo_ and in
|
28 |
+
GPT-NeoX_, GPT-NeoX was an inspiration
|
29 |
+
|
30 |
+
.. _RoFormer: https://arxiv.org/abs/2104.09864
|
31 |
+
.. _repo: https://github.com/ZhuiyiTechnology/roformer
|
32 |
+
.. _GPT-NeoX: https://github.com/EleutherAI/gpt-neox
|
33 |
+
|
34 |
+
|
35 |
+
.. warning: Please note that this embedding is not registered on purpose, as it is transformative
|
36 |
+
(it does not create the embedding dimension) and will likely be picked up (imported) on a ad-hoc basis
|
37 |
+
"""
|
38 |
+
|
39 |
+
def __init__(self, dim: int):
|
40 |
+
super().__init__()
|
41 |
+
# Generate and save the inverse frequency buffer (non trainable)
|
42 |
+
inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
|
43 |
+
self.register_buffer("inv_freq", inv_freq)
|
44 |
+
self._seq_len_cached = None
|
45 |
+
self._cos_cached = None
|
46 |
+
self._sin_cached = None
|
47 |
+
|
48 |
+
def _update_cos_sin_tables(self, x, seq_dimension=-2):
|
49 |
+
# expect input: B, H, L, D
|
50 |
+
seq_len = x.shape[seq_dimension]
|
51 |
+
|
52 |
+
# Reset the tables if the sequence length has changed,
|
53 |
+
# or if we're on a new device (possibly due to tracing for instance)
|
54 |
+
# also make sure dtype wont change
|
55 |
+
if (
|
56 |
+
seq_len != self._seq_len_cached
|
57 |
+
or self._cos_cached.device != x.device
|
58 |
+
or self._cos_cached.dtype != x.dtype
|
59 |
+
):
|
60 |
+
self._seq_len_cached = seq_len
|
61 |
+
t = torch.arange(
|
62 |
+
x.shape[seq_dimension], device=x.device, dtype=torch.float32
|
63 |
+
)
|
64 |
+
freqs = torch.einsum("i,j->ij", t, self.inv_freq.to(x.dtype))
|
65 |
+
emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
|
66 |
+
|
67 |
+
self._cos_cached = emb.cos()[None, None, :, :].to(x.dtype)
|
68 |
+
self._sin_cached = emb.sin()[None, None, :, :].to(x.dtype)
|
69 |
+
|
70 |
+
return self._cos_cached, self._sin_cached
|
71 |
+
|
72 |
+
def forward(self, q, k):
|
73 |
+
self._cos_cached, self._sin_cached = self._update_cos_sin_tables(
|
74 |
+
q.float(), seq_dimension=-2
|
75 |
+
)
|
76 |
+
if k is not None:
|
77 |
+
return (
|
78 |
+
apply_rotary_pos_emb(q.float(),
|
79 |
+
self._cos_cached,
|
80 |
+
self._sin_cached).type_as(q),
|
81 |
+
apply_rotary_pos_emb(k.float(),
|
82 |
+
self._cos_cached,
|
83 |
+
self._sin_cached).type_as(k),
|
84 |
+
)
|
85 |
+
else:
|
86 |
+
return (
|
87 |
+
apply_rotary_pos_emb(q.float(),
|
88 |
+
self._cos_cached,
|
89 |
+
self._sin_cached).type_as(q),
|
90 |
+
None
|
91 |
+
)
|
src/models/utils/.ipynb_checkpoints/span_mask-checkpoint.py
ADDED
@@ -0,0 +1,146 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import torch
|
3 |
+
from typing import Optional, Tuple
|
4 |
+
|
5 |
+
|
6 |
+
def compute_mask_indices(
|
7 |
+
shape: Tuple[int, int],
|
8 |
+
padding_mask: Optional[torch.Tensor],
|
9 |
+
mask_prob: float,
|
10 |
+
mask_length: int,
|
11 |
+
mask_type: str = "static",
|
12 |
+
mask_other: float = 0.0,
|
13 |
+
min_masks: int = 0,
|
14 |
+
no_overlap: bool = False,
|
15 |
+
min_space: int = 0,
|
16 |
+
) -> np.ndarray:
|
17 |
+
"""
|
18 |
+
Computes random mask spans for a given shape
|
19 |
+
|
20 |
+
Args:
|
21 |
+
shape: the the shape for which to compute masks.
|
22 |
+
should be of size 2 where first element is batch size and 2nd is timesteps
|
23 |
+
padding_mask: optional padding mask of the same size as shape, which will prevent masking padded elements
|
24 |
+
mask_prob: probability for each token to be chosen as start of the span to be masked. this will be multiplied by
|
25 |
+
number of timesteps divided by length of mask span to mask approximately this percentage of all elements.
|
26 |
+
however due to overlaps, the actual number will be smaller (unless no_overlap is True)
|
27 |
+
mask_type: how to compute mask lengths
|
28 |
+
static = fixed size
|
29 |
+
uniform = sample from uniform distribution [mask_other, mask_length*2]
|
30 |
+
normal = sample from normal distribution with mean mask_length and stdev mask_other. mask is min 1 element
|
31 |
+
poisson = sample from possion distribution with lambda = mask length
|
32 |
+
min_masks: minimum number of masked spans
|
33 |
+
no_overlap: if false, will switch to an alternative recursive algorithm that prevents spans from overlapping
|
34 |
+
min_space: only used if no_overlap is True, this is how many elements to keep unmasked between spans
|
35 |
+
"""
|
36 |
+
|
37 |
+
bsz, all_sz = shape
|
38 |
+
mask = np.full((bsz, all_sz), False)
|
39 |
+
|
40 |
+
# Convert mask_prob to a NumPy array
|
41 |
+
mask_prob = np.array(mask_prob)
|
42 |
+
|
43 |
+
# Calculate all_num_mask for each element in the batch
|
44 |
+
all_num_mask = np.floor(mask_prob * all_sz / float(mask_length) + np.random.rand(bsz)).astype(int)
|
45 |
+
|
46 |
+
# Apply the max operation with min_masks for each element
|
47 |
+
all_num_mask = np.maximum(min_masks, all_num_mask)
|
48 |
+
|
49 |
+
mask_idcs = []
|
50 |
+
for i in range(bsz):
|
51 |
+
if padding_mask is not None:
|
52 |
+
sz = all_sz - padding_mask[i].long().sum().item()
|
53 |
+
num_mask = int(
|
54 |
+
# add a random number for probabilistic rounding
|
55 |
+
mask_prob * sz / float(mask_length)
|
56 |
+
+ np.random.rand()
|
57 |
+
)
|
58 |
+
num_mask = max(min_masks, num_mask)
|
59 |
+
else:
|
60 |
+
sz = all_sz
|
61 |
+
num_mask = all_num_mask[i]
|
62 |
+
|
63 |
+
if mask_type == "static":
|
64 |
+
lengths = np.full(num_mask, mask_length)
|
65 |
+
elif mask_type == "uniform":
|
66 |
+
lengths = np.random.randint(mask_other, mask_length * 2 + 1, size=num_mask)
|
67 |
+
elif mask_type == "normal":
|
68 |
+
lengths = np.random.normal(mask_length, mask_other, size=num_mask)
|
69 |
+
lengths = [max(1, int(round(x))) for x in lengths]
|
70 |
+
elif mask_type == "poisson":
|
71 |
+
lengths = np.random.poisson(mask_length, size=num_mask)
|
72 |
+
lengths = [int(round(x)) for x in lengths]
|
73 |
+
else:
|
74 |
+
raise Exception("unknown mask selection " + mask_type)
|
75 |
+
|
76 |
+
if sum(lengths) == 0:
|
77 |
+
lengths[0] = min(mask_length, sz - 1)
|
78 |
+
|
79 |
+
if no_overlap:
|
80 |
+
mask_idc = []
|
81 |
+
|
82 |
+
def arrange(s, e, length, keep_length):
|
83 |
+
span_start = np.random.randint(s, e - length)
|
84 |
+
mask_idc.extend(span_start + i for i in range(length))
|
85 |
+
|
86 |
+
new_parts = []
|
87 |
+
if span_start - s - min_space >= keep_length:
|
88 |
+
new_parts.append((s, span_start - min_space + 1))
|
89 |
+
if e - span_start - keep_length - min_space > keep_length:
|
90 |
+
new_parts.append((span_start + length + min_space, e))
|
91 |
+
return new_parts
|
92 |
+
|
93 |
+
parts = [(0, sz)]
|
94 |
+
min_length = min(lengths)
|
95 |
+
for length in sorted(lengths, reverse=True):
|
96 |
+
lens = np.fromiter(
|
97 |
+
(e - s if e - s >= length + min_space else 0 for s, e in parts),
|
98 |
+
np.int,
|
99 |
+
)
|
100 |
+
l_sum = np.sum(lens)
|
101 |
+
if l_sum == 0:
|
102 |
+
break
|
103 |
+
probs = lens / np.sum(lens)
|
104 |
+
c = np.random.choice(len(parts), p=probs)
|
105 |
+
s, e = parts.pop(c)
|
106 |
+
parts.extend(arrange(s, e, length, min_length))
|
107 |
+
mask_idc = np.asarray(mask_idc)
|
108 |
+
else:
|
109 |
+
min_len = min(lengths)
|
110 |
+
if sz - min_len <= num_mask:
|
111 |
+
min_len = sz - num_mask - 1
|
112 |
+
|
113 |
+
mask_idc = np.random.choice(sz - min_len, num_mask, replace=False)
|
114 |
+
|
115 |
+
mask_idc = np.asarray(
|
116 |
+
[
|
117 |
+
mask_idc[j] + offset
|
118 |
+
for j in range(len(mask_idc))
|
119 |
+
for offset in range(lengths[j])
|
120 |
+
]
|
121 |
+
)
|
122 |
+
|
123 |
+
mask_idcs.append(np.unique(mask_idc[mask_idc < sz]))
|
124 |
+
# min_len = min([len(m) for m in mask_idcs])
|
125 |
+
for i, mask_idc in enumerate(mask_idcs):
|
126 |
+
# if len(mask_idc) > min_len:
|
127 |
+
# mask_idc = np.random.choice(mask_idc, min_len, replace=False)
|
128 |
+
mask[i, mask_idc] = True
|
129 |
+
|
130 |
+
return torch.tensor(mask)
|
131 |
+
|
132 |
+
|
133 |
+
if __name__ == '__main__':
|
134 |
+
mask = compute_mask_indices(
|
135 |
+
shape=[4, 500],
|
136 |
+
padding_mask=None,
|
137 |
+
mask_prob=[0.65, 0.5, 0.65, 0.65],
|
138 |
+
mask_length=10,
|
139 |
+
mask_type="static",
|
140 |
+
mask_other=0.0,
|
141 |
+
min_masks=1,
|
142 |
+
no_overlap=False,
|
143 |
+
min_space=0,
|
144 |
+
)
|
145 |
+
print(mask)
|
146 |
+
print(mask.sum(dim=1))
|
src/models/utils/.ipynb_checkpoints/timm-checkpoint.py
ADDED
@@ -0,0 +1,114 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# code from timm 0.3.2
|
2 |
+
import torch
|
3 |
+
import torch.nn as nn
|
4 |
+
import math
|
5 |
+
import warnings
|
6 |
+
|
7 |
+
|
8 |
+
def _no_grad_trunc_normal_(tensor, mean, std, a, b):
|
9 |
+
# Cut & paste from PyTorch official master until it's in a few official releases - RW
|
10 |
+
# Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
|
11 |
+
def norm_cdf(x):
|
12 |
+
# Computes standard normal cumulative distribution function
|
13 |
+
return (1. + math.erf(x / math.sqrt(2.))) / 2.
|
14 |
+
|
15 |
+
if (mean < a - 2 * std) or (mean > b + 2 * std):
|
16 |
+
warnings.warn("mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
|
17 |
+
"The distribution of values may be incorrect.",
|
18 |
+
stacklevel=2)
|
19 |
+
|
20 |
+
with torch.no_grad():
|
21 |
+
# Values are generated by using a truncated uniform distribution and
|
22 |
+
# then using the inverse CDF for the normal distribution.
|
23 |
+
# Get upper and lower cdf values
|
24 |
+
l = norm_cdf((a - mean) / std)
|
25 |
+
u = norm_cdf((b - mean) / std)
|
26 |
+
|
27 |
+
# Uniformly fill tensor with values from [l, u], then translate to
|
28 |
+
# [2l-1, 2u-1].
|
29 |
+
tensor.uniform_(2 * l - 1, 2 * u - 1)
|
30 |
+
|
31 |
+
# Use inverse cdf transform for normal distribution to get truncated
|
32 |
+
# standard normal
|
33 |
+
tensor.erfinv_()
|
34 |
+
|
35 |
+
# Transform to proper mean, std
|
36 |
+
tensor.mul_(std * math.sqrt(2.))
|
37 |
+
tensor.add_(mean)
|
38 |
+
|
39 |
+
# Clamp to ensure it's in the proper range
|
40 |
+
tensor.clamp_(min=a, max=b)
|
41 |
+
return tensor
|
42 |
+
|
43 |
+
|
44 |
+
def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
|
45 |
+
# type: (Tensor, float, float, float, float) -> Tensor
|
46 |
+
r"""Fills the input Tensor with values drawn from a truncated
|
47 |
+
normal distribution. The values are effectively drawn from the
|
48 |
+
normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
|
49 |
+
with values outside :math:`[a, b]` redrawn until they are within
|
50 |
+
the bounds. The method used for generating the random values works
|
51 |
+
best when :math:`a \leq \text{mean} \leq b`.
|
52 |
+
Args:
|
53 |
+
tensor: an n-dimensional `torch.Tensor`
|
54 |
+
mean: the mean of the normal distribution
|
55 |
+
std: the standard deviation of the normal distribution
|
56 |
+
a: the minimum cutoff value
|
57 |
+
b: the maximum cutoff value
|
58 |
+
Examples:
|
59 |
+
>>> w = torch.empty(3, 5)
|
60 |
+
>>> nn.init.trunc_normal_(w)
|
61 |
+
"""
|
62 |
+
return _no_grad_trunc_normal_(tensor, mean, std, a, b)
|
63 |
+
|
64 |
+
|
65 |
+
def drop_path(x, drop_prob: float = 0., training: bool = False):
|
66 |
+
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
|
67 |
+
|
68 |
+
This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
|
69 |
+
the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
|
70 |
+
See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
|
71 |
+
changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
|
72 |
+
'survival rate' as the argument.
|
73 |
+
|
74 |
+
"""
|
75 |
+
if drop_prob == 0. or not training:
|
76 |
+
return x
|
77 |
+
keep_prob = 1 - drop_prob
|
78 |
+
shape = (x.shape[0],) + (1,) * (x.ndim - 1) # work with diff dim tensors, not just 2D ConvNets
|
79 |
+
random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
|
80 |
+
random_tensor.floor_() # binarize
|
81 |
+
output = x.div(keep_prob) * random_tensor
|
82 |
+
return output
|
83 |
+
|
84 |
+
|
85 |
+
class DropPath(nn.Module):
|
86 |
+
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
|
87 |
+
"""
|
88 |
+
|
89 |
+
def __init__(self, drop_prob=None):
|
90 |
+
super(DropPath, self).__init__()
|
91 |
+
self.drop_prob = drop_prob
|
92 |
+
|
93 |
+
def forward(self, x):
|
94 |
+
return drop_path(x, self.drop_prob, self.training)
|
95 |
+
|
96 |
+
|
97 |
+
class Mlp(nn.Module):
|
98 |
+
def __init__(self, in_features, hidden_features=None, out_features=None,
|
99 |
+
act_layer=nn.GELU, drop=0.):
|
100 |
+
super().__init__()
|
101 |
+
out_features = out_features or in_features
|
102 |
+
hidden_features = hidden_features or in_features
|
103 |
+
self.fc1 = nn.Linear(in_features, hidden_features)
|
104 |
+
self.act = act_layer()
|
105 |
+
self.fc2 = nn.Linear(hidden_features, out_features)
|
106 |
+
self.drop = nn.Dropout(drop)
|
107 |
+
|
108 |
+
def forward(self, x):
|
109 |
+
x = self.fc1(x)
|
110 |
+
x = self.act(x)
|
111 |
+
x = self.drop(x)
|
112 |
+
x = self.fc2(x)
|
113 |
+
x = self.drop(x)
|
114 |
+
return x
|
src/models/utils/__init__.py
ADDED
File without changes
|
src/models/utils/__pycache__/__init__.cpython-310.pyc
ADDED
Binary file (178 Bytes). View file
|
|
src/models/utils/__pycache__/__init__.cpython-311.pyc
ADDED
Binary file (177 Bytes). View file
|
|
src/models/utils/__pycache__/attention.cpython-310.pyc
ADDED
Binary file (7.61 kB). View file
|
|
src/models/utils/__pycache__/attention.cpython-311.pyc
ADDED
Binary file (16.7 kB). View file
|
|
src/models/utils/__pycache__/modules.cpython-310.pyc
ADDED
Binary file (13.2 kB). View file
|
|
src/models/utils/__pycache__/modules.cpython-311.pyc
ADDED
Binary file (24 kB). View file
|
|
src/models/utils/__pycache__/rotary.cpython-310.pyc
ADDED
Binary file (2.81 kB). View file
|
|
src/models/utils/__pycache__/rotary.cpython-311.pyc
ADDED
Binary file (4.99 kB). View file
|
|
src/models/utils/__pycache__/span_mask.cpython-310.pyc
ADDED
Binary file (4.75 kB). View file
|
|
src/models/utils/__pycache__/span_mask.cpython-311.pyc
ADDED
Binary file (8.51 kB). View file
|
|
src/models/utils/__pycache__/timm.cpython-310.pyc
ADDED
Binary file (4.22 kB). View file
|
|
src/models/utils/__pycache__/timm.cpython-311.pyc
ADDED
Binary file (6.46 kB). View file
|
|
src/models/utils/attention.py
ADDED
@@ -0,0 +1,290 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import torch.utils.checkpoint
|
5 |
+
import einops
|
6 |
+
from einops import rearrange, repeat
|
7 |
+
from inspect import isfunction
|
8 |
+
from .rotary import RotaryEmbedding
|
9 |
+
from .modules import RMSNorm
|
10 |
+
|
11 |
+
|
12 |
+
if hasattr(nn.functional, 'scaled_dot_product_attention'):
|
13 |
+
ATTENTION_MODE = 'flash'
|
14 |
+
else:
|
15 |
+
ATTENTION_MODE = 'math'
|
16 |
+
print(f'attention mode is {ATTENTION_MODE}')
|
17 |
+
|
18 |
+
|
19 |
+
def add_mask(sim, mask):
|
20 |
+
b, ndim = sim.shape[0], mask.ndim
|
21 |
+
if ndim == 3:
|
22 |
+
mask = rearrange(mask, "b n m -> b 1 n m")
|
23 |
+
if ndim == 2:
|
24 |
+
mask = repeat(mask, "n m -> b 1 n m", b=b)
|
25 |
+
max_neg_value = -torch.finfo(sim.dtype).max
|
26 |
+
sim = sim.masked_fill(~mask, max_neg_value)
|
27 |
+
return sim
|
28 |
+
|
29 |
+
|
30 |
+
def create_mask(q_shape, k_shape, device, q_mask=None, k_mask=None):
|
31 |
+
def default(val, d):
|
32 |
+
return val if val is not None else (d() if isfunction(d) else d)
|
33 |
+
b, i, j, device = q_shape[0], q_shape[-2], k_shape[-2], device
|
34 |
+
q_mask = default(q_mask, torch.ones((b, i), device=device, dtype=torch.bool))
|
35 |
+
k_mask = default(k_mask, torch.ones((b, j), device=device, dtype=torch.bool))
|
36 |
+
attn_mask = rearrange(q_mask, 'b i -> b 1 i 1') * rearrange(k_mask, 'b j -> b 1 1 j')
|
37 |
+
return attn_mask
|
38 |
+
|
39 |
+
|
40 |
+
class Attention(nn.Module):
|
41 |
+
def __init__(self, dim, context_dim=None, num_heads=8,
|
42 |
+
qkv_bias=False, qk_scale=None, qk_norm=None,
|
43 |
+
attn_drop=0., proj_drop=0., rope_mode='none'):
|
44 |
+
super().__init__()
|
45 |
+
self.num_heads = num_heads
|
46 |
+
head_dim = dim // num_heads
|
47 |
+
self.scale = qk_scale or head_dim ** -0.5
|
48 |
+
|
49 |
+
if context_dim is None:
|
50 |
+
self.cross_attn = False
|
51 |
+
else:
|
52 |
+
self.cross_attn = True
|
53 |
+
|
54 |
+
context_dim = dim if context_dim is None else context_dim
|
55 |
+
|
56 |
+
self.to_q = nn.Linear(dim, dim, bias=qkv_bias)
|
57 |
+
self.to_k = nn.Linear(context_dim, dim, bias=qkv_bias)
|
58 |
+
self.to_v = nn.Linear(context_dim, dim, bias=qkv_bias)
|
59 |
+
|
60 |
+
if qk_norm is None:
|
61 |
+
self.norm_q = nn.Identity()
|
62 |
+
self.norm_k = nn.Identity()
|
63 |
+
elif qk_norm == 'layernorm':
|
64 |
+
self.norm_q = nn.LayerNorm(head_dim)
|
65 |
+
self.norm_k = nn.LayerNorm(head_dim)
|
66 |
+
elif qk_norm == 'rmsnorm':
|
67 |
+
self.norm_q = RMSNorm(head_dim)
|
68 |
+
self.norm_k = RMSNorm(head_dim)
|
69 |
+
else:
|
70 |
+
raise NotImplementedError
|
71 |
+
|
72 |
+
self.attn_drop_p = attn_drop
|
73 |
+
self.attn_drop = nn.Dropout(attn_drop)
|
74 |
+
self.proj = nn.Linear(dim, dim)
|
75 |
+
self.proj_drop = nn.Dropout(proj_drop)
|
76 |
+
|
77 |
+
if self.cross_attn:
|
78 |
+
assert rope_mode == 'none'
|
79 |
+
self.rope_mode = rope_mode
|
80 |
+
if self.rope_mode == 'shared' or self.rope_mode == 'x_only':
|
81 |
+
self.rotary = RotaryEmbedding(dim=head_dim)
|
82 |
+
elif self.rope_mode == 'dual':
|
83 |
+
self.rotary_x = RotaryEmbedding(dim=head_dim)
|
84 |
+
self.rotary_c = RotaryEmbedding(dim=head_dim)
|
85 |
+
|
86 |
+
def _rotary(self, q, k, extras):
|
87 |
+
if self.rope_mode == 'shared':
|
88 |
+
q, k = self.rotary(q=q, k=k)
|
89 |
+
elif self.rope_mode == 'x_only':
|
90 |
+
q_x, k_x = self.rotary(q=q[:, :, extras:, :], k=k[:, :, extras:, :])
|
91 |
+
q_c, k_c = q[:, :, :extras, :], k[:, :, :extras, :]
|
92 |
+
q = torch.cat((q_c, q_x), dim=2)
|
93 |
+
k = torch.cat((k_c, k_x), dim=2)
|
94 |
+
elif self.rope_mode == 'dual':
|
95 |
+
q_x, k_x = self.rotary_x(q=q[:, :, extras:, :], k=k[:, :, extras:, :])
|
96 |
+
q_c, k_c = self.rotary_c(q=q[:, :, :extras, :], k=k[:, :, :extras, :])
|
97 |
+
q = torch.cat((q_c, q_x), dim=2)
|
98 |
+
k = torch.cat((k_c, k_x), dim=2)
|
99 |
+
elif self.rope_mode == 'none':
|
100 |
+
pass
|
101 |
+
else:
|
102 |
+
raise NotImplementedError
|
103 |
+
return q, k
|
104 |
+
|
105 |
+
def _attn(self, q, k, v, mask_binary):
|
106 |
+
if ATTENTION_MODE == 'flash':
|
107 |
+
x = F.scaled_dot_product_attention(q, k, v,
|
108 |
+
dropout_p=self.attn_drop_p,
|
109 |
+
attn_mask=mask_binary)
|
110 |
+
x = einops.rearrange(x, 'B H L D -> B L (H D)')
|
111 |
+
elif ATTENTION_MODE == 'math':
|
112 |
+
attn = (q @ k.transpose(-2, -1)) * self.scale
|
113 |
+
attn = add_mask(attn, mask_binary) if mask_binary is not None else attn
|
114 |
+
attn = attn.softmax(dim=-1)
|
115 |
+
attn = self.attn_drop(attn)
|
116 |
+
x = (attn @ v).transpose(1, 2)
|
117 |
+
x = einops.rearrange(x, 'B H L D -> B L (H D)')
|
118 |
+
else:
|
119 |
+
raise NotImplementedError
|
120 |
+
return x
|
121 |
+
|
122 |
+
def forward(self, x, context=None, context_mask=None, extras=0):
|
123 |
+
B, L, C = x.shape
|
124 |
+
if context is None:
|
125 |
+
context = x
|
126 |
+
|
127 |
+
q = self.to_q(x)
|
128 |
+
k = self.to_k(context)
|
129 |
+
v = self.to_v(context)
|
130 |
+
|
131 |
+
if context_mask is not None:
|
132 |
+
mask_binary = create_mask(x.shape, context.shape,
|
133 |
+
x.device, None, context_mask)
|
134 |
+
else:
|
135 |
+
mask_binary = None
|
136 |
+
|
137 |
+
q = einops.rearrange(q, 'B L (H D) -> B H L D', H=self.num_heads)
|
138 |
+
k = einops.rearrange(k, 'B L (H D) -> B H L D', H=self.num_heads)
|
139 |
+
v = einops.rearrange(v, 'B L (H D) -> B H L D', H=self.num_heads)
|
140 |
+
|
141 |
+
q = self.norm_q(q)
|
142 |
+
k = self.norm_k(k)
|
143 |
+
|
144 |
+
q, k = self._rotary(q, k, extras)
|
145 |
+
|
146 |
+
x = self._attn(q, k, v, mask_binary)
|
147 |
+
|
148 |
+
x = self.proj(x)
|
149 |
+
x = self.proj_drop(x)
|
150 |
+
return x
|
151 |
+
|
152 |
+
|
153 |
+
class JointAttention(nn.Module):
|
154 |
+
def __init__(self, dim, num_heads=8,
|
155 |
+
qkv_bias=False, qk_scale=None, qk_norm=None,
|
156 |
+
attn_drop=0., proj_drop=0.,
|
157 |
+
rope_mode='none'):
|
158 |
+
super().__init__()
|
159 |
+
self.num_heads = num_heads
|
160 |
+
head_dim = dim // num_heads
|
161 |
+
self.scale = qk_scale or head_dim ** -0.5
|
162 |
+
|
163 |
+
self.to_qx, self.to_kx, self.to_vx = self._make_qkv_layers(dim, qkv_bias)
|
164 |
+
self.to_qc, self.to_kc, self.to_vc = self._make_qkv_layers(dim, qkv_bias)
|
165 |
+
|
166 |
+
self.norm_qx, self.norm_kx = self._make_norm_layers(qk_norm, head_dim)
|
167 |
+
self.norm_qc, self.norm_kc = self._make_norm_layers(qk_norm, head_dim)
|
168 |
+
|
169 |
+
self.attn_drop_p = attn_drop
|
170 |
+
self.attn_drop = nn.Dropout(attn_drop)
|
171 |
+
|
172 |
+
self.proj_x = nn.Linear(dim, dim)
|
173 |
+
self.proj_drop_x = nn.Dropout(proj_drop)
|
174 |
+
|
175 |
+
self.proj_c = nn.Linear(dim, dim)
|
176 |
+
self.proj_drop_c = nn.Dropout(proj_drop)
|
177 |
+
|
178 |
+
self.rope_mode = rope_mode
|
179 |
+
if self.rope_mode == 'shared' or self.rope_mode == 'x_only':
|
180 |
+
self.rotary = RotaryEmbedding(dim=head_dim)
|
181 |
+
elif self.rope_mode == 'dual':
|
182 |
+
self.rotary_x = RotaryEmbedding(dim=head_dim)
|
183 |
+
self.rotary_c = RotaryEmbedding(dim=head_dim)
|
184 |
+
|
185 |
+
def _make_qkv_layers(self, dim, qkv_bias):
|
186 |
+
return (nn.Linear(dim, dim, bias=qkv_bias),
|
187 |
+
nn.Linear(dim, dim, bias=qkv_bias),
|
188 |
+
nn.Linear(dim, dim, bias=qkv_bias))
|
189 |
+
|
190 |
+
def _make_norm_layers(self, qk_norm, head_dim):
|
191 |
+
if qk_norm is None:
|
192 |
+
norm_q = nn.Identity()
|
193 |
+
norm_k = nn.Identity()
|
194 |
+
elif qk_norm == 'layernorm':
|
195 |
+
norm_q = nn.LayerNorm(head_dim)
|
196 |
+
norm_k = nn.LayerNorm(head_dim)
|
197 |
+
elif qk_norm == 'rmsnorm':
|
198 |
+
norm_q = RMSNorm(head_dim)
|
199 |
+
norm_k = RMSNorm(head_dim)
|
200 |
+
else:
|
201 |
+
raise NotImplementedError
|
202 |
+
return norm_q, norm_k
|
203 |
+
|
204 |
+
def _rotary(self, q, k, extras):
|
205 |
+
if self.rope_mode == 'shared':
|
206 |
+
q, k = self.rotary(q=q, k=k)
|
207 |
+
elif self.rope_mode == 'x_only':
|
208 |
+
q_x, k_x = self.rotary(q=q[:, :, extras:, :], k=k[:, :, extras:, :])
|
209 |
+
q_c, k_c = q[:, :, :extras, :], k[:, :, :extras, :]
|
210 |
+
q = torch.cat((q_c, q_x), dim=2)
|
211 |
+
k = torch.cat((k_c, k_x), dim=2)
|
212 |
+
elif self.rope_mode == 'dual':
|
213 |
+
q_x, k_x = self.rotary_x(q=q[:, :, extras:, :], k=k[:, :, extras:, :])
|
214 |
+
q_c, k_c = self.rotary_c(q=q[:, :, :extras, :], k=k[:, :, :extras, :])
|
215 |
+
q = torch.cat((q_c, q_x), dim=2)
|
216 |
+
k = torch.cat((k_c, k_x), dim=2)
|
217 |
+
elif self.rope_mode == 'none':
|
218 |
+
pass
|
219 |
+
else:
|
220 |
+
raise NotImplementedError
|
221 |
+
return q, k
|
222 |
+
|
223 |
+
def _attn(self, q, k, v, mask_binary):
|
224 |
+
if ATTENTION_MODE == 'flash':
|
225 |
+
x = F.scaled_dot_product_attention(q, k, v,
|
226 |
+
dropout_p=self.attn_drop_p,
|
227 |
+
attn_mask=mask_binary)
|
228 |
+
x = einops.rearrange(x, 'B H L D -> B L (H D)')
|
229 |
+
elif ATTENTION_MODE == 'math':
|
230 |
+
attn = (q @ k.transpose(-2, -1)) * self.scale
|
231 |
+
attn = add_mask(attn, mask_binary) if mask_binary is not None else attn
|
232 |
+
attn = attn.softmax(dim=-1)
|
233 |
+
attn = self.attn_drop(attn)
|
234 |
+
x = (attn @ v).transpose(1, 2)
|
235 |
+
x = einops.rearrange(x, 'B H L D -> B L (H D)')
|
236 |
+
else:
|
237 |
+
raise NotImplementedError
|
238 |
+
return x
|
239 |
+
|
240 |
+
def _cat_mask(self, x, context, x_mask=None, context_mask=None):
|
241 |
+
B = x.shape[0]
|
242 |
+
if x_mask is None:
|
243 |
+
x_mask = torch.ones(B, x.shape[-2], device=x.device).bool()
|
244 |
+
if context_mask is None:
|
245 |
+
context_mask = torch.ones(B, context.shape[-2], device=context.device).bool()
|
246 |
+
mask = torch.cat([context_mask, x_mask], dim=1)
|
247 |
+
return mask
|
248 |
+
|
249 |
+
def forward(self, x, context, x_mask=None, context_mask=None, extras=0):
|
250 |
+
B, Lx, C = x.shape
|
251 |
+
_, Lc, _ = context.shape
|
252 |
+
if x_mask is not None or context_mask is not None:
|
253 |
+
mask = self._cat_mask(x, context,
|
254 |
+
x_mask=x_mask,
|
255 |
+
context_mask=context_mask)
|
256 |
+
shape = [B, Lx+Lc, C]
|
257 |
+
mask_binary = create_mask(q_shape=shape, k_shape=shape,
|
258 |
+
device=x.device,
|
259 |
+
q_mask=None, k_mask=mask)
|
260 |
+
else:
|
261 |
+
mask_binary = None
|
262 |
+
|
263 |
+
qx, kx, vx = self.to_qx(x), self.to_kx(x), self.to_vx(x)
|
264 |
+
qc, kc, vc = self.to_qc(context), self.to_kc(context), self.to_vc(context)
|
265 |
+
|
266 |
+
qx, kx, vx = map(lambda t: einops.rearrange(t, 'B L (H D) -> B H L D',
|
267 |
+
H=self.num_heads), [qx, kx, vx])
|
268 |
+
qc, kc, vc = map(lambda t: einops.rearrange(t, 'B L (H D) -> B H L D',
|
269 |
+
H=self.num_heads), [qc, kc, vc])
|
270 |
+
|
271 |
+
qx, kx = self.norm_qx(qx), self.norm_kx(kx)
|
272 |
+
qc, kc = self.norm_qc(qc), self.norm_kc(kc)
|
273 |
+
|
274 |
+
q, k, v = (torch.cat([qc, qx], dim=2),
|
275 |
+
torch.cat([kc, kx], dim=2),
|
276 |
+
torch.cat([vc, vx], dim=2))
|
277 |
+
|
278 |
+
q, k = self._rotary(q, k, extras)
|
279 |
+
|
280 |
+
x = self._attn(q, k, v, mask_binary)
|
281 |
+
|
282 |
+
context, x = x[:, :Lc, :], x[:, Lc:, :]
|
283 |
+
|
284 |
+
x = self.proj_x(x)
|
285 |
+
x = self.proj_drop_x(x)
|
286 |
+
|
287 |
+
context = self.proj_c(context)
|
288 |
+
context = self.proj_drop_c(context)
|
289 |
+
|
290 |
+
return x, context
|
src/models/utils/bk/.ipynb_checkpoints/attention-checkpoint.py
ADDED
@@ -0,0 +1,99 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.utils.checkpoint
|
4 |
+
import einops
|
5 |
+
from einops import rearrange, repeat
|
6 |
+
from inspect import isfunction
|
7 |
+
from .rotary import RotaryEmbedding
|
8 |
+
|
9 |
+
if hasattr(nn.functional, 'scaled_dot_product_attention'):
|
10 |
+
ATTENTION_MODE = 'flash'
|
11 |
+
else:
|
12 |
+
ATTENTION_MODE = 'math'
|
13 |
+
print(f'attention mode is {ATTENTION_MODE}')
|
14 |
+
|
15 |
+
|
16 |
+
def add_mask(sim, mask):
|
17 |
+
b, ndim = sim.shape[0], mask.ndim
|
18 |
+
if ndim == 3:
|
19 |
+
mask = rearrange(mask, "b n m -> b 1 n m")
|
20 |
+
if ndim == 2:
|
21 |
+
mask = repeat(mask, "n m -> b 1 n m", b=b)
|
22 |
+
max_neg_value = -torch.finfo(sim.dtype).max
|
23 |
+
sim = sim.masked_fill(~mask, max_neg_value)
|
24 |
+
return sim
|
25 |
+
|
26 |
+
|
27 |
+
def create_mask(q, k, q_mask=None, k_mask=None):
|
28 |
+
def default(val, d):
|
29 |
+
return val if val is not None else (d() if isfunction(d) else d)
|
30 |
+
|
31 |
+
b, i, j, device = q.shape[0], q.shape[-2], k.shape[-2], q.device
|
32 |
+
q_mask = default(q_mask, torch.ones((b, i), device=device, dtype=torch.bool))
|
33 |
+
k_mask = default(k_mask, torch.ones((b, j), device=device, dtype=torch.bool))
|
34 |
+
attn_mask = rearrange(q_mask, 'b i -> b 1 i 1') * rearrange(k_mask, 'b j -> b 1 1 j')
|
35 |
+
return attn_mask
|
36 |
+
|
37 |
+
|
38 |
+
class Attention(nn.Module):
|
39 |
+
def __init__(self, dim, context_dim=None, num_heads=8, qkv_bias=False, qk_scale=None,
|
40 |
+
attn_drop=0., proj_drop=0., use_rope=False):
|
41 |
+
super().__init__()
|
42 |
+
self.num_heads = num_heads
|
43 |
+
head_dim = dim // num_heads
|
44 |
+
self.scale = qk_scale or head_dim ** -0.5
|
45 |
+
|
46 |
+
context_dim = dim if context_dim is None else context_dim
|
47 |
+
|
48 |
+
self.to_q = nn.Linear(dim, dim, bias=qkv_bias)
|
49 |
+
self.to_k = nn.Linear(context_dim, dim, bias=qkv_bias)
|
50 |
+
self.to_v = nn.Linear(context_dim, dim, bias=qkv_bias)
|
51 |
+
self.attn_drop_p = attn_drop
|
52 |
+
self.attn_drop = nn.Dropout(attn_drop)
|
53 |
+
self.proj = nn.Linear(dim, dim)
|
54 |
+
self.proj_drop = nn.Dropout(proj_drop)
|
55 |
+
|
56 |
+
self.use_rope = use_rope
|
57 |
+
if self.use_rope:
|
58 |
+
self.rotary = RotaryEmbedding(dim=head_dim)
|
59 |
+
|
60 |
+
def forward(self, x, context=None, context_mask=None):
|
61 |
+
B, L, C = x.shape
|
62 |
+
q = self.to_q(x)
|
63 |
+
if context is None:
|
64 |
+
context = x
|
65 |
+
else:
|
66 |
+
assert self.use_rope is False
|
67 |
+
|
68 |
+
k = self.to_k(context)
|
69 |
+
v = self.to_v(context)
|
70 |
+
|
71 |
+
if context_mask is not None:
|
72 |
+
mask_binary = create_mask(x, context, None, context_mask)
|
73 |
+
else:
|
74 |
+
mask_binary = None
|
75 |
+
|
76 |
+
q = einops.rearrange(q, 'B L (H D) -> B H L D', H=self.num_heads).float()
|
77 |
+
k = einops.rearrange(k, 'B L (H D) -> B H L D', H=self.num_heads).float()
|
78 |
+
v = einops.rearrange(v, 'B L (H D) -> B H L D', H=self.num_heads).float()
|
79 |
+
|
80 |
+
if self.use_rope:
|
81 |
+
q, k = self.rotary(q=q, k=k)
|
82 |
+
|
83 |
+
if ATTENTION_MODE == 'flash':
|
84 |
+
x = torch.nn.functional.scaled_dot_product_attention(q, k, v,
|
85 |
+
dropout_p=self.attn_drop_p,
|
86 |
+
attn_mask=mask_binary)
|
87 |
+
x = einops.rearrange(x, 'B H L D -> B L (H D)')
|
88 |
+
elif ATTENTION_MODE == 'math':
|
89 |
+
attn = (q @ k.transpose(-2, -1)) * self.scale
|
90 |
+
attn = add_mask(attn, mask_binary) if mask_binary is not None else attn
|
91 |
+
attn = attn.softmax(dim=-1)
|
92 |
+
attn = self.attn_drop(attn)
|
93 |
+
x = (attn @ v).transpose(1, 2).reshape(B, L, C)
|
94 |
+
else:
|
95 |
+
raise NotImplementedError
|
96 |
+
|
97 |
+
x = self.proj(x)
|
98 |
+
x = self.proj_drop(x)
|
99 |
+
return x
|
src/models/utils/bk/.ipynb_checkpoints/llama_rotary-checkpoint.py
ADDED
@@ -0,0 +1,74 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from typing import Tuple
|
3 |
+
from rotary import RotaryEmbedding
|
4 |
+
import time
|
5 |
+
|
6 |
+
|
7 |
+
def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0):
|
8 |
+
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
|
9 |
+
t = torch.arange(end, device=freqs.device, dtype=torch.float32)
|
10 |
+
freqs = torch.outer(t, freqs)
|
11 |
+
freqs_cis = torch.polar(torch.ones_like(freqs), freqs) # complex64
|
12 |
+
return freqs_cis
|
13 |
+
|
14 |
+
|
15 |
+
def reshape_for_broadcast(freqs_cis: torch.Tensor,
|
16 |
+
x: torch.Tensor,):
|
17 |
+
ndim = x.ndim
|
18 |
+
assert 0 <= 1 < ndim
|
19 |
+
assert freqs_cis.shape == (x.shape[1], x.shape[-1])
|
20 |
+
shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
|
21 |
+
return freqs_cis.view(*shape)
|
22 |
+
|
23 |
+
|
24 |
+
def compute_rope(q, freqs_cis):
|
25 |
+
return q * freqs_cis
|
26 |
+
|
27 |
+
|
28 |
+
def apply_rotary_emb(
|
29 |
+
xq: torch.Tensor,
|
30 |
+
xk: torch.Tensor,
|
31 |
+
freqs_cis: torch.Tensor,
|
32 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
33 |
+
# xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
|
34 |
+
# xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
|
35 |
+
xq1, xq2 = xq.chunk(2, dim=-1)
|
36 |
+
xq_ = torch.view_as_complex(torch.stack((xq1, xq2), dim=-1).float())
|
37 |
+
|
38 |
+
xk1, xk2 = xk.chunk(2, dim=-1)
|
39 |
+
xk_ = torch.view_as_complex(torch.stack((xk1, xk2), dim=-1).float())
|
40 |
+
|
41 |
+
freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
|
42 |
+
xq_out = torch.view_as_real(compute_rope(xq_, freqs_cis)).flatten(3)
|
43 |
+
xk_out = torch.view_as_real(compute_rope(xk_, freqs_cis)).flatten(3)
|
44 |
+
return xq_out.type_as(xq), xk_out.type_as(xk)
|
45 |
+
|
46 |
+
|
47 |
+
if __name__ == '__main__':
|
48 |
+
# Move data to CUDA
|
49 |
+
freq_cis = precompute_freqs_cis(4, 5).cuda()
|
50 |
+
x = torch.rand(1, 5, 1, 4).cuda()
|
51 |
+
y = torch.rand(1, 5, 1, 4).cuda()
|
52 |
+
|
53 |
+
# First method
|
54 |
+
start_time = time.time()
|
55 |
+
for _ in range(20000):
|
56 |
+
x1, y1 = apply_rotary_emb(x, y, freq_cis)
|
57 |
+
end_time = time.time()
|
58 |
+
print(f"Method 1 time cost: {end_time - start_time} seconds")
|
59 |
+
|
60 |
+
# Prepare data for the second method
|
61 |
+
x = x.permute(0, 2, 1, 3)
|
62 |
+
y = y.permute(0, 2, 1, 3)
|
63 |
+
rope = RotaryEmbedding(4).cuda()
|
64 |
+
|
65 |
+
# Second method
|
66 |
+
start_time = time.time()
|
67 |
+
for _ in range(20000):
|
68 |
+
x2, y2 = rope(x, y)
|
69 |
+
end_time = time.time()
|
70 |
+
print(f"Method 2 time cost: {end_time - start_time} seconds")
|
71 |
+
|
72 |
+
# Print the results
|
73 |
+
print(x1)
|
74 |
+
print(x2)
|
src/models/utils/bk/__pycache__/rotary.cpython-311.pyc
ADDED
Binary file (4.8 kB). View file
|
|
src/models/utils/bk/attention.py
ADDED
@@ -0,0 +1,99 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.utils.checkpoint
|
4 |
+
import einops
|
5 |
+
from einops import rearrange, repeat
|
6 |
+
from inspect import isfunction
|
7 |
+
from .rotary import RotaryEmbedding
|
8 |
+
|
9 |
+
if hasattr(nn.functional, 'scaled_dot_product_attention'):
|
10 |
+
ATTENTION_MODE = 'flash'
|
11 |
+
else:
|
12 |
+
ATTENTION_MODE = 'math'
|
13 |
+
print(f'attention mode is {ATTENTION_MODE}')
|
14 |
+
|
15 |
+
|
16 |
+
def add_mask(sim, mask):
|
17 |
+
b, ndim = sim.shape[0], mask.ndim
|
18 |
+
if ndim == 3:
|
19 |
+
mask = rearrange(mask, "b n m -> b 1 n m")
|
20 |
+
if ndim == 2:
|
21 |
+
mask = repeat(mask, "n m -> b 1 n m", b=b)
|
22 |
+
max_neg_value = -torch.finfo(sim.dtype).max
|
23 |
+
sim = sim.masked_fill(~mask, max_neg_value)
|
24 |
+
return sim
|
25 |
+
|
26 |
+
|
27 |
+
def create_mask(q, k, q_mask=None, k_mask=None):
|
28 |
+
def default(val, d):
|
29 |
+
return val if val is not None else (d() if isfunction(d) else d)
|
30 |
+
|
31 |
+
b, i, j, device = q.shape[0], q.shape[-2], k.shape[-2], q.device
|
32 |
+
q_mask = default(q_mask, torch.ones((b, i), device=device, dtype=torch.bool))
|
33 |
+
k_mask = default(k_mask, torch.ones((b, j), device=device, dtype=torch.bool))
|
34 |
+
attn_mask = rearrange(q_mask, 'b i -> b 1 i 1') * rearrange(k_mask, 'b j -> b 1 1 j')
|
35 |
+
return attn_mask
|
36 |
+
|
37 |
+
|
38 |
+
class Attention(nn.Module):
|
39 |
+
def __init__(self, dim, context_dim=None, num_heads=8, qkv_bias=False, qk_scale=None,
|
40 |
+
attn_drop=0., proj_drop=0., use_rope=False):
|
41 |
+
super().__init__()
|
42 |
+
self.num_heads = num_heads
|
43 |
+
head_dim = dim // num_heads
|
44 |
+
self.scale = qk_scale or head_dim ** -0.5
|
45 |
+
|
46 |
+
context_dim = dim if context_dim is None else context_dim
|
47 |
+
|
48 |
+
self.to_q = nn.Linear(dim, dim, bias=qkv_bias)
|
49 |
+
self.to_k = nn.Linear(context_dim, dim, bias=qkv_bias)
|
50 |
+
self.to_v = nn.Linear(context_dim, dim, bias=qkv_bias)
|
51 |
+
self.attn_drop_p = attn_drop
|
52 |
+
self.attn_drop = nn.Dropout(attn_drop)
|
53 |
+
self.proj = nn.Linear(dim, dim)
|
54 |
+
self.proj_drop = nn.Dropout(proj_drop)
|
55 |
+
|
56 |
+
self.use_rope = use_rope
|
57 |
+
if self.use_rope:
|
58 |
+
self.rotary = RotaryEmbedding(dim=head_dim)
|
59 |
+
|
60 |
+
def forward(self, x, context=None, context_mask=None):
|
61 |
+
B, L, C = x.shape
|
62 |
+
q = self.to_q(x)
|
63 |
+
if context is None:
|
64 |
+
context = x
|
65 |
+
else:
|
66 |
+
assert self.use_rope is False
|
67 |
+
|
68 |
+
k = self.to_k(context)
|
69 |
+
v = self.to_v(context)
|
70 |
+
|
71 |
+
if context_mask is not None:
|
72 |
+
mask_binary = create_mask(x, context, None, context_mask)
|
73 |
+
else:
|
74 |
+
mask_binary = None
|
75 |
+
|
76 |
+
q = einops.rearrange(q, 'B L (H D) -> B H L D', H=self.num_heads).float()
|
77 |
+
k = einops.rearrange(k, 'B L (H D) -> B H L D', H=self.num_heads).float()
|
78 |
+
v = einops.rearrange(v, 'B L (H D) -> B H L D', H=self.num_heads).float()
|
79 |
+
|
80 |
+
if self.use_rope:
|
81 |
+
q, k = self.rotary(q=q, k=k)
|
82 |
+
|
83 |
+
if ATTENTION_MODE == 'flash':
|
84 |
+
x = torch.nn.functional.scaled_dot_product_attention(q, k, v,
|
85 |
+
dropout_p=self.attn_drop_p,
|
86 |
+
attn_mask=mask_binary)
|
87 |
+
x = einops.rearrange(x, 'B H L D -> B L (H D)')
|
88 |
+
elif ATTENTION_MODE == 'math':
|
89 |
+
attn = (q @ k.transpose(-2, -1)) * self.scale
|
90 |
+
attn = add_mask(attn, mask_binary) if mask_binary is not None else attn
|
91 |
+
attn = attn.softmax(dim=-1)
|
92 |
+
attn = self.attn_drop(attn)
|
93 |
+
x = (attn @ v).transpose(1, 2).reshape(B, L, C)
|
94 |
+
else:
|
95 |
+
raise NotImplementedError
|
96 |
+
|
97 |
+
x = self.proj(x)
|
98 |
+
x = self.proj_drop(x)
|
99 |
+
return x
|
src/models/utils/bk/llama_rotary.py
ADDED
@@ -0,0 +1,74 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from typing import Tuple
|
3 |
+
from rotary import RotaryEmbedding
|
4 |
+
import time
|
5 |
+
|
6 |
+
|
7 |
+
def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0):
|
8 |
+
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
|
9 |
+
t = torch.arange(end, device=freqs.device, dtype=torch.float32)
|
10 |
+
freqs = torch.outer(t, freqs)
|
11 |
+
freqs_cis = torch.polar(torch.ones_like(freqs), freqs) # complex64
|
12 |
+
return freqs_cis
|
13 |
+
|
14 |
+
|
15 |
+
def reshape_for_broadcast(freqs_cis: torch.Tensor,
|
16 |
+
x: torch.Tensor,):
|
17 |
+
ndim = x.ndim
|
18 |
+
assert 0 <= 1 < ndim
|
19 |
+
assert freqs_cis.shape == (x.shape[1], x.shape[-1])
|
20 |
+
shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
|
21 |
+
return freqs_cis.view(*shape)
|
22 |
+
|
23 |
+
|
24 |
+
def compute_rope(q, freqs_cis):
|
25 |
+
return q * freqs_cis
|
26 |
+
|
27 |
+
|
28 |
+
def apply_rotary_emb(
|
29 |
+
xq: torch.Tensor,
|
30 |
+
xk: torch.Tensor,
|
31 |
+
freqs_cis: torch.Tensor,
|
32 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
33 |
+
# xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
|
34 |
+
# xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
|
35 |
+
xq1, xq2 = xq.chunk(2, dim=-1)
|
36 |
+
xq_ = torch.view_as_complex(torch.stack((xq1, xq2), dim=-1).float())
|
37 |
+
|
38 |
+
xk1, xk2 = xk.chunk(2, dim=-1)
|
39 |
+
xk_ = torch.view_as_complex(torch.stack((xk1, xk2), dim=-1).float())
|
40 |
+
|
41 |
+
freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
|
42 |
+
xq_out = torch.view_as_real(compute_rope(xq_, freqs_cis)).flatten(3)
|
43 |
+
xk_out = torch.view_as_real(compute_rope(xk_, freqs_cis)).flatten(3)
|
44 |
+
return xq_out.type_as(xq), xk_out.type_as(xk)
|
45 |
+
|
46 |
+
|
47 |
+
if __name__ == '__main__':
|
48 |
+
# Move data to CUDA
|
49 |
+
freq_cis = precompute_freqs_cis(4, 5).cuda()
|
50 |
+
x = torch.rand(1, 5, 1, 4).cuda()
|
51 |
+
y = torch.rand(1, 5, 1, 4).cuda()
|
52 |
+
|
53 |
+
# First method
|
54 |
+
start_time = time.time()
|
55 |
+
for _ in range(20000):
|
56 |
+
x1, y1 = apply_rotary_emb(x, y, freq_cis)
|
57 |
+
end_time = time.time()
|
58 |
+
print(f"Method 1 time cost: {end_time - start_time} seconds")
|
59 |
+
|
60 |
+
# Prepare data for the second method
|
61 |
+
x = x.permute(0, 2, 1, 3)
|
62 |
+
y = y.permute(0, 2, 1, 3)
|
63 |
+
rope = RotaryEmbedding(4).cuda()
|
64 |
+
|
65 |
+
# Second method
|
66 |
+
start_time = time.time()
|
67 |
+
for _ in range(20000):
|
68 |
+
x2, y2 = rope(x, y)
|
69 |
+
end_time = time.time()
|
70 |
+
print(f"Method 2 time cost: {end_time - start_time} seconds")
|
71 |
+
|
72 |
+
# Print the results
|
73 |
+
print(x1)
|
74 |
+
print(x2)
|
src/models/utils/modules.py
ADDED
@@ -0,0 +1,374 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import torch.utils.checkpoint
|
5 |
+
from torch.cuda.amp import autocast
|
6 |
+
import math
|
7 |
+
import einops
|
8 |
+
from einops import rearrange, repeat
|
9 |
+
from inspect import isfunction
|
10 |
+
from .timm import trunc_normal_
|
11 |
+
|
12 |
+
|
13 |
+
# disable in checkpoint mode
|
14 |
+
# @torch.jit.script
|
15 |
+
def film_modulate(x, shift, scale):
|
16 |
+
return x * (1 + scale) + shift
|
17 |
+
|
18 |
+
|
19 |
+
def timestep_embedding(timesteps, dim, max_period=10000):
|
20 |
+
"""
|
21 |
+
Create sinusoidal timestep embeddings.
|
22 |
+
|
23 |
+
:param timesteps: a 1-D Tensor of N indices, one per batch element.
|
24 |
+
These may be fractional.
|
25 |
+
:param dim: the dimension of the output.
|
26 |
+
:param max_period: controls the minimum frequency of the embeddings.
|
27 |
+
:return: an [N x dim] Tensor of positional embeddings.
|
28 |
+
"""
|
29 |
+
half = dim // 2
|
30 |
+
freqs = torch.exp(
|
31 |
+
-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
|
32 |
+
).to(device=timesteps.device)
|
33 |
+
args = timesteps[:, None].float() * freqs[None]
|
34 |
+
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
|
35 |
+
if dim % 2:
|
36 |
+
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
|
37 |
+
return embedding
|
38 |
+
|
39 |
+
|
40 |
+
class TimestepEmbedder(nn.Module):
|
41 |
+
"""
|
42 |
+
Embeds scalar timesteps into vector representations.
|
43 |
+
"""
|
44 |
+
|
45 |
+
def __init__(self, hidden_size, frequency_embedding_size=256,
|
46 |
+
out_size=None):
|
47 |
+
super().__init__()
|
48 |
+
if out_size is None:
|
49 |
+
out_size = hidden_size
|
50 |
+
self.mlp = nn.Sequential(
|
51 |
+
nn.Linear(frequency_embedding_size, hidden_size, bias=True),
|
52 |
+
nn.SiLU(),
|
53 |
+
nn.Linear(hidden_size, out_size, bias=True),
|
54 |
+
)
|
55 |
+
self.frequency_embedding_size = frequency_embedding_size
|
56 |
+
|
57 |
+
def forward(self, t):
|
58 |
+
t_freq = timestep_embedding(t, self.frequency_embedding_size).type(
|
59 |
+
self.mlp[0].weight.dtype)
|
60 |
+
t_emb = self.mlp(t_freq)
|
61 |
+
return t_emb
|
62 |
+
|
63 |
+
|
64 |
+
def patchify(imgs, patch_size, input_type='2d'):
|
65 |
+
if input_type == '2d':
|
66 |
+
x = einops.rearrange(imgs, 'B C (h p1) (w p2) -> B (h w) (p1 p2 C)', p1=patch_size, p2=patch_size)
|
67 |
+
elif input_type == '1d':
|
68 |
+
x = einops.rearrange(imgs, 'B C (h p1) -> B h (p1 C)', p1=patch_size)
|
69 |
+
return x
|
70 |
+
|
71 |
+
|
72 |
+
def unpatchify(x, channels=3, input_type='2d', img_size=None):
|
73 |
+
if input_type == '2d':
|
74 |
+
patch_size = int((x.shape[2] // channels) ** 0.5)
|
75 |
+
# h = w = int(x.shape[1] ** .5)
|
76 |
+
h, w = img_size[0] // patch_size, img_size[1] // patch_size
|
77 |
+
assert h * w == x.shape[1] and patch_size ** 2 * channels == x.shape[2]
|
78 |
+
x = einops.rearrange(x, 'B (h w) (p1 p2 C) -> B C (h p1) (w p2)', h=h,
|
79 |
+
p1=patch_size, p2=patch_size)
|
80 |
+
elif input_type == '1d':
|
81 |
+
patch_size = int((x.shape[2] // channels))
|
82 |
+
h = x.shape[1]
|
83 |
+
assert patch_size * channels == x.shape[2]
|
84 |
+
x = einops.rearrange(x, 'B h (p1 C) -> B C (h p1)', h=h, p1=patch_size)
|
85 |
+
return x
|
86 |
+
|
87 |
+
|
88 |
+
class PatchEmbed(nn.Module):
|
89 |
+
"""
|
90 |
+
Image to Patch Embedding
|
91 |
+
"""
|
92 |
+
|
93 |
+
def __init__(self, patch_size, in_chans=3, embed_dim=768, input_type='2d'):
|
94 |
+
super().__init__()
|
95 |
+
self.patch_size = patch_size
|
96 |
+
self.input_type = input_type
|
97 |
+
if input_type == '2d':
|
98 |
+
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=True)
|
99 |
+
elif input_type == '1d':
|
100 |
+
self.proj = nn.Conv1d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=True)
|
101 |
+
|
102 |
+
def forward(self, x):
|
103 |
+
if self.input_type == '2d':
|
104 |
+
B, C, H, W = x.shape
|
105 |
+
assert H % self.patch_size == 0 and W % self.patch_size == 0
|
106 |
+
elif self.input_type == '1d':
|
107 |
+
B, C, H = x.shape
|
108 |
+
assert H % self.patch_size == 0
|
109 |
+
|
110 |
+
x = self.proj(x).flatten(2).transpose(1, 2)
|
111 |
+
return x
|
112 |
+
|
113 |
+
|
114 |
+
class PositionalConvEmbedding(nn.Module):
|
115 |
+
"""
|
116 |
+
Relative positional embedding used in HuBERT
|
117 |
+
"""
|
118 |
+
|
119 |
+
def __init__(self, dim=768, kernel_size=128, groups=16):
|
120 |
+
super().__init__()
|
121 |
+
self.conv = nn.Conv1d(
|
122 |
+
dim,
|
123 |
+
dim,
|
124 |
+
kernel_size=kernel_size,
|
125 |
+
padding=kernel_size // 2,
|
126 |
+
groups=groups,
|
127 |
+
bias=True
|
128 |
+
)
|
129 |
+
self.conv = nn.utils.parametrizations.weight_norm(self.conv, name="weight", dim=2)
|
130 |
+
|
131 |
+
def forward(self, x):
|
132 |
+
# B C T
|
133 |
+
x = self.conv(x)
|
134 |
+
x = F.gelu(x[:, :, :-1])
|
135 |
+
return x
|
136 |
+
|
137 |
+
|
138 |
+
class SinusoidalPositionalEncoding(nn.Module):
|
139 |
+
def __init__(self, dim, length):
|
140 |
+
super(SinusoidalPositionalEncoding, self).__init__()
|
141 |
+
self.length = length
|
142 |
+
self.dim = dim
|
143 |
+
self.register_buffer('pe', self._generate_positional_encoding(length, dim))
|
144 |
+
|
145 |
+
def _generate_positional_encoding(self, length, dim):
|
146 |
+
pe = torch.zeros(length, dim)
|
147 |
+
position = torch.arange(0, length, dtype=torch.float).unsqueeze(1)
|
148 |
+
div_term = torch.exp(torch.arange(0, dim, 2).float() * (-math.log(10000.0) / dim))
|
149 |
+
|
150 |
+
pe[:, 0::2] = torch.sin(position * div_term)
|
151 |
+
pe[:, 1::2] = torch.cos(position * div_term)
|
152 |
+
|
153 |
+
pe = pe.unsqueeze(0)
|
154 |
+
return pe
|
155 |
+
|
156 |
+
def forward(self, x):
|
157 |
+
x = x + self.pe[:, :x.size(1)]
|
158 |
+
return x
|
159 |
+
|
160 |
+
|
161 |
+
class PE_wrapper(nn.Module):
|
162 |
+
def __init__(self, dim=768, method='abs', length=None, **kwargs):
|
163 |
+
super().__init__()
|
164 |
+
self.method = method
|
165 |
+
if method == 'abs':
|
166 |
+
# init absolute pe like UViT
|
167 |
+
self.length = length
|
168 |
+
self.abs_pe = nn.Parameter(torch.zeros(1, length, dim))
|
169 |
+
trunc_normal_(self.abs_pe, std=.02)
|
170 |
+
elif method == 'conv':
|
171 |
+
self.conv_pe = PositionalConvEmbedding(dim=dim, **kwargs)
|
172 |
+
elif method == 'sinu':
|
173 |
+
self.sinu_pe = SinusoidalPositionalEncoding(dim=dim, length=length)
|
174 |
+
elif method == 'none':
|
175 |
+
# skip pe
|
176 |
+
self.id = nn.Identity()
|
177 |
+
else:
|
178 |
+
raise NotImplementedError
|
179 |
+
|
180 |
+
def forward(self, x):
|
181 |
+
if self.method == 'abs':
|
182 |
+
_, L, _ = x.shape
|
183 |
+
assert L <= self.length
|
184 |
+
x = x + self.abs_pe[:, :L, :]
|
185 |
+
elif self.method == 'conv':
|
186 |
+
x = x + self.conv_pe(x)
|
187 |
+
elif self.method == 'sinu':
|
188 |
+
x = self.sinu_pe(x)
|
189 |
+
elif self.method == 'none':
|
190 |
+
x = self.id(x)
|
191 |
+
else:
|
192 |
+
raise NotImplementedError
|
193 |
+
return x
|
194 |
+
|
195 |
+
|
196 |
+
class RMSNorm(torch.nn.Module):
|
197 |
+
def __init__(self, dim: int, eps: float = 1e-6):
|
198 |
+
"""
|
199 |
+
Initialize the RMSNorm normalization layer.
|
200 |
+
|
201 |
+
Args:
|
202 |
+
dim (int): The dimension of the input tensor.
|
203 |
+
eps (float, optional): A small value added to the denominator for numerical stability. Default is 1e-6.
|
204 |
+
|
205 |
+
Attributes:
|
206 |
+
eps (float): A small value added to the denominator for numerical stability.
|
207 |
+
weight (nn.Parameter): Learnable scaling parameter.
|
208 |
+
|
209 |
+
"""
|
210 |
+
super().__init__()
|
211 |
+
self.eps = eps
|
212 |
+
self.weight = nn.Parameter(torch.ones(dim))
|
213 |
+
|
214 |
+
def _norm(self, x):
|
215 |
+
"""
|
216 |
+
Apply the RMSNorm normalization to the input tensor.
|
217 |
+
|
218 |
+
Args:
|
219 |
+
x (torch.Tensor): The input tensor.
|
220 |
+
|
221 |
+
Returns:
|
222 |
+
torch.Tensor: The normalized tensor.
|
223 |
+
|
224 |
+
"""
|
225 |
+
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
|
226 |
+
|
227 |
+
def forward(self, x):
|
228 |
+
"""
|
229 |
+
Forward pass through the RMSNorm layer.
|
230 |
+
|
231 |
+
Args:
|
232 |
+
x (torch.Tensor): The input tensor.
|
233 |
+
|
234 |
+
Returns:
|
235 |
+
torch.Tensor: The output tensor after applying RMSNorm.
|
236 |
+
|
237 |
+
"""
|
238 |
+
output = self._norm(x.float()).type_as(x)
|
239 |
+
return output * self.weight
|
240 |
+
|
241 |
+
|
242 |
+
class GELU(nn.Module):
|
243 |
+
|
244 |
+
def __init__(self, dim_in: int, dim_out: int, approximate: str = "none",
|
245 |
+
bias: bool = True):
|
246 |
+
super().__init__()
|
247 |
+
self.proj = nn.Linear(dim_in, dim_out, bias=bias)
|
248 |
+
self.approximate = approximate
|
249 |
+
|
250 |
+
def gelu(self, gate: torch.Tensor) -> torch.Tensor:
|
251 |
+
if gate.device.type != "mps":
|
252 |
+
return F.gelu(gate, approximate=self.approximate)
|
253 |
+
# mps: gelu is not implemented for float16
|
254 |
+
return F.gelu(gate.to(dtype=torch.float32),
|
255 |
+
approximate=self.approximate).to(dtype=gate.dtype)
|
256 |
+
|
257 |
+
def forward(self, hidden_states):
|
258 |
+
hidden_states = self.proj(hidden_states)
|
259 |
+
hidden_states = self.gelu(hidden_states)
|
260 |
+
return hidden_states
|
261 |
+
|
262 |
+
|
263 |
+
class GEGLU(nn.Module):
|
264 |
+
def __init__(self, dim_in: int, dim_out: int, bias: bool = True):
|
265 |
+
super().__init__()
|
266 |
+
self.proj = nn.Linear(dim_in, dim_out * 2, bias=bias)
|
267 |
+
|
268 |
+
def gelu(self, gate: torch.Tensor) -> torch.Tensor:
|
269 |
+
if gate.device.type != "mps":
|
270 |
+
return F.gelu(gate)
|
271 |
+
# mps: gelu is not implemented for float16
|
272 |
+
return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)
|
273 |
+
|
274 |
+
def forward(self, hidden_states):
|
275 |
+
hidden_states = self.proj(hidden_states)
|
276 |
+
hidden_states, gate = hidden_states.chunk(2, dim=-1)
|
277 |
+
return hidden_states * self.gelu(gate)
|
278 |
+
|
279 |
+
|
280 |
+
class ApproximateGELU(nn.Module):
|
281 |
+
def __init__(self, dim_in: int, dim_out: int, bias: bool = True):
|
282 |
+
super().__init__()
|
283 |
+
self.proj = nn.Linear(dim_in, dim_out, bias=bias)
|
284 |
+
|
285 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
286 |
+
x = self.proj(x)
|
287 |
+
return x * torch.sigmoid(1.702 * x)
|
288 |
+
|
289 |
+
|
290 |
+
# disable in checkpoint mode
|
291 |
+
# @torch.jit.script
|
292 |
+
def snake_beta(x, alpha, beta):
|
293 |
+
return x + beta * torch.sin(x * alpha).pow(2)
|
294 |
+
|
295 |
+
|
296 |
+
class Snake(nn.Module):
|
297 |
+
def __init__(self, dim_in, dim_out, bias,
|
298 |
+
alpha_trainable=True):
|
299 |
+
super().__init__()
|
300 |
+
self.proj = nn.Linear(dim_in, dim_out, bias=bias)
|
301 |
+
self.alpha = nn.Parameter(torch.ones(1, 1, dim_out))
|
302 |
+
self.beta = nn.Parameter(torch.ones(1, 1, dim_out))
|
303 |
+
self.alpha.requires_grad = alpha_trainable
|
304 |
+
self.beta.requires_grad = alpha_trainable
|
305 |
+
|
306 |
+
def forward(self, x):
|
307 |
+
x = self.proj(x)
|
308 |
+
x = snake_beta(x, self.alpha, self.beta)
|
309 |
+
return x
|
310 |
+
|
311 |
+
|
312 |
+
class GESnake(nn.Module):
|
313 |
+
def __init__(self, dim_in, dim_out, bias,
|
314 |
+
alpha_trainable=True):
|
315 |
+
super().__init__()
|
316 |
+
self.proj = nn.Linear(dim_in, dim_out * 2, bias=bias)
|
317 |
+
self.alpha = nn.Parameter(torch.ones(1, 1, dim_out))
|
318 |
+
self.beta = nn.Parameter(torch.ones(1, 1, dim_out))
|
319 |
+
self.alpha.requires_grad = alpha_trainable
|
320 |
+
self.beta.requires_grad = alpha_trainable
|
321 |
+
|
322 |
+
def forward(self, x):
|
323 |
+
x = self.proj(x)
|
324 |
+
x, gate = x.chunk(2, dim=-1)
|
325 |
+
return x * snake_beta(gate, self.alpha, self.beta)
|
326 |
+
|
327 |
+
|
328 |
+
class FeedForward(nn.Module):
|
329 |
+
def __init__(
|
330 |
+
self,
|
331 |
+
dim,
|
332 |
+
dim_out=None,
|
333 |
+
mult=4,
|
334 |
+
dropout=0.0,
|
335 |
+
activation_fn="geglu",
|
336 |
+
final_dropout=False,
|
337 |
+
inner_dim=None,
|
338 |
+
bias=True,
|
339 |
+
):
|
340 |
+
super().__init__()
|
341 |
+
if inner_dim is None:
|
342 |
+
inner_dim = int(dim * mult)
|
343 |
+
dim_out = dim_out if dim_out is not None else dim
|
344 |
+
|
345 |
+
if activation_fn == "gelu":
|
346 |
+
act_fn = GELU(dim, inner_dim, bias=bias)
|
347 |
+
elif activation_fn == "gelu-approximate":
|
348 |
+
act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias)
|
349 |
+
elif activation_fn == "geglu":
|
350 |
+
act_fn = GEGLU(dim, inner_dim, bias=bias)
|
351 |
+
elif activation_fn == "geglu-approximate":
|
352 |
+
act_fn = ApproximateGELU(dim, inner_dim, bias=bias)
|
353 |
+
elif activation_fn == "snake":
|
354 |
+
act_fn = Snake(dim, inner_dim, bias=bias)
|
355 |
+
elif activation_fn == "gesnake":
|
356 |
+
act_fn = GESnake(dim, inner_dim, bias=bias)
|
357 |
+
else:
|
358 |
+
raise NotImplementedError
|
359 |
+
|
360 |
+
self.net = nn.ModuleList([])
|
361 |
+
# project in
|
362 |
+
self.net.append(act_fn)
|
363 |
+
# project dropout
|
364 |
+
self.net.append(nn.Dropout(dropout))
|
365 |
+
# project out
|
366 |
+
self.net.append(nn.Linear(inner_dim, dim_out, bias=bias))
|
367 |
+
# FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
|
368 |
+
if final_dropout:
|
369 |
+
self.net.append(nn.Dropout(dropout))
|
370 |
+
|
371 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
372 |
+
for module in self.net:
|
373 |
+
hidden_states = module(hidden_states)
|
374 |
+
return hidden_states
|
src/models/utils/rotary.py
ADDED
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
"this rope is faster than llama rope with jit script"
|
4 |
+
|
5 |
+
|
6 |
+
def rotate_half(x):
|
7 |
+
x1, x2 = x.chunk(2, dim=-1)
|
8 |
+
return torch.cat((-x2, x1), dim=-1)
|
9 |
+
|
10 |
+
|
11 |
+
# disable in checkpoint mode
|
12 |
+
# @torch.jit.script
|
13 |
+
def apply_rotary_pos_emb(x, cos, sin):
|
14 |
+
# NOTE: This could probably be moved to Triton
|
15 |
+
# Handle a possible sequence length mismatch in between q and k
|
16 |
+
cos = cos[:, :, : x.shape[-2], :]
|
17 |
+
sin = sin[:, :, : x.shape[-2], :]
|
18 |
+
return (x * cos) + (rotate_half(x) * sin)
|
19 |
+
|
20 |
+
|
21 |
+
class RotaryEmbedding(torch.nn.Module):
|
22 |
+
"""
|
23 |
+
The rotary position embeddings from RoFormer_ (Su et. al).
|
24 |
+
A crucial insight from the method is that the query and keys are
|
25 |
+
transformed by rotation matrices which depend on the relative positions.
|
26 |
+
|
27 |
+
Other implementations are available in the Rotary Transformer repo_ and in
|
28 |
+
GPT-NeoX_, GPT-NeoX was an inspiration
|
29 |
+
|
30 |
+
.. _RoFormer: https://arxiv.org/abs/2104.09864
|
31 |
+
.. _repo: https://github.com/ZhuiyiTechnology/roformer
|
32 |
+
.. _GPT-NeoX: https://github.com/EleutherAI/gpt-neox
|
33 |
+
|
34 |
+
|
35 |
+
.. warning: Please note that this embedding is not registered on purpose, as it is transformative
|
36 |
+
(it does not create the embedding dimension) and will likely be picked up (imported) on a ad-hoc basis
|
37 |
+
"""
|
38 |
+
|
39 |
+
def __init__(self, dim: int):
|
40 |
+
super().__init__()
|
41 |
+
# Generate and save the inverse frequency buffer (non trainable)
|
42 |
+
inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
|
43 |
+
self.register_buffer("inv_freq", inv_freq)
|
44 |
+
self._seq_len_cached = None
|
45 |
+
self._cos_cached = None
|
46 |
+
self._sin_cached = None
|
47 |
+
|
48 |
+
def _update_cos_sin_tables(self, x, seq_dimension=-2):
|
49 |
+
# expect input: B, H, L, D
|
50 |
+
seq_len = x.shape[seq_dimension]
|
51 |
+
|
52 |
+
# Reset the tables if the sequence length has changed,
|
53 |
+
# or if we're on a new device (possibly due to tracing for instance)
|
54 |
+
# also make sure dtype wont change
|
55 |
+
if (
|
56 |
+
seq_len != self._seq_len_cached
|
57 |
+
or self._cos_cached.device != x.device
|
58 |
+
or self._cos_cached.dtype != x.dtype
|
59 |
+
):
|
60 |
+
self._seq_len_cached = seq_len
|
61 |
+
t = torch.arange(
|
62 |
+
x.shape[seq_dimension], device=x.device, dtype=torch.float32
|
63 |
+
)
|
64 |
+
freqs = torch.einsum("i,j->ij", t, self.inv_freq.to(x.dtype))
|
65 |
+
emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
|
66 |
+
|
67 |
+
self._cos_cached = emb.cos()[None, None, :, :].to(x.dtype)
|
68 |
+
self._sin_cached = emb.sin()[None, None, :, :].to(x.dtype)
|
69 |
+
|
70 |
+
return self._cos_cached, self._sin_cached
|
71 |
+
|
72 |
+
def forward(self, q, k):
|
73 |
+
self._cos_cached, self._sin_cached = self._update_cos_sin_tables(
|
74 |
+
q.float(), seq_dimension=-2
|
75 |
+
)
|
76 |
+
if k is not None:
|
77 |
+
return (
|
78 |
+
apply_rotary_pos_emb(q.float(),
|
79 |
+
self._cos_cached,
|
80 |
+
self._sin_cached).type_as(q),
|
81 |
+
apply_rotary_pos_emb(k.float(),
|
82 |
+
self._cos_cached,
|
83 |
+
self._sin_cached).type_as(k),
|
84 |
+
)
|
85 |
+
else:
|
86 |
+
return (
|
87 |
+
apply_rotary_pos_emb(q.float(),
|
88 |
+
self._cos_cached,
|
89 |
+
self._sin_cached).type_as(q),
|
90 |
+
None
|
91 |
+
)
|
src/models/utils/span_mask.py
ADDED
@@ -0,0 +1,146 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import torch
|
3 |
+
from typing import Optional, Tuple
|
4 |
+
|
5 |
+
|
6 |
+
def compute_mask_indices(
|
7 |
+
shape: Tuple[int, int],
|
8 |
+
padding_mask: Optional[torch.Tensor],
|
9 |
+
mask_prob: float,
|
10 |
+
mask_length: int,
|
11 |
+
mask_type: str = "static",
|
12 |
+
mask_other: float = 0.0,
|
13 |
+
min_masks: int = 0,
|
14 |
+
no_overlap: bool = False,
|
15 |
+
min_space: int = 0,
|
16 |
+
) -> np.ndarray:
|
17 |
+
"""
|
18 |
+
Computes random mask spans for a given shape
|
19 |
+
|
20 |
+
Args:
|
21 |
+
shape: the the shape for which to compute masks.
|
22 |
+
should be of size 2 where first element is batch size and 2nd is timesteps
|
23 |
+
padding_mask: optional padding mask of the same size as shape, which will prevent masking padded elements
|
24 |
+
mask_prob: probability for each token to be chosen as start of the span to be masked. this will be multiplied by
|
25 |
+
number of timesteps divided by length of mask span to mask approximately this percentage of all elements.
|
26 |
+
however due to overlaps, the actual number will be smaller (unless no_overlap is True)
|
27 |
+
mask_type: how to compute mask lengths
|
28 |
+
static = fixed size
|
29 |
+
uniform = sample from uniform distribution [mask_other, mask_length*2]
|
30 |
+
normal = sample from normal distribution with mean mask_length and stdev mask_other. mask is min 1 element
|
31 |
+
poisson = sample from possion distribution with lambda = mask length
|
32 |
+
min_masks: minimum number of masked spans
|
33 |
+
no_overlap: if false, will switch to an alternative recursive algorithm that prevents spans from overlapping
|
34 |
+
min_space: only used if no_overlap is True, this is how many elements to keep unmasked between spans
|
35 |
+
"""
|
36 |
+
|
37 |
+
bsz, all_sz = shape
|
38 |
+
mask = np.full((bsz, all_sz), False)
|
39 |
+
|
40 |
+
# Convert mask_prob to a NumPy array
|
41 |
+
mask_prob = np.array(mask_prob)
|
42 |
+
|
43 |
+
# Calculate all_num_mask for each element in the batch
|
44 |
+
all_num_mask = np.floor(mask_prob * all_sz / float(mask_length) + np.random.rand(bsz)).astype(int)
|
45 |
+
|
46 |
+
# Apply the max operation with min_masks for each element
|
47 |
+
all_num_mask = np.maximum(min_masks, all_num_mask)
|
48 |
+
|
49 |
+
mask_idcs = []
|
50 |
+
for i in range(bsz):
|
51 |
+
if padding_mask is not None:
|
52 |
+
sz = all_sz - padding_mask[i].long().sum().item()
|
53 |
+
num_mask = int(
|
54 |
+
# add a random number for probabilistic rounding
|
55 |
+
mask_prob * sz / float(mask_length)
|
56 |
+
+ np.random.rand()
|
57 |
+
)
|
58 |
+
num_mask = max(min_masks, num_mask)
|
59 |
+
else:
|
60 |
+
sz = all_sz
|
61 |
+
num_mask = all_num_mask[i]
|
62 |
+
|
63 |
+
if mask_type == "static":
|
64 |
+
lengths = np.full(num_mask, mask_length)
|
65 |
+
elif mask_type == "uniform":
|
66 |
+
lengths = np.random.randint(mask_other, mask_length * 2 + 1, size=num_mask)
|
67 |
+
elif mask_type == "normal":
|
68 |
+
lengths = np.random.normal(mask_length, mask_other, size=num_mask)
|
69 |
+
lengths = [max(1, int(round(x))) for x in lengths]
|
70 |
+
elif mask_type == "poisson":
|
71 |
+
lengths = np.random.poisson(mask_length, size=num_mask)
|
72 |
+
lengths = [int(round(x)) for x in lengths]
|
73 |
+
else:
|
74 |
+
raise Exception("unknown mask selection " + mask_type)
|
75 |
+
|
76 |
+
if sum(lengths) == 0:
|
77 |
+
lengths[0] = min(mask_length, sz - 1)
|
78 |
+
|
79 |
+
if no_overlap:
|
80 |
+
mask_idc = []
|
81 |
+
|
82 |
+
def arrange(s, e, length, keep_length):
|
83 |
+
span_start = np.random.randint(s, e - length)
|
84 |
+
mask_idc.extend(span_start + i for i in range(length))
|
85 |
+
|
86 |
+
new_parts = []
|
87 |
+
if span_start - s - min_space >= keep_length:
|
88 |
+
new_parts.append((s, span_start - min_space + 1))
|
89 |
+
if e - span_start - keep_length - min_space > keep_length:
|
90 |
+
new_parts.append((span_start + length + min_space, e))
|
91 |
+
return new_parts
|
92 |
+
|
93 |
+
parts = [(0, sz)]
|
94 |
+
min_length = min(lengths)
|
95 |
+
for length in sorted(lengths, reverse=True):
|
96 |
+
lens = np.fromiter(
|
97 |
+
(e - s if e - s >= length + min_space else 0 for s, e in parts),
|
98 |
+
np.int,
|
99 |
+
)
|
100 |
+
l_sum = np.sum(lens)
|
101 |
+
if l_sum == 0:
|
102 |
+
break
|
103 |
+
probs = lens / np.sum(lens)
|
104 |
+
c = np.random.choice(len(parts), p=probs)
|
105 |
+
s, e = parts.pop(c)
|
106 |
+
parts.extend(arrange(s, e, length, min_length))
|
107 |
+
mask_idc = np.asarray(mask_idc)
|
108 |
+
else:
|
109 |
+
min_len = min(lengths)
|
110 |
+
if sz - min_len <= num_mask:
|
111 |
+
min_len = sz - num_mask - 1
|
112 |
+
|
113 |
+
mask_idc = np.random.choice(sz - min_len, num_mask, replace=False)
|
114 |
+
|
115 |
+
mask_idc = np.asarray(
|
116 |
+
[
|
117 |
+
mask_idc[j] + offset
|
118 |
+
for j in range(len(mask_idc))
|
119 |
+
for offset in range(lengths[j])
|
120 |
+
]
|
121 |
+
)
|
122 |
+
|
123 |
+
mask_idcs.append(np.unique(mask_idc[mask_idc < sz]))
|
124 |
+
# min_len = min([len(m) for m in mask_idcs])
|
125 |
+
for i, mask_idc in enumerate(mask_idcs):
|
126 |
+
# if len(mask_idc) > min_len:
|
127 |
+
# mask_idc = np.random.choice(mask_idc, min_len, replace=False)
|
128 |
+
mask[i, mask_idc] = True
|
129 |
+
|
130 |
+
return torch.tensor(mask)
|
131 |
+
|
132 |
+
|
133 |
+
if __name__ == '__main__':
|
134 |
+
mask = compute_mask_indices(
|
135 |
+
shape=[4, 500],
|
136 |
+
padding_mask=None,
|
137 |
+
mask_prob=[0.65, 0.5, 0.65, 0.65],
|
138 |
+
mask_length=10,
|
139 |
+
mask_type="static",
|
140 |
+
mask_other=0.0,
|
141 |
+
min_masks=1,
|
142 |
+
no_overlap=False,
|
143 |
+
min_space=0,
|
144 |
+
)
|
145 |
+
print(mask)
|
146 |
+
print(mask.sum(dim=1))
|
src/models/utils/timm.py
ADDED
@@ -0,0 +1,114 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# code from timm 0.3.2
|
2 |
+
import torch
|
3 |
+
import torch.nn as nn
|
4 |
+
import math
|
5 |
+
import warnings
|
6 |
+
|
7 |
+
|
8 |
+
def _no_grad_trunc_normal_(tensor, mean, std, a, b):
|
9 |
+
# Cut & paste from PyTorch official master until it's in a few official releases - RW
|
10 |
+
# Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
|
11 |
+
def norm_cdf(x):
|
12 |
+
# Computes standard normal cumulative distribution function
|
13 |
+
return (1. + math.erf(x / math.sqrt(2.))) / 2.
|
14 |
+
|
15 |
+
if (mean < a - 2 * std) or (mean > b + 2 * std):
|
16 |
+
warnings.warn("mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
|
17 |
+
"The distribution of values may be incorrect.",
|
18 |
+
stacklevel=2)
|
19 |
+
|
20 |
+
with torch.no_grad():
|
21 |
+
# Values are generated by using a truncated uniform distribution and
|
22 |
+
# then using the inverse CDF for the normal distribution.
|
23 |
+
# Get upper and lower cdf values
|
24 |
+
l = norm_cdf((a - mean) / std)
|
25 |
+
u = norm_cdf((b - mean) / std)
|
26 |
+
|
27 |
+
# Uniformly fill tensor with values from [l, u], then translate to
|
28 |
+
# [2l-1, 2u-1].
|
29 |
+
tensor.uniform_(2 * l - 1, 2 * u - 1)
|
30 |
+
|
31 |
+
# Use inverse cdf transform for normal distribution to get truncated
|
32 |
+
# standard normal
|
33 |
+
tensor.erfinv_()
|
34 |
+
|
35 |
+
# Transform to proper mean, std
|
36 |
+
tensor.mul_(std * math.sqrt(2.))
|
37 |
+
tensor.add_(mean)
|
38 |
+
|
39 |
+
# Clamp to ensure it's in the proper range
|
40 |
+
tensor.clamp_(min=a, max=b)
|
41 |
+
return tensor
|
42 |
+
|
43 |
+
|
44 |
+
def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
|
45 |
+
# type: (Tensor, float, float, float, float) -> Tensor
|
46 |
+
r"""Fills the input Tensor with values drawn from a truncated
|
47 |
+
normal distribution. The values are effectively drawn from the
|
48 |
+
normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
|
49 |
+
with values outside :math:`[a, b]` redrawn until they are within
|
50 |
+
the bounds. The method used for generating the random values works
|
51 |
+
best when :math:`a \leq \text{mean} \leq b`.
|
52 |
+
Args:
|
53 |
+
tensor: an n-dimensional `torch.Tensor`
|
54 |
+
mean: the mean of the normal distribution
|
55 |
+
std: the standard deviation of the normal distribution
|
56 |
+
a: the minimum cutoff value
|
57 |
+
b: the maximum cutoff value
|
58 |
+
Examples:
|
59 |
+
>>> w = torch.empty(3, 5)
|
60 |
+
>>> nn.init.trunc_normal_(w)
|
61 |
+
"""
|
62 |
+
return _no_grad_trunc_normal_(tensor, mean, std, a, b)
|
63 |
+
|
64 |
+
|
65 |
+
def drop_path(x, drop_prob: float = 0., training: bool = False):
|
66 |
+
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
|
67 |
+
|
68 |
+
This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
|
69 |
+
the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
|
70 |
+
See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
|
71 |
+
changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
|
72 |
+
'survival rate' as the argument.
|
73 |
+
|
74 |
+
"""
|
75 |
+
if drop_prob == 0. or not training:
|
76 |
+
return x
|
77 |
+
keep_prob = 1 - drop_prob
|
78 |
+
shape = (x.shape[0],) + (1,) * (x.ndim - 1) # work with diff dim tensors, not just 2D ConvNets
|
79 |
+
random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
|
80 |
+
random_tensor.floor_() # binarize
|
81 |
+
output = x.div(keep_prob) * random_tensor
|
82 |
+
return output
|
83 |
+
|
84 |
+
|
85 |
+
class DropPath(nn.Module):
|
86 |
+
"""Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
|
87 |
+
"""
|
88 |
+
|
89 |
+
def __init__(self, drop_prob=None):
|
90 |
+
super(DropPath, self).__init__()
|
91 |
+
self.drop_prob = drop_prob
|
92 |
+
|
93 |
+
def forward(self, x):
|
94 |
+
return drop_path(x, self.drop_prob, self.training)
|
95 |
+
|
96 |
+
|
97 |
+
class Mlp(nn.Module):
|
98 |
+
def __init__(self, in_features, hidden_features=None, out_features=None,
|
99 |
+
act_layer=nn.GELU, drop=0.):
|
100 |
+
super().__init__()
|
101 |
+
out_features = out_features or in_features
|
102 |
+
hidden_features = hidden_features or in_features
|
103 |
+
self.fc1 = nn.Linear(in_features, hidden_features)
|
104 |
+
self.act = act_layer()
|
105 |
+
self.fc2 = nn.Linear(hidden_features, out_features)
|
106 |
+
self.drop = nn.Dropout(drop)
|
107 |
+
|
108 |
+
def forward(self, x):
|
109 |
+
x = self.fc1(x)
|
110 |
+
x = self.act(x)
|
111 |
+
x = self.drop(x)
|
112 |
+
x = self.fc2(x)
|
113 |
+
x = self.drop(x)
|
114 |
+
return x
|
src/modules/autoencoder_wrapper.py
ADDED
@@ -0,0 +1,83 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
from .dac import DAC
|
4 |
+
from .stable_vae import load_vae
|
5 |
+
|
6 |
+
|
7 |
+
class Autoencoder(nn.Module):
|
8 |
+
def __init__(self, ckpt_path, model_type='dac', quantization_first=False):
|
9 |
+
super(Autoencoder, self).__init__()
|
10 |
+
self.model_type = model_type
|
11 |
+
if self.model_type == 'dac':
|
12 |
+
model = DAC.load(ckpt_path)
|
13 |
+
elif self.model_type == 'stable_vae':
|
14 |
+
model = load_vae(ckpt_path)
|
15 |
+
else:
|
16 |
+
raise NotImplementedError(f"Model type not implemented: {self.model_type}")
|
17 |
+
self.ae = model.eval()
|
18 |
+
self.quantization_first = quantization_first
|
19 |
+
print(f'Autoencoder quantization first mode: {quantization_first}')
|
20 |
+
|
21 |
+
@torch.no_grad()
|
22 |
+
def forward(self, audio=None, embedding=None):
|
23 |
+
if self.model_type == 'dac':
|
24 |
+
return self.process_dac(audio, embedding)
|
25 |
+
elif self.model_type == 'encodec':
|
26 |
+
return self.process_encodec(audio, embedding)
|
27 |
+
elif self.model_type == 'stable_vae':
|
28 |
+
return self.process_stable_vae(audio, embedding)
|
29 |
+
else:
|
30 |
+
raise NotImplementedError(f"Model type not implemented: {self.model_type}")
|
31 |
+
|
32 |
+
def process_dac(self, audio=None, embedding=None):
|
33 |
+
if audio is not None:
|
34 |
+
z = self.ae.encoder(audio)
|
35 |
+
if self.quantization_first:
|
36 |
+
z, *_ = self.ae.quantizer(z, None)
|
37 |
+
return z
|
38 |
+
elif embedding is not None:
|
39 |
+
z = embedding
|
40 |
+
if self.quantization_first:
|
41 |
+
audio = self.ae.decoder(z)
|
42 |
+
else:
|
43 |
+
z, *_ = self.ae.quantizer(z, None)
|
44 |
+
audio = self.ae.decoder(z)
|
45 |
+
return audio
|
46 |
+
else:
|
47 |
+
raise ValueError("Either audio or embedding must be provided.")
|
48 |
+
|
49 |
+
def process_encodec(self, audio=None, embedding=None):
|
50 |
+
if audio is not None:
|
51 |
+
z = self.ae.encoder(audio)
|
52 |
+
if self.quantization_first:
|
53 |
+
code = self.ae.quantizer.encode(z)
|
54 |
+
z = self.ae.quantizer.decode(code)
|
55 |
+
return z
|
56 |
+
elif embedding is not None:
|
57 |
+
z = embedding
|
58 |
+
if self.quantization_first:
|
59 |
+
audio = self.ae.decoder(z)
|
60 |
+
else:
|
61 |
+
code = self.ae.quantizer.encode(z)
|
62 |
+
z = self.ae.quantizer.decode(code)
|
63 |
+
audio = self.ae.decoder(z)
|
64 |
+
return audio
|
65 |
+
else:
|
66 |
+
raise ValueError("Either audio or embedding must be provided.")
|
67 |
+
|
68 |
+
def process_stable_vae(self, audio=None, embedding=None):
|
69 |
+
if audio is not None:
|
70 |
+
z = self.ae.encoder(audio)
|
71 |
+
if self.quantization_first:
|
72 |
+
z = self.ae.bottleneck.encode(z)
|
73 |
+
return z
|
74 |
+
if embedding is not None:
|
75 |
+
z = embedding
|
76 |
+
if self.quantization_first:
|
77 |
+
audio = self.ae.decoder(z)
|
78 |
+
else:
|
79 |
+
z = self.ae.bottleneck.encode(z)
|
80 |
+
audio = self.ae.decoder(z)
|
81 |
+
return audio
|
82 |
+
else:
|
83 |
+
raise ValueError("Either audio or embedding must be provided.")
|
src/modules/clap_wrapper.py
ADDED
File without changes
|
src/modules/dac/__init__.py
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
__version__ = "1.0.0"
|
2 |
+
|
3 |
+
# preserved here for legacy reasons
|
4 |
+
__model_version__ = "latest"
|
5 |
+
|
6 |
+
import audiotools
|
7 |
+
|
8 |
+
audiotools.ml.BaseModel.INTERN += ["dac.**"]
|
9 |
+
audiotools.ml.BaseModel.EXTERN += ["einops"]
|
10 |
+
|
11 |
+
|
12 |
+
from . import nn
|
13 |
+
from . import model
|
14 |
+
from . import utils
|
15 |
+
from .model import DAC
|
16 |
+
from .model import DACFile
|
src/modules/dac/__main__.py
ADDED
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import sys
|
2 |
+
|
3 |
+
import argbind
|
4 |
+
|
5 |
+
from dac.utils import download
|
6 |
+
from dac.utils.decode import decode
|
7 |
+
from dac.utils.encode import encode
|
8 |
+
|
9 |
+
STAGES = ["encode", "decode", "download"]
|
10 |
+
|
11 |
+
|
12 |
+
def run(stage: str):
|
13 |
+
"""Run stages.
|
14 |
+
|
15 |
+
Parameters
|
16 |
+
----------
|
17 |
+
stage : str
|
18 |
+
Stage to run
|
19 |
+
"""
|
20 |
+
if stage not in STAGES:
|
21 |
+
raise ValueError(f"Unknown command: {stage}. Allowed commands are {STAGES}")
|
22 |
+
stage_fn = globals()[stage]
|
23 |
+
|
24 |
+
if stage == "download":
|
25 |
+
stage_fn()
|
26 |
+
return
|
27 |
+
|
28 |
+
stage_fn()
|
29 |
+
|
30 |
+
|
31 |
+
if __name__ == "__main__":
|
32 |
+
group = sys.argv.pop(1)
|
33 |
+
args = argbind.parse_args(group=group)
|
34 |
+
|
35 |
+
with argbind.scope(args):
|
36 |
+
run(group)
|
src/modules/dac/compare/__init__.py
ADDED
File without changes
|
src/modules/dac/compare/encodec.py
ADDED
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from audiotools import AudioSignal
|
3 |
+
from audiotools.ml import BaseModel
|
4 |
+
from encodec import EncodecModel
|
5 |
+
|
6 |
+
|
7 |
+
class Encodec(BaseModel):
|
8 |
+
def __init__(self, sample_rate: int = 24000, bandwidth: float = 24.0):
|
9 |
+
super().__init__()
|
10 |
+
|
11 |
+
if sample_rate == 24000:
|
12 |
+
self.model = EncodecModel.encodec_model_24khz()
|
13 |
+
else:
|
14 |
+
self.model = EncodecModel.encodec_model_48khz()
|
15 |
+
self.model.set_target_bandwidth(bandwidth)
|
16 |
+
self.sample_rate = 44100
|
17 |
+
|
18 |
+
def forward(
|
19 |
+
self,
|
20 |
+
audio_data: torch.Tensor,
|
21 |
+
sample_rate: int = 44100,
|
22 |
+
n_quantizers: int = None,
|
23 |
+
):
|
24 |
+
signal = AudioSignal(audio_data, sample_rate)
|
25 |
+
signal.resample(self.model.sample_rate)
|
26 |
+
recons = self.model(signal.audio_data)
|
27 |
+
recons = AudioSignal(recons, self.model.sample_rate)
|
28 |
+
recons.resample(sample_rate)
|
29 |
+
return {"audio": recons.audio_data}
|
30 |
+
|
31 |
+
|
32 |
+
if __name__ == "__main__":
|
33 |
+
import numpy as np
|
34 |
+
from functools import partial
|
35 |
+
|
36 |
+
model = Encodec()
|
37 |
+
|
38 |
+
for n, m in model.named_modules():
|
39 |
+
o = m.extra_repr()
|
40 |
+
p = sum([np.prod(p.size()) for p in m.parameters()])
|
41 |
+
fn = lambda o, p: o + f" {p/1e6:<.3f}M params."
|
42 |
+
setattr(m, "extra_repr", partial(fn, o=o, p=p))
|
43 |
+
print(model)
|
44 |
+
print("Total # of params: ", sum([np.prod(p.size()) for p in model.parameters()]))
|
45 |
+
|
46 |
+
length = 88200 * 2
|
47 |
+
x = torch.randn(1, 1, length).to(model.device)
|
48 |
+
x.requires_grad_(True)
|
49 |
+
x.retain_grad()
|
50 |
+
|
51 |
+
# Make a forward pass
|
52 |
+
out = model(x)["audio"]
|
53 |
+
|
54 |
+
print(x.shape, out.shape)
|
src/modules/dac/model/__init__.py
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .base import CodecMixin
|
2 |
+
from .base import DACFile
|
3 |
+
from .dac import DAC
|
4 |
+
from .discriminator import Discriminator
|
src/modules/dac/model/base.py
ADDED
@@ -0,0 +1,294 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from dataclasses import dataclass
|
3 |
+
from pathlib import Path
|
4 |
+
from typing import Union
|
5 |
+
|
6 |
+
import numpy as np
|
7 |
+
import torch
|
8 |
+
import tqdm
|
9 |
+
from audiotools import AudioSignal
|
10 |
+
from torch import nn
|
11 |
+
|
12 |
+
SUPPORTED_VERSIONS = ["1.0.0"]
|
13 |
+
|
14 |
+
|
15 |
+
@dataclass
|
16 |
+
class DACFile:
|
17 |
+
codes: torch.Tensor
|
18 |
+
|
19 |
+
# Metadata
|
20 |
+
chunk_length: int
|
21 |
+
original_length: int
|
22 |
+
input_db: float
|
23 |
+
channels: int
|
24 |
+
sample_rate: int
|
25 |
+
padding: bool
|
26 |
+
dac_version: str
|
27 |
+
|
28 |
+
def save(self, path):
|
29 |
+
artifacts = {
|
30 |
+
"codes": self.codes.numpy().astype(np.uint16),
|
31 |
+
"metadata": {
|
32 |
+
"input_db": self.input_db.numpy().astype(np.float32),
|
33 |
+
"original_length": self.original_length,
|
34 |
+
"sample_rate": self.sample_rate,
|
35 |
+
"chunk_length": self.chunk_length,
|
36 |
+
"channels": self.channels,
|
37 |
+
"padding": self.padding,
|
38 |
+
"dac_version": SUPPORTED_VERSIONS[-1],
|
39 |
+
},
|
40 |
+
}
|
41 |
+
path = Path(path).with_suffix(".dac")
|
42 |
+
with open(path, "wb") as f:
|
43 |
+
np.save(f, artifacts)
|
44 |
+
return path
|
45 |
+
|
46 |
+
@classmethod
|
47 |
+
def load(cls, path):
|
48 |
+
artifacts = np.load(path, allow_pickle=True)[()]
|
49 |
+
codes = torch.from_numpy(artifacts["codes"].astype(int))
|
50 |
+
if artifacts["metadata"].get("dac_version", None) not in SUPPORTED_VERSIONS:
|
51 |
+
raise RuntimeError(
|
52 |
+
f"Given file {path} can't be loaded with this version of descript-audio-codec."
|
53 |
+
)
|
54 |
+
return cls(codes=codes, **artifacts["metadata"])
|
55 |
+
|
56 |
+
|
57 |
+
class CodecMixin:
|
58 |
+
@property
|
59 |
+
def padding(self):
|
60 |
+
if not hasattr(self, "_padding"):
|
61 |
+
self._padding = True
|
62 |
+
return self._padding
|
63 |
+
|
64 |
+
@padding.setter
|
65 |
+
def padding(self, value):
|
66 |
+
assert isinstance(value, bool)
|
67 |
+
|
68 |
+
layers = [
|
69 |
+
l for l in self.modules() if isinstance(l, (nn.Conv1d, nn.ConvTranspose1d))
|
70 |
+
]
|
71 |
+
|
72 |
+
for layer in layers:
|
73 |
+
if value:
|
74 |
+
if hasattr(layer, "original_padding"):
|
75 |
+
layer.padding = layer.original_padding
|
76 |
+
else:
|
77 |
+
layer.original_padding = layer.padding
|
78 |
+
layer.padding = tuple(0 for _ in range(len(layer.padding)))
|
79 |
+
|
80 |
+
self._padding = value
|
81 |
+
|
82 |
+
def get_delay(self):
|
83 |
+
# Any number works here, delay is invariant to input length
|
84 |
+
l_out = self.get_output_length(0)
|
85 |
+
L = l_out
|
86 |
+
|
87 |
+
layers = []
|
88 |
+
for layer in self.modules():
|
89 |
+
if isinstance(layer, (nn.Conv1d, nn.ConvTranspose1d)):
|
90 |
+
layers.append(layer)
|
91 |
+
|
92 |
+
for layer in reversed(layers):
|
93 |
+
d = layer.dilation[0]
|
94 |
+
k = layer.kernel_size[0]
|
95 |
+
s = layer.stride[0]
|
96 |
+
|
97 |
+
if isinstance(layer, nn.ConvTranspose1d):
|
98 |
+
L = ((L - d * (k - 1) - 1) / s) + 1
|
99 |
+
elif isinstance(layer, nn.Conv1d):
|
100 |
+
L = (L - 1) * s + d * (k - 1) + 1
|
101 |
+
|
102 |
+
L = math.ceil(L)
|
103 |
+
|
104 |
+
l_in = L
|
105 |
+
|
106 |
+
return (l_in - l_out) // 2
|
107 |
+
|
108 |
+
def get_output_length(self, input_length):
|
109 |
+
L = input_length
|
110 |
+
# Calculate output length
|
111 |
+
for layer in self.modules():
|
112 |
+
if isinstance(layer, (nn.Conv1d, nn.ConvTranspose1d)):
|
113 |
+
d = layer.dilation[0]
|
114 |
+
k = layer.kernel_size[0]
|
115 |
+
s = layer.stride[0]
|
116 |
+
|
117 |
+
if isinstance(layer, nn.Conv1d):
|
118 |
+
L = ((L - d * (k - 1) - 1) / s) + 1
|
119 |
+
elif isinstance(layer, nn.ConvTranspose1d):
|
120 |
+
L = (L - 1) * s + d * (k - 1) + 1
|
121 |
+
|
122 |
+
L = math.floor(L)
|
123 |
+
return L
|
124 |
+
|
125 |
+
@torch.no_grad()
|
126 |
+
def compress(
|
127 |
+
self,
|
128 |
+
audio_path_or_signal: Union[str, Path, AudioSignal],
|
129 |
+
win_duration: float = 1.0,
|
130 |
+
verbose: bool = False,
|
131 |
+
normalize_db: float = -16,
|
132 |
+
n_quantizers: int = None,
|
133 |
+
) -> DACFile:
|
134 |
+
"""Processes an audio signal from a file or AudioSignal object into
|
135 |
+
discrete codes. This function processes the signal in short windows,
|
136 |
+
using constant GPU memory.
|
137 |
+
|
138 |
+
Parameters
|
139 |
+
----------
|
140 |
+
audio_path_or_signal : Union[str, Path, AudioSignal]
|
141 |
+
audio signal to reconstruct
|
142 |
+
win_duration : float, optional
|
143 |
+
window duration in seconds, by default 5.0
|
144 |
+
verbose : bool, optional
|
145 |
+
by default False
|
146 |
+
normalize_db : float, optional
|
147 |
+
normalize db, by default -16
|
148 |
+
|
149 |
+
Returns
|
150 |
+
-------
|
151 |
+
DACFile
|
152 |
+
Object containing compressed codes and metadata
|
153 |
+
required for decompression
|
154 |
+
"""
|
155 |
+
audio_signal = audio_path_or_signal
|
156 |
+
if isinstance(audio_signal, (str, Path)):
|
157 |
+
audio_signal = AudioSignal.load_from_file_with_ffmpeg(str(audio_signal))
|
158 |
+
|
159 |
+
self.eval()
|
160 |
+
original_padding = self.padding
|
161 |
+
original_device = audio_signal.device
|
162 |
+
|
163 |
+
audio_signal = audio_signal.clone()
|
164 |
+
original_sr = audio_signal.sample_rate
|
165 |
+
|
166 |
+
resample_fn = audio_signal.resample
|
167 |
+
loudness_fn = audio_signal.loudness
|
168 |
+
|
169 |
+
# If audio is > 10 minutes long, use the ffmpeg versions
|
170 |
+
if audio_signal.signal_duration >= 10 * 60 * 60:
|
171 |
+
resample_fn = audio_signal.ffmpeg_resample
|
172 |
+
loudness_fn = audio_signal.ffmpeg_loudness
|
173 |
+
|
174 |
+
original_length = audio_signal.signal_length
|
175 |
+
resample_fn(self.sample_rate)
|
176 |
+
input_db = loudness_fn()
|
177 |
+
|
178 |
+
if normalize_db is not None:
|
179 |
+
audio_signal.normalize(normalize_db)
|
180 |
+
audio_signal.ensure_max_of_audio()
|
181 |
+
|
182 |
+
nb, nac, nt = audio_signal.audio_data.shape
|
183 |
+
audio_signal.audio_data = audio_signal.audio_data.reshape(nb * nac, 1, nt)
|
184 |
+
win_duration = (
|
185 |
+
audio_signal.signal_duration if win_duration is None else win_duration
|
186 |
+
)
|
187 |
+
|
188 |
+
if audio_signal.signal_duration <= win_duration:
|
189 |
+
# Unchunked compression (used if signal length < win duration)
|
190 |
+
self.padding = True
|
191 |
+
n_samples = nt
|
192 |
+
hop = nt
|
193 |
+
else:
|
194 |
+
# Chunked inference
|
195 |
+
self.padding = False
|
196 |
+
# Zero-pad signal on either side by the delay
|
197 |
+
audio_signal.zero_pad(self.delay, self.delay)
|
198 |
+
n_samples = int(win_duration * self.sample_rate)
|
199 |
+
# Round n_samples to nearest hop length multiple
|
200 |
+
n_samples = int(math.ceil(n_samples / self.hop_length) * self.hop_length)
|
201 |
+
hop = self.get_output_length(n_samples)
|
202 |
+
|
203 |
+
codes = []
|
204 |
+
range_fn = range if not verbose else tqdm.trange
|
205 |
+
|
206 |
+
for i in range_fn(0, nt, hop):
|
207 |
+
x = audio_signal[..., i : i + n_samples]
|
208 |
+
x = x.zero_pad(0, max(0, n_samples - x.shape[-1]))
|
209 |
+
|
210 |
+
audio_data = x.audio_data.to(self.device)
|
211 |
+
audio_data = self.preprocess(audio_data, self.sample_rate)
|
212 |
+
_, c, _, _, _ = self.encode(audio_data, n_quantizers)
|
213 |
+
codes.append(c.to(original_device))
|
214 |
+
chunk_length = c.shape[-1]
|
215 |
+
|
216 |
+
codes = torch.cat(codes, dim=-1)
|
217 |
+
|
218 |
+
dac_file = DACFile(
|
219 |
+
codes=codes,
|
220 |
+
chunk_length=chunk_length,
|
221 |
+
original_length=original_length,
|
222 |
+
input_db=input_db,
|
223 |
+
channels=nac,
|
224 |
+
sample_rate=original_sr,
|
225 |
+
padding=self.padding,
|
226 |
+
dac_version=SUPPORTED_VERSIONS[-1],
|
227 |
+
)
|
228 |
+
|
229 |
+
if n_quantizers is not None:
|
230 |
+
codes = codes[:, :n_quantizers, :]
|
231 |
+
|
232 |
+
self.padding = original_padding
|
233 |
+
return dac_file
|
234 |
+
|
235 |
+
@torch.no_grad()
|
236 |
+
def decompress(
|
237 |
+
self,
|
238 |
+
obj: Union[str, Path, DACFile],
|
239 |
+
verbose: bool = False,
|
240 |
+
) -> AudioSignal:
|
241 |
+
"""Reconstruct audio from a given .dac file
|
242 |
+
|
243 |
+
Parameters
|
244 |
+
----------
|
245 |
+
obj : Union[str, Path, DACFile]
|
246 |
+
.dac file location or corresponding DACFile object.
|
247 |
+
verbose : bool, optional
|
248 |
+
Prints progress if True, by default False
|
249 |
+
|
250 |
+
Returns
|
251 |
+
-------
|
252 |
+
AudioSignal
|
253 |
+
Object with the reconstructed audio
|
254 |
+
"""
|
255 |
+
self.eval()
|
256 |
+
if isinstance(obj, (str, Path)):
|
257 |
+
obj = DACFile.load(obj)
|
258 |
+
|
259 |
+
original_padding = self.padding
|
260 |
+
self.padding = obj.padding
|
261 |
+
|
262 |
+
range_fn = range if not verbose else tqdm.trange
|
263 |
+
codes = obj.codes
|
264 |
+
original_device = codes.device
|
265 |
+
chunk_length = obj.chunk_length
|
266 |
+
recons = []
|
267 |
+
|
268 |
+
for i in range_fn(0, codes.shape[-1], chunk_length):
|
269 |
+
c = codes[..., i : i + chunk_length].to(self.device)
|
270 |
+
z = self.quantizer.from_codes(c)[0]
|
271 |
+
r = self.decode(z)
|
272 |
+
recons.append(r.to(original_device))
|
273 |
+
|
274 |
+
recons = torch.cat(recons, dim=-1)
|
275 |
+
recons = AudioSignal(recons, self.sample_rate)
|
276 |
+
|
277 |
+
resample_fn = recons.resample
|
278 |
+
loudness_fn = recons.loudness
|
279 |
+
|
280 |
+
# If audio is > 10 minutes long, use the ffmpeg versions
|
281 |
+
if recons.signal_duration >= 10 * 60 * 60:
|
282 |
+
resample_fn = recons.ffmpeg_resample
|
283 |
+
loudness_fn = recons.ffmpeg_loudness
|
284 |
+
|
285 |
+
recons.normalize(obj.input_db)
|
286 |
+
resample_fn(obj.sample_rate)
|
287 |
+
recons = recons[..., : obj.original_length]
|
288 |
+
loudness_fn()
|
289 |
+
recons.audio_data = recons.audio_data.reshape(
|
290 |
+
-1, obj.channels, obj.original_length
|
291 |
+
)
|
292 |
+
|
293 |
+
self.padding = original_padding
|
294 |
+
return recons
|
src/modules/dac/model/dac.py
ADDED
@@ -0,0 +1,364 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from typing import List
|
3 |
+
from typing import Union
|
4 |
+
|
5 |
+
import numpy as np
|
6 |
+
import torch
|
7 |
+
from audiotools import AudioSignal
|
8 |
+
from audiotools.ml import BaseModel
|
9 |
+
from torch import nn
|
10 |
+
|
11 |
+
from .base import CodecMixin
|
12 |
+
from ..nn.layers import Snake1d
|
13 |
+
from ..nn.layers import WNConv1d
|
14 |
+
from ..nn.layers import WNConvTranspose1d
|
15 |
+
from ..nn.quantize import ResidualVectorQuantize
|
16 |
+
|
17 |
+
|
18 |
+
def init_weights(m):
|
19 |
+
if isinstance(m, nn.Conv1d):
|
20 |
+
nn.init.trunc_normal_(m.weight, std=0.02)
|
21 |
+
nn.init.constant_(m.bias, 0)
|
22 |
+
|
23 |
+
|
24 |
+
class ResidualUnit(nn.Module):
|
25 |
+
def __init__(self, dim: int = 16, dilation: int = 1):
|
26 |
+
super().__init__()
|
27 |
+
pad = ((7 - 1) * dilation) // 2
|
28 |
+
self.block = nn.Sequential(
|
29 |
+
Snake1d(dim),
|
30 |
+
WNConv1d(dim, dim, kernel_size=7, dilation=dilation, padding=pad),
|
31 |
+
Snake1d(dim),
|
32 |
+
WNConv1d(dim, dim, kernel_size=1),
|
33 |
+
)
|
34 |
+
|
35 |
+
def forward(self, x):
|
36 |
+
y = self.block(x)
|
37 |
+
pad = (x.shape[-1] - y.shape[-1]) // 2
|
38 |
+
if pad > 0:
|
39 |
+
x = x[..., pad:-pad]
|
40 |
+
return x + y
|
41 |
+
|
42 |
+
|
43 |
+
class EncoderBlock(nn.Module):
|
44 |
+
def __init__(self, dim: int = 16, stride: int = 1):
|
45 |
+
super().__init__()
|
46 |
+
self.block = nn.Sequential(
|
47 |
+
ResidualUnit(dim // 2, dilation=1),
|
48 |
+
ResidualUnit(dim // 2, dilation=3),
|
49 |
+
ResidualUnit(dim // 2, dilation=9),
|
50 |
+
Snake1d(dim // 2),
|
51 |
+
WNConv1d(
|
52 |
+
dim // 2,
|
53 |
+
dim,
|
54 |
+
kernel_size=2 * stride,
|
55 |
+
stride=stride,
|
56 |
+
padding=math.ceil(stride / 2),
|
57 |
+
),
|
58 |
+
)
|
59 |
+
|
60 |
+
def forward(self, x):
|
61 |
+
return self.block(x)
|
62 |
+
|
63 |
+
|
64 |
+
class Encoder(nn.Module):
|
65 |
+
def __init__(
|
66 |
+
self,
|
67 |
+
d_model: int = 64,
|
68 |
+
strides: list = [2, 4, 8, 8],
|
69 |
+
d_latent: int = 64,
|
70 |
+
):
|
71 |
+
super().__init__()
|
72 |
+
# Create first convolution
|
73 |
+
self.block = [WNConv1d(1, d_model, kernel_size=7, padding=3)]
|
74 |
+
|
75 |
+
# Create EncoderBlocks that double channels as they downsample by `stride`
|
76 |
+
for stride in strides:
|
77 |
+
d_model *= 2
|
78 |
+
self.block += [EncoderBlock(d_model, stride=stride)]
|
79 |
+
|
80 |
+
# Create last convolution
|
81 |
+
self.block += [
|
82 |
+
Snake1d(d_model),
|
83 |
+
WNConv1d(d_model, d_latent, kernel_size=3, padding=1),
|
84 |
+
]
|
85 |
+
|
86 |
+
# Wrap black into nn.Sequential
|
87 |
+
self.block = nn.Sequential(*self.block)
|
88 |
+
self.enc_dim = d_model
|
89 |
+
|
90 |
+
def forward(self, x):
|
91 |
+
return self.block(x)
|
92 |
+
|
93 |
+
|
94 |
+
class DecoderBlock(nn.Module):
|
95 |
+
def __init__(self, input_dim: int = 16, output_dim: int = 8, stride: int = 1):
|
96 |
+
super().__init__()
|
97 |
+
self.block = nn.Sequential(
|
98 |
+
Snake1d(input_dim),
|
99 |
+
WNConvTranspose1d(
|
100 |
+
input_dim,
|
101 |
+
output_dim,
|
102 |
+
kernel_size=2 * stride,
|
103 |
+
stride=stride,
|
104 |
+
padding=math.ceil(stride / 2),
|
105 |
+
),
|
106 |
+
ResidualUnit(output_dim, dilation=1),
|
107 |
+
ResidualUnit(output_dim, dilation=3),
|
108 |
+
ResidualUnit(output_dim, dilation=9),
|
109 |
+
)
|
110 |
+
|
111 |
+
def forward(self, x):
|
112 |
+
return self.block(x)
|
113 |
+
|
114 |
+
|
115 |
+
class Decoder(nn.Module):
|
116 |
+
def __init__(
|
117 |
+
self,
|
118 |
+
input_channel,
|
119 |
+
channels,
|
120 |
+
rates,
|
121 |
+
d_out: int = 1,
|
122 |
+
):
|
123 |
+
super().__init__()
|
124 |
+
|
125 |
+
# Add first conv layer
|
126 |
+
layers = [WNConv1d(input_channel, channels, kernel_size=7, padding=3)]
|
127 |
+
|
128 |
+
# Add upsampling + MRF blocks
|
129 |
+
for i, stride in enumerate(rates):
|
130 |
+
input_dim = channels // 2**i
|
131 |
+
output_dim = channels // 2 ** (i + 1)
|
132 |
+
layers += [DecoderBlock(input_dim, output_dim, stride)]
|
133 |
+
|
134 |
+
# Add final conv layer
|
135 |
+
layers += [
|
136 |
+
Snake1d(output_dim),
|
137 |
+
WNConv1d(output_dim, d_out, kernel_size=7, padding=3),
|
138 |
+
nn.Tanh(),
|
139 |
+
]
|
140 |
+
|
141 |
+
self.model = nn.Sequential(*layers)
|
142 |
+
|
143 |
+
def forward(self, x):
|
144 |
+
return self.model(x)
|
145 |
+
|
146 |
+
|
147 |
+
class DAC(BaseModel, CodecMixin):
|
148 |
+
def __init__(
|
149 |
+
self,
|
150 |
+
encoder_dim: int = 64,
|
151 |
+
encoder_rates: List[int] = [2, 4, 8, 8],
|
152 |
+
latent_dim: int = None,
|
153 |
+
decoder_dim: int = 1536,
|
154 |
+
decoder_rates: List[int] = [8, 8, 4, 2],
|
155 |
+
n_codebooks: int = 9,
|
156 |
+
codebook_size: int = 1024,
|
157 |
+
codebook_dim: Union[int, list] = 8,
|
158 |
+
quantizer_dropout: bool = False,
|
159 |
+
sample_rate: int = 44100,
|
160 |
+
):
|
161 |
+
super().__init__()
|
162 |
+
|
163 |
+
self.encoder_dim = encoder_dim
|
164 |
+
self.encoder_rates = encoder_rates
|
165 |
+
self.decoder_dim = decoder_dim
|
166 |
+
self.decoder_rates = decoder_rates
|
167 |
+
self.sample_rate = sample_rate
|
168 |
+
|
169 |
+
if latent_dim is None:
|
170 |
+
latent_dim = encoder_dim * (2 ** len(encoder_rates))
|
171 |
+
|
172 |
+
self.latent_dim = latent_dim
|
173 |
+
|
174 |
+
self.hop_length = np.prod(encoder_rates)
|
175 |
+
self.encoder = Encoder(encoder_dim, encoder_rates, latent_dim)
|
176 |
+
|
177 |
+
self.n_codebooks = n_codebooks
|
178 |
+
self.codebook_size = codebook_size
|
179 |
+
self.codebook_dim = codebook_dim
|
180 |
+
self.quantizer = ResidualVectorQuantize(
|
181 |
+
input_dim=latent_dim,
|
182 |
+
n_codebooks=n_codebooks,
|
183 |
+
codebook_size=codebook_size,
|
184 |
+
codebook_dim=codebook_dim,
|
185 |
+
quantizer_dropout=quantizer_dropout,
|
186 |
+
)
|
187 |
+
|
188 |
+
self.decoder = Decoder(
|
189 |
+
latent_dim,
|
190 |
+
decoder_dim,
|
191 |
+
decoder_rates,
|
192 |
+
)
|
193 |
+
self.sample_rate = sample_rate
|
194 |
+
self.apply(init_weights)
|
195 |
+
|
196 |
+
self.delay = self.get_delay()
|
197 |
+
|
198 |
+
def preprocess(self, audio_data, sample_rate):
|
199 |
+
if sample_rate is None:
|
200 |
+
sample_rate = self.sample_rate
|
201 |
+
assert sample_rate == self.sample_rate
|
202 |
+
|
203 |
+
length = audio_data.shape[-1]
|
204 |
+
right_pad = math.ceil(length / self.hop_length) * self.hop_length - length
|
205 |
+
audio_data = nn.functional.pad(audio_data, (0, right_pad))
|
206 |
+
|
207 |
+
return audio_data
|
208 |
+
|
209 |
+
def encode(
|
210 |
+
self,
|
211 |
+
audio_data: torch.Tensor,
|
212 |
+
n_quantizers: int = None,
|
213 |
+
):
|
214 |
+
"""Encode given audio data and return quantized latent codes
|
215 |
+
|
216 |
+
Parameters
|
217 |
+
----------
|
218 |
+
audio_data : Tensor[B x 1 x T]
|
219 |
+
Audio data to encode
|
220 |
+
n_quantizers : int, optional
|
221 |
+
Number of quantizers to use, by default None
|
222 |
+
If None, all quantizers are used.
|
223 |
+
|
224 |
+
Returns
|
225 |
+
-------
|
226 |
+
dict
|
227 |
+
A dictionary with the following keys:
|
228 |
+
"z" : Tensor[B x D x T]
|
229 |
+
Quantized continuous representation of input
|
230 |
+
"codes" : Tensor[B x N x T]
|
231 |
+
Codebook indices for each codebook
|
232 |
+
(quantized discrete representation of input)
|
233 |
+
"latents" : Tensor[B x N*D x T]
|
234 |
+
Projected latents (continuous representation of input before quantization)
|
235 |
+
"vq/commitment_loss" : Tensor[1]
|
236 |
+
Commitment loss to train encoder to predict vectors closer to codebook
|
237 |
+
entries
|
238 |
+
"vq/codebook_loss" : Tensor[1]
|
239 |
+
Codebook loss to update the codebook
|
240 |
+
"length" : int
|
241 |
+
Number of samples in input audio
|
242 |
+
"""
|
243 |
+
z = self.encoder(audio_data)
|
244 |
+
z, codes, latents, commitment_loss, codebook_loss = self.quantizer(
|
245 |
+
z, n_quantizers
|
246 |
+
)
|
247 |
+
return z, codes, latents, commitment_loss, codebook_loss
|
248 |
+
|
249 |
+
def decode(self, z: torch.Tensor):
|
250 |
+
"""Decode given latent codes and return audio data
|
251 |
+
|
252 |
+
Parameters
|
253 |
+
----------
|
254 |
+
z : Tensor[B x D x T]
|
255 |
+
Quantized continuous representation of input
|
256 |
+
length : int, optional
|
257 |
+
Number of samples in output audio, by default None
|
258 |
+
|
259 |
+
Returns
|
260 |
+
-------
|
261 |
+
dict
|
262 |
+
A dictionary with the following keys:
|
263 |
+
"audio" : Tensor[B x 1 x length]
|
264 |
+
Decoded audio data.
|
265 |
+
"""
|
266 |
+
return self.decoder(z)
|
267 |
+
|
268 |
+
def forward(
|
269 |
+
self,
|
270 |
+
audio_data: torch.Tensor,
|
271 |
+
sample_rate: int = None,
|
272 |
+
n_quantizers: int = None,
|
273 |
+
):
|
274 |
+
"""Model forward pass
|
275 |
+
|
276 |
+
Parameters
|
277 |
+
----------
|
278 |
+
audio_data : Tensor[B x 1 x T]
|
279 |
+
Audio data to encode
|
280 |
+
sample_rate : int, optional
|
281 |
+
Sample rate of audio data in Hz, by default None
|
282 |
+
If None, defaults to `self.sample_rate`
|
283 |
+
n_quantizers : int, optional
|
284 |
+
Number of quantizers to use, by default None.
|
285 |
+
If None, all quantizers are used.
|
286 |
+
|
287 |
+
Returns
|
288 |
+
-------
|
289 |
+
dict
|
290 |
+
A dictionary with the following keys:
|
291 |
+
"z" : Tensor[B x D x T]
|
292 |
+
Quantized continuous representation of input
|
293 |
+
"codes" : Tensor[B x N x T]
|
294 |
+
Codebook indices for each codebook
|
295 |
+
(quantized discrete representation of input)
|
296 |
+
"latents" : Tensor[B x N*D x T]
|
297 |
+
Projected latents (continuous representation of input before quantization)
|
298 |
+
"vq/commitment_loss" : Tensor[1]
|
299 |
+
Commitment loss to train encoder to predict vectors closer to codebook
|
300 |
+
entries
|
301 |
+
"vq/codebook_loss" : Tensor[1]
|
302 |
+
Codebook loss to update the codebook
|
303 |
+
"length" : int
|
304 |
+
Number of samples in input audio
|
305 |
+
"audio" : Tensor[B x 1 x length]
|
306 |
+
Decoded audio data.
|
307 |
+
"""
|
308 |
+
length = audio_data.shape[-1]
|
309 |
+
audio_data = self.preprocess(audio_data, sample_rate)
|
310 |
+
z, codes, latents, commitment_loss, codebook_loss = self.encode(
|
311 |
+
audio_data, n_quantizers
|
312 |
+
)
|
313 |
+
|
314 |
+
x = self.decode(z)
|
315 |
+
return {
|
316 |
+
"audio": x[..., :length],
|
317 |
+
"z": z,
|
318 |
+
"codes": codes,
|
319 |
+
"latents": latents,
|
320 |
+
"vq/commitment_loss": commitment_loss,
|
321 |
+
"vq/codebook_loss": codebook_loss,
|
322 |
+
}
|
323 |
+
|
324 |
+
|
325 |
+
if __name__ == "__main__":
|
326 |
+
import numpy as np
|
327 |
+
from functools import partial
|
328 |
+
|
329 |
+
model = DAC().to("cpu")
|
330 |
+
|
331 |
+
for n, m in model.named_modules():
|
332 |
+
o = m.extra_repr()
|
333 |
+
p = sum([np.prod(p.size()) for p in m.parameters()])
|
334 |
+
fn = lambda o, p: o + f" {p/1e6:<.3f}M params."
|
335 |
+
setattr(m, "extra_repr", partial(fn, o=o, p=p))
|
336 |
+
print(model)
|
337 |
+
print("Total # of params: ", sum([np.prod(p.size()) for p in model.parameters()]))
|
338 |
+
|
339 |
+
length = 88200 * 2
|
340 |
+
x = torch.randn(1, 1, length).to(model.device)
|
341 |
+
x.requires_grad_(True)
|
342 |
+
x.retain_grad()
|
343 |
+
|
344 |
+
# Make a forward pass
|
345 |
+
out = model(x)["audio"]
|
346 |
+
print("Input shape:", x.shape)
|
347 |
+
print("Output shape:", out.shape)
|
348 |
+
|
349 |
+
# Create gradient variable
|
350 |
+
grad = torch.zeros_like(out)
|
351 |
+
grad[:, :, grad.shape[-1] // 2] = 1
|
352 |
+
|
353 |
+
# Make a backward pass
|
354 |
+
out.backward(grad)
|
355 |
+
|
356 |
+
# Check non-zero values
|
357 |
+
gradmap = x.grad.squeeze(0)
|
358 |
+
gradmap = (gradmap != 0).sum(0) # sum across features
|
359 |
+
rf = (gradmap != 0).sum()
|
360 |
+
|
361 |
+
print(f"Receptive field: {rf.item()}")
|
362 |
+
|
363 |
+
x = AudioSignal(torch.randn(1, 1, 44100 * 60), 44100)
|
364 |
+
model.decompress(model.compress(x, verbose=True), verbose=True)
|