Jackflack09 YeOldHermit commited on
Commit
522606a
0 Parent(s):

Duplicate from YeOldHermit/Super-Resolution-Anime-Diffusion

Browse files

Co-authored-by: Elvin John Maat <YeOldHermit@users.noreply.huggingface.co>

This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +34 -0
  2. .gitignore +162 -0
  3. README.md +108 -0
  4. README_HG.md +12 -0
  5. Waifu2x/.gitattributes +1 -0
  6. Waifu2x/.gitignore +4 -0
  7. Waifu2x/Common.py +189 -0
  8. Waifu2x/Dataloader.py +215 -0
  9. Waifu2x/Img_to_Sqlite.py +115 -0
  10. Waifu2x/LICENSE +674 -0
  11. Waifu2x/Loss.py +44 -0
  12. Waifu2x/Models.py +316 -0
  13. Waifu2x/Readme.md +167 -0
  14. Waifu2x/__init__.py +9 -0
  15. Waifu2x/magnify.py +86 -0
  16. Waifu2x/model_check_points/CRAN_V2/CARN_adam_checkpoint.pt +3 -0
  17. Waifu2x/model_check_points/CRAN_V2/CARN_model_checkpoint.pt +3 -0
  18. Waifu2x/model_check_points/CRAN_V2/CARN_scheduler_last_iter.pt +3 -0
  19. Waifu2x/model_check_points/CRAN_V2/CRAN_V2_02_28_2019.pt +3 -0
  20. Waifu2x/model_check_points/CRAN_V2/ReadME.md +41 -0
  21. Waifu2x/model_check_points/CRAN_V2/test_loss.pt +3 -0
  22. Waifu2x/model_check_points/CRAN_V2/test_psnr.pt +3 -0
  23. Waifu2x/model_check_points/CRAN_V2/test_ssim.pt +3 -0
  24. Waifu2x/model_check_points/CRAN_V2/train_loss.pt +3 -0
  25. Waifu2x/model_check_points/CRAN_V2/train_psnr.pt +3 -0
  26. Waifu2x/model_check_points/CRAN_V2/train_ssim.pt +3 -0
  27. Waifu2x/train.py +174 -0
  28. Waifu2x/utils/Img_to_H5.py +50 -0
  29. Waifu2x/utils/__init__.py +8 -0
  30. Waifu2x/utils/cls.py +157 -0
  31. Waifu2x/utils/image_quality.py +173 -0
  32. Waifu2x/utils/prepare_images.py +120 -0
  33. app.py +344 -0
  34. diffusers/__init__.py +123 -0
  35. diffusers/commands/__init__.py +27 -0
  36. diffusers/commands/diffusers_cli.py +41 -0
  37. diffusers/commands/env.py +70 -0
  38. diffusers/configuration_utils.py +613 -0
  39. diffusers/dependency_versions_check.py +47 -0
  40. diffusers/dependency_versions_table.py +33 -0
  41. diffusers/dynamic_modules_utils.py +428 -0
  42. diffusers/experimental/README.md +5 -0
  43. diffusers/experimental/__init__.py +1 -0
  44. diffusers/experimental/rl/__init__.py +1 -0
  45. diffusers/experimental/rl/value_guided_sampling.py +130 -0
  46. diffusers/hub_utils.py +130 -0
  47. diffusers/modeling_flax_pytorch_utils.py +117 -0
  48. diffusers/modeling_flax_utils.py +526 -0
  49. diffusers/modeling_utils.py +764 -0
  50. diffusers/models/README.md +3 -0
.gitattributes ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ *.7z filter=lfs diff=lfs merge=lfs -text
2
+ *.arrow filter=lfs diff=lfs merge=lfs -text
3
+ *.bin filter=lfs diff=lfs merge=lfs -text
4
+ *.bz2 filter=lfs diff=lfs merge=lfs -text
5
+ *.ckpt filter=lfs diff=lfs merge=lfs -text
6
+ *.ftz filter=lfs diff=lfs merge=lfs -text
7
+ *.gz filter=lfs diff=lfs merge=lfs -text
8
+ *.h5 filter=lfs diff=lfs merge=lfs -text
9
+ *.joblib filter=lfs diff=lfs merge=lfs -text
10
+ *.lfs.* filter=lfs diff=lfs merge=lfs -text
11
+ *.mlmodel filter=lfs diff=lfs merge=lfs -text
12
+ *.model filter=lfs diff=lfs merge=lfs -text
13
+ *.msgpack filter=lfs diff=lfs merge=lfs -text
14
+ *.npy filter=lfs diff=lfs merge=lfs -text
15
+ *.npz filter=lfs diff=lfs merge=lfs -text
16
+ *.onnx filter=lfs diff=lfs merge=lfs -text
17
+ *.ot filter=lfs diff=lfs merge=lfs -text
18
+ *.parquet filter=lfs diff=lfs merge=lfs -text
19
+ *.pb filter=lfs diff=lfs merge=lfs -text
20
+ *.pickle filter=lfs diff=lfs merge=lfs -text
21
+ *.pkl filter=lfs diff=lfs merge=lfs -text
22
+ *.pt filter=lfs diff=lfs merge=lfs -text
23
+ *.pth filter=lfs diff=lfs merge=lfs -text
24
+ *.rar filter=lfs diff=lfs merge=lfs -text
25
+ *.safetensors filter=lfs diff=lfs merge=lfs -text
26
+ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
27
+ *.tar.* filter=lfs diff=lfs merge=lfs -text
28
+ *.tflite filter=lfs diff=lfs merge=lfs -text
29
+ *.tgz filter=lfs diff=lfs merge=lfs -text
30
+ *.wasm filter=lfs diff=lfs merge=lfs -text
31
+ *.xz filter=lfs diff=lfs merge=lfs -text
32
+ *.zip filter=lfs diff=lfs merge=lfs -text
33
+ *.zst filter=lfs diff=lfs merge=lfs -text
34
+ *tfevents* filter=lfs diff=lfs merge=lfs -text
.gitignore ADDED
@@ -0,0 +1,162 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # dev files
2
+ *.cache
3
+ *.dev.py
4
+ *.mv
5
+ state_dict/
6
+ integrated_datasets/
7
+ *.results
8
+ *.tokenizer
9
+ *.model
10
+ *.state_dict
11
+ *.config
12
+ *.args
13
+ *.zip
14
+ *.gz
15
+ *.bin
16
+ *.result.txt
17
+ *.DS_Store
18
+ *.tmp
19
+ *.args.txt
20
+ *.summary.txt
21
+ *.dat
22
+ *.graph
23
+ # Byte-compiled / optimized / DLL files
24
+ __pycache__/
25
+ *.py[cod]
26
+ *$py.class
27
+ *.pyc
28
+ experiments/
29
+ tests/
30
+ *.result.json
31
+ .idea/
32
+ imgs/
33
+
34
+ # Embedding
35
+ glove.840B.300d.txt
36
+ glove.42B.300d.txt
37
+ glove.twitter.27B.txt
38
+
39
+ # project main files
40
+ release_note.json
41
+
42
+ # C extensions
43
+ *.so
44
+
45
+ # Distribution / packaging
46
+ .Python
47
+ build/
48
+ develop-eggs/
49
+ dist/
50
+ downloads/
51
+ eggs/
52
+ .eggs/
53
+ lib64/
54
+ parts/
55
+ sdist/
56
+ var/
57
+ wheels/
58
+ pip-wheel-metadata/
59
+ share/python-wheels/
60
+ *.egg-info/
61
+ .installed.cfg
62
+ *.egg
63
+ MANIFEST
64
+
65
+ # PyInstaller
66
+ # Usually these files are written by a python script from a template
67
+ # before PyInstaller builds the exe, so as to inject date/other infos into it.
68
+ *.manifest
69
+ *.spec
70
+
71
+ # Installer training_logs
72
+ pip-log.txt
73
+ pip-delete-this-directory.txt
74
+
75
+ # Unit test / coverage reports
76
+ htmlcov/
77
+ .tox/
78
+ .nox/
79
+ .coverage
80
+ .coverage.*
81
+ .cache
82
+ nosetests.xml
83
+ coverage.xml
84
+ *.cover
85
+ *.py,cover
86
+ .hypothesis/
87
+ .pytest_cache/
88
+
89
+ # Translations
90
+ *.mo
91
+ *.pot
92
+
93
+ # Django stuff:
94
+ *.log
95
+ local_settings.py
96
+ db.sqlite3
97
+ db.sqlite3-journal
98
+
99
+ # Flask stuff:
100
+ instance/
101
+ .webassets-cache
102
+
103
+ # Scrapy stuff:
104
+ .scrapy
105
+
106
+ # Sphinx documentation
107
+ docs/_build/
108
+
109
+ # PyBuilder
110
+ target/
111
+
112
+ # Jupyter Notebook
113
+ .ipynb_checkpoints
114
+
115
+ # IPython
116
+ profile_default/
117
+ ipython_config.py
118
+
119
+ # pyenv
120
+ .python-version
121
+
122
+ # pipenv
123
+ # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
124
+ # However, in case of collaboration, if having platform-specific dependencies or dependencies
125
+ # having no cross-platform support, pipenv may install dependencies that don't work, or not
126
+ # install all needed dependencies.
127
+ #Pipfile.lock
128
+
129
+ # celery beat schedule file
130
+ celerybeat-schedule
131
+
132
+ # SageMath parsed files
133
+ *.sage.py
134
+
135
+ # Environments
136
+ .env
137
+ .venv
138
+ env/
139
+ venv/
140
+ ENV/
141
+ env.bak/
142
+ venv.bak/
143
+
144
+ # Spyder project settings
145
+ .spyderproject
146
+ .spyproject
147
+
148
+ # Rope project settings
149
+ .ropeproject
150
+
151
+ # mkdocs documentation
152
+ /site
153
+
154
+ # mypy
155
+ .mypy_cache/
156
+ .dmypy.json
157
+ dmypy.json
158
+
159
+ # Pyre type checker
160
+ .pyre/
161
+ .DS_Store
162
+ examples/.DS_Store
README.md ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ title: Super Resolution Anime Diffusion
3
+ emoji: 📊
4
+ colorFrom: yellow
5
+ colorTo: green
6
+ sdk: gradio
7
+ sdk_version: 3.12.0
8
+ app_file: app.py
9
+ pinned: false
10
+ license: mit
11
+ duplicated_from: YeOldHermit/Super-Resolution-Anime-Diffusion
12
+ ---
13
+
14
+ Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
15
+
16
+ # Super Resolution Anime Diffusion
17
+ This is demo forked from https://huggingface.co/Linaqruf/anything-v3.0.
18
+
19
+ ## Super Resolution Anime Diffusion
20
+ At this moment, many diffusion models can only generate <1024 width and length pictures.
21
+ I integrated the Super Resolution with [Anything diffusion model](https://huggingface.co/Linaqruf/anything-v3.0) to produce high resolution pictures.
22
+ Thanks to the open-source project: https://github.com/yu45020/Waifu2x
23
+
24
+ ## Modifications
25
+ 1. Disable the safety checker to save time and memory. You need to abide the original rules of the model.
26
+ 2. Add the Super Resolution function to the model.
27
+ 3. Add batch generation function to the model (see inference.py).
28
+ 4.
29
+ # Origin README
30
+ ---
31
+ language:
32
+ - en
33
+ license: creativeml-openrail-m
34
+ tags:
35
+ - stable-diffusion
36
+ - stable-diffusion-diffusers
37
+ - text-to-image
38
+ - diffusers
39
+ inference: true
40
+ ---
41
+
42
+ # Anything V3
43
+
44
+ Welcome to Anything V3 - a latent diffusion model for weebs. This model is intended to produce high-quality, highly detailed anime style with just a few prompts. Like other anime-style Stable Diffusion models, it also supports danbooru tags to generate images.
45
+
46
+ e.g. **_1girl, white hair, golden eyes, beautiful eyes, detail, flower meadow, cumulonimbus clouds, lighting, detailed sky, garden_**
47
+
48
+ ## Gradio
49
+
50
+ We support a [Gradio](https://github.com/gradio-app/gradio) Web UI to run Anything-V3.0:
51
+
52
+ [Open in Spaces](https://huggingface.co/spaces/akhaliq/anything-v3.0)
53
+
54
+
55
+
56
+ ## 🧨 Diffusers
57
+
58
+ This model can be used just like any other Stable Diffusion model. For more information,
59
+ please have a look at the [Stable Diffusion](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion).
60
+
61
+ You can also export the model to [ONNX](https://huggingface.co/docs/diffusers/optimization/onnx), [MPS](https://huggingface.co/docs/diffusers/optimization/mps) and/or [FLAX/JAX]().
62
+
63
+ ```python
64
+ from diffusers import StableDiffusionPipeline
65
+ import torch
66
+
67
+ model_id = "Linaqruf/anything-v3.0"
68
+ pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
69
+ pipe = pipe.to("cuda")
70
+
71
+ prompt = "pikachu"
72
+ image = pipe(prompt).images[0]
73
+
74
+ image.save("./pikachu.png")
75
+ ```
76
+
77
+ ## Examples
78
+
79
+ Below are some examples of images generated using this model:
80
+
81
+ **Anime Girl:**
82
+ ![Anime Girl](https://huggingface.co/Linaqruf/anything-v3.0/resolve/main/1girl.png)
83
+ ```
84
+ 1girl, brown hair, green eyes, colorful, autumn, cumulonimbus clouds, lighting, blue sky, falling leaves, garden
85
+ Steps: 50, Sampler: DDIM, CFG scale: 12
86
+ ```
87
+ **Anime Boy:**
88
+ ![Anime Boy](https://huggingface.co/Linaqruf/anything-v3.0/resolve/main/1boy.png)
89
+ ```
90
+ 1boy, medium hair, blonde hair, blue eyes, bishounen, colorful, autumn, cumulonimbus clouds, lighting, blue sky, falling leaves, garden
91
+ Steps: 50, Sampler: DDIM, CFG scale: 12
92
+ ```
93
+ **Scenery:**
94
+ ![Scenery](https://huggingface.co/Linaqruf/anything-v3.0/resolve/main/scenery.png)
95
+ ```
96
+ scenery, shibuya tokyo, post-apocalypse, ruins, rust, sky, skyscraper, abandoned, blue sky, broken window, building, cloud, crane machine, outdoors, overgrown, pillar, sunset
97
+ Steps: 50, Sampler: DDIM, CFG scale: 12
98
+ ```
99
+
100
+ ## License
101
+
102
+ This model is open access and available to all, with a CreativeML OpenRAIL-M license further specifying rights and usage.
103
+ The CreativeML OpenRAIL License specifies:
104
+
105
+ 1. You can't use the model to deliberately produce nor share illegal or harmful outputs or content
106
+ 2. The authors claims no rights on the outputs you generate, you are free to use them and are accountable for their use which must not go against the provisions set in the license
107
+ 3. You may re-distribute the weights and use the model commercially and/or as a service. If you do, please be aware you have to include the same use restrictions as the ones in the license and share a copy of the CreativeML OpenRAIL-M to all your users (please read the license entirely and carefully)
108
+ [Please read the full license here](https://huggingface.co/spaces/CompVis/stable-diffusion-license)
README_HG.md ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ title: Anything V3.0
3
+ emoji: 🏃
4
+ colorFrom: gray
5
+ colorTo: yellow
6
+ sdk: gradio
7
+ sdk_version: 3.10.1
8
+ app_file: app.py
9
+ pinned: false
10
+ ---
11
+
12
+ Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
Waifu2x/.gitattributes ADDED
@@ -0,0 +1 @@
 
 
1
+ Readme_imgs/* linguist-documentation
Waifu2x/.gitignore ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+
2
+ *.xml
3
+ *.iml
4
+ *.pyc
Waifu2x/Common.py ADDED
@@ -0,0 +1,189 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from contextlib import contextmanager
2
+ from math import sqrt, log
3
+
4
+ import torch
5
+ import torch.nn as nn
6
+
7
+
8
+ # import warnings
9
+ # warnings.simplefilter('ignore')
10
+
11
+
12
+ class BaseModule(nn.Module):
13
+ def __init__(self):
14
+ self.act_fn = None
15
+ super(BaseModule, self).__init__()
16
+
17
+ def selu_init_params(self):
18
+ for m in self.modules():
19
+ if isinstance(m, nn.Conv2d) and m.weight.requires_grad:
20
+ m.weight.data.normal_(0.0, 1.0 / sqrt(m.weight.numel()))
21
+ if m.bias is not None:
22
+ m.bias.data.fill_(0)
23
+ elif isinstance(m, nn.BatchNorm2d) and m.weight.requires_grad:
24
+ m.weight.data.fill_(1)
25
+ m.bias.data.zero_()
26
+
27
+ elif isinstance(m, nn.Linear) and m.weight.requires_grad:
28
+ m.weight.data.normal_(0, 1.0 / sqrt(m.weight.numel()))
29
+ m.bias.data.zero_()
30
+
31
+ def initialize_weights_xavier_uniform(self):
32
+ for m in self.modules():
33
+ if isinstance(m, nn.Conv2d) and m.weight.requires_grad:
34
+ # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='leaky_relu')
35
+ nn.init.xavier_uniform_(m.weight)
36
+ if m.bias is not None:
37
+ m.bias.data.zero_()
38
+ elif isinstance(m, nn.BatchNorm2d) and m.weight.requires_grad:
39
+ m.weight.data.fill_(1)
40
+ m.bias.data.zero_()
41
+
42
+ def load_state_dict(self, state_dict, strict=True, self_state=False):
43
+ own_state = self_state if self_state else self.state_dict()
44
+ for name, param in state_dict.items():
45
+ if name in own_state:
46
+ try:
47
+ own_state[name].copy_(param.data)
48
+ except Exception as e:
49
+ print("Parameter {} fails to load.".format(name))
50
+ print("-----------------------------------------")
51
+ print(e)
52
+ else:
53
+ print("Parameter {} is not in the model. ".format(name))
54
+
55
+ @contextmanager
56
+ def set_activation_inplace(self):
57
+ if hasattr(self, 'act_fn') and hasattr(self.act_fn, 'inplace'):
58
+ # save memory
59
+ self.act_fn.inplace = True
60
+ yield
61
+ self.act_fn.inplace = False
62
+ else:
63
+ yield
64
+
65
+ def total_parameters(self):
66
+ total = sum([i.numel() for i in self.parameters()])
67
+ trainable = sum([i.numel() for i in self.parameters() if i.requires_grad])
68
+ print("Total parameters : {}. Trainable parameters : {}".format(total, trainable))
69
+ return total
70
+
71
+ def forward(self, *x):
72
+ raise NotImplementedError
73
+
74
+
75
+ class ResidualFixBlock(BaseModule):
76
+ def __init__(self, in_channels, out_channels, kernel_size=3, padding=1, dilation=1,
77
+ groups=1, activation=nn.SELU(), conv=nn.Conv2d):
78
+ super(ResidualFixBlock, self).__init__()
79
+ self.act_fn = activation
80
+ self.m = nn.Sequential(
81
+ conv(in_channels, out_channels, kernel_size, padding=padding, dilation=dilation, groups=groups),
82
+ activation,
83
+ # conv(out_channels, out_channels, kernel_size, padding=(kernel_size - 1) // 2, dilation=1, groups=groups),
84
+ conv(in_channels, out_channels, kernel_size, padding=padding, dilation=dilation, groups=groups),
85
+ )
86
+
87
+ def forward(self, x):
88
+ out = self.m(x)
89
+ return self.act_fn(out + x)
90
+
91
+
92
+ class ConvBlock(BaseModule):
93
+ def __init__(self, in_channels, out_channels, kernel_size=3, padding=1, dilation=1, groups=1,
94
+ activation=nn.SELU(), conv=nn.Conv2d):
95
+ super(ConvBlock, self).__init__()
96
+ self.m = nn.Sequential(conv(in_channels, out_channels, kernel_size, padding=padding,
97
+ dilation=dilation, groups=groups),
98
+ activation)
99
+
100
+ def forward(self, x):
101
+ return self.m(x)
102
+
103
+
104
+ class UpSampleBlock(BaseModule):
105
+ def __init__(self, channels, scale, activation, atrous_rate=1, conv=nn.Conv2d):
106
+ assert scale in [2, 4, 8], "Currently UpSampleBlock supports 2, 4, 8 scaling"
107
+ super(UpSampleBlock, self).__init__()
108
+ m = nn.Sequential(
109
+ conv(channels, 4 * channels, kernel_size=3, padding=atrous_rate, dilation=atrous_rate),
110
+ activation,
111
+ nn.PixelShuffle(2)
112
+ )
113
+ self.m = nn.Sequential(*[m for _ in range(int(log(scale, 2)))])
114
+
115
+ def forward(self, x):
116
+ return self.m(x)
117
+
118
+
119
+ class SpatialChannelSqueezeExcitation(BaseModule):
120
+ # https://arxiv.org/abs/1709.01507
121
+ # https://arxiv.org/pdf/1803.02579v1.pdf
122
+ def __init__(self, in_channel, reduction=16, activation=nn.ReLU()):
123
+ super(SpatialChannelSqueezeExcitation, self).__init__()
124
+ linear_nodes = max(in_channel // reduction, 4) # avoid only 1 node case
125
+ self.avg_pool = nn.AdaptiveAvgPool2d(1)
126
+ self.channel_excite = nn.Sequential(
127
+ # check the paper for the number 16 in reduction. It is selected by experiment.
128
+ nn.Linear(in_channel, linear_nodes),
129
+ activation,
130
+ nn.Linear(linear_nodes, in_channel),
131
+ nn.Sigmoid()
132
+ )
133
+ self.spatial_excite = nn.Sequential(
134
+ nn.Conv2d(in_channel, 1, kernel_size=1, stride=1, padding=0, bias=False),
135
+ nn.Sigmoid()
136
+ )
137
+
138
+ def forward(self, x):
139
+ b, c, h, w = x.size()
140
+ #
141
+ channel = self.avg_pool(x).view(b, c)
142
+ # channel = F.avg_pool2d(x, kernel_size=(h,w)).view(b,c) # used for porting to other frameworks
143
+ cSE = self.channel_excite(channel).view(b, c, 1, 1)
144
+ x_cSE = torch.mul(x, cSE)
145
+
146
+ # spatial
147
+ sSE = self.spatial_excite(x)
148
+ x_sSE = torch.mul(x, sSE)
149
+ # return x_sSE
150
+ return torch.add(x_cSE, x_sSE)
151
+
152
+
153
+ class PartialConv(nn.Module):
154
+ # reference:
155
+ # Image Inpainting for Irregular Holes Using Partial Convolutions
156
+ # http://masc.cs.gmu.edu/wiki/partialconv/show?time=2018-05-24+21%3A41%3A10
157
+ # https://github.com/naoto0804/pytorch-inpainting-with-partial-conv/blob/master/net.py
158
+ # https://github.com/SeitaroShinagawa/chainer-partial_convolution_image_inpainting/blob/master/common/net.py
159
+ # partial based padding
160
+ # https: // github.com / NVIDIA / partialconv / blob / master / models / pd_resnet.py
161
+ def __init__(self, in_channels, out_channels, kernel_size, stride=1,
162
+ padding=0, dilation=1, groups=1, bias=True):
163
+
164
+ super(PartialConv, self).__init__()
165
+ self.feature_conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride,
166
+ padding, dilation, groups, bias)
167
+
168
+ self.mask_conv = nn.Conv2d(1, 1, kernel_size, stride,
169
+ padding, dilation, groups, bias=False)
170
+ self.window_size = self.mask_conv.kernel_size[0] * self.mask_conv.kernel_size[1]
171
+ torch.nn.init.constant_(self.mask_conv.weight, 1.0)
172
+
173
+ for param in self.mask_conv.parameters():
174
+ param.requires_grad = False
175
+
176
+ def forward(self, x):
177
+ output = self.feature_conv(x)
178
+ if self.feature_conv.bias is not None:
179
+ output_bias = self.feature_conv.bias.view(1, -1, 1, 1).expand_as(output)
180
+ else:
181
+ output_bias = torch.zeros_like(output, device=x.device)
182
+
183
+ with torch.no_grad():
184
+ ones = torch.ones(1, 1, x.size(2), x.size(3), device=x.device)
185
+ output_mask = self.mask_conv(ones)
186
+ output_mask = self.window_size / output_mask
187
+ output = (output - output_bias) * output_mask + output_bias
188
+
189
+ return output
Waifu2x/Dataloader.py ADDED
@@ -0,0 +1,215 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import glob
2
+ import io
3
+ import numpy as np
4
+ import re
5
+ import os
6
+ import random
7
+ from io import BytesIO
8
+ from uuid import uuid4
9
+ import sqlite3
10
+ import h5py
11
+ import torch
12
+ from PIL import Image
13
+ from torch.utils.data import Dataset
14
+ from torchvision.transforms import RandomCrop
15
+ from torchvision.transforms.functional import to_tensor
16
+
17
+
18
+ class ImageH5Data(Dataset):
19
+ def __init__(self, h5py_file, folder_name):
20
+ self.data = h5py.File(h5py_file, 'r')[folder_name]
21
+ self.data_hr = self.data['train_hr']
22
+ self.data_lr = self.data['train_lr']
23
+ self.len_imgs = len(self.data_hr)
24
+ self.h5py_file = h5py_file
25
+ self.folder_name = folder_name
26
+
27
+ def __len__(self):
28
+ # with h5py.File(self.h5py_file, 'r') as f:
29
+ # return len(f[self.folder_name]['train_lr'])
30
+ return self.len_imgs
31
+
32
+ def __getitem__(self, index):
33
+ # with h5py.File(self.h5py_file, 'r') as f:
34
+ # data_lr = f[self.folder_name]['train_lr'][index]
35
+ # data_hr = f[self.folder_name]['train_lr'][index]
36
+ #
37
+ # return data_lr, data_hr
38
+ return self.data_lr[index], self.data_hr[index]
39
+
40
+
41
+ class ImageData(Dataset):
42
+ def __init__(self,
43
+ img_folder,
44
+ patch_size=96,
45
+ shrink_size=2,
46
+ noise_level=1,
47
+ down_sample_method=None,
48
+ color_mod='RGB',
49
+ dummy_len=None):
50
+
51
+ self.img_folder = img_folder
52
+ all_img = glob.glob(self.img_folder + "/**", recursive=True)
53
+ self.img = list(filter(lambda x: x.endswith('png') or x.endswith("jpg") or x.endswith("jpeg"), all_img))
54
+ self.total_img = len(self.img)
55
+ self.dummy_len = dummy_len if dummy_len is not None else self.total_img
56
+ self.random_cropper = RandomCrop(size=patch_size)
57
+ self.color_mod = color_mod
58
+ self.img_augmenter = ImageAugment(shrink_size, noise_level, down_sample_method)
59
+
60
+ def get_img_patches(self, img_file):
61
+ img_pil = Image.open(img_file).convert("RGB")
62
+ img_patch = self.random_cropper(img_pil)
63
+ lr_hr_patches = self.img_augmenter.process(img_patch)
64
+ return lr_hr_patches
65
+
66
+ def __len__(self):
67
+ return self.dummy_len # len(self.img)
68
+
69
+ def __getitem__(self, index):
70
+ idx = random.choice(range(0, self.total_img))
71
+ img = self.img[idx]
72
+ patch = self.get_img_patches(img)
73
+ if self.color_mod == 'RGB':
74
+ lr_img = patch[0].convert("RGB")
75
+ hr_img = patch[1].convert("RGB")
76
+ elif self.color_mod == 'YCbCr':
77
+ lr_img, _, _ = patch[0].convert('YCbCr').split()
78
+ hr_img, _, _ = patch[1].convert('YCbCr').split()
79
+ else:
80
+ raise KeyError('Either RGB or YCbCr')
81
+ return to_tensor(lr_img), to_tensor(hr_img)
82
+
83
+
84
+ class Image2Sqlite(ImageData):
85
+ def __getitem__(self, item):
86
+ img = self.img[item]
87
+ lr_hr_patch = self.get_img_patches(img)
88
+ if self.color_mod == 'RGB':
89
+ lr_img = lr_hr_patch[0].convert("RGB")
90
+ hr_img = lr_hr_patch[1].convert("RGB")
91
+ elif self.color_mod == 'YCbCr':
92
+ lr_img, _, _ = lr_hr_patch[0].convert('YCbCr').split()
93
+ hr_img, _, _ = lr_hr_patch[1].convert('YCbCr').split()
94
+ else:
95
+ raise KeyError('Either RGB or YCbCr')
96
+ lr_byte = self.convert_to_bytevalue(lr_img)
97
+ hr_byte = self.convert_to_bytevalue(hr_img)
98
+ return [lr_byte, hr_byte]
99
+
100
+ @staticmethod
101
+ def convert_to_bytevalue(pil_img):
102
+ img_byte = io.BytesIO()
103
+ pil_img.save(img_byte, format='png')
104
+ return img_byte.getvalue()
105
+
106
+
107
+ class ImageDBData(Dataset):
108
+ def __init__(self, db_file, db_table="images", lr_col="lr_img", hr_col="hr_img", max_images=None):
109
+ self.db_file = db_file
110
+ self.db_table = db_table
111
+ self.lr_col = lr_col
112
+ self.hr_col = hr_col
113
+ self.total_images = self.get_num_rows(max_images)
114
+ # self.lr_hr_images = self.get_all_images()
115
+
116
+ def __len__(self):
117
+ return self.total_images
118
+
119
+ # def get_all_images(self):
120
+ # with sqlite3.connect(self.db_file) as conn:
121
+ # cursor = conn.cursor()
122
+ # cursor.execute(f"SELECT * FROM {self.db_table} LIMIT {self.total_images}")
123
+ # return cursor.fetchall()
124
+
125
+ def get_num_rows(self, max_images):
126
+ with sqlite3.connect(self.db_file) as conn:
127
+ cursor = conn.cursor()
128
+ cursor.execute(f"SELECT MAX(ROWID) FROM {self.db_table}")
129
+ db_rows = cursor.fetchone()[0]
130
+ if max_images:
131
+ return min(max_images, db_rows)
132
+ else:
133
+ return db_rows
134
+
135
+ def __getitem__(self, item):
136
+ # lr, hr = self.lr_hr_images[item]
137
+ # lr = Image.open(io.BytesIO(lr))
138
+ # hr = Image.open(io.BytesIO(hr))
139
+ # return to_tensor(lr), to_tensor(hr)
140
+ # note sqlite rowid starts with 1
141
+ with sqlite3.connect(self.db_file) as conn:
142
+ cursor = conn.cursor()
143
+ cursor.execute(f"SELECT {self.lr_col}, {self.hr_col} FROM {self.db_table} WHERE ROWID={item + 1}")
144
+ lr, hr = cursor.fetchone()
145
+ lr = Image.open(io.BytesIO(lr)).convert("RGB")
146
+ hr = Image.open(io.BytesIO(hr)).convert("RGB")
147
+ # lr = np.array(lr) # use scale [0, 255] instead of [0,1]
148
+ # hr = np.array(hr)
149
+ return to_tensor(lr), to_tensor(hr)
150
+
151
+
152
+ class ImagePatchData(Dataset):
153
+ def __init__(self, lr_folder, hr_folder):
154
+ self.lr_folder = lr_folder
155
+ self.hr_folder = hr_folder
156
+ self.lr_imgs = glob.glob(os.path.join(lr_folder, "**"))
157
+ self.total_imgs = len(self.lr_imgs)
158
+
159
+ def __len__(self):
160
+ return self.total_imgs
161
+
162
+ def __getitem__(self, item):
163
+ lr_file = self.lr_imgs[item]
164
+ hr_path = re.sub("lr", 'hr', os.path.dirname(lr_file))
165
+ filename = os.path.basename(lr_file)
166
+ hr_file = os.path.join(hr_path, filename)
167
+ return to_tensor(Image.open(lr_file)), to_tensor(Image.open(hr_file))
168
+
169
+
170
+ class ImageAugment:
171
+ def __init__(self,
172
+ shrink_size=2,
173
+ noise_level=1,
174
+ down_sample_method=None
175
+ ):
176
+ # noise_level (int): 0: no noise; 1: 75-95% quality; 2:50-75%
177
+ if noise_level == 0:
178
+ self.noise_level = [0, 0]
179
+ elif noise_level == 1:
180
+ self.noise_level = [5, 25]
181
+ elif noise_level == 2:
182
+ self.noise_level = [25, 50]
183
+ else:
184
+ raise KeyError("Noise level should be either 0, 1, 2")
185
+ self.shrink_size = shrink_size
186
+ self.down_sample_method = down_sample_method
187
+
188
+ def shrink_img(self, hr_img):
189
+
190
+ if self.down_sample_method is None:
191
+ resample_method = random.choice([Image.BILINEAR, Image.BICUBIC, Image.LANCZOS])
192
+ else:
193
+ resample_method = self.down_sample_method
194
+ img_w, img_h = tuple(map(lambda x: int(x / self.shrink_size), hr_img.size))
195
+ lr_img = hr_img.resize((img_w, img_h), resample_method)
196
+ return lr_img
197
+
198
+ def add_jpeg_noise(self, hr_img):
199
+ quality = 100 - round(random.uniform(*self.noise_level))
200
+ lr_img = BytesIO()
201
+ hr_img.save(lr_img, format='JPEG', quality=quality)
202
+ lr_img.seek(0)
203
+ lr_img = Image.open(lr_img)
204
+ return lr_img
205
+
206
+ def process(self, hr_patch_pil):
207
+ lr_patch_pil = self.shrink_img(hr_patch_pil)
208
+ if self.noise_level[1] > 0:
209
+ lr_patch_pil = self.add_jpeg_noise(lr_patch_pil)
210
+
211
+ return lr_patch_pil, hr_patch_pil
212
+
213
+ def up_sample(self, img, resample):
214
+ width, height = img.size
215
+ return img.resize((self.shrink_size * width, self.shrink_size * height), resample=resample)
Waifu2x/Img_to_Sqlite.py ADDED
@@ -0,0 +1,115 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ Split images into small patches and insert them into sqlite db. Reading and Inserting speeds are much better than
3
+ Ubuntu's (18.04) file system when the number of patches is larger than 20k. And it has smaller size than using h5 format
4
+
5
+ Recommend to check or filter out small size patches as their content vary little. 128x128 seems better than 64x64.
6
+
7
+
8
+ """
9
+ import sqlite3
10
+ from torch.utils.data import DataLoader
11
+ from tqdm import trange
12
+ from Dataloader import Image2Sqlite
13
+
14
+ conn = sqlite3.connect('dataset/image_yandere.db')
15
+ cursor = conn.cursor()
16
+
17
+ with conn:
18
+ cursor.execute("PRAGMA SYNCHRONOUS = OFF")
19
+
20
+ table_name = "train_images_size_128_noise_1_rgb"
21
+ lr_col = "lr_img"
22
+ hr_col = "hr_img"
23
+
24
+ with conn:
25
+ conn.execute(f"CREATE TABLE IF NOT EXISTS {table_name} ({lr_col} BLOB, {hr_col} BLOB)")
26
+
27
+ dat = Image2Sqlite(img_folder='./dataset/yande.re_test_shrink',
28
+ patch_size=256,
29
+ shrink_size=2,
30
+ noise_level=1,
31
+ down_sample_method=None,
32
+ color_mod='RGB',
33
+ dummy_len=None)
34
+ print(f"Total images {len(dat)}")
35
+
36
+ img_dat = DataLoader(dat, num_workers=6, batch_size=6, shuffle=True)
37
+
38
+ num_batches = 20
39
+ for i in trange(num_batches):
40
+ bulk = []
41
+ for lrs, hrs in img_dat:
42
+ patches = [(lrs[i], hrs[i]) for i in range(len(lrs))]
43
+ # patches = [(lrs[i], hrs[i]) for i in range(len(lrs)) if len(lrs[i]) > 14000]
44
+
45
+ bulk.extend(patches)
46
+
47
+ bulk = [i for i in bulk if len(i[0]) > 15000] # for 128x128, 14000 is fair. Around 20% of patches are filtered out
48
+ cursor.executemany(f"INSERT INTO {table_name}({lr_col}, {hr_col}) VALUES (?,?)", bulk)
49
+ conn.commit()
50
+
51
+ cursor.execute(f"select max(rowid) from {table_name}")
52
+ print(cursor.fetchall())
53
+ conn.commit()
54
+ # +++++++++++++++++++++++++++++++++++++
55
+ # Used for Create Test Database
56
+ # -------------------------------------
57
+
58
+ # cursor.execute(f"SELECT ROWID FROM {table_name} ORDER BY LENGTH({lr_col}) DESC LIMIT 400")
59
+ # rowdis = cursor.fetchall()
60
+ # rowdis = ",".join([str(i[0]) for i in rowdis])
61
+ #
62
+ # cursor.execute(f"DELETE FROM {table_name} WHERE ROWID NOT IN ({rowdis})")
63
+ # conn.commit()
64
+ # cursor.execute("vacuum")
65
+ #
66
+ # cursor.execute("""
67
+ # CREATE TABLE IF NOT EXISTS train_images_size_128_noise_1_rgb_small AS
68
+ # SELECT *
69
+ # FROM train_images_size_128_noise_1_rgb
70
+ # WHERE length(lr_img) < 14000;
71
+ # """)
72
+ #
73
+ # cursor.execute("""
74
+ # DELETE
75
+ # FROM train_images_size_128_noise_1_rgb
76
+ # WHERE length(lr_img) < 14000;
77
+ # """)
78
+
79
+ # reset index
80
+ cursor.execute("VACUUM")
81
+ conn.commit()
82
+
83
+ # +++++++++++++++++++++++++++++++++++++
84
+ # check image size
85
+ # -------------------------------------
86
+ #
87
+
88
+ from PIL import Image
89
+ import io
90
+
91
+ cursor.execute(
92
+ f"""
93
+ select {hr_col} from {table_name}
94
+ ORDER BY LENGTH({hr_col}) desc
95
+ limit 100
96
+ """
97
+ )
98
+ # WHERE LENGTH({lr_col}) BETWEEN 14000 AND 16000
99
+
100
+ # small = cursor.fetchall()
101
+ # print(len(small))
102
+ for idx, i in enumerate(cursor):
103
+ img = Image.open(io.BytesIO(i[0]))
104
+ img.save(f"dataset/check/{idx}.png")
105
+
106
+ # +++++++++++++++++++++++++++++++++++++
107
+ # Check Image Variance
108
+ # -------------------------------------
109
+
110
+ import pandas as pd
111
+ import matplotlib.pyplot as plt
112
+
113
+ dat = pd.read_sql(f"SELECT length({lr_col}) from {table_name}", conn)
114
+ dat.hist(bins=20)
115
+ plt.show()
Waifu2x/LICENSE ADDED
@@ -0,0 +1,674 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ GNU GENERAL PUBLIC LICENSE
2
+ Version 3, 29 June 2007
3
+
4
+ Copyright (C) 2007 Free Software Foundation, Inc. <http://fsf.org/>
5
+ Everyone is permitted to copy and distribute verbatim copies
6
+ of this license document, but changing it is not allowed.
7
+
8
+ Preamble
9
+
10
+ The GNU General Public License is a free, copyleft license for
11
+ software and other kinds of works.
12
+
13
+ The licenses for most software and other practical works are designed
14
+ to take away your freedom to share and change the works. By contrast,
15
+ the GNU General Public License is intended to guarantee your freedom to
16
+ share and change all versions of a program--to make sure it remains free
17
+ software for all its users. We, the Free Software Foundation, use the
18
+ GNU General Public License for most of our software; it applies also to
19
+ any other work released this way by its authors. You can apply it to
20
+ your programs, too.
21
+
22
+ When we speak of free software, we are referring to freedom, not
23
+ price. Our General Public Licenses are designed to make sure that you
24
+ have the freedom to distribute copies of free software (and charge for
25
+ them if you wish), that you receive source code or can get it if you
26
+ want it, that you can change the software or use pieces of it in new
27
+ free programs, and that you know you can do these things.
28
+
29
+ To protect your rights, we need to prevent others from denying you
30
+ these rights or asking you to surrender the rights. Therefore, you have
31
+ certain responsibilities if you distribute copies of the software, or if
32
+ you modify it: responsibilities to respect the freedom of others.
33
+
34
+ For example, if you distribute copies of such a program, whether
35
+ gratis or for a fee, you must pass on to the recipients the same
36
+ freedoms that you received. You must make sure that they, too, receive
37
+ or can get the source code. And you must show them these terms so they
38
+ know their rights.
39
+
40
+ Developers that use the GNU GPL protect your rights with two steps:
41
+ (1) assert copyright on the software, and (2) offer you this License
42
+ giving you legal permission to copy, distribute and/or modify it.
43
+
44
+ For the developers' and authors' protection, the GPL clearly explains
45
+ that there is no warranty for this free software. For both users' and
46
+ authors' sake, the GPL requires that modified versions be marked as
47
+ changed, so that their problems will not be attributed erroneously to
48
+ authors of previous versions.
49
+
50
+ Some devices are designed to deny users access to install or run
51
+ modified versions of the software inside them, although the manufacturer
52
+ can do so. This is fundamentally incompatible with the aim of
53
+ protecting users' freedom to change the software. The systematic
54
+ pattern of such abuse occurs in the area of products for individuals to
55
+ use, which is precisely where it is most unacceptable. Therefore, we
56
+ have designed this version of the GPL to prohibit the practice for those
57
+ products. If such problems arise substantially in other domains, we
58
+ stand ready to extend this provision to those domains in future versions
59
+ of the GPL, as needed to protect the freedom of users.
60
+
61
+ Finally, every program is threatened constantly by software patents.
62
+ States should not allow patents to restrict development and use of
63
+ software on general-purpose computers, but in those that do, we wish to
64
+ avoid the special danger that patents applied to a free program could
65
+ make it effectively proprietary. To prevent this, the GPL assures that
66
+ patents cannot be used to render the program non-free.
67
+
68
+ The precise terms and conditions for copying, distribution and
69
+ modification follow.
70
+
71
+ TERMS AND CONDITIONS
72
+
73
+ 0. Definitions.
74
+
75
+ "This License" refers to version 3 of the GNU General Public License.
76
+
77
+ "Copyright" also means copyright-like laws that apply to other kinds of
78
+ works, such as semiconductor masks.
79
+
80
+ "The Program" refers to any copyrightable work licensed under this
81
+ License. Each licensee is addressed as "you". "Licensees" and
82
+ "recipients" may be individuals or organizations.
83
+
84
+ To "modify" a work means to copy from or adapt all or part of the work
85
+ in a fashion requiring copyright permission, other than the making of an
86
+ exact copy. The resulting work is called a "modified version" of the
87
+ earlier work or a work "based on" the earlier work.
88
+
89
+ A "covered work" means either the unmodified Program or a work based
90
+ on the Program.
91
+
92
+ To "propagate" a work means to do anything with it that, without
93
+ permission, would make you directly or secondarily liable for
94
+ infringement under applicable copyright law, except executing it on a
95
+ computer or modifying a private copy. Propagation includes copying,
96
+ distribution (with or without modification), making available to the
97
+ public, and in some countries other activities as well.
98
+
99
+ To "convey" a work means any kind of propagation that enables other
100
+ parties to make or receive copies. Mere interaction with a user through
101
+ a computer network, with no transfer of a copy, is not conveying.
102
+
103
+ An interactive user interface displays "Appropriate Legal Notices"
104
+ to the extent that it includes a convenient and prominently visible
105
+ feature that (1) displays an appropriate copyright notice, and (2)
106
+ tells the user that there is no warranty for the work (except to the
107
+ extent that warranties are provided), that licensees may convey the
108
+ work under this License, and how to view a copy of this License. If
109
+ the interface presents a list of user commands or options, such as a
110
+ menu, a prominent item in the list meets this criterion.
111
+
112
+ 1. Source Code.
113
+
114
+ The "source code" for a work means the preferred form of the work
115
+ for making modifications to it. "Object code" means any non-source
116
+ form of a work.
117
+
118
+ A "Standard Interface" means an interface that either is an official
119
+ standard defined by a recognized standards body, or, in the case of
120
+ interfaces specified for a particular programming language, one that
121
+ is widely used among developers working in that language.
122
+
123
+ The "System Libraries" of an executable work include anything, other
124
+ than the work as a whole, that (a) is included in the normal form of
125
+ packaging a Major Component, but which is not part of that Major
126
+ Component, and (b) serves only to enable use of the work with that
127
+ Major Component, or to implement a Standard Interface for which an
128
+ implementation is available to the public in source code form. A
129
+ "Major Component", in this context, means a major essential component
130
+ (kernel, window system, and so on) of the specific operating system
131
+ (if any) on which the executable work runs, or a compiler used to
132
+ produce the work, or an object code interpreter used to run it.
133
+
134
+ The "Corresponding Source" for a work in object code form means all
135
+ the source code needed to generate, install, and (for an executable
136
+ work) run the object code and to modify the work, including scripts to
137
+ control those activities. However, it does not include the work's
138
+ System Libraries, or general-purpose tools or generally available free
139
+ programs which are used unmodified in performing those activities but
140
+ which are not part of the work. For example, Corresponding Source
141
+ includes interface definition files associated with source files for
142
+ the work, and the source code for shared libraries and dynamically
143
+ linked subprograms that the work is specifically designed to require,
144
+ such as by intimate data communication or control flow between those
145
+ subprograms and other parts of the work.
146
+
147
+ The Corresponding Source need not include anything that users
148
+ can regenerate automatically from other parts of the Corresponding
149
+ Source.
150
+
151
+ The Corresponding Source for a work in source code form is that
152
+ same work.
153
+
154
+ 2. Basic Permissions.
155
+
156
+ All rights granted under this License are granted for the term of
157
+ copyright on the Program, and are irrevocable provided the stated
158
+ conditions are met. This License explicitly affirms your unlimited
159
+ permission to run the unmodified Program. The output from running a
160
+ covered work is covered by this License only if the output, given its
161
+ content, constitutes a covered work. This License acknowledges your
162
+ rights of fair use or other equivalent, as provided by copyright law.
163
+
164
+ You may make, run and propagate covered works that you do not
165
+ convey, without conditions so long as your license otherwise remains
166
+ in force. You may convey covered works to others for the sole purpose
167
+ of having them make modifications exclusively for you, or provide you
168
+ with facilities for running those works, provided that you comply with
169
+ the terms of this License in conveying all material for which you do
170
+ not control copyright. Those thus making or running the covered works
171
+ for you must do so exclusively on your behalf, under your direction
172
+ and control, on terms that prohibit them from making any copies of
173
+ your copyrighted material outside their relationship with you.
174
+
175
+ Conveying under any other circumstances is permitted solely under
176
+ the conditions stated below. Sublicensing is not allowed; section 10
177
+ makes it unnecessary.
178
+
179
+ 3. Protecting Users' Legal Rights From Anti-Circumvention Law.
180
+
181
+ No covered work shall be deemed part of an effective technological
182
+ measure under any applicable law fulfilling obligations under article
183
+ 11 of the WIPO copyright treaty adopted on 20 December 1996, or
184
+ similar laws prohibiting or restricting circumvention of such
185
+ measures.
186
+
187
+ When you convey a covered work, you waive any legal power to forbid
188
+ circumvention of technological measures to the extent such circumvention
189
+ is effected by exercising rights under this License with respect to
190
+ the covered work, and you disclaim any intention to limit operation or
191
+ modification of the work as a means of enforcing, against the work's
192
+ users, your or third parties' legal rights to forbid circumvention of
193
+ technological measures.
194
+
195
+ 4. Conveying Verbatim Copies.
196
+
197
+ You may convey verbatim copies of the Program's source code as you
198
+ receive it, in any medium, provided that you conspicuously and
199
+ appropriately publish on each copy an appropriate copyright notice;
200
+ keep intact all notices stating that this License and any
201
+ non-permissive terms added in accord with section 7 apply to the code;
202
+ keep intact all notices of the absence of any warranty; and give all
203
+ recipients a copy of this License along with the Program.
204
+
205
+ You may charge any price or no price for each copy that you convey,
206
+ and you may offer support or warranty protection for a fee.
207
+
208
+ 5. Conveying Modified Source Versions.
209
+
210
+ You may convey a work based on the Program, or the modifications to
211
+ produce it from the Program, in the form of source code under the
212
+ terms of section 4, provided that you also meet all of these conditions:
213
+
214
+ a) The work must carry prominent notices stating that you modified
215
+ it, and giving a relevant date.
216
+
217
+ b) The work must carry prominent notices stating that it is
218
+ released under this License and any conditions added under section
219
+ 7. This requirement modifies the requirement in section 4 to
220
+ "keep intact all notices".
221
+
222
+ c) You must license the entire work, as a whole, under this
223
+ License to anyone who comes into possession of a copy. This
224
+ License will therefore apply, along with any applicable section 7
225
+ additional terms, to the whole of the work, and all its parts,
226
+ regardless of how they are packaged. This License gives no
227
+ permission to license the work in any other way, but it does not
228
+ invalidate such permission if you have separately received it.
229
+
230
+ d) If the work has interactive user interfaces, each must display
231
+ Appropriate Legal Notices; however, if the Program has interactive
232
+ interfaces that do not display Appropriate Legal Notices, your
233
+ work need not make them do so.
234
+
235
+ A compilation of a covered work with other separate and independent
236
+ works, which are not by their nature extensions of the covered work,
237
+ and which are not combined with it such as to form a larger program,
238
+ in or on a volume of a storage or distribution medium, is called an
239
+ "aggregate" if the compilation and its resulting copyright are not
240
+ used to limit the access or legal rights of the compilation's users
241
+ beyond what the individual works permit. Inclusion of a covered work
242
+ in an aggregate does not cause this License to apply to the other
243
+ parts of the aggregate.
244
+
245
+ 6. Conveying Non-Source Forms.
246
+
247
+ You may convey a covered work in object code form under the terms
248
+ of sections 4 and 5, provided that you also convey the
249
+ machine-readable Corresponding Source under the terms of this License,
250
+ in one of these ways:
251
+
252
+ a) Convey the object code in, or embodied in, a physical product
253
+ (including a physical distribution medium), accompanied by the
254
+ Corresponding Source fixed on a durable physical medium
255
+ customarily used for software interchange.
256
+
257
+ b) Convey the object code in, or embodied in, a physical product
258
+ (including a physical distribution medium), accompanied by a
259
+ written offer, valid for at least three years and valid for as
260
+ long as you offer spare parts or customer support for that product
261
+ model, to give anyone who possesses the object code either (1) a
262
+ copy of the Corresponding Source for all the software in the
263
+ product that is covered by this License, on a durable physical
264
+ medium customarily used for software interchange, for a price no
265
+ more than your reasonable cost of physically performing this
266
+ conveying of source, or (2) access to copy the
267
+ Corresponding Source from a network server at no charge.
268
+
269
+ c) Convey individual copies of the object code with a copy of the
270
+ written offer to provide the Corresponding Source. This
271
+ alternative is allowed only occasionally and noncommercially, and
272
+ only if you received the object code with such an offer, in accord
273
+ with subsection 6b.
274
+
275
+ d) Convey the object code by offering access from a designated
276
+ place (gratis or for a charge), and offer equivalent access to the
277
+ Corresponding Source in the same way through the same place at no
278
+ further charge. You need not require recipients to copy the
279
+ Corresponding Source along with the object code. If the place to
280
+ copy the object code is a network server, the Corresponding Source
281
+ may be on a different server (operated by you or a third party)
282
+ that supports equivalent copying facilities, provided you maintain
283
+ clear directions next to the object code saying where to find the
284
+ Corresponding Source. Regardless of what server hosts the
285
+ Corresponding Source, you remain obligated to ensure that it is
286
+ available for as long as needed to satisfy these requirements.
287
+
288
+ e) Convey the object code using peer-to-peer transmission, provided
289
+ you inform other peers where the object code and Corresponding
290
+ Source of the work are being offered to the general public at no
291
+ charge under subsection 6d.
292
+
293
+ A separable portion of the object code, whose source code is excluded
294
+ from the Corresponding Source as a System Library, need not be
295
+ included in conveying the object code work.
296
+
297
+ A "User Product" is either (1) a "consumer product", which means any
298
+ tangible personal property which is normally used for personal, family,
299
+ or household purposes, or (2) anything designed or sold for incorporation
300
+ into a dwelling. In determining whether a product is a consumer product,
301
+ doubtful cases shall be resolved in favor of coverage. For a particular
302
+ product received by a particular user, "normally used" refers to a
303
+ typical or common use of that class of product, regardless of the status
304
+ of the particular user or of the way in which the particular user
305
+ actually uses, or expects or is expected to use, the product. A product
306
+ is a consumer product regardless of whether the product has substantial
307
+ commercial, industrial or non-consumer uses, unless such uses represent
308
+ the only significant mode of use of the product.
309
+
310
+ "Installation Information" for a User Product means any methods,
311
+ procedures, authorization keys, or other information required to install
312
+ and execute modified versions of a covered work in that User Product from
313
+ a modified version of its Corresponding Source. The information must
314
+ suffice to ensure that the continued functioning of the modified object
315
+ code is in no case prevented or interfered with solely because
316
+ modification has been made.
317
+
318
+ If you convey an object code work under this section in, or with, or
319
+ specifically for use in, a User Product, and the conveying occurs as
320
+ part of a transaction in which the right of possession and use of the
321
+ User Product is transferred to the recipient in perpetuity or for a
322
+ fixed term (regardless of how the transaction is characterized), the
323
+ Corresponding Source conveyed under this section must be accompanied
324
+ by the Installation Information. But this requirement does not apply
325
+ if neither you nor any third party retains the ability to install
326
+ modified object code on the User Product (for example, the work has
327
+ been installed in ROM).
328
+
329
+ The requirement to provide Installation Information does not include a
330
+ requirement to continue to provide support service, warranty, or updates
331
+ for a work that has been modified or installed by the recipient, or for
332
+ the User Product in which it has been modified or installed. Access to a
333
+ network may be denied when the modification itself materially and
334
+ adversely affects the operation of the network or violates the rules and
335
+ protocols for communication across the network.
336
+
337
+ Corresponding Source conveyed, and Installation Information provided,
338
+ in accord with this section must be in a format that is publicly
339
+ documented (and with an implementation available to the public in
340
+ source code form), and must require no special password or key for
341
+ unpacking, reading or copying.
342
+
343
+ 7. Additional Terms.
344
+
345
+ "Additional permissions" are terms that supplement the terms of this
346
+ License by making exceptions from one or more of its conditions.
347
+ Additional permissions that are applicable to the entire Program shall
348
+ be treated as though they were included in this License, to the extent
349
+ that they are valid under applicable law. If additional permissions
350
+ apply only to part of the Program, that part may be used separately
351
+ under those permissions, but the entire Program remains governed by
352
+ this License without regard to the additional permissions.
353
+
354
+ When you convey a copy of a covered work, you may at your option
355
+ remove any additional permissions from that copy, or from any part of
356
+ it. (Additional permissions may be written to require their own
357
+ removal in certain cases when you modify the work.) You may place
358
+ additional permissions on material, added by you to a covered work,
359
+ for which you have or can give appropriate copyright permission.
360
+
361
+ Notwithstanding any other provision of this License, for material you
362
+ add to a covered work, you may (if authorized by the copyright holders of
363
+ that material) supplement the terms of this License with terms:
364
+
365
+ a) Disclaiming warranty or limiting liability differently from the
366
+ terms of sections 15 and 16 of this License; or
367
+
368
+ b) Requiring preservation of specified reasonable legal notices or
369
+ author attributions in that material or in the Appropriate Legal
370
+ Notices displayed by works containing it; or
371
+
372
+ c) Prohibiting misrepresentation of the origin of that material, or
373
+ requiring that modified versions of such material be marked in
374
+ reasonable ways as different from the original version; or
375
+
376
+ d) Limiting the use for publicity purposes of names of licensors or
377
+ authors of the material; or
378
+
379
+ e) Declining to grant rights under trademark law for use of some
380
+ trade names, trademarks, or service marks; or
381
+
382
+ f) Requiring indemnification of licensors and authors of that
383
+ material by anyone who conveys the material (or modified versions of
384
+ it) with contractual assumptions of liability to the recipient, for
385
+ any liability that these contractual assumptions directly impose on
386
+ those licensors and authors.
387
+
388
+ All other non-permissive additional terms are considered "further
389
+ restrictions" within the meaning of section 10. If the Program as you
390
+ received it, or any part of it, contains a notice stating that it is
391
+ governed by this License along with a term that is a further
392
+ restriction, you may remove that term. If a license document contains
393
+ a further restriction but permits relicensing or conveying under this
394
+ License, you may add to a covered work material governed by the terms
395
+ of that license document, provided that the further restriction does
396
+ not survive such relicensing or conveying.
397
+
398
+ If you add terms to a covered work in accord with this section, you
399
+ must place, in the relevant source files, a statement of the
400
+ additional terms that apply to those files, or a notice indicating
401
+ where to find the applicable terms.
402
+
403
+ Additional terms, permissive or non-permissive, may be stated in the
404
+ form of a separately written license, or stated as exceptions;
405
+ the above requirements apply either way.
406
+
407
+ 8. Termination.
408
+
409
+ You may not propagate or modify a covered work except as expressly
410
+ provided under this License. Any attempt otherwise to propagate or
411
+ modify it is void, and will automatically terminate your rights under
412
+ this License (including any patent licenses granted under the third
413
+ paragraph of section 11).
414
+
415
+ However, if you cease all violation of this License, then your
416
+ license from a particular copyright holder is reinstated (a)
417
+ provisionally, unless and until the copyright holder explicitly and
418
+ finally terminates your license, and (b) permanently, if the copyright
419
+ holder fails to notify you of the violation by some reasonable means
420
+ prior to 60 days after the cessation.
421
+
422
+ Moreover, your license from a particular copyright holder is
423
+ reinstated permanently if the copyright holder notifies you of the
424
+ violation by some reasonable means, this is the first time you have
425
+ received notice of violation of this License (for any work) from that
426
+ copyright holder, and you cure the violation prior to 30 days after
427
+ your receipt of the notice.
428
+
429
+ Termination of your rights under this section does not terminate the
430
+ licenses of parties who have received copies or rights from you under
431
+ this License. If your rights have been terminated and not permanently
432
+ reinstated, you do not qualify to receive new licenses for the same
433
+ material under section 10.
434
+
435
+ 9. Acceptance Not Required for Having Copies.
436
+
437
+ You are not required to accept this License in order to receive or
438
+ run a copy of the Program. Ancillary propagation of a covered work
439
+ occurring solely as a consequence of using peer-to-peer transmission
440
+ to receive a copy likewise does not require acceptance. However,
441
+ nothing other than this License grants you permission to propagate or
442
+ modify any covered work. These actions infringe copyright if you do
443
+ not accept this License. Therefore, by modifying or propagating a
444
+ covered work, you indicate your acceptance of this License to do so.
445
+
446
+ 10. Automatic Licensing of Downstream Recipients.
447
+
448
+ Each time you convey a covered work, the recipient automatically
449
+ receives a license from the original licensors, to run, modify and
450
+ propagate that work, subject to this License. You are not responsible
451
+ for enforcing compliance by third parties with this License.
452
+
453
+ An "entity transaction" is a transaction transferring control of an
454
+ organization, or substantially all assets of one, or subdividing an
455
+ organization, or merging organizations. If propagation of a covered
456
+ work results from an entity transaction, each party to that
457
+ transaction who receives a copy of the work also receives whatever
458
+ licenses to the work the party's predecessor in interest had or could
459
+ give under the previous paragraph, plus a right to possession of the
460
+ Corresponding Source of the work from the predecessor in interest, if
461
+ the predecessor has it or can get it with reasonable efforts.
462
+
463
+ You may not impose any further restrictions on the exercise of the
464
+ rights granted or affirmed under this License. For example, you may
465
+ not impose a license fee, royalty, or other charge for exercise of
466
+ rights granted under this License, and you may not initiate litigation
467
+ (including a cross-claim or counterclaim in a lawsuit) alleging that
468
+ any patent claim is infringed by making, using, selling, offering for
469
+ sale, or importing the Program or any portion of it.
470
+
471
+ 11. Patents.
472
+
473
+ A "contributor" is a copyright holder who authorizes use under this
474
+ License of the Program or a work on which the Program is based. The
475
+ work thus licensed is called the contributor's "contributor version".
476
+
477
+ A contributor's "essential patent claims" are all patent claims
478
+ owned or controlled by the contributor, whether already acquired or
479
+ hereafter acquired, that would be infringed by some manner, permitted
480
+ by this License, of making, using, or selling its contributor version,
481
+ but do not include claims that would be infringed only as a
482
+ consequence of further modification of the contributor version. For
483
+ purposes of this definition, "control" includes the right to grant
484
+ patent sublicenses in a manner consistent with the requirements of
485
+ this License.
486
+
487
+ Each contributor grants you a non-exclusive, worldwide, royalty-free
488
+ patent license under the contributor's essential patent claims, to
489
+ make, use, sell, offer for sale, import and otherwise run, modify and
490
+ propagate the contents of its contributor version.
491
+
492
+ In the following three paragraphs, a "patent license" is any express
493
+ agreement or commitment, however denominated, not to enforce a patent
494
+ (such as an express permission to practice a patent or covenant not to
495
+ sue for patent infringement). To "grant" such a patent license to a
496
+ party means to make such an agreement or commitment not to enforce a
497
+ patent against the party.
498
+
499
+ If you convey a covered work, knowingly relying on a patent license,
500
+ and the Corresponding Source of the work is not available for anyone
501
+ to copy, free of charge and under the terms of this License, through a
502
+ publicly available network server or other readily accessible means,
503
+ then you must either (1) cause the Corresponding Source to be so
504
+ available, or (2) arrange to deprive yourself of the benefit of the
505
+ patent license for this particular work, or (3) arrange, in a manner
506
+ consistent with the requirements of this License, to extend the patent
507
+ license to downstream recipients. "Knowingly relying" means you have
508
+ actual knowledge that, but for the patent license, your conveying the
509
+ covered work in a country, or your recipient's use of the covered work
510
+ in a country, would infringe one or more identifiable patents in that
511
+ country that you have reason to believe are valid.
512
+
513
+ If, pursuant to or in connection with a single transaction or
514
+ arrangement, you convey, or propagate by procuring conveyance of, a
515
+ covered work, and grant a patent license to some of the parties
516
+ receiving the covered work authorizing them to use, propagate, modify
517
+ or convey a specific copy of the covered work, then the patent license
518
+ you grant is automatically extended to all recipients of the covered
519
+ work and works based on it.
520
+
521
+ A patent license is "discriminatory" if it does not include within
522
+ the scope of its coverage, prohibits the exercise of, or is
523
+ conditioned on the non-exercise of one or more of the rights that are
524
+ specifically granted under this License. You may not convey a covered
525
+ work if you are a party to an arrangement with a third party that is
526
+ in the business of distributing software, under which you make payment
527
+ to the third party based on the extent of your activity of conveying
528
+ the work, and under which the third party grants, to any of the
529
+ parties who would receive the covered work from you, a discriminatory
530
+ patent license (a) in connection with copies of the covered work
531
+ conveyed by you (or copies made from those copies), or (b) primarily
532
+ for and in connection with specific products or compilations that
533
+ contain the covered work, unless you entered into that arrangement,
534
+ or that patent license was granted, prior to 28 March 2007.
535
+
536
+ Nothing in this License shall be construed as excluding or limiting
537
+ any implied license or other defenses to infringement that may
538
+ otherwise be available to you under applicable patent law.
539
+
540
+ 12. No Surrender of Others' Freedom.
541
+
542
+ If conditions are imposed on you (whether by court order, agreement or
543
+ otherwise) that contradict the conditions of this License, they do not
544
+ excuse you from the conditions of this License. If you cannot convey a
545
+ covered work so as to satisfy simultaneously your obligations under this
546
+ License and any other pertinent obligations, then as a consequence you may
547
+ not convey it at all. For example, if you agree to terms that obligate you
548
+ to collect a royalty for further conveying from those to whom you convey
549
+ the Program, the only way you could satisfy both those terms and this
550
+ License would be to refrain entirely from conveying the Program.
551
+
552
+ 13. Use with the GNU Affero General Public License.
553
+
554
+ Notwithstanding any other provision of this License, you have
555
+ permission to link or combine any covered work with a work licensed
556
+ under version 3 of the GNU Affero General Public License into a single
557
+ combined work, and to convey the resulting work. The terms of this
558
+ License will continue to apply to the part which is the covered work,
559
+ but the special requirements of the GNU Affero General Public License,
560
+ section 13, concerning interaction through a network will apply to the
561
+ combination as such.
562
+
563
+ 14. Revised Versions of this License.
564
+
565
+ The Free Software Foundation may publish revised and/or new versions of
566
+ the GNU General Public License from time to time. Such new versions will
567
+ be similar in spirit to the present version, but may differ in detail to
568
+ address new problems or concerns.
569
+
570
+ Each version is given a distinguishing version number. If the
571
+ Program specifies that a certain numbered version of the GNU General
572
+ Public License "or any later version" applies to it, you have the
573
+ option of following the terms and conditions either of that numbered
574
+ version or of any later version published by the Free Software
575
+ Foundation. If the Program does not specify a version number of the
576
+ GNU General Public License, you may choose any version ever published
577
+ by the Free Software Foundation.
578
+
579
+ If the Program specifies that a proxy can decide which future
580
+ versions of the GNU General Public License can be used, that proxy's
581
+ public statement of acceptance of a version permanently authorizes you
582
+ to choose that version for the Program.
583
+
584
+ Later license versions may give you additional or different
585
+ permissions. However, no additional obligations are imposed on any
586
+ author or copyright holder as a result of your choosing to follow a
587
+ later version.
588
+
589
+ 15. Disclaimer of Warranty.
590
+
591
+ THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
592
+ APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
593
+ HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
594
+ OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
595
+ THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
596
+ PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
597
+ IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
598
+ ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
599
+
600
+ 16. Limitation of Liability.
601
+
602
+ IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
603
+ WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
604
+ THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
605
+ GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
606
+ USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
607
+ DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
608
+ PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
609
+ EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
610
+ SUCH DAMAGES.
611
+
612
+ 17. Interpretation of Sections 15 and 16.
613
+
614
+ If the disclaimer of warranty and limitation of liability provided
615
+ above cannot be given local legal effect according to their terms,
616
+ reviewing courts shall apply local law that most closely approximates
617
+ an absolute waiver of all civil liability in connection with the
618
+ Program, unless a warranty or assumption of liability accompanies a
619
+ copy of the Program in return for a fee.
620
+
621
+ END OF TERMS AND CONDITIONS
622
+
623
+ How to Apply These Terms to Your New Programs
624
+
625
+ If you develop a new program, and you want it to be of the greatest
626
+ possible use to the public, the best way to achieve this is to make it
627
+ free software which everyone can redistribute and change under these terms.
628
+
629
+ To do so, attach the following notices to the program. It is safest
630
+ to attach them to the start of each source file to most effectively
631
+ state the exclusion of warranty; and each file should have at least
632
+ the "copyright" line and a pointer to where the full notice is found.
633
+
634
+ <one line to give the program's name and a brief idea of what it does.>
635
+ Copyright (C) <year> <name of author>
636
+
637
+ This program is free software: you can redistribute it and/or modify
638
+ it under the terms of the GNU General Public License as published by
639
+ the Free Software Foundation, either version 3 of the License, or
640
+ (at your option) any later version.
641
+
642
+ This program is distributed in the hope that it will be useful,
643
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
644
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
645
+ GNU General Public License for more details.
646
+
647
+ You should have received a copy of the GNU General Public License
648
+ along with this program. If not, see <http://www.gnu.org/licenses/>.
649
+
650
+ Also add information on how to contact you by electronic and paper mail.
651
+
652
+ If the program does terminal interaction, make it output a short
653
+ notice like this when it starts in an interactive mode:
654
+
655
+ <program> Copyright (C) <year> <name of author>
656
+ This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
657
+ This is free software, and you are welcome to redistribute it
658
+ under certain conditions; type `show c' for details.
659
+
660
+ The hypothetical commands `show w' and `show c' should show the appropriate
661
+ parts of the General Public License. Of course, your program's commands
662
+ might be different; for a GUI interface, you would use an "about box".
663
+
664
+ You should also get your employer (if you work as a programmer) or school,
665
+ if any, to sign a "copyright disclaimer" for the program, if necessary.
666
+ For more information on this, and how to apply and follow the GNU GPL, see
667
+ <http://www.gnu.org/licenses/>.
668
+
669
+ The GNU General Public License does not permit incorporating your program
670
+ into proprietary programs. If your program is a subroutine library, you
671
+ may consider it more useful to permit linking proprietary applications with
672
+ the library. If this is what you want to do, use the GNU Lesser General
673
+ Public License instead of this License. But first, please read
674
+ <http://www.gnu.org/philosophy/why-not-lgpl.html>.
Waifu2x/Loss.py ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch import nn
3
+ from torch.nn.functional import _pointwise_loss
4
+
5
+ rgb_weights = [0.29891 * 3, 0.58661 * 3, 0.11448 * 3]
6
+ # RGB have different weights
7
+ # https://github.com/nagadomi/waifu2x/blob/master/train.lua#L109
8
+ use_cuda = torch.cuda.is_available()
9
+ FloatTensor = torch.cuda.FloatTensor if use_cuda else torch.FloatTensor
10
+ LongTensor = torch.cuda.LongTensor if use_cuda else torch.LongTensor
11
+ Tensor = FloatTensor
12
+
13
+
14
+ class WeightedHuberLoss(nn.SmoothL1Loss):
15
+ def __init__(self, weights=rgb_weights):
16
+ super(WeightedHuberLoss, self).__init__(size_average=True, reduce=True)
17
+ self.weights = torch.FloatTensor(weights).view(3, 1, 1)
18
+
19
+ def forward(self, input_data, target):
20
+ diff = torch.abs(input_data - target)
21
+ z = torch.where(diff < 1, 0.5 * torch.pow(diff, 2), (diff - 0.5))
22
+ out = z * self.weights.expand_as(diff)
23
+ return out.mean()
24
+
25
+
26
+ def weighted_mse_loss(input, target, weights):
27
+ out = (input - target) ** 2
28
+ out = out * weights.expand_as(out)
29
+ loss = out.sum(0) # or sum over whatever dimensions
30
+ return loss / out.size(0)
31
+
32
+
33
+ class WeightedL1Loss(nn.SmoothL1Loss):
34
+ def __init__(self, weights=rgb_weights):
35
+ super(WeightedHuberLoss, self).__init__(size_average=True, reduce=True)
36
+ self.weights = torch.FloatTensor(weights).view(3, 1, 1)
37
+
38
+ def forward(self, input_data, target):
39
+ return self.l1_loss(input_data, target, size_average=self.size_average,
40
+ reduce=self.reduce)
41
+
42
+ def l1_loss(self, input_data, target, size_average=True, reduce=True):
43
+ return _pointwise_loss(lambda a, b: torch.abs(a - b) * self.weights.expand_as(a),
44
+ torch._C._nn.l1_loss, input_data, target, size_average, reduce)
Waifu2x/Models.py ADDED
@@ -0,0 +1,316 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import json
2
+ from collections import OrderedDict
3
+ from math import exp
4
+
5
+ from .Common import *
6
+
7
+
8
+ # +++++++++++++++++++++++++++++++++++++
9
+ # FP16 Training
10
+ # -------------------------------------
11
+ # Modified from Nvidia/Apex
12
+ # https://github.com/NVIDIA/apex/blob/master/apex/fp16_utils/fp16util.py
13
+
14
+ class tofp16(nn.Module):
15
+ def __init__(self):
16
+ super(tofp16, self).__init__()
17
+
18
+ def forward(self, input):
19
+ if input.is_cuda:
20
+ return input.half()
21
+ else: # PyTorch 1.0 doesn't support fp16 in CPU
22
+ return input.float()
23
+
24
+
25
+ def BN_convert_float(module):
26
+ if isinstance(module, torch.nn.modules.batchnorm._BatchNorm):
27
+ module.float()
28
+ for child in module.children():
29
+ BN_convert_float(child)
30
+ return module
31
+
32
+
33
+ def network_to_half(network):
34
+ return nn.Sequential(tofp16(), BN_convert_float(network.half()))
35
+
36
+
37
+ # warnings.simplefilter('ignore')
38
+
39
+ # +++++++++++++++++++++++++++++++++++++
40
+ # DCSCN
41
+ # -------------------------------------
42
+
43
+ class DCSCN(BaseModule):
44
+ # https://github.com/jiny2001/dcscn-super-resolution
45
+ def __init__(self,
46
+ color_channel=3,
47
+ up_scale=2,
48
+ feature_layers=12,
49
+ first_feature_filters=196,
50
+ last_feature_filters=48,
51
+ reconstruction_filters=128,
52
+ up_sampler_filters=32
53
+ ):
54
+ super(DCSCN, self).__init__()
55
+ self.total_feature_channels = 0
56
+ self.total_reconstruct_filters = 0
57
+ self.upscale = up_scale
58
+
59
+ self.act_fn = nn.SELU(inplace=False)
60
+ self.feature_block = self.make_feature_extraction_block(color_channel,
61
+ feature_layers,
62
+ first_feature_filters,
63
+ last_feature_filters)
64
+
65
+ self.reconstruction_block = self.make_reconstruction_block(reconstruction_filters)
66
+ self.up_sampler = self.make_upsampler(up_sampler_filters, color_channel)
67
+ self.selu_init_params()
68
+
69
+ def selu_init_params(self):
70
+ for i in self.modules():
71
+ if isinstance(i, nn.Conv2d):
72
+ i.weight.data.normal_(0.0, 1.0 / sqrt(i.weight.numel()))
73
+ if i.bias is not None:
74
+ i.bias.data.fill_(0)
75
+
76
+ def conv_block(self, in_channel, out_channel, kernel_size):
77
+ m = OrderedDict([
78
+ # ("Padding", nn.ReplicationPad2d((kernel_size - 1) // 2)),
79
+ ('Conv2d', nn.Conv2d(in_channel, out_channel, kernel_size=kernel_size, padding=(kernel_size - 1) // 2)),
80
+ ('Activation', self.act_fn)
81
+ ])
82
+
83
+ return nn.Sequential(m)
84
+
85
+ def make_feature_extraction_block(self, color_channel, num_layers, first_filters, last_filters):
86
+ # input layer
87
+ feature_block = [("Feature 1", self.conv_block(color_channel, first_filters, 3))]
88
+ # exponential decay
89
+ # rest layers
90
+ alpha_rate = log(first_filters / last_filters) / (num_layers - 1)
91
+ filter_nums = [round(first_filters * exp(-alpha_rate * i)) for i in range(num_layers)]
92
+
93
+ self.total_feature_channels = sum(filter_nums)
94
+
95
+ layer_filters = [[filter_nums[i], filter_nums[i + 1], 3] for i in range(num_layers - 1)]
96
+
97
+ feature_block.extend([("Feature {}".format(index + 2), self.conv_block(*x))
98
+ for index, x in enumerate(layer_filters)])
99
+ return nn.Sequential(OrderedDict(feature_block))
100
+
101
+ def make_reconstruction_block(self, num_filters):
102
+ B1 = self.conv_block(self.total_feature_channels, num_filters // 2, 1)
103
+ B2 = self.conv_block(num_filters // 2, num_filters, 3)
104
+ m = OrderedDict([
105
+ ("A", self.conv_block(self.total_feature_channels, num_filters, 1)),
106
+ ("B", nn.Sequential(*[B1, B2]))
107
+ ])
108
+ self.total_reconstruct_filters = num_filters * 2
109
+ return nn.Sequential(m)
110
+
111
+ def make_upsampler(self, out_channel, color_channel):
112
+ out = out_channel * self.upscale ** 2
113
+ m = OrderedDict([
114
+ ('Conv2d_block', self.conv_block(self.total_reconstruct_filters, out, kernel_size=3)),
115
+ ('PixelShuffle', nn.PixelShuffle(self.upscale)),
116
+ ("Conv2d", nn.Conv2d(out_channel, color_channel, kernel_size=3, padding=1, bias=False))
117
+ ])
118
+
119
+ return nn.Sequential(m)
120
+
121
+ def forward(self, x):
122
+ # residual learning
123
+ lr, lr_up = x
124
+ feature = []
125
+ for layer in self.feature_block.children():
126
+ lr = layer(lr)
127
+ feature.append(lr)
128
+ feature = torch.cat(feature, dim=1)
129
+
130
+ reconstruction = [layer(feature) for layer in self.reconstruction_block.children()]
131
+ reconstruction = torch.cat(reconstruction, dim=1)
132
+
133
+ lr = self.up_sampler(reconstruction)
134
+ return lr + lr_up
135
+
136
+
137
+ # +++++++++++++++++++++++++++++++++++++
138
+ # CARN
139
+ # -------------------------------------
140
+
141
+ class CARN_Block(BaseModule):
142
+ def __init__(self, channels, kernel_size=3, padding=1, dilation=1,
143
+ groups=1, activation=nn.SELU(), repeat=3,
144
+ SEBlock=False, conv=nn.Conv2d,
145
+ single_conv_size=1, single_conv_group=1):
146
+ super(CARN_Block, self).__init__()
147
+ m = []
148
+ for i in range(repeat):
149
+ m.append(ResidualFixBlock(channels, channels, kernel_size=kernel_size, padding=padding, dilation=dilation,
150
+ groups=groups, activation=activation, conv=conv))
151
+ if SEBlock:
152
+ m.append(SpatialChannelSqueezeExcitation(channels, reduction=channels))
153
+ self.blocks = nn.Sequential(*m)
154
+ self.singles = nn.Sequential(
155
+ *[ConvBlock(channels * (i + 2), channels, kernel_size=single_conv_size,
156
+ padding=(single_conv_size - 1) // 2, groups=single_conv_group,
157
+ activation=activation, conv=conv)
158
+ for i in range(repeat)])
159
+
160
+ def forward(self, x):
161
+ c0 = x
162
+ for block, single in zip(self.blocks, self.singles):
163
+ b = block(x)
164
+ c0 = c = torch.cat([c0, b], dim=1)
165
+ x = single(c)
166
+
167
+ return x
168
+
169
+
170
+ class CARN(BaseModule):
171
+ # Fast, Accurate, and Lightweight Super-Resolution with Cascading Residual Network
172
+ # https://github.com/nmhkahn/CARN-pytorch
173
+ def __init__(self,
174
+ color_channels=3,
175
+ mid_channels=64,
176
+ scale=2,
177
+ activation=nn.SELU(),
178
+ num_blocks=3,
179
+ conv=nn.Conv2d):
180
+ super(CARN, self).__init__()
181
+
182
+ self.color_channels = color_channels
183
+ self.mid_channels = mid_channels
184
+ self.scale = scale
185
+
186
+ self.entry_block = ConvBlock(color_channels, mid_channels, kernel_size=3, padding=1, activation=activation,
187
+ conv=conv)
188
+ self.blocks = nn.Sequential(
189
+ *[CARN_Block(mid_channels, kernel_size=3, padding=1, activation=activation, conv=conv,
190
+ single_conv_size=1, single_conv_group=1)
191
+ for _ in range(num_blocks)])
192
+ self.singles = nn.Sequential(
193
+ *[ConvBlock(mid_channels * (i + 2), mid_channels, kernel_size=1, padding=0,
194
+ activation=activation, conv=conv)
195
+ for i in range(num_blocks)])
196
+
197
+ self.upsampler = UpSampleBlock(mid_channels, scale=scale, activation=activation, conv=conv)
198
+ self.exit_conv = conv(mid_channels, color_channels, kernel_size=3, padding=1)
199
+
200
+ def forward(self, x):
201
+ x = self.entry_block(x)
202
+ c0 = x
203
+ for block, single in zip(self.blocks, self.singles):
204
+ b = block(x)
205
+ c0 = c = torch.cat([c0, b], dim=1)
206
+ x = single(c)
207
+ x = self.upsampler(x)
208
+ out = self.exit_conv(x)
209
+ return out
210
+
211
+
212
+ class CARN_V2(CARN):
213
+ def __init__(self, color_channels=3, mid_channels=64,
214
+ scale=2, activation=nn.LeakyReLU(0.1),
215
+ SEBlock=True, conv=nn.Conv2d,
216
+ atrous=(1, 1, 1), repeat_blocks=3,
217
+ single_conv_size=3, single_conv_group=1):
218
+ super(CARN_V2, self).__init__(color_channels=color_channels, mid_channels=mid_channels, scale=scale,
219
+ activation=activation, conv=conv)
220
+
221
+ num_blocks = len(atrous)
222
+ m = []
223
+ for i in range(num_blocks):
224
+ m.append(CARN_Block(mid_channels, kernel_size=3, padding=1, dilation=1,
225
+ activation=activation, SEBlock=SEBlock, conv=conv, repeat=repeat_blocks,
226
+ single_conv_size=single_conv_size, single_conv_group=single_conv_group))
227
+
228
+ self.blocks = nn.Sequential(*m)
229
+
230
+ self.singles = nn.Sequential(
231
+ *[ConvBlock(mid_channels * (i + 2), mid_channels, kernel_size=single_conv_size,
232
+ padding=(single_conv_size - 1) // 2, groups=single_conv_group,
233
+ activation=activation, conv=conv)
234
+ for i in range(num_blocks)])
235
+
236
+ def forward(self, x):
237
+ x = self.entry_block(x)
238
+ c0 = x
239
+ res = x
240
+ for block, single in zip(self.blocks, self.singles):
241
+ b = block(x)
242
+ c0 = c = torch.cat([c0, b], dim=1)
243
+ x = single(c)
244
+ x = x + res
245
+ x = self.upsampler(x)
246
+ out = self.exit_conv(x)
247
+ return out
248
+
249
+
250
+ # +++++++++++++++++++++++++++++++++++++
251
+ # original Waifu2x model
252
+ # -------------------------------------
253
+
254
+
255
+ class UpConv_7(BaseModule):
256
+ # https://github.com/nagadomi/waifu2x/blob/3c46906cb78895dbd5a25c3705994a1b2e873199/lib/srcnn.lua#L311
257
+ def __init__(self):
258
+ super(UpConv_7, self).__init__()
259
+ self.act_fn = nn.LeakyReLU(0.1, inplace=False)
260
+ self.offset = 7 # because of 0 padding
261
+ from torch.nn import ZeroPad2d
262
+ self.pad = ZeroPad2d(self.offset)
263
+ m = [nn.Conv2d(3, 16, 3, 1, 0),
264
+ self.act_fn,
265
+ nn.Conv2d(16, 32, 3, 1, 0),
266
+ self.act_fn,
267
+ nn.Conv2d(32, 64, 3, 1, 0),
268
+ self.act_fn,
269
+ nn.Conv2d(64, 128, 3, 1, 0),
270
+ self.act_fn,
271
+ nn.Conv2d(128, 128, 3, 1, 0),
272
+ self.act_fn,
273
+ nn.Conv2d(128, 256, 3, 1, 0),
274
+ self.act_fn,
275
+ # in_channels, out_channels, kernel_size, stride=1, padding=0, output_padding=
276
+ nn.ConvTranspose2d(256, 3, kernel_size=4, stride=2, padding=3, bias=False)
277
+ ]
278
+ self.Sequential = nn.Sequential(*m)
279
+
280
+ def load_pre_train_weights(self, json_file):
281
+ with open(json_file) as f:
282
+ weights = json.load(f)
283
+ box = []
284
+ for i in weights:
285
+ box.append(i['weight'])
286
+ box.append(i['bias'])
287
+ own_state = self.state_dict()
288
+ for index, (name, param) in enumerate(own_state.items()):
289
+ own_state[name].copy_(torch.FloatTensor(box[index]))
290
+
291
+ def forward(self, x):
292
+ x = self.pad(x)
293
+ return self.Sequential.forward(x)
294
+
295
+
296
+
297
+ class Vgg_7(UpConv_7):
298
+ def __init__(self):
299
+ super(Vgg_7, self).__init__()
300
+ self.act_fn = nn.LeakyReLU(0.1, inplace=False)
301
+ self.offset = 7
302
+ m = [nn.Conv2d(3, 32, 3, 1, 0),
303
+ self.act_fn,
304
+ nn.Conv2d(32, 32, 3, 1, 0),
305
+ self.act_fn,
306
+ nn.Conv2d(32, 64, 3, 1, 0),
307
+ self.act_fn,
308
+ nn.Conv2d(64, 64, 3, 1, 0),
309
+ self.act_fn,
310
+ nn.Conv2d(64, 128, 3, 1, 0),
311
+ self.act_fn,
312
+ nn.Conv2d(128, 128, 3, 1, 0),
313
+ self.act_fn,
314
+ nn.Conv2d(128, 3, 3, 1, 0)
315
+ ]
316
+ self.Sequential = nn.Sequential(*m)
Waifu2x/Readme.md ADDED
@@ -0,0 +1,167 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Waifu2x
2
+
3
+ Re-implementation on the original [waifu2x](https://github.com/nagadomi/waifu2x) in PyTorch with additional super resolution models. This repo is mainly used to explore interesting super resolution models. User-friendly tools may not be available now ><.
4
+
5
+ ## Dependencies
6
+ * Python 3x
7
+ * [PyTorch](https://pytorch.org/) >= 1 ( > 0.41 shall also work, but not guarantee)
8
+ * [Nvidia/Apex](https://github.com/NVIDIA/apex/) (used for mixed precision training, you may use the [python codes](https://github.com/NVIDIA/apex/tree/master/apex/fp16_utils) directly)
9
+
10
+ Optinal: Nvidia GPU. Model inference (32 fp only) can run in cpu only.
11
+
12
+ ## What's New
13
+ * Add [CARN Model (Fast, Accurate, and Lightweight Super-Resolution with Cascading Residual Network)](https://github.com/nmhkahn/CARN-pytorch). Model Codes are adapted from the authors's [github repo](https://github.com/nmhkahn/CARN-pytorch). I add [Spatial Channel Squeeze Excitation](https://arxiv.org/abs/1709.01507) and swap all 1x1 convolution with 3x3 standard convolutions. The model is trained in fp 16 with Nvidia's [apex](https://github.com/NVIDIA/apex). Details and plots on model variant can be found in [docs/CARN](./docs/CARN)
14
+
15
+ * Dilated Convolution seems less effective (if not make the model worse) in super resolution, though it brings some improvement in image segmentation, especially when dilated rate increases and then decreases. Further investigation is needed.
16
+
17
+ ## How to Use
18
+ Compare the input image and upscaled image
19
+ ```python
20
+ from utils.prepare_images import *
21
+ from Models import *
22
+ from torchvision.utils import save_image
23
+ model_cran_v2 = CARN_V2(color_channels=3, mid_channels=64, conv=nn.Conv2d,
24
+ single_conv_size=3, single_conv_group=1,
25
+ scale=2, activation=nn.LeakyReLU(0.1),
26
+ SEBlock=True, repeat_blocks=3, atrous=(1, 1, 1))
27
+
28
+ model_cran_v2 = network_to_half(model_cran_v2)
29
+ checkpoint = "model_check_points/CRAN_V2/CARN_model_checkpoint.pt"
30
+ model_cran_v2.load_state_dict(torch.load(checkpoint, 'cpu'))
31
+ # if use GPU, then comment out the next line so it can use fp16.
32
+ model_cran_v2 = model_cran_v2.float()
33
+
34
+ demo_img = "input_image.png"
35
+ img = Image.open(demo_img).convert("RGB")
36
+
37
+ # origin
38
+ img_t = to_tensor(img).unsqueeze(0)
39
+
40
+ # used to compare the origin
41
+ img = img.resize((img.size[0] // 2, img.size[1] // 2), Image.BICUBIC)
42
+
43
+ # overlapping split
44
+ # if input image is too large, then split it into overlapped patches
45
+ # details can be found at [here](https://github.com/nagadomi/waifu2x/issues/238)
46
+ img_splitter = ImageSplitter(seg_size=64, scale_factor=2, boarder_pad_size=3)
47
+ img_patches = img_splitter.split_img_tensor(img, scale_method=None, img_pad=0)
48
+ with torch.no_grad():
49
+ out = [model_cran_v2(i) for i in img_patches]
50
+ img_upscale = img_splitter.merge_img_tensor(out)
51
+
52
+ final = torch.cat([img_t, img_upscale])
53
+ save_image(final, 'out.png', nrow=2)
54
+ ```
55
+
56
+ ## Training
57
+
58
+ If possible, fp16 training is preferred because it is much faster with minimal quality decrease.
59
+
60
+ Sample training script is available in `train.py`, but you may need to change some liens.
61
+
62
+ ### Image Processing
63
+ Original images are all at least 3k x 3K. I downsample them by LANCZOS so that one side has at most 2048, then I randomly cut them into 256x256 patches as target and use 128x128 with jpeg noise as input images. All input patches have at least 14 kb, and they are stored in SQLite with BLOB format. SQlite seems to have [better performance](https://www.sqlite.org/intern-v-extern-blob.html) than file system for small objects. H5 file format may not be optimal because of its larger size.
64
+
65
+ Although convolutions can take in any sizes of images, the content of image matters. For real life images, small patches may maintain color,brightness, etc variances in small regions, but for digital drawn images, colors are added in block areas. A small patch may end up showing entirely one color, and the model has little to learn.
66
+
67
+ For example, the following two plots come from CARN and have the same settings, including initial parameters. Both training loss and ssim are lower for 64x64, but they perform worse in test time compared to 128x128.
68
+
69
+ ![loss](docs/CARN/plots/128_vs_64_model_loss.png)
70
+ ![ssim](docs/CARN/plots/128_vs_64_model_ssim.png)
71
+
72
+
73
+ Downsampling methods are uniformly chosen among ```[PIL.Image.BILINEAR, PIL.Image.BICUBIC, PIL.Image.LANCZOS]``` , so different patches in the same image might be down-scaled in different ways.
74
+
75
+ Image noise are from JPEG format only. They are added by re-encoding PNG images into PIL's JPEG data with various quality. Noise level 1 means quality ranges uniformly from [75, 95]; level 2 means quality ranges uniformly from [50, 75].
76
+
77
+
78
+ ## Models
79
+ Models are tuned and modified with extra features.
80
+
81
+
82
+ * [DCSCN 12](https://github.com/jiny2001/dcscn-super-resolution)
83
+
84
+ * [CRAN](https://github.com/nmhkahn/CARN-pytorch)
85
+
86
+ #### From [Waifu2x](https://github.com/nagadomi/waifu2x)
87
+ * [Upconv7](https://github.com/nagadomi/waifu2x/blob/7d156917ae1113ab847dab15c75db7642231e7fa/lib/srcnn.lua#L360)
88
+
89
+ * [Vgg_7](https://github.com/nagadomi/waifu2x/blob/7d156917ae1113ab847dab15c75db7642231e7fa/lib/srcnn.lua#L334)
90
+
91
+ * [Cascaded Residual U-Net with SEBlock](https://github.com/nagadomi/waifu2x/blob/7d156917ae1113ab847dab15c75db7642231e7fa/lib/srcnn.lua#L514) (PyTorch codes are not available and under testing)
92
+
93
+ #### Models Comparison
94
+ Images are from [Key: サマボケ(Summer Pocket)](http://key.visualarts.gr.jp/summer/).
95
+
96
+ The left column is the original image, and the right column is bicubic, DCSCN, CRAN_V2
97
+
98
+ ![img](docs/demo_bicubic_model_comparison.png)
99
+
100
+
101
+ ![img](docs/demo_true_bicubic_dcscn_upconv.png)
102
+
103
+
104
+
105
+ ##### Scores
106
+ The list will be updated after I add more models.
107
+
108
+ Images are twitter icons (PNG) from [Key: サマボケ(Summer Pocket)](http://key.visualarts.gr.jp/summer/). They are cropped into non-overlapping 96x96 patches and down-scaled by 2. Then images are re-encoded into JPEG format with quality from [75, 95]. Scores are PSNR and MS-SSIM.
109
+
110
+ | | Total Parameters | BICUBIC | Random* |
111
+ | :---: | :---: | :---: | :---: |
112
+ | CRAN V2| 2,149,607 | 34.0985 (0.9924) | 34.0509 (0.9922) |
113
+ | DCSCN 12 |1,889,974 | 31.5358 (0.9851) | 31.1457 (0.9834) |
114
+ | Upconv 7| 552,480| 31.4566 (0.9788) | 30.9492 (0.9772) |
115
+
116
+ *uniformly select down scale methods from Image.BICUBIC, Image.BILINEAR, Image.LANCZOS.
117
+
118
+
119
+
120
+
121
+
122
+ #### DCSCN
123
+ [Fast and Accurate Image Super Resolution by Deep CNN with Skip Connection and Network in Network](https://github.com/jiny2001/dcscn-super-resolution#fast-and-accurate-image-super-resolution-by-deep-cnn-with-skip-connection-and-network-in-network)
124
+
125
+ DCSCN is very interesting as it has relatively quick forward computation, and both the shallow model (layerr 8) and deep model (layer 12) are quick to train. The settings are different from the paper.
126
+
127
+ * I use exponential decay to decrease the number of feature filters in each layer. [Here](https://github.com/jiny2001/dcscn-super-resolution/blob/a868775930c6b36922897b0203468f3f1481e935/DCSCN.py#L204) is the original filter decay method.
128
+
129
+ * I also increase the reconstruction filters from 48 to 128.
130
+
131
+ * All activations are replaced by SELU. Dropout and weight decay are not added neither because they significantly increase the training time.
132
+
133
+ * The loss function is changed from MSE to L1.
134
+ According to [Loss Functions for Image Restoration with Neural
135
+ Networks](https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=4&cad=rja&uact=8&ved=0ahUKEwi7kuGt_7_bAhXrqVQKHRqhCcUQFghUMAM&url=http%3A%2F%2Fresearch.nvidia.com%2Fsites%2Fdefault%2Ffiles%2Fpubs%2F2017-03_Loss-Functions-for%2Fcomparison_tci.pdf&usg=AOvVaw1p0ndOKRH2ZaEsumO7d_bA), L1 seems to be more robust and converges faster than MSE. But the authors find the results from L1 and MSE are [similar](https://github.com/jiny2001/dcscn-super-resolution/issues/29).
136
+
137
+
138
+ I need to thank jiny2001 (one of the paper's author) to test the difference of SELU and PRELU. SELU seems more stable and has fewer parameters to train. It is a good drop in replacement
139
+ >layers=8, filters=96 and dataset=yang91+bsd200.
140
+ ![](docs/DCSCN_comparison/selu_prelu.png)
141
+ The details can be found in [here]( https://github.com/jiny2001/dcscn-super-resolution/issues/29).
142
+
143
+
144
+
145
+ A pre-trained 12-layer model as well as model parameters are available. The model run time is around 3-5 times of Waifu2x. The output quality is usually visually indistinguishable, but its PSNR and SSIM are bit higher. Though, such comparison is not fair since the 12-layer model has around 1,889,974 parameters, 5 times more than waifu2x's Upconv_7 model.
146
+
147
+ #### CARN
148
+ Channels are set to 64 across all blocks, so residual adds are very effective. Increase the channels to 128 lower the loss curve a little bit but doubles the total parameters from 0.9 Millions to 3 Millions. 32 Channels has much worse performance. Increasing the number of cascaded blocks from 3 to 5 doesn't lower the loss a lot.
149
+
150
+ SE Blocks seems to have the most obvious improvement without increasing the computation a lot. Partial based padding seems have little effect if not decrease the quality. Atrous convolution is slower about 10%-20% than normal convolution in Pytorch 1.0, but there are no obvious improvement.
151
+
152
+ Another more effective model is to add upscaled input image to the final convolution. A simple bilinear upscaled image seems sufficient.
153
+
154
+ More examples on model configurations can be found in [docs/CARN folder](./docs/CARN/carn_plot_loss.md)
155
+
156
+ ![img](docs/CARN/plots/CARN_Compare.png)
157
+
158
+ ![img](docs/CARN/plots/CARN_Compare_Res_Add.png)
159
+
160
+ ### Waifu2x Original Models
161
+ Models can load waifu2x's pre-trained weights. The function ```forward_checkpoint``` sets the ```nn.LeakyReLU``` to compute data inplace.
162
+
163
+ #### Upconv_7
164
+ Original waifu2x's model. PyTorch's implementation with cpu only is around 5 times longer for large images. The output images have very close PSNR and SSIM scores compared to images generated from the [caffe version](https://github.com/lltcggie/waifu2x-caffe) , thought they are not identical.
165
+
166
+ #### Vgg_7
167
+ Not tested yet, but it is ready to use.
Waifu2x/__init__.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ # -*- coding: utf-8 -*-
2
+ # file: __init__.py
3
+ # time: 05/12/2022
4
+ # author: yangheng <hy345@exeter.ac.uk>
5
+ # github: https://github.com/yangheng95
6
+ # huggingface: https://huggingface.co/yangheng
7
+ # google scholar: https://scholar.google.com/citations?user=NPq5a_0AAAAJ&hl=en
8
+ # Copyright (C) 2021. All Rights Reserved.
9
+ from .magnify import ImageMagnifier
Waifu2x/magnify.py ADDED
@@ -0,0 +1,86 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # -*- coding: utf-8 -*-
2
+ # file: test.py
3
+ # time: 05/12/2022
4
+ # author: yangheng <hy345@exeter.ac.uk>
5
+ # github: https://github.com/yangheng95
6
+ # huggingface: https://huggingface.co/yangheng
7
+ # google scholar: https://scholar.google.com/citations?user=NPq5a_0AAAAJ&hl=en
8
+ # Copyright (C) 2021. All Rights Reserved.
9
+ from pathlib import Path
10
+ from typing import Union
11
+
12
+ import autocuda
13
+ import findfile
14
+ from pyabsa.utils.pyabsa_utils import fprint
15
+ from torchvision import transforms
16
+ from .utils.prepare_images import *
17
+ from .Models import *
18
+
19
+
20
+ class ImageMagnifier:
21
+
22
+ def __init__(self):
23
+ self.device = autocuda.auto_cuda()
24
+ self.model_cran_v2 = CARN_V2(color_channels=3, mid_channels=64, conv=nn.Conv2d,
25
+ single_conv_size=3, single_conv_group=1,
26
+ scale=2, activation=nn.LeakyReLU(0.1),
27
+ SEBlock=True, repeat_blocks=3, atrous=(1, 1, 1))
28
+
29
+ self.model_cran_v2 = network_to_half(self.model_cran_v2)
30
+ self.checkpoint = findfile.find_cwd_file("CARN_model_checkpoint.pt")
31
+ self.model_cran_v2.load_state_dict(torch.load(self.checkpoint, map_location='cpu'))
32
+ # if use GPU, then comment out the next line so it can use fp16.
33
+ self.model_cran_v2 = self.model_cran_v2.float().to(self.device)
34
+ self.model_cran_v2.to(self.device)
35
+
36
+ def __image_scale(self, img, scale_factor: int = 2):
37
+ img_splitter = ImageSplitter(seg_size=64, scale_factor=scale_factor, boarder_pad_size=3)
38
+ img_patches = img_splitter.split_img_tensor(img, scale_method=None, img_pad=0)
39
+ with torch.no_grad():
40
+ if self.device != 'cpu':
41
+ with torch.cuda.amp.autocast():
42
+ out = [self.model_cran_v2(i.to(self.device)) for i in img_patches]
43
+ else:
44
+ with torch.cpu.amp.autocast():
45
+ out = [self.model_cran_v2(i) for i in img_patches]
46
+ img_upscale = img_splitter.merge_img_tensor(out)
47
+
48
+ final = torch.cat([img_upscale])
49
+
50
+ return transforms.ToPILImage()(final[0])
51
+
52
+ def magnify(self, img, scale_factor: int = 2):
53
+ fprint("scale factor reset to:", scale_factor//2*2)
54
+ _scale_factor = scale_factor
55
+ while _scale_factor // 2 > 0:
56
+ img = self.__image_scale(img, scale_factor=2)
57
+ _scale_factor = _scale_factor // 2
58
+ return img
59
+
60
+ def magnify_from_file(self, img_path: Union[str, Path], scale_factor: int = 2, save_img: bool = True):
61
+
62
+ if not os.path.exists(img_path):
63
+ raise FileNotFoundError("Path is not found.")
64
+ if os.path.isfile(img_path):
65
+ try:
66
+ img = Image.open(img_path)
67
+ img = self.magnify(img, scale_factor)
68
+ if save_img:
69
+ img.save(os.path.join(img_path))
70
+ except Exception as e:
71
+ fprint(img_path, e)
72
+ fprint(img_path, "Done.")
73
+
74
+ elif os.path.isdir(img_path):
75
+ for path in os.listdir(img_path):
76
+ try:
77
+ img = Image.open(os.path.join(img_path, path))
78
+ img = self.magnify(img, scale_factor)
79
+ if save_img:
80
+ img.save(os.path.join(img_path, path))
81
+ except Exception as e:
82
+ fprint(path, e)
83
+ continue
84
+ fprint(path, "Done.")
85
+ else:
86
+ raise TypeError("Path is not a file or directory.")
Waifu2x/model_check_points/CRAN_V2/CARN_adam_checkpoint.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:292f2be9ea173861e4a7f6cf580f04fe9a1fc6c78fdac6f182cbc051ea50791e
3
+ size 31734614
Waifu2x/model_check_points/CRAN_V2/CARN_model_checkpoint.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3a0ef4fdea5b0b2a91b05b7a39fe63df3e27e1d18df477065e7f268a7feaaad2
3
+ size 4329165
Waifu2x/model_check_points/CRAN_V2/CARN_scheduler_last_iter.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ba2302e523d32bfeb9b542a9dc6aa5ecdb45babc793892153245d6c69ae23433
3
+ size 151
Waifu2x/model_check_points/CRAN_V2/CRAN_V2_02_28_2019.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b74e163d829f6f587e3fdb0b645342e494416accb1962cf0973354de5ec157ea
3
+ size 49895595
Waifu2x/model_check_points/CRAN_V2/ReadME.md ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Model Specifications
2
+
3
+
4
+ ```python
5
+ model_cran_v2 = CARN_V2(color_channels=3, mid_channels=64, conv=nn.Conv2d,
6
+ single_conv_size=3, single_conv_group=1,
7
+ scale=2, activation=nn.LeakyReLU(0.1),
8
+ SEBlock=True, repeat_blocks=3, atrous=(1, 1, 1))
9
+
10
+ model_cran_v2 = network_to_half(model_cran_v2)
11
+ checkpoint = "CARN_model_checkpoint.pt"
12
+ model_cran_v2.load_state_dict(torch.load(checkpoint, 'cpu'))
13
+ model_cran_v2 = model_cran_v2.float() # if use cpu
14
+
15
+ ````
16
+
17
+ To use pre-trained model for training
18
+
19
+ ```python
20
+
21
+ model = CARN_V2(color_channels=3, mid_channels=64, conv=nn.Conv2d,
22
+ single_conv_size=3, single_conv_group=1,
23
+ scale=2, activation=nn.LeakyReLU(0.1),
24
+ SEBlock=True, repeat_blocks=3, atrous=(1, 1, 1))
25
+
26
+ model = network_to_half(model)
27
+ model = model.cuda()
28
+ model.load_state_dict(torch.load("CARN_model_checkpoint.pt"))
29
+
30
+ learning_rate = 1e-4
31
+ weight_decay = 1e-6
32
+ optimizer = optim.Adam(model.parameters(), lr=learning_rate, weight_decay=weight_decay, amsgrad=True)
33
+ optimizer = FP16_Optimizer(optimizer, static_loss_scale=128.0, verbose=False)
34
+ optimizer.load_state_dict(torch.load("CARN_adam_checkpoint.pt"))
35
+
36
+ last_iter = torch.load("CARN_scheduler_last_iter") # -1 if start from new
37
+ scheduler = CyclicLR(optimizer.optimizer, base_lr=1e-4, max_lr=4e-4,
38
+ step_size=3 * total_batch, mode="triangular",
39
+ last_batch_iteration=last_iter)
40
+
41
+ ```
Waifu2x/model_check_points/CRAN_V2/test_loss.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:93f644a6a3f6636035980855f56ef3dbc8784679371b06b81e0e4d06067c142d
3
+ size 43507
Waifu2x/model_check_points/CRAN_V2/test_psnr.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ae8f8d1a3d175e76dcbcdcf0cede898e8f2cf169f3eec14eeb75a4e19d8e2d6b
3
+ size 42563
Waifu2x/model_check_points/CRAN_V2/test_ssim.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:763ff936f536b12b37b351c09f3c1290fb2188399aea3d9ce3cf069bd0d135e7
3
+ size 43515
Waifu2x/model_check_points/CRAN_V2/train_loss.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:85a86e94cd689adff04c4b22bf2534d17aa52af5e7309a82bc2a4f5c6c144900
3
+ size 15564175
Waifu2x/model_check_points/CRAN_V2/train_psnr.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2d1e88646b74a054ddf20ba41368a01162e35d9c88ac72f392a6ba08a5c7ef3b
3
+ size 15564175
Waifu2x/model_check_points/CRAN_V2/train_ssim.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6b8da8bc73f64997c5b2d15d6161b11dbd172258a62c88572c032feb73bd022b
3
+ size 15564175
Waifu2x/train.py ADDED
@@ -0,0 +1,174 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ from torch import optim
3
+ from torch.utils.data import DataLoader
4
+ from torchvision.utils import save_image
5
+ from tqdm import trange
6
+
7
+ from Dataloader import *
8
+ from .utils import image_quality
9
+ from .utils.cls import CyclicLR
10
+ from .utils.prepare_images import *
11
+
12
+ train_folder = './dataset/train'
13
+ test_folder = "./dataset/test"
14
+
15
+ img_dataset = ImageDBData(db_file='dataset/images.db', db_table="train_images_size_128_noise_1_rgb", max_images=24)
16
+ img_data = DataLoader(img_dataset, batch_size=6, shuffle=True, num_workers=6)
17
+
18
+ total_batch = len(img_data)
19
+ print(len(img_dataset))
20
+
21
+ test_dataset = ImageDBData(db_file='dataset/test2.db', db_table="test_images_size_128_noise_1_rgb", max_images=None)
22
+ num_test = len(test_dataset)
23
+ test_data = DataLoader(test_dataset, batch_size=1, shuffle=False, num_workers=1)
24
+
25
+ criteria = nn.L1Loss()
26
+
27
+ model = CARN_V2(color_channels=3, mid_channels=64, conv=nn.Conv2d,
28
+ single_conv_size=3, single_conv_group=1,
29
+ scale=2, activation=nn.LeakyReLU(0.1),
30
+ SEBlock=True, repeat_blocks=3, atrous=(1, 1, 1))
31
+
32
+ model.total_parameters()
33
+
34
+
35
+ # model.initialize_weights_xavier_uniform()
36
+
37
+ # fp16 training is available in GPU only
38
+ model = network_to_half(model)
39
+ model = model.cuda()
40
+ model.load_state_dict(torch.load("CARN_model_checkpoint.pt"))
41
+
42
+ learning_rate = 1e-4
43
+ weight_decay = 1e-6
44
+ optimizer = optim.Adam(model.parameters(), lr=learning_rate, weight_decay=weight_decay, amsgrad=True)
45
+ # optimizer = optim.SGD(model.parameters(), momentum=0.9, nesterov=True, weight_decay=weight_decay, lr=learning_rate)
46
+
47
+ # optimizer = FP16_Optimizer(optimizer, static_loss_scale=128.0, verbose=False)
48
+ # optimizer.load_state_dict(torch.load("CARN_adam_checkpoint.pt"))
49
+
50
+ last_iter = -1 # torch.load("CARN_scheduler_last_iter")
51
+ scheduler = CyclicLR(optimizer, base_lr=1e-4, max_lr=1e-4,
52
+ step_size=3 * total_batch, mode="triangular",
53
+ last_batch_iteration=last_iter)
54
+ train_loss = []
55
+ train_ssim = []
56
+ train_psnr = []
57
+
58
+ test_loss = []
59
+ test_ssim = []
60
+ test_psnr = []
61
+
62
+ # train_loss = torch.load("train_loss.pt")
63
+ # train_ssim = torch.load("train_ssim.pt")
64
+ # train_psnr = torch.load("train_psnr.pt")
65
+ #
66
+ # test_loss = torch.load("test_loss.pt")
67
+ # test_ssim = torch.load("test_ssim.pt")
68
+ # test_psnr = torch.load("test_psnr.pt")
69
+
70
+
71
+ counter = 0
72
+ iteration = 2
73
+ ibar = trange(iteration, ascii=True, maxinterval=1, postfix={"avg_loss": 0, "train_ssim": 0, "test_ssim": 0})
74
+ for i in ibar:
75
+ # batch_loss = []
76
+ # insample_ssim = []
77
+ # insample_psnr = []
78
+ for index, batch in enumerate(img_data):
79
+ scheduler.batch_step()
80
+ lr_img, hr_img = batch
81
+ lr_img = lr_img.cuda().half()
82
+ hr_img = hr_img.cuda()
83
+
84
+ # model.zero_grad()
85
+ optimizer.zero_grad()
86
+ outputs = model.forward(lr_img)
87
+ outputs = outputs.float()
88
+ loss = criteria(outputs, hr_img)
89
+ # loss.backward()
90
+ optimizer.backward(loss)
91
+ # nn.utils.clip_grad_norm_(model.parameters(), 5)
92
+ optimizer.step()
93
+
94
+ counter += 1
95
+ # train_loss.append(loss.item())
96
+
97
+ ssim = image_quality.msssim(outputs, hr_img).item()
98
+ psnr = image_quality.psnr(outputs, hr_img).item()
99
+
100
+ ibar.set_postfix(ratio=index / total_batch, loss=loss.item(),
101
+ ssim=ssim, batch=index,
102
+ psnr=psnr,
103
+ lr=scheduler.current_lr
104
+ )
105
+ train_loss.append(loss.item())
106
+ train_ssim.append(ssim)
107
+ train_psnr.append(psnr)
108
+
109
+ # +++++++++++++++++++++++++++++++++++++
110
+ # save checkpoints by iterations
111
+ # -------------------------------------
112
+
113
+ if (counter + 1) % 500 == 0:
114
+ torch.save(model.state_dict(), 'CARN_model_checkpoint.pt')
115
+ torch.save(optimizer.state_dict(), 'CARN_adam_checkpoint.pt')
116
+ torch.save(train_loss, 'train_loss.pt')
117
+ torch.save(train_ssim, "train_ssim.pt")
118
+ torch.save(train_psnr, 'train_psnr.pt')
119
+ torch.save(scheduler.last_batch_iteration, "CARN_scheduler_last_iter.pt")
120
+
121
+ # +++++++++++++++++++++++++++++++++++++
122
+ # End of One Epoch
123
+ # -------------------------------------
124
+
125
+ # one_ite_loss = np.mean(batch_loss)
126
+ # one_ite_ssim = np.mean(insample_ssim)
127
+ # one_ite_psnr = np.mean(insample_psnr)
128
+
129
+ # print(f"One iteration loss {one_ite_loss}, ssim {one_ite_ssim}, psnr {one_ite_psnr}")
130
+ # train_loss.append(one_ite_loss)
131
+ # train_ssim.append(one_ite_ssim)
132
+ # train_psnr.append(one_ite_psnr)
133
+
134
+ torch.save(model.state_dict(), 'CARN_model_checkpoint.pt')
135
+ # torch.save(scheduler, "CARN_scheduler_optim.pt")
136
+ torch.save(optimizer.state_dict(), 'CARN_adam_checkpoint.pt')
137
+ torch.save(train_loss, 'train_loss.pt')
138
+ torch.save(train_ssim, "train_ssim.pt")
139
+ torch.save(train_psnr, 'train_psnr.pt')
140
+ # torch.save(scheduler.last_batch_iteration, "CARN_scheduler_last_iter.pt")
141
+
142
+ # +++++++++++++++++++++++++++++++++++++
143
+ # Test
144
+ # -------------------------------------
145
+
146
+ with torch.no_grad():
147
+ ssim = []
148
+ batch_loss = []
149
+ psnr = []
150
+ for index, test_batch in enumerate(test_data):
151
+ lr_img, hr_img = test_batch
152
+ lr_img = lr_img.cuda()
153
+ hr_img = hr_img.cuda()
154
+
155
+ lr_img_up = model(lr_img)
156
+ lr_img_up = lr_img_up.float()
157
+ loss = criteria(lr_img_up, hr_img)
158
+
159
+ save_image([lr_img_up[0], hr_img[0]], f"check_test_imgs/{index}.png")
160
+ batch_loss.append(loss.item())
161
+ ssim.append(image_quality.msssim(lr_img_up, hr_img).item())
162
+ psnr.append(image_quality.psnr(lr_img_up, hr_img).item())
163
+
164
+ test_ssim.append(np.mean(ssim))
165
+ test_loss.append(np.mean(batch_loss))
166
+ test_psnr.append(np.mean(psnr))
167
+
168
+ torch.save(test_loss, 'test_loss.pt')
169
+ torch.save(test_ssim, "test_ssim.pt")
170
+ torch.save(test_psnr, "test_psnr.pt")
171
+
172
+ # import subprocess
173
+
174
+ # subprocess.call(["shutdown", "/s"])
Waifu2x/utils/Img_to_H5.py ADDED
@@ -0,0 +1,50 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import glob
2
+
3
+ import h5py
4
+ from PIL import Image
5
+ from torchvision.transforms import RandomCrop
6
+ from torchvision.transforms.functional import to_tensor
7
+ from tqdm import tqdm
8
+
9
+ from Dataloader import ImageAugment
10
+
11
+ patch_size = 128
12
+ shrink_size = 2
13
+ noise_level = 1
14
+ patches_per_img = 20
15
+ images = glob.glob("dataset/train/*")
16
+
17
+ database = h5py.File("train_images.hdf5", 'w')
18
+
19
+ dat_group = database.create_group("shrink_2_noise_level_1_downsample_random_rgb")
20
+ # del database['shrink_2_noise_level_1_downsample_random']
21
+ storage_lr = dat_group.create_dataset("train_lr", shape=(patches_per_img * len(images), 3,
22
+ patch_size // shrink_size,
23
+ patch_size // shrink_size),
24
+ dtype='float32',
25
+ # compression='lzf',
26
+ )
27
+ storage_hr = dat_group.create_dataset("train_hr", shape=(patches_per_img * len(images), 3,
28
+ patch_size, patch_size),
29
+ # compression='lzf',
30
+ dtype='float32')
31
+
32
+ random_cropper = RandomCrop(size=patch_size)
33
+ img_augmenter = ImageAugment(shrink_size, noise_level, down_sample_method=None)
34
+
35
+
36
+ def get_img_patches(img_pil):
37
+ img_patch = random_cropper(img_pil)
38
+ lr_hr_patches = img_augmenter.process(img_patch)
39
+ return lr_hr_patches
40
+
41
+
42
+ counter = 0
43
+ for img in tqdm(images):
44
+ img_pil = Image.open(img).convert("RGB")
45
+ for i in range(patches_per_img):
46
+ patch = get_img_patches(img_pil)
47
+ storage_lr[counter] = to_tensor(patch[0].convert("RGB")).numpy()
48
+ storage_hr[counter] = to_tensor(patch[1].convert("RGB")).numpy()
49
+ counter += 1
50
+ database.close()
Waifu2x/utils/__init__.py ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ # -*- coding: utf-8 -*-
2
+ # file: __init__.py
3
+ # time: 05/12/2022
4
+ # author: yangheng <hy345@exeter.ac.uk>
5
+ # github: https://github.com/yangheng95
6
+ # huggingface: https://huggingface.co/yangheng
7
+ # google scholar: https://scholar.google.com/citations?user=NPq5a_0AAAAJ&hl=en
8
+ # Copyright (C) 2021. All Rights Reserved.
Waifu2x/utils/cls.py ADDED
@@ -0,0 +1,157 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # This code is copied from https://github.com/thomasjpfan/pytorch/blob/401ec389db2c9d2978917a6e4d1101b20340d7e7/torch/optim/lr_scheduler.py
2
+
3
+
4
+ # This code is under review at PyTorch and is to be merged eventually to make CLR available to all.
5
+ # Tested with pytorch 0.2.0
6
+
7
+ import numpy as np
8
+
9
+
10
+ class CyclicLR(object):
11
+ """Sets the learning rate of each parameter group according to
12
+ cyclical learning rate policy (CLR). The policy cycles the learning
13
+ rate between two boundaries with a constant frequency, as detailed in
14
+ the paper `Cyclical Learning Rates for Training Neural Networks`_.
15
+ The distance between the two boundaries can be scaled on a per-iteration
16
+ or per-cycle basis.
17
+ Cyclical learning rate policy changes the learning rate after every batch.
18
+ `batch_step` should be called after a batch has been used for training.
19
+ To resume training, save `last_batch_iteration` and use it to instantiate `CycleLR`.
20
+ This class has three built-in policies, as put forth in the paper:
21
+ "triangular":
22
+ A basic triangular cycle w/ no amplitude scaling.
23
+ "triangular2":
24
+ A basic triangular cycle that scales initial amplitude by half each cycle.
25
+ "exp_range":
26
+ A cycle that scales initial amplitude by gamma**(cycle iterations) at each
27
+ cycle iteration.
28
+ This implementation was adapted from the github repo: `bckenstler/CLR`_
29
+ Args:
30
+ optimizer (Optimizer): Wrapped optimizer.
31
+ base_lr (float or list): Initial learning rate which is the
32
+ lower boundary in the cycle for eachparam groups.
33
+ Default: 0.001
34
+ max_lr (float or list): Upper boundaries in the cycle for
35
+ each parameter group. Functionally,
36
+ it defines the cycle amplitude (max_lr - base_lr).
37
+ The lr at any cycle is the sum of base_lr
38
+ and some scaling of the amplitude; therefore
39
+ max_lr may not actually be reached depending on
40
+ scaling function. Default: 0.006
41
+ step_size (int): Number of training iterations per
42
+ half cycle. Authors suggest setting step_size
43
+ 2-8 x training iterations in epoch. Default: 2000
44
+ mode (str): One of {triangular, triangular2, exp_range}.
45
+ Values correspond to policies detailed above.
46
+ If scale_fn is not None, this argument is ignored.
47
+ Default: 'triangular'
48
+ gamma (float): Constant in 'exp_range' scaling function:
49
+ gamma**(cycle iterations)
50
+ Default: 1.0
51
+ scale_fn (function): Custom scaling policy defined by a single
52
+ argument lambda function, where
53
+ 0 <= scale_fn(x) <= 1 for all x >= 0.
54
+ mode paramater is ignored
55
+ Default: None
56
+ scale_mode (str): {'cycle', 'iterations'}.
57
+ Defines whether scale_fn is evaluated on
58
+ cycle number or cycle iterations (training
59
+ iterations since start of cycle).
60
+ Default: 'cycle'
61
+ last_batch_iteration (int): The index of the last batch. Default: -1
62
+ Example:
63
+ >>> optimizer = torch.optim.SGD(model.parameters(), lr=0.1, momentum=0.9)
64
+ >>> scheduler = torch.optim.CyclicLR(optimizer)
65
+ >>> data_loader = torch.utils.data.DataLoader(...)
66
+ >>> for epoch in range(10):
67
+ >>> for batch in data_loader:
68
+ >>> scheduler.batch_step()
69
+ >>> train_batch(...)
70
+ .. _Cyclical Learning Rates for Training Neural Networks: https://arxiv.org/abs/1506.01186
71
+ .. _bckenstler/CLR: https://github.com/bckenstler/CLR
72
+ """
73
+
74
+ def __init__(self, optimizer, base_lr=1e-3, max_lr=6e-3,
75
+ step_size=2000, mode='triangular', gamma=1.,
76
+ scale_fn=None, scale_mode='cycle', last_batch_iteration=-1):
77
+
78
+ # if not isinstance(optimizer, Optimizer):
79
+ # raise TypeError('{} is not an Optimizer'.format(
80
+ # type(optimizer).__name__))
81
+ self.optimizer = optimizer
82
+
83
+ if isinstance(base_lr, list) or isinstance(base_lr, tuple):
84
+ if len(base_lr) != len(optimizer.param_groups):
85
+ raise ValueError("expected {} base_lr, got {}".format(
86
+ len(optimizer.param_groups), len(base_lr)))
87
+ self.base_lrs = list(base_lr)
88
+ else:
89
+ self.base_lrs = [base_lr] * len(optimizer.param_groups)
90
+
91
+ if isinstance(max_lr, list) or isinstance(max_lr, tuple):
92
+ if len(max_lr) != len(optimizer.param_groups):
93
+ raise ValueError("expected {} max_lr, got {}".format(
94
+ len(optimizer.param_groups), len(max_lr)))
95
+ self.max_lrs = list(max_lr)
96
+ else:
97
+ self.max_lrs = [max_lr] * len(optimizer.param_groups)
98
+
99
+ self.step_size = step_size
100
+
101
+ if mode not in ['triangular', 'triangular2', 'exp_range'] \
102
+ and scale_fn is None:
103
+ raise ValueError('mode is invalid and scale_fn is None')
104
+
105
+ self.mode = mode
106
+ self.gamma = gamma
107
+ self.current_lr = None
108
+
109
+ if scale_fn is None:
110
+ if self.mode == 'triangular':
111
+ self.scale_fn = self._triangular_scale_fn
112
+ self.scale_mode = 'cycle'
113
+ elif self.mode == 'triangular2':
114
+ self.scale_fn = self._triangular2_scale_fn
115
+ self.scale_mode = 'cycle'
116
+ elif self.mode == 'exp_range':
117
+ self.scale_fn = self._exp_range_scale_fn
118
+ self.scale_mode = 'iterations'
119
+ else:
120
+ self.scale_fn = scale_fn
121
+ self.scale_mode = scale_mode
122
+
123
+ self.batch_step(last_batch_iteration + 1)
124
+ self.last_batch_iteration = last_batch_iteration
125
+
126
+ def batch_step(self, batch_iteration=None):
127
+ if batch_iteration is None:
128
+ batch_iteration = self.last_batch_iteration + 1
129
+ self.last_batch_iteration = batch_iteration
130
+ for param_group, lr in zip(self.optimizer.param_groups, self.get_lr()):
131
+ param_group['lr'] = lr
132
+ self.current_lr = lr
133
+
134
+ def _triangular_scale_fn(self, x):
135
+ return 1.
136
+
137
+ def _triangular2_scale_fn(self, x):
138
+ return 1 / (2. ** (x - 1))
139
+
140
+ def _exp_range_scale_fn(self, x):
141
+ return self.gamma ** (x)
142
+
143
+ def get_lr(self):
144
+ step_size = float(self.step_size)
145
+ cycle = np.floor(1 + self.last_batch_iteration / (2 * step_size))
146
+ x = np.abs(self.last_batch_iteration / step_size - 2 * cycle + 1)
147
+
148
+ lrs = []
149
+ param_lrs = zip(self.optimizer.param_groups, self.base_lrs, self.max_lrs)
150
+ for param_group, base_lr, max_lr in param_lrs:
151
+ base_height = (max_lr - base_lr) * np.maximum(0, (1 - x))
152
+ if self.scale_mode == 'cycle':
153
+ lr = base_lr + base_height * self.scale_fn(cycle)
154
+ else:
155
+ lr = base_lr + base_height * self.scale_fn(self.last_batch_iteration)
156
+ lrs.append(lr)
157
+ return lrs
Waifu2x/utils/image_quality.py ADDED
@@ -0,0 +1,173 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Pytorch Multi-Scale Structural Similarity Index (SSIM)
2
+ # This code is written by jorge-pessoa (https://github.com/jorge-pessoa/pytorch-msssim)
3
+ # MIT licence
4
+ import math
5
+ from math import exp
6
+
7
+ import torch
8
+ import torch.nn.functional as F
9
+ from torch.autograd import Variable
10
+
11
+
12
+ # +++++++++++++++++++++++++++++++++++++
13
+ # SSIM
14
+ # -------------------------------------
15
+
16
+ def gaussian(window_size, sigma):
17
+ gauss = torch.Tensor([exp(-(x - window_size // 2) ** 2 / float(2 * sigma ** 2)) for x in range(window_size)])
18
+ return gauss / gauss.sum()
19
+
20
+
21
+ def create_window(window_size, channel):
22
+ _1D_window = gaussian(window_size, 1.5).unsqueeze(1)
23
+ _2D_window = _1D_window.mm(_1D_window.t()).float().unsqueeze(0).unsqueeze(0)
24
+ window = Variable(_2D_window.expand(channel, 1, window_size, window_size).contiguous())
25
+ return window
26
+
27
+
28
+ def _ssim(img1, img2, window, window_size, channel, size_average=True, full=False):
29
+ padd = 0
30
+
31
+ mu1 = F.conv2d(img1, window, padding=padd, groups=channel)
32
+ mu2 = F.conv2d(img2, window, padding=padd, groups=channel)
33
+
34
+ mu1_sq = mu1.pow(2)
35
+ mu2_sq = mu2.pow(2)
36
+ mu1_mu2 = mu1 * mu2
37
+
38
+ sigma1_sq = F.conv2d(img1 * img1, window, padding=padd, groups=channel) - mu1_sq
39
+ sigma2_sq = F.conv2d(img2 * img2, window, padding=padd, groups=channel) - mu2_sq
40
+ sigma12 = F.conv2d(img1 * img2, window, padding=padd, groups=channel) - mu1_mu2
41
+
42
+ C1 = 0.01 ** 2
43
+ C2 = 0.03 ** 2
44
+
45
+ ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2))
46
+
47
+ v1 = 2.0 * sigma12 + C2
48
+ v2 = sigma1_sq + sigma2_sq + C2
49
+ cs = torch.mean(v1 / v2)
50
+
51
+ if size_average:
52
+ ret = ssim_map.mean()
53
+ else:
54
+ ret = ssim_map.mean(1).mean(1).mean(1)
55
+
56
+ if full:
57
+ return ret, cs
58
+ return ret
59
+
60
+
61
+ class SSIM(torch.nn.Module):
62
+ def __init__(self, window_size=11, size_average=True):
63
+ super(SSIM, self).__init__()
64
+ self.window_size = window_size
65
+ self.size_average = size_average
66
+ self.channel = 1
67
+ self.window = create_window(window_size, self.channel)
68
+
69
+ def forward(self, img1, img2):
70
+ (_, channel, _, _) = img1.size()
71
+
72
+ if channel == self.channel and self.window.data.type() == img1.data.type():
73
+ window = self.window
74
+ else:
75
+ window = create_window(self.window_size, channel)
76
+
77
+ if img1.is_cuda:
78
+ window = window.cuda(img1.get_device())
79
+ window = window.type_as(img1)
80
+
81
+ self.window = window
82
+ self.channel = channel
83
+
84
+ return _ssim(img1, img2, window, self.window_size, channel, self.size_average)
85
+
86
+
87
+ def ssim(img1, img2, window_size=11, size_average=True, full=False):
88
+ (_, channel, height, width) = img1.size()
89
+
90
+ real_size = min(window_size, height, width)
91
+ window = create_window(real_size, channel)
92
+
93
+ if img1.is_cuda:
94
+ window = window.cuda(img1.get_device())
95
+ window = window.type_as(img1)
96
+
97
+ return _ssim(img1, img2, window, real_size, channel, size_average, full=full)
98
+
99
+
100
+ def msssim(img1, img2, window_size=11, size_average=True):
101
+ # TODO: fix NAN results
102
+ if img1.size() != img2.size():
103
+ raise RuntimeError('Input images must have the same shape (%s vs. %s).' %
104
+ (img1.size(), img2.size()))
105
+ if len(img1.size()) != 4:
106
+ raise RuntimeError('Input images must have four dimensions, not %d' %
107
+ len(img1.size()))
108
+
109
+ weights = torch.tensor([0.0448, 0.2856, 0.3001, 0.2363, 0.1333], dtype=img1.dtype)
110
+ if img1.is_cuda:
111
+ weights = weights.cuda(img1.get_device())
112
+
113
+ levels = weights.size()[0]
114
+ mssim = []
115
+ mcs = []
116
+ for _ in range(levels):
117
+ sim, cs = ssim(img1, img2, window_size=window_size, size_average=size_average, full=True)
118
+ mssim.append(sim)
119
+ mcs.append(cs)
120
+
121
+ img1 = F.avg_pool2d(img1, (2, 2))
122
+ img2 = F.avg_pool2d(img2, (2, 2))
123
+
124
+ mssim = torch.stack(mssim)
125
+ mcs = torch.stack(mcs)
126
+ return (torch.prod(mcs[0:levels - 1] ** weights[0:levels - 1]) *
127
+ (mssim[levels - 1] ** weights[levels - 1]))
128
+
129
+
130
+ class MSSSIM(torch.nn.Module):
131
+ def __init__(self, window_size=11, size_average=True, channel=3):
132
+ super(MSSSIM, self).__init__()
133
+ self.window_size = window_size
134
+ self.size_average = size_average
135
+ self.channel = channel
136
+
137
+ def forward(self, img1, img2):
138
+ # TODO: store window between calls if possible
139
+ return msssim(img1, img2, window_size=self.window_size, size_average=self.size_average)
140
+
141
+
142
+ def calc_psnr(sr, hr, scale=0, benchmark=False):
143
+ # adapt from EDSR: https://github.com/thstkdgus35/EDSR-PyTorch
144
+ diff = (sr - hr).data
145
+ if benchmark:
146
+ shave = scale
147
+ if diff.size(1) > 1:
148
+ convert = diff.new(1, 3, 1, 1)
149
+ convert[0, 0, 0, 0] = 65.738
150
+ convert[0, 1, 0, 0] = 129.057
151
+ convert[0, 2, 0, 0] = 25.064
152
+ diff.mul_(convert).div_(256)
153
+ diff = diff.sum(dim=1, keepdim=True)
154
+ else:
155
+ shave = scale + 6
156
+
157
+ valid = diff[:, :, shave:-shave, shave:-shave]
158
+ mse = valid.pow(2).mean()
159
+
160
+ return -10 * math.log10(mse)
161
+
162
+
163
+ # +++++++++++++++++++++++++++++++++++++
164
+ # PSNR
165
+ # -------------------------------------
166
+ from torch import nn
167
+
168
+
169
+ def psnr(predict, target):
170
+ with torch.no_grad():
171
+ criteria = nn.MSELoss()
172
+ mse = criteria(predict, target)
173
+ return -10 * torch.log10(mse)
Waifu2x/utils/prepare_images.py ADDED
@@ -0,0 +1,120 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import copy
2
+ import glob
3
+ import os
4
+ from multiprocessing.dummy import Pool as ThreadPool
5
+
6
+ from PIL import Image
7
+ from torchvision.transforms.functional import to_tensor
8
+
9
+ from ..Models import *
10
+
11
+
12
+ class ImageSplitter:
13
+ # key points:
14
+ # Boarder padding and over-lapping img splitting to avoid the instability of edge value
15
+ # Thanks Waifu2x's autorh nagadomi for suggestions (https://github.com/nagadomi/waifu2x/issues/238)
16
+
17
+ def __init__(self, seg_size=48, scale_factor=2, boarder_pad_size=3):
18
+ self.seg_size = seg_size
19
+ self.scale_factor = scale_factor
20
+ self.pad_size = boarder_pad_size
21
+ self.height = 0
22
+ self.width = 0
23
+ self.upsampler = nn.Upsample(scale_factor=scale_factor, mode='bilinear')
24
+
25
+ def split_img_tensor(self, pil_img, scale_method=Image.BILINEAR, img_pad=0):
26
+ # resize image and convert them into tensor
27
+ img_tensor = to_tensor(pil_img).unsqueeze(0)
28
+ img_tensor = nn.ReplicationPad2d(self.pad_size)(img_tensor)
29
+ batch, channel, height, width = img_tensor.size()
30
+ self.height = height
31
+ self.width = width
32
+
33
+ if scale_method is not None:
34
+ img_up = pil_img.resize((2 * pil_img.size[0], 2 * pil_img.size[1]), scale_method)
35
+ img_up = to_tensor(img_up).unsqueeze(0)
36
+ img_up = nn.ReplicationPad2d(self.pad_size * self.scale_factor)(img_up)
37
+
38
+ patch_box = []
39
+ # avoid the residual part is smaller than the padded size
40
+ if height % self.seg_size < self.pad_size or width % self.seg_size < self.pad_size:
41
+ self.seg_size += self.scale_factor * self.pad_size
42
+
43
+ # split image into over-lapping pieces
44
+ for i in range(self.pad_size, height, self.seg_size):
45
+ for j in range(self.pad_size, width, self.seg_size):
46
+ part = img_tensor[:, :,
47
+ (i - self.pad_size):min(i + self.pad_size + self.seg_size, height),
48
+ (j - self.pad_size):min(j + self.pad_size + self.seg_size, width)]
49
+ if img_pad > 0:
50
+ part = nn.ZeroPad2d(img_pad)(part)
51
+ if scale_method is not None:
52
+ # part_up = self.upsampler(part)
53
+ part_up = img_up[:, :,
54
+ self.scale_factor * (i - self.pad_size):min(i + self.pad_size + self.seg_size,
55
+ height) * self.scale_factor,
56
+ self.scale_factor * (j - self.pad_size):min(j + self.pad_size + self.seg_size,
57
+ width) * self.scale_factor]
58
+
59
+ patch_box.append((part, part_up))
60
+ else:
61
+ patch_box.append(part)
62
+ return patch_box
63
+
64
+ def merge_img_tensor(self, list_img_tensor):
65
+ out = torch.zeros((1, 3, self.height * self.scale_factor, self.width * self.scale_factor))
66
+ img_tensors = copy.copy(list_img_tensor)
67
+ rem = self.pad_size * 2
68
+
69
+ pad_size = self.scale_factor * self.pad_size
70
+ seg_size = self.scale_factor * self.seg_size
71
+ height = self.scale_factor * self.height
72
+ width = self.scale_factor * self.width
73
+ for i in range(pad_size, height, seg_size):
74
+ for j in range(pad_size, width, seg_size):
75
+ part = img_tensors.pop(0)
76
+ part = part[:, :, rem:-rem, rem:-rem]
77
+ # might have error
78
+ if len(part.size()) > 3:
79
+ _, _, p_h, p_w = part.size()
80
+ out[:, :, i:i + p_h, j:j + p_w] = part
81
+ # out[:,:,
82
+ # self.scale_factor*i:self.scale_factor*i+p_h,
83
+ # self.scale_factor*j:self.scale_factor*j+p_w] = part
84
+ out = out[:, :, rem:-rem, rem:-rem]
85
+ return out
86
+
87
+
88
+ def load_single_image(img_file,
89
+ up_scale=False,
90
+ up_scale_factor=2,
91
+ up_scale_method=Image.BILINEAR,
92
+ zero_padding=False):
93
+ img = Image.open(img_file).convert("RGB")
94
+ out = to_tensor(img).unsqueeze(0)
95
+ if zero_padding:
96
+ out = nn.ZeroPad2d(zero_padding)(out)
97
+ if up_scale:
98
+ size = tuple(map(lambda x: x * up_scale_factor, img.size))
99
+ img_up = img.resize(size, up_scale_method)
100
+ img_up = to_tensor(img_up).unsqueeze(0)
101
+ out = (out, img_up)
102
+
103
+ return out
104
+
105
+
106
+ def standardize_img_format(img_folder):
107
+ def process(img_file):
108
+ img_path = os.path.dirname(img_file)
109
+ img_name, _ = os.path.basename(img_file).split(".")
110
+ out = os.path.join(img_path, img_name + ".JPEG")
111
+ os.rename(img_file, out)
112
+
113
+ list_imgs = []
114
+ for i in ['png', "jpeg", 'jpg']:
115
+ list_imgs.extend(glob.glob(img_folder + "**/*." + i, recursive=True))
116
+ print("Found {} images.".format(len(list_imgs)))
117
+ pool = ThreadPool(4)
118
+ pool.map(process, list_imgs)
119
+ pool.close()
120
+ pool.join()
app.py ADDED
@@ -0,0 +1,344 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import random
3
+
4
+ import autocuda
5
+ from pyabsa.utils.pyabsa_utils import fprint
6
+
7
+ from diffusers import AutoencoderKL, UNet2DConditionModel, StableDiffusionPipeline, StableDiffusionImg2ImgPipeline, \
8
+ DPMSolverMultistepScheduler
9
+ import gradio as gr
10
+ import torch
11
+ from PIL import Image
12
+ import utils
13
+ import datetime
14
+ import time
15
+ import psutil
16
+ from Waifu2x.magnify import ImageMagnifier
17
+
18
+ magnifier = ImageMagnifier()
19
+
20
+ start_time = time.time()
21
+ is_colab = utils.is_google_colab()
22
+
23
+ CUDA_VISIBLE_DEVICES = ''
24
+ device = autocuda.auto_cuda()
25
+
26
+ dtype = torch.float16 if device != 'cpu' else torch.float32
27
+
28
+
29
+ class Model:
30
+ def __init__(self, name, path="", prefix=""):
31
+ self.name = name
32
+ self.path = path
33
+ self.prefix = prefix
34
+ self.pipe_t2i = None
35
+ self.pipe_i2i = None
36
+
37
+
38
+ models = [
39
+ Model("anything v3", "Linaqruf/anything-v3.0", "anything v3 style"),
40
+ ]
41
+ # Model("Spider-Verse", "nitrosocke/spider-verse-diffusion", "spiderverse style "),
42
+ # Model("Balloon Art", "Fictiverse/Stable_Diffusion_BalloonArt_Model", "BalloonArt "),
43
+ # Model("Elden Ring", "nitrosocke/elden-ring-diffusion", "elden ring style "),
44
+ # Model("Tron Legacy", "dallinmackay/Tron-Legacy-diffusion", "trnlgcy ")
45
+ # Model("Pokémon", "lambdalabs/sd-pokemon-diffusers", ""),
46
+ # Model("Pony Diffusion", "AstraliteHeart/pony-diffusion", ""),
47
+ # Model("Robo Diffusion", "nousr/robo-diffusion", ""),
48
+
49
+ scheduler = DPMSolverMultistepScheduler(
50
+ beta_start=0.00085,
51
+ beta_end=0.012,
52
+ beta_schedule="scaled_linear",
53
+ num_train_timesteps=1000,
54
+ trained_betas=None,
55
+ predict_epsilon=True,
56
+ thresholding=False,
57
+ algorithm_type="dpmsolver++",
58
+ solver_type="midpoint",
59
+ lower_order_final=True,
60
+ )
61
+
62
+ custom_model = None
63
+ if is_colab:
64
+ models.insert(0, Model("Custom model"))
65
+ custom_model = models[0]
66
+
67
+ last_mode = "txt2img"
68
+ current_model = models[1] if is_colab else models[0]
69
+ current_model_path = current_model.path
70
+
71
+ if is_colab:
72
+ pipe = StableDiffusionPipeline.from_pretrained(current_model.path, torch_dtype=dtype, scheduler=scheduler,
73
+ safety_checker=lambda images, clip_input: (images, False))
74
+
75
+ else: # download all models
76
+ print(f"{datetime.datetime.now()} Downloading vae...")
77
+ vae = AutoencoderKL.from_pretrained(current_model.path, subfolder="vae", torch_dtype=dtype)
78
+ for model in models:
79
+ try:
80
+ print(f"{datetime.datetime.now()} Downloading {model.name} model...")
81
+ unet = UNet2DConditionModel.from_pretrained(model.path, subfolder="unet", torch_dtype=dtype)
82
+ model.pipe_t2i = StableDiffusionPipeline.from_pretrained(model.path, unet=unet, vae=vae,
83
+ torch_dtype=dtype, scheduler=scheduler,
84
+ safety_checker=None)
85
+ model.pipe_i2i = StableDiffusionImg2ImgPipeline.from_pretrained(model.path, unet=unet, vae=vae,
86
+ torch_dtype=dtype,
87
+ scheduler=scheduler, safety_checker=None)
88
+ except Exception as e:
89
+ print(f"{datetime.datetime.now()} Failed to load model " + model.name + ": " + str(e))
90
+ models.remove(model)
91
+ pipe = models[0].pipe_t2i
92
+
93
+ # model.pipe_i2i = torch.compile(model.pipe_i2i)
94
+ # model.pipe_t2i = torch.compile(model.pipe_t2i)
95
+ if torch.cuda.is_available():
96
+ pipe = pipe.to(device)
97
+
98
+
99
+ # device = "GPU 🔥" if torch.cuda.is_available() else "CPU 🥶"
100
+
101
+
102
+ def error_str(error, title="Error"):
103
+ return f"""#### {title}
104
+ {error}""" if error else ""
105
+
106
+
107
+ def custom_model_changed(path):
108
+ models[0].path = path
109
+ global current_model
110
+ current_model = models[0]
111
+
112
+
113
+ def on_model_change(model_name):
114
+ prefix = "Enter prompt. \"" + next((m.prefix for m in models if m.name == model_name),
115
+ None) + "\" is prefixed automatically" if model_name != models[
116
+ 0].name else "Don't forget to use the custom model prefix in the prompt!"
117
+
118
+ return gr.update(visible=model_name == models[0].name), gr.update(placeholder=prefix)
119
+
120
+
121
+ def inference(model_name, prompt, guidance, steps, width=512, height=512, seed=0, img=None, strength=0.5,
122
+ neg_prompt="", scale_factor=2):
123
+ fprint(psutil.virtual_memory()) # print memory usage
124
+ prompt = 'detailed fingers, beautiful hands,' + prompt
125
+ fprint(f"Prompt: {prompt}")
126
+ global current_model
127
+ for model in models:
128
+ if model.name == model_name:
129
+ current_model = model
130
+ model_path = current_model.path
131
+
132
+ generator = torch.Generator(device).manual_seed(seed) if seed != 0 else None
133
+
134
+ try:
135
+ if img is not None:
136
+ return img_to_img(model_path, prompt, neg_prompt, img, strength, guidance, steps, width, height,
137
+ generator, scale_factor), None
138
+ else:
139
+ return txt_to_img(model_path, prompt, neg_prompt, guidance, steps, width, height, generator,
140
+ scale_factor), None
141
+ except Exception as e:
142
+ return None, error_str(e)
143
+ # if img is not None:
144
+ # return img_to_img(model_path, prompt, neg_prompt, img, strength, guidance, steps, width, height,
145
+ # generator, scale_factor), None
146
+ # else:
147
+ # return txt_to_img(model_path, prompt, neg_prompt, guidance, steps, width, height, generator, scale_factor), None
148
+
149
+
150
+ def txt_to_img(model_path, prompt, neg_prompt, guidance, steps, width, height, generator, scale_factor):
151
+ print(f"{datetime.datetime.now()} txt_to_img, model: {current_model.name}")
152
+
153
+ global last_mode
154
+ global pipe
155
+ global current_model_path
156
+ if model_path != current_model_path or last_mode != "txt2img":
157
+ current_model_path = model_path
158
+
159
+ if is_colab or current_model == custom_model:
160
+ pipe = StableDiffusionPipeline.from_pretrained(current_model_path, torch_dtype=dtype,
161
+ scheduler=scheduler,
162
+ safety_checker=lambda images, clip_input: (images, False))
163
+ else:
164
+ # pipe = pipe.to("cpu")
165
+ pipe = current_model.pipe_t2i
166
+
167
+ if torch.cuda.is_available():
168
+ pipe = pipe.to(device)
169
+ last_mode = "txt2img"
170
+
171
+ prompt = current_model.prefix + prompt
172
+ result = pipe(
173
+ prompt,
174
+ negative_prompt=neg_prompt,
175
+ # num_images_per_prompt=n_images,
176
+ num_inference_steps=int(steps),
177
+ guidance_scale=guidance,
178
+ width=width,
179
+ height=height,
180
+ generator=generator)
181
+ result.images[0] = magnifier.magnify(result.images[0], scale_factor=scale_factor)
182
+
183
+ # save image
184
+ result.images[0].save("imgs/result-{}.png".format(datetime.datetime.now().strftime("%Y%m%d-%H%M%S")))
185
+ return replace_nsfw_images(result)
186
+
187
+
188
+ def img_to_img(model_path, prompt, neg_prompt, img, strength, guidance, steps, width, height, generator, scale_factor):
189
+ fprint(f"{datetime.datetime.now()} img_to_img, model: {model_path}")
190
+
191
+ global last_mode
192
+ global pipe
193
+ global current_model_path
194
+ if model_path != current_model_path or last_mode != "img2img":
195
+ current_model_path = model_path
196
+
197
+ if is_colab or current_model == custom_model:
198
+ pipe = StableDiffusionImg2ImgPipeline.from_pretrained(current_model_path, torch_dtype=dtype,
199
+ scheduler=scheduler,
200
+ safety_checker=lambda images, clip_input: (
201
+ images, False))
202
+ else:
203
+ # pipe = pipe.to("cpu")
204
+ pipe = current_model.pipe_i2i
205
+
206
+ if torch.cuda.is_available():
207
+ pipe = pipe.to(device)
208
+ last_mode = "img2img"
209
+
210
+ prompt = current_model.prefix + prompt
211
+ ratio = min(height / img.height, width / img.width)
212
+ img = img.resize((int(img.width * ratio), int(img.height * ratio)), Image.LANCZOS)
213
+ result = pipe(
214
+ prompt,
215
+ negative_prompt=neg_prompt,
216
+ # num_images_per_prompt=n_images,
217
+ image=img,
218
+ num_inference_steps=int(steps),
219
+ strength=strength,
220
+ guidance_scale=guidance,
221
+ # width=width,
222
+ # height=height,
223
+ generator=generator)
224
+ result.images[0] = magnifier.magnify(result.images[0], scale_factor=scale_factor)
225
+
226
+ # save image
227
+ result.images[0].save("imgs/result-{}.png".format(datetime.datetime.now().strftime("%Y%m%d-%H%M%S")))
228
+ return replace_nsfw_images(result)
229
+
230
+
231
+ def replace_nsfw_images(results):
232
+ if is_colab:
233
+ return results.images[0]
234
+ if hasattr(results, "nsfw_content_detected") and results.nsfw_content_detected:
235
+ for i in range(len(results.images)):
236
+ if results.nsfw_content_detected[i]:
237
+ results.images[i] = Image.open("nsfw.png")
238
+ return results.images[0]
239
+
240
+
241
+ css = """.finetuned-diffusion-div div{display:inline-flex;align-items:center;gap:.8rem;font-size:1.75rem}.finetuned-diffusion-div div h1{font-weight:900;margin-bottom:7px}.finetuned-diffusion-div p{margin-bottom:10px;font-size:94%}a{text-decoration:underline}.tabs{margin-top:0;margin-bottom:0}#gallery{min-height:20rem}
242
+ """
243
+ with gr.Blocks(css=css) as demo:
244
+ if not os.path.exists('imgs'):
245
+ os.mkdir('imgs')
246
+
247
+ gr.Markdown('# Super Resolution Anime Diffusion')
248
+ gr.Markdown(
249
+ "## Author: [yangheng95](https://github.com/yangheng95) Github:[Github](https://github.com/yangheng95/SuperResolutionAnimeDiffusion)")
250
+ gr.Markdown("### This demo is running on a CPU, so it will take at least 20 minutes. "
251
+ "If you have a GPU, you can clone from [Github](https://github.com/yangheng95/SuperResolutionAnimeDiffusion) and run it locally.")
252
+ gr.Markdown("### FYI: to generate a 512*512 image and magnify 4x, it only takes 5~8 seconds on a RTX 2080 GPU")
253
+ gr.Markdown(
254
+ "### You can duplicate this demo on HuggingFace Spaces, click [here](https://huggingface.co/spaces/yangheng/Super-Resolution-Anime-Diffusion?duplicate=true)")
255
+
256
+ with gr.Row():
257
+ with gr.Column(scale=55):
258
+ with gr.Group():
259
+ gr.Markdown("Text to image")
260
+
261
+ model_name = gr.Dropdown(label="Model", choices=[m.name for m in models], value=current_model.name)
262
+
263
+ with gr.Box(visible=False) as custom_model_group:
264
+ custom_model_path = gr.Textbox(label="Custom model path",
265
+ placeholder="Path to model, e.g. nitrosocke/Arcane-Diffusion",
266
+ interactive=True)
267
+ gr.HTML(
268
+ "<div><font size='2'>Custom models have to be downloaded first, so give it some time.</font></div>")
269
+
270
+ with gr.Row():
271
+ prompt = gr.Textbox(label="Prompt", show_label=False, max_lines=2,
272
+ placeholder="Enter prompt. Style applied automatically").style(container=False)
273
+ with gr.Row():
274
+ generate = gr.Button(value="Generate")
275
+
276
+ with gr.Row():
277
+ with gr.Group():
278
+ neg_prompt = gr.Textbox(label="Negative prompt", placeholder="What to exclude from the image")
279
+
280
+ image_out = gr.Image(height=512)
281
+ # gallery = gr.Gallery(
282
+ # label="Generated images", show_label=False, elem_id="gallery"
283
+ # ).style(grid=[1], height="auto")
284
+ error_output = gr.Markdown()
285
+
286
+ with gr.Column(scale=45):
287
+ with gr.Group():
288
+ gr.Markdown("Image to Image")
289
+
290
+ with gr.Row():
291
+ with gr.Group():
292
+ image = gr.Image(label="Image", height=256, tool="editor", type="pil")
293
+ strength = gr.Slider(label="Transformation strength", minimum=0, maximum=1, step=0.01,
294
+ value=0.5)
295
+
296
+ with gr.Row():
297
+ with gr.Group():
298
+ # n_images = gr.Slider(label="Images", value=1, minimum=1, maximum=4, step=1)
299
+
300
+ with gr.Row():
301
+ guidance = gr.Slider(label="Guidance scale", value=7.5, maximum=15)
302
+ steps = gr.Slider(label="Steps", value=15, minimum=2, maximum=75, step=1)
303
+
304
+ with gr.Row():
305
+ width = gr.Slider(label="Width", value=512, minimum=64, maximum=1024, step=8)
306
+ height = gr.Slider(label="Height", value=512, minimum=64, maximum=1024, step=8)
307
+ with gr.Row():
308
+ scale_factor = gr.Slider(1, 8, label='Scale factor (to magnify image) (1, 2, 4, 8)',
309
+ value=2,
310
+ step=1)
311
+
312
+ seed = gr.Slider(0, 2147483647, label='Seed (0 = random)', value=0, step=1)
313
+
314
+ if is_colab:
315
+ model_name.change(on_model_change, inputs=model_name, outputs=[custom_model_group, prompt], queue=False)
316
+ custom_model_path.change(custom_model_changed, inputs=custom_model_path, outputs=None)
317
+ # n_images.change(lambda n: gr.Gallery().style(grid=[2 if n > 1 else 1], height="auto"), inputs=n_images, outputs=gallery)
318
+
319
+ gr.Markdown("### based on [Anything V3](https://huggingface.co/Linaqruf/anything-v3.0)")
320
+
321
+ inputs = [model_name, prompt, guidance, steps, width, height, seed, image, strength, neg_prompt, scale_factor]
322
+ outputs = [image_out, error_output]
323
+ prompt.submit(inference, inputs=inputs, outputs=outputs)
324
+ generate.click(inference, inputs=inputs, outputs=outputs, api_name="generate")
325
+
326
+ prompt_keys = [
327
+ 'girl', 'lovely', 'cute', 'beautiful eyes', 'cumulonimbus clouds', 'detailed fingers',
328
+ random.choice(['dress']),
329
+ random.choice(['white hair']),
330
+ random.choice(['blue eyes']),
331
+ random.choice(['flower meadow']),
332
+ random.choice(['Elif', 'Angel'])
333
+ ]
334
+ prompt.value = ','.join(prompt_keys)
335
+ ex = gr.Examples([
336
+ [models[0].name, prompt.value, 7.5, 15],
337
+
338
+ ], inputs=[model_name, prompt, guidance, steps, seed], outputs=outputs, fn=inference, cache_examples=False)
339
+
340
+ print(f"Space built in {time.time() - start_time:.2f} seconds")
341
+
342
+ if not is_colab:
343
+ demo.queue(concurrency_count=2)
344
+ demo.launch(debug=is_colab, enable_queue=True, share=is_colab)
diffusers/__init__.py ADDED
@@ -0,0 +1,123 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from .utils import (
2
+ is_flax_available,
3
+ is_inflect_available,
4
+ is_onnx_available,
5
+ is_scipy_available,
6
+ is_torch_available,
7
+ is_transformers_available,
8
+ is_unidecode_available,
9
+ )
10
+
11
+
12
+ __version__ = "0.10.0.dev0"
13
+
14
+ from .configuration_utils import ConfigMixin
15
+ from .onnx_utils import OnnxRuntimeModel
16
+ from .utils import logging
17
+
18
+
19
+ if is_torch_available():
20
+ from .modeling_utils import ModelMixin
21
+ from .models import AutoencoderKL, Transformer2DModel, UNet1DModel, UNet2DConditionModel, UNet2DModel, VQModel
22
+ from .optimization import (
23
+ get_constant_schedule,
24
+ get_constant_schedule_with_warmup,
25
+ get_cosine_schedule_with_warmup,
26
+ get_cosine_with_hard_restarts_schedule_with_warmup,
27
+ get_linear_schedule_with_warmup,
28
+ get_polynomial_decay_schedule_with_warmup,
29
+ get_scheduler,
30
+ )
31
+ from .pipeline_utils import DiffusionPipeline
32
+ from .pipelines import (
33
+ DanceDiffusionPipeline,
34
+ DDIMPipeline,
35
+ DDPMPipeline,
36
+ KarrasVePipeline,
37
+ LDMPipeline,
38
+ LDMSuperResolutionPipeline,
39
+ PNDMPipeline,
40
+ RePaintPipeline,
41
+ ScoreSdeVePipeline,
42
+ )
43
+ from .schedulers import (
44
+ DDIMScheduler,
45
+ DDPMScheduler,
46
+ DPMSolverMultistepScheduler,
47
+ EulerAncestralDiscreteScheduler,
48
+ EulerDiscreteScheduler,
49
+ HeunDiscreteScheduler,
50
+ IPNDMScheduler,
51
+ KarrasVeScheduler,
52
+ KDPM2AncestralDiscreteScheduler,
53
+ KDPM2DiscreteScheduler,
54
+ PNDMScheduler,
55
+ RePaintScheduler,
56
+ SchedulerMixin,
57
+ ScoreSdeVeScheduler,
58
+ VQDiffusionScheduler,
59
+ )
60
+ from .training_utils import EMAModel
61
+ else:
62
+ from .utils.dummy_pt_objects import * # noqa F403
63
+
64
+ if is_torch_available() and is_scipy_available():
65
+ from .schedulers import LMSDiscreteScheduler
66
+ else:
67
+ from .utils.dummy_torch_and_scipy_objects import * # noqa F403
68
+
69
+ if is_torch_available() and is_transformers_available():
70
+ from .pipelines import (
71
+ AltDiffusionImg2ImgPipeline,
72
+ AltDiffusionPipeline,
73
+ CycleDiffusionPipeline,
74
+ LDMTextToImagePipeline,
75
+ StableDiffusionImageVariationPipeline,
76
+ StableDiffusionImg2ImgPipeline,
77
+ StableDiffusionInpaintPipeline,
78
+ StableDiffusionInpaintPipelineLegacy,
79
+ StableDiffusionPipeline,
80
+ StableDiffusionPipelineSafe,
81
+ StableDiffusionUpscalePipeline,
82
+ VersatileDiffusionDualGuidedPipeline,
83
+ VersatileDiffusionImageVariationPipeline,
84
+ VersatileDiffusionPipeline,
85
+ VersatileDiffusionTextToImagePipeline,
86
+ VQDiffusionPipeline,
87
+ )
88
+ else:
89
+ from .utils.dummy_torch_and_transformers_objects import * # noqa F403
90
+
91
+ if is_torch_available() and is_transformers_available() and is_onnx_available():
92
+ from .pipelines import (
93
+ OnnxStableDiffusionImg2ImgPipeline,
94
+ OnnxStableDiffusionInpaintPipeline,
95
+ OnnxStableDiffusionInpaintPipelineLegacy,
96
+ OnnxStableDiffusionPipeline,
97
+ StableDiffusionOnnxPipeline,
98
+ )
99
+ else:
100
+ from .utils.dummy_torch_and_transformers_and_onnx_objects import * # noqa F403
101
+
102
+ if is_flax_available():
103
+ from .modeling_flax_utils import FlaxModelMixin
104
+ from .models.unet_2d_condition_flax import FlaxUNet2DConditionModel
105
+ from .models.vae_flax import FlaxAutoencoderKL
106
+ from .pipeline_flax_utils import FlaxDiffusionPipeline
107
+ from .schedulers import (
108
+ FlaxDDIMScheduler,
109
+ FlaxDDPMScheduler,
110
+ FlaxDPMSolverMultistepScheduler,
111
+ FlaxKarrasVeScheduler,
112
+ FlaxLMSDiscreteScheduler,
113
+ FlaxPNDMScheduler,
114
+ FlaxSchedulerMixin,
115
+ FlaxScoreSdeVeScheduler,
116
+ )
117
+ else:
118
+ from .utils.dummy_flax_objects import * # noqa F403
119
+
120
+ if is_flax_available() and is_transformers_available():
121
+ from .pipelines import FlaxStableDiffusionPipeline
122
+ else:
123
+ from .utils.dummy_flax_and_transformers_objects import * # noqa F403
diffusers/commands/__init__.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2022 The HuggingFace Team. All rights reserved.
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ from abc import ABC, abstractmethod
16
+ from argparse import ArgumentParser
17
+
18
+
19
+ class BaseDiffusersCLICommand(ABC):
20
+ @staticmethod
21
+ @abstractmethod
22
+ def register_subcommand(parser: ArgumentParser):
23
+ raise NotImplementedError()
24
+
25
+ @abstractmethod
26
+ def run(self):
27
+ raise NotImplementedError()
diffusers/commands/diffusers_cli.py ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ # Copyright 2022 The HuggingFace Team. All rights reserved.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ from argparse import ArgumentParser
17
+
18
+ from .env import EnvironmentCommand
19
+
20
+
21
+ def main():
22
+ parser = ArgumentParser("Diffusers CLI tool", usage="diffusers-cli <command> [<args>]")
23
+ commands_parser = parser.add_subparsers(help="diffusers-cli command helpers")
24
+
25
+ # Register commands
26
+ EnvironmentCommand.register_subcommand(commands_parser)
27
+
28
+ # Let's go
29
+ args = parser.parse_args()
30
+
31
+ if not hasattr(args, "func"):
32
+ parser.print_help()
33
+ exit(1)
34
+
35
+ # Run
36
+ service = args.func(args)
37
+ service.run()
38
+
39
+
40
+ if __name__ == "__main__":
41
+ main()
diffusers/commands/env.py ADDED
@@ -0,0 +1,70 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2022 The HuggingFace Team. All rights reserved.
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ import platform
16
+ from argparse import ArgumentParser
17
+
18
+ import huggingface_hub
19
+
20
+ from .. import __version__ as version
21
+ from ..utils import is_torch_available, is_transformers_available
22
+ from . import BaseDiffusersCLICommand
23
+
24
+
25
+ def info_command_factory(_):
26
+ return EnvironmentCommand()
27
+
28
+
29
+ class EnvironmentCommand(BaseDiffusersCLICommand):
30
+ @staticmethod
31
+ def register_subcommand(parser: ArgumentParser):
32
+ download_parser = parser.add_parser("env")
33
+ download_parser.set_defaults(func=info_command_factory)
34
+
35
+ def run(self):
36
+ hub_version = huggingface_hub.__version__
37
+
38
+ pt_version = "not installed"
39
+ pt_cuda_available = "NA"
40
+ if is_torch_available():
41
+ import torch
42
+
43
+ pt_version = torch.__version__
44
+ pt_cuda_available = torch.cuda.is_available()
45
+
46
+ transformers_version = "not installed"
47
+ if is_transformers_available:
48
+ import transformers
49
+
50
+ transformers_version = transformers.__version__
51
+
52
+ info = {
53
+ "`diffusers` version": version,
54
+ "Platform": platform.platform(),
55
+ "Python version": platform.python_version(),
56
+ "PyTorch version (GPU?)": f"{pt_version} ({pt_cuda_available})",
57
+ "Huggingface_hub version": hub_version,
58
+ "Transformers version": transformers_version,
59
+ "Using GPU in script?": "<fill in>",
60
+ "Using distributed or parallel set-up in script?": "<fill in>",
61
+ }
62
+
63
+ print("\nCopy-and-paste the text below in your GitHub issue and FILL OUT the two last points.\n")
64
+ print(self.format_dict(info))
65
+
66
+ return info
67
+
68
+ @staticmethod
69
+ def format_dict(d):
70
+ return "\n".join([f"- {prop}: {val}" for prop, val in d.items()]) + "\n"
diffusers/configuration_utils.py ADDED
@@ -0,0 +1,613 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 The HuggingFace Inc. team.
3
+ # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+ # limitations under the License.
16
+ """ ConfigMixin base class and utilities."""
17
+ import dataclasses
18
+ import functools
19
+ import importlib
20
+ import inspect
21
+ import json
22
+ import os
23
+ import re
24
+ from collections import OrderedDict
25
+ from typing import Any, Dict, Tuple, Union
26
+
27
+ import numpy as np
28
+
29
+ from huggingface_hub import hf_hub_download
30
+ from huggingface_hub.utils import EntryNotFoundError, RepositoryNotFoundError, RevisionNotFoundError
31
+ from requests import HTTPError
32
+
33
+ from . import __version__
34
+ from .utils import DIFFUSERS_CACHE, HUGGINGFACE_CO_RESOLVE_ENDPOINT, DummyObject, deprecate, logging
35
+
36
+
37
+ logger = logging.get_logger(__name__)
38
+
39
+ _re_configuration_file = re.compile(r"config\.(.*)\.json")
40
+
41
+
42
+ class FrozenDict(OrderedDict):
43
+ def __init__(self, *args, **kwargs):
44
+ super().__init__(*args, **kwargs)
45
+
46
+ for key, value in self.items():
47
+ setattr(self, key, value)
48
+
49
+ self.__frozen = True
50
+
51
+ def __delitem__(self, *args, **kwargs):
52
+ raise Exception(f"You cannot use ``__delitem__`` on a {self.__class__.__name__} instance.")
53
+
54
+ def setdefault(self, *args, **kwargs):
55
+ raise Exception(f"You cannot use ``setdefault`` on a {self.__class__.__name__} instance.")
56
+
57
+ def pop(self, *args, **kwargs):
58
+ raise Exception(f"You cannot use ``pop`` on a {self.__class__.__name__} instance.")
59
+
60
+ def update(self, *args, **kwargs):
61
+ raise Exception(f"You cannot use ``update`` on a {self.__class__.__name__} instance.")
62
+
63
+ def __setattr__(self, name, value):
64
+ if hasattr(self, "__frozen") and self.__frozen:
65
+ raise Exception(f"You cannot use ``__setattr__`` on a {self.__class__.__name__} instance.")
66
+ super().__setattr__(name, value)
67
+
68
+ def __setitem__(self, name, value):
69
+ if hasattr(self, "__frozen") and self.__frozen:
70
+ raise Exception(f"You cannot use ``__setattr__`` on a {self.__class__.__name__} instance.")
71
+ super().__setitem__(name, value)
72
+
73
+
74
+ class ConfigMixin:
75
+ r"""
76
+ Base class for all configuration classes. Stores all configuration parameters under `self.config` Also handles all
77
+ methods for loading/downloading/saving classes inheriting from [`ConfigMixin`] with
78
+ - [`~ConfigMixin.from_config`]
79
+ - [`~ConfigMixin.save_config`]
80
+
81
+ Class attributes:
82
+ - **config_name** (`str`) -- A filename under which the config should stored when calling
83
+ [`~ConfigMixin.save_config`] (should be overridden by parent class).
84
+ - **ignore_for_config** (`List[str]`) -- A list of attributes that should not be saved in the config (should be
85
+ overridden by subclass).
86
+ - **has_compatibles** (`bool`) -- Whether the class has compatible classes (should be overridden by subclass).
87
+ - **_deprecated_kwargs** (`List[str]`) -- Keyword arguments that are deprecated. Note that the init function
88
+ should only have a `kwargs` argument if at least one argument is deprecated (should be overridden by
89
+ subclass).
90
+ """
91
+ config_name = None
92
+ ignore_for_config = []
93
+ has_compatibles = False
94
+
95
+ _deprecated_kwargs = []
96
+
97
+ def register_to_config(self, **kwargs):
98
+ if self.config_name is None:
99
+ raise NotImplementedError(f"Make sure that {self.__class__} has defined a class name `config_name`")
100
+ # Special case for `kwargs` used in deprecation warning added to schedulers
101
+ # TODO: remove this when we remove the deprecation warning, and the `kwargs` argument,
102
+ # or solve in a more general way.
103
+ kwargs.pop("kwargs", None)
104
+ for key, value in kwargs.items():
105
+ try:
106
+ setattr(self, key, value)
107
+ except AttributeError as err:
108
+ logger.error(f"Can't set {key} with value {value} for {self}")
109
+ raise err
110
+
111
+ if not hasattr(self, "_internal_dict"):
112
+ internal_dict = kwargs
113
+ else:
114
+ previous_dict = dict(self._internal_dict)
115
+ internal_dict = {**self._internal_dict, **kwargs}
116
+ logger.debug(f"Updating config from {previous_dict} to {internal_dict}")
117
+
118
+ self._internal_dict = FrozenDict(internal_dict)
119
+
120
+ def save_config(self, save_directory: Union[str, os.PathLike], push_to_hub: bool = False, **kwargs):
121
+ """
122
+ Save a configuration object to the directory `save_directory`, so that it can be re-loaded using the
123
+ [`~ConfigMixin.from_config`] class method.
124
+
125
+ Args:
126
+ save_directory (`str` or `os.PathLike`):
127
+ Directory where the configuration JSON file will be saved (will be created if it does not exist).
128
+ """
129
+ if os.path.isfile(save_directory):
130
+ raise AssertionError(f"Provided path ({save_directory}) should be a directory, not a file")
131
+
132
+ os.makedirs(save_directory, exist_ok=True)
133
+
134
+ # If we save using the predefined names, we can load using `from_config`
135
+ output_config_file = os.path.join(save_directory, self.config_name)
136
+
137
+ self.to_json_file(output_config_file)
138
+ logger.info(f"Configuration saved in {output_config_file}")
139
+
140
+ @classmethod
141
+ def from_config(cls, config: Union[FrozenDict, Dict[str, Any]] = None, return_unused_kwargs=False, **kwargs):
142
+ r"""
143
+ Instantiate a Python class from a config dictionary
144
+
145
+ Parameters:
146
+ config (`Dict[str, Any]`):
147
+ A config dictionary from which the Python class will be instantiated. Make sure to only load
148
+ configuration files of compatible classes.
149
+ return_unused_kwargs (`bool`, *optional*, defaults to `False`):
150
+ Whether kwargs that are not consumed by the Python class should be returned or not.
151
+
152
+ kwargs (remaining dictionary of keyword arguments, *optional*):
153
+ Can be used to update the configuration object (after it being loaded) and initiate the Python class.
154
+ `**kwargs` will be directly passed to the underlying scheduler/model's `__init__` method and eventually
155
+ overwrite same named arguments of `config`.
156
+
157
+ Examples:
158
+
159
+ ```python
160
+ >>> from diffusers import DDPMScheduler, DDIMScheduler, PNDMScheduler
161
+
162
+ >>> # Download scheduler from huggingface.co and cache.
163
+ >>> scheduler = DDPMScheduler.from_pretrained("google/ddpm-cifar10-32")
164
+
165
+ >>> # Instantiate DDIM scheduler class with same config as DDPM
166
+ >>> scheduler = DDIMScheduler.from_config(scheduler.config)
167
+
168
+ >>> # Instantiate PNDM scheduler class with same config as DDPM
169
+ >>> scheduler = PNDMScheduler.from_config(scheduler.config)
170
+ ```
171
+ """
172
+ # <===== TO BE REMOVED WITH DEPRECATION
173
+ # TODO(Patrick) - make sure to remove the following lines when config=="model_path" is deprecated
174
+ if "pretrained_model_name_or_path" in kwargs:
175
+ config = kwargs.pop("pretrained_model_name_or_path")
176
+
177
+ if config is None:
178
+ raise ValueError("Please make sure to provide a config as the first positional argument.")
179
+ # ======>
180
+
181
+ if not isinstance(config, dict):
182
+ deprecation_message = "It is deprecated to pass a pretrained model name or path to `from_config`."
183
+ if "Scheduler" in cls.__name__:
184
+ deprecation_message += (
185
+ f"If you were trying to load a scheduler, please use {cls}.from_pretrained(...) instead."
186
+ " Otherwise, please make sure to pass a configuration dictionary instead. This functionality will"
187
+ " be removed in v1.0.0."
188
+ )
189
+ elif "Model" in cls.__name__:
190
+ deprecation_message += (
191
+ f"If you were trying to load a model, please use {cls}.load_config(...) followed by"
192
+ f" {cls}.from_config(...) instead. Otherwise, please make sure to pass a configuration dictionary"
193
+ " instead. This functionality will be removed in v1.0.0."
194
+ )
195
+ deprecate("config-passed-as-path", "1.0.0", deprecation_message, standard_warn=False)
196
+ config, kwargs = cls.load_config(pretrained_model_name_or_path=config, return_unused_kwargs=True, **kwargs)
197
+
198
+ init_dict, unused_kwargs, hidden_dict = cls.extract_init_dict(config, **kwargs)
199
+
200
+ # Allow dtype to be specified on initialization
201
+ if "dtype" in unused_kwargs:
202
+ init_dict["dtype"] = unused_kwargs.pop("dtype")
203
+
204
+ # add possible deprecated kwargs
205
+ for deprecated_kwarg in cls._deprecated_kwargs:
206
+ if deprecated_kwarg in unused_kwargs:
207
+ init_dict[deprecated_kwarg] = unused_kwargs.pop(deprecated_kwarg)
208
+
209
+ # Return model and optionally state and/or unused_kwargs
210
+ model = cls(**init_dict)
211
+
212
+ # make sure to also save config parameters that might be used for compatible classes
213
+ model.register_to_config(**hidden_dict)
214
+
215
+ # add hidden kwargs of compatible classes to unused_kwargs
216
+ unused_kwargs = {**unused_kwargs, **hidden_dict}
217
+
218
+ if return_unused_kwargs:
219
+ return (model, unused_kwargs)
220
+ else:
221
+ return model
222
+
223
+ @classmethod
224
+ def get_config_dict(cls, *args, **kwargs):
225
+ deprecation_message = (
226
+ f" The function get_config_dict is deprecated. Please use {cls}.load_config instead. This function will be"
227
+ " removed in version v1.0.0"
228
+ )
229
+ deprecate("get_config_dict", "1.0.0", deprecation_message, standard_warn=False)
230
+ return cls.load_config(*args, **kwargs)
231
+
232
+ @classmethod
233
+ def load_config(
234
+ cls, pretrained_model_name_or_path: Union[str, os.PathLike], return_unused_kwargs=False, **kwargs
235
+ ) -> Tuple[Dict[str, Any], Dict[str, Any]]:
236
+ r"""
237
+ Instantiate a Python class from a config dictionary
238
+
239
+ Parameters:
240
+ pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
241
+ Can be either:
242
+
243
+ - A string, the *model id* of a model repo on huggingface.co. Valid model ids should have an
244
+ organization name, like `google/ddpm-celebahq-256`.
245
+ - A path to a *directory* containing model weights saved using [`~ConfigMixin.save_config`], e.g.,
246
+ `./my_model_directory/`.
247
+
248
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
249
+ Path to a directory in which a downloaded pretrained model configuration should be cached if the
250
+ standard cache should not be used.
251
+ force_download (`bool`, *optional*, defaults to `False`):
252
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
253
+ cached versions if they exist.
254
+ resume_download (`bool`, *optional*, defaults to `False`):
255
+ Whether or not to delete incompletely received files. Will attempt to resume the download if such a
256
+ file exists.
257
+ proxies (`Dict[str, str]`, *optional*):
258
+ A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
259
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
260
+ output_loading_info(`bool`, *optional*, defaults to `False`):
261
+ Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
262
+ local_files_only(`bool`, *optional*, defaults to `False`):
263
+ Whether or not to only look at local files (i.e., do not try to download the model).
264
+ use_auth_token (`str` or *bool*, *optional*):
265
+ The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
266
+ when running `transformers-cli login` (stored in `~/.huggingface`).
267
+ revision (`str`, *optional*, defaults to `"main"`):
268
+ The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
269
+ git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
270
+ identifier allowed by git.
271
+ subfolder (`str`, *optional*, defaults to `""`):
272
+ In case the relevant files are located inside a subfolder of the model repo (either remote in
273
+ huggingface.co or downloaded locally), you can specify the folder name here.
274
+
275
+ <Tip>
276
+
277
+ It is required to be logged in (`huggingface-cli login`) when you want to use private or [gated
278
+ models](https://huggingface.co/docs/hub/models-gated#gated-models).
279
+
280
+ </Tip>
281
+
282
+ <Tip>
283
+
284
+ Activate the special ["offline-mode"](https://huggingface.co/transformers/installation.html#offline-mode) to
285
+ use this method in a firewalled environment.
286
+
287
+ </Tip>
288
+ """
289
+ cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
290
+ force_download = kwargs.pop("force_download", False)
291
+ resume_download = kwargs.pop("resume_download", False)
292
+ proxies = kwargs.pop("proxies", None)
293
+ use_auth_token = kwargs.pop("use_auth_token", None)
294
+ local_files_only = kwargs.pop("local_files_only", False)
295
+ revision = kwargs.pop("revision", None)
296
+ _ = kwargs.pop("mirror", None)
297
+ subfolder = kwargs.pop("subfolder", None)
298
+
299
+ user_agent = {"file_type": "config"}
300
+
301
+ pretrained_model_name_or_path = str(pretrained_model_name_or_path)
302
+
303
+ if cls.config_name is None:
304
+ raise ValueError(
305
+ "`self.config_name` is not defined. Note that one should not load a config from "
306
+ "`ConfigMixin`. Please make sure to define `config_name` in a class inheriting from `ConfigMixin`"
307
+ )
308
+
309
+ if os.path.isfile(pretrained_model_name_or_path):
310
+ config_file = pretrained_model_name_or_path
311
+ elif os.path.isdir(pretrained_model_name_or_path):
312
+ if os.path.isfile(os.path.join(pretrained_model_name_or_path, cls.config_name)):
313
+ # Load from a PyTorch checkpoint
314
+ config_file = os.path.join(pretrained_model_name_or_path, cls.config_name)
315
+ elif subfolder is not None and os.path.isfile(
316
+ os.path.join(pretrained_model_name_or_path, subfolder, cls.config_name)
317
+ ):
318
+ config_file = os.path.join(pretrained_model_name_or_path, subfolder, cls.config_name)
319
+ else:
320
+ raise EnvironmentError(
321
+ f"Error no file named {cls.config_name} found in directory {pretrained_model_name_or_path}."
322
+ )
323
+ else:
324
+ try:
325
+ # Load from URL or cache if already cached
326
+ config_file = hf_hub_download(
327
+ pretrained_model_name_or_path,
328
+ filename=cls.config_name,
329
+ cache_dir=cache_dir,
330
+ force_download=force_download,
331
+ proxies=proxies,
332
+ resume_download=resume_download,
333
+ local_files_only=local_files_only,
334
+ use_auth_token=use_auth_token,
335
+ user_agent=user_agent,
336
+ subfolder=subfolder,
337
+ revision=revision,
338
+ )
339
+
340
+ except RepositoryNotFoundError:
341
+ raise EnvironmentError(
342
+ f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier"
343
+ " listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a"
344
+ " token having permission to this repo with `use_auth_token` or log in with `huggingface-cli"
345
+ " login`."
346
+ )
347
+ except RevisionNotFoundError:
348
+ raise EnvironmentError(
349
+ f"{revision} is not a valid git identifier (branch name, tag name or commit id) that exists for"
350
+ " this model name. Check the model page at"
351
+ f" 'https://huggingface.co/{pretrained_model_name_or_path}' for available revisions."
352
+ )
353
+ except EntryNotFoundError:
354
+ raise EnvironmentError(
355
+ f"{pretrained_model_name_or_path} does not appear to have a file named {cls.config_name}."
356
+ )
357
+ except HTTPError as err:
358
+ raise EnvironmentError(
359
+ "There was a specific connection error when trying to load"
360
+ f" {pretrained_model_name_or_path}:\n{err}"
361
+ )
362
+ except ValueError:
363
+ raise EnvironmentError(
364
+ f"We couldn't connect to '{HUGGINGFACE_CO_RESOLVE_ENDPOINT}' to load this model, couldn't find it"
365
+ f" in the cached files and it looks like {pretrained_model_name_or_path} is not the path to a"
366
+ f" directory containing a {cls.config_name} file.\nCheckout your internet connection or see how to"
367
+ " run the library in offline mode at"
368
+ " 'https://huggingface.co/docs/diffusers/installation#offline-mode'."
369
+ )
370
+ except EnvironmentError:
371
+ raise EnvironmentError(
372
+ f"Can't load config for '{pretrained_model_name_or_path}'. If you were trying to load it from "
373
+ "'https://huggingface.co/models', make sure you don't have a local directory with the same name. "
374
+ f"Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a directory "
375
+ f"containing a {cls.config_name} file"
376
+ )
377
+
378
+ try:
379
+ # Load config dict
380
+ config_dict = cls._dict_from_json_file(config_file)
381
+ except (json.JSONDecodeError, UnicodeDecodeError):
382
+ raise EnvironmentError(f"It looks like the config file at '{config_file}' is not a valid JSON file.")
383
+
384
+ if return_unused_kwargs:
385
+ return config_dict, kwargs
386
+
387
+ return config_dict
388
+
389
+ @staticmethod
390
+ def _get_init_keys(cls):
391
+ return set(dict(inspect.signature(cls.__init__).parameters).keys())
392
+
393
+ @classmethod
394
+ def extract_init_dict(cls, config_dict, **kwargs):
395
+ # 0. Copy origin config dict
396
+ original_dict = {k: v for k, v in config_dict.items()}
397
+
398
+ # 1. Retrieve expected config attributes from __init__ signature
399
+ expected_keys = cls._get_init_keys(cls)
400
+ expected_keys.remove("self")
401
+ # remove general kwargs if present in dict
402
+ if "kwargs" in expected_keys:
403
+ expected_keys.remove("kwargs")
404
+ # remove flax internal keys
405
+ if hasattr(cls, "_flax_internal_args"):
406
+ for arg in cls._flax_internal_args:
407
+ expected_keys.remove(arg)
408
+
409
+ # 2. Remove attributes that cannot be expected from expected config attributes
410
+ # remove keys to be ignored
411
+ if len(cls.ignore_for_config) > 0:
412
+ expected_keys = expected_keys - set(cls.ignore_for_config)
413
+
414
+ # load diffusers library to import compatible and original scheduler
415
+ diffusers_library = importlib.import_module(__name__.split(".")[0])
416
+
417
+ if cls.has_compatibles:
418
+ compatible_classes = [c for c in cls._get_compatibles() if not isinstance(c, DummyObject)]
419
+ else:
420
+ compatible_classes = []
421
+
422
+ expected_keys_comp_cls = set()
423
+ for c in compatible_classes:
424
+ expected_keys_c = cls._get_init_keys(c)
425
+ expected_keys_comp_cls = expected_keys_comp_cls.union(expected_keys_c)
426
+ expected_keys_comp_cls = expected_keys_comp_cls - cls._get_init_keys(cls)
427
+ config_dict = {k: v for k, v in config_dict.items() if k not in expected_keys_comp_cls}
428
+
429
+ # remove attributes from orig class that cannot be expected
430
+ orig_cls_name = config_dict.pop("_class_name", cls.__name__)
431
+ if orig_cls_name != cls.__name__ and hasattr(diffusers_library, orig_cls_name):
432
+ orig_cls = getattr(diffusers_library, orig_cls_name)
433
+ unexpected_keys_from_orig = cls._get_init_keys(orig_cls) - expected_keys
434
+ config_dict = {k: v for k, v in config_dict.items() if k not in unexpected_keys_from_orig}
435
+
436
+ # remove private attributes
437
+ config_dict = {k: v for k, v in config_dict.items() if not k.startswith("_")}
438
+
439
+ # 3. Create keyword arguments that will be passed to __init__ from expected keyword arguments
440
+ init_dict = {}
441
+ for key in expected_keys:
442
+ # if config param is passed to kwarg and is present in config dict
443
+ # it should overwrite existing config dict key
444
+ if key in kwargs and key in config_dict:
445
+ config_dict[key] = kwargs.pop(key)
446
+
447
+ if key in kwargs:
448
+ # overwrite key
449
+ init_dict[key] = kwargs.pop(key)
450
+ elif key in config_dict:
451
+ # use value from config dict
452
+ init_dict[key] = config_dict.pop(key)
453
+
454
+ # 4. Give nice warning if unexpected values have been passed
455
+ if len(config_dict) > 0:
456
+ logger.warning(
457
+ f"The config attributes {config_dict} were passed to {cls.__name__}, "
458
+ "but are not expected and will be ignored. Please verify your "
459
+ f"{cls.config_name} configuration file."
460
+ )
461
+
462
+ # 5. Give nice info if config attributes are initiliazed to default because they have not been passed
463
+ passed_keys = set(init_dict.keys())
464
+ if len(expected_keys - passed_keys) > 0:
465
+ logger.info(
466
+ f"{expected_keys - passed_keys} was not found in config. Values will be initialized to default values."
467
+ )
468
+
469
+ # 6. Define unused keyword arguments
470
+ unused_kwargs = {**config_dict, **kwargs}
471
+
472
+ # 7. Define "hidden" config parameters that were saved for compatible classes
473
+ hidden_config_dict = {k: v for k, v in original_dict.items() if k not in init_dict}
474
+
475
+ return init_dict, unused_kwargs, hidden_config_dict
476
+
477
+ @classmethod
478
+ def _dict_from_json_file(cls, json_file: Union[str, os.PathLike]):
479
+ with open(json_file, "r", encoding="utf-8") as reader:
480
+ text = reader.read()
481
+ return json.loads(text)
482
+
483
+ def __repr__(self):
484
+ return f"{self.__class__.__name__} {self.to_json_string()}"
485
+
486
+ @property
487
+ def config(self) -> Dict[str, Any]:
488
+ """
489
+ Returns the config of the class as a frozen dictionary
490
+
491
+ Returns:
492
+ `Dict[str, Any]`: Config of the class.
493
+ """
494
+ return self._internal_dict
495
+
496
+ def to_json_string(self) -> str:
497
+ """
498
+ Serializes this instance to a JSON string.
499
+
500
+ Returns:
501
+ `str`: String containing all the attributes that make up this configuration instance in JSON format.
502
+ """
503
+ config_dict = self._internal_dict if hasattr(self, "_internal_dict") else {}
504
+ config_dict["_class_name"] = self.__class__.__name__
505
+ config_dict["_diffusers_version"] = __version__
506
+
507
+ def to_json_saveable(value):
508
+ if isinstance(value, np.ndarray):
509
+ value = value.tolist()
510
+ return value
511
+
512
+ config_dict = {k: to_json_saveable(v) for k, v in config_dict.items()}
513
+ return json.dumps(config_dict, indent=2, sort_keys=True) + "\n"
514
+
515
+ def to_json_file(self, json_file_path: Union[str, os.PathLike]):
516
+ """
517
+ Save this instance to a JSON file.
518
+
519
+ Args:
520
+ json_file_path (`str` or `os.PathLike`):
521
+ Path to the JSON file in which this configuration instance's parameters will be saved.
522
+ """
523
+ with open(json_file_path, "w", encoding="utf-8") as writer:
524
+ writer.write(self.to_json_string())
525
+
526
+
527
+ def register_to_config(init):
528
+ r"""
529
+ Decorator to apply on the init of classes inheriting from [`ConfigMixin`] so that all the arguments are
530
+ automatically sent to `self.register_for_config`. To ignore a specific argument accepted by the init but that
531
+ shouldn't be registered in the config, use the `ignore_for_config` class variable
532
+
533
+ Warning: Once decorated, all private arguments (beginning with an underscore) are trashed and not sent to the init!
534
+ """
535
+
536
+ @functools.wraps(init)
537
+ def inner_init(self, *args, **kwargs):
538
+ # Ignore private kwargs in the init.
539
+ init_kwargs = {k: v for k, v in kwargs.items() if not k.startswith("_")}
540
+ config_init_kwargs = {k: v for k, v in kwargs.items() if k.startswith("_")}
541
+ if not isinstance(self, ConfigMixin):
542
+ raise RuntimeError(
543
+ f"`@register_for_config` was applied to {self.__class__.__name__} init method, but this class does "
544
+ "not inherit from `ConfigMixin`."
545
+ )
546
+
547
+ ignore = getattr(self, "ignore_for_config", [])
548
+ # Get positional arguments aligned with kwargs
549
+ new_kwargs = {}
550
+ signature = inspect.signature(init)
551
+ parameters = {
552
+ name: p.default for i, (name, p) in enumerate(signature.parameters.items()) if i > 0 and name not in ignore
553
+ }
554
+ for arg, name in zip(args, parameters.keys()):
555
+ new_kwargs[name] = arg
556
+
557
+ # Then add all kwargs
558
+ new_kwargs.update(
559
+ {
560
+ k: init_kwargs.get(k, default)
561
+ for k, default in parameters.items()
562
+ if k not in ignore and k not in new_kwargs
563
+ }
564
+ )
565
+ new_kwargs = {**config_init_kwargs, **new_kwargs}
566
+ getattr(self, "register_to_config")(**new_kwargs)
567
+ init(self, *args, **init_kwargs)
568
+
569
+ return inner_init
570
+
571
+
572
+ def flax_register_to_config(cls):
573
+ original_init = cls.__init__
574
+
575
+ @functools.wraps(original_init)
576
+ def init(self, *args, **kwargs):
577
+ if not isinstance(self, ConfigMixin):
578
+ raise RuntimeError(
579
+ f"`@register_for_config` was applied to {self.__class__.__name__} init method, but this class does "
580
+ "not inherit from `ConfigMixin`."
581
+ )
582
+
583
+ # Ignore private kwargs in the init. Retrieve all passed attributes
584
+ init_kwargs = {k: v for k, v in kwargs.items()}
585
+
586
+ # Retrieve default values
587
+ fields = dataclasses.fields(self)
588
+ default_kwargs = {}
589
+ for field in fields:
590
+ # ignore flax specific attributes
591
+ if field.name in self._flax_internal_args:
592
+ continue
593
+ if type(field.default) == dataclasses._MISSING_TYPE:
594
+ default_kwargs[field.name] = None
595
+ else:
596
+ default_kwargs[field.name] = getattr(self, field.name)
597
+
598
+ # Make sure init_kwargs override default kwargs
599
+ new_kwargs = {**default_kwargs, **init_kwargs}
600
+ # dtype should be part of `init_kwargs`, but not `new_kwargs`
601
+ if "dtype" in new_kwargs:
602
+ new_kwargs.pop("dtype")
603
+
604
+ # Get positional arguments aligned with kwargs
605
+ for i, arg in enumerate(args):
606
+ name = fields[i].name
607
+ new_kwargs[name] = arg
608
+
609
+ getattr(self, "register_to_config")(**new_kwargs)
610
+ original_init(self, *args, **kwargs)
611
+
612
+ cls.__init__ = init
613
+ return cls
diffusers/dependency_versions_check.py ADDED
@@ -0,0 +1,47 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2020 The HuggingFace Team. All rights reserved.
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+ import sys
15
+
16
+ from .dependency_versions_table import deps
17
+ from .utils.versions import require_version, require_version_core
18
+
19
+
20
+ # define which module versions we always want to check at run time
21
+ # (usually the ones defined in `install_requires` in setup.py)
22
+ #
23
+ # order specific notes:
24
+ # - tqdm must be checked before tokenizers
25
+
26
+ pkgs_to_check_at_runtime = "python tqdm regex requests packaging filelock numpy tokenizers".split()
27
+ if sys.version_info < (3, 7):
28
+ pkgs_to_check_at_runtime.append("dataclasses")
29
+ if sys.version_info < (3, 8):
30
+ pkgs_to_check_at_runtime.append("importlib_metadata")
31
+
32
+ for pkg in pkgs_to_check_at_runtime:
33
+ if pkg in deps:
34
+ if pkg == "tokenizers":
35
+ # must be loaded here, or else tqdm check may fail
36
+ from .utils import is_tokenizers_available
37
+
38
+ if not is_tokenizers_available():
39
+ continue # not required, check version only if installed
40
+
41
+ require_version_core(deps[pkg])
42
+ else:
43
+ raise ValueError(f"can't find {pkg} in {deps.keys()}, check dependency_versions_table.py")
44
+
45
+
46
+ def dep_version_check(pkg, hint=None):
47
+ require_version(deps[pkg], hint)
diffusers/dependency_versions_table.py ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # THIS FILE HAS BEEN AUTOGENERATED. To update:
2
+ # 1. modify the `_deps` dict in setup.py
3
+ # 2. run `make deps_table_update``
4
+ deps = {
5
+ "Pillow": "Pillow",
6
+ "accelerate": "accelerate>=0.11.0",
7
+ "black": "black==22.8",
8
+ "datasets": "datasets",
9
+ "filelock": "filelock",
10
+ "flake8": "flake8>=3.8.3",
11
+ "flax": "flax>=0.4.1",
12
+ "hf-doc-builder": "hf-doc-builder>=0.3.0",
13
+ "huggingface-hub": "huggingface-hub>=0.10.0",
14
+ "importlib_metadata": "importlib_metadata",
15
+ "isort": "isort>=5.5.4",
16
+ "jax": "jax>=0.2.8,!=0.3.2",
17
+ "jaxlib": "jaxlib>=0.1.65",
18
+ "modelcards": "modelcards>=0.1.4",
19
+ "numpy": "numpy",
20
+ "parameterized": "parameterized",
21
+ "pytest": "pytest",
22
+ "pytest-timeout": "pytest-timeout",
23
+ "pytest-xdist": "pytest-xdist",
24
+ "safetensors": "safetensors",
25
+ "sentencepiece": "sentencepiece>=0.1.91,!=0.1.92",
26
+ "scipy": "scipy",
27
+ "regex": "regex!=2019.12.17",
28
+ "requests": "requests",
29
+ "tensorboard": "tensorboard",
30
+ "torch": "torch>=1.4",
31
+ "torchvision": "torchvision",
32
+ "transformers": "transformers>=4.21.0",
33
+ }
diffusers/dynamic_modules_utils.py ADDED
@@ -0,0 +1,428 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 The HuggingFace Inc. team.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+ """Utilities to dynamically load objects from the Hub."""
16
+
17
+ import importlib
18
+ import inspect
19
+ import os
20
+ import re
21
+ import shutil
22
+ import sys
23
+ from pathlib import Path
24
+ from typing import Dict, Optional, Union
25
+
26
+ from huggingface_hub import HfFolder, cached_download, hf_hub_download, model_info
27
+
28
+ from .utils import DIFFUSERS_DYNAMIC_MODULE_NAME, HF_MODULES_CACHE, logging
29
+
30
+
31
+ COMMUNITY_PIPELINES_URL = (
32
+ "https://raw.githubusercontent.com/huggingface/diffusers/main/examples/community/{pipeline}.py"
33
+ )
34
+
35
+
36
+ logger = logging.get_logger(__name__) # pylint: disable=invalid-name
37
+
38
+
39
+ def init_hf_modules():
40
+ """
41
+ Creates the cache directory for modules with an init, and adds it to the Python path.
42
+ """
43
+ # This function has already been executed if HF_MODULES_CACHE already is in the Python path.
44
+ if HF_MODULES_CACHE in sys.path:
45
+ return
46
+
47
+ sys.path.append(HF_MODULES_CACHE)
48
+ os.makedirs(HF_MODULES_CACHE, exist_ok=True)
49
+ init_path = Path(HF_MODULES_CACHE) / "__init__.py"
50
+ if not init_path.exists():
51
+ init_path.touch()
52
+
53
+
54
+ def create_dynamic_module(name: Union[str, os.PathLike]):
55
+ """
56
+ Creates a dynamic module in the cache directory for modules.
57
+ """
58
+ init_hf_modules()
59
+ dynamic_module_path = Path(HF_MODULES_CACHE) / name
60
+ # If the parent module does not exist yet, recursively create it.
61
+ if not dynamic_module_path.parent.exists():
62
+ create_dynamic_module(dynamic_module_path.parent)
63
+ os.makedirs(dynamic_module_path, exist_ok=True)
64
+ init_path = dynamic_module_path / "__init__.py"
65
+ if not init_path.exists():
66
+ init_path.touch()
67
+
68
+
69
+ def get_relative_imports(module_file):
70
+ """
71
+ Get the list of modules that are relatively imported in a module file.
72
+
73
+ Args:
74
+ module_file (`str` or `os.PathLike`): The module file to inspect.
75
+ """
76
+ with open(module_file, "r", encoding="utf-8") as f:
77
+ content = f.read()
78
+
79
+ # Imports of the form `import .xxx`
80
+ relative_imports = re.findall("^\s*import\s+\.(\S+)\s*$", content, flags=re.MULTILINE)
81
+ # Imports of the form `from .xxx import yyy`
82
+ relative_imports += re.findall("^\s*from\s+\.(\S+)\s+import", content, flags=re.MULTILINE)
83
+ # Unique-ify
84
+ return list(set(relative_imports))
85
+
86
+
87
+ def get_relative_import_files(module_file):
88
+ """
89
+ Get the list of all files that are needed for a given module. Note that this function recurses through the relative
90
+ imports (if a imports b and b imports c, it will return module files for b and c).
91
+
92
+ Args:
93
+ module_file (`str` or `os.PathLike`): The module file to inspect.
94
+ """
95
+ no_change = False
96
+ files_to_check = [module_file]
97
+ all_relative_imports = []
98
+
99
+ # Let's recurse through all relative imports
100
+ while not no_change:
101
+ new_imports = []
102
+ for f in files_to_check:
103
+ new_imports.extend(get_relative_imports(f))
104
+
105
+ module_path = Path(module_file).parent
106
+ new_import_files = [str(module_path / m) for m in new_imports]
107
+ new_import_files = [f for f in new_import_files if f not in all_relative_imports]
108
+ files_to_check = [f"{f}.py" for f in new_import_files]
109
+
110
+ no_change = len(new_import_files) == 0
111
+ all_relative_imports.extend(files_to_check)
112
+
113
+ return all_relative_imports
114
+
115
+
116
+ def check_imports(filename):
117
+ """
118
+ Check if the current Python environment contains all the libraries that are imported in a file.
119
+ """
120
+ with open(filename, "r", encoding="utf-8") as f:
121
+ content = f.read()
122
+
123
+ # Imports of the form `import xxx`
124
+ imports = re.findall("^\s*import\s+(\S+)\s*$", content, flags=re.MULTILINE)
125
+ # Imports of the form `from xxx import yyy`
126
+ imports += re.findall("^\s*from\s+(\S+)\s+import", content, flags=re.MULTILINE)
127
+ # Only keep the top-level module
128
+ imports = [imp.split(".")[0] for imp in imports if not imp.startswith(".")]
129
+
130
+ # Unique-ify and test we got them all
131
+ imports = list(set(imports))
132
+ missing_packages = []
133
+ for imp in imports:
134
+ try:
135
+ importlib.import_module(imp)
136
+ except ImportError:
137
+ missing_packages.append(imp)
138
+
139
+ if len(missing_packages) > 0:
140
+ raise ImportError(
141
+ "This modeling file requires the following packages that were not found in your environment: "
142
+ f"{', '.join(missing_packages)}. Run `pip install {' '.join(missing_packages)}`"
143
+ )
144
+
145
+ return get_relative_imports(filename)
146
+
147
+
148
+ def get_class_in_module(class_name, module_path):
149
+ """
150
+ Import a module on the cache directory for modules and extract a class from it.
151
+ """
152
+ module_path = module_path.replace(os.path.sep, ".")
153
+ module = importlib.import_module(module_path)
154
+
155
+ if class_name is None:
156
+ return find_pipeline_class(module)
157
+ return getattr(module, class_name)
158
+
159
+
160
+ def find_pipeline_class(loaded_module):
161
+ """
162
+ Retrieve pipeline class that inherits from `DiffusionPipeline`. Note that there has to be exactly one class
163
+ inheriting from `DiffusionPipeline`.
164
+ """
165
+ from .pipeline_utils import DiffusionPipeline
166
+
167
+ cls_members = dict(inspect.getmembers(loaded_module, inspect.isclass))
168
+
169
+ pipeline_class = None
170
+ for cls_name, cls in cls_members.items():
171
+ if (
172
+ cls_name != DiffusionPipeline.__name__
173
+ and issubclass(cls, DiffusionPipeline)
174
+ and cls.__module__.split(".")[0] != "diffusers"
175
+ ):
176
+ if pipeline_class is not None:
177
+ raise ValueError(
178
+ f"Multiple classes that inherit from {DiffusionPipeline.__name__} have been found:"
179
+ f" {pipeline_class.__name__}, and {cls_name}. Please make sure to define only one in"
180
+ f" {loaded_module}."
181
+ )
182
+ pipeline_class = cls
183
+
184
+ return pipeline_class
185
+
186
+
187
+ def get_cached_module_file(
188
+ pretrained_model_name_or_path: Union[str, os.PathLike],
189
+ module_file: str,
190
+ cache_dir: Optional[Union[str, os.PathLike]] = None,
191
+ force_download: bool = False,
192
+ resume_download: bool = False,
193
+ proxies: Optional[Dict[str, str]] = None,
194
+ use_auth_token: Optional[Union[bool, str]] = None,
195
+ revision: Optional[str] = None,
196
+ local_files_only: bool = False,
197
+ ):
198
+ """
199
+ Prepares Downloads a module from a local folder or a distant repo and returns its path inside the cached
200
+ Transformers module.
201
+
202
+ Args:
203
+ pretrained_model_name_or_path (`str` or `os.PathLike`):
204
+ This can be either:
205
+
206
+ - a string, the *model id* of a pretrained model configuration hosted inside a model repo on
207
+ huggingface.co. Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced
208
+ under a user or organization name, like `dbmdz/bert-base-german-cased`.
209
+ - a path to a *directory* containing a configuration file saved using the
210
+ [`~PreTrainedTokenizer.save_pretrained`] method, e.g., `./my_model_directory/`.
211
+
212
+ module_file (`str`):
213
+ The name of the module file containing the class to look for.
214
+ cache_dir (`str` or `os.PathLike`, *optional*):
215
+ Path to a directory in which a downloaded pretrained model configuration should be cached if the standard
216
+ cache should not be used.
217
+ force_download (`bool`, *optional*, defaults to `False`):
218
+ Whether or not to force to (re-)download the configuration files and override the cached versions if they
219
+ exist.
220
+ resume_download (`bool`, *optional*, defaults to `False`):
221
+ Whether or not to delete incompletely received file. Attempts to resume the download if such a file exists.
222
+ proxies (`Dict[str, str]`, *optional*):
223
+ A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
224
+ 'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
225
+ use_auth_token (`str` or *bool*, *optional*):
226
+ The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
227
+ when running `transformers-cli login` (stored in `~/.huggingface`).
228
+ revision (`str`, *optional*, defaults to `"main"`):
229
+ The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
230
+ git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
231
+ identifier allowed by git.
232
+ local_files_only (`bool`, *optional*, defaults to `False`):
233
+ If `True`, will only try to load the tokenizer configuration from local files.
234
+
235
+ <Tip>
236
+
237
+ You may pass a token in `use_auth_token` if you are not logged in (`huggingface-cli long`) and want to use private
238
+ or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models).
239
+
240
+ </Tip>
241
+
242
+ Returns:
243
+ `str`: The path to the module inside the cache.
244
+ """
245
+ # Download and cache module_file from the repo `pretrained_model_name_or_path` of grab it if it's a local file.
246
+ pretrained_model_name_or_path = str(pretrained_model_name_or_path)
247
+
248
+ module_file_or_url = os.path.join(pretrained_model_name_or_path, module_file)
249
+
250
+ if os.path.isfile(module_file_or_url):
251
+ resolved_module_file = module_file_or_url
252
+ submodule = "local"
253
+ elif pretrained_model_name_or_path.count("/") == 0:
254
+ # community pipeline on GitHub
255
+ github_url = COMMUNITY_PIPELINES_URL.format(pipeline=pretrained_model_name_or_path)
256
+ try:
257
+ resolved_module_file = cached_download(
258
+ github_url,
259
+ cache_dir=cache_dir,
260
+ force_download=force_download,
261
+ proxies=proxies,
262
+ resume_download=resume_download,
263
+ local_files_only=local_files_only,
264
+ use_auth_token=False,
265
+ )
266
+ submodule = "git"
267
+ module_file = pretrained_model_name_or_path + ".py"
268
+ except EnvironmentError:
269
+ logger.error(f"Could not locate the {module_file} inside {pretrained_model_name_or_path}.")
270
+ raise
271
+ else:
272
+ try:
273
+ # Load from URL or cache if already cached
274
+ resolved_module_file = hf_hub_download(
275
+ pretrained_model_name_or_path,
276
+ module_file,
277
+ cache_dir=cache_dir,
278
+ force_download=force_download,
279
+ proxies=proxies,
280
+ resume_download=resume_download,
281
+ local_files_only=local_files_only,
282
+ use_auth_token=use_auth_token,
283
+ )
284
+ submodule = os.path.join("local", "--".join(pretrained_model_name_or_path.split("/")))
285
+ except EnvironmentError:
286
+ logger.error(f"Could not locate the {module_file} inside {pretrained_model_name_or_path}.")
287
+ raise
288
+
289
+ # Check we have all the requirements in our environment
290
+ modules_needed = check_imports(resolved_module_file)
291
+
292
+ # Now we move the module inside our cached dynamic modules.
293
+ full_submodule = DIFFUSERS_DYNAMIC_MODULE_NAME + os.path.sep + submodule
294
+ create_dynamic_module(full_submodule)
295
+ submodule_path = Path(HF_MODULES_CACHE) / full_submodule
296
+ if submodule == "local" or submodule == "git":
297
+ # We always copy local files (we could hash the file to see if there was a change, and give them the name of
298
+ # that hash, to only copy when there is a modification but it seems overkill for now).
299
+ # The only reason we do the copy is to avoid putting too many folders in sys.path.
300
+ shutil.copy(resolved_module_file, submodule_path / module_file)
301
+ for module_needed in modules_needed:
302
+ module_needed = f"{module_needed}.py"
303
+ shutil.copy(os.path.join(pretrained_model_name_or_path, module_needed), submodule_path / module_needed)
304
+ else:
305
+ # Get the commit hash
306
+ # TODO: we will get this info in the etag soon, so retrieve it from there and not here.
307
+ if isinstance(use_auth_token, str):
308
+ token = use_auth_token
309
+ elif use_auth_token is True:
310
+ token = HfFolder.get_token()
311
+ else:
312
+ token = None
313
+
314
+ commit_hash = model_info(pretrained_model_name_or_path, revision=revision, token=token).sha
315
+
316
+ # The module file will end up being placed in a subfolder with the git hash of the repo. This way we get the
317
+ # benefit of versioning.
318
+ submodule_path = submodule_path / commit_hash
319
+ full_submodule = full_submodule + os.path.sep + commit_hash
320
+ create_dynamic_module(full_submodule)
321
+
322
+ if not (submodule_path / module_file).exists():
323
+ shutil.copy(resolved_module_file, submodule_path / module_file)
324
+ # Make sure we also have every file with relative
325
+ for module_needed in modules_needed:
326
+ if not (submodule_path / module_needed).exists():
327
+ get_cached_module_file(
328
+ pretrained_model_name_or_path,
329
+ f"{module_needed}.py",
330
+ cache_dir=cache_dir,
331
+ force_download=force_download,
332
+ resume_download=resume_download,
333
+ proxies=proxies,
334
+ use_auth_token=use_auth_token,
335
+ revision=revision,
336
+ local_files_only=local_files_only,
337
+ )
338
+ return os.path.join(full_submodule, module_file)
339
+
340
+
341
+ def get_class_from_dynamic_module(
342
+ pretrained_model_name_or_path: Union[str, os.PathLike],
343
+ module_file: str,
344
+ class_name: Optional[str] = None,
345
+ cache_dir: Optional[Union[str, os.PathLike]] = None,
346
+ force_download: bool = False,
347
+ resume_download: bool = False,
348
+ proxies: Optional[Dict[str, str]] = None,
349
+ use_auth_token: Optional[Union[bool, str]] = None,
350
+ revision: Optional[str] = None,
351
+ local_files_only: bool = False,
352
+ **kwargs,
353
+ ):
354
+ """
355
+ Extracts a class from a module file, present in the local folder or repository of a model.
356
+
357
+ <Tip warning={true}>
358
+
359
+ Calling this function will execute the code in the module file found locally or downloaded from the Hub. It should
360
+ therefore only be called on trusted repos.
361
+
362
+ </Tip>
363
+
364
+ Args:
365
+ pretrained_model_name_or_path (`str` or `os.PathLike`):
366
+ This can be either:
367
+
368
+ - a string, the *model id* of a pretrained model configuration hosted inside a model repo on
369
+ huggingface.co. Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced
370
+ under a user or organization name, like `dbmdz/bert-base-german-cased`.
371
+ - a path to a *directory* containing a configuration file saved using the
372
+ [`~PreTrainedTokenizer.save_pretrained`] method, e.g., `./my_model_directory/`.
373
+
374
+ module_file (`str`):
375
+ The name of the module file containing the class to look for.
376
+ class_name (`str`):
377
+ The name of the class to import in the module.
378
+ cache_dir (`str` or `os.PathLike`, *optional*):
379
+ Path to a directory in which a downloaded pretrained model configuration should be cached if the standard
380
+ cache should not be used.
381
+ force_download (`bool`, *optional*, defaults to `False`):
382
+ Whether or not to force to (re-)download the configuration files and override the cached versions if they
383
+ exist.
384
+ resume_download (`bool`, *optional*, defaults to `False`):
385
+ Whether or not to delete incompletely received file. Attempts to resume the download if such a file exists.
386
+ proxies (`Dict[str, str]`, *optional*):
387
+ A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
388
+ 'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
389
+ use_auth_token (`str` or `bool`, *optional*):
390
+ The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
391
+ when running `transformers-cli login` (stored in `~/.huggingface`).
392
+ revision (`str`, *optional*, defaults to `"main"`):
393
+ The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
394
+ git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
395
+ identifier allowed by git.
396
+ local_files_only (`bool`, *optional*, defaults to `False`):
397
+ If `True`, will only try to load the tokenizer configuration from local files.
398
+
399
+ <Tip>
400
+
401
+ You may pass a token in `use_auth_token` if you are not logged in (`huggingface-cli long`) and want to use private
402
+ or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models).
403
+
404
+ </Tip>
405
+
406
+ Returns:
407
+ `type`: The class, dynamically imported from the module.
408
+
409
+ Examples:
410
+
411
+ ```python
412
+ # Download module `modeling.py` from huggingface.co and cache then extract the class `MyBertModel` from this
413
+ # module.
414
+ cls = get_class_from_dynamic_module("sgugger/my-bert-model", "modeling.py", "MyBertModel")
415
+ ```"""
416
+ # And lastly we get the class inside our newly created module
417
+ final_module = get_cached_module_file(
418
+ pretrained_model_name_or_path,
419
+ module_file,
420
+ cache_dir=cache_dir,
421
+ force_download=force_download,
422
+ resume_download=resume_download,
423
+ proxies=proxies,
424
+ use_auth_token=use_auth_token,
425
+ revision=revision,
426
+ local_files_only=local_files_only,
427
+ )
428
+ return get_class_in_module(class_name, final_module.replace(".py", ""))
diffusers/experimental/README.md ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ # 🧨 Diffusers Experimental
2
+
3
+ We are adding experimental code to support novel applications and usages of the Diffusers library.
4
+ Currently, the following experiments are supported:
5
+ * Reinforcement learning via an implementation of the [Diffuser](https://arxiv.org/abs/2205.09991) model.
diffusers/experimental/__init__.py ADDED
@@ -0,0 +1 @@
 
 
1
+ from .rl import ValueGuidedRLPipeline
diffusers/experimental/rl/__init__.py ADDED
@@ -0,0 +1 @@
 
 
1
+ from .value_guided_sampling import ValueGuidedRLPipeline
diffusers/experimental/rl/value_guided_sampling.py ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright 2022 The HuggingFace Team. All rights reserved.
2
+ #
3
+ # Licensed under the Apache License, Version 2.0 (the "License");
4
+ # you may not use this file except in compliance with the License.
5
+ # You may obtain a copy of the License at
6
+ #
7
+ # http://www.apache.org/licenses/LICENSE-2.0
8
+ #
9
+ # Unless required by applicable law or agreed to in writing, software
10
+ # distributed under the License is distributed on an "AS IS" BASIS,
11
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12
+ # See the License for the specific language governing permissions and
13
+ # limitations under the License.
14
+
15
+ import numpy as np
16
+ import torch
17
+
18
+ import tqdm
19
+
20
+ from ...models.unet_1d import UNet1DModel
21
+ from ...pipeline_utils import DiffusionPipeline
22
+ from ...utils.dummy_pt_objects import DDPMScheduler
23
+
24
+
25
+ class ValueGuidedRLPipeline(DiffusionPipeline):
26
+ def __init__(
27
+ self,
28
+ value_function: UNet1DModel,
29
+ unet: UNet1DModel,
30
+ scheduler: DDPMScheduler,
31
+ env,
32
+ ):
33
+ super().__init__()
34
+ self.value_function = value_function
35
+ self.unet = unet
36
+ self.scheduler = scheduler
37
+ self.env = env
38
+ self.data = env.get_dataset()
39
+ self.means = dict()
40
+ for key in self.data.keys():
41
+ try:
42
+ self.means[key] = self.data[key].mean()
43
+ except:
44
+ pass
45
+ self.stds = dict()
46
+ for key in self.data.keys():
47
+ try:
48
+ self.stds[key] = self.data[key].std()
49
+ except:
50
+ pass
51
+ self.state_dim = env.observation_space.shape[0]
52
+ self.action_dim = env.action_space.shape[0]
53
+
54
+ def normalize(self, x_in, key):
55
+ return (x_in - self.means[key]) / self.stds[key]
56
+
57
+ def de_normalize(self, x_in, key):
58
+ return x_in * self.stds[key] + self.means[key]
59
+
60
+ def to_torch(self, x_in):
61
+ if type(x_in) is dict:
62
+ return {k: self.to_torch(v) for k, v in x_in.items()}
63
+ elif torch.is_tensor(x_in):
64
+ return x_in.to(self.unet.device)
65
+ return torch.tensor(x_in, device=self.unet.device)
66
+
67
+ def reset_x0(self, x_in, cond, act_dim):
68
+ for key, val in cond.items():
69
+ x_in[:, key, act_dim:] = val.clone()
70
+ return x_in
71
+
72
+ def run_diffusion(self, x, conditions, n_guide_steps, scale):
73
+ batch_size = x.shape[0]
74
+ y = None
75
+ for i in tqdm.tqdm(self.scheduler.timesteps):
76
+ # create batch of timesteps to pass into model
77
+ timesteps = torch.full((batch_size,), i, device=self.unet.device, dtype=torch.long)
78
+ for _ in range(n_guide_steps):
79
+ with torch.enable_grad():
80
+ x.requires_grad_()
81
+ y = self.value_function(x.permute(0, 2, 1), timesteps).sample
82
+ grad = torch.autograd.grad([y.sum()], [x])[0]
83
+
84
+ posterior_variance = self.scheduler._get_variance(i)
85
+ model_std = torch.exp(0.5 * posterior_variance)
86
+ grad = model_std * grad
87
+ grad[timesteps < 2] = 0
88
+ x = x.detach()
89
+ x = x + scale * grad
90
+ x = self.reset_x0(x, conditions, self.action_dim)
91
+ prev_x = self.unet(x.permute(0, 2, 1), timesteps).sample.permute(0, 2, 1)
92
+ # TODO: set prediction_type when instantiating the model
93
+ x = self.scheduler.step(prev_x, i, x, predict_epsilon=False)["prev_sample"]
94
+
95
+ # apply conditions to the trajectory
96
+ x = self.reset_x0(x, conditions, self.action_dim)
97
+ x = self.to_torch(x)
98
+ return x, y
99
+
100
+ def __call__(self, obs, batch_size=64, planning_horizon=32, n_guide_steps=2, scale=0.1):
101
+ # normalize the observations and create batch dimension
102
+ obs = self.normalize(obs, "observations")
103
+ obs = obs[None].repeat(batch_size, axis=0)
104
+
105
+ conditions = {0: self.to_torch(obs)}
106
+ shape = (batch_size, planning_horizon, self.state_dim + self.action_dim)
107
+
108
+ # generate initial noise and apply our conditions (to make the trajectories start at current state)
109
+ x1 = torch.randn(shape, device=self.unet.device)
110
+ x = self.reset_x0(x1, conditions, self.action_dim)
111
+ x = self.to_torch(x)
112
+
113
+ # run the diffusion process
114
+ x, y = self.run_diffusion(x, conditions, n_guide_steps, scale)
115
+
116
+ # sort output trajectories by value
117
+ sorted_idx = y.argsort(0, descending=True).squeeze()
118
+ sorted_values = x[sorted_idx]
119
+ actions = sorted_values[:, :, : self.action_dim]
120
+ actions = actions.detach().cpu().numpy()
121
+ denorm_actions = self.de_normalize(actions, key="actions")
122
+
123
+ # select the action with the highest value
124
+ if y is not None:
125
+ selected_index = 0
126
+ else:
127
+ # if we didn't run value guiding, select a random action
128
+ selected_index = np.random.randint(0, batch_size)
129
+ denorm_actions = denorm_actions[selected_index, 0]
130
+ return denorm_actions
diffusers/hub_utils.py ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 The HuggingFace Inc. team.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+
17
+ import os
18
+ import sys
19
+ from pathlib import Path
20
+ from typing import Dict, Optional, Union
21
+ from uuid import uuid4
22
+
23
+ from huggingface_hub import HfFolder, whoami
24
+
25
+ from . import __version__
26
+ from .utils import ENV_VARS_TRUE_VALUES, logging
27
+ from .utils.import_utils import (
28
+ _flax_version,
29
+ _jax_version,
30
+ _onnxruntime_version,
31
+ _torch_version,
32
+ is_flax_available,
33
+ is_modelcards_available,
34
+ is_onnx_available,
35
+ is_torch_available,
36
+ )
37
+
38
+
39
+ if is_modelcards_available():
40
+ from modelcards import CardData, ModelCard
41
+
42
+
43
+ logger = logging.get_logger(__name__)
44
+
45
+
46
+ MODEL_CARD_TEMPLATE_PATH = Path(__file__).parent / "utils" / "model_card_template.md"
47
+ SESSION_ID = uuid4().hex
48
+ DISABLE_TELEMETRY = os.getenv("DISABLE_TELEMETRY", "").upper() in ENV_VARS_TRUE_VALUES
49
+
50
+
51
+ def http_user_agent(user_agent: Union[Dict, str, None] = None) -> str:
52
+ """
53
+ Formats a user-agent string with basic info about a request.
54
+ """
55
+ ua = f"diffusers/{__version__}; python/{sys.version.split()[0]}; session_id/{SESSION_ID}"
56
+ if DISABLE_TELEMETRY:
57
+ return ua + "; telemetry/off"
58
+ if is_torch_available():
59
+ ua += f"; torch/{_torch_version}"
60
+ if is_flax_available():
61
+ ua += f"; jax/{_jax_version}"
62
+ ua += f"; flax/{_flax_version}"
63
+ if is_onnx_available():
64
+ ua += f"; onnxruntime/{_onnxruntime_version}"
65
+ # CI will set this value to True
66
+ if os.environ.get("DIFFUSERS_IS_CI", "").upper() in ENV_VARS_TRUE_VALUES:
67
+ ua += "; is_ci/true"
68
+ if isinstance(user_agent, dict):
69
+ ua += "; " + "; ".join(f"{k}/{v}" for k, v in user_agent.items())
70
+ elif isinstance(user_agent, str):
71
+ ua += "; " + user_agent
72
+ return ua
73
+
74
+
75
+ def get_full_repo_name(model_id: str, organization: Optional[str] = None, token: Optional[str] = None):
76
+ if token is None:
77
+ token = HfFolder.get_token()
78
+ if organization is None:
79
+ username = whoami(token)["name"]
80
+ return f"{username}/{model_id}"
81
+ else:
82
+ return f"{organization}/{model_id}"
83
+
84
+
85
+ def create_model_card(args, model_name):
86
+ if not is_modelcards_available:
87
+ raise ValueError(
88
+ "Please make sure to have `modelcards` installed when using the `create_model_card` function. You can"
89
+ " install the package with `pip install modelcards`."
90
+ )
91
+
92
+ if hasattr(args, "local_rank") and args.local_rank not in [-1, 0]:
93
+ return
94
+
95
+ hub_token = args.hub_token if hasattr(args, "hub_token") else None
96
+ repo_name = get_full_repo_name(model_name, token=hub_token)
97
+
98
+ model_card = ModelCard.from_template(
99
+ card_data=CardData( # Card metadata object that will be converted to YAML block
100
+ language="en",
101
+ license="apache-2.0",
102
+ library_name="diffusers",
103
+ tags=[],
104
+ datasets=args.dataset_name,
105
+ metrics=[],
106
+ ),
107
+ template_path=MODEL_CARD_TEMPLATE_PATH,
108
+ model_name=model_name,
109
+ repo_name=repo_name,
110
+ dataset_name=args.dataset_name if hasattr(args, "dataset_name") else None,
111
+ learning_rate=args.learning_rate,
112
+ train_batch_size=args.train_batch_size,
113
+ eval_batch_size=args.eval_batch_size,
114
+ gradient_accumulation_steps=args.gradient_accumulation_steps
115
+ if hasattr(args, "gradient_accumulation_steps")
116
+ else None,
117
+ adam_beta1=args.adam_beta1 if hasattr(args, "adam_beta1") else None,
118
+ adam_beta2=args.adam_beta2 if hasattr(args, "adam_beta2") else None,
119
+ adam_weight_decay=args.adam_weight_decay if hasattr(args, "adam_weight_decay") else None,
120
+ adam_epsilon=args.adam_epsilon if hasattr(args, "adam_epsilon") else None,
121
+ lr_scheduler=args.lr_scheduler if hasattr(args, "lr_scheduler") else None,
122
+ lr_warmup_steps=args.lr_warmup_steps if hasattr(args, "lr_warmup_steps") else None,
123
+ ema_inv_gamma=args.ema_inv_gamma if hasattr(args, "ema_inv_gamma") else None,
124
+ ema_power=args.ema_power if hasattr(args, "ema_power") else None,
125
+ ema_max_decay=args.ema_max_decay if hasattr(args, "ema_max_decay") else None,
126
+ mixed_precision=args.mixed_precision,
127
+ )
128
+
129
+ card_path = os.path.join(args.output_dir, "README.md")
130
+ model_card.save(card_path)
diffusers/modeling_flax_pytorch_utils.py ADDED
@@ -0,0 +1,117 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 The HuggingFace Inc. team.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+ """ PyTorch - Flax general utilities."""
16
+ import re
17
+
18
+ import jax.numpy as jnp
19
+ from flax.traverse_util import flatten_dict, unflatten_dict
20
+ from jax.random import PRNGKey
21
+
22
+ from .utils import logging
23
+
24
+
25
+ logger = logging.get_logger(__name__)
26
+
27
+
28
+ def rename_key(key):
29
+ regex = r"\w+[.]\d+"
30
+ pats = re.findall(regex, key)
31
+ for pat in pats:
32
+ key = key.replace(pat, "_".join(pat.split(".")))
33
+ return key
34
+
35
+
36
+ #####################
37
+ # PyTorch => Flax #
38
+ #####################
39
+
40
+ # Adapted from https://github.com/huggingface/transformers/blob/c603c80f46881ae18b2ca50770ef65fa4033eacd/src/transformers/modeling_flax_pytorch_utils.py#L69
41
+ # and https://github.com/patil-suraj/stable-diffusion-jax/blob/main/stable_diffusion_jax/convert_diffusers_to_jax.py
42
+ def rename_key_and_reshape_tensor(pt_tuple_key, pt_tensor, random_flax_state_dict):
43
+ """Rename PT weight names to corresponding Flax weight names and reshape tensor if necessary"""
44
+
45
+ # conv norm or layer norm
46
+ renamed_pt_tuple_key = pt_tuple_key[:-1] + ("scale",)
47
+ if (
48
+ any("norm" in str_ for str_ in pt_tuple_key)
49
+ and (pt_tuple_key[-1] == "bias")
50
+ and (pt_tuple_key[:-1] + ("bias",) not in random_flax_state_dict)
51
+ and (pt_tuple_key[:-1] + ("scale",) in random_flax_state_dict)
52
+ ):
53
+ renamed_pt_tuple_key = pt_tuple_key[:-1] + ("scale",)
54
+ return renamed_pt_tuple_key, pt_tensor
55
+ elif pt_tuple_key[-1] in ["weight", "gamma"] and pt_tuple_key[:-1] + ("scale",) in random_flax_state_dict:
56
+ renamed_pt_tuple_key = pt_tuple_key[:-1] + ("scale",)
57
+ return renamed_pt_tuple_key, pt_tensor
58
+
59
+ # embedding
60
+ if pt_tuple_key[-1] == "weight" and pt_tuple_key[:-1] + ("embedding",) in random_flax_state_dict:
61
+ pt_tuple_key = pt_tuple_key[:-1] + ("embedding",)
62
+ return renamed_pt_tuple_key, pt_tensor
63
+
64
+ # conv layer
65
+ renamed_pt_tuple_key = pt_tuple_key[:-1] + ("kernel",)
66
+ if pt_tuple_key[-1] == "weight" and pt_tensor.ndim == 4:
67
+ pt_tensor = pt_tensor.transpose(2, 3, 1, 0)
68
+ return renamed_pt_tuple_key, pt_tensor
69
+
70
+ # linear layer
71
+ renamed_pt_tuple_key = pt_tuple_key[:-1] + ("kernel",)
72
+ if pt_tuple_key[-1] == "weight":
73
+ pt_tensor = pt_tensor.T
74
+ return renamed_pt_tuple_key, pt_tensor
75
+
76
+ # old PyTorch layer norm weight
77
+ renamed_pt_tuple_key = pt_tuple_key[:-1] + ("weight",)
78
+ if pt_tuple_key[-1] == "gamma":
79
+ return renamed_pt_tuple_key, pt_tensor
80
+
81
+ # old PyTorch layer norm bias
82
+ renamed_pt_tuple_key = pt_tuple_key[:-1] + ("bias",)
83
+ if pt_tuple_key[-1] == "beta":
84
+ return renamed_pt_tuple_key, pt_tensor
85
+
86
+ return pt_tuple_key, pt_tensor
87
+
88
+
89
+ def convert_pytorch_state_dict_to_flax(pt_state_dict, flax_model, init_key=42):
90
+ # Step 1: Convert pytorch tensor to numpy
91
+ pt_state_dict = {k: v.numpy() for k, v in pt_state_dict.items()}
92
+
93
+ # Step 2: Since the model is stateless, get random Flax params
94
+ random_flax_params = flax_model.init_weights(PRNGKey(init_key))
95
+
96
+ random_flax_state_dict = flatten_dict(random_flax_params)
97
+ flax_state_dict = {}
98
+
99
+ # Need to change some parameters name to match Flax names
100
+ for pt_key, pt_tensor in pt_state_dict.items():
101
+ renamed_pt_key = rename_key(pt_key)
102
+ pt_tuple_key = tuple(renamed_pt_key.split("."))
103
+
104
+ # Correctly rename weight parameters
105
+ flax_key, flax_tensor = rename_key_and_reshape_tensor(pt_tuple_key, pt_tensor, random_flax_state_dict)
106
+
107
+ if flax_key in random_flax_state_dict:
108
+ if flax_tensor.shape != random_flax_state_dict[flax_key].shape:
109
+ raise ValueError(
110
+ f"PyTorch checkpoint seems to be incorrect. Weight {pt_key} was expected to be of shape "
111
+ f"{random_flax_state_dict[flax_key].shape}, but is {flax_tensor.shape}."
112
+ )
113
+
114
+ # also add unexpected weight so that warning is thrown
115
+ flax_state_dict[flax_key] = jnp.asarray(flax_tensor)
116
+
117
+ return unflatten_dict(flax_state_dict)
diffusers/modeling_flax_utils.py ADDED
@@ -0,0 +1,526 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 The HuggingFace Inc. team.
3
+ #
4
+ # Licensed under the Apache License, Version 2.0 (the "License");
5
+ # you may not use this file except in compliance with the License.
6
+ # You may obtain a copy of the License at
7
+ #
8
+ # http://www.apache.org/licenses/LICENSE-2.0
9
+ #
10
+ # Unless required by applicable law or agreed to in writing, software
11
+ # distributed under the License is distributed on an "AS IS" BASIS,
12
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
+ # See the License for the specific language governing permissions and
14
+ # limitations under the License.
15
+
16
+ import os
17
+ from pickle import UnpicklingError
18
+ from typing import Any, Dict, Union
19
+
20
+ import jax
21
+ import jax.numpy as jnp
22
+ import msgpack.exceptions
23
+ from flax.core.frozen_dict import FrozenDict, unfreeze
24
+ from flax.serialization import from_bytes, to_bytes
25
+ from flax.traverse_util import flatten_dict, unflatten_dict
26
+ from huggingface_hub import hf_hub_download
27
+ from huggingface_hub.utils import EntryNotFoundError, RepositoryNotFoundError, RevisionNotFoundError
28
+ from requests import HTTPError
29
+
30
+ from . import __version__, is_torch_available
31
+ from .modeling_flax_pytorch_utils import convert_pytorch_state_dict_to_flax
32
+ from .utils import (
33
+ CONFIG_NAME,
34
+ DIFFUSERS_CACHE,
35
+ FLAX_WEIGHTS_NAME,
36
+ HUGGINGFACE_CO_RESOLVE_ENDPOINT,
37
+ WEIGHTS_NAME,
38
+ logging,
39
+ )
40
+
41
+
42
+ logger = logging.get_logger(__name__)
43
+
44
+
45
+ class FlaxModelMixin:
46
+ r"""
47
+ Base class for all flax models.
48
+
49
+ [`FlaxModelMixin`] takes care of storing the configuration of the models and handles methods for loading,
50
+ downloading and saving models.
51
+ """
52
+ config_name = CONFIG_NAME
53
+ _automatically_saved_args = ["_diffusers_version", "_class_name", "_name_or_path"]
54
+ _flax_internal_args = ["name", "parent", "dtype"]
55
+
56
+ @classmethod
57
+ def _from_config(cls, config, **kwargs):
58
+ """
59
+ All context managers that the model should be initialized under go here.
60
+ """
61
+ return cls(config, **kwargs)
62
+
63
+ def _cast_floating_to(self, params: Union[Dict, FrozenDict], dtype: jnp.dtype, mask: Any = None) -> Any:
64
+ """
65
+ Helper method to cast floating-point values of given parameter `PyTree` to given `dtype`.
66
+ """
67
+
68
+ # taken from https://github.com/deepmind/jmp/blob/3a8318abc3292be38582794dbf7b094e6583b192/jmp/_src/policy.py#L27
69
+ def conditional_cast(param):
70
+ if isinstance(param, jnp.ndarray) and jnp.issubdtype(param.dtype, jnp.floating):
71
+ param = param.astype(dtype)
72
+ return param
73
+
74
+ if mask is None:
75
+ return jax.tree_map(conditional_cast, params)
76
+
77
+ flat_params = flatten_dict(params)
78
+ flat_mask, _ = jax.tree_flatten(mask)
79
+
80
+ for masked, key in zip(flat_mask, flat_params.keys()):
81
+ if masked:
82
+ param = flat_params[key]
83
+ flat_params[key] = conditional_cast(param)
84
+
85
+ return unflatten_dict(flat_params)
86
+
87
+ def to_bf16(self, params: Union[Dict, FrozenDict], mask: Any = None):
88
+ r"""
89
+ Cast the floating-point `params` to `jax.numpy.bfloat16`. This returns a new `params` tree and does not cast
90
+ the `params` in place.
91
+
92
+ This method can be used on TPU to explicitly convert the model parameters to bfloat16 precision to do full
93
+ half-precision training or to save weights in bfloat16 for inference in order to save memory and improve speed.
94
+
95
+ Arguments:
96
+ params (`Union[Dict, FrozenDict]`):
97
+ A `PyTree` of model parameters.
98
+ mask (`Union[Dict, FrozenDict]`):
99
+ A `PyTree` with same structure as the `params` tree. The leaves should be booleans, `True` for params
100
+ you want to cast, and should be `False` for those you want to skip.
101
+
102
+ Examples:
103
+
104
+ ```python
105
+ >>> from diffusers import FlaxUNet2DConditionModel
106
+
107
+ >>> # load model
108
+ >>> model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")
109
+ >>> # By default, the model parameters will be in fp32 precision, to cast these to bfloat16 precision
110
+ >>> params = model.to_bf16(params)
111
+ >>> # If you don't want to cast certain parameters (for example layer norm bias and scale)
112
+ >>> # then pass the mask as follows
113
+ >>> from flax import traverse_util
114
+
115
+ >>> model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")
116
+ >>> flat_params = traverse_util.flatten_dict(params)
117
+ >>> mask = {
118
+ ... path: (path[-2] != ("LayerNorm", "bias") and path[-2:] != ("LayerNorm", "scale"))
119
+ ... for path in flat_params
120
+ ... }
121
+ >>> mask = traverse_util.unflatten_dict(mask)
122
+ >>> params = model.to_bf16(params, mask)
123
+ ```"""
124
+ return self._cast_floating_to(params, jnp.bfloat16, mask)
125
+
126
+ def to_fp32(self, params: Union[Dict, FrozenDict], mask: Any = None):
127
+ r"""
128
+ Cast the floating-point `params` to `jax.numpy.float32`. This method can be used to explicitly convert the
129
+ model parameters to fp32 precision. This returns a new `params` tree and does not cast the `params` in place.
130
+
131
+ Arguments:
132
+ params (`Union[Dict, FrozenDict]`):
133
+ A `PyTree` of model parameters.
134
+ mask (`Union[Dict, FrozenDict]`):
135
+ A `PyTree` with same structure as the `params` tree. The leaves should be booleans, `True` for params
136
+ you want to cast, and should be `False` for those you want to skip
137
+
138
+ Examples:
139
+
140
+ ```python
141
+ >>> from diffusers import FlaxUNet2DConditionModel
142
+
143
+ >>> # Download model and configuration from huggingface.co
144
+ >>> model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")
145
+ >>> # By default, the model params will be in fp32, to illustrate the use of this method,
146
+ >>> # we'll first cast to fp16 and back to fp32
147
+ >>> params = model.to_f16(params)
148
+ >>> # now cast back to fp32
149
+ >>> params = model.to_fp32(params)
150
+ ```"""
151
+ return self._cast_floating_to(params, jnp.float32, mask)
152
+
153
+ def to_fp16(self, params: Union[Dict, FrozenDict], mask: Any = None):
154
+ r"""
155
+ Cast the floating-point `params` to `jax.numpy.float16`. This returns a new `params` tree and does not cast the
156
+ `params` in place.
157
+
158
+ This method can be used on GPU to explicitly convert the model parameters to float16 precision to do full
159
+ half-precision training or to save weights in float16 for inference in order to save memory and improve speed.
160
+
161
+ Arguments:
162
+ params (`Union[Dict, FrozenDict]`):
163
+ A `PyTree` of model parameters.
164
+ mask (`Union[Dict, FrozenDict]`):
165
+ A `PyTree` with same structure as the `params` tree. The leaves should be booleans, `True` for params
166
+ you want to cast, and should be `False` for those you want to skip
167
+
168
+ Examples:
169
+
170
+ ```python
171
+ >>> from diffusers import FlaxUNet2DConditionModel
172
+
173
+ >>> # load model
174
+ >>> model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")
175
+ >>> # By default, the model params will be in fp32, to cast these to float16
176
+ >>> params = model.to_fp16(params)
177
+ >>> # If you want don't want to cast certain parameters (for example layer norm bias and scale)
178
+ >>> # then pass the mask as follows
179
+ >>> from flax import traverse_util
180
+
181
+ >>> model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")
182
+ >>> flat_params = traverse_util.flatten_dict(params)
183
+ >>> mask = {
184
+ ... path: (path[-2] != ("LayerNorm", "bias") and path[-2:] != ("LayerNorm", "scale"))
185
+ ... for path in flat_params
186
+ ... }
187
+ >>> mask = traverse_util.unflatten_dict(mask)
188
+ >>> params = model.to_fp16(params, mask)
189
+ ```"""
190
+ return self._cast_floating_to(params, jnp.float16, mask)
191
+
192
+ def init_weights(self, rng: jax.random.PRNGKey) -> Dict:
193
+ raise NotImplementedError(f"init_weights method has to be implemented for {self}")
194
+
195
+ @classmethod
196
+ def from_pretrained(
197
+ cls,
198
+ pretrained_model_name_or_path: Union[str, os.PathLike],
199
+ dtype: jnp.dtype = jnp.float32,
200
+ *model_args,
201
+ **kwargs,
202
+ ):
203
+ r"""
204
+ Instantiate a pretrained flax model from a pre-trained model configuration.
205
+
206
+ The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
207
+ pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
208
+ task.
209
+
210
+ The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
211
+ weights are discarded.
212
+
213
+ Parameters:
214
+ pretrained_model_name_or_path (`str` or `os.PathLike`):
215
+ Can be either:
216
+
217
+ - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
218
+ Valid model ids are namespaced under a user or organization name, like
219
+ `runwayml/stable-diffusion-v1-5`.
220
+ - A path to a *directory* containing model weights saved using [`~ModelMixin.save_pretrained`],
221
+ e.g., `./my_model_directory/`.
222
+ dtype (`jax.numpy.dtype`, *optional*, defaults to `jax.numpy.float32`):
223
+ The data type of the computation. Can be one of `jax.numpy.float32`, `jax.numpy.float16` (on GPUs) and
224
+ `jax.numpy.bfloat16` (on TPUs).
225
+
226
+ This can be used to enable mixed-precision training or half-precision inference on GPUs or TPUs. If
227
+ specified all the computation will be performed with the given `dtype`.
228
+
229
+ **Note that this only specifies the dtype of the computation and does not influence the dtype of model
230
+ parameters.**
231
+
232
+ If you wish to change the dtype of the model parameters, see [`~ModelMixin.to_fp16`] and
233
+ [`~ModelMixin.to_bf16`].
234
+ model_args (sequence of positional arguments, *optional*):
235
+ All remaining positional arguments will be passed to the underlying model's `__init__` method.
236
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
237
+ Path to a directory in which a downloaded pretrained model configuration should be cached if the
238
+ standard cache should not be used.
239
+ force_download (`bool`, *optional*, defaults to `False`):
240
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
241
+ cached versions if they exist.
242
+ resume_download (`bool`, *optional*, defaults to `False`):
243
+ Whether or not to delete incompletely received files. Will attempt to resume the download if such a
244
+ file exists.
245
+ proxies (`Dict[str, str]`, *optional*):
246
+ A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
247
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
248
+ local_files_only(`bool`, *optional*, defaults to `False`):
249
+ Whether or not to only look at local files (i.e., do not try to download the model).
250
+ revision (`str`, *optional*, defaults to `"main"`):
251
+ The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
252
+ git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
253
+ identifier allowed by git.
254
+ from_pt (`bool`, *optional*, defaults to `False`):
255
+ Load the model weights from a PyTorch checkpoint save file.
256
+ kwargs (remaining dictionary of keyword arguments, *optional*):
257
+ Can be used to update the configuration object (after it being loaded) and initiate the model (e.g.,
258
+ `output_attentions=True`). Behaves differently depending on whether a `config` is provided or
259
+ automatically loaded:
260
+
261
+ - If a configuration is provided with `config`, `**kwargs` will be directly passed to the
262
+ underlying model's `__init__` method (we assume all relevant updates to the configuration have
263
+ already been done)
264
+ - If a configuration is not provided, `kwargs` will be first passed to the configuration class
265
+ initialization function ([`~ConfigMixin.from_config`]). Each key of `kwargs` that corresponds to
266
+ a configuration attribute will be used to override said attribute with the supplied `kwargs`
267
+ value. Remaining keys that do not correspond to any configuration attribute will be passed to the
268
+ underlying model's `__init__` function.
269
+
270
+ Examples:
271
+
272
+ ```python
273
+ >>> from diffusers import FlaxUNet2DConditionModel
274
+
275
+ >>> # Download model and configuration from huggingface.co and cache.
276
+ >>> model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")
277
+ >>> # Model was saved using *save_pretrained('./test/saved_model/')* (for example purposes, not runnable).
278
+ >>> model, params = FlaxUNet2DConditionModel.from_pretrained("./test/saved_model/")
279
+ ```"""
280
+ config = kwargs.pop("config", None)
281
+ cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
282
+ force_download = kwargs.pop("force_download", False)
283
+ from_pt = kwargs.pop("from_pt", False)
284
+ resume_download = kwargs.pop("resume_download", False)
285
+ proxies = kwargs.pop("proxies", None)
286
+ local_files_only = kwargs.pop("local_files_only", False)
287
+ use_auth_token = kwargs.pop("use_auth_token", None)
288
+ revision = kwargs.pop("revision", None)
289
+ subfolder = kwargs.pop("subfolder", None)
290
+
291
+ user_agent = {
292
+ "diffusers": __version__,
293
+ "file_type": "model",
294
+ "framework": "flax",
295
+ }
296
+
297
+ # Load config if we don't provide a configuration
298
+ config_path = config if config is not None else pretrained_model_name_or_path
299
+ model, model_kwargs = cls.from_config(
300
+ config_path,
301
+ cache_dir=cache_dir,
302
+ return_unused_kwargs=True,
303
+ force_download=force_download,
304
+ resume_download=resume_download,
305
+ proxies=proxies,
306
+ local_files_only=local_files_only,
307
+ use_auth_token=use_auth_token,
308
+ revision=revision,
309
+ subfolder=subfolder,
310
+ # model args
311
+ dtype=dtype,
312
+ **kwargs,
313
+ )
314
+
315
+ # Load model
316
+ pretrained_path_with_subfolder = (
317
+ pretrained_model_name_or_path
318
+ if subfolder is None
319
+ else os.path.join(pretrained_model_name_or_path, subfolder)
320
+ )
321
+ if os.path.isdir(pretrained_path_with_subfolder):
322
+ if from_pt:
323
+ if not os.path.isfile(os.path.join(pretrained_path_with_subfolder, WEIGHTS_NAME)):
324
+ raise EnvironmentError(
325
+ f"Error no file named {WEIGHTS_NAME} found in directory {pretrained_path_with_subfolder} "
326
+ )
327
+ model_file = os.path.join(pretrained_path_with_subfolder, WEIGHTS_NAME)
328
+ elif os.path.isfile(os.path.join(pretrained_path_with_subfolder, FLAX_WEIGHTS_NAME)):
329
+ # Load from a Flax checkpoint
330
+ model_file = os.path.join(pretrained_path_with_subfolder, FLAX_WEIGHTS_NAME)
331
+ # Check if pytorch weights exist instead
332
+ elif os.path.isfile(os.path.join(pretrained_path_with_subfolder, WEIGHTS_NAME)):
333
+ raise EnvironmentError(
334
+ f"{WEIGHTS_NAME} file found in directory {pretrained_path_with_subfolder}. Please load the model"
335
+ " using `from_pt=True`."
336
+ )
337
+ else:
338
+ raise EnvironmentError(
339
+ f"Error no file named {FLAX_WEIGHTS_NAME} or {WEIGHTS_NAME} found in directory "
340
+ f"{pretrained_path_with_subfolder}."
341
+ )
342
+ else:
343
+ try:
344
+ model_file = hf_hub_download(
345
+ pretrained_model_name_or_path,
346
+ filename=FLAX_WEIGHTS_NAME if not from_pt else WEIGHTS_NAME,
347
+ cache_dir=cache_dir,
348
+ force_download=force_download,
349
+ proxies=proxies,
350
+ resume_download=resume_download,
351
+ local_files_only=local_files_only,
352
+ use_auth_token=use_auth_token,
353
+ user_agent=user_agent,
354
+ subfolder=subfolder,
355
+ revision=revision,
356
+ )
357
+
358
+ except RepositoryNotFoundError:
359
+ raise EnvironmentError(
360
+ f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier "
361
+ "listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a "
362
+ "token having permission to this repo with `use_auth_token` or log in with `huggingface-cli "
363
+ "login`."
364
+ )
365
+ except RevisionNotFoundError:
366
+ raise EnvironmentError(
367
+ f"{revision} is not a valid git identifier (branch name, tag name or commit id) that exists for "
368
+ "this model name. Check the model page at "
369
+ f"'https://huggingface.co/{pretrained_model_name_or_path}' for available revisions."
370
+ )
371
+ except EntryNotFoundError:
372
+ raise EnvironmentError(
373
+ f"{pretrained_model_name_or_path} does not appear to have a file named {FLAX_WEIGHTS_NAME}."
374
+ )
375
+ except HTTPError as err:
376
+ raise EnvironmentError(
377
+ f"There was a specific connection error when trying to load {pretrained_model_name_or_path}:\n"
378
+ f"{err}"
379
+ )
380
+ except ValueError:
381
+ raise EnvironmentError(
382
+ f"We couldn't connect to '{HUGGINGFACE_CO_RESOLVE_ENDPOINT}' to load this model, couldn't find it"
383
+ f" in the cached files and it looks like {pretrained_model_name_or_path} is not the path to a"
384
+ f" directory containing a file named {FLAX_WEIGHTS_NAME} or {WEIGHTS_NAME}.\nCheckout your"
385
+ " internet connection or see how to run the library in offline mode at"
386
+ " 'https://huggingface.co/docs/transformers/installation#offline-mode'."
387
+ )
388
+ except EnvironmentError:
389
+ raise EnvironmentError(
390
+ f"Can't load the model for '{pretrained_model_name_or_path}'. If you were trying to load it from "
391
+ "'https://huggingface.co/models', make sure you don't have a local directory with the same name. "
392
+ f"Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a directory "
393
+ f"containing a file named {FLAX_WEIGHTS_NAME} or {WEIGHTS_NAME}."
394
+ )
395
+
396
+ if from_pt:
397
+ if is_torch_available():
398
+ from .modeling_utils import load_state_dict
399
+ else:
400
+ raise EnvironmentError(
401
+ "Can't load the model in PyTorch format because PyTorch is not installed. "
402
+ "Please, install PyTorch or use native Flax weights."
403
+ )
404
+
405
+ # Step 1: Get the pytorch file
406
+ pytorch_model_file = load_state_dict(model_file)
407
+
408
+ # Step 2: Convert the weights
409
+ state = convert_pytorch_state_dict_to_flax(pytorch_model_file, model)
410
+ else:
411
+ try:
412
+ with open(model_file, "rb") as state_f:
413
+ state = from_bytes(cls, state_f.read())
414
+ except (UnpicklingError, msgpack.exceptions.ExtraData) as e:
415
+ try:
416
+ with open(model_file) as f:
417
+ if f.read().startswith("version"):
418
+ raise OSError(
419
+ "You seem to have cloned a repository without having git-lfs installed. Please"
420
+ " install git-lfs and run `git lfs install` followed by `git lfs pull` in the"
421
+ " folder you cloned."
422
+ )
423
+ else:
424
+ raise ValueError from e
425
+ except (UnicodeDecodeError, ValueError):
426
+ raise EnvironmentError(f"Unable to convert {model_file} to Flax deserializable object. ")
427
+ # make sure all arrays are stored as jnp.ndarray
428
+ # NOTE: This is to prevent a bug this will be fixed in Flax >= v0.3.4:
429
+ # https://github.com/google/flax/issues/1261
430
+ state = jax.tree_util.tree_map(lambda x: jax.device_put(x, jax.devices("cpu")[0]), state)
431
+
432
+ # flatten dicts
433
+ state = flatten_dict(state)
434
+
435
+ params_shape_tree = jax.eval_shape(model.init_weights, rng=jax.random.PRNGKey(0))
436
+ required_params = set(flatten_dict(unfreeze(params_shape_tree)).keys())
437
+
438
+ shape_state = flatten_dict(unfreeze(params_shape_tree))
439
+
440
+ missing_keys = required_params - set(state.keys())
441
+ unexpected_keys = set(state.keys()) - required_params
442
+
443
+ if missing_keys:
444
+ logger.warning(
445
+ f"The checkpoint {pretrained_model_name_or_path} is missing required keys: {missing_keys}. "
446
+ "Make sure to call model.init_weights to initialize the missing weights."
447
+ )
448
+ cls._missing_keys = missing_keys
449
+
450
+ for key in state.keys():
451
+ if key in shape_state and state[key].shape != shape_state[key].shape:
452
+ raise ValueError(
453
+ f"Trying to load the pretrained weight for {key} failed: checkpoint has shape "
454
+ f"{state[key].shape} which is incompatible with the model shape {shape_state[key].shape}. "
455
+ )
456
+
457
+ # remove unexpected keys to not be saved again
458
+ for unexpected_key in unexpected_keys:
459
+ del state[unexpected_key]
460
+
461
+ if len(unexpected_keys) > 0:
462
+ logger.warning(
463
+ f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when"
464
+ f" initializing {model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are"
465
+ f" initializing {model.__class__.__name__} from the checkpoint of a model trained on another task or"
466
+ " with another architecture."
467
+ )
468
+ else:
469
+ logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")
470
+
471
+ if len(missing_keys) > 0:
472
+ logger.warning(
473
+ f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
474
+ f" {pretrained_model_name_or_path} and are newly initialized: {missing_keys}\nYou should probably"
475
+ " TRAIN this model on a down-stream task to be able to use it for predictions and inference."
476
+ )
477
+ else:
478
+ logger.info(
479
+ f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at"
480
+ f" {pretrained_model_name_or_path}.\nIf your task is similar to the task the model of the checkpoint"
481
+ f" was trained on, you can already use {model.__class__.__name__} for predictions without further"
482
+ " training."
483
+ )
484
+
485
+ return model, unflatten_dict(state)
486
+
487
+ def save_pretrained(
488
+ self,
489
+ save_directory: Union[str, os.PathLike],
490
+ params: Union[Dict, FrozenDict],
491
+ is_main_process: bool = True,
492
+ ):
493
+ """
494
+ Save a model and its configuration file to a directory, so that it can be re-loaded using the
495
+ `[`~FlaxModelMixin.from_pretrained`]` class method
496
+
497
+ Arguments:
498
+ save_directory (`str` or `os.PathLike`):
499
+ Directory to which to save. Will be created if it doesn't exist.
500
+ params (`Union[Dict, FrozenDict]`):
501
+ A `PyTree` of model parameters.
502
+ is_main_process (`bool`, *optional*, defaults to `True`):
503
+ Whether the process calling this is the main process or not. Useful when in distributed training like
504
+ TPUs and need to call this function on all processes. In this case, set `is_main_process=True` only on
505
+ the main process to avoid race conditions.
506
+ """
507
+ if os.path.isfile(save_directory):
508
+ logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
509
+ return
510
+
511
+ os.makedirs(save_directory, exist_ok=True)
512
+
513
+ model_to_save = self
514
+
515
+ # Attach architecture to the config
516
+ # Save the config
517
+ if is_main_process:
518
+ model_to_save.save_config(save_directory)
519
+
520
+ # save model
521
+ output_model_file = os.path.join(save_directory, FLAX_WEIGHTS_NAME)
522
+ with open(output_model_file, "wb") as f:
523
+ model_bytes = to_bytes(params)
524
+ f.write(model_bytes)
525
+
526
+ logger.info(f"Model weights saved in {output_model_file}")
diffusers/modeling_utils.py ADDED
@@ -0,0 +1,764 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2022 The HuggingFace Inc. team.
3
+ # Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+ # limitations under the License.
16
+
17
+ import os
18
+ from functools import partial
19
+ from typing import Callable, List, Optional, Tuple, Union
20
+
21
+ import torch
22
+ from torch import Tensor, device
23
+
24
+ from huggingface_hub import hf_hub_download
25
+ from huggingface_hub.utils import EntryNotFoundError, RepositoryNotFoundError, RevisionNotFoundError
26
+ from requests import HTTPError
27
+
28
+ from . import __version__
29
+ from .utils import (
30
+ CONFIG_NAME,
31
+ DIFFUSERS_CACHE,
32
+ HUGGINGFACE_CO_RESOLVE_ENDPOINT,
33
+ SAFETENSORS_WEIGHTS_NAME,
34
+ WEIGHTS_NAME,
35
+ is_accelerate_available,
36
+ is_safetensors_available,
37
+ is_torch_version,
38
+ logging,
39
+ )
40
+
41
+
42
+ logger = logging.get_logger(__name__)
43
+
44
+
45
+ if is_torch_version(">=", "1.9.0"):
46
+ _LOW_CPU_MEM_USAGE_DEFAULT = True
47
+ else:
48
+ _LOW_CPU_MEM_USAGE_DEFAULT = False
49
+
50
+
51
+ if is_accelerate_available():
52
+ import accelerate
53
+ from accelerate.utils import set_module_tensor_to_device
54
+ from accelerate.utils.versions import is_torch_version
55
+
56
+ if is_safetensors_available():
57
+ import safetensors
58
+
59
+
60
+ def get_parameter_device(parameter: torch.nn.Module):
61
+ try:
62
+ return next(parameter.parameters()).device
63
+ except StopIteration:
64
+ # For torch.nn.DataParallel compatibility in PyTorch 1.5
65
+
66
+ def find_tensor_attributes(module: torch.nn.Module) -> List[Tuple[str, Tensor]]:
67
+ tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
68
+ return tuples
69
+
70
+ gen = parameter._named_members(get_members_fn=find_tensor_attributes)
71
+ first_tuple = next(gen)
72
+ return first_tuple[1].device
73
+
74
+
75
+ def get_parameter_dtype(parameter: torch.nn.Module):
76
+ try:
77
+ return next(parameter.parameters()).dtype
78
+ except StopIteration:
79
+ # For torch.nn.DataParallel compatibility in PyTorch 1.5
80
+
81
+ def find_tensor_attributes(module: torch.nn.Module) -> List[Tuple[str, Tensor]]:
82
+ tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
83
+ return tuples
84
+
85
+ gen = parameter._named_members(get_members_fn=find_tensor_attributes)
86
+ first_tuple = next(gen)
87
+ return first_tuple[1].dtype
88
+
89
+
90
+ def load_state_dict(checkpoint_file: Union[str, os.PathLike]):
91
+ """
92
+ Reads a checkpoint file, returning properly formatted errors if they arise.
93
+ """
94
+ try:
95
+ if os.path.basename(checkpoint_file) == WEIGHTS_NAME:
96
+ return torch.load(checkpoint_file, map_location="cpu")
97
+ else:
98
+ return safetensors.torch.load_file(checkpoint_file, device="cpu")
99
+ except Exception as e:
100
+ try:
101
+ with open(checkpoint_file) as f:
102
+ if f.read().startswith("version"):
103
+ raise OSError(
104
+ "You seem to have cloned a repository without having git-lfs installed. Please install "
105
+ "git-lfs and run `git lfs install` followed by `git lfs pull` in the folder "
106
+ "you cloned."
107
+ )
108
+ else:
109
+ raise ValueError(
110
+ f"Unable to locate the file {checkpoint_file} which is necessary to load this pretrained "
111
+ "model. Make sure you have saved the model properly."
112
+ ) from e
113
+ except (UnicodeDecodeError, ValueError):
114
+ raise OSError(
115
+ f"Unable to load weights from checkpoint file for '{checkpoint_file}' "
116
+ f"at '{checkpoint_file}'. "
117
+ "If you tried to load a PyTorch model from a TF 2.0 checkpoint, please set from_tf=True."
118
+ )
119
+
120
+
121
+ def _load_state_dict_into_model(model_to_load, state_dict):
122
+ # Convert old format to new format if needed from a PyTorch state_dict
123
+ # copy state_dict so _load_from_state_dict can modify it
124
+ state_dict = state_dict.copy()
125
+ error_msgs = []
126
+
127
+ # PyTorch's `_load_from_state_dict` does not copy parameters in a module's descendants
128
+ # so we need to apply the function recursively.
129
+ def load(module: torch.nn.Module, prefix=""):
130
+ args = (state_dict, prefix, {}, True, [], [], error_msgs)
131
+ module._load_from_state_dict(*args)
132
+
133
+ for name, child in module._modules.items():
134
+ if child is not None:
135
+ load(child, prefix + name + ".")
136
+
137
+ load(model_to_load)
138
+
139
+ return error_msgs
140
+
141
+
142
+ class ModelMixin(torch.nn.Module):
143
+ r"""
144
+ Base class for all models.
145
+
146
+ [`ModelMixin`] takes care of storing the configuration of the models and handles methods for loading, downloading
147
+ and saving models.
148
+
149
+ - **config_name** ([`str`]) -- A filename under which the model should be stored when calling
150
+ [`~modeling_utils.ModelMixin.save_pretrained`].
151
+ """
152
+ config_name = CONFIG_NAME
153
+ _automatically_saved_args = ["_diffusers_version", "_class_name", "_name_or_path"]
154
+ _supports_gradient_checkpointing = False
155
+
156
+ def __init__(self):
157
+ super().__init__()
158
+
159
+ @property
160
+ def is_gradient_checkpointing(self) -> bool:
161
+ """
162
+ Whether gradient checkpointing is activated for this model or not.
163
+
164
+ Note that in other frameworks this feature can be referred to as "activation checkpointing" or "checkpoint
165
+ activations".
166
+ """
167
+ return any(hasattr(m, "gradient_checkpointing") and m.gradient_checkpointing for m in self.modules())
168
+
169
+ def enable_gradient_checkpointing(self):
170
+ """
171
+ Activates gradient checkpointing for the current model.
172
+
173
+ Note that in other frameworks this feature can be referred to as "activation checkpointing" or "checkpoint
174
+ activations".
175
+ """
176
+ if not self._supports_gradient_checkpointing:
177
+ raise ValueError(f"{self.__class__.__name__} does not support gradient checkpointing.")
178
+ self.apply(partial(self._set_gradient_checkpointing, value=True))
179
+
180
+ def disable_gradient_checkpointing(self):
181
+ """
182
+ Deactivates gradient checkpointing for the current model.
183
+
184
+ Note that in other frameworks this feature can be referred to as "activation checkpointing" or "checkpoint
185
+ activations".
186
+ """
187
+ if self._supports_gradient_checkpointing:
188
+ self.apply(partial(self._set_gradient_checkpointing, value=False))
189
+
190
+ def save_pretrained(
191
+ self,
192
+ save_directory: Union[str, os.PathLike],
193
+ is_main_process: bool = True,
194
+ save_function: Callable = None,
195
+ safe_serialization: bool = False,
196
+ ):
197
+ """
198
+ Save a model and its configuration file to a directory, so that it can be re-loaded using the
199
+ `[`~modeling_utils.ModelMixin.from_pretrained`]` class method.
200
+
201
+ Arguments:
202
+ save_directory (`str` or `os.PathLike`):
203
+ Directory to which to save. Will be created if it doesn't exist.
204
+ is_main_process (`bool`, *optional*, defaults to `True`):
205
+ Whether the process calling this is the main process or not. Useful when in distributed training like
206
+ TPUs and need to call this function on all processes. In this case, set `is_main_process=True` only on
207
+ the main process to avoid race conditions.
208
+ save_function (`Callable`):
209
+ The function to use to save the state dictionary. Useful on distributed training like TPUs when one
210
+ need to replace `torch.save` by another method. Can be configured with the environment variable
211
+ `DIFFUSERS_SAVE_MODE`.
212
+ safe_serialization (`bool`, *optional*, defaults to `False`):
213
+ Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
214
+ """
215
+ if safe_serialization and not is_safetensors_available():
216
+ raise ImportError("`safe_serialization` requires the `safetensors library: `pip install safetensors`.")
217
+
218
+ if os.path.isfile(save_directory):
219
+ logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
220
+ return
221
+
222
+ if save_function is None:
223
+ save_function = safetensors.torch.save_file if safe_serialization else torch.save
224
+
225
+ os.makedirs(save_directory, exist_ok=True)
226
+
227
+ model_to_save = self
228
+
229
+ # Attach architecture to the config
230
+ # Save the config
231
+ if is_main_process:
232
+ model_to_save.save_config(save_directory)
233
+
234
+ # Save the model
235
+ state_dict = model_to_save.state_dict()
236
+
237
+ weights_name = SAFETENSORS_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
238
+
239
+ # Clean the folder from a previous save
240
+ for filename in os.listdir(save_directory):
241
+ full_filename = os.path.join(save_directory, filename)
242
+ # If we have a shard file that is not going to be replaced, we delete it, but only from the main process
243
+ # in distributed settings to avoid race conditions.
244
+ weights_no_suffix = weights_name.replace(".bin", "").replace(".safetensors", "")
245
+ if filename.startswith(weights_no_suffix) and os.path.isfile(full_filename) and is_main_process:
246
+ os.remove(full_filename)
247
+
248
+ # Save the model
249
+ save_function(state_dict, os.path.join(save_directory, weights_name))
250
+
251
+ logger.info(f"Model weights saved in {os.path.join(save_directory, weights_name)}")
252
+
253
+ @classmethod
254
+ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
255
+ r"""
256
+ Instantiate a pretrained pytorch model from a pre-trained model configuration.
257
+
258
+ The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated). To train
259
+ the model, you should first set it back in training mode with `model.train()`.
260
+
261
+ The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
262
+ pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
263
+ task.
264
+
265
+ The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
266
+ weights are discarded.
267
+
268
+ Parameters:
269
+ pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
270
+ Can be either:
271
+
272
+ - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.
273
+ Valid model ids should have an organization name, like `google/ddpm-celebahq-256`.
274
+ - A path to a *directory* containing model weights saved using [`~ModelMixin.save_config`], e.g.,
275
+ `./my_model_directory/`.
276
+
277
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
278
+ Path to a directory in which a downloaded pretrained model configuration should be cached if the
279
+ standard cache should not be used.
280
+ torch_dtype (`str` or `torch.dtype`, *optional*):
281
+ Override the default `torch.dtype` and load the model under this dtype. If `"auto"` is passed the dtype
282
+ will be automatically derived from the model's weights.
283
+ force_download (`bool`, *optional*, defaults to `False`):
284
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
285
+ cached versions if they exist.
286
+ resume_download (`bool`, *optional*, defaults to `False`):
287
+ Whether or not to delete incompletely received files. Will attempt to resume the download if such a
288
+ file exists.
289
+ proxies (`Dict[str, str]`, *optional*):
290
+ A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
291
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
292
+ output_loading_info(`bool`, *optional*, defaults to `False`):
293
+ Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
294
+ local_files_only(`bool`, *optional*, defaults to `False`):
295
+ Whether or not to only look at local files (i.e., do not try to download the model).
296
+ use_auth_token (`str` or *bool*, *optional*):
297
+ The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
298
+ when running `diffusers-cli login` (stored in `~/.huggingface`).
299
+ revision (`str`, *optional*, defaults to `"main"`):
300
+ The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
301
+ git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
302
+ identifier allowed by git.
303
+ subfolder (`str`, *optional*, defaults to `""`):
304
+ In case the relevant files are located inside a subfolder of the model repo (either remote in
305
+ huggingface.co or downloaded locally), you can specify the folder name here.
306
+
307
+ mirror (`str`, *optional*):
308
+ Mirror source to accelerate downloads in China. If you are from China and have an accessibility
309
+ problem, you can set this option to resolve it. Note that we do not guarantee the timeliness or safety.
310
+ Please refer to the mirror site for more information.
311
+ device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
312
+ A map that specifies where each submodule should go. It doesn't need to be refined to each
313
+ parameter/buffer name, once a given module name is inside, every submodule of it will be sent to the
314
+ same device.
315
+
316
+ To have Accelerate compute the most optimized `device_map` automatically, set `device_map="auto"`. For
317
+ more information about each option see [designing a device
318
+ map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
319
+ low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
320
+ Speed up model loading by not initializing the weights and only loading the pre-trained weights. This
321
+ also tries to not use more than 1x model size in CPU memory (including peak memory) while loading the
322
+ model. This is only supported when torch version >= 1.9.0. If you are using an older version of torch,
323
+ setting this argument to `True` will raise an error.
324
+
325
+ <Tip>
326
+
327
+ It is required to be logged in (`huggingface-cli login`) when you want to use private or [gated
328
+ models](https://huggingface.co/docs/hub/models-gated#gated-models).
329
+
330
+ </Tip>
331
+
332
+ <Tip>
333
+
334
+ Activate the special ["offline-mode"](https://huggingface.co/diffusers/installation.html#offline-mode) to use
335
+ this method in a firewalled environment.
336
+
337
+ </Tip>
338
+
339
+ """
340
+ cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
341
+ ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)
342
+ force_download = kwargs.pop("force_download", False)
343
+ resume_download = kwargs.pop("resume_download", False)
344
+ proxies = kwargs.pop("proxies", None)
345
+ output_loading_info = kwargs.pop("output_loading_info", False)
346
+ local_files_only = kwargs.pop("local_files_only", False)
347
+ use_auth_token = kwargs.pop("use_auth_token", None)
348
+ revision = kwargs.pop("revision", None)
349
+ torch_dtype = kwargs.pop("torch_dtype", None)
350
+ subfolder = kwargs.pop("subfolder", None)
351
+ device_map = kwargs.pop("device_map", None)
352
+ low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
353
+
354
+ if low_cpu_mem_usage and not is_accelerate_available():
355
+ low_cpu_mem_usage = False
356
+ logger.warning(
357
+ "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
358
+ " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
359
+ " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
360
+ " install accelerate\n```\n."
361
+ )
362
+
363
+ if device_map is not None and not is_accelerate_available():
364
+ raise NotImplementedError(
365
+ "Loading and dispatching requires `accelerate`. Please make sure to install accelerate or set"
366
+ " `device_map=None`. You can install accelerate with `pip install accelerate`."
367
+ )
368
+
369
+ # Check if we can handle device_map and dispatching the weights
370
+ if device_map is not None and not is_torch_version(">=", "1.9.0"):
371
+ raise NotImplementedError(
372
+ "Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set"
373
+ " `device_map=None`."
374
+ )
375
+
376
+ if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
377
+ raise NotImplementedError(
378
+ "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
379
+ " `low_cpu_mem_usage=False`."
380
+ )
381
+
382
+ if low_cpu_mem_usage is False and device_map is not None:
383
+ raise ValueError(
384
+ f"You cannot set `low_cpu_mem_usage` to `False` while using device_map={device_map} for loading and"
385
+ " dispatching. Please make sure to set `low_cpu_mem_usage=True`."
386
+ )
387
+
388
+ user_agent = {
389
+ "diffusers": __version__,
390
+ "file_type": "model",
391
+ "framework": "pytorch",
392
+ }
393
+
394
+ # Load config if we don't provide a configuration
395
+ config_path = pretrained_model_name_or_path
396
+
397
+ # This variable will flag if we're loading a sharded checkpoint. In this case the archive file is just the
398
+ # Load model
399
+
400
+ model_file = None
401
+ if is_safetensors_available():
402
+ try:
403
+ model_file = _get_model_file(
404
+ pretrained_model_name_or_path,
405
+ weights_name=SAFETENSORS_WEIGHTS_NAME,
406
+ cache_dir=cache_dir,
407
+ force_download=force_download,
408
+ resume_download=resume_download,
409
+ proxies=proxies,
410
+ local_files_only=local_files_only,
411
+ use_auth_token=use_auth_token,
412
+ revision=revision,
413
+ subfolder=subfolder,
414
+ user_agent=user_agent,
415
+ )
416
+ except:
417
+ pass
418
+ if model_file is None:
419
+ model_file = _get_model_file(
420
+ pretrained_model_name_or_path,
421
+ weights_name=WEIGHTS_NAME,
422
+ cache_dir=cache_dir,
423
+ force_download=force_download,
424
+ resume_download=resume_download,
425
+ proxies=proxies,
426
+ local_files_only=local_files_only,
427
+ use_auth_token=use_auth_token,
428
+ revision=revision,
429
+ subfolder=subfolder,
430
+ user_agent=user_agent,
431
+ )
432
+
433
+ if low_cpu_mem_usage:
434
+ # Instantiate model with empty weights
435
+ with accelerate.init_empty_weights():
436
+ config, unused_kwargs = cls.load_config(
437
+ config_path,
438
+ cache_dir=cache_dir,
439
+ return_unused_kwargs=True,
440
+ force_download=force_download,
441
+ resume_download=resume_download,
442
+ proxies=proxies,
443
+ local_files_only=local_files_only,
444
+ use_auth_token=use_auth_token,
445
+ revision=revision,
446
+ subfolder=subfolder,
447
+ device_map=device_map,
448
+ **kwargs,
449
+ )
450
+ model = cls.from_config(config, **unused_kwargs)
451
+
452
+ # if device_map is Non,e load the state dict on move the params from meta device to the cpu
453
+ if device_map is None:
454
+ param_device = "cpu"
455
+ state_dict = load_state_dict(model_file)
456
+ # move the parms from meta device to cpu
457
+ for param_name, param in state_dict.items():
458
+ set_module_tensor_to_device(model, param_name, param_device, value=param)
459
+ else: # else let accelerate handle loading and dispatching.
460
+ # Load weights and dispatch according to the device_map
461
+ # by deafult the device_map is None and the weights are loaded on the CPU
462
+ accelerate.load_checkpoint_and_dispatch(model, model_file, device_map)
463
+
464
+ loading_info = {
465
+ "missing_keys": [],
466
+ "unexpected_keys": [],
467
+ "mismatched_keys": [],
468
+ "error_msgs": [],
469
+ }
470
+ else:
471
+ config, unused_kwargs = cls.load_config(
472
+ config_path,
473
+ cache_dir=cache_dir,
474
+ return_unused_kwargs=True,
475
+ force_download=force_download,
476
+ resume_download=resume_download,
477
+ proxies=proxies,
478
+ local_files_only=local_files_only,
479
+ use_auth_token=use_auth_token,
480
+ revision=revision,
481
+ subfolder=subfolder,
482
+ device_map=device_map,
483
+ **kwargs,
484
+ )
485
+ model = cls.from_config(config, **unused_kwargs)
486
+
487
+ state_dict = load_state_dict(model_file)
488
+ dtype = set(v.dtype for v in state_dict.values())
489
+
490
+ if len(dtype) > 1 and torch.float32 not in dtype:
491
+ raise ValueError(
492
+ f"The weights of the model file {model_file} have a mixture of incompatible dtypes {dtype}. Please"
493
+ f" make sure that {model_file} weights have only one dtype."
494
+ )
495
+ elif len(dtype) > 1 and torch.float32 in dtype:
496
+ dtype = torch.float32
497
+ else:
498
+ dtype = dtype.pop()
499
+
500
+ # move model to correct dtype
501
+ model = model.to(dtype)
502
+
503
+ model, missing_keys, unexpected_keys, mismatched_keys, error_msgs = cls._load_pretrained_model(
504
+ model,
505
+ state_dict,
506
+ model_file,
507
+ pretrained_model_name_or_path,
508
+ ignore_mismatched_sizes=ignore_mismatched_sizes,
509
+ )
510
+
511
+ loading_info = {
512
+ "missing_keys": missing_keys,
513
+ "unexpected_keys": unexpected_keys,
514
+ "mismatched_keys": mismatched_keys,
515
+ "error_msgs": error_msgs,
516
+ }
517
+
518
+ if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype):
519
+ raise ValueError(
520
+ f"{torch_dtype} needs to be of type `torch.dtype`, e.g. `torch.float16`, but is {type(torch_dtype)}."
521
+ )
522
+ elif torch_dtype is not None:
523
+ model = model.to(torch_dtype)
524
+
525
+ model.register_to_config(_name_or_path=pretrained_model_name_or_path)
526
+
527
+ # Set model in evaluation mode to deactivate DropOut modules by default
528
+ model.eval()
529
+ if output_loading_info:
530
+ return model, loading_info
531
+
532
+ return model
533
+
534
+ @classmethod
535
+ def _load_pretrained_model(
536
+ cls,
537
+ model,
538
+ state_dict,
539
+ resolved_archive_file,
540
+ pretrained_model_name_or_path,
541
+ ignore_mismatched_sizes=False,
542
+ ):
543
+ # Retrieve missing & unexpected_keys
544
+ model_state_dict = model.state_dict()
545
+ loaded_keys = [k for k in state_dict.keys()]
546
+
547
+ expected_keys = list(model_state_dict.keys())
548
+
549
+ original_loaded_keys = loaded_keys
550
+
551
+ missing_keys = list(set(expected_keys) - set(loaded_keys))
552
+ unexpected_keys = list(set(loaded_keys) - set(expected_keys))
553
+
554
+ # Make sure we are able to load base models as well as derived models (with heads)
555
+ model_to_load = model
556
+
557
+ def _find_mismatched_keys(
558
+ state_dict,
559
+ model_state_dict,
560
+ loaded_keys,
561
+ ignore_mismatched_sizes,
562
+ ):
563
+ mismatched_keys = []
564
+ if ignore_mismatched_sizes:
565
+ for checkpoint_key in loaded_keys:
566
+ model_key = checkpoint_key
567
+
568
+ if (
569
+ model_key in model_state_dict
570
+ and state_dict[checkpoint_key].shape != model_state_dict[model_key].shape
571
+ ):
572
+ mismatched_keys.append(
573
+ (checkpoint_key, state_dict[checkpoint_key].shape, model_state_dict[model_key].shape)
574
+ )
575
+ del state_dict[checkpoint_key]
576
+ return mismatched_keys
577
+
578
+ if state_dict is not None:
579
+ # Whole checkpoint
580
+ mismatched_keys = _find_mismatched_keys(
581
+ state_dict,
582
+ model_state_dict,
583
+ original_loaded_keys,
584
+ ignore_mismatched_sizes,
585
+ )
586
+ error_msgs = _load_state_dict_into_model(model_to_load, state_dict)
587
+
588
+ if len(error_msgs) > 0:
589
+ error_msg = "\n\t".join(error_msgs)
590
+ if "size mismatch" in error_msg:
591
+ error_msg += (
592
+ "\n\tYou may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method."
593
+ )
594
+ raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}")
595
+
596
+ if len(unexpected_keys) > 0:
597
+ logger.warning(
598
+ f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when"
599
+ f" initializing {model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are"
600
+ f" initializing {model.__class__.__name__} from the checkpoint of a model trained on another task"
601
+ " or with another architecture (e.g. initializing a BertForSequenceClassification model from a"
602
+ " BertForPreTraining model).\n- This IS NOT expected if you are initializing"
603
+ f" {model.__class__.__name__} from the checkpoint of a model that you expect to be exactly"
604
+ " identical (initializing a BertForSequenceClassification model from a"
605
+ " BertForSequenceClassification model)."
606
+ )
607
+ else:
608
+ logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")
609
+ if len(missing_keys) > 0:
610
+ logger.warning(
611
+ f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
612
+ f" {pretrained_model_name_or_path} and are newly initialized: {missing_keys}\nYou should probably"
613
+ " TRAIN this model on a down-stream task to be able to use it for predictions and inference."
614
+ )
615
+ elif len(mismatched_keys) == 0:
616
+ logger.info(
617
+ f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at"
618
+ f" {pretrained_model_name_or_path}.\nIf your task is similar to the task the model of the"
619
+ f" checkpoint was trained on, you can already use {model.__class__.__name__} for predictions"
620
+ " without further training."
621
+ )
622
+ if len(mismatched_keys) > 0:
623
+ mismatched_warning = "\n".join(
624
+ [
625
+ f"- {key}: found shape {shape1} in the checkpoint and {shape2} in the model instantiated"
626
+ for key, shape1, shape2 in mismatched_keys
627
+ ]
628
+ )
629
+ logger.warning(
630
+ f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
631
+ f" {pretrained_model_name_or_path} and are newly initialized because the shapes did not"
632
+ f" match:\n{mismatched_warning}\nYou should probably TRAIN this model on a down-stream task to be"
633
+ " able to use it for predictions and inference."
634
+ )
635
+
636
+ return model, missing_keys, unexpected_keys, mismatched_keys, error_msgs
637
+
638
+ @property
639
+ def device(self) -> device:
640
+ """
641
+ `torch.device`: The device on which the module is (assuming that all the module parameters are on the same
642
+ device).
643
+ """
644
+ return get_parameter_device(self)
645
+
646
+ @property
647
+ def dtype(self) -> torch.dtype:
648
+ """
649
+ `torch.dtype`: The dtype of the module (assuming that all the module parameters have the same dtype).
650
+ """
651
+ return get_parameter_dtype(self)
652
+
653
+ def num_parameters(self, only_trainable: bool = False, exclude_embeddings: bool = False) -> int:
654
+ """
655
+ Get number of (optionally, trainable or non-embeddings) parameters in the module.
656
+
657
+ Args:
658
+ only_trainable (`bool`, *optional*, defaults to `False`):
659
+ Whether or not to return only the number of trainable parameters
660
+
661
+ exclude_embeddings (`bool`, *optional*, defaults to `False`):
662
+ Whether or not to return only the number of non-embeddings parameters
663
+
664
+ Returns:
665
+ `int`: The number of parameters.
666
+ """
667
+
668
+ if exclude_embeddings:
669
+ embedding_param_names = [
670
+ f"{name}.weight"
671
+ for name, module_type in self.named_modules()
672
+ if isinstance(module_type, torch.nn.Embedding)
673
+ ]
674
+ non_embedding_parameters = [
675
+ parameter for name, parameter in self.named_parameters() if name not in embedding_param_names
676
+ ]
677
+ return sum(p.numel() for p in non_embedding_parameters if p.requires_grad or not only_trainable)
678
+ else:
679
+ return sum(p.numel() for p in self.parameters() if p.requires_grad or not only_trainable)
680
+
681
+
682
+ def _get_model_file(
683
+ pretrained_model_name_or_path,
684
+ *,
685
+ weights_name,
686
+ subfolder,
687
+ cache_dir,
688
+ force_download,
689
+ proxies,
690
+ resume_download,
691
+ local_files_only,
692
+ use_auth_token,
693
+ user_agent,
694
+ revision,
695
+ ):
696
+ pretrained_model_name_or_path = str(pretrained_model_name_or_path)
697
+ if os.path.isdir(pretrained_model_name_or_path):
698
+ if os.path.isfile(os.path.join(pretrained_model_name_or_path, weights_name)):
699
+ # Load from a PyTorch checkpoint
700
+ model_file = os.path.join(pretrained_model_name_or_path, weights_name)
701
+ return model_file
702
+ elif subfolder is not None and os.path.isfile(
703
+ os.path.join(pretrained_model_name_or_path, subfolder, weights_name)
704
+ ):
705
+ model_file = os.path.join(pretrained_model_name_or_path, subfolder, weights_name)
706
+ return model_file
707
+ else:
708
+ raise EnvironmentError(
709
+ f"Error no file named {weights_name} found in directory {pretrained_model_name_or_path}."
710
+ )
711
+ else:
712
+ try:
713
+ # Load from URL or cache if already cached
714
+ model_file = hf_hub_download(
715
+ pretrained_model_name_or_path,
716
+ filename=weights_name,
717
+ cache_dir=cache_dir,
718
+ force_download=force_download,
719
+ proxies=proxies,
720
+ resume_download=resume_download,
721
+ local_files_only=local_files_only,
722
+ use_auth_token=use_auth_token,
723
+ user_agent=user_agent,
724
+ subfolder=subfolder,
725
+ revision=revision,
726
+ )
727
+ return model_file
728
+
729
+ except RepositoryNotFoundError:
730
+ raise EnvironmentError(
731
+ f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier "
732
+ "listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a "
733
+ "token having permission to this repo with `use_auth_token` or log in with `huggingface-cli "
734
+ "login`."
735
+ )
736
+ except RevisionNotFoundError:
737
+ raise EnvironmentError(
738
+ f"{revision} is not a valid git identifier (branch name, tag name or commit id) that exists for "
739
+ "this model name. Check the model page at "
740
+ f"'https://huggingface.co/{pretrained_model_name_or_path}' for available revisions."
741
+ )
742
+ except EntryNotFoundError:
743
+ raise EnvironmentError(
744
+ f"{pretrained_model_name_or_path} does not appear to have a file named {weights_name}."
745
+ )
746
+ except HTTPError as err:
747
+ raise EnvironmentError(
748
+ f"There was a specific connection error when trying to load {pretrained_model_name_or_path}:\n{err}"
749
+ )
750
+ except ValueError:
751
+ raise EnvironmentError(
752
+ f"We couldn't connect to '{HUGGINGFACE_CO_RESOLVE_ENDPOINT}' to load this model, couldn't find it"
753
+ f" in the cached files and it looks like {pretrained_model_name_or_path} is not the path to a"
754
+ f" directory containing a file named {weights_name} or"
755
+ " \nCheckout your internet connection or see how to run the library in"
756
+ " offline mode at 'https://huggingface.co/docs/diffusers/installation#offline-mode'."
757
+ )
758
+ except EnvironmentError:
759
+ raise EnvironmentError(
760
+ f"Can't load the model for '{pretrained_model_name_or_path}'. If you were trying to load it from "
761
+ "'https://huggingface.co/models', make sure you don't have a local directory with the same name. "
762
+ f"Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a directory "
763
+ f"containing a file named {weights_name}"
764
+ )
diffusers/models/README.md ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ # Models
2
+
3
+ For more detail on the models, please refer to the [docs](https://huggingface.co/docs/diffusers/api/models).