Spaces:
Build error
Build error
Artiprocher
commited on
Commit
•
4822be4
1
Parent(s):
be9a42a
support new model
Browse files- LdmZhPipeline.py +0 -1036
- app.py +9 -11
- requirements.txt +1 -1
LdmZhPipeline.py
DELETED
@@ -1,1036 +0,0 @@
|
|
1 |
-
# coding=utf-8
|
2 |
-
|
3 |
-
import importlib
|
4 |
-
import inspect
|
5 |
-
import os
|
6 |
-
from dataclasses import dataclass
|
7 |
-
from typing import Any, Dict, List, Optional, Union
|
8 |
-
from collections import OrderedDict
|
9 |
-
|
10 |
-
import numpy as np
|
11 |
-
import torch
|
12 |
-
from torch import nn
|
13 |
-
import functools
|
14 |
-
|
15 |
-
import diffusers
|
16 |
-
import PIL
|
17 |
-
from accelerate.utils.versions import is_torch_version
|
18 |
-
from huggingface_hub import snapshot_download
|
19 |
-
from packaging import version
|
20 |
-
from PIL import Image
|
21 |
-
from tqdm.auto import tqdm
|
22 |
-
|
23 |
-
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
24 |
-
from diffusers.dynamic_modules_utils import get_class_from_dynamic_module
|
25 |
-
from diffusers.modeling_utils import ModelMixin
|
26 |
-
from diffusers.hub_utils import http_user_agent
|
27 |
-
from diffusers.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT
|
28 |
-
from diffusers.schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME
|
29 |
-
from diffusers.utils import (
|
30 |
-
CONFIG_NAME,
|
31 |
-
DIFFUSERS_CACHE,
|
32 |
-
ONNX_WEIGHTS_NAME,
|
33 |
-
WEIGHTS_NAME,
|
34 |
-
BaseOutput,
|
35 |
-
deprecate,
|
36 |
-
is_transformers_available,
|
37 |
-
logging,
|
38 |
-
)
|
39 |
-
|
40 |
-
|
41 |
-
if is_transformers_available():
|
42 |
-
import transformers
|
43 |
-
from transformers import PreTrainedModel
|
44 |
-
|
45 |
-
|
46 |
-
INDEX_FILE = "diffusion_pytorch_model.bin"
|
47 |
-
CUSTOM_PIPELINE_FILE_NAME = "pipeline.py"
|
48 |
-
DUMMY_MODULES_FOLDER = "diffusers.utils"
|
49 |
-
|
50 |
-
|
51 |
-
logger = logging.get_logger(__name__)
|
52 |
-
|
53 |
-
|
54 |
-
LOADABLE_CLASSES = {
|
55 |
-
"diffusers": {
|
56 |
-
"ModelMixin": ["save_pretrained", "from_pretrained"],
|
57 |
-
"SchedulerMixin": ["save_config", "from_config"],
|
58 |
-
"DiffusionPipeline": ["save_pretrained", "from_pretrained"],
|
59 |
-
"OnnxRuntimeModel": ["save_pretrained", "from_pretrained"],
|
60 |
-
},
|
61 |
-
"transformers": {
|
62 |
-
"PreTrainedTokenizer": ["save_pretrained", "from_pretrained"],
|
63 |
-
"PreTrainedTokenizerFast": ["save_pretrained", "from_pretrained"],
|
64 |
-
"PreTrainedModel": ["save_pretrained", "from_pretrained"],
|
65 |
-
"FeatureExtractionMixin": ["save_pretrained", "from_pretrained"],
|
66 |
-
},
|
67 |
-
"LdmZhPipeline": {
|
68 |
-
"WukongClipTextEncoder": ["save_pretrained", "from_pretrained"],
|
69 |
-
"ESRGAN": ["save_pretrained", "from_pretrained"],
|
70 |
-
},
|
71 |
-
}
|
72 |
-
|
73 |
-
ALL_IMPORTABLE_CLASSES = {}
|
74 |
-
for library in LOADABLE_CLASSES:
|
75 |
-
ALL_IMPORTABLE_CLASSES.update(LOADABLE_CLASSES[library])
|
76 |
-
|
77 |
-
|
78 |
-
@dataclass
|
79 |
-
class ImagePipelineOutput(BaseOutput):
|
80 |
-
"""
|
81 |
-
Output class for image pipelines.
|
82 |
-
|
83 |
-
Args:
|
84 |
-
images (`List[PIL.Image.Image]` or `np.ndarray`)
|
85 |
-
List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
|
86 |
-
num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
|
87 |
-
"""
|
88 |
-
|
89 |
-
images: Union[List[PIL.Image.Image], np.ndarray]
|
90 |
-
|
91 |
-
|
92 |
-
@dataclass
|
93 |
-
class AudioPipelineOutput(BaseOutput):
|
94 |
-
"""
|
95 |
-
Output class for audio pipelines.
|
96 |
-
|
97 |
-
Args:
|
98 |
-
audios (`np.ndarray`)
|
99 |
-
List of denoised samples of shape `(batch_size, num_channels, sample_rate)`. Numpy array present the
|
100 |
-
denoised audio samples of the diffusion pipeline.
|
101 |
-
"""
|
102 |
-
|
103 |
-
audios: np.ndarray
|
104 |
-
|
105 |
-
|
106 |
-
class DiffusionPipeline(ConfigMixin):
|
107 |
-
r"""
|
108 |
-
Base class for all models.
|
109 |
-
|
110 |
-
[`DiffusionPipeline`] takes care of storing all components (models, schedulers, processors) for diffusion pipelines
|
111 |
-
and handles methods for loading, downloading and saving models as well as a few methods common to all pipelines to:
|
112 |
-
|
113 |
-
- move all PyTorch modules to the device of your choice
|
114 |
-
- enabling/disabling the progress bar for the denoising iteration
|
115 |
-
|
116 |
-
Class attributes:
|
117 |
-
|
118 |
-
- **config_name** ([`str`]) -- name of the config file that will store the class and module names of all
|
119 |
-
components of the diffusion pipeline.
|
120 |
-
"""
|
121 |
-
config_name = "model_index.json"
|
122 |
-
|
123 |
-
def register_modules(self, **kwargs):
|
124 |
-
# import it here to avoid circular import
|
125 |
-
from diffusers import pipelines
|
126 |
-
|
127 |
-
for name, module in kwargs.items():
|
128 |
-
# retrieve library
|
129 |
-
if module is None:
|
130 |
-
register_dict = {name: (None, None)}
|
131 |
-
else:
|
132 |
-
library = module.__module__.split(".")[0]
|
133 |
-
|
134 |
-
# check if the module is a pipeline module
|
135 |
-
pipeline_dir = module.__module__.split(".")[-2] if len(module.__module__.split(".")) > 2 else None
|
136 |
-
path = module.__module__.split(".")
|
137 |
-
is_pipeline_module = pipeline_dir in path and hasattr(pipelines, pipeline_dir)
|
138 |
-
|
139 |
-
# if library is not in LOADABLE_CLASSES, then it is a custom module.
|
140 |
-
# Or if it's a pipeline module, then the module is inside the pipeline
|
141 |
-
# folder so we set the library to module name.
|
142 |
-
if library not in LOADABLE_CLASSES or is_pipeline_module:
|
143 |
-
library = pipeline_dir
|
144 |
-
|
145 |
-
# retrieve class_name
|
146 |
-
class_name = module.__class__.__name__
|
147 |
-
|
148 |
-
register_dict = {name: (library, class_name)}
|
149 |
-
|
150 |
-
# save model index config
|
151 |
-
self.register_to_config(**register_dict)
|
152 |
-
|
153 |
-
# set models
|
154 |
-
setattr(self, name, module)
|
155 |
-
|
156 |
-
def save_pretrained(self, save_directory: Union[str, os.PathLike]):
|
157 |
-
"""
|
158 |
-
Save all variables of the pipeline that can be saved and loaded as well as the pipelines configuration file to
|
159 |
-
a directory. A pipeline variable can be saved and loaded if its class implements both a save and loading
|
160 |
-
method. The pipeline can easily be re-loaded using the `[`~DiffusionPipeline.from_pretrained`]` class method.
|
161 |
-
|
162 |
-
Arguments:
|
163 |
-
save_directory (`str` or `os.PathLike`):
|
164 |
-
Directory to which to save. Will be created if it doesn't exist.
|
165 |
-
"""
|
166 |
-
self.save_config(save_directory)
|
167 |
-
|
168 |
-
model_index_dict = dict(self.config)
|
169 |
-
model_index_dict.pop("_class_name")
|
170 |
-
model_index_dict.pop("_diffusers_version")
|
171 |
-
model_index_dict.pop("_module", None)
|
172 |
-
|
173 |
-
for pipeline_component_name in model_index_dict.keys():
|
174 |
-
sub_model = getattr(self, pipeline_component_name)
|
175 |
-
if sub_model is None:
|
176 |
-
# edge case for saving a pipeline with safety_checker=None
|
177 |
-
continue
|
178 |
-
|
179 |
-
model_cls = sub_model.__class__
|
180 |
-
|
181 |
-
save_method_name = None
|
182 |
-
# search for the model's base class in LOADABLE_CLASSES
|
183 |
-
for library_name, library_classes in LOADABLE_CLASSES.items():
|
184 |
-
library = importlib.import_module(library_name)
|
185 |
-
for base_class, save_load_methods in library_classes.items():
|
186 |
-
class_candidate = getattr(library, base_class)
|
187 |
-
if issubclass(model_cls, class_candidate):
|
188 |
-
# if we found a suitable base class in LOADABLE_CLASSES then grab its save method
|
189 |
-
save_method_name = save_load_methods[0]
|
190 |
-
break
|
191 |
-
if save_method_name is not None:
|
192 |
-
break
|
193 |
-
|
194 |
-
save_method = getattr(sub_model, save_method_name)
|
195 |
-
save_method(os.path.join(save_directory, pipeline_component_name))
|
196 |
-
|
197 |
-
def to(self, torch_device: Optional[Union[str, torch.device]] = None):
|
198 |
-
if torch_device is None:
|
199 |
-
return self
|
200 |
-
|
201 |
-
module_names, _ = self.extract_init_dict(dict(self.config))
|
202 |
-
for name in module_names.keys():
|
203 |
-
module = getattr(self, name)
|
204 |
-
if isinstance(module, torch.nn.Module):
|
205 |
-
if module.dtype == torch.float16 and str(torch_device) in ["cpu", "mps"]:
|
206 |
-
logger.warning(
|
207 |
-
"Pipelines loaded with `torch_dtype=torch.float16` cannot run with `cpu` or `mps` device. It"
|
208 |
-
" is not recommended to move them to `cpu` or `mps` as running them will fail. Please make"
|
209 |
-
" sure to use a `cuda` device to run the pipeline in inference. due to the lack of support for"
|
210 |
-
" `float16` operations on those devices in PyTorch. Please remove the"
|
211 |
-
" `torch_dtype=torch.float16` argument, or use a `cuda` device to run inference."
|
212 |
-
)
|
213 |
-
module.to(torch_device)
|
214 |
-
return self
|
215 |
-
|
216 |
-
@property
|
217 |
-
def device(self) -> torch.device:
|
218 |
-
r"""
|
219 |
-
Returns:
|
220 |
-
`torch.device`: The torch device on which the pipeline is located.
|
221 |
-
"""
|
222 |
-
module_names, _ = self.extract_init_dict(dict(self.config))
|
223 |
-
for name in module_names.keys():
|
224 |
-
module = getattr(self, name)
|
225 |
-
if isinstance(module, torch.nn.Module):
|
226 |
-
# if module.device == torch.device("meta"):
|
227 |
-
# return torch.device("cpu")
|
228 |
-
return module.device
|
229 |
-
return torch.device("cpu")
|
230 |
-
|
231 |
-
@classmethod
|
232 |
-
def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
|
233 |
-
r"""
|
234 |
-
Instantiate a PyTorch diffusion pipeline from pre-trained pipeline weights.
|
235 |
-
|
236 |
-
The pipeline is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated).
|
237 |
-
|
238 |
-
The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
|
239 |
-
pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
|
240 |
-
task.
|
241 |
-
|
242 |
-
The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
|
243 |
-
weights are discarded.
|
244 |
-
|
245 |
-
Parameters:
|
246 |
-
pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
|
247 |
-
Can be either:
|
248 |
-
|
249 |
-
- A string, the *repo id* of a pretrained pipeline hosted inside a model repo on
|
250 |
-
https://huggingface.co/ Valid repo ids have to be located under a user or organization name, like
|
251 |
-
`CompVis/ldm-text2im-large-256`.
|
252 |
-
- A path to a *directory* containing pipeline weights saved using
|
253 |
-
[`~DiffusionPipeline.save_pretrained`], e.g., `./my_pipeline_directory/`.
|
254 |
-
torch_dtype (`str` or `torch.dtype`, *optional*):
|
255 |
-
Override the default `torch.dtype` and load the model under this dtype. If `"auto"` is passed the dtype
|
256 |
-
will be automatically derived from the model's weights.
|
257 |
-
custom_pipeline (`str`, *optional*):
|
258 |
-
|
259 |
-
<Tip warning={true}>
|
260 |
-
|
261 |
-
This is an experimental feature and is likely to change in the future.
|
262 |
-
|
263 |
-
</Tip>
|
264 |
-
|
265 |
-
Can be either:
|
266 |
-
|
267 |
-
- A string, the *repo id* of a custom pipeline hosted inside a model repo on
|
268 |
-
https://huggingface.co/. Valid repo ids have to be located under a user or organization name,
|
269 |
-
like `hf-internal-testing/diffusers-dummy-pipeline`.
|
270 |
-
|
271 |
-
<Tip>
|
272 |
-
|
273 |
-
It is required that the model repo has a file, called `pipeline.py` that defines the custom
|
274 |
-
pipeline.
|
275 |
-
|
276 |
-
</Tip>
|
277 |
-
|
278 |
-
- A string, the *file name* of a community pipeline hosted on GitHub under
|
279 |
-
https://github.com/huggingface/diffusers/tree/main/examples/community. Valid file names have to
|
280 |
-
match exactly the file name without `.py` located under the above link, *e.g.*
|
281 |
-
`clip_guided_stable_diffusion`.
|
282 |
-
|
283 |
-
<Tip>
|
284 |
-
|
285 |
-
Community pipelines are always loaded from the current `main` branch of GitHub.
|
286 |
-
|
287 |
-
</Tip>
|
288 |
-
|
289 |
-
- A path to a *directory* containing a custom pipeline, e.g., `./my_pipeline_directory/`.
|
290 |
-
|
291 |
-
<Tip>
|
292 |
-
|
293 |
-
It is required that the directory has a file, called `pipeline.py` that defines the custom
|
294 |
-
pipeline.
|
295 |
-
|
296 |
-
</Tip>
|
297 |
-
|
298 |
-
For more information on how to load and create custom pipelines, please have a look at [Loading and
|
299 |
-
Creating Custom
|
300 |
-
Pipelines](https://huggingface.co/docs/diffusers/main/en/using-diffusers/custom_pipelines)
|
301 |
-
|
302 |
-
torch_dtype (`str` or `torch.dtype`, *optional*):
|
303 |
-
force_download (`bool`, *optional*, defaults to `False`):
|
304 |
-
Whether or not to force the (re-)download of the model weights and configuration files, overriding the
|
305 |
-
cached versions if they exist.
|
306 |
-
resume_download (`bool`, *optional*, defaults to `False`):
|
307 |
-
Whether or not to delete incompletely received files. Will attempt to resume the download if such a
|
308 |
-
file exists.
|
309 |
-
proxies (`Dict[str, str]`, *optional*):
|
310 |
-
A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
|
311 |
-
'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
|
312 |
-
output_loading_info(`bool`, *optional*, defaults to `False`):
|
313 |
-
Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
|
314 |
-
local_files_only(`bool`, *optional*, defaults to `False`):
|
315 |
-
Whether or not to only look at local files (i.e., do not try to download the model).
|
316 |
-
use_auth_token (`str` or *bool*, *optional*):
|
317 |
-
The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
|
318 |
-
when running `huggingface-cli login` (stored in `~/.huggingface`).
|
319 |
-
revision (`str`, *optional*, defaults to `"main"`):
|
320 |
-
The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
|
321 |
-
git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
|
322 |
-
identifier allowed by git.
|
323 |
-
mirror (`str`, *optional*):
|
324 |
-
Mirror source to accelerate downloads in China. If you are from China and have an accessibility
|
325 |
-
problem, you can set this option to resolve it. Note that we do not guarantee the timeliness or safety.
|
326 |
-
Please refer to the mirror site for more information. specify the folder name here.
|
327 |
-
device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
|
328 |
-
A map that specifies where each submodule should go. It doesn't need to be refined to each
|
329 |
-
parameter/buffer name, once a given module name is inside, every submodule of it will be sent to the
|
330 |
-
same device.
|
331 |
-
|
332 |
-
To have Accelerate compute the most optimized `device_map` automatically, set `device_map="auto"`. For
|
333 |
-
more information about each option see [designing a device
|
334 |
-
map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
|
335 |
-
low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
|
336 |
-
Speed up model loading by not initializing the weights and only loading the pre-trained weights. This
|
337 |
-
also tries to not use more than 1x model size in CPU memory (including peak memory) while loading the
|
338 |
-
model. This is only supported when torch version >= 1.9.0. If you are using an older version of torch,
|
339 |
-
setting this argument to `True` will raise an error.
|
340 |
-
|
341 |
-
kwargs (remaining dictionary of keyword arguments, *optional*):
|
342 |
-
Can be used to overwrite load - and saveable variables - *i.e.* the pipeline components - of the
|
343 |
-
specific pipeline class. The overwritten components are then directly passed to the pipelines
|
344 |
-
`__init__` method. See example below for more information.
|
345 |
-
|
346 |
-
<Tip>
|
347 |
-
|
348 |
-
It is required to be logged in (`huggingface-cli login`) when you want to use private or [gated
|
349 |
-
models](https://huggingface.co/docs/hub/models-gated#gated-models), *e.g.* `"runwayml/stable-diffusion-v1-5"`
|
350 |
-
|
351 |
-
</Tip>
|
352 |
-
|
353 |
-
<Tip>
|
354 |
-
|
355 |
-
Activate the special ["offline-mode"](https://huggingface.co/diffusers/installation.html#offline-mode) to use
|
356 |
-
this method in a firewalled environment.
|
357 |
-
|
358 |
-
</Tip>
|
359 |
-
|
360 |
-
Examples:
|
361 |
-
|
362 |
-
```py
|
363 |
-
>>> from diffusers import DiffusionPipeline
|
364 |
-
|
365 |
-
>>> # Download pipeline from huggingface.co and cache.
|
366 |
-
>>> pipeline = DiffusionPipeline.from_pretrained("CompVis/ldm-text2im-large-256")
|
367 |
-
|
368 |
-
>>> # Download pipeline that requires an authorization token
|
369 |
-
>>> # For more information on access tokens, please refer to this section
|
370 |
-
>>> # of the documentation](https://huggingface.co/docs/hub/security-tokens)
|
371 |
-
>>> pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
|
372 |
-
|
373 |
-
>>> # Download pipeline, but overwrite scheduler
|
374 |
-
>>> from diffusers import LMSDiscreteScheduler
|
375 |
-
|
376 |
-
>>> scheduler = LMSDiscreteScheduler.from_config("runwayml/stable-diffusion-v1-5", subfolder="scheduler")
|
377 |
-
>>> pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", scheduler=scheduler)
|
378 |
-
```
|
379 |
-
"""
|
380 |
-
cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
|
381 |
-
resume_download = kwargs.pop("resume_download", False)
|
382 |
-
force_download = kwargs.pop("force_download", False)
|
383 |
-
proxies = kwargs.pop("proxies", None)
|
384 |
-
local_files_only = kwargs.pop("local_files_only", False)
|
385 |
-
use_auth_token = kwargs.pop("use_auth_token", None)
|
386 |
-
revision = kwargs.pop("revision", None)
|
387 |
-
torch_dtype = kwargs.pop("torch_dtype", None)
|
388 |
-
custom_pipeline = kwargs.pop("custom_pipeline", None)
|
389 |
-
provider = kwargs.pop("provider", None)
|
390 |
-
sess_options = kwargs.pop("sess_options", None)
|
391 |
-
device_map = kwargs.pop("device_map", None)
|
392 |
-
low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
|
393 |
-
|
394 |
-
if device_map is not None and not is_torch_version(">=", "1.9.0"):
|
395 |
-
raise NotImplementedError(
|
396 |
-
"Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set"
|
397 |
-
" `device_map=None`."
|
398 |
-
)
|
399 |
-
|
400 |
-
if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
|
401 |
-
raise NotImplementedError(
|
402 |
-
"Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
|
403 |
-
" `low_cpu_mem_usage=False`."
|
404 |
-
)
|
405 |
-
|
406 |
-
if low_cpu_mem_usage is False and device_map is not None:
|
407 |
-
raise ValueError(
|
408 |
-
f"You cannot set `low_cpu_mem_usage` to False while using device_map={device_map} for loading and"
|
409 |
-
" dispatching. Please make sure to set `low_cpu_mem_usage=True`."
|
410 |
-
)
|
411 |
-
|
412 |
-
# 1. Download the checkpoints and configs
|
413 |
-
# use snapshot download here to get it working from from_pretrained
|
414 |
-
if not os.path.isdir(pretrained_model_name_or_path):
|
415 |
-
config_dict = cls.get_config_dict(
|
416 |
-
pretrained_model_name_or_path,
|
417 |
-
cache_dir=cache_dir,
|
418 |
-
resume_download=resume_download,
|
419 |
-
force_download=force_download,
|
420 |
-
proxies=proxies,
|
421 |
-
local_files_only=local_files_only,
|
422 |
-
use_auth_token=use_auth_token,
|
423 |
-
revision=revision,
|
424 |
-
)
|
425 |
-
# make sure we only download sub-folders and `diffusers` filenames
|
426 |
-
folder_names = [k for k in config_dict.keys() if not k.startswith("_")]
|
427 |
-
allow_patterns = [os.path.join(k, "*") for k in folder_names]
|
428 |
-
allow_patterns += [WEIGHTS_NAME, SCHEDULER_CONFIG_NAME, CONFIG_NAME, ONNX_WEIGHTS_NAME, cls.config_name]
|
429 |
-
|
430 |
-
# make sure we don't download flax weights
|
431 |
-
ignore_patterns = "*.msgpack"
|
432 |
-
|
433 |
-
if custom_pipeline is not None:
|
434 |
-
allow_patterns += [CUSTOM_PIPELINE_FILE_NAME]
|
435 |
-
|
436 |
-
if cls != DiffusionPipeline:
|
437 |
-
requested_pipeline_class = cls.__name__
|
438 |
-
else:
|
439 |
-
requested_pipeline_class = config_dict.get("_class_name", cls.__name__)
|
440 |
-
user_agent = {"pipeline_class": requested_pipeline_class}
|
441 |
-
if custom_pipeline is not None:
|
442 |
-
user_agent["custom_pipeline"] = custom_pipeline
|
443 |
-
user_agent = http_user_agent(user_agent)
|
444 |
-
|
445 |
-
# download all allow_patterns
|
446 |
-
cached_folder = snapshot_download(
|
447 |
-
pretrained_model_name_or_path,
|
448 |
-
cache_dir=cache_dir,
|
449 |
-
resume_download=resume_download,
|
450 |
-
proxies=proxies,
|
451 |
-
local_files_only=local_files_only,
|
452 |
-
use_auth_token=use_auth_token,
|
453 |
-
revision=revision,
|
454 |
-
allow_patterns=allow_patterns,
|
455 |
-
ignore_patterns=ignore_patterns,
|
456 |
-
user_agent=user_agent,
|
457 |
-
)
|
458 |
-
else:
|
459 |
-
cached_folder = pretrained_model_name_or_path
|
460 |
-
|
461 |
-
config_dict = cls.get_config_dict(cached_folder)
|
462 |
-
|
463 |
-
# 2. Load the pipeline class, if using custom module then load it from the hub
|
464 |
-
# if we load from explicit class, let's use it
|
465 |
-
if custom_pipeline is not None:
|
466 |
-
pipeline_class = get_class_from_dynamic_module(
|
467 |
-
custom_pipeline, module_file=CUSTOM_PIPELINE_FILE_NAME, cache_dir=custom_pipeline
|
468 |
-
)
|
469 |
-
elif cls != DiffusionPipeline:
|
470 |
-
pipeline_class = cls
|
471 |
-
else:
|
472 |
-
diffusers_module = importlib.import_module(cls.__module__.split(".")[0])
|
473 |
-
pipeline_class = getattr(diffusers_module, config_dict["_class_name"])
|
474 |
-
|
475 |
-
# To be removed in 1.0.0
|
476 |
-
if pipeline_class.__name__ == "StableDiffusionInpaintPipeline" and version.parse(
|
477 |
-
version.parse(config_dict["_diffusers_version"]).base_version
|
478 |
-
) <= version.parse("0.5.1"):
|
479 |
-
from diffusers import StableDiffusionInpaintPipeline, StableDiffusionInpaintPipelineLegacy
|
480 |
-
|
481 |
-
pipeline_class = StableDiffusionInpaintPipelineLegacy
|
482 |
-
|
483 |
-
deprecation_message = (
|
484 |
-
"You are using a legacy checkpoint for inpainting with Stable Diffusion, therefore we are loading the"
|
485 |
-
f" {StableDiffusionInpaintPipelineLegacy} class instead of {StableDiffusionInpaintPipeline}. For"
|
486 |
-
" better inpainting results, we strongly suggest using Stable Diffusion's official inpainting"
|
487 |
-
" checkpoint: https://huggingface.co/runwayml/stable-diffusion-inpainting instead or adapting your"
|
488 |
-
f" checkpoint {pretrained_model_name_or_path} to the format of"
|
489 |
-
" https://huggingface.co/runwayml/stable-diffusion-inpainting. Note that we do not actively maintain"
|
490 |
-
" the {StableDiffusionInpaintPipelineLegacy} class and will likely remove it in version 1.0.0."
|
491 |
-
)
|
492 |
-
deprecate("StableDiffusionInpaintPipelineLegacy", "1.0.0", deprecation_message, standard_warn=False)
|
493 |
-
|
494 |
-
# some modules can be passed directly to the init
|
495 |
-
# in this case they are already instantiated in `kwargs`
|
496 |
-
# extract them here
|
497 |
-
expected_modules = set(inspect.signature(pipeline_class.__init__).parameters.keys()) - set(["self"])
|
498 |
-
passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs}
|
499 |
-
|
500 |
-
init_dict, unused_kwargs = pipeline_class.extract_init_dict(config_dict, **kwargs)
|
501 |
-
|
502 |
-
if len(unused_kwargs) > 0:
|
503 |
-
logger.warning(f"Keyword arguments {unused_kwargs} not recognized.")
|
504 |
-
|
505 |
-
init_kwargs = {}
|
506 |
-
|
507 |
-
# import it here to avoid circular import
|
508 |
-
from diffusers import pipelines
|
509 |
-
|
510 |
-
# 3. Load each module in the pipeline
|
511 |
-
for name, (library_name, class_name) in init_dict.items():
|
512 |
-
if class_name is None:
|
513 |
-
# edge case for when the pipeline was saved with safety_checker=None
|
514 |
-
init_kwargs[name] = None
|
515 |
-
continue
|
516 |
-
|
517 |
-
# 3.1 - now that JAX/Flax is an official framework of the library, we might load from Flax names
|
518 |
-
if class_name.startswith("Flax"):
|
519 |
-
class_name = class_name[4:]
|
520 |
-
|
521 |
-
is_pipeline_module = hasattr(pipelines, library_name)
|
522 |
-
loaded_sub_model = None
|
523 |
-
sub_model_should_be_defined = True
|
524 |
-
|
525 |
-
# if the model is in a pipeline module, then we load it from the pipeline
|
526 |
-
if name in passed_class_obj:
|
527 |
-
# 1. check that passed_class_obj has correct parent class
|
528 |
-
if not is_pipeline_module:
|
529 |
-
library = importlib.import_module(library_name)
|
530 |
-
class_obj = getattr(library, class_name)
|
531 |
-
importable_classes = LOADABLE_CLASSES[library_name]
|
532 |
-
class_candidates = {c: getattr(library, c) for c in importable_classes.keys()}
|
533 |
-
|
534 |
-
expected_class_obj = None
|
535 |
-
for class_name, class_candidate in class_candidates.items():
|
536 |
-
if issubclass(class_obj, class_candidate):
|
537 |
-
expected_class_obj = class_candidate
|
538 |
-
|
539 |
-
if not issubclass(passed_class_obj[name].__class__, expected_class_obj):
|
540 |
-
raise ValueError(
|
541 |
-
f"{passed_class_obj[name]} is of type: {type(passed_class_obj[name])}, but should be"
|
542 |
-
f" {expected_class_obj}"
|
543 |
-
)
|
544 |
-
elif passed_class_obj[name] is None:
|
545 |
-
logger.warn(
|
546 |
-
f"You have passed `None` for {name} to disable its functionality in {pipeline_class}. Note"
|
547 |
-
f" that this might lead to problems when using {pipeline_class} and is not recommended."
|
548 |
-
)
|
549 |
-
sub_model_should_be_defined = False
|
550 |
-
else:
|
551 |
-
logger.warn(
|
552 |
-
f"You have passed a non-standard module {passed_class_obj[name]}. We cannot verify whether it"
|
553 |
-
" has the correct type"
|
554 |
-
)
|
555 |
-
|
556 |
-
# set passed class object
|
557 |
-
loaded_sub_model = passed_class_obj[name]
|
558 |
-
elif is_pipeline_module:
|
559 |
-
pipeline_module = getattr(pipelines, library_name)
|
560 |
-
class_obj = getattr(pipeline_module, class_name)
|
561 |
-
importable_classes = ALL_IMPORTABLE_CLASSES
|
562 |
-
class_candidates = {c: class_obj for c in importable_classes.keys()}
|
563 |
-
else:
|
564 |
-
# else we just import it from the library.
|
565 |
-
library = importlib.import_module(library_name)
|
566 |
-
class_obj = getattr(library, class_name)
|
567 |
-
importable_classes = LOADABLE_CLASSES[library_name]
|
568 |
-
class_candidates = {c: getattr(library, c) for c in importable_classes.keys()}
|
569 |
-
|
570 |
-
if loaded_sub_model is None and sub_model_should_be_defined:
|
571 |
-
load_method_name = None
|
572 |
-
for class_name, class_candidate in class_candidates.items():
|
573 |
-
if issubclass(class_obj, class_candidate):
|
574 |
-
load_method_name = importable_classes[class_name][1]
|
575 |
-
|
576 |
-
if load_method_name is None:
|
577 |
-
none_module = class_obj.__module__
|
578 |
-
if none_module.startswith(DUMMY_MODULES_FOLDER) and "dummy" in none_module:
|
579 |
-
# call class_obj for nice error message of missing requirements
|
580 |
-
class_obj()
|
581 |
-
|
582 |
-
raise ValueError(
|
583 |
-
f"The component {class_obj} of {pipeline_class} cannot be loaded as it does not seem to have"
|
584 |
-
f" any of the loading methods defined in {ALL_IMPORTABLE_CLASSES}."
|
585 |
-
)
|
586 |
-
|
587 |
-
load_method = getattr(class_obj, load_method_name)
|
588 |
-
loading_kwargs = {}
|
589 |
-
|
590 |
-
if issubclass(class_obj, torch.nn.Module):
|
591 |
-
loading_kwargs["torch_dtype"] = torch_dtype
|
592 |
-
if issubclass(class_obj, diffusers.OnnxRuntimeModel):
|
593 |
-
loading_kwargs["provider"] = provider
|
594 |
-
loading_kwargs["sess_options"] = sess_options
|
595 |
-
|
596 |
-
is_diffusers_model = issubclass(class_obj, diffusers.ModelMixin)
|
597 |
-
is_transformers_model = (
|
598 |
-
is_transformers_available()
|
599 |
-
and issubclass(class_obj, PreTrainedModel)
|
600 |
-
and version.parse(version.parse(transformers.__version__).base_version) >= version.parse("4.20.0")
|
601 |
-
)
|
602 |
-
|
603 |
-
# When loading a transformers model, if the device_map is None, the weights will be initialized as opposed to diffusers.
|
604 |
-
# To make default loading faster we set the `low_cpu_mem_usage=low_cpu_mem_usage` flag which is `True` by default.
|
605 |
-
# This makes sure that the weights won't be initialized which significantly speeds up loading.
|
606 |
-
if is_diffusers_model or is_transformers_model:
|
607 |
-
loading_kwargs["device_map"] = device_map
|
608 |
-
loading_kwargs["low_cpu_mem_usage"] = low_cpu_mem_usage
|
609 |
-
|
610 |
-
# check if the module is in a subdirectory
|
611 |
-
if os.path.isdir(os.path.join(cached_folder, name)):
|
612 |
-
loaded_sub_model = load_method(os.path.join(cached_folder, name), **loading_kwargs)
|
613 |
-
else:
|
614 |
-
# else load from the root directory
|
615 |
-
loaded_sub_model = load_method(cached_folder, **loading_kwargs)
|
616 |
-
|
617 |
-
init_kwargs[name] = loaded_sub_model # UNet(...), # DiffusionSchedule(...)
|
618 |
-
|
619 |
-
# 4. Potentially add passed objects if expected
|
620 |
-
missing_modules = set(expected_modules) - set(init_kwargs.keys())
|
621 |
-
if len(missing_modules) > 0 and missing_modules <= set(passed_class_obj.keys()):
|
622 |
-
for module in missing_modules:
|
623 |
-
init_kwargs[module] = passed_class_obj[module]
|
624 |
-
elif len(missing_modules) > 0:
|
625 |
-
passed_modules = set(list(init_kwargs.keys()) + list(passed_class_obj.keys()))
|
626 |
-
raise ValueError(
|
627 |
-
f"Pipeline {pipeline_class} expected {expected_modules}, but only {passed_modules} were passed."
|
628 |
-
)
|
629 |
-
|
630 |
-
# 5. Instantiate the pipeline
|
631 |
-
model = pipeline_class(**init_kwargs)
|
632 |
-
return model
|
633 |
-
|
634 |
-
@property
|
635 |
-
def components(self) -> Dict[str, Any]:
|
636 |
-
r"""
|
637 |
-
|
638 |
-
The `self.components` property can be useful to run different pipelines with the same weights and
|
639 |
-
configurations to not have to re-allocate memory.
|
640 |
-
|
641 |
-
Examples:
|
642 |
-
|
643 |
-
```py
|
644 |
-
>>> from diffusers import (
|
645 |
-
... StableDiffusionPipeline,
|
646 |
-
... StableDiffusionImg2ImgPipeline,
|
647 |
-
... StableDiffusionInpaintPipeline,
|
648 |
-
... )
|
649 |
-
|
650 |
-
>>> img2text = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
|
651 |
-
>>> img2img = StableDiffusionImg2ImgPipeline(**img2text.components)
|
652 |
-
>>> inpaint = StableDiffusionInpaintPipeline(**img2text.components)
|
653 |
-
```
|
654 |
-
|
655 |
-
Returns:
|
656 |
-
A dictionaly containing all the modules needed to initialize the pipeline.
|
657 |
-
"""
|
658 |
-
components = {k: getattr(self, k) for k in self.config.keys() if not k.startswith("_")}
|
659 |
-
expected_modules = set(inspect.signature(self.__init__).parameters.keys()) - set(["self"])
|
660 |
-
|
661 |
-
if set(components.keys()) != expected_modules:
|
662 |
-
raise ValueError(
|
663 |
-
f"{self} has been incorrectly initialized or {self.__class__} is incorrectly implemented. Expected"
|
664 |
-
f" {expected_modules} to be defined, but {components} are defined."
|
665 |
-
)
|
666 |
-
|
667 |
-
return components
|
668 |
-
|
669 |
-
@staticmethod
|
670 |
-
def numpy_to_pil(images):
|
671 |
-
"""
|
672 |
-
Convert a numpy image or a batch of images to a PIL image.
|
673 |
-
"""
|
674 |
-
if images.ndim == 3:
|
675 |
-
images = images[None, ...]
|
676 |
-
images = (images * 255).round().astype("uint8")
|
677 |
-
if images.shape[-1] == 1:
|
678 |
-
# special case for grayscale (single channel) images
|
679 |
-
pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images]
|
680 |
-
else:
|
681 |
-
pil_images = [Image.fromarray(image) for image in images]
|
682 |
-
|
683 |
-
return pil_images
|
684 |
-
|
685 |
-
def progress_bar(self, iterable):
|
686 |
-
if not hasattr(self, "_progress_bar_config"):
|
687 |
-
self._progress_bar_config = {}
|
688 |
-
elif not isinstance(self._progress_bar_config, dict):
|
689 |
-
raise ValueError(
|
690 |
-
f"`self._progress_bar_config` should be of type `dict`, but is {type(self._progress_bar_config)}."
|
691 |
-
)
|
692 |
-
|
693 |
-
return tqdm(iterable, **self._progress_bar_config)
|
694 |
-
|
695 |
-
def set_progress_bar_config(self, **kwargs):
|
696 |
-
self._progress_bar_config = kwargs
|
697 |
-
|
698 |
-
|
699 |
-
class LDMZhTextToImagePipeline(DiffusionPipeline):
|
700 |
-
|
701 |
-
def __init__(
|
702 |
-
self,
|
703 |
-
vqvae,
|
704 |
-
bert,
|
705 |
-
tokenizer,
|
706 |
-
unet,
|
707 |
-
scheduler,
|
708 |
-
sr,
|
709 |
-
):
|
710 |
-
super().__init__()
|
711 |
-
self.register_modules(vqvae=vqvae, bert=bert, tokenizer=tokenizer, unet=unet, scheduler=scheduler, sr=sr)
|
712 |
-
|
713 |
-
@torch.no_grad()
|
714 |
-
def __call__(
|
715 |
-
self,
|
716 |
-
prompt: Union[str, List[str]],
|
717 |
-
height: Optional[int] = 256,
|
718 |
-
width: Optional[int] = 256,
|
719 |
-
num_inference_steps: Optional[int] = 50,
|
720 |
-
guidance_scale: Optional[float] = 5.0,
|
721 |
-
eta: Optional[float] = 0.0,
|
722 |
-
generator: Optional[torch.Generator] = None,
|
723 |
-
output_type: Optional[str] = "pil",
|
724 |
-
return_dict: bool = True,
|
725 |
-
use_sr: bool = False,
|
726 |
-
**kwargs,
|
727 |
-
):
|
728 |
-
r"""
|
729 |
-
Args:
|
730 |
-
prompt (`str` or `List[str]`):
|
731 |
-
The prompt or prompts to guide the image generation.
|
732 |
-
height (`int`, *optional*, defaults to 256):
|
733 |
-
The height in pixels of the generated image.
|
734 |
-
width (`int`, *optional*, defaults to 256):
|
735 |
-
The width in pixels of the generated image.
|
736 |
-
num_inference_steps (`int`, *optional*, defaults to 50):
|
737 |
-
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
738 |
-
expense of slower inference.
|
739 |
-
guidance_scale (`float`, *optional*, defaults to 1.0):
|
740 |
-
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
|
741 |
-
`guidance_scale` is defined as `w` of equation 2. of [Imagen
|
742 |
-
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
|
743 |
-
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt` at
|
744 |
-
the, usually at the expense of lower image quality.
|
745 |
-
generator (`torch.Generator`, *optional*):
|
746 |
-
A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
|
747 |
-
deterministic.
|
748 |
-
output_type (`str`, *optional*, defaults to `"pil"`):
|
749 |
-
The output format of the generate image. Choose between
|
750 |
-
[PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
|
751 |
-
return_dict (`bool`, *optional*):
|
752 |
-
Whether or not to return a [`~pipeline_utils.ImagePipelineOutput`] instead of a plain tuple.
|
753 |
-
|
754 |
-
Returns:
|
755 |
-
[`~pipeline_utils.ImagePipelineOutput`] or `tuple`: [`~pipelines.utils.ImagePipelineOutput`] if
|
756 |
-
`return_dict` is True, otherwise a `tuple. When returning a tuple, the first element is a list with the
|
757 |
-
generated images.
|
758 |
-
"""
|
759 |
-
|
760 |
-
if isinstance(prompt, str):
|
761 |
-
batch_size = 1
|
762 |
-
elif isinstance(prompt, list):
|
763 |
-
batch_size = len(prompt)
|
764 |
-
else:
|
765 |
-
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
|
766 |
-
|
767 |
-
if height % 8 != 0 or width % 8 != 0:
|
768 |
-
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
|
769 |
-
|
770 |
-
# get unconditional embeddings for classifier free guidance
|
771 |
-
if guidance_scale != 1.0:
|
772 |
-
uncond_input = self.tokenizer([""] * batch_size, padding="max_length", max_length=32, return_tensors="pt")
|
773 |
-
uncond_embeddings = self.bert(uncond_input.input_ids.to(self.device))
|
774 |
-
|
775 |
-
# get prompt text embeddings
|
776 |
-
text_input = self.tokenizer(prompt, padding="max_length", max_length=32, return_tensors="pt")
|
777 |
-
text_embeddings = self.bert(text_input.input_ids.to(self.device))
|
778 |
-
|
779 |
-
latents = torch.randn(
|
780 |
-
(batch_size, self.unet.in_channels, height // 8, width // 8),
|
781 |
-
generator=generator,
|
782 |
-
)
|
783 |
-
latents = latents.to(self.device)
|
784 |
-
|
785 |
-
self.scheduler.set_timesteps(num_inference_steps)
|
786 |
-
|
787 |
-
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
|
788 |
-
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
789 |
-
|
790 |
-
extra_kwargs = {}
|
791 |
-
if accepts_eta:
|
792 |
-
extra_kwargs["eta"] = eta
|
793 |
-
|
794 |
-
for t in self.progress_bar(self.scheduler.timesteps):
|
795 |
-
if guidance_scale == 1.0:
|
796 |
-
# guidance_scale of 1 means no guidance
|
797 |
-
latents_input = latents
|
798 |
-
context = text_embeddings
|
799 |
-
else:
|
800 |
-
# For classifier free guidance, we need to do two forward passes.
|
801 |
-
# Here we concatenate the unconditional and text embeddings into a single batch
|
802 |
-
# to avoid doing two forward passes
|
803 |
-
latents_input = torch.cat([latents] * 2)
|
804 |
-
context = torch.cat([uncond_embeddings, text_embeddings])
|
805 |
-
|
806 |
-
# predict the noise residual
|
807 |
-
noise_pred = self.unet(latents_input, t, encoder_hidden_states=context).sample
|
808 |
-
# perform guidance
|
809 |
-
if guidance_scale != 1.0:
|
810 |
-
noise_pred_uncond, noise_prediction_text = noise_pred.chunk(2)
|
811 |
-
noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond)
|
812 |
-
|
813 |
-
# compute the previous noisy sample x_t -> x_t-1
|
814 |
-
latents = self.scheduler.step(noise_pred, t, latents, **extra_kwargs).prev_sample
|
815 |
-
|
816 |
-
# scale and decode the image latents with vae
|
817 |
-
latents = 1 / 0.18215 * latents
|
818 |
-
image = self.vqvae.decode(latents).sample
|
819 |
-
|
820 |
-
image = (image / 2 + 0.5).clamp(0, 1)
|
821 |
-
if use_sr:
|
822 |
-
image = self.sr(image)
|
823 |
-
image = image.cpu().permute(0, 2, 3, 1).numpy()
|
824 |
-
if output_type == "pil":
|
825 |
-
image = self.numpy_to_pil(image)
|
826 |
-
|
827 |
-
if not return_dict:
|
828 |
-
return (image,)
|
829 |
-
|
830 |
-
return ImagePipelineOutput(images=image)
|
831 |
-
|
832 |
-
|
833 |
-
class QuickGELU(nn.Module):
|
834 |
-
def forward(self, x: torch.Tensor):
|
835 |
-
return x * torch.sigmoid(1.702 * x)
|
836 |
-
|
837 |
-
|
838 |
-
class ResidualAttentionBlock(nn.Module):
|
839 |
-
def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):
|
840 |
-
super().__init__()
|
841 |
-
self.attn = nn.MultiheadAttention(d_model, n_head)
|
842 |
-
self.ln_1 = nn.LayerNorm(d_model,eps=1e-07)
|
843 |
-
self.mlp = nn.Sequential(OrderedDict([
|
844 |
-
("c_fc", nn.Linear(d_model, d_model * 4)),
|
845 |
-
("gelu", QuickGELU()),
|
846 |
-
("c_proj", nn.Linear(d_model * 4, d_model))
|
847 |
-
]))
|
848 |
-
self.ln_2 = nn.LayerNorm(d_model,eps=1e-07)
|
849 |
-
self.attn_mask = attn_mask
|
850 |
-
def attention(self, x: torch.Tensor):
|
851 |
-
self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None
|
852 |
-
return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
|
853 |
-
def forward(self, x: torch.Tensor):
|
854 |
-
x = x + self.attention(self.ln_1(x))
|
855 |
-
x = x + self.mlp(self.ln_2(x))
|
856 |
-
return x
|
857 |
-
|
858 |
-
|
859 |
-
class Transformer(nn.Module):
|
860 |
-
def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):
|
861 |
-
super().__init__()
|
862 |
-
self.width = width
|
863 |
-
self.layers = layers
|
864 |
-
self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])
|
865 |
-
|
866 |
-
def forward(self, x: torch.Tensor):
|
867 |
-
return self.resblocks(x)
|
868 |
-
|
869 |
-
|
870 |
-
class TextTransformer(nn.Module):
|
871 |
-
def __init__(self,
|
872 |
-
context_length = 32,
|
873 |
-
vocab_size = 21128,
|
874 |
-
output_dim = 768,
|
875 |
-
width = 768,
|
876 |
-
layers = 12,
|
877 |
-
heads = 12,
|
878 |
-
return_full_embed = False):
|
879 |
-
super(TextTransformer, self).__init__()
|
880 |
-
self.width = width
|
881 |
-
self.layers = layers
|
882 |
-
self.vocab_size = vocab_size
|
883 |
-
self.return_full_embed = return_full_embed
|
884 |
-
|
885 |
-
self.transformer = Transformer(width, layers, heads, self.build_attntion_mask(context_length))
|
886 |
-
self.text_projection = torch.nn.Parameter(
|
887 |
-
torch.tensor(np.random.normal(0, self.width ** -0.5, size=(self.width, output_dim)).astype(np.float32)))
|
888 |
-
self.ln_final = nn.LayerNorm(width,eps=1e-07)
|
889 |
-
|
890 |
-
# https://discuss.pytorch.org/t/implementing-truncated-normal-initializer/4778/27
|
891 |
-
# https://github.com/pytorch/pytorch/blob/a40812de534b42fcf0eb57a5cecbfdc7a70100cf/torch/nn/init.py#L22
|
892 |
-
self.embedding_table = nn.Parameter(nn.init.trunc_normal_(torch.empty(vocab_size, width),std=0.02))
|
893 |
-
# self.embedding_table = nn.Embedding.from_pretrained(nn.init.trunc_normal_(torch.empty(vocab_size, width),std=0.02))
|
894 |
-
self.positional_embedding = nn.Parameter(nn.init.trunc_normal_(torch.empty(context_length, width),std=0.01))
|
895 |
-
# self.positional_embedding = nn.Embedding.from_pretrained(nn.init.trunc_normal_(torch.empty(context_length, width),std=0.01))
|
896 |
-
|
897 |
-
self.index_select=torch.index_select
|
898 |
-
self.reshape=torch.reshape
|
899 |
-
|
900 |
-
@staticmethod
|
901 |
-
def build_attntion_mask(context_length):
|
902 |
-
mask = np.triu(np.full((context_length, context_length), -np.inf).astype(np.float32), 1)
|
903 |
-
mask = torch.tensor(mask)
|
904 |
-
return mask
|
905 |
-
|
906 |
-
def forward(self, x: torch.Tensor):
|
907 |
-
|
908 |
-
tail_token=(x==102).nonzero(as_tuple=True)
|
909 |
-
bsz, ctx_len = x.shape
|
910 |
-
flatten_id = x.flatten()
|
911 |
-
index_select_result = self.index_select(self.embedding_table,0, flatten_id)
|
912 |
-
x = self.reshape(index_select_result, (bsz, ctx_len, -1))
|
913 |
-
x = x + self.positional_embedding
|
914 |
-
x = x.permute(1, 0, 2) # NLD -> LND
|
915 |
-
x = self.transformer(x)
|
916 |
-
x = x.permute(1, 0, 2) # LND -> NLD
|
917 |
-
x = self.ln_final(x)
|
918 |
-
x=x[tail_token]
|
919 |
-
x = x @ self.text_projection
|
920 |
-
return x
|
921 |
-
|
922 |
-
|
923 |
-
class WukongClipTextEncoder(ModelMixin, ConfigMixin):
|
924 |
-
|
925 |
-
@register_to_config
|
926 |
-
def __init__(
|
927 |
-
self,
|
928 |
-
):
|
929 |
-
super().__init__()
|
930 |
-
self.model = TextTransformer()
|
931 |
-
|
932 |
-
def forward(
|
933 |
-
self,
|
934 |
-
tokens
|
935 |
-
):
|
936 |
-
z = self.model(tokens)
|
937 |
-
z = z / torch.linalg.norm(z, dim=-1, keepdim=True)
|
938 |
-
if z.ndim==2:
|
939 |
-
z = z.view((z.shape[0], 1, z.shape[1]))
|
940 |
-
return z
|
941 |
-
|
942 |
-
|
943 |
-
def make_layer(block, n_layers):
|
944 |
-
layers = []
|
945 |
-
for _ in range(n_layers):
|
946 |
-
layers.append(block())
|
947 |
-
return nn.Sequential(*layers)
|
948 |
-
|
949 |
-
|
950 |
-
class ResidualDenseBlock_5C(nn.Module):
|
951 |
-
def __init__(self, nf=64, gc=32, bias=True):
|
952 |
-
super(ResidualDenseBlock_5C, self).__init__()
|
953 |
-
# gc: growth channel, i.e. intermediate channels
|
954 |
-
self.conv1 = nn.Conv2d(nf, gc, 3, 1, 1, bias=bias)
|
955 |
-
self.conv2 = nn.Conv2d(nf + gc, gc, 3, 1, 1, bias=bias)
|
956 |
-
self.conv3 = nn.Conv2d(nf + 2 * gc, gc, 3, 1, 1, bias=bias)
|
957 |
-
self.conv4 = nn.Conv2d(nf + 3 * gc, gc, 3, 1, 1, bias=bias)
|
958 |
-
self.conv5 = nn.Conv2d(nf + 4 * gc, nf, 3, 1, 1, bias=bias)
|
959 |
-
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
|
960 |
-
|
961 |
-
# initialization
|
962 |
-
# mutil.initialize_weights([self.conv1, self.conv2, self.conv3, self.conv4, self.conv5], 0.1)
|
963 |
-
|
964 |
-
def forward(self, x):
|
965 |
-
x1 = self.lrelu(self.conv1(x))
|
966 |
-
x2 = self.lrelu(self.conv2(torch.cat((x, x1), 1)))
|
967 |
-
x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), 1)))
|
968 |
-
x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), 1)))
|
969 |
-
x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1))
|
970 |
-
return x5 * 0.2 + x
|
971 |
-
|
972 |
-
|
973 |
-
class RRDB(nn.Module):
|
974 |
-
'''Residual in Residual Dense Block'''
|
975 |
-
|
976 |
-
def __init__(self, nf, gc=32):
|
977 |
-
super(RRDB, self).__init__()
|
978 |
-
self.RDB1 = ResidualDenseBlock_5C(nf, gc)
|
979 |
-
self.RDB2 = ResidualDenseBlock_5C(nf, gc)
|
980 |
-
self.RDB3 = ResidualDenseBlock_5C(nf, gc)
|
981 |
-
|
982 |
-
def forward(self, x):
|
983 |
-
out = self.RDB1(x)
|
984 |
-
out = self.RDB2(out)
|
985 |
-
out = self.RDB3(out)
|
986 |
-
return out * 0.2 + x
|
987 |
-
|
988 |
-
|
989 |
-
class RRDBNet(nn.Module):
|
990 |
-
def __init__(self, in_nc, out_nc, nf, nb, gc=32):
|
991 |
-
super(RRDBNet, self).__init__()
|
992 |
-
RRDB_block_f = functools.partial(RRDB, nf=nf, gc=gc)
|
993 |
-
|
994 |
-
self.conv_first = nn.Conv2d(in_nc, nf, 3, 1, 1, bias=True)
|
995 |
-
self.RRDB_trunk = make_layer(RRDB_block_f, nb)
|
996 |
-
self.trunk_conv = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
|
997 |
-
#### upsampling
|
998 |
-
self.upconv1 = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
|
999 |
-
self.upconv2 = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
|
1000 |
-
self.HRconv = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
|
1001 |
-
self.conv_last = nn.Conv2d(nf, out_nc, 3, 1, 1, bias=True)
|
1002 |
-
|
1003 |
-
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
|
1004 |
-
|
1005 |
-
def forward(self, x):
|
1006 |
-
fea = self.conv_first(x)
|
1007 |
-
trunk = self.trunk_conv(self.RRDB_trunk(fea))
|
1008 |
-
fea = fea + trunk
|
1009 |
-
|
1010 |
-
fea = self.lrelu(self.upconv1(torch.nn.functional.interpolate(fea, scale_factor=2, mode='nearest')))
|
1011 |
-
fea = self.lrelu(self.upconv2(torch.nn.functional.interpolate(fea, scale_factor=2, mode='nearest')))
|
1012 |
-
out = self.conv_last(self.lrelu(self.HRconv(fea)))
|
1013 |
-
|
1014 |
-
return out
|
1015 |
-
|
1016 |
-
|
1017 |
-
class ESRGAN(ModelMixin, ConfigMixin):
|
1018 |
-
|
1019 |
-
@register_to_config
|
1020 |
-
def __init__(
|
1021 |
-
self,
|
1022 |
-
):
|
1023 |
-
super().__init__()
|
1024 |
-
self.model = RRDBNet(3, 3, 64, 23, gc=32)
|
1025 |
-
|
1026 |
-
def forward(
|
1027 |
-
self,
|
1028 |
-
img_LR
|
1029 |
-
):
|
1030 |
-
img_LR = img_LR[:,[2,1,0],:,:]
|
1031 |
-
img_LR = img_LR.to(self.device)
|
1032 |
-
with torch.no_grad():
|
1033 |
-
output = self.model(img_LR)
|
1034 |
-
output = output.data.float().clamp_(0, 1)
|
1035 |
-
output = output[:,[2,1,0],:,:]
|
1036 |
-
return output
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
app.py
CHANGED
@@ -1,24 +1,22 @@
|
|
1 |
import gradio as gr
|
2 |
-
from
|
3 |
import torch
|
4 |
-
import numpy as np
|
5 |
from PIL import Image
|
6 |
|
7 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
8 |
model_id = "alibaba-pai/pai-diffusion-food-large-zh"
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
|
13 |
def infer_text2img(prompt, guide, steps):
|
14 |
-
|
15 |
-
|
16 |
-
return images
|
17 |
|
18 |
with gr.Blocks() as demo:
|
19 |
examples = [
|
20 |
-
["
|
21 |
-
["韩式炸鸡"]
|
|
|
22 |
]
|
23 |
with gr.Row():
|
24 |
with gr.Column(scale=0.5, ):
|
|
|
1 |
import gradio as gr
|
2 |
+
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler
|
3 |
import torch
|
|
|
4 |
from PIL import Image
|
5 |
|
|
|
6 |
model_id = "alibaba-pai/pai-diffusion-food-large-zh"
|
7 |
+
pipe = StableDiffusionPipeline.from_pretrained(model_id)
|
8 |
+
pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
|
9 |
+
pipe = pipe.to("cuda")
|
10 |
|
11 |
def infer_text2img(prompt, guide, steps):
|
12 |
+
image = pipe([prompt], guidance_scale=guide, num_inference_steps=steps).images[0]
|
13 |
+
return image
|
|
|
14 |
|
15 |
with gr.Blocks() as demo:
|
16 |
examples = [
|
17 |
+
["过桥米线"],
|
18 |
+
["韩式炸鸡"],
|
19 |
+
["小炒黄牛肉"],
|
20 |
]
|
21 |
with gr.Row():
|
22 |
with gr.Column(scale=0.5, ):
|
requirements.txt
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
--extra-index-url https://download.pytorch.org/whl/cu113
|
2 |
torch
|
3 |
torchvision
|
4 |
-
diffusers==0.
|
5 |
transformers
|
6 |
accelerate
|
|
|
1 |
--extra-index-url https://download.pytorch.org/whl/cu113
|
2 |
torch
|
3 |
torchvision
|
4 |
+
diffusers==0.14.0
|
5 |
transformers
|
6 |
accelerate
|