# coding=utf-8
import importlib
import inspect
import os
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Union
from collections import OrderedDict
import numpy as np
import torch
from torch import nn
import functools
import diffusers
import PIL
from accelerate.utils.versions import is_torch_version
from huggingface_hub import snapshot_download
from packaging import version
from PIL import Image
from tqdm.auto import tqdm
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.dynamic_modules_utils import get_class_from_dynamic_module
from diffusers.modeling_utils import ModelMixin
from diffusers.hub_utils import http_user_agent
from diffusers.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT
from diffusers.schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME
from diffusers.utils import (
CONFIG_NAME,
DIFFUSERS_CACHE,
ONNX_WEIGHTS_NAME,
WEIGHTS_NAME,
BaseOutput,
deprecate,
is_transformers_available,
logging,
)
if is_transformers_available():
import transformers
from transformers import PreTrainedModel
INDEX_FILE = "diffusion_pytorch_model.bin"
CUSTOM_PIPELINE_FILE_NAME = "pipeline.py"
DUMMY_MODULES_FOLDER = "diffusers.utils"
logger = logging.get_logger(__name__)
LOADABLE_CLASSES = {
"diffusers": {
"ModelMixin": ["save_pretrained", "from_pretrained"],
"SchedulerMixin": ["save_config", "from_config"],
"DiffusionPipeline": ["save_pretrained", "from_pretrained"],
"OnnxRuntimeModel": ["save_pretrained", "from_pretrained"],
},
"transformers": {
"PreTrainedTokenizer": ["save_pretrained", "from_pretrained"],
"PreTrainedTokenizerFast": ["save_pretrained", "from_pretrained"],
"PreTrainedModel": ["save_pretrained", "from_pretrained"],
"FeatureExtractionMixin": ["save_pretrained", "from_pretrained"],
},
"LdmZhPipeline": {
"WukongClipTextEncoder": ["save_pretrained", "from_pretrained"],
"ESRGAN": ["save_pretrained", "from_pretrained"],
},
}
ALL_IMPORTABLE_CLASSES = {}
for library in LOADABLE_CLASSES:
ALL_IMPORTABLE_CLASSES.update(LOADABLE_CLASSES[library])
@dataclass
class ImagePipelineOutput(BaseOutput):
"""
Output class for image pipelines.
Args:
images (`List[PIL.Image.Image]` or `np.ndarray`)
List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
"""
images: Union[List[PIL.Image.Image], np.ndarray]
@dataclass
class AudioPipelineOutput(BaseOutput):
"""
Output class for audio pipelines.
Args:
audios (`np.ndarray`)
List of denoised samples of shape `(batch_size, num_channels, sample_rate)`. Numpy array present the
denoised audio samples of the diffusion pipeline.
"""
audios: np.ndarray
class DiffusionPipeline(ConfigMixin):
r"""
Base class for all models.
[`DiffusionPipeline`] takes care of storing all components (models, schedulers, processors) for diffusion pipelines
and handles methods for loading, downloading and saving models as well as a few methods common to all pipelines to:
- move all PyTorch modules to the device of your choice
- enabling/disabling the progress bar for the denoising iteration
Class attributes:
- **config_name** ([`str`]) -- name of the config file that will store the class and module names of all
components of the diffusion pipeline.
"""
config_name = "model_index.json"
def register_modules(self, **kwargs):
# import it here to avoid circular import
from diffusers import pipelines
for name, module in kwargs.items():
# retrieve library
if module is None:
register_dict = {name: (None, None)}
else:
library = module.__module__.split(".")[0]
# check if the module is a pipeline module
pipeline_dir = module.__module__.split(".")[-2] if len(module.__module__.split(".")) > 2 else None
path = module.__module__.split(".")
is_pipeline_module = pipeline_dir in path and hasattr(pipelines, pipeline_dir)
# if library is not in LOADABLE_CLASSES, then it is a custom module.
# Or if it's a pipeline module, then the module is inside the pipeline
# folder so we set the library to module name.
if library not in LOADABLE_CLASSES or is_pipeline_module:
library = pipeline_dir
# retrieve class_name
class_name = module.__class__.__name__
register_dict = {name: (library, class_name)}
# save model index config
self.register_to_config(**register_dict)
# set models
setattr(self, name, module)
def save_pretrained(self, save_directory: Union[str, os.PathLike]):
"""
Save all variables of the pipeline that can be saved and loaded as well as the pipelines configuration file to
a directory. A pipeline variable can be saved and loaded if its class implements both a save and loading
method. The pipeline can easily be re-loaded using the `[`~DiffusionPipeline.from_pretrained`]` class method.
Arguments:
save_directory (`str` or `os.PathLike`):
Directory to which to save. Will be created if it doesn't exist.
"""
self.save_config(save_directory)
model_index_dict = dict(self.config)
model_index_dict.pop("_class_name")
model_index_dict.pop("_diffusers_version")
model_index_dict.pop("_module", None)
for pipeline_component_name in model_index_dict.keys():
sub_model = getattr(self, pipeline_component_name)
if sub_model is None:
# edge case for saving a pipeline with safety_checker=None
continue
model_cls = sub_model.__class__
save_method_name = None
# search for the model's base class in LOADABLE_CLASSES
for library_name, library_classes in LOADABLE_CLASSES.items():
library = importlib.import_module(library_name)
for base_class, save_load_methods in library_classes.items():
class_candidate = getattr(library, base_class)
if issubclass(model_cls, class_candidate):
# if we found a suitable base class in LOADABLE_CLASSES then grab its save method
save_method_name = save_load_methods[0]
break
if save_method_name is not None:
break
save_method = getattr(sub_model, save_method_name)
save_method(os.path.join(save_directory, pipeline_component_name))
def to(self, torch_device: Optional[Union[str, torch.device]] = None):
if torch_device is None:
return self
module_names, _ = self.extract_init_dict(dict(self.config))
for name in module_names.keys():
module = getattr(self, name)
if isinstance(module, torch.nn.Module):
if module.dtype == torch.float16 and str(torch_device) in ["cpu", "mps"]:
logger.warning(
"Pipelines loaded with `torch_dtype=torch.float16` cannot run with `cpu` or `mps` device. It"
" is not recommended to move them to `cpu` or `mps` as running them will fail. Please make"
" sure to use a `cuda` device to run the pipeline in inference. due to the lack of support for"
" `float16` operations on those devices in PyTorch. Please remove the"
" `torch_dtype=torch.float16` argument, or use a `cuda` device to run inference."
)
module.to(torch_device)
return self
@property
def device(self) -> torch.device:
r"""
Returns:
`torch.device`: The torch device on which the pipeline is located.
"""
module_names, _ = self.extract_init_dict(dict(self.config))
for name in module_names.keys():
module = getattr(self, name)
if isinstance(module, torch.nn.Module):
# if module.device == torch.device("meta"):
# return torch.device("cpu")
return module.device
return torch.device("cpu")
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
r"""
Instantiate a PyTorch diffusion pipeline from pre-trained pipeline weights.
The pipeline is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated).
The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
task.
The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
weights are discarded.
Parameters:
pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
Can be either:
- A string, the *repo id* of a pretrained pipeline hosted inside a model repo on
https://huggingface.co/ Valid repo ids have to be located under a user or organization name, like
`CompVis/ldm-text2im-large-256`.
- A path to a *directory* containing pipeline weights saved using
[`~DiffusionPipeline.save_pretrained`], e.g., `./my_pipeline_directory/`.
torch_dtype (`str` or `torch.dtype`, *optional*):
Override the default `torch.dtype` and load the model under this dtype. If `"auto"` is passed the dtype
will be automatically derived from the model's weights.
custom_pipeline (`str`, *optional*):
This is an experimental feature and is likely to change in the future.
Can be either:
- A string, the *repo id* of a custom pipeline hosted inside a model repo on
https://huggingface.co/. Valid repo ids have to be located under a user or organization name,
like `hf-internal-testing/diffusers-dummy-pipeline`.
It is required that the model repo has a file, called `pipeline.py` that defines the custom
pipeline.
- A string, the *file name* of a community pipeline hosted on GitHub under
https://github.com/huggingface/diffusers/tree/main/examples/community. Valid file names have to
match exactly the file name without `.py` located under the above link, *e.g.*
`clip_guided_stable_diffusion`.
Community pipelines are always loaded from the current `main` branch of GitHub.
- A path to a *directory* containing a custom pipeline, e.g., `./my_pipeline_directory/`.
It is required that the directory has a file, called `pipeline.py` that defines the custom
pipeline.
For more information on how to load and create custom pipelines, please have a look at [Loading and
Creating Custom
Pipelines](https://huggingface.co/docs/diffusers/main/en/using-diffusers/custom_pipelines)
torch_dtype (`str` or `torch.dtype`, *optional*):
force_download (`bool`, *optional*, defaults to `False`):
Whether or not to force the (re-)download of the model weights and configuration files, overriding the
cached versions if they exist.
resume_download (`bool`, *optional*, defaults to `False`):
Whether or not to delete incompletely received files. Will attempt to resume the download if such a
file exists.
proxies (`Dict[str, str]`, *optional*):
A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
output_loading_info(`bool`, *optional*, defaults to `False`):
Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
local_files_only(`bool`, *optional*, defaults to `False`):
Whether or not to only look at local files (i.e., do not try to download the model).
use_auth_token (`str` or *bool*, *optional*):
The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
when running `huggingface-cli login` (stored in `~/.huggingface`).
revision (`str`, *optional*, defaults to `"main"`):
The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
identifier allowed by git.
mirror (`str`, *optional*):
Mirror source to accelerate downloads in China. If you are from China and have an accessibility
problem, you can set this option to resolve it. Note that we do not guarantee the timeliness or safety.
Please refer to the mirror site for more information. specify the folder name here.
device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
A map that specifies where each submodule should go. It doesn't need to be refined to each
parameter/buffer name, once a given module name is inside, every submodule of it will be sent to the
same device.
To have Accelerate compute the most optimized `device_map` automatically, set `device_map="auto"`. For
more information about each option see [designing a device
map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
Speed up model loading by not initializing the weights and only loading the pre-trained weights. This
also tries to not use more than 1x model size in CPU memory (including peak memory) while loading the
model. This is only supported when torch version >= 1.9.0. If you are using an older version of torch,
setting this argument to `True` will raise an error.
kwargs (remaining dictionary of keyword arguments, *optional*):
Can be used to overwrite load - and saveable variables - *i.e.* the pipeline components - of the
specific pipeline class. The overwritten components are then directly passed to the pipelines
`__init__` method. See example below for more information.
It is required to be logged in (`huggingface-cli login`) when you want to use private or [gated
models](https://huggingface.co/docs/hub/models-gated#gated-models), *e.g.* `"runwayml/stable-diffusion-v1-5"`
Activate the special ["offline-mode"](https://huggingface.co/diffusers/installation.html#offline-mode) to use
this method in a firewalled environment.
Examples:
```py
>>> from diffusers import DiffusionPipeline
>>> # Download pipeline from huggingface.co and cache.
>>> pipeline = DiffusionPipeline.from_pretrained("CompVis/ldm-text2im-large-256")
>>> # Download pipeline that requires an authorization token
>>> # For more information on access tokens, please refer to this section
>>> # of the documentation](https://huggingface.co/docs/hub/security-tokens)
>>> pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
>>> # Download pipeline, but overwrite scheduler
>>> from diffusers import LMSDiscreteScheduler
>>> scheduler = LMSDiscreteScheduler.from_config("runwayml/stable-diffusion-v1-5", subfolder="scheduler")
>>> pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", scheduler=scheduler)
```
"""
cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
resume_download = kwargs.pop("resume_download", False)
force_download = kwargs.pop("force_download", False)
proxies = kwargs.pop("proxies", None)
local_files_only = kwargs.pop("local_files_only", False)
use_auth_token = kwargs.pop("use_auth_token", None)
revision = kwargs.pop("revision", None)
torch_dtype = kwargs.pop("torch_dtype", None)
custom_pipeline = kwargs.pop("custom_pipeline", None)
provider = kwargs.pop("provider", None)
sess_options = kwargs.pop("sess_options", None)
device_map = kwargs.pop("device_map", None)
low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
if device_map is not None and not is_torch_version(">=", "1.9.0"):
raise NotImplementedError(
"Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set"
" `device_map=None`."
)
if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
raise NotImplementedError(
"Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
" `low_cpu_mem_usage=False`."
)
if low_cpu_mem_usage is False and device_map is not None:
raise ValueError(
f"You cannot set `low_cpu_mem_usage` to False while using device_map={device_map} for loading and"
" dispatching. Please make sure to set `low_cpu_mem_usage=True`."
)
# 1. Download the checkpoints and configs
# use snapshot download here to get it working from from_pretrained
if not os.path.isdir(pretrained_model_name_or_path):
config_dict = cls.get_config_dict(
pretrained_model_name_or_path,
cache_dir=cache_dir,
resume_download=resume_download,
force_download=force_download,
proxies=proxies,
local_files_only=local_files_only,
use_auth_token=use_auth_token,
revision=revision,
)
# make sure we only download sub-folders and `diffusers` filenames
folder_names = [k for k in config_dict.keys() if not k.startswith("_")]
allow_patterns = [os.path.join(k, "*") for k in folder_names]
allow_patterns += [WEIGHTS_NAME, SCHEDULER_CONFIG_NAME, CONFIG_NAME, ONNX_WEIGHTS_NAME, cls.config_name]
# make sure we don't download flax weights
ignore_patterns = "*.msgpack"
if custom_pipeline is not None:
allow_patterns += [CUSTOM_PIPELINE_FILE_NAME]
if cls != DiffusionPipeline:
requested_pipeline_class = cls.__name__
else:
requested_pipeline_class = config_dict.get("_class_name", cls.__name__)
user_agent = {"pipeline_class": requested_pipeline_class}
if custom_pipeline is not None:
user_agent["custom_pipeline"] = custom_pipeline
user_agent = http_user_agent(user_agent)
# download all allow_patterns
cached_folder = snapshot_download(
pretrained_model_name_or_path,
cache_dir=cache_dir,
resume_download=resume_download,
proxies=proxies,
local_files_only=local_files_only,
use_auth_token=use_auth_token,
revision=revision,
allow_patterns=allow_patterns,
ignore_patterns=ignore_patterns,
user_agent=user_agent,
)
else:
cached_folder = pretrained_model_name_or_path
config_dict = cls.get_config_dict(cached_folder)
# 2. Load the pipeline class, if using custom module then load it from the hub
# if we load from explicit class, let's use it
if custom_pipeline is not None:
pipeline_class = get_class_from_dynamic_module(
custom_pipeline, module_file=CUSTOM_PIPELINE_FILE_NAME, cache_dir=custom_pipeline
)
elif cls != DiffusionPipeline:
pipeline_class = cls
else:
diffusers_module = importlib.import_module(cls.__module__.split(".")[0])
pipeline_class = getattr(diffusers_module, config_dict["_class_name"])
# To be removed in 1.0.0
if pipeline_class.__name__ == "StableDiffusionInpaintPipeline" and version.parse(
version.parse(config_dict["_diffusers_version"]).base_version
) <= version.parse("0.5.1"):
from diffusers import StableDiffusionInpaintPipeline, StableDiffusionInpaintPipelineLegacy
pipeline_class = StableDiffusionInpaintPipelineLegacy
deprecation_message = (
"You are using a legacy checkpoint for inpainting with Stable Diffusion, therefore we are loading the"
f" {StableDiffusionInpaintPipelineLegacy} class instead of {StableDiffusionInpaintPipeline}. For"
" better inpainting results, we strongly suggest using Stable Diffusion's official inpainting"
" checkpoint: https://huggingface.co/runwayml/stable-diffusion-inpainting instead or adapting your"
f" checkpoint {pretrained_model_name_or_path} to the format of"
" https://huggingface.co/runwayml/stable-diffusion-inpainting. Note that we do not actively maintain"
" the {StableDiffusionInpaintPipelineLegacy} class and will likely remove it in version 1.0.0."
)
deprecate("StableDiffusionInpaintPipelineLegacy", "1.0.0", deprecation_message, standard_warn=False)
# some modules can be passed directly to the init
# in this case they are already instantiated in `kwargs`
# extract them here
expected_modules = set(inspect.signature(pipeline_class.__init__).parameters.keys()) - set(["self"])
passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs}
init_dict, unused_kwargs = pipeline_class.extract_init_dict(config_dict, **kwargs)
if len(unused_kwargs) > 0:
logger.warning(f"Keyword arguments {unused_kwargs} not recognized.")
init_kwargs = {}
# import it here to avoid circular import
from diffusers import pipelines
# 3. Load each module in the pipeline
for name, (library_name, class_name) in init_dict.items():
if class_name is None:
# edge case for when the pipeline was saved with safety_checker=None
init_kwargs[name] = None
continue
# 3.1 - now that JAX/Flax is an official framework of the library, we might load from Flax names
if class_name.startswith("Flax"):
class_name = class_name[4:]
is_pipeline_module = hasattr(pipelines, library_name)
loaded_sub_model = None
sub_model_should_be_defined = True
# if the model is in a pipeline module, then we load it from the pipeline
if name in passed_class_obj:
# 1. check that passed_class_obj has correct parent class
if not is_pipeline_module:
library = importlib.import_module(library_name)
class_obj = getattr(library, class_name)
importable_classes = LOADABLE_CLASSES[library_name]
class_candidates = {c: getattr(library, c) for c in importable_classes.keys()}
expected_class_obj = None
for class_name, class_candidate in class_candidates.items():
if issubclass(class_obj, class_candidate):
expected_class_obj = class_candidate
if not issubclass(passed_class_obj[name].__class__, expected_class_obj):
raise ValueError(
f"{passed_class_obj[name]} is of type: {type(passed_class_obj[name])}, but should be"
f" {expected_class_obj}"
)
elif passed_class_obj[name] is None:
logger.warn(
f"You have passed `None` for {name} to disable its functionality in {pipeline_class}. Note"
f" that this might lead to problems when using {pipeline_class} and is not recommended."
)
sub_model_should_be_defined = False
else:
logger.warn(
f"You have passed a non-standard module {passed_class_obj[name]}. We cannot verify whether it"
" has the correct type"
)
# set passed class object
loaded_sub_model = passed_class_obj[name]
elif is_pipeline_module:
pipeline_module = getattr(pipelines, library_name)
class_obj = getattr(pipeline_module, class_name)
importable_classes = ALL_IMPORTABLE_CLASSES
class_candidates = {c: class_obj for c in importable_classes.keys()}
else:
# else we just import it from the library.
library = importlib.import_module(library_name)
class_obj = getattr(library, class_name)
importable_classes = LOADABLE_CLASSES[library_name]
class_candidates = {c: getattr(library, c) for c in importable_classes.keys()}
if loaded_sub_model is None and sub_model_should_be_defined:
load_method_name = None
for class_name, class_candidate in class_candidates.items():
if issubclass(class_obj, class_candidate):
load_method_name = importable_classes[class_name][1]
if load_method_name is None:
none_module = class_obj.__module__
if none_module.startswith(DUMMY_MODULES_FOLDER) and "dummy" in none_module:
# call class_obj for nice error message of missing requirements
class_obj()
raise ValueError(
f"The component {class_obj} of {pipeline_class} cannot be loaded as it does not seem to have"
f" any of the loading methods defined in {ALL_IMPORTABLE_CLASSES}."
)
load_method = getattr(class_obj, load_method_name)
loading_kwargs = {}
if issubclass(class_obj, torch.nn.Module):
loading_kwargs["torch_dtype"] = torch_dtype
if issubclass(class_obj, diffusers.OnnxRuntimeModel):
loading_kwargs["provider"] = provider
loading_kwargs["sess_options"] = sess_options
is_diffusers_model = issubclass(class_obj, diffusers.ModelMixin)
is_transformers_model = (
is_transformers_available()
and issubclass(class_obj, PreTrainedModel)
and version.parse(version.parse(transformers.__version__).base_version) >= version.parse("4.20.0")
)
# When loading a transformers model, if the device_map is None, the weights will be initialized as opposed to diffusers.
# To make default loading faster we set the `low_cpu_mem_usage=low_cpu_mem_usage` flag which is `True` by default.
# This makes sure that the weights won't be initialized which significantly speeds up loading.
if is_diffusers_model or is_transformers_model:
loading_kwargs["device_map"] = device_map
loading_kwargs["low_cpu_mem_usage"] = low_cpu_mem_usage
# check if the module is in a subdirectory
if os.path.isdir(os.path.join(cached_folder, name)):
loaded_sub_model = load_method(os.path.join(cached_folder, name), **loading_kwargs)
else:
# else load from the root directory
loaded_sub_model = load_method(cached_folder, **loading_kwargs)
init_kwargs[name] = loaded_sub_model # UNet(...), # DiffusionSchedule(...)
# 4. Potentially add passed objects if expected
missing_modules = set(expected_modules) - set(init_kwargs.keys())
if len(missing_modules) > 0 and missing_modules <= set(passed_class_obj.keys()):
for module in missing_modules:
init_kwargs[module] = passed_class_obj[module]
elif len(missing_modules) > 0:
passed_modules = set(list(init_kwargs.keys()) + list(passed_class_obj.keys()))
raise ValueError(
f"Pipeline {pipeline_class} expected {expected_modules}, but only {passed_modules} were passed."
)
# 5. Instantiate the pipeline
model = pipeline_class(**init_kwargs)
return model
@property
def components(self) -> Dict[str, Any]:
r"""
The `self.components` property can be useful to run different pipelines with the same weights and
configurations to not have to re-allocate memory.
Examples:
```py
>>> from diffusers import (
... StableDiffusionPipeline,
... StableDiffusionImg2ImgPipeline,
... StableDiffusionInpaintPipeline,
... )
>>> img2text = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
>>> img2img = StableDiffusionImg2ImgPipeline(**img2text.components)
>>> inpaint = StableDiffusionInpaintPipeline(**img2text.components)
```
Returns:
A dictionaly containing all the modules needed to initialize the pipeline.
"""
components = {k: getattr(self, k) for k in self.config.keys() if not k.startswith("_")}
expected_modules = set(inspect.signature(self.__init__).parameters.keys()) - set(["self"])
if set(components.keys()) != expected_modules:
raise ValueError(
f"{self} has been incorrectly initialized or {self.__class__} is incorrectly implemented. Expected"
f" {expected_modules} to be defined, but {components} are defined."
)
return components
@staticmethod
def numpy_to_pil(images):
"""
Convert a numpy image or a batch of images to a PIL image.
"""
if images.ndim == 3:
images = images[None, ...]
images = (images * 255).round().astype("uint8")
if images.shape[-1] == 1:
# special case for grayscale (single channel) images
pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images]
else:
pil_images = [Image.fromarray(image) for image in images]
return pil_images
def progress_bar(self, iterable):
if not hasattr(self, "_progress_bar_config"):
self._progress_bar_config = {}
elif not isinstance(self._progress_bar_config, dict):
raise ValueError(
f"`self._progress_bar_config` should be of type `dict`, but is {type(self._progress_bar_config)}."
)
return tqdm(iterable, **self._progress_bar_config)
def set_progress_bar_config(self, **kwargs):
self._progress_bar_config = kwargs
class LDMZhTextToImagePipeline(DiffusionPipeline):
def __init__(
self,
vqvae,
bert,
tokenizer,
unet,
scheduler,
sr,
):
super().__init__()
self.register_modules(vqvae=vqvae, bert=bert, tokenizer=tokenizer, unet=unet, scheduler=scheduler, sr=sr)
@torch.no_grad()
def __call__(
self,
prompt: Union[str, List[str]],
height: Optional[int] = 256,
width: Optional[int] = 256,
num_inference_steps: Optional[int] = 50,
guidance_scale: Optional[float] = 5.0,
eta: Optional[float] = 0.0,
generator: Optional[torch.Generator] = None,
output_type: Optional[str] = "pil",
return_dict: bool = True,
use_sr: bool = False,
**kwargs,
):
r"""
Args:
prompt (`str` or `List[str]`):
The prompt or prompts to guide the image generation.
height (`int`, *optional*, defaults to 256):
The height in pixels of the generated image.
width (`int`, *optional*, defaults to 256):
The width in pixels of the generated image.
num_inference_steps (`int`, *optional*, defaults to 50):
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
expense of slower inference.
guidance_scale (`float`, *optional*, defaults to 1.0):
Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
`guidance_scale` is defined as `w` of equation 2. of [Imagen
Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt` at
the, usually at the expense of lower image quality.
generator (`torch.Generator`, *optional*):
A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation
deterministic.
output_type (`str`, *optional*, defaults to `"pil"`):
The output format of the generate image. Choose between
[PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
return_dict (`bool`, *optional*):
Whether or not to return a [`~pipeline_utils.ImagePipelineOutput`] instead of a plain tuple.
Returns:
[`~pipeline_utils.ImagePipelineOutput`] or `tuple`: [`~pipelines.utils.ImagePipelineOutput`] if
`return_dict` is True, otherwise a `tuple. When returning a tuple, the first element is a list with the
generated images.
"""
if isinstance(prompt, str):
batch_size = 1
elif isinstance(prompt, list):
batch_size = len(prompt)
else:
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
if height % 8 != 0 or width % 8 != 0:
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
# get unconditional embeddings for classifier free guidance
if guidance_scale != 1.0:
uncond_input = self.tokenizer([""] * batch_size, padding="max_length", max_length=32, return_tensors="pt")
uncond_embeddings = self.bert(uncond_input.input_ids.to(self.device))
# get prompt text embeddings
text_input = self.tokenizer(prompt, padding="max_length", max_length=32, return_tensors="pt")
text_embeddings = self.bert(text_input.input_ids.to(self.device))
latents = torch.randn(
(batch_size, self.unet.in_channels, height // 8, width // 8),
generator=generator,
)
latents = latents.to(self.device)
self.scheduler.set_timesteps(num_inference_steps)
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
extra_kwargs = {}
if accepts_eta:
extra_kwargs["eta"] = eta
for t in self.progress_bar(self.scheduler.timesteps):
if guidance_scale == 1.0:
# guidance_scale of 1 means no guidance
latents_input = latents
context = text_embeddings
else:
# For classifier free guidance, we need to do two forward passes.
# Here we concatenate the unconditional and text embeddings into a single batch
# to avoid doing two forward passes
latents_input = torch.cat([latents] * 2)
context = torch.cat([uncond_embeddings, text_embeddings])
# predict the noise residual
noise_pred = self.unet(latents_input, t, encoder_hidden_states=context).sample
# perform guidance
if guidance_scale != 1.0:
noise_pred_uncond, noise_prediction_text = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond)
# compute the previous noisy sample x_t -> x_t-1
latents = self.scheduler.step(noise_pred, t, latents, **extra_kwargs).prev_sample
# scale and decode the image latents with vae
latents = 1 / 0.18215 * latents
image = self.vqvae.decode(latents).sample
image = (image / 2 + 0.5).clamp(0, 1)
if use_sr:
image = self.sr(image)
image = image.cpu().permute(0, 2, 3, 1).numpy()
if output_type == "pil":
image = self.numpy_to_pil(image)
if not return_dict:
return (image,)
return ImagePipelineOutput(images=image)
class QuickGELU(nn.Module):
def forward(self, x: torch.Tensor):
return x * torch.sigmoid(1.702 * x)
class ResidualAttentionBlock(nn.Module):
def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):
super().__init__()
self.attn = nn.MultiheadAttention(d_model, n_head)
self.ln_1 = nn.LayerNorm(d_model,eps=1e-07)
self.mlp = nn.Sequential(OrderedDict([
("c_fc", nn.Linear(d_model, d_model * 4)),
("gelu", QuickGELU()),
("c_proj", nn.Linear(d_model * 4, d_model))
]))
self.ln_2 = nn.LayerNorm(d_model,eps=1e-07)
self.attn_mask = attn_mask
def attention(self, x: torch.Tensor):
self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None
return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
def forward(self, x: torch.Tensor):
x = x + self.attention(self.ln_1(x))
x = x + self.mlp(self.ln_2(x))
return x
class Transformer(nn.Module):
def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):
super().__init__()
self.width = width
self.layers = layers
self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])
def forward(self, x: torch.Tensor):
return self.resblocks(x)
class TextTransformer(nn.Module):
def __init__(self,
context_length = 32,
vocab_size = 21128,
output_dim = 768,
width = 768,
layers = 12,
heads = 12,
return_full_embed = False):
super(TextTransformer, self).__init__()
self.width = width
self.layers = layers
self.vocab_size = vocab_size
self.return_full_embed = return_full_embed
self.transformer = Transformer(width, layers, heads, self.build_attntion_mask(context_length))
self.text_projection = torch.nn.Parameter(
torch.tensor(np.random.normal(0, self.width ** -0.5, size=(self.width, output_dim)).astype(np.float32)))
self.ln_final = nn.LayerNorm(width,eps=1e-07)
# https://discuss.pytorch.org/t/implementing-truncated-normal-initializer/4778/27
# https://github.com/pytorch/pytorch/blob/a40812de534b42fcf0eb57a5cecbfdc7a70100cf/torch/nn/init.py#L22
self.embedding_table = nn.Parameter(nn.init.trunc_normal_(torch.empty(vocab_size, width),std=0.02))
# self.embedding_table = nn.Embedding.from_pretrained(nn.init.trunc_normal_(torch.empty(vocab_size, width),std=0.02))
self.positional_embedding = nn.Parameter(nn.init.trunc_normal_(torch.empty(context_length, width),std=0.01))
# self.positional_embedding = nn.Embedding.from_pretrained(nn.init.trunc_normal_(torch.empty(context_length, width),std=0.01))
self.index_select=torch.index_select
self.reshape=torch.reshape
@staticmethod
def build_attntion_mask(context_length):
mask = np.triu(np.full((context_length, context_length), -np.inf).astype(np.float32), 1)
mask = torch.tensor(mask)
return mask
def forward(self, x: torch.Tensor):
tail_token=(x==102).nonzero(as_tuple=True)
bsz, ctx_len = x.shape
flatten_id = x.flatten()
index_select_result = self.index_select(self.embedding_table,0, flatten_id)
x = self.reshape(index_select_result, (bsz, ctx_len, -1))
x = x + self.positional_embedding
x = x.permute(1, 0, 2) # NLD -> LND
x = self.transformer(x)
x = x.permute(1, 0, 2) # LND -> NLD
x = self.ln_final(x)
x=x[tail_token]
x = x @ self.text_projection
return x
class WukongClipTextEncoder(ModelMixin, ConfigMixin):
@register_to_config
def __init__(
self,
):
super().__init__()
self.model = TextTransformer()
def forward(
self,
tokens
):
z = self.model(tokens)
z = z / torch.linalg.norm(z, dim=-1, keepdim=True)
if z.ndim==2:
z = z.view((z.shape[0], 1, z.shape[1]))
return z
def make_layer(block, n_layers):
layers = []
for _ in range(n_layers):
layers.append(block())
return nn.Sequential(*layers)
class ResidualDenseBlock_5C(nn.Module):
def __init__(self, nf=64, gc=32, bias=True):
super(ResidualDenseBlock_5C, self).__init__()
# gc: growth channel, i.e. intermediate channels
self.conv1 = nn.Conv2d(nf, gc, 3, 1, 1, bias=bias)
self.conv2 = nn.Conv2d(nf + gc, gc, 3, 1, 1, bias=bias)
self.conv3 = nn.Conv2d(nf + 2 * gc, gc, 3, 1, 1, bias=bias)
self.conv4 = nn.Conv2d(nf + 3 * gc, gc, 3, 1, 1, bias=bias)
self.conv5 = nn.Conv2d(nf + 4 * gc, nf, 3, 1, 1, bias=bias)
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
# initialization
# mutil.initialize_weights([self.conv1, self.conv2, self.conv3, self.conv4, self.conv5], 0.1)
def forward(self, x):
x1 = self.lrelu(self.conv1(x))
x2 = self.lrelu(self.conv2(torch.cat((x, x1), 1)))
x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), 1)))
x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), 1)))
x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1))
return x5 * 0.2 + x
class RRDB(nn.Module):
'''Residual in Residual Dense Block'''
def __init__(self, nf, gc=32):
super(RRDB, self).__init__()
self.RDB1 = ResidualDenseBlock_5C(nf, gc)
self.RDB2 = ResidualDenseBlock_5C(nf, gc)
self.RDB3 = ResidualDenseBlock_5C(nf, gc)
def forward(self, x):
out = self.RDB1(x)
out = self.RDB2(out)
out = self.RDB3(out)
return out * 0.2 + x
class RRDBNet(nn.Module):
def __init__(self, in_nc, out_nc, nf, nb, gc=32):
super(RRDBNet, self).__init__()
RRDB_block_f = functools.partial(RRDB, nf=nf, gc=gc)
self.conv_first = nn.Conv2d(in_nc, nf, 3, 1, 1, bias=True)
self.RRDB_trunk = make_layer(RRDB_block_f, nb)
self.trunk_conv = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
#### upsampling
self.upconv1 = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
self.upconv2 = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
self.HRconv = nn.Conv2d(nf, nf, 3, 1, 1, bias=True)
self.conv_last = nn.Conv2d(nf, out_nc, 3, 1, 1, bias=True)
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
def forward(self, x):
fea = self.conv_first(x)
trunk = self.trunk_conv(self.RRDB_trunk(fea))
fea = fea + trunk
fea = self.lrelu(self.upconv1(torch.nn.functional.interpolate(fea, scale_factor=2, mode='nearest')))
fea = self.lrelu(self.upconv2(torch.nn.functional.interpolate(fea, scale_factor=2, mode='nearest')))
out = self.conv_last(self.lrelu(self.HRconv(fea)))
return out
class ESRGAN(ModelMixin, ConfigMixin):
@register_to_config
def __init__(
self,
):
super().__init__()
self.model = RRDBNet(3, 3, 64, 23, gc=32)
def forward(
self,
img_LR
):
img_LR = img_LR[:,[2,1,0],:,:]
img_LR = img_LR.to(self.device)
with torch.no_grad():
output = self.model(img_LR)
output = output.data.float().clamp_(0, 1)
output = output[:,[2,1,0],:,:]
return output