declare-lab/tango2
92
1# coding=utf-82# Copyright 2023 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 at8#9# http://www.apache.org/licenses/LICENSE-2.010#11# Unless required by applicable law or agreed to in writing, software12# 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 and15# limitations under the License.16 17import inspect18import os19from functools import partial20from typing import Callable, List, Optional, Tuple, Union21 22import torch23from torch import Tensor, device24 25from .. import __version__26from ..utils import (27 CONFIG_NAME,28 DIFFUSERS_CACHE,29 FLAX_WEIGHTS_NAME,30 HF_HUB_OFFLINE,31 SAFETENSORS_WEIGHTS_NAME,32 WEIGHTS_NAME,33 _add_variant,34 _get_model_file,35 is_accelerate_available,36 is_safetensors_available,37 is_torch_version,38 logging,39)40 41 42logger = logging.get_logger(__name__)43 44 45if is_torch_version(">=", "1.9.0"):46 _LOW_CPU_MEM_USAGE_DEFAULT = True47else:48 _LOW_CPU_MEM_USAGE_DEFAULT = False49 50 51if is_accelerate_available():52 import accelerate53 from accelerate.utils import set_module_tensor_to_device54 from accelerate.utils.versions import is_torch_version55 56if is_safetensors_available():57 import safetensors58 59 60def get_parameter_device(parameter: torch.nn.Module):61 try:62 return next(parameter.parameters()).device63 except StopIteration:64 # For torch.nn.DataParallel compatibility in PyTorch 1.565 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 tuples69 70 gen = parameter._named_members(get_members_fn=find_tensor_attributes)71 first_tuple = next(gen)72 return first_tuple[1].device73 74 75def get_parameter_dtype(parameter: torch.nn.Module):76 try:77 return next(parameter.parameters()).dtype78 except StopIteration:79 # For torch.nn.DataParallel compatibility in PyTorch 1.580 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 tuples84 85 gen = parameter._named_members(get_members_fn=find_tensor_attributes)86 first_tuple = next(gen)87 return first_tuple[1].dtype88 89 90def load_state_dict(checkpoint_file: Union[str, os.PathLike], variant: Optional[str] = None):91 """92 Reads a checkpoint file, returning properly formatted errors if they arise.93 """94 try:95 if os.path.basename(checkpoint_file) == _add_variant(WEIGHTS_NAME, variant):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 e113 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 121def _load_state_dict_into_model(model_to_load, state_dict):122 # Convert old format to new format if needed from a PyTorch state_dict123 # copy state_dict so _load_from_state_dict can modify it124 state_dict = state_dict.copy()125 error_msgs = []126 127 # PyTorch's `_load_from_state_dict` does not copy parameters in a module's descendants128 # 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_msgs140 141 142class 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, downloading147 and saving models.148 149 - **config_name** ([`str`]) -- A filename under which the model should be stored when calling150 [`~models.ModelMixin.save_pretrained`].151 """152 config_name = CONFIG_NAME153 _automatically_saved_args = ["_diffusers_version", "_class_name", "_name_or_path"]154 _supports_gradient_checkpointing = False155 156 def __init__(self):157 super().__init__()158 159 @property160 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 "checkpoint165 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 "checkpoint174 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 "checkpoint185 activations".186 """187 if self._supports_gradient_checkpointing:188 self.apply(partial(self._set_gradient_checkpointing, value=False))189 190 def set_use_memory_efficient_attention_xformers(191 self, valid: bool, attention_op: Optional[Callable] = None192 ) -> None:193 # Recursively walk through all the children.194 # Any children which exposes the set_use_memory_efficient_attention_xformers method195 # gets the message196 def fn_recursive_set_mem_eff(module: torch.nn.Module):197 if hasattr(module, "set_use_memory_efficient_attention_xformers"):198 module.set_use_memory_efficient_attention_xformers(valid, attention_op)199 200 for child in module.children():201 fn_recursive_set_mem_eff(child)202 203 for module in self.children():204 if isinstance(module, torch.nn.Module):205 fn_recursive_set_mem_eff(module)206 207 def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):208 r"""209 Enable memory efficient attention as implemented in xformers.210 211 When this option is enabled, you should observe lower GPU memory usage and a potential speed up at inference212 time. Speed up at training time is not guaranteed.213 214 Warning: When Memory Efficient Attention and Sliced attention are both enabled, the Memory Efficient Attention215 is used.216 217 Parameters:218 attention_op (`Callable`, *optional*):219 Override the default `None` operator for use as `op` argument to the220 [`memory_efficient_attention()`](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.memory_efficient_attention)221 function of xFormers.222 223 Examples:224 225 ```py226 >>> import torch227 >>> from diffusers import UNet2DConditionModel228 >>> from xformers.ops import MemoryEfficientAttentionFlashAttentionOp229 230 >>> model = UNet2DConditionModel.from_pretrained(231 ... "stabilityai/stable-diffusion-2-1", subfolder="unet", torch_dtype=torch.float16232 ... )233 >>> model = model.to("cuda")234 >>> model.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp)235 ```236 """237 self.set_use_memory_efficient_attention_xformers(True, attention_op)238 239 def disable_xformers_memory_efficient_attention(self):240 r"""241 Disable memory efficient attention as implemented in xformers.242 """243 self.set_use_memory_efficient_attention_xformers(False)244 245 def save_pretrained(246 self,247 save_directory: Union[str, os.PathLike],248 is_main_process: bool = True,249 save_function: Callable = None,250 safe_serialization: bool = False,251 variant: Optional[str] = None,252 ):253 """254 Save a model and its configuration file to a directory, so that it can be re-loaded using the255 `[`~models.ModelMixin.from_pretrained`]` class method.256 257 Arguments:258 save_directory (`str` or `os.PathLike`):259 Directory to which to save. Will be created if it doesn't exist.260 is_main_process (`bool`, *optional*, defaults to `True`):261 Whether the process calling this is the main process or not. Useful when in distributed training like262 TPUs and need to call this function on all processes. In this case, set `is_main_process=True` only on263 the main process to avoid race conditions.264 save_function (`Callable`):265 The function to use to save the state dictionary. Useful on distributed training like TPUs when one266 need to replace `torch.save` by another method. Can be configured with the environment variable267 `DIFFUSERS_SAVE_MODE`.268 safe_serialization (`bool`, *optional*, defaults to `False`):269 Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).270 variant (`str`, *optional*):271 If specified, weights are saved in the format pytorch_model.<variant>.bin.272 """273 if safe_serialization and not is_safetensors_available():274 raise ImportError("`safe_serialization` requires the `safetensors library: `pip install safetensors`.")275 276 if os.path.isfile(save_directory):277 logger.error(f"Provided path ({save_directory}) should be a directory, not a file")278 return279 280 os.makedirs(save_directory, exist_ok=True)281 282 model_to_save = self283 284 # Attach architecture to the config285 # Save the config286 if is_main_process:287 model_to_save.save_config(save_directory)288 289 # Save the model290 state_dict = model_to_save.state_dict()291 292 weights_name = SAFETENSORS_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME293 weights_name = _add_variant(weights_name, variant)294 295 # Save the model296 if safe_serialization:297 safetensors.torch.save_file(298 state_dict, os.path.join(save_directory, weights_name), metadata={"format": "pt"}299 )300 else:301 torch.save(state_dict, os.path.join(save_directory, weights_name))302 303 logger.info(f"Model weights saved in {os.path.join(save_directory, weights_name)}")304 305 @classmethod306 def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):307 r"""308 Instantiate a pretrained pytorch model from a pre-trained model configuration.309 310 The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated). To train311 the model, you should first set it back in training mode with `model.train()`.312 313 The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come314 pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning315 task.316 317 The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those318 weights are discarded.319 320 Parameters:321 pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):322 Can be either:323 324 - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co.325 Valid model ids should have an organization name, like `google/ddpm-celebahq-256`.326 - A path to a *directory* containing model weights saved using [`~ModelMixin.save_config`], e.g.,327 `./my_model_directory/`.328 329 cache_dir (`Union[str, os.PathLike]`, *optional*):330 Path to a directory in which a downloaded pretrained model configuration should be cached if the331 standard cache should not be used.332 torch_dtype (`str` or `torch.dtype`, *optional*):333 Override the default `torch.dtype` and load the model under this dtype. If `"auto"` is passed the dtype334 will be automatically derived from the model's weights.335 force_download (`bool`, *optional*, defaults to `False`):336 Whether or not to force the (re-)download of the model weights and configuration files, overriding the337 cached versions if they exist.338 resume_download (`bool`, *optional*, defaults to `False`):339 Whether or not to delete incompletely received files. Will attempt to resume the download if such a340 file exists.341 proxies (`Dict[str, str]`, *optional*):342 A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',343 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.344 output_loading_info(`bool`, *optional*, defaults to `False`):345 Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.346 local_files_only(`bool`, *optional*, defaults to `False`):347 Whether or not to only look at local files (i.e., do not try to download the model).348 use_auth_token (`str` or *bool*, *optional*):349 The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated350 when running `diffusers-cli login` (stored in `~/.huggingface`).351 revision (`str`, *optional*, defaults to `"main"`):352 The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a353 git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any354 identifier allowed by git.355 from_flax (`bool`, *optional*, defaults to `False`):356 Load the model weights from a Flax checkpoint save file.357 subfolder (`str`, *optional*, defaults to `""`):358 In case the relevant files are located inside a subfolder of the model repo (either remote in359 huggingface.co or downloaded locally), you can specify the folder name here.360 361 mirror (`str`, *optional*):362 Mirror source to accelerate downloads in China. If you are from China and have an accessibility363 problem, you can set this option to resolve it. Note that we do not guarantee the timeliness or safety.364 Please refer to the mirror site for more information.365 device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):366 A map that specifies where each submodule should go. It doesn't need to be refined to each367 parameter/buffer name, once a given module name is inside, every submodule of it will be sent to the368 same device.369 370 To have Accelerate compute the most optimized `device_map` automatically, set `device_map="auto"`. For371 more information about each option see [designing a device372 map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).373 low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):374 Speed up model loading by not initializing the weights and only loading the pre-trained weights. This375 also tries to not use more than 1x model size in CPU memory (including peak memory) while loading the376 model. This is only supported when torch version >= 1.9.0. If you are using an older version of torch,377 setting this argument to `True` will raise an error.378 variant (`str`, *optional*):379 If specified load weights from `variant` filename, *e.g.* pytorch_model.<variant>.bin. `variant` is380 ignored when using `from_flax`.381 use_safetensors (`bool`, *optional* ):382 If set to `True`, the pipeline will forcibly load the models from `safetensors` weights. If set to383 `None` (the default). The pipeline will load using `safetensors` if safetensors weights are available384 *and* if `safetensors` is installed. If the to `False` the pipeline will *not* use `safetensors`.385 386 <Tip>387 388 It is required to be logged in (`huggingface-cli login`) when you want to use private or [gated389 models](https://huggingface.co/docs/hub/models-gated#gated-models).390 391 </Tip>392 393 <Tip>394 395 Activate the special ["offline-mode"](https://huggingface.co/diffusers/installation.html#offline-mode) to use396 this method in a firewalled environment.397 398 </Tip>399 400 """401 cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)402 ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)403 force_download = kwargs.pop("force_download", False)404 from_flax = kwargs.pop("from_flax", False)405 resume_download = kwargs.pop("resume_download", False)406 proxies = kwargs.pop("proxies", None)407 output_loading_info = kwargs.pop("output_loading_info", False)408 local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)409 use_auth_token = kwargs.pop("use_auth_token", None)410 revision = kwargs.pop("revision", None)411 torch_dtype = kwargs.pop("torch_dtype", None)412 subfolder = kwargs.pop("subfolder", None)413 device_map = kwargs.pop("device_map", None)414 low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)415 variant = kwargs.pop("variant", None)416 use_safetensors = kwargs.pop("use_safetensors", None)417 418 if use_safetensors and not is_safetensors_available():419 raise ValueError(420 "`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetenstors"421 )422 423 allow_pickle = False424 if use_safetensors is None:425 use_safetensors = is_safetensors_available()426 allow_pickle = True427 428 if low_cpu_mem_usage and not is_accelerate_available():429 low_cpu_mem_usage = False430 logger.warning(431 "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"432 " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"433 " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"434 " install accelerate\n```\n."435 )436 437 if device_map is not None and not is_accelerate_available():438 raise NotImplementedError(439 "Loading and dispatching requires `accelerate`. Please make sure to install accelerate or set"440 " `device_map=None`. You can install accelerate with `pip install accelerate`."441 )442 443 # Check if we can handle device_map and dispatching the weights444 if device_map is not None and not is_torch_version(">=", "1.9.0"):445 raise NotImplementedError(446 "Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set"447 " `device_map=None`."448 )449 450 if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):451 raise NotImplementedError(452 "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"453 " `low_cpu_mem_usage=False`."454 )455 456 if low_cpu_mem_usage is False and device_map is not None:457 raise ValueError(458 f"You cannot set `low_cpu_mem_usage` to `False` while using device_map={device_map} for loading and"459 " dispatching. Please make sure to set `low_cpu_mem_usage=True`."460 )461 462 # Load config if we don't provide a configuration463 config_path = pretrained_model_name_or_path464 465 user_agent = {466 "diffusers": __version__,467 "file_type": "model",468 "framework": "pytorch",469 }470 471 # load config472 config, unused_kwargs, commit_hash = cls.load_config(473 config_path,474 cache_dir=cache_dir,475 return_unused_kwargs=True,476 return_commit_hash=True,477 force_download=force_download,478 resume_download=resume_download,479 proxies=proxies,480 local_files_only=local_files_only,481 use_auth_token=use_auth_token,482 revision=revision,483 subfolder=subfolder,484 device_map=device_map,485 user_agent=user_agent,486 **kwargs,487 )488 489 # load model490 model_file = None491 if from_flax:492 model_file = _get_model_file(493 pretrained_model_name_or_path,494 weights_name=FLAX_WEIGHTS_NAME,495 cache_dir=cache_dir,496 force_download=force_download,497 resume_download=resume_download,498 proxies=proxies,499 local_files_only=local_files_only,500 use_auth_token=use_auth_token,501 revision=revision,502 subfolder=subfolder,503 user_agent=user_agent,504 commit_hash=commit_hash,505 )506 model = cls.from_config(config, **unused_kwargs)507 508 # Convert the weights509 from .modeling_pytorch_flax_utils import load_flax_checkpoint_in_pytorch_model510 511 model = load_flax_checkpoint_in_pytorch_model(model, model_file)512 else:513 if use_safetensors:514 try:515 model_file = _get_model_file(516 pretrained_model_name_or_path,517 weights_name=_add_variant(SAFETENSORS_WEIGHTS_NAME, variant),518 cache_dir=cache_dir,519 force_download=force_download,520 resume_download=resume_download,521 proxies=proxies,522 local_files_only=local_files_only,523 use_auth_token=use_auth_token,524 revision=revision,525 subfolder=subfolder,526 user_agent=user_agent,527 commit_hash=commit_hash,528 )529 except IOError as e:530 if not allow_pickle:531 raise e532 pass533 if model_file is None:534 model_file = _get_model_file(535 pretrained_model_name_or_path,536 weights_name=_add_variant(WEIGHTS_NAME, variant),537 cache_dir=cache_dir,538 force_download=force_download,539 resume_download=resume_download,540 proxies=proxies,541 local_files_only=local_files_only,542 use_auth_token=use_auth_token,543 revision=revision,544 subfolder=subfolder,545 user_agent=user_agent,546 commit_hash=commit_hash,547 )548 549 if low_cpu_mem_usage:550 # Instantiate model with empty weights551 with accelerate.init_empty_weights():552 model = cls.from_config(config, **unused_kwargs)553 554 # if device_map is None, load the state dict and move the params from meta device to the cpu555 if device_map is None:556 param_device = "cpu"557 state_dict = load_state_dict(model_file, variant=variant)558 # move the params from meta device to cpu559 missing_keys = set(model.state_dict().keys()) - set(state_dict.keys())560 if len(missing_keys) > 0:561 raise ValueError(562 f"Cannot load {cls} from {pretrained_model_name_or_path} because the following keys are"563 f" missing: \n {', '.join(missing_keys)}. \n Please make sure to pass"564 " `low_cpu_mem_usage=False` and `device_map=None` if you want to randomly initialize"565 " those weights or else make sure your checkpoint file is correct."566 )567 568 empty_state_dict = model.state_dict()569 for param_name, param in state_dict.items():570 accepts_dtype = "dtype" in set(571 inspect.signature(set_module_tensor_to_device).parameters.keys()572 )573 574 if empty_state_dict[param_name].shape != param.shape:575 raise ValueError(576 f"Cannot load {pretrained_model_name_or_path} because {param_name} expected shape {empty_state_dict[param_name]}, but got {param.shape}. If you want to instead overwrite randomly initialized weights, please make sure to pass both `low_cpu_mem_usage=False` and `ignore_mismatched_sizes=True`. For more information, see also: https://github.com/huggingface/diffusers/issues/1619#issuecomment-1345604389 as an example."577 )578 579 if accepts_dtype:580 set_module_tensor_to_device(581 model, param_name, param_device, value=param, dtype=torch_dtype582 )583 else:584 set_module_tensor_to_device(model, param_name, param_device, value=param)585 else: # else let accelerate handle loading and dispatching.586 # Load weights and dispatch according to the device_map587 # by default the device_map is None and the weights are loaded on the CPU588 accelerate.load_checkpoint_and_dispatch(model, model_file, device_map, dtype=torch_dtype)589 590 loading_info = {591 "missing_keys": [],592 "unexpected_keys": [],593 "mismatched_keys": [],594 "error_msgs": [],595 }596 else:597 model = cls.from_config(config, **unused_kwargs)598 599 state_dict = load_state_dict(model_file, variant=variant)600 601 model, missing_keys, unexpected_keys, mismatched_keys, error_msgs = cls._load_pretrained_model(602 model,603 state_dict,604 model_file,605 pretrained_model_name_or_path,606 ignore_mismatched_sizes=ignore_mismatched_sizes,607 )608 609 loading_info = {610 "missing_keys": missing_keys,611 "unexpected_keys": unexpected_keys,612 "mismatched_keys": mismatched_keys,613 "error_msgs": error_msgs,614 }615 616 if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype):617 raise ValueError(618 f"{torch_dtype} needs to be of type `torch.dtype`, e.g. `torch.float16`, but is {type(torch_dtype)}."619 )620 elif torch_dtype is not None:621 model = model.to(torch_dtype)622 623 model.register_to_config(_name_or_path=pretrained_model_name_or_path)624 625 # Set model in evaluation mode to deactivate DropOut modules by default626 model.eval()627 if output_loading_info:628 return model, loading_info629 630 return model631 632 @classmethod633 def _load_pretrained_model(634 cls,635 model,636 state_dict,637 resolved_archive_file,638 pretrained_model_name_or_path,639 ignore_mismatched_sizes=False,640 ):641 # Retrieve missing & unexpected_keys642 model_state_dict = model.state_dict()643 loaded_keys = list(state_dict.keys())644 645 expected_keys = list(model_state_dict.keys())646 647 original_loaded_keys = loaded_keys648 649 missing_keys = list(set(expected_keys) - set(loaded_keys))650 unexpected_keys = list(set(loaded_keys) - set(expected_keys))651 652 # Make sure we are able to load base models as well as derived models (with heads)653 model_to_load = model654 655 def _find_mismatched_keys(656 state_dict,657 model_state_dict,658 loaded_keys,659 ignore_mismatched_sizes,660 ):661 mismatched_keys = []662 if ignore_mismatched_sizes:663 for checkpoint_key in loaded_keys:664 model_key = checkpoint_key665 666 if (667 model_key in model_state_dict668 and state_dict[checkpoint_key].shape != model_state_dict[model_key].shape669 ):670 mismatched_keys.append(671 (checkpoint_key, state_dict[checkpoint_key].shape, model_state_dict[model_key].shape)672 )673 del state_dict[checkpoint_key]674 return mismatched_keys675 676 if state_dict is not None:677 # Whole checkpoint678 mismatched_keys = _find_mismatched_keys(679 state_dict,680 model_state_dict,681 original_loaded_keys,682 ignore_mismatched_sizes,683 )684 error_msgs = _load_state_dict_into_model(model_to_load, state_dict)685 686 if len(error_msgs) > 0:687 error_msg = "\n\t".join(error_msgs)688 if "size mismatch" in error_msg:689 error_msg += (690 "\n\tYou may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method."691 )692 raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}")693 694 if len(unexpected_keys) > 0:695 logger.warning(696 f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when"697 f" initializing {model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are"698 f" initializing {model.__class__.__name__} from the checkpoint of a model trained on another task"699 " or with another architecture (e.g. initializing a BertForSequenceClassification model from a"700 " BertForPreTraining model).\n- This IS NOT expected if you are initializing"701 f" {model.__class__.__name__} from the checkpoint of a model that you expect to be exactly"702 " identical (initializing a BertForSequenceClassification model from a"703 " BertForSequenceClassification model)."704 )705 else:706 logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")707 if len(missing_keys) > 0:708 logger.warning(709 f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"710 f" {pretrained_model_name_or_path} and are newly initialized: {missing_keys}\nYou should probably"711 " TRAIN this model on a down-stream task to be able to use it for predictions and inference."712 )713 elif len(mismatched_keys) == 0:714 logger.info(715 f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at"716 f" {pretrained_model_name_or_path}.\nIf your task is similar to the task the model of the"717 f" checkpoint was trained on, you can already use {model.__class__.__name__} for predictions"718 " without further training."719 )720 if len(mismatched_keys) > 0:721 mismatched_warning = "\n".join(722 [723 f"- {key}: found shape {shape1} in the checkpoint and {shape2} in the model instantiated"724 for key, shape1, shape2 in mismatched_keys725 ]726 )727 logger.warning(728 f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"729 f" {pretrained_model_name_or_path} and are newly initialized because the shapes did not"730 f" match:\n{mismatched_warning}\nYou should probably TRAIN this model on a down-stream task to be"731 " able to use it for predictions and inference."732 )733 734 return model, missing_keys, unexpected_keys, mismatched_keys, error_msgs735 736 @property737 def device(self) -> device:738 """739 `torch.device`: The device on which the module is (assuming that all the module parameters are on the same740 device).741 """742 return get_parameter_device(self)743 744 @property745 def dtype(self) -> torch.dtype:746 """747 `torch.dtype`: The dtype of the module (assuming that all the module parameters have the same dtype).748 """749 return get_parameter_dtype(self)750 751 def num_parameters(self, only_trainable: bool = False, exclude_embeddings: bool = False) -> int:752 """753 Get number of (optionally, trainable or non-embeddings) parameters in the module.754 755 Args:756 only_trainable (`bool`, *optional*, defaults to `False`):757 Whether or not to return only the number of trainable parameters758 759 exclude_embeddings (`bool`, *optional*, defaults to `False`):760 Whether or not to return only the number of non-embeddings parameters761 762 Returns:763 `int`: The number of parameters.764 """765 766 if exclude_embeddings:767 embedding_param_names = [768 f"{name}.weight"769 for name, module_type in self.named_modules()770 if isinstance(module_type, torch.nn.Embedding)771 ]772 non_embedding_parameters = [773 parameter for name, parameter in self.named_parameters() if name not in embedding_param_names774 ]775 return sum(p.numel() for p in non_embedding_parameters if p.requires_grad or not only_trainable)776 else:777 return sum(p.numel() for p in self.parameters() if p.requires_grad or not only_trainable)778