CoolFace
Apppublic

Aluode/PerceptionLabPortable

sourceHugging Faceupdated 9mo agoView on Hugging Face
0likes
configuration_utils.py1381 linesDownload Raw Back to transformers
1# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.2# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.3#4# Licensed under the Apache License, Version 2.0 (the "License");5# you may not use this file except in compliance with the License.6# You may obtain a copy of the License at7#8#     http://www.apache.org/licenses/LICENSE-2.09#10# Unless required by applicable law or agreed to in writing, software11# distributed under the License is distributed on an "AS IS" BASIS,12# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.13# See the License for the specific language governing permissions and14# limitations under the License.15"""Configuration base class and utilities."""16 17import copy18import json19import os20import warnings21from typing import TYPE_CHECKING, Any, Optional, TypeVar, Union22 23from packaging import version24 25from . import __version__26from .dynamic_module_utils import custom_object_save27from .modeling_gguf_pytorch_utils import load_gguf_checkpoint28from .utils import (29    CONFIG_NAME,30    PushToHubMixin,31    cached_file,32    copy_func,33    download_url,34    extract_commit_hash,35    is_remote_url,36    is_torch_available,37    logging,38)39from .utils.generic import is_timm_config_dict40 41 42if TYPE_CHECKING:43    import torch44 45 46logger = logging.get_logger(__name__)47 48 49# type hinting: specifying the type of config class that inherits from PretrainedConfig50SpecificPretrainedConfigType = TypeVar("SpecificPretrainedConfigType", bound="PretrainedConfig")51 52 53class PretrainedConfig(PushToHubMixin):54    # no-format55    r"""56    Base class for all configuration classes. Handles a few parameters common to all models' configurations as well as57    methods for loading/downloading/saving configurations.58 59    <Tip>60 61    A configuration file can be loaded and saved to disk. Loading the configuration file and using this file to62    initialize a model does **not** load the model weights. It only affects the model's configuration.63 64    </Tip>65 66    Class attributes (overridden by derived classes):67 68    - **model_type** (`str`) -- An identifier for the model type, serialized into the JSON file, and used to recreate69      the correct object in [`~transformers.AutoConfig`].70    - **has_no_defaults_at_init** (`bool`) -- Whether the config class can be initialized without providing input arguments.71      Some configurations requires inputs to be defined at init and have no default values, usually these are composite configs,72      (but not necessarily) such as [`~transformers.EncoderDecoderConfig`] or [`~RagConfig`]. They have to be initialized from73      two or more configs of type [`~transformers.PretrainedConfig`].74    - **keys_to_ignore_at_inference** (`list[str]`) -- A list of keys to ignore by default when looking at dictionary75      outputs of the model during inference.76    - **attribute_map** (`dict[str, str]`) -- A dict that maps model specific attribute names to the standardized77      naming of attributes.78    - **base_model_tp_plan** (`dict[str, Any]`) -- A dict that maps sub-modules FQNs of a base model to a tensor79      parallel plan applied to the sub-module when `model.tensor_parallel` is called.80    - **base_model_pp_plan** (`dict[str, tuple[list[str]]]`) -- A dict that maps child-modules of a base model to a81      pipeline parallel plan that enables users to place the child-module on the appropriate device.82 83    Common attributes (present in all subclasses):84 85    - **vocab_size** (`int`) -- The number of tokens in the vocabulary, which is also the first dimension of the86      embeddings matrix (this attribute may be missing for models that don't have a text modality like ViT).87    - **hidden_size** (`int`) -- The hidden size of the model.88    - **num_attention_heads** (`int`) -- The number of attention heads used in the multi-head attention layers of the89      model.90    - **num_hidden_layers** (`int`) -- The number of blocks in the model.91 92    <Tip warning={true}>93 94    Setting parameters for sequence generation in the model config is deprecated. For backward compatibility, loading95    some of them will still be possible, but attempting to overwrite them will throw an exception -- you should set96    them in a [~transformers.GenerationConfig]. Check the documentation of [~transformers.GenerationConfig] for more97    information about the individual parameters.98 99    </Tip>100 101    Arg:102        name_or_path (`str`, *optional*, defaults to `""`):103            Store the string that was passed to [`PreTrainedModel.from_pretrained`] or104            [`TFPreTrainedModel.from_pretrained`] as `pretrained_model_name_or_path` if the configuration was created105            with such a method.106        output_hidden_states (`bool`, *optional*, defaults to `False`):107            Whether or not the model should return all hidden-states.108        output_attentions (`bool`, *optional*, defaults to `False`):109            Whether or not the model should returns all attentions.110        return_dict (`bool`, *optional*, defaults to `True`):111            Whether or not the model should return a [`~transformers.utils.ModelOutput`] instead of a plain tuple.112        is_encoder_decoder (`bool`, *optional*, defaults to `False`):113            Whether the model is used as an encoder/decoder or not.114        is_decoder (`bool`, *optional*, defaults to `False`):115            Whether to only use the decoder in an encoder-decoder architecture, otherwise it has no effect on116            decoder-only or encoder-only architectures.117        cross_attention_hidden_size (`bool`, *optional*):118            The hidden size of the cross-attention layer in case the model is used as a decoder in an encoder-decoder119            setting and the cross-attention hidden dimension differs from `self.config.hidden_size`.120        add_cross_attention (`bool`, *optional*, defaults to `False`):121            Whether cross-attention layers should be added to the model. Note, this option is only relevant for models122            that can be used as decoder models within the [`EncoderDecoderModel`] class, which consists of all models123            in `AUTO_MODELS_FOR_CAUSAL_LM`.124        tie_encoder_decoder (`bool`, *optional*, defaults to `False`):125            Whether all encoder weights should be tied to their equivalent decoder weights. This requires the encoder126            and decoder model to have the exact same parameter names.127        prune_heads (`dict[int, list[int]]`, *optional*, defaults to `{}`):128            Pruned heads of the model. The keys are the selected layer indices and the associated values, the list of129            heads to prune in said layer.130 131            For instance `{1: [0, 2], 2: [2, 3]}` will prune heads 0 and 2 on layer 1 and heads 2 and 3 on layer 2.132        chunk_size_feed_forward (`int`, *optional*, defaults to `0`):133            The chunk size of all feed forward layers in the residual attention blocks. A chunk size of `0` means that134            the feed forward layer is not chunked. A chunk size of n means that the feed forward layer processes `n` <135            sequence_length embeddings at a time. For more information on feed forward chunking, see [How does Feed136            Forward Chunking work?](../glossary.html#feed-forward-chunking).137 138        > Parameters for fine-tuning tasks139 140        architectures (`list[str]`, *optional*):141            Model architectures that can be used with the model pretrained weights.142        finetuning_task (`str`, *optional*):143            Name of the task used to fine-tune the model. This can be used when converting from an original (TensorFlow144            or PyTorch) checkpoint.145        id2label (`dict[int, str]`, *optional*):146            A map from index (for instance prediction index, or target index) to label.147        label2id (`dict[str, int]`, *optional*):148            A map from label to index for the model.149        num_labels (`int`, *optional*):150            Number of labels to use in the last layer added to the model, typically for a classification task.151        task_specific_params (`dict[str, Any]`, *optional*):152            Additional keyword arguments to store for the current task.153        problem_type (`str`, *optional*):154            Problem type for `XxxForSequenceClassification` models. Can be one of `"regression"`,155            `"single_label_classification"` or `"multi_label_classification"`.156 157        > Parameters linked to the tokenizer158 159        tokenizer_class (`str`, *optional*):160            The name of the associated tokenizer class to use (if none is set, will use the tokenizer associated to the161            model by default).162        prefix (`str`, *optional*):163            A specific prompt that should be added at the beginning of each text before calling the model.164        bos_token_id (`int`, *optional*):165            The id of the _beginning-of-stream_ token.166        pad_token_id (`int`, *optional*):167            The id of the _padding_ token.168        eos_token_id (`int`, *optional*):169            The id of the _end-of-stream_ token.170        decoder_start_token_id (`int`, *optional*):171            If an encoder-decoder model starts decoding with a different token than _bos_, the id of that token.172        sep_token_id (`int`, *optional*):173            The id of the _separation_ token.174 175        > PyTorch specific parameters176 177        torchscript (`bool`, *optional*, defaults to `False`):178            Whether or not the model should be used with Torchscript.179        tie_word_embeddings (`bool`, *optional*, defaults to `True`):180            Whether the model's input and output word embeddings should be tied. Note that this is only relevant if the181            model has a output word embedding layer.182        dtype (`str`, *optional*):183            The `dtype` of the weights. This attribute can be used to initialize the model to a non-default `dtype`184            (which is normally `float32`) and thus allow for optimal storage allocation. For example, if the saved185            model is `float16`, ideally we want to load it back using the minimal amount of memory needed to load186            `float16` weights.187    """188 189    model_type: str = ""190    base_config_key: str = ""191    sub_configs: dict[str, type["PretrainedConfig"]] = {}192    has_no_defaults_at_init: bool = False193    attribute_map: dict[str, str] = {}194    base_model_tp_plan: Optional[dict[str, Any]] = None195    base_model_pp_plan: Optional[dict[str, tuple[list[str]]]] = None196    base_model_ep_plan: Optional[dict[str, tuple[list[str]]]] = None197    _auto_class: Optional[str] = None198 199    def __setattr__(self, key, value):200        if key in super().__getattribute__("attribute_map"):201            key = super().__getattribute__("attribute_map")[key]202        super().__setattr__(key, value)203 204    def __getattribute__(self, key):205        if key != "attribute_map" and key in super().__getattribute__("attribute_map"):206            key = super().__getattribute__("attribute_map")[key]207        return super().__getattribute__(key)208 209    def __init__(210        self,211        *,212        # All models common arguments213        output_hidden_states: bool = False,214        output_attentions: bool = False,215        return_dict: bool = True,216        torchscript: bool = False,217        dtype: Optional[Union[str, "torch.dtype"]] = None,218        # Common arguments219        pruned_heads: Optional[dict[int, list[int]]] = None,220        tie_word_embeddings: bool = True,221        chunk_size_feed_forward: int = 0,222        is_encoder_decoder: bool = False,223        is_decoder: bool = False,224        cross_attention_hidden_size: Optional[int] = None,225        add_cross_attention: bool = False,226        tie_encoder_decoder: bool = False,227        # Fine-tuning task arguments228        architectures: Optional[list[str]] = None,229        finetuning_task: Optional[str] = None,230        id2label: Optional[dict[int, str]] = None,231        label2id: Optional[dict[str, int]] = None,232        num_labels: Optional[int] = None,233        task_specific_params: Optional[dict[str, Any]] = None,234        problem_type: Optional[str] = None,235        # Tokenizer kwargs236        tokenizer_class: Optional[str] = None,237        prefix: Optional[str] = None,238        bos_token_id: Optional[int] = None,239        pad_token_id: Optional[int] = None,240        eos_token_id: Optional[int] = None,241        sep_token_id: Optional[int] = None,242        decoder_start_token_id: Optional[int] = None,243        **kwargs,244    ):245        # Validation for some arguments246        if label2id is not None and not isinstance(label2id, dict):247            raise ValueError("Argument label2id should be a dictionary.")248        if id2label is not None and not isinstance(id2label, dict):249            raise ValueError("Argument id2label should be a dictionary.")250        if num_labels is not None and id2label is not None and len(id2label) != num_labels:251            logger.warning(252                f"You passed `num_labels={num_labels}` which is incompatible to "253                f"the `id2label` map of length `{len(id2label)}`."254            )255        if problem_type is not None and problem_type not in (256            "regression",257            "single_label_classification",258            "multi_label_classification",259        ):260            raise ValueError(261                f"The config parameter `problem_type` was not understood: received {problem_type} "262                "but only 'regression', 'single_label_classification' and 'multi_label_classification' are valid."263            )264        # BC for the `torch_dtype` argument instead of the simpler `dtype`265        # Do not warn, as it would otherwise always be triggered since most configs on the hub have `torch_dtype`266        if (torch_dtype := kwargs.pop("torch_dtype", None)) is not None:267            # If both are provided, keep `dtype`268            dtype = dtype if dtype is not None else torch_dtype269        if dtype is not None and isinstance(dtype, str) and is_torch_available():270            # we will start using self.dtype in v5, but to be consistent with271            # from_pretrained's dtype arg convert it to an actual torch.dtype object272            import torch273 274            dtype = getattr(torch, dtype)275 276        # Attributes common for all models277        self.return_dict = return_dict278        self.output_hidden_states = output_hidden_states279        self.torchscript = torchscript280        self.dtype = dtype281        self._output_attentions = output_attentions  # has public property282 283        # Less common kwargs, only used by some models284        self.pruned_heads = pruned_heads if pruned_heads is not None else {}285        self.tie_word_embeddings = tie_word_embeddings286        self.chunk_size_feed_forward = chunk_size_feed_forward287 288        # Encoder-decoder models attributes289        self.is_encoder_decoder = is_encoder_decoder290        self.is_decoder = is_decoder  # used in encoder-decoder models to differentiate encoder from decoder291        self.cross_attention_hidden_size = cross_attention_hidden_size292        self.add_cross_attention = add_cross_attention293        self.tie_encoder_decoder = tie_encoder_decoder294 295        # Fine-tuning task attributes296        self.architectures = architectures297        self.finetuning_task = finetuning_task298        self.id2label = id2label299        self.label2id = label2id300        self.task_specific_params = task_specific_params301        self.problem_type = problem_type302 303        if self.id2label is None:304            self._create_id_label_maps(num_labels if num_labels is not None else 2)305        else:306            # Keys are always strings in JSON so convert ids to int here.307            self.id2label = {int(key): value for key, value in self.id2label.items()}308 309        # Tokenizer attributes310        self.tokenizer_class = tokenizer_class311        self.prefix = prefix312        self.bos_token_id = bos_token_id313        self.pad_token_id = pad_token_id314        self.eos_token_id = eos_token_id315        self.sep_token_id = sep_token_id316        self.decoder_start_token_id = decoder_start_token_id317 318        # Retrocompatibility: Parameters for sequence generation. While we will keep the ability to load these319        # parameters, saving them will be deprecated. In a distant future, we won't need to load them.320        for parameter_name, default_value in self._get_global_generation_defaults().items():321            setattr(self, parameter_name, kwargs.pop(parameter_name, default_value))322 323        # Name or path to the pretrained checkpoint324        self._name_or_path = str(kwargs.pop("name_or_path", ""))325        self._commit_hash = kwargs.pop("_commit_hash", None)326 327        # Attention implementation to use, if relevant (it sets it recursively on sub-configs)328        self._attn_implementation = kwargs.pop("attn_implementation", None)329 330        # Drop the transformers version info331        self.transformers_version = kwargs.pop("transformers_version", None)332 333        # Deal with gradient checkpointing334        if kwargs.get("gradient_checkpointing", False):335            warnings.warn(336                "Passing `gradient_checkpointing` to a config initialization is deprecated and will be removed in v5 "337                "Transformers. Using `model.gradient_checkpointing_enable()` instead, or if you are using the "338                "`Trainer` API, pass `gradient_checkpointing=True` in your `TrainingArguments`."339            )340 341        # Additional attributes without default values342        for key, value in kwargs.items():343            try:344                setattr(self, key, value)345            except AttributeError as err:346                logger.error(f"Can't set {key} with value {value} for {self}")347                raise err348 349        # TODO: remove later, deprecated arguments for TF models350        self.tf_legacy_loss = kwargs.pop("tf_legacy_loss", False)351        self.use_bfloat16 = kwargs.pop("use_bfloat16", False)352 353    def _create_id_label_maps(self, num_labels: int):354        self.id2label = {i: f"LABEL_{i}" for i in range(num_labels)}355        self.label2id = dict(zip(self.id2label.values(), self.id2label.keys()))356 357    @property358    def name_or_path(self) -> Optional[str]:359        return getattr(self, "_name_or_path", None)360 361    @name_or_path.setter362    def name_or_path(self, value):363        self._name_or_path = str(value)  # Make sure that name_or_path is a string (for JSON encoding)364 365    @property366    def output_attentions(self):367        """368        `bool`: Whether or not the model should returns all attentions.369        """370        return self._output_attentions371 372    @output_attentions.setter373    def output_attentions(self, value: bool):374        # If we set `output_attentions` explicitly before the attn implementation, dispatch eager375        if value and self._attn_implementation is None:376            self._attn_implementation = "eager"377        if value and self._attn_implementation != "eager":378            raise ValueError(379                "The `output_attentions` attribute is not supported when using the `attn_implementation` set to "380                f"{self._attn_implementation}. Please set it to 'eager' instead."381            )382        self._output_attentions = value383 384    @property385    def use_return_dict(self) -> bool:386        """387        `bool`: Whether or not return [`~utils.ModelOutput`] instead of tuples.388        """389        # If torchscript is set, force `return_dict=False` to avoid jit errors390        return self.return_dict and not self.torchscript391 392    @property393    def num_labels(self) -> int:394        """395        `int`: The number of labels for classification models.396        """397        return len(self.id2label)398 399    @num_labels.setter400    def num_labels(self, num_labels: int):401        # we do not store `num_labels` attribute in config, but instead402        # compute it based on the length of the `id2label` map403        if self.id2label is None or self.num_labels != num_labels:404            self._create_id_label_maps(num_labels)405 406    @property407    def _attn_implementation(self):408        return self._attn_implementation_internal409 410    @_attn_implementation.setter411    def _attn_implementation(self, value: Optional[Union[str, dict]]):412        """We set it recursively on the sub-configs as well"""413        # Set if for current config414        current_attn = getattr(self, "_attn_implementation", None)415        attn_implementation = value if not isinstance(value, dict) else value.get("", current_attn)416        self._attn_implementation_internal = attn_implementation417 418        # Set it recursively on the subconfigs419        for subconfig_key in self.sub_configs:420            subconfig = getattr(self, subconfig_key, None)421            if subconfig is not None:422                current_subconfig_attn = getattr(subconfig, "_attn_implementation", None)423                sub_implementation = (424                    value if not isinstance(value, dict) else value.get(subconfig_key, current_subconfig_attn)425                )426                subconfig._attn_implementation = sub_implementation427 428    @property429    def torch_dtype(self):430        logger.warning_once("`torch_dtype` is deprecated! Use `dtype` instead!")431        return self.dtype432 433    @torch_dtype.setter434    def torch_dtype(self, value):435        logger.warning_once("`torch_dtype` is deprecated! Use `dtype` instead!")436        self.dtype = value437 438    def save_pretrained(self, save_directory: Union[str, os.PathLike], push_to_hub: bool = False, **kwargs):439        """440        Save a configuration object to the directory `save_directory`, so that it can be re-loaded using the441        [`~PretrainedConfig.from_pretrained`] class method.442 443        Args:444            save_directory (`str` or `os.PathLike`):445                Directory where the configuration JSON file will be saved (will be created if it does not exist).446            push_to_hub (`bool`, *optional*, defaults to `False`):447                Whether or not to push your model to the Hugging Face model hub after saving it. You can specify the448                repository you want to push to with `repo_id` (will default to the name of `save_directory` in your449                namespace).450            kwargs (`dict[str, Any]`, *optional*):451                Additional key word arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.452        """453        self._set_token_in_kwargs(kwargs)454 455        if os.path.isfile(save_directory):456            raise AssertionError(f"Provided path ({save_directory}) should be a directory, not a file")457 458        non_default_generation_parameters = self._get_non_default_generation_parameters()459        if len(non_default_generation_parameters) > 0:460            # TODO (joao): this should be an exception if the user has modified the loaded config. See #33886461            warnings.warn(462                "Some non-default generation parameters are set in the model config. These should go into either a) "463                "`model.generation_config` (as opposed to `model.config`); OR b) a GenerationConfig file "464                "(https://huggingface.co/docs/transformers/generation_strategies#save-a-custom-decoding-strategy-with-your-model)."465                "This warning will become an exception in the future."466                f"\nNon-default generation parameters: {str(non_default_generation_parameters)}",467                UserWarning,468            )469 470        os.makedirs(save_directory, exist_ok=True)471 472        if push_to_hub:473            commit_message = kwargs.pop("commit_message", None)474            repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])475            repo_id = self._create_repo(repo_id, **kwargs)476            files_timestamps = self._get_files_timestamps(save_directory)477 478        # This attribute is important to know on load, but should not be serialized on save.479        if "transformers_weights" in self:480            delattr(self, "transformers_weights")481 482        # If we have a custom config, we copy the file defining it in the folder and set the attributes so it can be483        # loaded from the Hub.484        if self._auto_class is not None:485            custom_object_save(self, save_directory, config=self)486 487        # If we save using the predefined names, we can load using `from_pretrained`488        output_config_file = os.path.join(save_directory, CONFIG_NAME)489 490        self.to_json_file(output_config_file, use_diff=True)491        logger.info(f"Configuration saved in {output_config_file}")492 493        if push_to_hub:494            self._upload_modified_files(495                save_directory,496                repo_id,497                files_timestamps,498                commit_message=commit_message,499                token=kwargs.get("token"),500            )501 502    @staticmethod503    def _set_token_in_kwargs(kwargs, token=None):504        """Temporary method to deal with `token` and `use_auth_token`.505 506        This method is to avoid apply the same changes in all model config classes that overwrite `from_pretrained`.507 508        Need to clean up `use_auth_token` in a follow PR.509        """510        # Some model config classes like CLIP define their own `from_pretrained` without the new argument `token` yet.511        if token is None:512            token = kwargs.pop("token", None)513        use_auth_token = kwargs.pop("use_auth_token", None)514 515        if use_auth_token is not None:516            warnings.warn(517                "The `use_auth_token` argument is deprecated and will be removed in v5 of Transformers. Please use `token` instead.",518                FutureWarning,519            )520            if token is not None:521                raise ValueError(522                    "`token` and `use_auth_token` are both specified. Please set only the argument `token`."523                )524            token = use_auth_token525 526        if token is not None:527            kwargs["token"] = token528 529    @classmethod530    def from_pretrained(531        cls: type[SpecificPretrainedConfigType],532        pretrained_model_name_or_path: Union[str, os.PathLike],533        cache_dir: Optional[Union[str, os.PathLike]] = None,534        force_download: bool = False,535        local_files_only: bool = False,536        token: Optional[Union[str, bool]] = None,537        revision: str = "main",538        **kwargs,539    ) -> SpecificPretrainedConfigType:540        r"""541        Instantiate a [`PretrainedConfig`] (or a derived class) from a pretrained model configuration.542 543        Args:544            pretrained_model_name_or_path (`str` or `os.PathLike`):545                This can be either:546 547                - a string, the *model id* of a pretrained model configuration hosted inside a model repo on548                  huggingface.co.549                - a path to a *directory* containing a configuration file saved using the550                  [`~PretrainedConfig.save_pretrained`] method, e.g., `./my_model_directory/`.551                - a path or url to a saved configuration JSON *file*, e.g., `./my_model_directory/configuration.json`.552            cache_dir (`str` or `os.PathLike`, *optional*):553                Path to a directory in which a downloaded pretrained model configuration should be cached if the554                standard cache should not be used.555            force_download (`bool`, *optional*, defaults to `False`):556                Whether or not to force to (re-)download the configuration files and override the cached versions if557                they exist.558            resume_download:559                Deprecated and ignored. All downloads are now resumed by default when possible.560                Will be removed in v5 of Transformers.561            proxies (`dict[str, str]`, *optional*):562                A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',563                'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.564            token (`str` or `bool`, *optional*):565                The token to use as HTTP bearer authorization for remote files. If `True`, or not specified, will use566                the token generated when running `hf auth login` (stored in `~/.huggingface`).567            revision (`str`, *optional*, defaults to `"main"`):568                The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a569                git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any570                identifier allowed by git.571 572                <Tip>573 574                To test a pull request you made on the Hub, you can pass `revision="refs/pr/<pr_number>"`.575 576                </Tip>577 578            return_unused_kwargs (`bool`, *optional*, defaults to `False`):579                If `False`, then this function returns just the final configuration object.580 581                If `True`, then this functions returns a `Tuple(config, unused_kwargs)` where *unused_kwargs* is a582                dictionary consisting of the key/value pairs whose keys are not configuration attributes: i.e., the583                part of `kwargs` which has not been used to update `config` and is otherwise ignored.584            subfolder (`str`, *optional*, defaults to `""`):585                In case the relevant files are located inside a subfolder of the model repo on huggingface.co, you can586                specify the folder name here.587            kwargs (`dict[str, Any]`, *optional*):588                The values in kwargs of any keys which are configuration attributes will be used to override the loaded589                values. Behavior concerning key/value pairs whose keys are *not* configuration attributes is controlled590                by the `return_unused_kwargs` keyword parameter.591 592        Returns:593            [`PretrainedConfig`]: The configuration object instantiated from this pretrained model.594 595        Examples:596 597        ```python598        # We can't instantiate directly the base class *PretrainedConfig* so let's show the examples on a599        # derived class: BertConfig600        config = BertConfig.from_pretrained(601            "google-bert/bert-base-uncased"602        )  # Download configuration from huggingface.co and cache.603        config = BertConfig.from_pretrained(604            "./test/saved_model/"605        )  # E.g. config (or model) was saved using *save_pretrained('./test/saved_model/')*606        config = BertConfig.from_pretrained("./test/saved_model/my_configuration.json")607        config = BertConfig.from_pretrained("google-bert/bert-base-uncased", output_attentions=True, foo=False)608        assert config.output_attentions == True609        config, unused_kwargs = BertConfig.from_pretrained(610            "google-bert/bert-base-uncased", output_attentions=True, foo=False, return_unused_kwargs=True611        )612        assert config.output_attentions == True613        assert unused_kwargs == {"foo": False}614        ```"""615        kwargs["cache_dir"] = cache_dir616        kwargs["force_download"] = force_download617        kwargs["local_files_only"] = local_files_only618        kwargs["revision"] = revision619 620        cls._set_token_in_kwargs(kwargs, token)621 622        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)623        if cls.base_config_key and cls.base_config_key in config_dict:624            config_dict = config_dict[cls.base_config_key]625 626        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:627            # sometimes the config has no `base_config_key` if the config is used in several composite models628            # e.g. LlamaConfig. In that case we try to see if there is match in `model_type` before raising a warning629            for v in config_dict.values():630                if isinstance(v, dict) and v.get("model_type") == cls.model_type:631                    config_dict = v632 633            # raise warning only if we still can't see a match in `model_type`634            if config_dict["model_type"] != cls.model_type:635                logger.warning(636                    f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "637                    f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."638                )639 640        return cls.from_dict(config_dict, **kwargs)641 642    @classmethod643    def get_config_dict(644        cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs645    ) -> tuple[dict[str, Any], dict[str, Any]]:646        """647        From a `pretrained_model_name_or_path`, resolve to a dictionary of parameters, to be used for instantiating a648        [`PretrainedConfig`] using `from_dict`.649 650        Parameters:651            pretrained_model_name_or_path (`str` or `os.PathLike`):652                The identifier of the pre-trained checkpoint from which we want the dictionary of parameters.653 654        Returns:655            `tuple[Dict, Dict]`: The dictionary(ies) that will be used to instantiate the configuration object.656 657        """658        cls._set_token_in_kwargs(kwargs)659 660        original_kwargs = copy.deepcopy(kwargs)661        # Get config dict associated with the base config file662        config_dict, kwargs = cls._get_config_dict(pretrained_model_name_or_path, **kwargs)663        if config_dict is None:664            return {}, kwargs665        if "_commit_hash" in config_dict:666            original_kwargs["_commit_hash"] = config_dict["_commit_hash"]667 668        # That config file may point us toward another config file to use.669        if "configuration_files" in config_dict:670            configuration_file = get_configuration_file(config_dict["configuration_files"])671            config_dict, kwargs = cls._get_config_dict(672                pretrained_model_name_or_path, _configuration_file=configuration_file, **original_kwargs673            )674 675        return config_dict, kwargs676 677    @classmethod678    def _get_config_dict(679        cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs680    ) -> tuple[dict[str, Any], dict[str, Any]]:681        cache_dir = kwargs.pop("cache_dir", None)682        force_download = kwargs.pop("force_download", False)683        resume_download = kwargs.pop("resume_download", None)684        proxies = kwargs.pop("proxies", None)685        token = kwargs.pop("token", None)686        local_files_only = kwargs.pop("local_files_only", False)687        revision = kwargs.pop("revision", None)688        trust_remote_code = kwargs.pop("trust_remote_code", None)689        subfolder = kwargs.pop("subfolder", "")690        from_pipeline = kwargs.pop("_from_pipeline", None)691        from_auto_class = kwargs.pop("_from_auto", False)692        commit_hash = kwargs.pop("_commit_hash", None)693 694        gguf_file = kwargs.get("gguf_file")695 696        if trust_remote_code is True:697            logger.warning(698                "The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is"699                " ignored."700            )701 702        user_agent = {"file_type": "config", "from_auto_class": from_auto_class}703        if from_pipeline is not None:704            user_agent["using_pipeline"] = from_pipeline705 706        pretrained_model_name_or_path = str(pretrained_model_name_or_path)707 708        is_local = os.path.isdir(pretrained_model_name_or_path)709        if os.path.isfile(os.path.join(subfolder, pretrained_model_name_or_path)):710            # Special case when pretrained_model_name_or_path is a local file711            resolved_config_file = pretrained_model_name_or_path712            is_local = True713        elif is_remote_url(pretrained_model_name_or_path):714            configuration_file = pretrained_model_name_or_path if gguf_file is None else gguf_file715            resolved_config_file = download_url(pretrained_model_name_or_path)716        else:717            configuration_file = kwargs.pop("_configuration_file", CONFIG_NAME) if gguf_file is None else gguf_file718 719            try:720                # Load from local folder or from cache or download from model Hub and cache721                resolved_config_file = cached_file(722                    pretrained_model_name_or_path,723                    configuration_file,724                    cache_dir=cache_dir,725                    force_download=force_download,726                    proxies=proxies,727                    resume_download=resume_download,728                    local_files_only=local_files_only,729                    token=token,730                    user_agent=user_agent,731                    revision=revision,732                    subfolder=subfolder,733                    _commit_hash=commit_hash,734                )735                if resolved_config_file is None:736                    return None, kwargs737                commit_hash = extract_commit_hash(resolved_config_file, commit_hash)738            except OSError:739                # Raise any environment error raise by `cached_file`. It will have a helpful error message adapted to740                # the original exception.741                raise742            except Exception:743                # For any other exception, we throw a generic error.744                raise OSError(745                    f"Can't load the configuration of '{pretrained_model_name_or_path}'. If you were trying to load it"746                    " from 'https://huggingface.co/models', make sure you don't have a local directory with the same"747                    f" name. Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a directory"748                    f" containing a {configuration_file} file"749                )750 751        try:752            if gguf_file:753                config_dict = load_gguf_checkpoint(resolved_config_file, return_tensors=False)["config"]754            else:755                # Load config dict756                config_dict = cls._dict_from_json_file(resolved_config_file)757 758            config_dict["_commit_hash"] = commit_hash759        except (json.JSONDecodeError, UnicodeDecodeError):760            raise OSError(f"It looks like the config file at '{resolved_config_file}' is not a valid JSON file.")761 762        if is_local:763            logger.info(f"loading configuration file {resolved_config_file}")764        else:765            logger.info(f"loading configuration file {configuration_file} from cache at {resolved_config_file}")766 767        # timm models are not saved with the model_type in the config file768        if "model_type" not in config_dict and is_timm_config_dict(config_dict):769            config_dict["model_type"] = "timm_wrapper"770 771        return config_dict, kwargs772 773    @classmethod774    def from_dict(775        cls: type[SpecificPretrainedConfigType], config_dict: dict[str, Any], **kwargs776    ) -> SpecificPretrainedConfigType:777        """778        Instantiates a [`PretrainedConfig`] from a Python dictionary of parameters.779 780        Args:781            config_dict (`dict[str, Any]`):782                Dictionary that will be used to instantiate the configuration object. Such a dictionary can be783                retrieved from a pretrained checkpoint by leveraging the [`~PretrainedConfig.get_config_dict`] method.784            kwargs (`dict[str, Any]`):785                Additional parameters from which to initialize the configuration object.786 787        Returns:788            [`PretrainedConfig`]: The configuration object instantiated from those parameters.789        """790        return_unused_kwargs = kwargs.pop("return_unused_kwargs", False)791        # Those arguments may be passed along for our internal telemetry.792        # We remove them so they don't appear in `return_unused_kwargs`.793        kwargs.pop("_from_auto", None)794        kwargs.pop("_from_pipeline", None)795        # The commit hash might have been updated in the `config_dict`, we don't want the kwargs to erase that update.796        if "_commit_hash" in kwargs and "_commit_hash" in config_dict:797            kwargs["_commit_hash"] = config_dict["_commit_hash"]798 799        # For BC on the old `torch_dtype`800        if (torch_dtype := kwargs.pop("torch_dtype", None)) is not None:801            logger.warning_once("`torch_dtype` is deprecated! Use `dtype` instead!")802            # If both are present, use `dtype`803            kwargs["dtype"] = kwargs.get("dtype", torch_dtype)804 805        # We remove it from kwargs so that it does not appear in `return_unused_kwargs`.806        config_dict["attn_implementation"] = kwargs.pop("attn_implementation", None)807 808        config = cls(**config_dict)809 810        if hasattr(config, "pruned_heads"):811            config.pruned_heads = {int(key): value for key, value in config.pruned_heads.items()}812 813        # Update config with kwargs if needed814        if "num_labels" in kwargs and "id2label" in kwargs:815            num_labels = kwargs["num_labels"]816            id2label = kwargs["id2label"] if kwargs["id2label"] is not None else []817            if len(id2label) != num_labels:818                raise ValueError(819                    f"You passed along `num_labels={num_labels}` with an incompatible id to label map: "820                    f"{kwargs['id2label']}. Since those arguments are inconsistent with each other, you should remove "821                    "one of them."822                )823        to_remove = []824        for key, value in kwargs.items():825            if hasattr(config, key):826                current_attr = getattr(config, key)827                # To authorize passing a custom subconfig as kwarg in models that have nested configs.828                # We need to update only custom kwarg values instead and keep other attributes in subconfig.829                if isinstance(current_attr, PretrainedConfig) and isinstance(value, dict):830                    current_attr_updated = current_attr.to_dict()831                    current_attr_updated.update(value)832                    value = current_attr.__class__(**current_attr_updated)833                setattr(config, key, value)834                if key != "dtype":835                    to_remove.append(key)836        for key in to_remove:837            kwargs.pop(key, None)838 839        logger.info(f"Model config {config}")840        if return_unused_kwargs:841            return config, kwargs842        else:843            return config844 845    @classmethod846    def from_json_file(847        cls: type[SpecificPretrainedConfigType], json_file: Union[str, os.PathLike]848    ) -> SpecificPretrainedConfigType:849        """850        Instantiates a [`PretrainedConfig`] from the path to a JSON file of parameters.851 852        Args:853            json_file (`str` or `os.PathLike`):854                Path to the JSON file containing the parameters.855 856        Returns:857            [`PretrainedConfig`]: The configuration object instantiated from that JSON file.858 859        """860        config_dict = cls._dict_from_json_file(json_file)861        return cls(**config_dict)862 863    @classmethod864    def _dict_from_json_file(cls, json_file: Union[str, os.PathLike]):865        with open(json_file, encoding="utf-8") as reader:866            text = reader.read()867        return json.loads(text)868 869    def __eq__(self, other):870        return isinstance(other, PretrainedConfig) and (self.__dict__ == other.__dict__)871 872    def __repr__(self):873        return f"{self.__class__.__name__} {self.to_json_string()}"874 875    def __iter__(self):876        yield from self.__dict__877 878    def to_diff_dict(self) -> dict[str, Any]:879        """880        Removes all attributes from the configuration that correspond to the default config attributes for881        better readability, while always retaining the `config` attribute from the class. Serializes to a882        Python dictionary.883 884        Returns:885            dict[str, Any]: Dictionary of all the attributes that make up this configuration instance.886        """887        config_dict = self.to_dict()888 889        # Get the default config dict (from a fresh PreTrainedConfig instance)890        default_config_dict = PretrainedConfig().to_dict()891 892        # get class specific config dict893        class_config_dict = self.__class__().to_dict() if not self.has_no_defaults_at_init else {}894 895        serializable_config_dict = {}896 897        # Only serialize values that differ from the default config,898        # except always keep the 'config' attribute.899        for key, value in config_dict.items():900            if (901                isinstance(getattr(self, key, None), PretrainedConfig)902                and key in class_config_dict903                and isinstance(class_config_dict[key], dict)904                or key in self.sub_configs905            ):906                # For nested configs we need to clean the diff recursively907                diff = recursive_diff_dict(value, default_config_dict, config_obj=getattr(self, key, None))908                if "model_type" in value:909                    # Needs to be set even if it's not in the diff910                    diff["model_type"] = value["model_type"]911 912                serializable_config_dict[key] = diff913            elif (914                key not in default_config_dict915                or key == "transformers_version"916                or key == "vocab_file"917                or value != default_config_dict[key]918                or (key in default_config_dict and value != class_config_dict.get(key, value))919            ):920                serializable_config_dict[key] = value921 922        self._remove_keys_not_serialized(serializable_config_dict)923 924        # Key removed only in diff dict925        if "_name_or_path" in serializable_config_dict:926            del serializable_config_dict["_name_or_path"]927 928        if hasattr(self, "quantization_config"):929            serializable_config_dict["quantization_config"] = (930                self.quantization_config.to_dict()931                if not isinstance(self.quantization_config, dict)932                else self.quantization_config933            )934        self.dict_dtype_to_str(serializable_config_dict)935 936        return serializable_config_dict937 938    def to_dict(self) -> dict[str, Any]:939        """940        Serializes this instance to a Python dictionary.941 942        Returns:943            `dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance.944        """945        output = copy.deepcopy(self.__dict__)946        if hasattr(self.__class__, "model_type"):947            output["model_type"] = self.__class__.model_type948 949        # Transformers version when serializing the model950        output["transformers_version"] = __version__951 952        for key, value in output.items():953            # Deal with nested configs like CLIP954            if isinstance(value, PretrainedConfig):955                value = value.to_dict()956                del value["transformers_version"]957 958            output[key] = value959 960        self._remove_keys_not_serialized(output)961 962        if hasattr(self, "quantization_config"):963            output["quantization_config"] = (964                self.quantization_config.to_dict()965                if not isinstance(self.quantization_config, dict)966                else self.quantization_config967            )968        self.dict_dtype_to_str(output)969 970        return output971 972    def to_json_string(self, use_diff: bool = True) -> str:973        """974        Serializes this instance to a JSON string.975 976        Args:977            use_diff (`bool`, *optional*, defaults to `True`):978                If set to `True`, only the difference between the config instance and the default `PretrainedConfig()`979                is serialized to JSON string.980 981        Returns:982            `str`: String containing all the attributes that make up this configuration instance in JSON format.983        """984        if use_diff is True:985            config_dict = self.to_diff_dict()986        else:987            config_dict = self.to_dict()988        return json.dumps(config_dict, indent=2, sort_keys=True) + "\n"989 990    def to_json_file(self, json_file_path: Union[str, os.PathLike], use_diff: bool = True):991        """992        Save this instance to a JSON file.993 994        Args:995            json_file_path (`str` or `os.PathLike`):996                Path to the JSON file in which this configuration instance's parameters will be saved.997            use_diff (`bool`, *optional*, defaults to `True`):998                If set to `True`, only the difference between the config instance and the default `PretrainedConfig()`999                is serialized to JSON file.1000        """1001        with open(json_file_path, "w", encoding="utf-8") as writer:1002            writer.write(self.to_json_string(use_diff=use_diff))1003 1004    def update(self, config_dict: dict[str, Any]):1005        """1006        Updates attributes of this class with attributes from `config_dict`.1007 1008        Args:1009            config_dict (`dict[str, Any]`): Dictionary of attributes that should be updated for this class.1010        """1011        for key, value in config_dict.items():1012            setattr(self, key, value)1013 1014    def update_from_string(self, update_str: str):1015        """1016        Updates attributes of this class with attributes from `update_str`.1017 1018        The expected format is ints, floats and strings as is, and for booleans use `true` or `false`. For example:1019        "n_embd=10,resid_pdrop=0.2,scale_attn_weights=false,summary_type=cls_index"1020 1021        The keys to change have to already exist in the config object.1022 1023        Args:1024            update_str (`str`): String with attributes that should be updated for this class.1025 1026        """1027 1028        d = dict(x.split("=") for x in update_str.split(","))1029        for k, v in d.items():1030            if not hasattr(self, k):1031                raise ValueError(f"key {k} isn't in the original config dict")1032 1033            old_v = getattr(self, k)1034            if isinstance(old_v, bool):1035                if v.lower() in ["true", "1", "y", "yes"]:1036                    v = True1037                elif v.lower() in ["false", "0", "n", "no"]:1038                    v = False1039                else:1040                    raise ValueError(f"can't derive true or false from {v} (key {k})")1041            elif isinstance(old_v, int):1042                v = int(v)1043            elif isinstance(old_v, float):1044                v = float(v)1045            elif not isinstance(old_v, str):1046                raise TypeError(1047                    f"You can only update int, float, bool or string values in the config, got {v} for key {k}"1048                )1049 1050            setattr(self, k, v)1051 1052    def dict_dtype_to_str(self, d: dict[str, Any]) -> None:1053        """1054        Checks whether the passed dictionary and its nested dicts have a *dtype* key and if it's not None,1055        converts torch.dtype to a string of just the type. For example, `torch.float32` get converted into *"float32"*1056        string, which can then be stored in the json format.1057        """1058        if d.get("dtype") is not None:1059            if isinstance(d["dtype"], dict):1060                d["dtype"] = {k: str(v).split(".")[-1] for k, v in d["dtype"].items()}1061            # models like Emu3 can have "dtype" as token in config's vocabulary map,1062            # so we also exclude int type here to avoid error in this special case.1063            elif not isinstance(d["dtype"], (str, int)):1064                d["dtype"] = str(d["dtype"]).split(".")[1]1065        for value in d.values():1066            if isinstance(value, dict):1067                self.dict_dtype_to_str(value)1068 1069    def _remove_keys_not_serialized(self, d: dict[str, Any]) -> None:1070        """1071        Checks and removes if there are any keys in the dict that should not be serialized when saving the config.1072        Runs recursive check on the dict, to remove from all sub configs.1073        """1074        if hasattr(self, "quantization_config"):1075            # Pop the `_pre_quantization_dtype` as torch.dtypes are not serializable.1076            _ = d.pop("_pre_quantization_dtype", None)1077 1078        if "_auto_class" in d:1079            del d["_auto_class"]1080        if "_output_attentions" in d:1081            d["output_attentions"] = d.pop("_output_attentions")1082        if "_commit_hash" in d:1083            del d["_commit_hash"]1084        if "_attn_implementation_internal" in d:1085            del d["_attn_implementation_internal"]1086        # Do not serialize `base_model_tp_plan` for now1087        if "base_model_tp_plan" in d:1088            del d["base_model_tp_plan"]1089        # Do not serialize `base_model_pp_plan` for now1090        if "base_model_pp_plan" in d:1091            del d["base_model_pp_plan"]1092        for value in d.values():1093            if isinstance(value, dict):1094                self._remove_keys_not_serialized(value)1095 1096    @classmethod1097    def register_for_auto_class(cls, auto_class="AutoConfig"):1098        """1099        Register this class with a given auto class. This should only be used for custom configurations as the ones in1100        the library are already mapped with `AutoConfig`.1101 1102 1103 1104        Args:1105            auto_class (`str` or `type`, *optional*, defaults to `"AutoConfig"`):1106                The auto class to register this new configuration with.1107        """1108        if not isinstance(auto_class, str):1109            auto_class = auto_class.__name__1110 1111        import transformers.models.auto as auto_module1112 1113        if not hasattr(auto_module, auto_class):1114            raise ValueError(f"{auto_class} is not a valid auto class.")1115 1116        cls._auto_class = auto_class1117 1118    @staticmethod1119    def _get_global_generation_defaults() -> dict[str, Any]:1120        return {1121            "max_length": 20,1122            "min_length": 0,1123            "do_sample": False,1124            "early_stopping": False,1125            "num_beams": 1,1126            "temperature": 1.0,1127            "top_k": 50,1128            "top_p": 1.0,1129            "typical_p": 1.0,1130            "repetition_penalty": 1.0,1131            "length_penalty": 1.0,1132            "no_repeat_ngram_size": 0,1133            "encoder_no_repeat_ngram_size": 0,1134            "bad_words_ids": None,1135            "num_return_sequences": 1,1136            "output_scores": False,1137            "return_dict_in_generate": False,1138            "forced_bos_token_id": None,1139            "forced_eos_token_id": None,1140            "remove_invalid_values": False,1141            "exponential_decay_length_penalty": None,1142            "suppress_tokens": None,1143            "begin_suppress_tokens": None,1144            # Deprecated arguments (moved to the Hub). TODO joao, manuel: remove in v4.62.01145            "num_beam_groups": 1,1146            "diversity_penalty": 0.0,1147        }1148 1149    def _get_non_default_generation_parameters(self) -> dict[str, Any]:1150        """1151        Gets the non-default generation parameters on the PretrainedConfig instance1152        """1153        non_default_generation_parameters = {}1154        decoder_attribute_name = None1155 1156        # Composite models don't have a default config, use their decoder config as a fallback for default values1157        # If no known pattern is matched, then `default_config = None` -> check against the global generation defaults1158        try:1159            default_config = self.__class__()1160        except ValueError:1161            decoder_config = self.get_text_config(decoder=True)1162            if decoder_config is not self:1163                default_config = decoder_config.__class__()1164            else:1165                default_config = None1166 1167        # If it is a composite model, we want to check the subconfig that will be used for generation1168        self_decoder_config = self if decoder_attribute_name is None else getattr(self, decoder_attribute_name)1169 1170        for parameter_name, default_global_value in self._get_global_generation_defaults().items():1171            if hasattr(self_decoder_config, parameter_name):1172                is_default_in_config = is_default_generation_value = None1173                parameter_value = getattr(self_decoder_config, parameter_name)1174                # Three cases in which is okay for the model config to hold generation config parameters:1175                # 1. The parameter is set to `None`, effectively delegating its value to the generation config1176                if parameter_value is None:1177                    continue1178                # 2. If we have a default config, then the instance should hold the same generation defaults1179                if default_config is not None:1180                    is_default_in_config = parameter_value == getattr(default_config, parameter_name)1181                # 3. if we don't have a default config, then the instance should hold the global generation defaults1182                else:1183                    is_default_generation_value = parameter_value == default_global_value1184 1185                is_non_default = (is_default_in_config is False) or (1186                    is_default_in_config is None and is_default_generation_value is False1187                )1188                if is_non_default:1189                    non_default_generation_parameters[parameter_name] = getattr(self_decoder_config, parameter_name)1190 1191        return non_default_generation_parameters1192 1193    def get_text_config(self, decoder=None, encoder=None) -> "PretrainedConfig":1194        """1195        Returns the text config related to the text input (encoder) or text output (decoder) of the model. The1196        `decoder` and `encoder` input arguments can be used to specify which end of the model we are interested in,1197        which is useful on models that have both text input and output modalities.1198 1199        There are three possible outcomes of using this method:1200        1. On most models, it returns the original config instance itself.

Showing the first 1,200 of 1381 lines. Download the file for the rest.

Aluode/PerceptionLabPortable · CoolFace