Models¶

The base classes PreTrainedModel, TFPreTrainedModel, and FlaxPreTrainedModel implement the common methods for loading/saving a model either from a local file or directory, or from a pretrained model configuration provided by the library (downloaded from HuggingFace’s AWS S3 repository).

PreTrainedModel and TFPreTrainedModel also implement a few methods which are common among all the models to:

  • resize the input token embeddings when new tokens are added to the vocabulary

  • prune the attention heads of the model.

The other methods that are common to each model are defined in ModuleUtilsMixin (for the PyTorch models) and TFModuleUtilsMixin (for the TensorFlow models) or for text generation, GenerationMixin (for the PyTorch models) and TFGenerationMixin (for the TensorFlow models)

PreTrainedModel¶

class transformers.PreTrainedModel(config: transformers.configuration_utils.PretrainedConfig, *inputs, **kwargs)[source]¶

Base class for all models.

PreTrainedModel takes care of storing the configuration of the models and handles methods for loading, downloading and saving models as well as a few methods common to all models to:

  • resize the input embeddings,

  • prune heads in the self-attention heads.

Class attributes (overridden by derived classes):

  • config_class (PretrainedConfig) – A subclass of PretrainedConfig to use as configuration class for this model architecture.

  • load_tf_weights (Callable) – A python method for loading a TensorFlow checkpoint in a PyTorch model, taking as arguments:

    • model (PreTrainedModel) – An instance of the model on which to load the TensorFlow checkpoint.

    • config (PreTrainedConfig) – An instance of the configuration associated to the model.

    • path (str) – A path to the TensorFlow checkpoint.

  • base_model_prefix (str) – A string indicating the attribute associated to the base model in derived classes of the same architecture adding modules on top of the base model.

  • is_parallelizable (bool) – A flag indicating whether this model supports model parallelization.

property base_model¶

The main body of the model.

Type

torch.nn.Module

property dummy_inputs¶

Dummy inputs to do a forward pass in the network.

Type

Dict[str, torch.Tensor]

classmethod from_pretrained(pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs)[source]¶

Instantiate a pretrained pytorch model from a pre-trained model configuration.

The model is set in evaluation mode by default using model.eval() (Dropout modules are deactivated). To train the model, you should first set it back in training mode with model.train().

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 model id of a pretrained model hosted inside a model repo on huggingface.co. Valid model ids can be located at the root-level, like bert-base-uncased, or namespaced under a user or organization name, like dbmdz/bert-base-german-cased.

    • A path to a directory containing model weights saved using save_pretrained(), e.g., ./my_model_directory/.

    • A path or url to a tensorflow index checkpoint file (e.g, ./tf_model/model.ckpt.index). In this case, from_tf should be set to True and a configuration object should be provided as config argument. This loading path is slower than converting the TensorFlow checkpoint in a PyTorch model using the provided conversion scripts and loading the PyTorch model afterwards.

    • None if you are both providing the configuration and state dictionary (resp. with keyword arguments config and state_dict).

  • model_args (sequence of positional arguments, optional) – All remaning positional arguments will be passed to the underlying model’s __init__ method.

  • config (Union[PretrainedConfig, str, os.PathLike], optional) –

    Can be either:

    Configuration for the model to use instead of an automatically loaded configuation. Configuration can be automatically loaded when:

    • The model is a model provided by the library (loaded with the model id string of a pretrained model).

    • The model was saved using save_pretrained() and is reloaded by supplying the save directory.

    • The model is loaded by supplying a local directory as pretrained_model_name_or_path and a configuration JSON file named config.json is found in the directory.

  • state_dict (Dict[str, torch.Tensor], optional) –

    A state dictionary to use instead of a state dictionary loaded from saved weights file.

    This option can be used if you want to create a model from a pretrained configuration but load your own weights. In this case though, you should check if using save_pretrained() and from_pretrained() is not a simpler option.

  • cache_dir (Union[str, os.PathLike], optional) – Path to a directory in which a downloaded pretrained model configuration should be cached if the standard cache should not be used.

  • from_tf (bool, optional, defaults to False) – Load the model weights from a TensorFlow checkpoint save file (see docstring of pretrained_model_name_or_path argument).

  • 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 ot 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 transformers-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, defaults to None) – 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.

  • kwargs (remaining dictionary of keyword arguments, optional) –

    Can be used to update the configuration object (after it being loaded) and initiate the model (e.g., output_attentions=True). Behaves differently depending on whether a config is provided or automatically loaded:

    • If a configuration is provided with config, **kwargs will be directly passed to the underlying model’s __init__ method (we assume all relevant updates to the configuration have already been done)

    • If a configuration is not provided, kwargs will be first passed to the configuration class initialization function (from_pretrained()). Each key of kwargs that corresponds to a configuration attribute will be used to override said attribute with the supplied kwargs value. Remaining keys that do not correspond to any configuration attribute will be passed to the underlying model’s __init__ function.

Note

Passing use_auth_token=True is required when you want to use a private model.

Examples:

>>> from transformers import BertConfig, BertModel
>>> # Download model and configuration from huggingface.co and cache.
>>> model = BertModel.from_pretrained('bert-base-uncased')
>>> # Model was saved using `save_pretrained('./test/saved_model/')` (for example purposes, not runnable).
>>> model = BertModel.from_pretrained('./test/saved_model/')
>>> # Update configuration during loading.
>>> model = BertModel.from_pretrained('bert-base-uncased', output_attentions=True)
>>> assert model.config.output_attentions == True
>>> # Loading from a TF checkpoint file instead of a PyTorch model (slower, for example purposes, not runnable).
>>> config = BertConfig.from_json_file('./tf_model/my_tf_model_config.json')
>>> model = BertModel.from_pretrained('./tf_model/my_tf_checkpoint.ckpt.index', from_tf=True, config=config)
get_input_embeddings() → torch.nn.modules.module.Module[source]¶

Returns the model’s input embeddings.

Returns

A torch module mapping vocabulary to hidden states.

Return type

nn.Module

get_output_embeddings() → torch.nn.modules.module.Module[source]¶

Returns the model’s output embeddings.

Returns

A torch module mapping hidden states to vocabulary.

Return type

nn.Module

init_weights()[source]¶

Initializes and prunes weights if needed.

prune_heads(heads_to_prune: Dict[int, List[int]])[source]¶

Prunes heads of the base model.

Parameters

heads_to_prune (Dict[int, List[int]]) – Dictionary with keys being selected layer indices (int) and associated values being the list of heads to prune in said layer (list of int). 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.

resize_token_embeddings(new_num_tokens: Optional[int] = None) → torch.nn.modules.sparse.Embedding[source]¶

Resizes input token embeddings matrix of the model if new_num_tokens != config.vocab_size.

Takes care of tying weights embeddings afterwards if the model class has a tie_weights() method.

Parameters

new_num_tokens (int, optional) – The number of new tokens in the embedding matrix. Increasing the size will add newly initialized vectors at the end. Reducing the size will remove vectors from the end. If not provided or None, just returns a pointer to the input tokens torch.nn.Embedding module of the model without doing anything.

Returns

Pointer to the input tokens Embeddings Module of the model.

Return type

torch.nn.Embedding

save_pretrained(save_directory: Union[str, os.PathLike], save_config: bool = True, state_dict: Optional[dict] = None, save_function: Callable = <function save>)[source]¶

Save a model and its configuration file to a directory, so that it can be re-loaded using the :func:`~transformers.PreTrainedModel.from_pretrained` class method.

Parameters
  • save_directory (str or os.PathLike) – Directory to which to save. Will be created if it doesn’t exist.

  • save_config (bool, optional, defaults to True) – Whether or not to save the config of the model. Useful when in distributed training like TPUs and need to call this function on all processes. In this case, set save_config=True only on the main process to avoid race conditions.

  • state_dict (nested dictionary of torch.Tensor) – The state dictionary of the model to save. Will default to self.state_dict(), but can be used to only save parts of the model or if special precautions need to be taken when recovering the state dictionary of a model (like when using model parallelism).

  • save_function (Callable) – The function to use to save the state dictionary. Useful on distributed training like TPUs when one need to replace torch.save by another method.

set_input_embeddings(value: torch.nn.modules.module.Module)[source]¶

Set model’s input embeddings.

Parameters

value (nn.Module) – A module mapping vocabulary to hidden states.

tie_weights()[source]¶

Tie the weights between the input embeddings and the output embeddings.

If the torchscript flag is set in the configuration, can’t handle parameter sharing so we are cloning the weights instead.

ModuleUtilsMixin¶

class transformers.modeling_utils.ModuleUtilsMixin[source]¶

A few utilities for torch.nn.Modules, to be used as a mixin.

add_memory_hooks()[source]¶

Add a memory hook before and after each sub-module forward pass to record increase in memory consumption.

Increase in memory consumption is stored in a mem_rss_diff attribute for each module and can be reset to zero with model.reset_memory_hooks_state().

property device¶

The device on which the module is (assuming that all the module parameters are on the same device).

Type

torch.device

property dtype¶

The dtype of the module (assuming that all the module parameters have the same dtype).

Type

torch.dtype

estimate_tokens(input_dict: Dict[str, Union[torch.Tensor, Any]]) → int[source]¶

Helper function to estimate the total number of tokens from the model inputs.

Parameters

inputs (dict) – The model inputs.

Returns

The total number of tokens.

Return type

int

floating_point_ops(input_dict: Dict[str, Union[torch.Tensor, Any]], exclude_embeddings: bool = True) → int[source]¶

Get number of (optionally, non-embeddings) floating-point operations for the forward and backward passes of a batch with this transformer model. Default approximation neglects the quadratic dependency on the number of tokens (valid if 12 * d_model << sequence_length) as laid out in this paper section 2.1. Should be overridden for transformers with parameter re-use e.g. Albert or Universal Transformers, or if doing long-range modeling with very high sequence lengths.

Parameters
  • batch_size (int) – The batch size for the forward pass.

  • sequence_length (int) – The number of tokens in each line of the batch.

  • exclude_embeddings (bool, optional, defaults to True) – Whether or not to count embedding and softmax operations.

Returns

The number of floating-point operations.

Return type

int

get_extended_attention_mask(attention_mask: torch.Tensor, input_shape: Tuple[int], device: <property object at 0x7f3242450278>) → torch.Tensor[source]¶

Makes broadcastable attention and causal masks so that future and masked tokens are ignored.

Parameters
  • attention_mask (torch.Tensor) – Mask with ones indicating tokens to attend to, zeros for tokens to ignore.

  • input_shape (Tuple[int]) – The shape of the input to the model.

  • device – (torch.device): The device of the input to the model.

Returns

torch.Tensor The extended attention mask, with a the same dtype as attention_mask.dtype.

get_head_mask(head_mask: Optional[torch.Tensor], num_hidden_layers: int, is_attention_chunked: bool = False) → torch.Tensor[source]¶

Prepare the head mask if needed.

Parameters
  • head_mask (torch.Tensor with shape [num_heads] or [num_hidden_layers x num_heads], optional) – The mask indicating if we should keep the heads or not (1.0 for keep, 0.0 for discard).

  • num_hidden_layers (int) – The number of hidden layers in the model.

  • is_attention_chunked – (bool, optional, defaults to :obj:`False): Whether or not the attentions scores are computed by chunks or not.

Returns

torch.Tensor with shape [num_hidden_layers x batch x num_heads x seq_length x seq_length] or list with [None] for each layer.

invert_attention_mask(encoder_attention_mask: torch.Tensor) → torch.Tensor[source]¶

Invert an attention mask (e.g., switches 0. and 1.).

Parameters

encoder_attention_mask (torch.Tensor) – An attention mask.

Returns

The inverted attention mask.

Return type

torch.Tensor

num_parameters(only_trainable: bool = False, exclude_embeddings: bool = False) → int[source]¶

Get number of (optionally, trainable or non-embeddings) parameters in the module.

Parameters
  • only_trainable (bool, optional, defaults to False) – Whether or not to return only the number of trainable parameters

  • exclude_embeddings (bool, optional, defaults to False) – Whether or not to return only the number of non-embeddings parameters

Returns

The number of parameters.

Return type

int

reset_memory_hooks_state()[source]¶

Reset the mem_rss_diff attribute of each module (see add_memory_hooks()).

TFPreTrainedModel¶

class transformers.TFPreTrainedModel(*args, **kwargs)[source]¶

Base class for all TF models.

TFPreTrainedModel takes care of storing the configuration of the models and handles methods for loading, downloading and saving models as well as a few methods common to all models to:

  • resize the input embeddings,

  • prune heads in the self-attention heads.

Class attributes (overridden by derived classes):

  • config_class (PretrainedConfig) – A subclass of PretrainedConfig to use as configuration class for this model architecture.

  • base_model_prefix (str) – A string indicating the attribute associated to the base model in derived classes of the same architecture adding modules on top of the base model.

property dummy_inputs¶

Dummy inputs to build the network.

Returns

The dummy inputs.

Return type

Dict[str, tf.Tensor]

classmethod from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)[source]¶

Instantiate a pretrained TF 2.0 model from a pre-trained model configuration.

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, optional) –

    Can be either:

    • A string, the model id of a pretrained model hosted inside a model repo on huggingface.co. Valid model ids can be located at the root-level, like bert-base-uncased, or namespaced under a user or organization name, like dbmdz/bert-base-german-cased.

    • A path to a directory containing model weights saved using save_pretrained(), e.g., ./my_model_directory/.

    • A path or url to a PyTorch state_dict save file (e.g, ./pt_model/pytorch_model.bin). In this case, from_pt should be set to True and a configuration object should be provided as config argument. This loading path is slower than converting the PyTorch model in a TensorFlow model using the provided conversion scripts and loading the TensorFlow model afterwards.

    • None if you are both providing the configuration and state dictionary (resp. with keyword arguments config and state_dict).

  • model_args (sequence of positional arguments, optional) – All remaning positional arguments will be passed to the underlying model’s __init__ method.

  • config (Union[PretrainedConfig, str], optional) –

    Can be either:

    Configuration for the model to use instead of an automatically loaded configuation. Configuration can be automatically loaded when:

    • The model is a model provided by the library (loaded with the model id string of a pretrained model).

    • The model was saved using save_pretrained() and is reloaded by supplying the save directory.

    • The model is loaded by supplying a local directory as pretrained_model_name_or_path and a configuration JSON file named config.json is found in the directory.

  • from_pt – (bool, optional, defaults to False): Load the model weights from a PyTorch state_dict save file (see docstring of pretrained_model_name_or_path argument).

  • cache_dir (str, optional) – Path to a directory in which a downloaded pretrained model configuration should be cached if the standard cache should not be used.

  • 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 ot 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 (e.g., not try doanloading 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 transformers-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, defaults to None) – 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.

  • kwargs (remaining dictionary of keyword arguments, optional) –

    Can be used to update the configuration object (after it being loaded) and initiate the model (e.g., output_attentions=True). Behaves differently depending on whether a config is provided or automatically loaded:

    • If a configuration is provided with config, **kwargs will be directly passed to the underlying model’s __init__ method (we assume all relevant updates to the configuration have already been done)

    • If a configuration is not provided, kwargs will be first passed to the configuration class initialization function (from_pretrained()). Each key of kwargs that corresponds to a configuration attribute will be used to override said attribute with the supplied kwargs value. Remaining keys that do not correspond to any configuration attribute will be passed to the underlying model’s __init__ function.

Note

Passing use_auth_token=True is required when you want to use a private model.

Examples:

>>> from transformers import BertConfig, TFBertModel
>>> # Download model and configuration from huggingface.co and cache.
>>> model = TFBertModel.from_pretrained('bert-base-uncased')
>>> # Model was saved using `save_pretrained('./test/saved_model/')` (for example purposes, not runnable).
>>> model = TFBertModel.from_pretrained('./test/saved_model/')
>>> # Update configuration during loading.
>>> model = TFBertModel.from_pretrained('bert-base-uncased', output_attentions=True)
>>> assert model.config.output_attentions == True
>>> # Loading from a Pytorch model file instead of a TensorFlow checkpoint (slower, for example purposes, not runnable).
>>> config = BertConfig.from_json_file('./pt_model/my_pt_model_config.json')
>>> model = TFBertModel.from_pretrained('./pt_model/my_pytorch_model.bin', from_pt=True, config=config)
get_bias() → Union[None, Dict[str, tensorflow.python.ops.variables.Variable]][source]¶

Dict of bias attached to an LM head. The key represents the name of the bias attribute.

Returns

The weights representing the bias, None if not an LM model.

Return type

tf.Variable

get_input_embeddings() → tensorflow.python.keras.engine.base_layer.Layer[source]¶

Returns the model’s input embeddings layer.

Returns

The embeddings layer mapping vocabulary to hidden states.

Return type

tf.Variable

get_lm_head() → tensorflow.python.keras.engine.base_layer.Layer[source]¶

The LM Head layer. This method must be overwritten by all the models that have a lm head.

Returns

The LM head layer if the model has one, None if not.

Return type

tf.keras.layers.Layer

get_output_embeddings() → Union[None, tensorflow.python.keras.engine.base_layer.Layer][source]¶

Returns the model’s output embeddings

Returns

The new weights mapping vocabulary to hidden states.

Return type

tf.Variable

get_output_layer_with_bias() → Union[None, tensorflow.python.keras.engine.base_layer.Layer][source]¶

Get the layer that handles a bias attribute in case the model has an LM head with weights tied to the embeddings

Returns

The layer that handles the bias, None if not an LM model.

Return type

tf.keras.layers.Layer

get_prefix_bias_name() → Union[None, str][source]¶

Get the concatenated _prefix name of the bias from the model name to the parent layer

Returns

The _prefix name of the bias.

Return type

str

prune_heads(heads_to_prune)[source]¶

Prunes heads of the base model.

Parameters

heads_to_prune (Dict[int, List[int]]) – Dictionary with keys being selected layer indices (int) and associated values being the list of heads to prune in said layer (list of int). 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.

resize_token_embeddings(new_num_tokens=None) → tensorflow.python.ops.variables.Variable[source]¶

Resizes input token embeddings matrix of the model if new_num_tokens != config.vocab_size.

Takes care of tying weights embeddings afterwards if the model class has a tie_weights() method.

Parameters

new_num_tokens (int, optional) – The number of new tokens in the embedding matrix. Increasing the size will add newly initialized vectors at the end. Reducing the size will remove vectors from the end. If not provided or None, just returns a pointer to the input tokens tf.Variable module of the model without doing anything.

Returns

Pointer to the input tokens Embeddings Module of the model.

Return type

tf.Variable

save_pretrained(save_directory, saved_model=False, version=1)[source]¶

Save a model and its configuration file to a directory, so that it can be re-loaded using the from_pretrained() class method.

Parameters
  • save_directory (str) – Directory to which to save. Will be created if it doesn’t exist.

  • saved_model (bool, optional, defaults to False) – If the model has to be saved in saved model format as well or not.

  • version (int, optional, defaults to 1) – The version of the saved model. A saved model needs to be versioned in order to be properly loaded by TensorFlow Serving as detailed in the official documentation https://www.tensorflow.org/tfx/serving/serving_basic

serving(inputs)[source]¶

Method used for serving the model.

Parameters

inputs (Dict[str, tf.Tensor]) – The input of the saved model as a dictionnary of tensors.

serving_output()[source]¶

Prepare the output of the saved model. Each model must implement this function.

Parameters

output (TFBaseModelOutput) – The output returned by the model.

set_bias(value)[source]¶

Set all the bias in the LM head.

Parameters

value (Dict[tf.Variable]) – All the new bias attached to an LM head.

set_input_embeddings(value)[source]¶

Set model’s input embeddings

Parameters

value (tf.Variable) – The new weights mapping hidden states to vocabulary.

set_output_embeddings(value)[source]¶

Set model’s output embeddings

Parameters

value (tf.Variable) – The new weights mapping hidden states to vocabulary.

TFModelUtilsMixin¶

class transformers.modeling_tf_utils.TFModelUtilsMixin[source]¶

A few utilities for tf.keras.Model, to be used as a mixin.

num_parameters(only_trainable: bool = False) → int[source]¶

Get the number of (optionally, trainable) parameters in the model.

Parameters

only_trainable (bool, optional, defaults to False) – Whether or not to return only the number of trainable parameters

Returns

The number of parameters.

Return type

int

FlaxPreTrainedModel¶

class transformers.FlaxPreTrainedModel(config: transformers.configuration_utils.PretrainedConfig, module: flax.linen.module.Module, input_shape: Tuple = (1, 1), seed: int = 0, dtype: numpy.dtype = <class 'jax._src.numpy.lax_numpy.float32'>)[source]¶

Base class for all models.

FlaxPreTrainedModel takes care of storing the configuration of the models and handles methods for loading, downloading and saving models.

Class attributes (overridden by derived classes):

  • config_class (PretrainedConfig) – A subclass of PretrainedConfig to use as configuration class for this model architecture.

  • base_model_prefix (str) – A string indicating the attribute associated to the base model in derived classes of the same architecture adding modules on top of the base model.

classmethod from_pretrained(pretrained_model_name_or_path: Union[str, os.PathLike], dtype: numpy.dtype = <class 'jax._src.numpy.lax_numpy.float32'>, *model_args, **kwargs)[source]¶

Instantiate a pretrained flax model from a pre-trained model configuration.

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) –

    Can be either:

    • A string, the model id of a pretrained model hosted inside a model repo on huggingface.co. Valid model ids can be located at the root-level, like bert-base-uncased, or namespaced under a user or organization name, like dbmdz/bert-base-german-cased.

    • A path to a directory containing model weights saved using save_pretrained(), e.g., ./my_model_directory/.

    • A path or url to a pt index checkpoint file (e.g, ./tf_model/model.ckpt.index). In this case, from_pt should be set to True.

  • model_args (sequence of positional arguments, optional) – All remaning positional arguments will be passed to the underlying model’s __init__ method.

  • config (Union[PretrainedConfig, str, os.PathLike], optional) –

    Can be either:

    Configuration for the model to use instead of an automatically loaded configuation. Configuration can be automatically loaded when:

    • The model is a model provided by the library (loaded with the model id string of a pretrained model).

    • The model was saved using save_pretrained() and is reloaded by supplying the save directory.

    • The model is loaded by supplying a local directory as pretrained_model_name_or_path and a configuration JSON file named config.json is found in the directory.

  • cache_dir (Union[str, os.PathLike], optional) – Path to a directory in which a downloaded pretrained model configuration should be cached if the standard cache should not be used.

  • from_pt (bool, optional, defaults to False) – Load the model weights from a PyTorch checkpoint save file (see docstring of pretrained_model_name_or_path argument).

  • 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.

  • 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).

  • 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.

  • kwargs (remaining dictionary of keyword arguments, optional) –

    Can be used to update the configuration object (after it being loaded) and initiate the model (e.g., output_attentions=True). Behaves differently depending on whether a config is provided or automatically loaded:

    • If a configuration is provided with config, **kwargs will be directly passed to the underlying model’s __init__ method (we assume all relevant updates to the configuration have already been done)

    • If a configuration is not provided, kwargs will be first passed to the configuration class initialization function (from_pretrained()). Each key of kwargs that corresponds to a configuration attribute will be used to override said attribute with the supplied kwargs value. Remaining keys that do not correspond to any configuration attribute will be passed to the underlying model’s __init__ function.

Examples:

>>> from transformers import BertConfig, FlaxBertModel
>>> # Download model and configuration from huggingface.co and cache.
>>> model = FlaxBertModel.from_pretrained('bert-base-cased')
>>> # Model was saved using `save_pretrained('./test/saved_model/')` (for example purposes, not runnable).
>>> model = FlaxBertModel.from_pretrained('./test/saved_model/')
>>> # Loading from a PyTorch checkpoint file instead of a PyTorch model (slower, for example purposes, not runnable).
>>> config = BertConfig.from_json_file('./pt_model/config.json')
>>> model = FlaxBertModel.from_pretrained('./pt_model/pytorch_model.bin', from_pt=True, config=config)
save_pretrained(save_directory: Union[str, os.PathLike])[source]¶

Save a model and its configuration file to a directory, so that it can be re-loaded using the :func:`~transformers.FlaxPreTrainedModel.from_pretrained` class method

Parameters

save_directory (str or os.PathLike) – Directory to which to save. Will be created if it doesn’t exist.

Generation¶

class transformers.generation_utils.GenerationMixin[source]¶

A class containing all of the functions supporting generation, to be used as a mixin in PreTrainedModel.

adjust_logits_during_generation(logits: torch.FloatTensor, **kwargs) → torch.FloatTensor[source]¶

Implement in subclasses of PreTrainedModel for custom behavior to adjust the logits in the generate method.

beam_sample(input_ids: torch.LongTensor, beam_scorer: transformers.generation_beam_search.BeamScorer, logits_processor: Optional[transformers.generation_logits_process.LogitsProcessorList] = None, stopping_criteria: Optional[transformers.generation_stopping_criteria.StoppingCriteriaList] = None, logits_warper: Optional[transformers.generation_logits_process.LogitsProcessorList] = None, max_length: Optional[int] = None, pad_token_id: Optional[int] = None, eos_token_id: Optional[int] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, output_scores: Optional[bool] = None, return_dict_in_generate: Optional[bool] = None, **model_kwargs) → Union[transformers.generation_utils.BeamSampleEncoderDecoderOutput, transformers.generation_utils.BeamSampleDecoderOnlyOutput, torch.LongTensor][source]¶

Generates sequences for models with a language modeling head using beam search with multinomial sampling.

Parameters
  • input_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) – The sequence used as a prompt for the generation. If None the method initializes it as an empty torch.LongTensor of shape (1,).

  • beam_scorer (BeamScorer) – A derived instance of BeamScorer that defines how beam hypotheses are constructed, stored and sorted during generation. For more information, the documentation of BeamScorer should be read.

  • logits_processor (LogitsProcessorList, optional) – An instance of LogitsProcessorList. List of instances of class derived from LogitsProcessor used to modify the prediction scores of the language modeling head applied at each generation step.

  • stopping_criteria (StoppingCriteriaList, optional) – An instance of StoppingCriteriaList. List of instances of class derived from StoppingCriteria used to tell if the generation loop should stop.

  • logits_warper (LogitsProcessorList, optional) – An instance of LogitsProcessorList. List of instances of class derived from LogitsWarper used to warp the prediction score distribution of the language modeling head applied before multinomial sampling at each generation step.

  • max_length (int, optional, defaults to 20) – The maximum length of the sequence to be generated.

  • pad_token_id (int, optional) – The id of the padding token.

  • eos_token_id (int, optional) – The id of the end-of-sequence token.

  • output_attentions (bool, optional, defaults to False) – Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more details.

  • output_hidden_states (bool, optional, defaults to False) – Whether or not to return trhe hidden states of all layers. See hidden_states under returned tensors for more details.

  • output_scores (bool, optional, defaults to False) – Whether or not to return the prediction scores. See scores under returned tensors for more details.

  • return_dict_in_generate (bool, optional, defaults to False) – Whether or not to return a ModelOutput instead of a plain tuple.

  • model_kwargs – Additional model specific kwargs will be forwarded to the forward function of the model. If model is an encoder-decoder model the kwargs should include encoder_outputs.

Returns

BeamSampleDecoderOnlyOutput, BeamSampleEncoderDecoderOutput or obj:torch.LongTensor: A torch.LongTensor containing the generated tokens (default behaviour) or a BeamSampleDecoderOnlyOutput if model.config.is_encoder_decoder=False and return_dict_in_generate=True or a BeamSampleEncoderDecoderOutput if model.config.is_encoder_decoder=True.

Examples:

>>> from transformers import (
...     AutoTokenizer,
...     AutoModelForSeq2SeqLM,
...     LogitsProcessorList,
...     MinLengthLogitsProcessor,
...     TopKLogitsWarper,
...     TemperatureLogitsWarper,
...     BeamSearchScorer,
... )
>>> import torch

>>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
>>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")

>>> encoder_input_str = "translate English to German: How old are you?"
>>> encoder_input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids

>>> # lets run beam search using 3 beams
>>> num_beams = 3
>>> # define decoder start token ids
>>> input_ids = torch.ones((num_beams, 1), device=model.device, dtype=torch.long)
>>> input_ids = input_ids * model.config.decoder_start_token_id

>>> # add encoder_outputs to model keyword arguments
>>> model_kwargs = {
...     "encoder_outputs": model.get_encoder()(encoder_input_ids.repeat_interleave(num_beams, dim=0), return_dict=True)
... }

>>> # instantiate beam scorer
>>> beam_scorer = BeamSearchScorer(
...     batch_size=1,
...     max_length=model.config.max_length,
...     num_beams=num_beams,
...     device=model.device,
... )

>>> # instantiate logits processors
>>> logits_processor = LogitsProcessorList([
...     MinLengthLogitsProcessor(5, eos_token_id=model.config.eos_token_id)
... ])
>>> # instantiate logits processors
>>> logits_warper = LogitsProcessorList([
...     TopKLogitsWarper(50),
...     TemperatureLogitsWarper(0.7),
... ])

>>> outputs = model.beam_sample(
...     input_ids, beam_scorer, logits_processor=logits_processor, logits_warper=logits_warper, **model_kwargs
... )

>>> print("Generated:", tokenizer.batch_decode(outputs, skip_special_tokens=True))

Generates sequences for models with a language modeling head using beam search decoding.

Parameters
  • input_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) – The sequence used as a prompt for the generation. If None the method initializes it as an empty torch.LongTensor of shape (1,).

  • beam_scorer (BeamScorer) – An derived instance of BeamScorer that defines how beam hypotheses are constructed, stored and sorted during generation. For more information, the documentation of BeamScorer should be read.

  • logits_processor (LogitsProcessorList, optional) – An instance of LogitsProcessorList. List of instances of class derived from LogitsProcessor used to modify the prediction scores of the language modeling head applied at each generation step.

  • stopping_criteria (StoppingCriteriaList, optional) – An instance of StoppingCriteriaList. List of instances of class derived from StoppingCriteria used to tell if the generation loop should stop.

  • max_length (int, optional, defaults to 20) – The maximum length of the sequence to be generated.

  • pad_token_id (int, optional) – The id of the padding token.

  • eos_token_id (int, optional) – The id of the end-of-sequence token.

  • output_attentions (bool, optional, defaults to False) – Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more details.

  • output_hidden_states (bool, optional, defaults to False) – Whether or not to return trhe hidden states of all layers. See hidden_states under returned tensors for more details.

  • output_scores (bool, optional, defaults to False) – Whether or not to return the prediction scores. See scores under returned tensors for more details.

  • return_dict_in_generate (bool, optional, defaults to False) – Whether or not to return a ModelOutput instead of a plain tuple.

  • model_kwargs – Additional model specific kwargs will be forwarded to the forward function of the model. If model is an encoder-decoder model the kwargs should include encoder_outputs.

Returns

generation_utilsBeamSearchDecoderOnlyOutput, BeamSearchEncoderDecoderOutput or obj:torch.LongTensor: A torch.LongTensor containing the generated tokens (default behaviour) or a BeamSearchDecoderOnlyOutput if model.config.is_encoder_decoder=False and return_dict_in_generate=True or a BeamSearchEncoderDecoderOutput if model.config.is_encoder_decoder=True.

Examples:

>>> from transformers import (
...    AutoTokenizer,
...    AutoModelForSeq2SeqLM,
...    LogitsProcessorList,
...    MinLengthLogitsProcessor,
...    BeamSearchScorer,
... )
>>> import torch

>>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
>>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")

>>> encoder_input_str = "translate English to German: How old are you?"
>>> encoder_input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids


>>> # lets run beam search using 3 beams
>>> num_beams = 3
>>> # define decoder start token ids
>>> input_ids = torch.ones((num_beams, 1), device=model.device, dtype=torch.long)
>>> input_ids = input_ids * model.config.decoder_start_token_id

>>> # add encoder_outputs to model keyword arguments
>>> model_kwargs = {
...     "encoder_outputs": model.get_encoder()(encoder_input_ids.repeat_interleave(num_beams, dim=0), return_dict=True)
... }

>>> # instantiate beam scorer
>>> beam_scorer = BeamSearchScorer(
...     batch_size=1,
...     max_length=model.config.max_length,
...     num_beams=num_beams,
...     device=model.device,
... )

>>> # instantiate logits processors
>>> logits_processor = LogitsProcessorList([
...     MinLengthLogitsProcessor(5, eos_token_id=model.config.eos_token_id),
... ])

>>> outputs = model.beam_search(input_ids, beam_scorer, logits_processor=logits_processor, **model_kwargs)

>>> print("Generated:", tokenizer.batch_decode(outputs, skip_special_tokens=True))
generate(input_ids: Optional[torch.LongTensor] = None, max_length: Optional[int] = None, min_length: Optional[int] = None, do_sample: Optional[bool] = None, early_stopping: Optional[bool] = None, num_beams: Optional[int] = None, temperature: Optional[float] = None, top_k: Optional[int] = None, top_p: Optional[float] = None, repetition_penalty: Optional[float] = None, bad_words_ids: Optional[Iterable[int]] = None, bos_token_id: Optional[int] = None, pad_token_id: Optional[int] = None, eos_token_id: Optional[int] = None, length_penalty: Optional[float] = None, no_repeat_ngram_size: Optional[int] = None, encoder_no_repeat_ngram_size: Optional[int] = None, num_return_sequences: Optional[int] = None, max_time: Optional[float] = None, decoder_start_token_id: Optional[int] = None, use_cache: Optional[bool] = None, num_beam_groups: Optional[int] = None, diversity_penalty: Optional[float] = None, prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, output_scores: Optional[bool] = None, return_dict_in_generate: Optional[bool] = None, forced_bos_token_id: Optional[int] = None, forced_eos_token_id: Optional[int] = None, **model_kwargs) → Union[transformers.generation_utils.GreedySearchEncoderDecoderOutput, transformers.generation_utils.GreedySearchDecoderOnlyOutput, transformers.generation_utils.SampleEncoderDecoderOutput, transformers.generation_utils.SampleDecoderOnlyOutput, transformers.generation_utils.BeamSearchEncoderDecoderOutput, transformers.generation_utils.BeamSearchDecoderOnlyOutput, transformers.generation_utils.BeamSampleEncoderDecoderOutput, transformers.generation_utils.BeamSampleDecoderOnlyOutput, torch.LongTensor][source]¶

Generates sequences for models with a language modeling head. The method currently supports greedy decoding, multinomial sampling, beam-search decoding, and beam-search multinomial sampling.

Apart from input_ids and attention_mask, all the arguments below will default to the value of the attribute of the same name inside the PretrainedConfig of the model. The default values indicated are the default values of those config.

Most of these parameters are explained in more detail in this blog post.

Parameters
  • input_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) – The sequence used as a prompt for the generation. If None the method initializes it as an empty torch.LongTensor of shape (1,).

  • max_length (int, optional, defaults to 20) – The maximum length of the sequence to be generated.

  • min_length (int, optional, defaults to 10) – The minimum length of the sequence to be generated.

  • do_sample (bool, optional, defaults to False) – Whether or not to use sampling ; use greedy decoding otherwise.

  • early_stopping (bool, optional, defaults to False) – Whether to stop the beam search when at least num_beams sentences are finished per batch or not.

  • num_beams (int, optional, defaults to 1) – Number of beams for beam search. 1 means no beam search.

  • temperature (float, optional, defaults tp 1.0) – The value used to module the next token probabilities.

  • top_k (int, optional, defaults to 50) – The number of highest probability vocabulary tokens to keep for top-k-filtering.

  • top_p (float, optional, defaults to 1.0) – If set to float < 1, only the most probable tokens with probabilities that add up to top_p or higher are kept for generation.

  • repetition_penalty (float, optional, defaults to 1.0) – The parameter for repetition penalty. 1.0 means no penalty. See this paper for more details.

  • pad_token_id (int, optional) – The id of the padding token.

  • bos_token_id (int, optional) – The id of the beginning-of-sequence token.

  • eos_token_id (int, optional) – The id of the end-of-sequence token.

  • length_penalty (float, optional, defaults to 1.0) – Exponential penalty to the length. 1.0 means no penalty. Set to values < 1.0 in order to encourage the model to generate shorter sequences, to a value > 1.0 in order to encourage the model to produce longer sequences.

  • no_repeat_ngram_size (int, optional, defaults to 0) – If set to int > 0, all ngrams of that size can only occur once.

  • encoder_no_repeat_ngram_size (int, optional, defaults to 0) – If set to int > 0, all ngrams of that size that occur in the encoder_input_ids cannot occur in the decoder_input_ids.

  • bad_words_ids (List[List[int]], optional) – List of token ids that are not allowed to be generated. In order to get the tokens of the words that should not appear in the generated text, use tokenizer(bad_word, add_prefix_space=True).input_ids.

  • num_return_sequences (int, optional, defaults to 1) – The number of independently computed returned sequences for each element in the batch.

  • max_time (float, optional, defaults to None) – The maximum amount of time you allow the computation to run for in seconds. generation will still finish the current pass after allocated time has been passed.

  • attention_mask (torch.LongTensor of shape (batch_size, sequence_length), optional) – Mask to avoid performing attention on padding token indices. Mask values are in [0, 1], 1 for tokens that are not masked, and 0 for masked tokens. If not provided, will default to a tensor the same shape as input_ids that masks the pad token. What are attention masks?

  • decoder_start_token_id (int, optional) – If an encoder-decoder model starts decoding with a different token than bos, the id of that token.

  • use_cache – (bool, optional, defaults to True): Whether or not the model should use the past last key/values attentions (if applicable to the model) to speed up decoding.

  • num_beam_groups (int, optional, defaults to 1) – Number of groups to divide num_beams into in order to ensure diversity among different groups of beams. this paper for more details.

  • diversity_penalty (float, optional, defaults to 0.0) – This value is subtracted from a beam’s score if it generates a token same as any beam from other group at a particular time. Note that diversity_penalty is only effective if group beam search is enabled.

  • prefix_allowed_tokens_fn – (Callable[[int, torch.Tensor], List[int]], optional): If provided, this function constraints the beam search to allowed tokens only at each step. If not provided no constraint is applied. This function takes 2 arguments inputs_ids and the batch ID batch_id. It has to return a list with the allowed tokens for the next generation step conditioned on the previously generated tokens inputs_ids and the batch ID batch_id. This argument is useful for constrained generation conditioned on the prefix, as described in Autoregressive Entity Retrieval.

  • output_attentions (bool, optional, defaults to False) – Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more details.

  • output_hidden_states (bool, optional, defaults to False) – Whether or not to return trhe hidden states of all layers. See hidden_states under returned tensors for more details.

  • output_scores (bool, optional, defaults to False) – Whether or not to return the prediction scores. See scores under returned tensors for more details.

  • return_dict_in_generate (bool, optional, defaults to False) – Whether or not to return a ModelOutput instead of a plain tuple.

  • forced_bos_token_id (int, optional) – The id of the token to force as the first generated token after the decoder_start_token_id. Useful for multilingual models like mBART where the first generated token needs to be the target language token.

  • forced_eos_token_id (int, optional) – The id of the token to force as the last generated token when max_length is reached.

  • model_kwargs – Additional model specific kwargs will be forwarded to the forward function of the model. If the model is an encoder-decoder model, encoder specific kwargs should not be prefixed and decoder specific kwargs should be prefixed with decoder_.

Returns

A ModelOutput (if return_dict_in_generate=True or when config.return_dict_in_generate=True) or a torch.FloatTensor.

If the model is not an encoder-decoder model (model.config.is_encoder_decoder=False), the possible ModelOutput types are:

If the model is an encoder-decoder model (model.config.is_encoder_decoder=True), the possible ModelOutput types are:

Return type

ModelOutput or torch.LongTensor

Examples::
>>> from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModelForSeq2SeqLM
>>> tokenizer = AutoTokenizer.from_pretrained("distilgpt2")
>>> model = AutoModelForCausalLM.from_pretrained("distilgpt2")
>>> # do greedy decoding without providing a prompt
>>> outputs = model.generate(max_length=40)
>>> print("Generated:", tokenizer.decode(outputs[0], skip_special_tokens=True))
>>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
>>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")
>>> document = (
... "at least two people were killed in a suspected bomb attack on a passenger bus "
... "in the strife-torn southern philippines on monday , the military said."
... )
>>> # encode input contex
>>> input_ids = tokenizer(document, return_tensors="pt").input_ids
>>> # generate 3 independent sequences using beam search decoding (5 beams)
>>> # with T5 encoder-decoder model conditioned on short news article.
>>> outputs = model.generate(input_ids=input_ids, num_beams=5, num_return_sequences=3)
>>> print("Generated:", tokenizer.batch_decode(outputs, skip_special_tokens=True))
>>> tokenizer = AutoTokenizer.from_pretrained("distilgpt2")
>>> model = AutoModelForCausalLM.from_pretrained("distilgpt2")
>>> input_context = "The dog"
>>> # encode input context
>>> input_ids = tokenizer(input_context, return_tensors="pt").input_ids
>>> # generate 3 candidates using sampling
>>> outputs = model.generate(input_ids=input_ids, max_length=20, num_return_sequences=3, do_sample=True)
>>> print("Generated:", tokenizer.batch_decode(outputs, skip_special_tokens=True))
>>> tokenizer = AutoTokenizer.from_pretrained("ctrl")
>>> model = AutoModelForCausalLM.from_pretrained("ctrl")
>>> # "Legal" is one of the control codes for ctrl
>>> input_context = "Legal My neighbor is"
>>> # encode input context
>>> input_ids = tokenizer(input_context, return_tensors="pt").input_ids
>>> outputs = model.generate(input_ids=input_ids, max_length=20, repetition_penalty=1.2)
>>> print("Generated:", tokenizer.decode(outputs[0], skip_special_tokens=True))
>>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
>>> model = AutoModelForCausalLM.from_pretrained("gpt2")
>>> input_context = "My cute dog"
>>> # get tokens of words that should not be generated
>>> bad_words_ids = [tokenizer(bad_word, add_prefix_space=True).input_ids for bad_word in ["idiot", "stupid", "shut up"]]
>>> # encode input context
>>> input_ids = tokenizer(input_context, return_tensors="pt").input_ids
>>> # generate sequences without allowing bad_words to be generated
>>> outputs = model.generate(input_ids=input_ids, max_length=20, do_sample=True, bad_words_ids=bad_words_ids)
>>> print("Generated:", tokenizer.decode(outputs[0], skip_special_tokens=True))

Generates sequences for models with a language modeling head using greedy decoding.

Parameters
  • input_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) – The sequence used as a prompt for the generation. If None the method initializes it as an empty torch.LongTensor of shape (1,).

  • logits_processor (LogitsProcessorList, optional) – An instance of LogitsProcessorList. List of instances of class derived from LogitsProcessor used to modify the prediction scores of the language modeling head applied at each generation step.

  • stopping_criteria (StoppingCriteriaList, optional) – An instance of StoppingCriteriaList. List of instances of class derived from StoppingCriteria used to tell if the generation loop should stop.

  • max_length (int, optional, defaults to 20) – The maximum length of the sequence to be generated.

  • pad_token_id (int, optional) – The id of the padding token.

  • eos_token_id (int, optional) – The id of the end-of-sequence token.

  • output_attentions (bool, optional, defaults to False) – Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more details.

  • output_hidden_states (bool, optional, defaults to False) – Whether or not to return trhe hidden states of all layers. See hidden_states under returned tensors for more details.

  • output_scores (bool, optional, defaults to False) – Whether or not to return the prediction scores. See scores under returned tensors for more details.

  • return_dict_in_generate (bool, optional, defaults to False) – Whether or not to return a ModelOutput instead of a plain tuple.

  • model_kwargs – Additional model specific keyword arguments will be forwarded to the forward function of the model. If model is an encoder-decoder model the kwargs should include encoder_outputs.

Returns

GreedySearchDecoderOnlyOutput, GreedySearchEncoderDecoderOutput or obj:torch.LongTensor: A torch.LongTensor containing the generated tokens (default behaviour) or a GreedySearchDecoderOnlyOutput if model.config.is_encoder_decoder=False and return_dict_in_generate=True or a GreedySearchEncoderDecoderOutput if model.config.is_encoder_decoder=True.

Examples:

>>> from transformers import (
... AutoTokenizer,
... AutoModelForCausalLM,
... LogitsProcessorList,
... MinLengthLogitsProcessor,
... )

>>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
>>> model = AutoModelForCausalLM.from_pretrained("gpt2")

>>> # set pad_token_id to eos_token_id because GPT2 does not have a EOS token
>>> model.config.pad_token_id = model.config.eos_token_id

>>> input_prompt = "Today is a beautiful day, and"
>>> input_ids = tokenizer(input_prompt, return_tensors="pt").input_ids

>>> # instantiate logits processors
>>> logits_processor = LogitsProcessorList([
...     MinLengthLogitsProcessor(15, eos_token_id=model.config.eos_token_id),
... ])

>>> outputs = model.greedy_search(input_ids, logits_processor=logits_processor)

>>> print("Generated:", tokenizer.batch_decode(outputs, skip_special_tokens=True))

Generates sequences for models with a language modeling head using beam search decoding.

Parameters
  • input_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) – The sequence used as a prompt for the generation. If None the method initializes it as an empty torch.LongTensor of shape (1,).

  • beam_scorer (BeamScorer) – An derived instance of BeamScorer that defines how beam hypotheses are constructed, stored and sorted during generation. For more information, the documentation of BeamScorer should be read.

  • logits_processor (LogitsProcessorList, optional) – An instance of LogitsProcessorList. List of instances of class derived from LogitsProcessor used to modify the prediction scores of the language modeling head applied at each generation step.

  • stopping_criteria (StoppingCriteriaList, optional) – An instance of StoppingCriteriaList. List of instances of class derived from StoppingCriteria used to tell if the generation loop should stop.

  • max_length (int, optional, defaults to 20) – The maximum length of the sequence to be generated.

  • pad_token_id (int, optional) – The id of the padding token.

  • eos_token_id (int, optional) – The id of the end-of-sequence token.

  • output_attentions (bool, optional, defaults to False) – Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more details.

  • output_hidden_states (bool, optional, defaults to False) – Whether or not to return trhe hidden states of all layers. See hidden_states under returned tensors for more details.

  • output_scores (bool, optional, defaults to False) – Whether or not to return the prediction scores. See scores under returned tensors for more details.

  • return_dict_in_generate (bool, optional, defaults to False) – Whether or not to return a ModelOutput instead of a plain tuple.

  • model_kwargs – Additional model specific kwargs that will be forwarded to the forward function of the model. If model is an encoder-decoder model the kwargs should include encoder_outputs.

Returns

BeamSearchDecoderOnlyOutput, BeamSearchEncoderDecoderOutput or obj:torch.LongTensor: A torch.LongTensor containing the generated tokens (default behaviour) or a BeamSearchDecoderOnlyOutput if BeamSearchDecoderOnlyOutput if model.config.is_encoder_decoder=False and return_dict_in_generate=True or a BeamSearchEncoderDecoderOutput if model.config.is_encoder_decoder=True.

Examples:

>>> from transformers import (
...    AutoTokenizer,
...    AutoModelForSeq2SeqLM,
...    LogitsProcessorList,
...    MinLengthLogitsProcessor,
...    HammingDiversityLogitsProcessor,
...    BeamSearchScorer,
... )
>>> import torch

>>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
>>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")

>>> encoder_input_str = "translate English to German: How old are you?"
>>> encoder_input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids


>>> # lets run diverse beam search using 6 beams
>>> num_beams = 6
>>> # define decoder start token ids
>>> input_ids = torch.ones((num_beams, 1), device=model.device, dtype=torch.long)
>>> input_ids = input_ids * model.config.decoder_start_token_id

>>> # add encoder_outputs to model keyword arguments
>>> model_kwargs = {
...     "encoder_outputs": model.get_encoder()(encoder_input_ids.repeat_interleave(num_beams, dim=0), return_dict=True)
... }

>>> # instantiate beam scorer
>>> beam_scorer = BeamSearchScorer(
...     batch_size=1,
...     max_length=model.config.max_length,
...     num_beams=num_beams,
...     device=model.device,
...     num_beam_groups=3
... )

>>> # instantiate logits processors
>>> logits_processor = LogitsProcessorList([
...     HammingDiversityLogitsProcessor(5.5, num_beams=6, num_beam_groups=3),
...     MinLengthLogitsProcessor(5, eos_token_id=model.config.eos_token_id),
... ])

>>> outputs = model.group_beam_search(input_ids, beam_scorer, logits_processor=logits_processor, **model_kwargs)

>>> print("Generated:", tokenizer.batch_decode(outputs, skip_special_tokens=True))
prepare_inputs_for_generation(input_ids: torch.LongTensor, **kwargs) → Dict[str, Any][source]¶

Implement in subclasses of PreTrainedModel for custom behavior to prepare inputs in the generate method.

sample(input_ids: torch.LongTensor, logits_processor: Optional[transformers.generation_logits_process.LogitsProcessorList] = None, stopping_criteria: Optional[transformers.generation_stopping_criteria.StoppingCriteriaList] = None, logits_warper: Optional[transformers.generation_logits_process.LogitsProcessorList] = None, max_length: Optional[int] = None, pad_token_id: Optional[int] = None, eos_token_id: Optional[int] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, output_scores: Optional[bool] = None, return_dict_in_generate: Optional[bool] = None, **model_kwargs) → Union[transformers.generation_utils.SampleEncoderDecoderOutput, transformers.generation_utils.SampleDecoderOnlyOutput, torch.LongTensor][source]¶

Generates sequences for models with a language modeling head using multinomial sampling.

Parameters
  • input_ids (torch.LongTensor of shape (batch_size, sequence_length), optional) – The sequence used as a prompt for the generation. If None the method initializes it as an empty torch.LongTensor of shape (1,).

  • logits_processor (LogitsProcessorList, optional) – An instance of LogitsProcessorList. List of instances of class derived from LogitsProcessor used to modify the prediction scores of the language modeling head applied at each generation step.

  • stopping_criteria (StoppingCriteriaList, optional) – An instance of StoppingCriteriaList. List of instances of class derived from StoppingCriteria used to tell if the generation loop should stop.

  • logits_warper (LogitsProcessorList, optional) – An instance of LogitsProcessorList. List of instances of class derived from LogitsWarper used to warp the prediction score distribution of the language modeling head applied before multinomial sampling at each generation step.

  • max_length (int, optional, defaults to 20) – The maximum length of the sequence to be generated.

  • pad_token_id (int, optional) – The id of the padding token.

  • eos_token_id (int, optional) – The id of the end-of-sequence token.

  • output_attentions (bool, optional, defaults to False) – Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more details.

  • output_hidden_states (bool, optional, defaults to False) – Whether or not to return trhe hidden states of all layers. See hidden_states under returned tensors for more details.

  • output_scores (bool, optional, defaults to False) – Whether or not to return the prediction scores. See scores under returned tensors for more details.

  • return_dict_in_generate (bool, optional, defaults to False) – Whether or not to return a ModelOutput instead of a plain tuple.

  • model_kwargs – Additional model specific kwargs will be forwarded to the forward function of the model. If model is an encoder-decoder model the kwargs should include encoder_outputs.

Returns

SampleDecoderOnlyOutput, SampleEncoderDecoderOutput or obj:torch.LongTensor: A torch.LongTensor containing the generated tokens (default behaviour) or a SampleDecoderOnlyOutput if model.config.is_encoder_decoder=False and return_dict_in_generate=True or a SampleEncoderDecoderOutput if model.config.is_encoder_decoder=True.

Examples:

>>> from transformers import (
...    AutoTokenizer,
...    AutoModelForCausalLM,
...    LogitsProcessorList,
...    MinLengthLogitsProcessor,
...    TopKLogitsWarper,
...    TemperatureLogitsWarper,
... )

>>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
>>> model = AutoModelForCausalLM.from_pretrained("gpt2")

>>> # set pad_token_id to eos_token_id because GPT2 does not have a EOS token
>>> model.config.pad_token_id = model.config.eos_token_id

>>> input_prompt = "Today is a beautiful day, and"
>>> input_ids = tokenizer(input_prompt, return_tensors="pt").input_ids

>>> # instantiate logits processors
>>> logits_processor = LogitsProcessorList([
...     MinLengthLogitsProcessor(15, eos_token_id=model.config.eos_token_id),
... ])
>>> # instantiate logits processors
>>> logits_warper = LogitsProcessorList([
...     TopKLogitsWarper(50),
...     TemperatureLogitsWarper(0.7),
... ])

>>> outputs = model.sample(input_ids, logits_processor=logits_processor, logits_warper=logits_warper)

>>> print("Generated:", tokenizer.batch_decode(outputs, skip_special_tokens=True))
class transformers.generation_tf_utils.TFGenerationMixin[source]¶

A class containing all of the functions supporting generation, to be used as a mixin in TFPreTrainedModel.

adjust_logits_during_generation(logits, cur_len, max_length, forced_bos_token_id, forced_eos_token_id, **kwargs)[source]¶

Implement in subclasses of PreTrainedModel for custom behavior to adjust the logits in the generate method.

generate(input_ids=None, max_length=None, min_length=None, do_sample=None, early_stopping=None, num_beams=None, temperature=None, top_k=None, top_p=None, repetition_penalty=None, bad_words_ids=None, bos_token_id=None, pad_token_id=None, eos_token_id=None, length_penalty=None, no_repeat_ngram_size=None, num_return_sequences=None, attention_mask=None, decoder_start_token_id=None, use_cache=None, forced_bos_token_id=None, forced_eos_token_id=None)[source]¶

Generates sequences for models with a language modeling head. The method currently supports greedy decoding, beam-search decoding, sampling with temperature, sampling with top-k or nucleus sampling.

Adapted in part from Facebook’s XLM beam search code.

Apart from input_ids and attention_mask, all the arguments below will default to the value of the attribute of the same name inside the PretrainedConfig of the model. The default values indicated are the default values of those config.

Most of these parameters are explained in more detail in this blog post.

Parameters
  • input_ids (tf.Tensor of dtype=tf.int32 and shape (batch_size, sequence_length), optional) – The sequence used as a prompt for the generation. If None the method initializes it as an empty tf.Tensor of shape (1,).

  • max_length (int, optional, defaults to 20) – The maximum length of the sequence to be generated.

  • min_length (int, optional, defaults to 10) – The minimum length of the sequence to be generated.

  • do_sample (bool, optional, defaults to False) – Whether or not to use sampling ; use greedy decoding otherwise.

  • early_stopping (bool, optional, defaults to False) – Whether to stop the beam search when at least num_beams sentences are finished per batch or not.

  • num_beams (int, optional, defaults to 1) – Number of beams for beam search. 1 means no beam search.

  • temperature (float, optional, defaults to 1.0) – The value used to module the next token probabilities.

  • top_k (int, optional, defaults to 50) – The number of highest probability vocabulary tokens to keep for top-k-filtering.

  • top_p (float, optional, defaults to 1.0) – If set to float < 1, only the most probable tokens with probabilities that add up to top_p or higher are kept for generation.

  • repetition_penalty (float, optional, defaults to 1.0) – The parameter for repetition penalty. 1.0 means no penalty. See this paper for more details.

  • pad_token_id (int, optional) – The id of the padding token.

  • bos_token_id (int, optional) – The id of the beginning-of-sequence token.

  • eos_token_id (int, optional) – The id of the end-of-sequence token.

  • length_penalty (float, optional, defaults to 1.0) –

    Exponential penalty to the length. 1.0 means no penalty.

    Set to values < 1.0 in order to encourage the model to generate shorter sequences, to a value > 1.0 in order to encourage the model to produce longer sequences.

  • no_repeat_ngram_size (int, optional, defaults to 0) – If set to int > 0, all ngrams of that size can only occur once.

  • bad_words_ids (List[int], optional) – List of token ids that are not allowed to be generated. In order to get the tokens of the words that should not appear in the generated text, use tokenizer.encode(bad_word, add_prefix_space=True).

  • num_return_sequences (int, optional, defaults to 1) – The number of independently computed returned sequences for each element in the batch.

  • attention_mask (tf.Tensor of dtype=tf.int32 and shape (batch_size, sequence_length), optional) –

    Mask to avoid performing attention on padding token indices. Mask values are in [0, 1], 1 for tokens that are not masked, and 0 for masked tokens.

    If not provided, will default to a tensor the same shape as input_ids that masks the pad token.

    What are attention masks?

  • decoder_start_token_id (int, optional) – If an encoder-decoder model starts decoding with a different token than bos, the id of that token.

  • use_cache – (bool, optional, defaults to True): Whether or not the model should use the past last key/values attentions (if applicable to the model) to speed up decoding.

  • forced_bos_token_id (int, optional) – The id of the token to force as the first generated token after the decoder_start_token_id. Useful for multilingual models like mBART where the first generated token needs to be the target language token.

  • forced_eos_token_id (int, optional) – The id of the token to force as the last generated token when max_length is reached.

  • model_specific_kwargs – Additional model specific kwargs will be forwarded to the forward function of the model.

Returns

obj:(batch_size * num_return_sequences, sequence_length): The generated sequences. The second dimension (sequence_length) is either equal to max_length or shorter if all batches finished early due to the eos_token_id.

Return type

tf.Tensor of dtype=tf.int32 and shape

Examples:

tokenizer = AutoTokenizer.from_pretrained('distilgpt2')   # Initialize tokenizer
model = TFAutoModelWithLMHead.from_pretrained('distilgpt2')    # Download model and configuration from huggingface.co and cache.
outputs = model.generate(max_length=40)  # do greedy decoding
print('Generated: {}'.format(tokenizer.decode(outputs[0], skip_special_tokens=True)))

tokenizer = AutoTokenizer.from_pretrained('openai-gpt')   # Initialize tokenizer
model = TFAutoModelWithLMHead.from_pretrained('openai-gpt')    # Download model and configuration from huggingface.co and cache.
input_context = 'The dog'
input_ids = tokenizer.encode(input_context, return_tensors='tf')  # encode input context
outputs = model.generate(input_ids=input_ids, num_beams=5, num_return_sequences=3, temperature=1.5)  # generate 3 independent sequences using beam search decoding (5 beams) with sampling from initial context 'The dog'
for i in range(3): #  3 output sequences were generated
    print('Generated {}: {}'.format(i, tokenizer.decode(outputs[i], skip_special_tokens=True)))

tokenizer = AutoTokenizer.from_pretrained('distilgpt2')   # Initialize tokenizer
model = TFAutoModelWithLMHead.from_pretrained('distilgpt2')    # Download model and configuration from huggingface.co and cache.
input_context = 'The dog'
input_ids = tokenizer.encode(input_context, return_tensors='tf')  # encode input context
outputs = model.generate(input_ids=input_ids, max_length=40, temperature=0.7, num_return_sequences=3, do_sample=True)  # generate 3 candidates using sampling
for i in range(3): #  3 output sequences were generated
    print('Generated {}: {}'.format(i, tokenizer.decode(outputs[i], skip_special_tokens=True)))

tokenizer = AutoTokenizer.from_pretrained('ctrl')   # Initialize tokenizer
model = TFAutoModelWithLMHead.from_pretrained('ctrl')    # Download model and configuration from huggingface.co and cache.
input_context = 'Legal My neighbor is'  # "Legal" is one of the control codes for ctrl
input_ids = tokenizer.encode(input_context, return_tensors='tf')  # encode input context
outputs = model.generate(input_ids=input_ids, max_length=50, temperature=0.7, repetition_penalty=1.2)  # generate sequences
print('Generated: {}'.format(tokenizer.decode(outputs[0], skip_special_tokens=True)))

tokenizer = AutoTokenizer.from_pretrained('gpt2')   # Initialize tokenizer
model = TFAutoModelWithLMHead.from_pretrained('gpt2')    # Download model and configuration from huggingface.co and cache.
input_context = 'My cute dog'
bad_words_ids = [tokenizer.encode(bad_word, add_prefix_space=True) for bad_word in ['idiot', 'stupid', 'shut up']]
input_ids = tokenizer.encode(input_context, return_tensors='tf')  # encode input context
outputs = model.generate(input_ids=input_ids, max_length=100, do_sample=True, bad_words_ids=bad_words_ids)  # generate sequences without allowing bad_words to be generated
prepare_inputs_for_generation(inputs, **kwargs)[source]¶

Implement in subclasses of TFPreTrainedModel for custom behavior to prepare inputs in the generate method.