diff --git a/docs/source/model_doc/auto.rst b/docs/source/model_doc/auto.rst new file mode 100644 index 0000000000..ad439fff03 --- /dev/null +++ b/docs/source/model_doc/auto.rst @@ -0,0 +1,26 @@ +AutoModel, AutoConfig and AutoTokenizer - Standard derived classes +--------------------------------------------------------------- + +In many case, the architecture you want to use can be guessed from the name or the path of the pretrained model you are supplying to the ``from_pretrained`` method. + +Auto classes are here to do this job for you so that you automatically retreive the relevant model given the name/path to the pretrained weights/config/vocabulary. + +``AutoConfig`` +~~~~~~~~~~~~~~~~~~~~~ + +.. autoclass:: pytorch_transformers.AutoConfig + :members: + + +``AutoModel`` +~~~~~~~~~~~~~~~~~~~~~ + +.. autoclass:: pytorch_transformers.AutoModel + :members: + + +``AutoTokenizer`` +~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ + +.. autoclass:: pytorch_transformers.AutoTokenizer + :members: diff --git a/examples/run_squad.py b/examples/run_squad.py index 36e03fb012..937aa154ea 100644 --- a/examples/run_squad.py +++ b/examples/run_squad.py @@ -134,7 +134,7 @@ def train(args, train_dataset, model, tokenizer): 'end_positions': batch[4]} if args.model_type in ['xlnet', 'xlm']: inputs.update({'cls_index': batch[5], - 'p_mask': batch[6]}) + 'p_mask': batch[6]}) outputs = model(**inputs) loss = outputs[0] # model outputs are always tuple in pytorch-transformers (see doc) diff --git a/pytorch_transformers/modeling_auto.py b/pytorch_transformers/modeling_auto.py index 3e28fbd0a9..7d3ea7ec60 100644 --- a/pytorch_transformers/modeling_auto.py +++ b/pytorch_transformers/modeling_auto.py @@ -18,6 +18,11 @@ from __future__ import absolute_import, division, print_function, unicode_litera import logging +import torch +import torch.nn as nn +from torch.nn import CrossEntropyLoss, MSELoss +from torch.nn.parameter import Parameter + from .modeling_bert import BertConfig, BertModel from .modeling_openai import OpenAIGPTConfig, OpenAIGPTModel from .modeling_gpt2 import GPT2Config, GPT2Model @@ -25,6 +30,8 @@ from .modeling_transfo_xl import TransfoXLConfig, TransfoXLModel from .modeling_xlnet import XLNetConfig, XLNetModel from .modeling_xlm import XLMConfig, XLMModel +from .modeling_utils import PreTrainedModel, SequenceSummary + logger = logging.getLogger(__name__) class AutoConfig(object): @@ -228,3 +235,245 @@ class AutoModel(object): "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', " "'xlm'".format(pretrained_model_name_or_path)) + +class DerivedAutoModel(PreTrainedModel): + r""" + :class:`~pytorch_transformers.DerivedAutoModel` is a base class for building + standardized derived models on top of :class:`~pytorch_transformers.AutoModel` by adding heads + + The `from_pretrained()` method take care of using the correct base model class instance + using pattern matching on the `pretrained_model_name_or_path` string. + + The base model class to instantiate is selected as the first pattern matching + in the `pretrained_model_name_or_path` string (in the following order): + - contains `bert`: BertConfig (Bert model) + - contains `openai-gpt`: OpenAIGPTConfig (OpenAI GPT model) + - contains `gpt2`: GPT2Config (OpenAI GPT-2 model) + - contains `transfo-xl`: TransfoXLConfig (Transformer-XL model) + - contains `xlnet`: XLNetConfig (XLNet model) + - contains `xlm`: XLMConfig (XLM model) + + This class should usually not be instantiated using `__init__()` but `from_pretrained()`. + """ + config_class = None + pretrained_model_archive_map = {} + load_tf_weights = lambda model, config, path: None + base_model_prefix = "transformer" + + def __init__(self, base_model): + super(DerivedAutoModel, self).__init__(base_model.config) + self.transformer = base_model + + def init_weights(self, module): + """ Initialize the weights. Use the base model initialization function. + """ + self.transformer.init_weights(module) + + @classmethod + def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs): + r""" Instantiate a :class:`~pytorch_transformers.DerivedAutoModel` with one of the base model classes of the library + from a pre-trained model configuration. + + The base model class to instantiate is selected as the first pattern matching + in the `pretrained_model_name_or_path` string (in the following order): + - contains `bert`: BertConfig (Bert model) + - contains `openai-gpt`: OpenAIGPTConfig (OpenAI GPT model) + - contains `gpt2`: GPT2Config (OpenAI GPT-2 model) + - contains `transfo-xl`: TransfoXLConfig (Transformer-XL model) + - contains `xlnet`: XLNetConfig (XLNet model) + - contains `xlm`: XLMConfig (XLM model) + + 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()` + + Params: + **pretrained_model_name_or_path**: either: + - a string with the `shortcut name` of a pre-trained model to load from cache + or download and cache if not already stored in cache (e.g. 'bert-base-uncased'). + - a path to a `directory` containing a configuration file saved + using the `save_pretrained(save_directory)` method. + - 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 option is slower than converting the TensorFlow + checkpoint in a PyTorch model using the provided conversion scripts and loading + the PyTorch model afterwards. + **model_args**: (`optional`) Sequence: + All remaning positional arguments will be passed to the underlying model's __init__ function + **config**: an optional 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 a `shortcut name` of a pre-trained model), or + - the model was saved using the `save_pretrained(save_directory)` (loaded by suppling the save directory). + **state_dict**: an optional state dictionnary for the model 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(dir)` and `from_pretrained(save_directory)` is not + a simpler option. + **cache_dir**: (`optional`) string: + Path to a directory in which a downloaded pre-trained model + configuration should be cached if the standard cache should not be used. + **output_loading_info**: (`optional`) boolean: + Set to ``True`` to also return a dictionnary containing missing keys, unexpected keys and error messages. + **kwargs**: (`optional`) dict: + Dictionary of key, values to update the configuration object after loading. + Can be used to override selected configuration parameters. E.g. ``output_attention=True``. + + - If a configuration is provided with `config`, **kwargs will be directly passed + to the underlying model's __init__ method. + - If a configuration is not provided, **kwargs will be first passed to the pretrained + model configuration class loading function (`PretrainedConfig.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:: + + model = AutoModel.from_pretrained('bert-base-uncased') # Download model and configuration from S3 and cache. + model = AutoModel.from_pretrained('./test/bert_model/') # E.g. model was saved using `save_pretrained('./test/saved_model/')` + model = AutoModel.from_pretrained('bert-base-uncased', output_attention=True) # Update configuration during loading + assert model.config.output_attention == True + # Loading from a TF checkpoint file instead of a PyTorch model (slower) + config = AutoConfig.from_json_file('./tf_model/bert_tf_model_config.json') + model = AutoModel.from_pretrained('./tf_model/bert_tf_checkpoint.ckpt.index', from_tf=True, config=config) + + """ + if 'bert' in pretrained_model_name_or_path: + base_model_class = BertModel + elif 'openai-gpt' in pretrained_model_name_or_path: + base_model_class = OpenAIGPTModel + elif 'gpt2' in pretrained_model_name_or_path: + base_model_class = GPT2Model + elif 'transfo-xl' in pretrained_model_name_or_path: + base_model_class = TransfoXLModel + elif 'xlnet' in pretrained_model_name_or_path: + base_model_class = XLNetModel + elif 'xlm' in pretrained_model_name_or_path: + base_model_class = XLMModel + else: + raise ValueError("Unrecognized model identifier in {}. Should contains one of " + "'bert', 'openai-gpt', 'gpt2', 'transfo-xl', 'xlnet', " + "'xlm'".format(pretrained_model_name_or_path)) + + # Get a pretrained base_model + base_model = base_model_class.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs) + + # Create our derived model + model = cls(base_model) + + # Setup class attribute from the base model class + model.config_class = base_model.config_class + model.pretrained_model_archive_map = base_model.pretrained_model_archive_map + model.load_tf_weights = base_model.load_tf_weights + + return model + + +class AutoModelWithLMHead(DerivedAutoModel): + r""" + :class:`~pytorch_transformers.AutoModelWithLMHead` is a base class for language modeling + that contains + + - a base model instantiated as one of the base model classes of the library when created with the `AutoModel.from_pretrained(pretrained_model_name_or_path)` class method, and . + - a language modeling head on top of the base model. + + The `from_pretrained()` method take care of using the correct base model class instance + using pattern matching on the `pretrained_model_name_or_path` string. + + The base model class to instantiate is selected as the first pattern matching + in the `pretrained_model_name_or_path` string (in the following order): + - contains `bert`: BertConfig (Bert model) + - contains `openai-gpt`: OpenAIGPTConfig (OpenAI GPT model) + - contains `gpt2`: GPT2Config (OpenAI GPT-2 model) + - contains `transfo-xl`: TransfoXLConfig (Transformer-XL model) + - contains `xlnet`: XLNetConfig (XLNet model) + - contains `xlm`: XLMConfig (XLM model) + + This class should usually not be instantiated using `__init__()` but `from_pretrained()`. + """ + + def __init__(self, base_model): + super(AutoModelWithLMHead, self).__init__(base_model) + self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) + + self.apply(self.init_weights) + self.tie_weights() + + def tie_weights(self): + """ Make sure we are sharing the input and output embeddings. + Export to TorchScript can't handle parameter sharing so we are cloning them instead. + """ + # get input embeddings - whatever the model is + input_embeddings = self.transformer.resize_token_embeddings(new_num_tokens=None) + + # tie of clone (torchscript) embeddings + self._tie_or_clone_weights(self.lm_head, input_embeddings) + + def forward(self, input_ids, **kwargs): + labels = kwargs.pop('labels', None) # Python 2 compatibility... + + transformer_outputs = self.transformer(input_ids, **kwargs) + hidden_states = transformer_outputs[0] + + lm_logits = self.lm_head(hidden_states) + + outputs = (lm_logits,) + transformer_outputs[1:] + if labels is not None: + loss_fct = CrossEntropyLoss(ignore_index=-1) + loss = loss_fct(lm_logits.view(-1, lm_logits.size(-1)), + labels.view(-1)) + outputs = (loss,) + outputs + + return outputs # (loss), lm_logits, presents, (all hidden_states), (attentions) + + +class AutoModelForSequenceClassification(DerivedAutoModel): + r""" + :class:`~pytorch_transformers.AutoModelForSequenceClassification` is a class for sequence classification + that contains + + - a base model instantiated as one of the base model classes of the library when created with the `AutoModel.from_pretrained(pretrained_model_name_or_path)` class method, and . + - a classification head on top of the base model. + + The `from_pretrained()` method take care of using the correct base model class instance + using pattern matching on the `pretrained_model_name_or_path` string. + + The base model class to instantiate is selected as the first pattern matching + in the `pretrained_model_name_or_path` string (in the following order): + - contains `bert`: BertConfig (Bert model) + - contains `openai-gpt`: OpenAIGPTConfig (OpenAI GPT model) + - contains `gpt2`: GPT2Config (OpenAI GPT-2 model) + - contains `transfo-xl`: TransfoXLConfig (Transformer-XL model) + - contains `xlnet`: XLNetConfig (XLNet model) + - contains `xlm`: XLMConfig (XLM model) + + This class should usually not be instantiated using `__init__()` but `from_pretrained()`. + """ + + def __init__(self, base_model): + super(AutoModelForSequenceClassification, self).__init__(base_model) + self.num_labels = base_model.config.num_labels + self.sequence_summary = SequenceSummary(base_model.config) + + self.apply(self.init_weights) + + def forward(self, input_ids, cls_index, **kwargs): + labels = kwargs.pop('labels', None) # Python 2 compatibility... + + transformer_outputs = self.transformer(input_ids, **kwargs) + + output = transformer_outputs[0] + logits = self.sequence_summary(output, cls_index=cls_index) + + outputs = (logits,) + transformer_outputs[1:] # Keep new_mems and attention/hidden states if they are here + + if labels is not None: + if self.num_labels == 1: + # We are doing regression + loss_fct = MSELoss() + loss = loss_fct(logits.view(-1), labels.view(-1)) + else: + loss_fct = CrossEntropyLoss() + loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1)) + outputs = (loss,) + outputs + + return outputs diff --git a/pytorch_transformers/modeling_gpt2.py b/pytorch_transformers/modeling_gpt2.py index 4341f0d8a1..5268c5de7d 100644 --- a/pytorch_transformers/modeling_gpt2.py +++ b/pytorch_transformers/modeling_gpt2.py @@ -137,7 +137,7 @@ class GPT2Config(PretrainedConfig): initializer_range=0.02, num_labels=1, - summary_type='token_ids', + summary_type='cls_index', summary_use_proj=True, summary_activation=None, summary_proj_to_labels=True, diff --git a/pytorch_transformers/modeling_openai.py b/pytorch_transformers/modeling_openai.py index a6cb6212ef..c51023444d 100644 --- a/pytorch_transformers/modeling_openai.py +++ b/pytorch_transformers/modeling_openai.py @@ -171,7 +171,7 @@ class OpenAIGPTConfig(PretrainedConfig): predict_special_tokens=True, num_labels=1, - summary_type='token_ids', + summary_type='cls_index', summary_use_proj=True, summary_activation=None, summary_proj_to_labels=True, diff --git a/pytorch_transformers/modeling_utils.py b/pytorch_transformers/modeling_utils.py index 8970cd56f8..2664c542e0 100644 --- a/pytorch_transformers/modeling_utils.py +++ b/pytorch_transformers/modeling_utils.py @@ -765,7 +765,7 @@ class SequenceSummary(nn.Module): - 'last' => [default] take the last token hidden state (like XLNet) - 'first' => take the first token hidden state (like Bert) - 'mean' => take the mean of all tokens hidden states - - 'token_ids' => supply a Tensor of classification token indices (GPT/GPT-2) + - 'cls_index' => supply a Tensor of classification token position (GPT/GPT-2) - 'attn' => Not implemented now, use multi-head attention summary_use_proj: Add a projection after the vector extraction summary_proj_to_labels: If True, the projection outputs to config.num_labels classes (otherwise to hidden_size). Default: False. @@ -803,11 +803,11 @@ class SequenceSummary(nn.Module): if hasattr(config, 'summary_last_dropout') and config.summary_last_dropout > 0: self.last_dropout = nn.Dropout(config.summary_last_dropout) - def forward(self, hidden_states, token_ids=None): + def forward(self, hidden_states, cls_index=None): """ hidden_states: float Tensor in shape [bsz, seq_len, hidden_size], the hidden-states of the last layer. - token_ids: [optional] index of the classification token if summary_type == 'token_ids', + cls_index: [optional] position of the classification token if summary_type == 'cls_index', shape (bsz,) or more generally (bsz, ...) where ... are optional leading dimensions of hidden_states. - if summary_type == 'token_ids' and token_ids is None: + if summary_type == 'cls_index' and cls_index is None: we take the last token of the sequence as classification token """ if self.summary_type == 'last': @@ -816,14 +816,14 @@ class SequenceSummary(nn.Module): output = hidden_states[:, 0] elif self.summary_type == 'mean': output = hidden_states.mean(dim=1) - elif self.summary_type == 'token_ids': - if token_ids is None: - token_ids = torch.full_like(hidden_states[..., :1, :], hidden_states.shape[-2]-1, dtype=torch.long) + elif self.summary_type == 'cls_index': + if cls_index is None: + cls_index = torch.full_like(hidden_states[..., :1, :], hidden_states.shape[-2]-1, dtype=torch.long) else: - token_ids = token_ids.unsqueeze(-1).unsqueeze(-1) - token_ids = token_ids.expand((-1,) * (token_ids.dim()-1) + (hidden_states.size(-1),)) - # shape of token_ids: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states - output = hidden_states.gather(-2, token_ids).squeeze(-2) # shape (bsz, XX, hidden_size) + cls_index = cls_index.unsqueeze(-1).unsqueeze(-1) + cls_index = cls_index.expand((-1,) * (cls_index.dim()-1) + (hidden_states.size(-1),)) + # shape of cls_index: (bsz, XX, 1, hidden_size) where XX are optional leading dim of hidden_states + output = hidden_states.gather(-2, cls_index).squeeze(-2) # shape (bsz, XX, hidden_size) elif self.summary_type == 'attn': raise NotImplementedError