LanguageModel(vocab: allennlp.data.vocabulary.Vocabulary, text_field_embedder: allennlp.modules.text_field_embedders.text_field_embedder.TextFieldEmbedder, contextualizer: allennlp.modules.seq2seq_encoders.seq2seq_encoder.Seq2SeqEncoder, dropout: float = None, num_samples: int = None, sparse_embeddings: bool = False, bidirectional: bool = False, initializer: allennlp.nn.initializers.InitializerApplicator = None, regularizer: Optional[allennlp.nn.regularizers.regularizer_applicator.RegularizerApplicator] = None)¶
LanguageModelapplies a “contextualizing”
Seq2SeqEncoderto uncontextualized embeddings, using a
SoftmaxLossmodule (defined above) to compute the language modeling loss.
If bidirectional is True, the language model is trained to predict the next and previous tokens for each token in the input. In this case, the contextualizer must be bidirectional. If bidirectional is False, the language model is trained to only predict the next token for each token in the input; the contextualizer should also be unidirectional.
If your language model is bidirectional, it is IMPORTANT that your bidirectional
Seq2SeqEncodercontextualizer does not do any “peeking ahead”. That is, for its forward direction it should only consider embeddings at previous timesteps, and for its backward direction only embeddings at subsequent timesteps. Similarly, if your language model is unidirectional, the unidirectional contextualizer should only consider embeddings at previous timesteps. If this condition is not met, your language model is cheating.
- vocab: ``Vocabulary``
- text_field_embedder: ``TextFieldEmbedder``
Used to embed the indexed tokens we get in
- contextualizer: ``Seq2SeqEncoder``
Used to “contextualize” the embeddings. As described above, this encoder must not cheat by peeking ahead.
- dropout: ``float``, optional (default: None)
If specified, dropout is applied to the contextualized embeddings before computation of the softmax. The contextualized embeddings themselves are returned without dropout.
- num_samples: ``int``, optional (default: None)
If provided, the model will use
SampledSoftmaxLosswith the specified number of samples. Otherwise, it will use the full
- sparse_embeddings: ``bool``, optional (default: False)
Passed on to
- bidirectional: ``bool``, optional (default: False)
Train a bidirectional language model, where the contextualizer is used to predict the next and previous token for each input token. This must match the bidirectionality of the contextualizer.
RegularizerApplicator, optional (default=``None``)
If provided, will be used to calculate the regularization penalty during training.
delete_softmax(self) → None¶
Remove the softmax weights. Useful for saving memory when calculating the loss is not necessary, e.g. in an embedder.
forward(self, source:Dict[str, torch.LongTensor]) → Dict[str, torch.Tensor]¶
Computes the averaged forward (and backward, if language model is bidirectional) LM loss from the batch.
- source: ``Dict[str, torch.LongTensor]``, required.
The output of
Batch.as_tensor_dict()for a batch of sentences. By convention, it’s required to have at least a
"tokens"entry that’s the output of a
SingleIdTokenIndexer, which is used to compute the language model targets.
- Dict with keys:
forward negative log likelihood, or the average of forward/backward if language model is bidirectional
forward direction negative log likelihood
backward direction negative log likelihood. If language model is not bidirectional, this is
(batch_size, timesteps, embed_dim) tensor of top layer contextual representations or list of all layers. No dropout applied.
(batch_size, timesteps, token_embed_dim) tensor of bottom layer noncontextual representations
(batch_size, timesteps) mask for the embeddings
Returns a dictionary of metrics. This method will be called by
allennlp.training.Trainerin order to compute and use model metrics for early stopping and model serialization. We return an empty dictionary here rather than raising as it is not required to implement metrics for a new model. A boolean reset parameter is passed, as frequently a metric accumulator will have some state which should be reset between epochs. This is also compatible with
Metrics should be populated during the call to ``forward`, with the
Metrichandling the accumulation of the metric until this method is called.
num_layers(self) → int¶
Returns the depth of this LM. That is, how many layers the contextualizer has plus one for the non-contextual layer.