allennlp.modules.elmo

class allennlp.modules.elmo.Elmo(options_file: str, weight_file: str, num_output_representations: int, requires_grad: bool = False, do_layer_norm: bool = False, dropout: float = 0.5, vocab_to_cache: typing.List[str] = None, module: torch.nn.modules.module.Module = None) → None

Bases: torch.nn.modules.module.Module

Compute ELMo representations using a pre-trained bidirectional language model.

See “Deep contextualized word representations”, Peters et al. for details.

This module takes character id input and computes num_output_representations different layers of ELMo representations. Typically num_output_representations is 1 or 2. For example, in the case of the SRL model in the above paper, num_output_representations=1 where ELMo was included at the input token representation layer. In the case of the SQuAD model, num_output_representations=2 as ELMo was also included at the GRU output layer.

In the implementation below, we learn separate scalar weights for each output layer, but only run the biLM once on each input sequence for efficiency.

Parameters:

options_file : str, required.

ELMo JSON options file

weight_file : str, required.

ELMo hdf5 weight file

num_output_representations: ``int``, required.

The number of ELMo representation layers to output.

requires_grad: ``bool``, optional

If True, compute gradient of ELMo parameters for fine tuning.

do_layer_norm : bool, optional, (default=False).

Should we apply layer normalization (passed to ScalarMix)?

dropout : float, optional, (default = 0.5).

The dropout to be applied to the ELMo representations.

vocab_to_cache : List[str], optional, (default = 0.5).

A list of words to pre-compute and cache character convolutions for. If you use this option, Elmo expects that you pass word indices of shape (batch_size, timesteps) to forward, instead of character indices. If you use this option and pass a word which wasn’t pre-cached, this will break.

module : torch.nn.Module, optional, (default = None).

If provided, then use this module instead of the pre-trained ELMo biLM. If using this option, then pass None for both options_file and weight_file. The module must provide a public attribute num_layers with the number of internal layers and its forward method must return a dict with activations and mask keys (see _ElmoBilm` for an example). Note that requires_grad is also ignored with this option.

forward(inputs: torch.Tensor, word_inputs: torch.Tensor = None) → typing.Dict[str, typing.Union[torch.Tensor, typing.List[torch.Tensor]]]

Parameters:

inputs: ``torch.Tensor``, required.

Shape ``(batch_size, timesteps, 50)`` of character ids representing the current batch.

word_inputs : torch.Tensor, required.

If you passed a cached vocab, you can in addition pass a tensor of shape (batch_size, timesteps), which represent word ids which have been pre-cached.

Returns:

Dict with keys:

``’elmo_representations’``: ``List[torch.Tensor]``

A num_output_representations list of ELMo representations for the input sequence. Each representation is shape (batch_size, timesteps, embedding_dim)

``’mask’``: ``torch.Tensor``

Shape (batch_size, timesteps) long tensor with sequence mask.

classmethod from_params(params: allennlp.common.params.Params) → allennlp.modules.elmo.Elmo

get_output_dim()

allennlp.modules.elmo.batch_to_ids(batch: typing.List[typing.List[str]]) → torch.Tensor

Converts a batch of tokenized sentences to a tensor representing the sentences with encoded characters (len(batch), max sentence length, max word length).

Parameters:

batch : List[List[str]], required

A list of tokenized sentences.

Returns:

A tensor of padded character ids.