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: List[str] = None, keep_sentence_boundaries: bool = False, scalar_mix_parameters: List[float] = None, module: torch.nn.modules.module.Module = None)¶
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_representationsdifferent layers of ELMo representations. Typically
num_output_representationsis 1 or 2. For example, in the case of the SRL model in the above paper,
num_output_representations=1where ELMo was included at the input token representation layer. In the case of the SQuAD model,
num_output_representations=2as 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.
ELMo JSON options file
ELMo hdf5 weight file
- num_output_representations: ``int``, required.
The number of ELMo representation to output with different linear weighted combination of the 3 layers (i.e., character-convnet output, 1st lstm output, 2nd lstm output).
- requires_grad: ``bool``, optional
If True, compute gradient of ELMo parameters for fine tuning.
bool, optional, (default = False).
Should we apply layer normalization (passed to
float, optional, (default = 0.5).
The dropout to be applied to the ELMo representations.
List[str], optional, (default = None).
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.
bool, optional, (default = False)
If True, the representation of the sentence boundary tokens are not removed.
List[float], optional, (default = None)
None, use these scalar mix parameters to weight the representations produced by different layers. These mixing weights are not updated during training.
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
weight_file. The module must provide a public attribute
num_layerswith the number of internal layers and its
forwardmethod must return a
maskkeys (see _ElmoBilm` for an example). Note that
requires_gradis also ignored with this option.
forward(self, inputs:torch.Tensor, word_inputs:torch.Tensor=None) → Dict[str, Union[torch.Tensor, List[torch.Tensor]]]¶
- inputs: ``torch.Tensor``, required.
- Shape ``(batch_size, timesteps, 50)`` of character ids representing the current batch.
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.
- Dict with keys:
num_output_representationslist of ELMo representations for the input sequence. Each representation is shape
(batch_size, timesteps, embedding_dim)
(batch_size, timesteps)long tensor with sequence mask.
from_params(params:allennlp.common.params.Params) → 'Elmo'¶
batch_to_ids(batch:List[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).
A list of tokenized sentences.
- A tensor of padded character ids.