espnet2.lm package

espnet2.lm.espnet_model

class espnet2.lm.espnet_model.ESPnetLanguageModel(lm: espnet2.lm.abs_model.AbsLM, vocab_size: int, ignore_id: int = 0)

Bases: espnet2.train.abs_espnet_model.AbsESPnetModel

batchify_nll(text: torch.Tensor, text_lengths: torch.Tensor, batch_size: int = 100) → Tuple[torch.Tensor, torch.Tensor]

Compute negative log likelihood(nll) from transformer language model

To avoid OOM, this fuction seperate the input into batches. Then call nll for each batch and combine and return results. :param text: (Batch, Length) :param text_lengths: (Batch,) :param batch_size: int, samples each batch contain when computing nll,

you may change this to avoid OOM or increase

collect_feats(text: torch.Tensor, text_lengths: torch.Tensor, **kwargs) → Dict[str, torch.Tensor]

forward(text: torch.Tensor, text_lengths: torch.Tensor, **kwargs) → Tuple[torch.Tensor, Dict[str, torch.Tensor], torch.Tensor]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

nll(text: torch.Tensor, text_lengths: torch.Tensor, max_length: Optional[int] = None) → Tuple[torch.Tensor, torch.Tensor]

Compute negative log likelihood(nll)

Normally, this function is called in batchify_nll. :param text: (Batch, Length) :param text_lengths: (Batch,) :param max_lengths: int

espnet2.lm.seq_rnn_lm

Sequential implementation of Recurrent Neural Network Language Model.

class espnet2.lm.seq_rnn_lm.SequentialRNNLM(vocab_size: int, unit: int = 650, nhid: int = None, nlayers: int = 2, dropout_rate: float = 0.0, tie_weights: bool = False, rnn_type: str = 'lstm', ignore_id: int = 0)

Bases: espnet2.lm.abs_model.AbsLM

Sequential RNNLM.

See also

https://github.com/pytorch/examples/blob/4581968193699de14b56527296262dd76ab43557/word_language_model/model.py

batch_score(ys: torch.Tensor, states: torch.Tensor, xs: torch.Tensor) → Tuple[torch.Tensor, torch.Tensor]

Score new token batch.

Parameters:

• ys (torch.Tensor) – torch.int64 prefix tokens (n_batch, ylen).

• states (List[Any]) – Scorer states for prefix tokens.

• xs (torch.Tensor) – The encoder feature that generates ys (n_batch, xlen, n_feat).

Returns:

Tuple of

batchfied scores for next token with shape of (n_batch, n_vocab) and next state list for ys.

Return type:

tuple[torch.Tensor, List[Any]]

forward(input: torch.Tensor, hidden: torch.Tensor) → Tuple[torch.Tensor, torch.Tensor]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

score(y: torch.Tensor, state: Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]], x: torch.Tensor) → Tuple[torch.Tensor, Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]]

Score new token.

Parameters:

• y – 1D torch.int64 prefix tokens.

• state – Scorer state for prefix tokens

• x – 2D encoder feature that generates ys.

Returns:

Tuple of

torch.float32 scores for next token (n_vocab) and next state for ys

zero_state()

Initialize LM state filled with zero values.

espnet2.lm.transformer_lm

class espnet2.lm.transformer_lm.TransformerLM(vocab_size: int, pos_enc: str = None, embed_unit: int = 128, att_unit: int = 256, head: int = 2, unit: int = 1024, layer: int = 4, dropout_rate: float = 0.5)

Bases: espnet2.lm.abs_model.AbsLM

batch_score(ys: torch.Tensor, states: List[Any], xs: torch.Tensor) → Tuple[torch.Tensor, List[Any]]

Score new token batch.

Parameters:

• ys (torch.Tensor) – torch.int64 prefix tokens (n_batch, ylen).

• states (List[Any]) – Scorer states for prefix tokens.

• xs (torch.Tensor) – The encoder feature that generates ys (n_batch, xlen, n_feat).

Returns:

Tuple of

batchfied scores for next token with shape of (n_batch, vocab_size) and next state list for ys.

Return type:

tuple[torch.Tensor, List[Any]]

forward(input: torch.Tensor, hidden: None) → Tuple[torch.Tensor, None]

Compute LM loss value from buffer sequences.

Parameters:

• input (torch.Tensor) – Input ids. (batch, len)

• hidden (torch.Tensor) – Target ids. (batch, len)

score(y: torch.Tensor, state: Any, x: torch.Tensor) → Tuple[torch.Tensor, Any]

Score new token.

Parameters:

• y (torch.Tensor) – 1D torch.int64 prefix tokens.

• state – Scorer state for prefix tokens

• x (torch.Tensor) – encoder feature that generates ys.

Returns:

Tuple of

torch.float32 scores for next token (vocab_size) and next state for ys

Return type:

tuple[torch.Tensor, Any]

espnet2.lm.abs_model

class espnet2.lm.abs_model.AbsLM

Bases: torch.nn.modules.module.Module, espnet.nets.scorer_interface.BatchScorerInterface, abc.ABC

The abstract LM class

To share the loss calculation way among different models, We uses delegate pattern here: The instance of this class should be passed to “LanguageModel”

>>> from espnet2.lm.abs_model import AbsLM
>>> lm = AbsLM()
>>> model = LanguageESPnetModel(lm=lm)

This “model” is one of mediator objects for “Task” class.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

abstract forward(input: torch.Tensor, hidden: torch.Tensor) → Tuple[torch.Tensor, torch.Tensor]

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

espnet2.lm.__init__