espnet2.legacy.nets.scorers.ctc.CTCPrefixScorer

source

class espnet2.legacy.nets.scorers.ctc.CTCPrefixScorer(ctc: Module, eos: int, window_margin: int = 0)

Bases: BatchPartialScorerInterface

Decoder interface wrapper for CTCPrefixScore.

Initialize class.

• Parameters:
  • ctc (torch.nn.Module) – The CTC implementation. For example, espnet2.legacy.nets.pytorch_backend.ctc.CTC
  • eos (int) – The end-of-sequence id.
  • window_margin (int) – Half-width, in encoder frames, of the window the CTC forward recursion is restricted to. 0 (default) walks the whole utterance at every decoding step, which is exact but costs O(T) per step. A positive value centres a window on the frame the prefix has reached and is an approximation.

batch_init_state(x: Tensor, xs_lengths: Tensor = None)

Get an initial state for decoding.

• Parameters:
  • x (torch.Tensor) – The encoded feature tensor. Either a single utterance (T, D) or a batch (B, T, D).
  • xs_lengths (torch.Tensor) – Encoder output lengths (B,). Required when x is a padded batch; the padded frames are excluded from the CTC prefix scores.

Returns: initial state

batch_score_partial(y, ids, state, x)

Score new token.

• Parameters:
  • y (torch.Tensor) – 1D prefix token
  • ids (torch.Tensor) – torch.int64 next token to score
  • state – decoder state for prefix tokens
  • x (torch.Tensor) – 2D encoder feature that generates ys
• Returns: Tuple of a score tensor for y that has a shape (len(next_tokens),) and next state for ys
• Return type: tuple[torch.Tensor, Any]

batch_select_state(state, best_ids: Tensor)

Select states of a whole (n_utt, beam) hypothesis grid at once.

This is the vectorized counterpart of calling select_state() once per hypothesis, used by espnet2.legacy.nets.batch_beam_search.BatchBeamSearch.

• Parameters:
  • state – The state returned by batch_score_partial().
  • best_ids (torch.Tensor) – (n_utt, beam) tensor of pruned candidates encoded as beam_index * odim + token_index.

Returns: the reordered state, in the batched form that : batch_score_partial() accepts directly.

extend_prob(x: Tensor)

Extend probs for decoding.

This extension is for streaming decoding as in Eq (14) in https://arxiv.org/abs/2006.14941

• Parameters:x (torch.Tensor) – The encoded feature tensor

extend_state(state)

Extend state for decoding.

This extension is for streaming decoding as in Eq (14) in https://arxiv.org/abs/2006.14941

• Parameters:state – The states of hyps

Returns: exteded state

init_state(x: Tensor)

Get an initial state for decoding.

• Parameters:x (torch.Tensor) – The encoded feature tensor

Returns: initial state

score_partial(y, ids, state, x)

Score new token.

• Parameters:
  • y (torch.Tensor) – 1D prefix token
  • next_tokens (torch.Tensor) – torch.int64 next token to score
  • state – decoder state for prefix tokens
  • x (torch.Tensor) – 2D encoder feature that generates ys
• Returns: Tuple of a score tensor for y that has a shape (len(next_tokens),) and next state for ys
• Return type: tuple[torch.Tensor, Any]

select_state(state, i, new_id=None)

Select state with relative ids in the main beam search.

• Parameters:
  • state – Decoder state for prefix tokens
  • i (int) – Index to select a state in the main beam search
  • new_id (int) – New label id to select a state if necessary
• Returns: pruned state
• Return type: state