espnet.nets.pytorch_backend.e2e_asr.E2E

source

class espnet.nets.pytorch_backend.e2e_asr.E2E(idim, odim, args)

Bases: ASRInterface, Module

E2E module.

• Parameters:
  • idim (int) – dimension of inputs
  • odim (int) – dimension of outputs
  • args (Namespace) – argument Namespace containing options

Construct an E2E object.

• Parameters:
  • idim (int) – dimension of inputs
  • odim (int) – dimension of outputs
  • args (Namespace) – argument Namespace containing options

static add_arguments(parser)

Add arguments.

static attention_add_arguments(parser)

Add arguments for the attention.

calculate_all_attentions(xs_pad, ilens, ys_pad)

E2E attention calculation.

• Parameters:
  • xs_pad (torch.Tensor) – batch of padded input sequences (B, Tmax, idim)
  • ilens (torch.Tensor) – batch of lengths of input sequences (B)
  • ys_pad (torch.Tensor) – batch of padded token id sequence tensor (B, Lmax)
• Returns: attention weights with the following shape,
  1. multi-head case => attention weights (B, H, Lmax, Tmax),
  2. other case => attention weights (B, Lmax, Tmax).
• Return type: float ndarray

calculate_all_ctc_probs(xs_pad, ilens, ys_pad)

E2E CTC probability calculation.

• Parameters:
  • xs_pad (torch.Tensor) – batch of padded input sequences (B, Tmax)
  • ilens (torch.Tensor) – batch of lengths of input sequences (B)
  • ys_pad (torch.Tensor) – batch of padded token id sequence tensor (B, Lmax)
• Returns: CTC probability (B, Tmax, vocab)
• Return type: float ndarray

static decoder_add_arguments(parser)

Add arguments for the decoder.

encode(x)

Encode acoustic features.

• Parameters:x (ndarray) – input acoustic feature (T, D)
• Returns: encoder outputs
• Return type: torch.Tensor

static encoder_add_arguments(parser)

Add arguments for the encoder.

enhance(xs)

Forward only in the frontend stage.

• Parameters:xs (ndarray) – input acoustic feature (T, C, F)
• Returns: enhaned feature
• Return type: torch.Tensor

forward(xs_pad, ilens, ys_pad)

E2E forward.

• Parameters:
  • xs_pad (torch.Tensor) – batch of padded input sequences (B, Tmax, idim)
  • ilens (torch.Tensor) – batch of lengths of input sequences (B)
  • ys_pad (torch.Tensor) – batch of padded token id sequence tensor (B, Lmax)
• Returns: loss value
• Return type: torch.Tensor

get_total_subsampling_factor()

Get total subsampling factor.

init_like_chainer()

Initialize weight like chainer.

chainer basically uses LeCun way: W ~ Normal(0, fan_in ** -0.5), b = 0 pytorch basically uses W, b ~ Uniform(-fan_in**-0.5, fan_in**-0.5) however, there are two exceptions as far as I know.

• EmbedID.W ~ Normal(0, 1)
• LSTM.upward.b[forget_gate_range] = 1 (but not used in NStepLSTM)

recognize(x, recog_args, char_list, rnnlm=None)

E2E beam search.

• Parameters:
  • x (ndarray) – input acoustic feature (T, D)
  • recog_args (Namespace) – argument Namespace containing options
  • char_list (list) – list of characters
  • rnnlm (torch.nn.Module) – language model module
• Returns: N-best decoding results
• Return type: list

recognize_batch(xs, recog_args, char_list, rnnlm=None)

E2E batch beam search.

• Parameters:
  • xs (list) – list of input acoustic feature arrays [(T_1, D), (T_2, D), …]
  • recog_args (Namespace) – argument Namespace containing options
  • char_list (list) – list of characters
  • rnnlm (torch.nn.Module) – language model module
• Returns: N-best decoding results
• Return type: list

scorers()

Scorers.

subsample_frames(x)

Subsample speeh frames in the encoder.