espnet.nets.pytorch_backend.transducer.vgg2l.VGG2L

source

class espnet.nets.pytorch_backend.transducer.vgg2l.VGG2L(idim: int, odim: int, pos_enc: Module | None = None)

Bases: Module

VGG2L module for custom encoder.

• Parameters:
  • idim – Input dimension.
  • odim – Output dimension.
  • pos_enc – Positional encoding class.

Construct a VGG2L object.

create_new_mask(feats_mask: Tensor) → Tensor

Create a subsampled mask of feature sequences.

• Parameters:feats_mask – Mask of feature sequences. (B, 1, F)
• Returns: Mask of VGG2L output sequences. (B, 1, sub(F))
• Return type: vgg_mask

forward(feats: Tensor, feats_mask: Tensor) → Tuple[Tensor, Tensor] | Tuple[Tuple[Tensor, Tensor], Tensor]

Forward VGG2L bottleneck.

• Parameters:

  • feats – Feature sequences. (B, F, D_feats)
  • feats_mask – Mask of feature sequences. (B, 1, F)

• Returns: VGG output sequences. : (B, sub(F), D_out) or ((B, sub(F), D_out), (B, sub(F), D_att))

  vgg_mask: Mask of VGG output sequences. (B, 1, sub(F))

• Return type: vgg_output