espnet2.legacy.nets.pytorch_backend.conformer.convolution.ConvolutionModule

source

class espnet2.legacy.nets.pytorch_backend.conformer.convolution.ConvolutionModule(channels, kernel_size, activation=ReLU(), bias=True)

Bases: Module

ConvolutionModule in Conformer model.

• Parameters:
  • channels (int) – The number of channels of conv layers.
  • kernel_size (int) – Kernerl size of conv layers.

Construct an ConvolutionModule object.

forward(x, mask_pad=None)

Compute convolution module.

• Parameters:
  • x (torch.Tensor) – Input tensor (#batch, time, channels).
  • mask_pad (torch.Tensor) – Non-padding mask (#batch, 1, time), True or 1 where a frame is real. When given, the padded frames are zeroed right before the depthwise convolution, so the last real frames of a padded utterance see the same neighbourhood they see when the utterance is encoded on its own (a padded frame is not zero at this point: the pointwise convolution and the GLU have already put their biases into it).
• Returns: Output tensor (#batch, time, channels).
• Return type: torch.Tensor