espnet2.enh.layers.dpmulcat.MulCatBlock

source

class espnet2.enh.layers.dpmulcat.MulCatBlock(input_size: int, hidden_size: int, dropout: float = 0.0, bidirectional: bool = True)

Bases: Module

The MulCat block.

This module implements a multiplicative concatenation block using LSTM layers. It processes input sequences through two separate LSTM networks: one for the main processing and another to create a gating mechanism. The outputs are combined to enhance the feature representation.

rnn

The primary LSTM layer for feature extraction.

• Type: nn.LSTM

rnn

_proj

Linear layer to project the output of the RNN.

• Type: nn.Linear

gate_rnn

The gating LSTM layer.

• Type: nn.LSTM

gate_rnn

_proj

Linear layer to project the output of the gate RNN.

• Type: nn.Linear

block_projection

Final linear projection to match input size.

• Type: nn.Linear

• Parameters:

  • input_size (int) – Dimension of the input feature. The input should have shape (batch, seq_len, input_size).
  • hidden_size (int) – Dimension of the hidden state.
  • dropout (float , optional) – The dropout rate in the LSTM layer. Defaults to 0.0.
  • bidirectional (bool , optional) – Whether the RNN layers are bidirectional. Defaults to True.

####### Examples

>>> input_tensor = torch.randn(32, 10, 64)  # (batch_size, seq_len, input_size)
>>> mulcat_block = MulCatBlock(input_size=64, hidden_size=32)
>>> output_tensor = mulcat_block(input_tensor)
>>> output_tensor.shape
torch.Size([32, 10, 64])  # Output has the same shape as input

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

forward(input)

The MulCat block.

This module implements a MulCat block that processes input features through LSTM layers and applies gating mechanisms to produce an output feature.

• Parameters:
  • input_size (int) – Dimension of the input feature. The input should have shape (batch, seq_len, input_size).
  • hidden_size (int) – Dimension of the hidden state.
  • dropout (float , optional) – The dropout rate in the LSTM layer. (Default: 0.0)
  • bidirectional (bool , optional) – Whether the RNN layers are bidirectional. (Default: True)

####### Examples

>>> mul_cat_block = MulCatBlock(input_size=128, hidden_size=64)
>>> input_tensor = torch.randn(32, 10, 128)  # (batch, seq_len, input_size)
>>> output_tensor = mul_cat_block(input_tensor)
>>> print(output_tensor.shape)  # (32, 10, 128)