Merge pull request #95 from alibaba-damo-academy/dev_dzh

zhifu gao

2023-02-10 60aef2aa96750e86158d83e52429750cf4b626e3

 funasr/modules/multi_layer_conv.py

@@ -63,6 +63,58 @@
        return self.w_2(self.dropout(x).transpose(-1, 1)).transpose(-1, 1)


class FsmnFeedForward(torch.nn.Module):
    """Position-wise feed forward for FSMN blocks.

    This is a module of multi-leyered conv1d designed
    to replace position-wise feed-forward network
    in FSMN block.
    """

    def __init__(self, in_chans, hidden_chans, out_chans, kernel_size, dropout_rate):
        """Initialize FsmnFeedForward module.

        Args:
            in_chans (int): Number of input channels.
            hidden_chans (int): Number of hidden channels.
            out_chans (int): Number of output channels.
            kernel_size (int): Kernel size of conv1d.
            dropout_rate (float): Dropout rate.

        """
        super(FsmnFeedForward, self).__init__()
        self.w_1 = torch.nn.Conv1d(
            in_chans,
            hidden_chans,
            kernel_size,
            stride=1,
            padding=(kernel_size - 1) // 2,
        )
        self.w_2 = torch.nn.Conv1d(
            hidden_chans,
            out_chans,
            kernel_size,
            stride=1,
            padding=(kernel_size - 1) // 2,
            bias=False
        )
        self.norm = torch.nn.LayerNorm(hidden_chans)
        self.dropout = torch.nn.Dropout(dropout_rate)

    def forward(self, x, ilens=None):
        """Calculate forward propagation.

        Args:
            x (torch.Tensor): Batch of input tensors (B, T, in_chans).

        Returns:
            torch.Tensor: Batch of output tensors (B, T, out_chans).

        """
        x = torch.relu(self.w_1(x.transpose(-1, 1))).transpose(-1, 1)
        return self.w_2(self.norm(self.dropout(x)).transpose(-1, 1)).transpose(-1, 1), ilens


class Conv1dLinear(torch.nn.Module):
    """Conv1D + Linear for Transformer block.