modules/wenet_extractor/efficient_conformer/convolution.py

# This module is from [WeNet](https://github.com/wenet-e2e/wenet).

# ## Citations

# ```bibtex
# @inproceedings{yao2021wenet,
#   title={WeNet: Production oriented Streaming and Non-streaming End-to-End Speech Recognition Toolkit},
#   author={Yao, Zhuoyuan and Wu, Di and Wang, Xiong and Zhang, Binbin and Yu, Fan and Yang, Chao and Peng, Zhendong and Chen, Xiaoyu and Xie, Lei and Lei, Xin},
#   booktitle={Proc. Interspeech},
#   year={2021},
#   address={Brno, Czech Republic },
#   organization={IEEE}
# }

# @article{zhang2022wenet,
#   title={WeNet 2.0: More Productive End-to-End Speech Recognition Toolkit},
#   author={Zhang, Binbin and Wu, Di and Peng, Zhendong and Song, Xingchen and Yao, Zhuoyuan and Lv, Hang and Xie, Lei and Yang, Chao and Pan, Fuping and Niu, Jianwei},
#   journal={arXiv preprint arXiv:2203.15455},
#   year={2022}
# }
#

"""ConvolutionModule definition."""
from typing import Tuple

import torch
from torch import nn


class ConvolutionModule(nn.Module):
    """ConvolutionModule in Conformer model."""

    def __init__(
        self,
        channels: int,
        kernel_size: int = 15,
        activation: nn.Module = nn.ReLU(),
        norm: str = "batch_norm",
        causal: bool = False,
        bias: bool = True,
        stride: int = 1,
    ):
        """Construct an ConvolutionModule object.
        Args:
            channels (int): The number of channels of conv layers.
            kernel_size (int): Kernel size of conv layers.
            causal (int): Whether use causal convolution or not
            stride (int): Stride Convolution, for efficient Conformer
        """
        super().__init__()

        self.pointwise_conv1 = nn.Conv1d(
            channels,
            2 * channels,
            kernel_size=1,
            stride=1,
            padding=0,
            bias=bias,
        )
        # self.lorder is used to distinguish if it's a causal convolution,
        # if self.lorder > 0: it's a causal convolution, the input will be
        #    padded with self.lorder frames on the left in forward.
        # else: it's a symmetrical convolution
        if causal:
            padding = 0
            self.lorder = kernel_size - 1
        else:
            # kernel_size should be an odd number for none causal convolution
            assert (kernel_size - 1) % 2 == 0
            padding = (kernel_size - 1) // 2
            self.lorder = 0

        self.depthwise_conv = nn.Conv1d(
            channels,
            channels,
            kernel_size,
            stride=stride,  # for depthwise_conv in StrideConv
            padding=padding,
            groups=channels,
            bias=bias,
        )

        assert norm in ["batch_norm", "layer_norm"]
        if norm == "batch_norm":
            self.use_layer_norm = False
            self.norm = nn.BatchNorm1d(channels)
        else:
            self.use_layer_norm = True
            self.norm = nn.LayerNorm(channels)

        self.pointwise_conv2 = nn.Conv1d(
            channels,
            channels,
            kernel_size=1,
            stride=1,
            padding=0,
            bias=bias,
        )
        self.activation = activation
        self.stride = stride

    def forward(
        self,
        x: torch.Tensor,
        mask_pad: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
        cache: torch.Tensor = torch.zeros((0, 0, 0)),
    ) -> Tuple[torch.Tensor, torch.Tensor]:
        """Compute convolution module.
        Args:
            x (torch.Tensor): Input tensor (#batch, time, channels).
            mask_pad (torch.Tensor): used for batch padding (#batch, 1, time),
                (0, 0, 0) means fake mask.
            cache (torch.Tensor): left context cache, it is only
                used in causal convolution (#batch, channels, cache_t),
                (0, 0, 0) meas fake cache.
        Returns:
            torch.Tensor: Output tensor (#batch, time, channels).
        """
        # exchange the temporal dimension and the feature dimension
        x = x.transpose(1, 2)  # (#batch, channels, time)

        # mask batch padding
        if mask_pad.size(2) > 0:  # time > 0
            x.masked_fill_(~mask_pad, 0.0)

        if self.lorder > 0:
            if cache.size(2) == 0:  # cache_t == 0
                x = nn.functional.pad(x, (self.lorder, 0), "constant", 0.0)
            else:
                # When export ONNX，the first cache is not None but all-zero,
                # cause shape error in residual block,
                #   eg. cache14 + x9 = 23, 23-7+1=17 != 9
                cache = cache[:, :, -self.lorder :]
                assert cache.size(0) == x.size(0)  # equal batch
                assert cache.size(1) == x.size(1)  # equal channel
                x = torch.cat((cache, x), dim=2)
            assert x.size(2) > self.lorder
            new_cache = x[:, :, -self.lorder :]
        else:
            # It's better we just return None if no cache is requried,
            # However, for JIT export, here we just fake one tensor instead of
            # None.
            new_cache = torch.zeros((0, 0, 0), dtype=x.dtype, device=x.device)

        # GLU mechanism
        x = self.pointwise_conv1(x)  # (batch, 2*channel, dim)
        x = nn.functional.glu(x, dim=1)  # (batch, channel, dim)

        # 1D Depthwise Conv
        x = self.depthwise_conv(x)
        if self.use_layer_norm:
            x = x.transpose(1, 2)
        x = self.activation(self.norm(x))
        if self.use_layer_norm:
            x = x.transpose(1, 2)
        x = self.pointwise_conv2(x)
        # mask batch padding
        if mask_pad.size(2) > 0:  # time > 0
            if mask_pad.size(2) != x.size(2):
                mask_pad = mask_pad[:, :, :: self.stride]
            x.masked_fill_(~mask_pad, 0.0)

        return x.transpose(1, 2), new_cache