Method forward

acestep/models/lyrics_utils/lyric_encoder.py:75–130

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

(
        self,
        x: torch.Tensor,
        mask_pad: torch.Tensor = torch.ones((0, 0, 0), dtype=torch.bool),
        cache: torch.Tensor = torch.zeros((0, 0, 0)),
    )

        self.activation = activation

    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:
                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 required,
            # 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
            x.masked_fill_(~mask_pad, 0.0)

        return x.transpose(1, 2), new_cache