Method forward

inspiremusic/flow/decoder.py:202–277

Forward pass of the UNet1DConditional model. Args: x (torch.Tensor): shape (batch_size, in_channels, time) mask (_type_): shape (batch_size, 1, time) t (_type_): shape (batch_size) spks (_type_, optional): shape: (batch_size, condition_channel

(self, x, mask, mu, t, spks=None, cond=None)

                    nn.init.constant_(m.bias, 0)

    def forward(self, x, mask, mu, t, spks=None, cond=None):
        """Forward pass of the UNet1DConditional model.

        Args:
            x (torch.Tensor): shape (batch_size, in_channels, time)
            mask (_type_): shape (batch_size, 1, time)
            t (_type_): shape (batch_size)
            spks (_type_, optional): shape: (batch_size, condition_channels). Defaults to None.
            cond (_type_, optional): placeholder for future use. Defaults to None.

        Raises:
            ValueError: _description_
            ValueError: _description_

        Returns:
            _type_: _description_
        """

        t = self.time_embeddings(t).to(t.dtype)
        t = self.time_mlp(t)
        x = pack([x, mu], "b * t")[0]
        if spks is not None:
            spks = repeat(spks, "b c -> b c t", t=x.shape[-1])
            x = pack([x, spks], "b * t")[0]
        if cond is not None:
            x = pack([x, cond], "b * t")[0]
        hiddens = []
        masks = [mask]
        for resnet, transformer_blocks, downsample in self.down_blocks:
            mask_down = masks[-1]
            x = resnet(x, mask_down, t)
            x = rearrange(x, "b c t -> b t c").contiguous()
            attn_mask = torch.matmul(mask_down.transpose(1, 2).contiguous(), mask_down)
            for transformer_block in transformer_blocks:
                x = transformer_block(
                    hidden_states=x,
                    attention_mask=attn_mask,
                    timestep=t,
                )
            x = rearrange(x, "b t c -> b c t").contiguous()
            hiddens.append(x)  # Save hidden states for skip connections
            x = downsample(x * mask_down)
            masks.append(mask_down[:, :, ::2])
        masks = masks[:-1]
        mask_mid = masks[-1]

        for resnet, transformer_blocks in self.mid_blocks:
            x = resnet(x, mask_mid, t)
            x = rearrange(x, "b c t -> b t c").contiguous()
            attn_mask = torch.matmul(mask_mid.transpose(1, 2).contiguous(), mask_mid)
            for transformer_block in transformer_blocks:
                x = transformer_block(
                    hidden_states=x,
                    attention_mask=attn_mask,
                    timestep=t,
                )
            x = rearrange(x, "b t c -> b c t").contiguous()