Class ConditionalDecoder

inspiremusic/flow/decoder.py:80–277

        return x

class ConditionalDecoder(nn.Module):
    def __init__(
        self,
        in_channels,
        out_channels,
        channels=(256, 256),
        dropout=0.05,
        attention_head_dim=64,
        n_blocks=1,
        num_mid_blocks=2,
        num_heads=4,
        act_fn="snake",
    ):
        """
        This decoder requires an input with the same shape of the target. So, if your text content
        is shorter or longer than the outputs, please re-sampling it before feeding to the decoder.
        """
        super().__init__()
        channels = tuple(channels)
        self.in_channels = in_channels
        self.out_channels = out_channels

        self.time_embeddings = SinusoidalPosEmb(in_channels)
        time_embed_dim = channels[0] * 4
        self.time_mlp = TimestepEmbedding(
            in_channels=in_channels,
            time_embed_dim=time_embed_dim,
            act_fn="silu",
        )
        self.down_blocks = nn.ModuleList([])
        self.mid_blocks = nn.ModuleList([])
        self.up_blocks = nn.ModuleList([])

        output_channel = in_channels
        for i in range(len(channels)):  # pylint: disable=consider-using-enumerate
            input_channel = output_channel
            output_channel = channels[i]
            is_last = i == len(channels) - 1
            resnet = ResnetBlock1D(dim=input_channel, dim_out=output_channel, time_emb_dim=time_embed_dim)
            transformer_blocks = nn.ModuleList(
                [
                    BasicTransformerBlock(
                        dim=output_channel,
                        num_attention_heads=num_heads,
                        attention_head_dim=attention_head_dim,
                        dropout=dropout,
                        activation_fn=act_fn,
                    )
                    for _ in range(n_blocks)
                ]
            )
            downsample = (
                Downsample1D(output_channel) if not is_last else nn.Conv1d(output_channel, output_channel, 3, padding=1)
            )
            self.down_blocks.append(nn.ModuleList([resnet, transformer_blocks, downsample]))

        for _ in range(num_mid_blocks):
            input_channel = channels[-1]