Class FeedForward

monai/networks/blocks/cablock.py:27–75

Gated-DConv Feed-Forward Network (GDFN) that controls feature flow using gating mechanism. Uses depth-wise convolutions for local context mixing and GELU-activated gating for refined feature selection. Args: spatial_dims: Number of spatial dimensions (2D or 3D) dim: Number o

class FeedForward(nn.Module):
    """Gated-DConv Feed-Forward Network (GDFN) that controls feature flow using gating mechanism.
    Uses depth-wise convolutions for local context mixing and GELU-activated gating for refined feature selection.

    Args:
        spatial_dims: Number of spatial dimensions (2D or 3D)
        dim: Number of input channels
        ffn_expansion_factor: Factor to expand hidden features dimension
        bias: Whether to use bias in convolution layers
    """

    def __init__(self, spatial_dims: int, dim: int, ffn_expansion_factor: float, bias: bool):
        super().__init__()
        hidden_features = int(dim * ffn_expansion_factor)

        self.project_in = Convolution(
            spatial_dims=spatial_dims,
            in_channels=dim,
            out_channels=hidden_features * 2,
            kernel_size=1,
            bias=bias,
            conv_only=True,
        )

        self.dwconv = Convolution(
            spatial_dims=spatial_dims,
            in_channels=hidden_features * 2,
            out_channels=hidden_features * 2,
            kernel_size=3,
            strides=1,
            padding=1,
            groups=hidden_features * 2,
            bias=bias,
            conv_only=True,
        )

        self.project_out = Convolution(
            spatial_dims=spatial_dims,
            in_channels=hidden_features,
            out_channels=dim,
            kernel_size=1,
            bias=bias,
            conv_only=True,
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.project_in(x)
        x1, x2 = self.dwconv(x).chunk(2, dim=1)
        return cast(torch.Tensor, self.project_out(F.gelu(x1) * x2))


class CABlock(nn.Module):