Class OverlapPatchEmbed

monai/networks/nets/restormer.py:61–87

Initial feature extraction using overlapped convolutions. Unlike standard patch embeddings that use non-overlapping patches, this approach maintains spatial continuity through 3x3 convolutions. Args: spatial_dims: Number of spatial dimensions (2D or 3D) in_channels: Numb

class OverlapPatchEmbed(Convolution):
    """Initial feature extraction using overlapped convolutions.
    Unlike standard patch embeddings that use non-overlapping patches,
    this approach maintains spatial continuity through 3x3 convolutions.

    Args:
        spatial_dims: Number of spatial dimensions (2D or 3D)
        in_channels: Number of input channels
        embed_dim: Dimension of embedded features. Defaults to 48.
        bias: Whether to use bias in convolution layer. Defaults to False.
    """

    def __init__(self, spatial_dims: int, in_channels: int = 3, embed_dim: int = 48, bias: bool = False):
        super().__init__(
            spatial_dims=spatial_dims,
            in_channels=in_channels,
            out_channels=embed_dim,
            kernel_size=3,
            strides=1,
            padding=1,
            bias=bias,
            conv_only=True,
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = super().forward(x)
        return x


class Restormer(nn.Module):