Class MaskDownSampler

sam2_train/modeling/memory_encoder.py:17–58

Progressively downsample a mask by total_stride, each time by stride. Note that LayerNorm is applied per *token*, like in ViT. With each downsample (by a factor stride**2), channel capacity increases by the same factor. In the end, we linearly project to embed_dim channels.

class MaskDownSampler(nn.Module):
    """
    Progressively downsample a mask by total_stride, each time by stride.
    Note that LayerNorm is applied per *token*, like in ViT.

    With each downsample (by a factor stride**2), channel capacity increases by the same factor.
    In the end, we linearly project to embed_dim channels.
    """

    def __init__(
        self,
        embed_dim=256,
        kernel_size=4,
        stride=4,
        padding=0,
        total_stride=16,
        activation=nn.GELU,
    ):
        super().__init__()
        num_layers = int(math.log2(total_stride) // math.log2(stride))
        assert stride**num_layers == total_stride
        self.encoder = nn.Sequential()
        mask_in_chans, mask_out_chans = 1, 1
        for _ in range(num_layers):
            mask_out_chans = mask_in_chans * (stride**2)
            self.encoder.append(
                nn.Conv2d(
                    mask_in_chans,
                    mask_out_chans,
                    kernel_size=kernel_size,
                    stride=stride,
                    padding=padding,
                )
            )
            self.encoder.append(LayerNorm2d(mask_out_chans))
            self.encoder.append(activation())
            mask_in_chans = mask_out_chans

        self.encoder.append(nn.Conv2d(mask_out_chans, embed_dim, kernel_size=1))

    def forward(self, x):
        return self.encoder(x)


# Lightly adapted from ConvNext (https://github.com/facebookresearch/ConvNeXt)