The source code reads:
self.down = nn.Sequential(
nn.Conv2d(channels[level-1], channels[level], kernel_size=2, stride=2),
LayerNorm(channels[level], eps=1e-6, data_format="channels_first"),
) if level in [1, 2, 3] else nn.Identity()
So when level=0 and first_col=True, how do we implement downsampling?
The source code reads:
self.down = nn.Sequential(
nn.Conv2d(channels[level-1], channels[level], kernel_size=2, stride=2),
LayerNorm(channels[level], eps=1e-6, data_format="channels_first"),
) if level in [1, 2, 3] else nn.Identity()
So when level=0 and first_col=True, how do we implement downsampling?