Some questions about the code

[HC-2016]

1 why do you construct the val sampler as a random sampler rather than a sequence sampler? It may be inconvenient when inferencing on test data. https://github.com/microsoft/Swin-Transformer/blob/777f6c66604bb5579086c4447efe3620344d95a9/data/build.py#L41 https://github.com/microsoft/Swin-Transformer/blob/777f6c66604bb5579086c4447efe3620344d95a9/data/samplers.py#L23

2 The organization of the stage in the code seems to be different with the organization illustrateed in the paper. The organization of the stage in the code:

    def forward(self, x):
        for blk in self.blocks:
            if self.use_checkpoint:
                x = checkpoint.checkpoint(blk, x)
            else:
                x = blk(x)
        if self.downsample is not None:
            x = self.downsample(x)
        return x

[kingkung2016]

i have the same question, the outputs of each stage are [bs, H/8, W/8, 2C], [bs, H/16, W/16, 4C], [bs, H/32, W/32, 8C], [bs, H/32, W/32, 8C]

[ancientmooner]

Thanks for point out these issues @HC-2016 . We will fix both of them.

[k-zha14]

Wow, the same question with Q2. To be honest, the code implementation is quite confusing. WHY put the patch merging module at last? It's really inconvinient to extract multi-level features based on this repo.