How does the decoder produce prediction in the same resolution as original input?

[edizhuang]

Hi,

I'm checking how the decoder produce prediction in the same resolution as original input, for my case only partial region is the interest and therefore needs customization.

For example in mmsegmentation, take Swin Transformer backbone as encoder and UperNet as decoder, only the output features (in reduced resolution) from the encoder are passed to the decoder, where is the step to resize the prediction to the original resolution? This helps model customization.

Thanks,

[lorinczszabolcs]

See #1474 and https://github.com/open-mmlab/mmsegmentation/issues/1474#issuecomment-1099827006 .