An Inquiry about Patch Embedding

[Qingyong Hu]

Dear authors,

Thanks for your great work! I have a question about the 'Patch Embed' class. It seems that there is no projection operation on the input image data. Maybe I am wrong, I think this code cannot work when we want to add the positional encodings without the projection operation.

I would appreciate too much if you could kindly help me.

Best Regards,
Qy

[Jingyun Liang]

Yes, there is no projection. The positional encodings are added at

SwinIR/models/network_swinir.py

Line 130 in 9b1a9bf

 attn = attn + relative_position_bias.unsqueeze(0)