Open zhetongliang opened 4 years ago
Thanks for pointing a potential error, theoretically, it could bring about issues when channels of dimension 1 and 2 are equal, but the chances are slim in practical. We have modified the code and tested the results on Vid4, the PSNR was 27.404891 dB and is 27.405332 dB now, and the change is about 0.000441 dB. We will report the new results on the homepage.
Hi, in the implementation of non-local module, the phi_x should be transposed first before reshaping.