P2T replaces PVT trunk bug

[liu-tianxiang]

When I replaced the PVT trunk with P2T in my code, I encountered an error ：
RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [16, 512, 3, 3]], which is output 0 of AdaptiveAvgPool2DBackward, is at version 1; expected version 0 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

[yuhuan-wu]

Thanks for you notice. In some higher torch versions, inplace operation after adaptive_avg_pool is not supported.

I have disabled the inplace operation in the newest version. See the right change here:

P2T/p2t.py

Line 74 in 4d52408

pool = pool + l(pool) # fix backward bug in higher torch versions when training

[liu-tianxiang]

Thanks! problem solved