Torch Version UNetEncoderBlock causes multi-card training error
Yeeesir opened this issue · comments
When I use maxim_pytorch provided by link, and try multi-gpu training, the following error occurred
File "/home/miniconda3/lib/python3.8/site-packages/torch/autograd/__init__.py", line 154, in backward
Variable._execution_engine.run_backward(
RuntimeError: Function BroadcastBackward returned an invalid gradient at index 76 - got [0] but expected shape compatible with [0, 32, 2, 2]
I found that the key problem lies in the torch implementation of UNetEncoderBlock, other network structures did not introduce errors