[pytorch 에러해결] RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [3, 4]], which is output 0 of ReluBackward0, is at version 1; expected version 0 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!

github에서 가져온 EfficientDet의 pytorch 버전 코드를 돌리다가 발생한 문제!

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [3, 4]], which is output 0 of ReluBackward0, is at version 1; expected version 0 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!

에러 참 길다 ㅎㅎ 특히 마지막의 Good Luck! 너무 화가난다 🤪🤪

구글링한 결과, 두 가지 해결방법을 찾을 수 있었다.

1. torch.autograd.set_detect_anomaly(True) 추가해주기

2. torchvision version 다르게 설치하기

먼저 1번 방법대로 코드를 추가해주었다. => 참고한 블로그

안된다! 포기!

다음은 2번의 방법을 위해서 아래와 같은 버전으로 깔아 주어야 한다고 하는데. 나는 해당 버전이 존재하지 않아 깔지 못한다고 한다..

pip install torch==1.4.0 torchvision==0.5.0

이런 워닝 문구~

또 열심히 구글링해서 깔 수 있었던 방법.. “=” 이 3개나 필요하다

pip install torch===1.4.0 torchvision===0.5.0

이래도 안되면

conda install pytorch==1.4.0
pip install tochvision===0.5.0

해보기!