Closed jgyy4775 closed 4 years ago
I never face this problem before. I'm afraid this may be some unknown reasons after upgrading PyTorch/ CUDA/ etc. If you'd like to try, there are a tons of feature normalization trick nowadays can help prevent nan problems.
@yaohungt The reason for the "indexError" is that the feature become "nan" in "i3d.py". This problem occurs very randomly. Sometimes Mixed_3c, and sometimes Conv3d_2c_3*3.
Is there a cause or solution for this problem?