Closed chenchr closed 5 years ago
We do find that our implementation has some numeric issue. It seems to be less numerically stable than the pytorch built-in BN. But in our experiments, empirically, it does not hurt the performance.
Closing for now. Feel free to reopen if you have further questions.
Hello. Thanks for the awesome code. I read this code base and the adaptation from zhanghang1989 . His code customs two operators for saving gpu memory. I've just combined the cuda extension for PyTorch 0.4.1. Although I have past the gradcheck of the operator bn and the opeartor sum_square, and for each operator I've compared the operator ouput with the ouput from imperative implementation using PyTorch, I can not past the test case provided here... May you have any suggestion about the numeric stability ?