Hello everyone,I ran into this issue when I was reproducing the experiment:reflection_pad2d_backward_cuda does not have a deterministic implementation, but you set ‘torch.use_deterministic_algorithms(True)’. You can turn off determinism just for this operation, or you can use the ‘warn_only=True’ option, if that’s acceptable for your application. You can also file an issue at Issues · pytorch/pytorch · GitHub to help us prioritize adding deterministic support for this operation.
This seems to show that ReflectPad2D does not implement a deterministic algorithm when backpropagating, so if the network contains ReflectPad2D, it will not be able to fully reproduce the previous results(Even if I’ve fixed the random number seed)?
Or how can I fully reproduce the results of a network experiment that includes ReflectPad2D？