Torch.randn().normal_() returns nans

1

When running for example

torch.randn(64, 512, 1, 1).normal_(mean=0.0007720401627011597, std=0.01006225310266018))

there are sometimes nans in the tensor.

I’m working on pytorch 0.1.12 with python 3.5 and the “Bug” is reproducible. Am I doing sth wrong or is it already fixed ?

2

did you figure out how to solve it?

3

I am running this on pytorch 0.4 and I haven’t seen any nans yet. @isalirezag what version of pytorch are you on?

4

@richard I am using 0.3 version