Pytorch 1.1.0 gave different results comparing with pytorch 1.0.0?

I train a CNN.
Experiment1: pytorch 1.1.0, loss is NaN
Experiment2: pytorch 1.0.0, loss is around 1.0

What are the differents between pytorch 1.1.0 and pytorch 1.0.0?
I think it is related to the initialization method.

Any other thoughts? Thanks!

It’s hard to tell without seeing any code.
Could you update to the latest stable release (1.3.1) and check, if you still get a NaN loss?

I tried pytorch 1.2.0, the loss is also around 1.0 .

You might have hit a legacy bug. If your code is working with the current version, I wouldn’t care too much about it. Is there a specific reason you need to use older PyTorch versions? I would still recommend to use the latest release.