nan loss when training on Windows

[rose-jinyang]

Hello
How are you?
Thanks for contributing to this project.
I am going to train a model on Windows.
But I get nan loss values at the beginning.
I checked the same training project on Google Colab (Ubuntu) and it works well.
I think that the there are some issues in the current project for Windows.

[RizwanMunawar]

Hi @rose-jinyang! I am fine.
At the start of training, the loss can be zero. There can be a possibility that the current batch loaded for training will provide zero loss value on the window machine. It should be much better if you will compare the results after 5 to 6 epochs at least. Also, I will recommend you use Linux/Google Colab, as it could speed up the training process.

[rose-jinyang]

Thanks for your quick reply.
I tested on Google Colab but It works well even at the first epoch.
But when training on Windows, the losses are nan even after 5 epochs.

[RizwanMunawar]

@rose-jinyang! I will test the results on windows and explore the topic in detail to optimize code as much as possible especially for windows.

[rose-jinyang]

Thanks