L-BFGS-B optimizer reaches to NaN errors with PyTorch Backend

[ehsanngh]

Hello,

I have faced an issue when I'm using PyTorch Backend. When I switch my optimizer from Adam to "L-BFGS-B," after a few thousands iterations, I get NaN for my losses. Then, I cannot even use the best-achieved trained model for the prediction. model.predict returns nothing when this happens. It's worth mentioning that I don't face this issue when I have TF2 as my Backend. I was wondering maybe someone else had faced this problem before. Thanks.

[lululxvi]

If TF2 works, but not pytorch, then it seems an issue of the L-BFGS implementation in pytorch.