Nan train loss values for ResNet-50 trained on ImageNet dataset

Thank you for your work!

I tried reproducing the results on your paper. Anytime I run the code it keeps returning train loss values as nan for the resnet-50 model trained on the Imagenet dataset. All the default parameters as provided in the code are still maintained. Please could you help on this?

Thank you once again!

Start by reproducing the default ResNet-50 training run on the ImageNet dataset and inspect where the training loss first becomes NaN. Compare the run against the default parameters; the issue is done when the default configuration produces finite training-loss values.

Written by the indexing model from the issue text.