Training with clipped_grad_norm value of NaN #38
Description
When I train on coco dataset, I find that the clipped_grad_norm value is NaN and total_loss is difficult to decrease, what is wrong or what might be the reason for this？
{{ message }}
When I train on coco dataset, I find that the clipped_grad_norm value is NaN and total_loss is difficult to decrease, what is wrong or what might be the reason for this？