GitLab

Create DataParallel model if several GPUs

I seriously don't understand why they defined num_train_epochs as a float in the originial tf code.
I Will change it at the end to avoir merge conflicts for now.

Please review @thomwolf but i think this is equivqlent (and it mimics the loss computation of the original loss)

Error was coming from "modeling_pytorch.py", line 484, in forward: start_loss = loss_fct(start_logits, start_positions) --> ValueError: Expected target size (12, 1), got torch.Size([12])