Reproducing repository accuracy

[OmarDahroug]

Dear Anant,

I'm have problems reproducing the accuracies you mentioned in the README file. I tried training the model on several tasks but I receive similar behaviour. The loss always reaches a loss of around 1,7 and stays mostly constant.

I train the model with the following hyperparameters:

• epochs = 200
• StepLR scheduler with:
  • step size =10
  • base_lr = 1e-2
  • base_lr = 2e-4
• batchsize = 32

I'm looking forward to your reply.

Best,
Omar