Cudnn RNN backward can only be called in training mode

autograd
2

Hi

You call model.eval() at line 81 but forget to set the model to train mode again when you run start the second epoch.

Adding model.train() after line 47 will remove the error.

5 Likes