Using batch_size greater than 1 when calculating test losses and accuracies

[talrub]

Hi,

In train_distributed.py you are using 'batch_size=1' when calculating test losses and accuracies.
Why not to use a bigger batch_size to speed up the calculations?

Thanks!