fix torch no grad

[jpata]

no_grad is required around the model forward pass in the validation loop to allow larger batch sizes not to OOM.