Getting out of memory in validation even using torch.no_grad()

[Oktai15 (Oktai Tatanov)]

Please, check this thread: Increase the CUDA memory twice then stop increasing