When we make “torch.backends.cudnn.benchmark = True”，it need many loops then the machine will test the time more stably.
So we are worried about how to accurately test time in the test phase?
Whether use your code in the same setting to measure? （only need our model and checkpoint and make sure it can be executable）
No need to worry about this. I will use the same setting for all the models and may set “torch.backends.cudnn.benchmark = False”.Posted by: cszn @ July 2, 2020, 11:40 a.m.