I find that in the testing code of MSRResNet the torch.backends.cudnn.benchmark is set as true. However, in the test when I set the torch.backends.cudnn.benchmark=False in the code, the test is much faster.( ~0.13s for the False case and ~0.5s for the Ture case with GTX 2080). So could I set backends.cudnn.benchmark=False for both the baseline model and my model?
Posted by: wy-sun16 @ Aug. 26, 2019, 8:24 a.m.Yes, you can set backends.cudnn.benchmark=False for both the baseline model and your model.
We will finally report the running times of different models on the same machine.