Questions regarding inference time measurement

[yiming0416]

Hi,

Thanks for your great work! I have a question regarding the time measurement during inference. Since here you are using GPU and CUDA, I don't think using time.time() is a correct approach because CUDA is asynchronous. Instead, I think the following code is the correct way to measure the inference time when using PyTorch and GPU:

start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
start.record()
_, _, enhanced_img = model(low_img)
end.record()
torch.cuda.synchronize()
total_time += (start.elapsed_time(end))

See this refernece: https://discuss.pytorch.org/t/how-to-measure-time-in-pytorch/26964

If I use the code snippet above to measure the inference time, it is way slower than the speed you claim in the paper. I would like to get your opinion on this. Thanks.

[cuiziteng]

Thanks so much for your advice, I got it ~

Also, its OK that you can use this way to measure time and make fair comparisons.