I have some questions.On Gopro datasets, I used a single deblurnet(yours)to train the parameters. The experimental setup is that the train_batch_size is 2, the test_bs is 1, the initial lr is 1e-4, and lr is decayed by 0.1 in [80, 160, 240], and the other parameters are consistent with your code.
I have a few problems :
(1) Such a small batch, Memory-Usage and GPU-Util are small, is this a waste of resources? I ran 200 epochs in a day and a half, and the PSNR only reached 28.8 not 30.55 reported in your paper, but there is no sign of rising.
(2) If I use a large batch_size, the PSNR performance is worse.
What do I need to pay attention to in terms of training speed and code tricks to reproduce the results of around 30.5 in your paper? If you can give some guidance, I really appreciate you!
Hello, your work is really amazing!
I have some questions.On Gopro datasets, I used a single deblurnet(yours)to train the parameters. The experimental setup is that the train_batch_size is 2, the test_bs is 1, the initial lr is 1e-4, and lr is decayed by 0.1 in [80, 160, 240], and the other parameters are consistent with your code.
I have a few problems : (1) Such a small batch, Memory-Usage and GPU-Util are small, is this a waste of resources? I ran 200 epochs in a day and a half, and the PSNR only reached 28.8 not 30.55 reported in your paper, but there is no sign of rising. (2) If I use a large batch_size, the PSNR performance is worse.
What do I need to pay attention to in terms of training speed and code tricks to reproduce the results of around 30.5 in your paper? If you can give some guidance, I really appreciate you!
Thanks for your nice work again!