search

Zheng222 / IMDN

Lightweight Image Super-Resolution with Information Multi-distillation Network (ACM MM 2019, Winner Award of ICCVW AIM 2019 Constrained SR Track1&Track2)
425 stars 70 forks source link

about train and checkpoint #6

Closed zymize closed 4 years ago

zymize commented 4 years ago

hello,nice to meet you I want to ask that is checkpoint a pre-train model? and how can i train it? thanks looking forward your reply

zymize commented 4 years ago

what's more,I found my testing time has a problem as following:

Namespace(checkpoint='checkpoints/IMDN_x2.pth', cuda=True, is_y=True, output_folder='results/Set5/x2', test_hr_folder='Test_Datasets/Set5/', test_lr_folder='Test_Datasets/Set5_LR/x2/', upscale_factor=2) Mean PSNR: 37.91055018265467, SSIM: 0.9594468667890528, TIME: 954.8178810119629 ms

Zheng222 commented 4 years ago

@zymize

  1. The training phase will be released. Now, you can use EDSR-pytorch to training our IMDN model.

  2. Please check out your device. https://github.com/Zheng222/IMDN/blob/0dd1c373e7ecb2d518acc049f5640ac554cd3714/test_IMDN.py#L30 We test the inference snippet on Nvidia TITAN Xp. About the training time, you can also refer to #4.

zymize commented 4 years ago

thank you for your kind reply. what's more,I have another question about your IDN work. I couldn't send 800 photos to train this work beause I just have 16G CPU. Could I send 800 photos dividing two parts? How to achieve it? looking forword to your reply. thank you verch again.

Zheng222 commented 4 years ago

@zymize you can rewrite the pytorch code for IDN, 16G RAM is enough for training IDN.

zymize commented 4 years ago

Thanks for your reply. I resolve this question for adding CPU. But someone use your caffe code IDN's X2 PSNR result of the paper in compared table is 37.75 . And I use tensorflow code, the same is 37.75. When I compare your IDN's work, ,can I write that the IDN's X2 PSNR result of the paper is 37.75 ? If my result is higher 37.75 the same environment, can I think that my work is effective?

Zheng222 commented 4 years ago

@zymize Hello, I have listed the result in IDN-tensorflow. You can directly copy these results. For fair comparison, you should use the matlab evaluation code to test PSNR, not use python skimage.measure.compute_psnr(). Thank you!

zymize commented 4 years ago

@Zheng222 OK thanks for your reminding. I am sorry to forget to test your code on Matlab. I use your original code to test on Matlab . I cannot acheive your results by my training . Could you give me some suggestions to get your results?

Zheng222 commented 4 years ago

Hello, I recommend that you train a few more epochs at learning rates of 2e-4 and 1e-4.