search

creotiv / hdrnet-pytorch

Unofficial PyTorch implementation of 'Deep Bilateral Learning for Real-Time Image Enhancement', SIGGRAPH 2017 https://groups.csail.mit.edu/graphics/hdrnet/
220 stars 43 forks source link

Have you reproduced the results shown in the paper? #30

Open Jian-danai opened 1 year ago

Jian-danai commented 1 year ago

Hi, have you reproduced the results on the FiveK dataset?

Ssakura-go commented 1 year ago

Hi, have you reproduced the results on the FiveK dataset?

The size shown in the ds (input and output) are not match, but the results seems great. Have you get what you want by this code? I'd like to get the results and compare it with my own method. If you are convenient, I am very glad to make a discussion with you. (#^.^#)

Jian-danai commented 1 year ago

The PSNR on fivek is not stable even after 20~30 epochs training, typically +-0.3dB, maybe the lr is too large. Do you have any idea on it? Thanks.

Sissuire commented 1 year ago

What is the batch size and learning rate ? I cannot reproduce the result and the gap is large. I'm seeking for the reason.

Jian-danai commented 1 year ago

What is the batch size and learning rate ? I cannot reproduce the result and the gap is large. I'm seeking for the reason.

just try default? batch_size=6

Sissuire commented 1 year ago

@Jian-danai Currently I set lr=1e-4, batch_size=24, and directly resize the input into 256x256/512x512 without any augmentation. And PSNR on fivek is 21.96dB. I don;t know whether it is good or bad :(

Sissuire commented 1 year ago

Further, I have also encountered the discontinuity problem. Even with global features, the result seems not good as the difference among grids is visible.

Ssakura-go commented 1 year ago

Further, I have also encountered the discontinuity problem. Even with global features, the result seems not good as the difference among grids is visible.

Maybe we can talk on WeChat or other app where we are convenient to discuss.

Ssakura-go commented 1 year ago

Maybe we can talk on WeChat or other app where we are convenient to discuss.

Maybe we can talk on WeChat or other app where we are convenient to discuss.