search

zzr-idam / 4KDehazing

78 stars 15 forks source link

Questions on the shared pretrained model #7

Open Andy-YX opened 2 years ago

Andy-YX commented 2 years ago

Hello, thanks for sharing your dehazing work.

It seems that the shared pretrained model does not exactly match the network structure described in the paper. I have also conducted some tests and the inference quality of the shared pretrained model is not quite desirable when applied for 4KID or O-Haze dataset.

So could you please briefly explain how the shared pre-trained model is generated? Is it for 4KID or for O-Haze dataset? Are there any fine-tunes needed for models used for O-Haze dataset?

Thanks!

17328-wu commented 2 years ago

Hi, have you solved this problem?

delldu commented 2 years ago

Oh MY God !!! 1) Code does not match the paper 2) Effect does not match the paper (pre-trained model performance too bad) 3) Run time does not match the paper 4) Question answer say nothing

Andy-YX commented 2 years ago

Hi, have you solved this problem?

@98-xiaowu Nope, something seems not correct with the shared pre-trained model. Need to retrain it from scratch...

Andy-YX commented 2 years ago

Oh MY God !!!

  1. Code does not match the paper
  2. Effect does not match the paper (pre-trained model performance too bad)
  3. Run time does not match the paper
  4. Question answer say nothing

@delldu Nice summary, lol

zzr-idam commented 2 years ago

Sorry, 1) our model has been improved to improve the performance, such as using Unet instead of convolutional blocks; 2) the pre-trained model we will retrain.

17328-wu commented 2 years ago

Thank you for sharing，will you release the optimized code and pre-trained model?

17328-wu commented 2 years ago

@zzr-idam And, I used 3 Quadro RTX 6000 (24GB) GPU cards for training. Why can I only train over 100 Epochs a day ？ Is this normal speed? May I ask what is the reason for this?