search

proteus1991 / GridDehazeNet

This repo contains the official training and testing codes for our paper: GridDehazeNet: Attention-Based Multi-Scale Network for Image Dehazing.
https://jhc.sjtu.edu.cn/~xiaohongliu/
243 stars 54 forks source link

why do u use pretrained best network weight to retrain the network? #2

Closed leochangzliao closed 5 years ago

leochangzliao commented 5 years ago

As the title mentions, the code (https://github.com/proteus1991/GridDehazeNet/blob/master/train.py#L98) use the best weight (indoor_haze_best_3_6) to retrain the network with indoor training dataset. It seems that it is not reasonable.

proteus1991 commented 5 years ago

Hi, this is just an example to show how to retrain/finetune the GridDehazeNet. You can modify the code accordingly to match your use case.

leochangzliao commented 5 years ago

If I want to reproduce the experiment results, I should comment the line mentioned above, is that right? @proteus1991

leochangzliao commented 5 years ago

After reading your paper, I have some questions.

  1. Do u just use the ITS dataset to train the network and test on SOTS?
  2. Do u just use the OTS dataset to train the network and test on SOTS?
  3. If I use pretrained network (using OTS to train) to finetune on ITS, will I get better test results on SOTS?
proteus1991 commented 5 years ago

If I want to reproduce the experiment results, I should comment the line mentioned above, is that right? @proteus1991

Yes. You should not load my well-trained weights.

proteus1991 commented 5 years ago

After reading your paper, I have some questions.

  1. Do u just use the ITS dataset to train the network and test on SOTS?
  2. Do u just use the OTS dataset to train the network and test on SOTS?
  3. If I use pretrained network (using OTS to train) to finetune on ITS, will I get better test results on SOTS?

For question 1 and 2, yes. I used ITS dataset to train the network and test on SOTS indoor and used OTS dataset for outdoor.

For question 3, I didn't train it in that way. You can read the paper which introduces RESIDE dataset for more details.

leochangzliao commented 5 years ago

After reading your paper, I have some questions.

  1. Do u just use the ITS dataset to train the network and test on SOTS?
  2. Do u just use the OTS dataset to train the network and test on SOTS?
  3. If I use pretrained network (using OTS to train) to finetune on ITS, will I get better test results on SOTS?

For question 1 and 2, yes. I used ITS dataset to train the network and test on SOTS indoor and used OTS dataset for outdoor.

For question 3, I didn't train it in that way. You can read the paper which introduces RESIDE dataset for more details.

Thanks for ur reply!!! Great work!

proteus1991 commented 5 years ago

After reading your paper, I have some questions.

  1. Do u just use the ITS dataset to train the network and test on SOTS?
  2. Do u just use the OTS dataset to train the network and test on SOTS?
  3. If I use pretrained network (using OTS to train) to finetune on ITS, will I get better test results on SOTS?

For question 1 and 2, yes. I used ITS dataset to train the network and test on SOTS indoor and used OTS dataset for outdoor. For question 3, I didn't train it in that way. You can read the paper which introduces RESIDE dataset for more details.

Thanks for ur reply!!! Great work!

Thanks. This issue is closed.