search

RaphaelMeudec / deblur-gan

Keras implementation of "DeblurGAN: Blind Motion Deblurring Using Conditional Adversarial Networks"
588 stars 244 forks source link

Model Performance #2

Open Karthik-Suresh93 opened 6 years ago

Karthik-Suresh93 commented 6 years ago

Hi!

I am curious if you were able to reproduce the paper results with this implementation? I just want to use pre-trained weights of this model on a downstream task and the performance of this model will be an important factor for me.

Thanks and Regards, Karthik

RaphaelMeudec commented 6 years ago

Hi! I haven't found the time yet to remove the checkerboard artifacts, I'll first check this issue before trying to apply it to object detection. Will keep you updated!

Karthik-Suresh93 commented 6 years ago

Did you check the average PSNR on the GOPRO test set?

RaphaelMeudec commented 6 years ago

It is planned, I want to check this checkerboard artifact before computing PSNR!

Karthik-Suresh93 commented 6 years ago

Thanks a lot for the responses. Please do update the repo with your findings

caiya55 commented 6 years ago

Hi， In the case of checkboard artifact. I did some research about this. It seems like a common phenomenon and also occurs in original deblur. The reason is the deconvlution/tranconvolution layers from generator model. details are here checkboard I found several suggestions from the author and the link.

  1. change the kernel size from 3 to 4 like the author did. (check the newest update from torch project)
  2. change the Conv2DTranspose layers into UpSampling2D+Conv2D layer in generator
  3. use sub-pixel convolution from the paper subpixel
  4. more time training.

I tried the first one, results dont change too much. I am trying on the second one. the results seems good and I am still working on it.

RaphaelMeudec commented 6 years ago

Pretty nice! If you have a breakthrough, please open a PR!

shivam2296 commented 6 years ago

Hello there, as you have worked on fine tuning the parameters. Please let me know the best hyper parameters. The problem is that the code doesn't give the user an option to the user. We have to change the code to do so. Please help!

wiwengweng commented 6 years ago

@caiya55 Hi, caiya55, how can I change the layers? Is there any work that can share with us? Thanks.

caiya55 commented 6 years ago

Do you mean Upsampling2D ? Please check the PR that I put in this project. And Raphael also add some codes in his branch. Actually I am working on other image deforming correction not blurring. I am still fine tuning and try to get better results. I would share my work maybe in the next month, if everything goes well

wiwengweng commented 6 years ago

thanks, I saw that branch and PR. I will try that, maybe in two days when I finish my recent job.

RoshanSalian commented 3 years ago

The PSNR and SSIM of model trained on 2103 images of GOPRO dataset are very low compared to original paper. PSNR I obtained is only 25.2. Anyone else observing this?

ChefLiutao commented 3 years ago

The PSNR and SSIM of model trained on 2103 images of GOPRO dataset are very low compared to original paper. PSNR I obtained is only 25.2. Anyone else observing this?

I find this as well. The highest PSNR I have obtained now is just 25.7.

1z1y commented 1 year ago

The PSNR and SSIM of model trained on 2103 images of GOPRO dataset are very low compared to original paper. PSNR I obtained is only 25.2. Anyone else observing this?与原始论文相比，在 GOPRO 数据集的 2103 张图像上训练的模型的 PSNR 和 SSIM 非常低。我获得的PSNR仅为25.2。还有其他人观察到这一点吗？

I find this as well. The highest PSNR I have obtained now is just 25.7.我也发现了这一点。我现在获得的最高PSNR仅为25.7。

i also find this , how did you solve this problem