Alvin2112
commented
3 months ago Thank you very much for your attention to this work. I have graduated and left school, so I need some time to find the pre-trained model. If I find it, I will upload it to the GitHub homepage as soon as possible.
发件人: jingyuSong @.> 发送时间: 2024年8月2日 14:16 收件人: Alvin2112/PSIDNet @.> 抄送: Subscribed @.***> 主题: [Alvin2112/PSIDNet] Could you please share your trained weights? (Issue #1)
Hi, thank you for sharing your great results. I am training PSIDNet using your code base. I am training PSIDNet using KITTI Odometry Dataset. However, I keep experiencing learning failure around 10epochs. In particular, I can see that the loss value of stage 3 increases sharply around 10epochs. I think it may be because the normalization using batch normalization is not properly used throughout the network. Or maybe I am doing something wrong. In any case, this training is too painful for me.
Could you please share your trained weights? I would like to use them to compare your results with mine (Haze removal method).
PSIDNet trained for about 10 epochs works well. (Compared to other SOTA Stereo-based haze removal methods, e.g. SRDNet, BidNet,) Below is a haze image and a image of haze removal using a trained PSIDNet. left_i0.03.png (view on web)https://github.com/user-attachments/assets/68c6245e-e8ff-4212-b4c4-87f3ff99df45 left_p0.03.png (view on web)https://github.com/user-attachments/assets/9df665da-a91f-4286-9e3b-04c1bce7a654 However, according to my own evaluation results(Using around 4500 images), PSNR dB is only 30. I think it is probably a problem caused by not being able to continue training for a long time. Below is my training log. image.png (view on web)https://github.com/user-attachments/assets/2322391c-905b-483b-bea7-2368e79acc31 Thank you.
― Reply to this email directly, view it on GitHubhttps://github.com/Alvin2112/PSIDNet/issues/1, or unsubscribehttps://github.com/notifications/unsubscribe-auth/A63JJODZTZSMZE7IUMLIYO3ZPMP4HAVCNFSM6AAAAABL3442XKVHI2DSMVQWIX3LMV43ASLTON2WKOZSGQ2DIMJVGIZTEOI. You are receiving this because you are subscribed to this thread.Message ID: @.***>
Hi, thank you for sharing your great results. I am training PSIDNet using your code base. I am training PSIDNet using KITTI Odometry Dataset. However, I keep experiencing learning failure around 10epochs. In particular, I can see that the loss value of stage 3 increases sharply around 10epochs. I think it may be because the normalization using batch normalization is not properly used throughout the network. Or maybe I am doing something wrong. In any case, this training is too painful for me.
Could you please share your trained weights? I would like to use them to compare your results with mine (Haze removal method).
PSIDNet trained for about 10 epochs works well. (Compared to other SOTA Stereo-based haze removal methods, e.g. SRDNet, BidNet,) Below is a haze image and a image of haze removal using a trained PSIDNet.
However, according to my own evaluation results(Using around 4500 images), PSNR dB is only 30.
I think it is probably a problem caused by not being able to continue training for a long time.
Below is my training logs.
Thank you.