deep-photo-enhancer
deep-photo-enhancer copied to clipboard
mtics
Solvd some issues on the implementation
Hi I have solved some issues with the implementation and got it working properly with 2-way gan . Take a look and let's see if we can merge efforts https://github.com/MrRobot2211/pytorch-deep-photo-enhancer
Hi I have solved some issues with the implementation and got it working properly with 2-way gan . Take a look and let's see if we can merge efforts https://github.com/MrRobot2211/pytorch-deep-photo-enhancer
Thanks a lot! Could you please show the result and losses? For now,I have used it to train 155 epochs, but the result is awful, and D loss is about -400000000, G loss is about -150000000.
Hi I am testing it on a reduced set I am geting
[Epoch 300/300] [Batch 2/2] [D loss: -8273.702557] [G loss: 1007.653687]
Loss loss: 0.003262 PSNR Avg: 24.864868 Loss loss: 0.003182 PSNR Avg: 24.919203 Final PSNR Avg: 24.919203
And the results are as expected, I think the psnr is more infomative than the actual losses. In any case I am taining on a reduced set. Can you upload your train/test data (as they are in the directories) somewhere ? so I can download it and certify that we are actually doing the same.
Hi I am testing it on a reduced set I am geting
[Epoch 300/300] [Batch 2/2] [D loss: -8273.702557] [G loss: 1007.653687]
Loss loss: 0.003262 PSNR Avg: 24.864868 Loss loss: 0.003182 PSNR Avg: 24.919203 Final PSNR Avg: 24.919203
And the results are as expected, I think the psnr is more infomative than the actual losses. In any case I am taining on a reduced set. Can you upload your train/test data (as they are in the directories) somewhere ? so I can download it and certify that we are actually doing the same.
when learning_rate was 1e-3, I got these: [Epoch 229/300] [Batch 9/29] [D loss: -560770361.040283] [G loss: -275414976.000000]
so I changed it to 1e-5, then got these: [Epoch 300/300] [Batch 29/29] [D loss: -173917.736206] [G loss: -56673.386719]
PSNR Avg: 12.403313 PSNR Avg: 11.698874 PSNR Avg: 11.964977 PSNR Avg: 11.624313 PSNR Avg: 11.389729 PSNR Avg: 11.009916 PSNR Avg: 11.019134 Final PSNR Avg: 11.019134
I upload them on dropbox. It contains train_checkpoints, train_images, constant.py, and the dataset I used (in mini.tar)
Great thank you I'll try to run it as soon as possible so that we can compare. In any case you have some differences with the original code, which may be related to the bad performance, the main one being they use instancenorm instead of batchnorm.
Great thank you I'll try to run it as soon as possible so that we can compare. In any case you have some differences with the original code, which may be related to the bad performance, the main one being they use instancenorm instead of batchnorm.
I used your code to train, maybe there is something wrong in my dataset or hyperparameters?
when lr = 1e-3, the result is really awful...
Yes you should set lr to 1e-5 . At 1e-3 the losses fall nre ally fast but the psnr does not improve. Try pulling it now. I just made another commit
Yes you should set lr to 1e-5 . At 1e-3 the losses fall nre ally fast but the psnr does not improve. Try pulling it now. I just made another commit
I used my mini dataset, and have trained 60 epochs, but the d loss is -20,000, and G loss is -3,000.
is there something wrong in my dataset?
HAve you set lr to 1e-5 ? Run it 300 epochs and then check the quality of the results, and the psnr.
HAve you set lr to 1e-5 ? Run it 300 epochs and then check the quality of the results, and the psnr.
yes, i set it. my dataset may take a long time. For now, it just trained 105 epochs. I will wait until it is finished.
We may have to revise batch size, afterwards. I am also running it.
Could you upload your dataset? I want to compare them to see if there is any thing wrong in my dataset.
Yes, when this run end I'll upload it . In any case with your dataset and a batch size of 6 I am geting this at epoch 22 Done training discriminator on iteration: 4 [Epoch 24/300] [Batch 6/14] [D loss: -2097.180758] [G loss: 4727.187500] Done training discriminator on iteration: 5 So i would rather think it was the batchize. That insome way affect the D/G ratio
Another possible issue is the pretrain .... If you are using my pretrain it probably differs a lot from yours because of the dataset difference
Yes, when this run end I'll upload it . In any case with your dataset and a batch size of 6 I am geting this at epoch 22 Done training discriminator on iteration: 4 [Epoch 24/300] [Batch 6/14] [D loss: -2097.180758] [G loss: 4727.187500] Done training discriminator on iteration: 5 So i would rather think it was the batchize. That insome way affect the D/G ratio
I used my pretrain models before, and the psnr was alwasy about 11.
I finished this train, and got these: [Epoch 300/300] [Batch 29/29] [D loss: -152075.056702] [G loss: -52065.578125]
PSNR Avg: 10.446398 PSNR Avg: 10.814289 PSNR Avg: 10.811638 PSNR Avg: 10.657595 PSNR Avg: 10.770314 PSNR Avg: 10.719786 PSNR Avg: 11.297231 Final PSNR Avg: 11.297231
it seems to be worse than before...
Hi Let's try debugging. I have trained my 1way gan on your dataset and I am geting good output results. Can you try it? I supect we are doing a small amount of generator passes in 2waygan
I tried 1WayGan_train on my mini dataset, but the result is not so good.
[Epoch 300/300] [Batch 28/29] [D loss: 150.137268] [G loss: 28.260000] [Epoch 300/300] [Batch 29/29] [D loss: 195.740005] [G loss: 28.260000]
Final PSNR Avg: 10.712687
I really don't know what makes the difference. I even didn't use multiple gpus.
Sorry I meant the 1way pretrain. I have not changed anything in the 1waygan train
Oh, I also used it to pre-train my dataset, I will train on 2 way gan.
I saw you update the code last night, which code should I use, 2WayGan_Train or V2?
Hi none, I was just saying that the 1waygan pretrain results seem okey. I Tried running on your dataset changing btchsize and the results are not good either. It may have been that my dataset was too small and the network was just memorizing. I am exploring making the LAMBDA parameter adaptive as in the paper. Do you have any leads on how to do it?
I don't know either. The author tried his model on 5,000 pics, I only use 1/10 of them.
I have changed LAMBDA to 1, the result was awful.
Yes but ...it has to vary as we train you ensure the lipschitz condition. And get the nice swirls of the paper. Do not worry I think I have got it. Have you checked if the croppings are compatible with the original code and paper?
On Wed, Apr 22, 2020, 20:46 Zhiwei Li [email protected] wrote:
I don't know either. The author tried his model on 5,000 pics, I only use 1/10 of them.
I have changed LAMBDA to 1, the result was awful.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/mtics/deep-photo-enhancer/issues/4#issuecomment-618095832, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFTMVKBCHXU433L5E4HUBI3RN56TVANCNFSM4MNQM6TA .
Yes but ...it has to vary as we train you ensure the lipschitz condition. And get the nice swirls of the paper. Do not worry I think I have got it. Have you checked if the croppings are compatible with the original code and paper? … On Wed, Apr 22, 2020, 20:46 Zhiwei Li @.***> wrote: I don't know either. The author tried his model on 5,000 pics, I only use 1/10 of them. I have changed LAMBDA to 1, the result was awful. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#4 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFTMVKBCHXU433L5E4HUBI3RN56TVANCNFSM4MNQM6TA .
not yet. I'm new at deep learning, there are a lot of things I don't understand really....
I merged my master branch and dev branch, you can take a look at master. With the help of my teacher and seniors, the code has been greater than the version yours based on.
Yes, when this run end I'll upload it . In any case with your dataset and a batch size of 6 I am geting this at epoch 22 Done training discriminator on iteration: 4 [Epoch 24/300] [Batch 6/14] [D loss: -2097.180758] [G loss: 4727.187500] Done training discriminator on iteration: 5 So i would rather think it was the batchize. That insome way affect the D/G ratio
Could you upload your dataset?
