search

aiff22 / PyNET-Bokeh

Rendering Realistic Bokeh Images with PyNET
http://people.ee.ethz.ch/~ihnatova/pynet-bokeh.html
Other
170 stars 29 forks source link

Training from scratch. #11

Open make-j64 opened 3 years ago

make-j64 commented 3 years ago

Hi @aiff22, Im training from scratch with EBB dataset, result at train_iters 40000 of level 1 is wrong, like this: https://ibb.co/bByGP4T

Is it normally? Should i continue training to 100000 iters as your recommend?

Thanks.

C-water commented 3 years ago

I meet the same problem, the color of the image was distorted at level 1.

C-water commented 3 years ago

@unknow-dev hi, have you found a solution? I tried to modify loss function into L1*100, and adjust the lr of level-1, but it did not work. If you find the reason, could you share it with me? Thank you.

C-water commented 3 years ago

@unknow-dev Hi, I recommend using the PyTorch code for training, and the author provides a basic version that requires modifications to level1 and level2. Unfortunately, the Tensorflow code I trained three times has gradient explosion that caused color distortion at level 1.

make-j64 commented 3 years ago

@unknow-dev Hi, I recommend using the PyTorch code for training, and the author provides a basic version that requires modifications to level1 and level2. Unfortunately, the Tensorflow code I trained three times has gradient explosion that caused color distortion at level 1. Hi @chenquan-hdu , Thanks for your information, i will try.

JialeHu97 commented 2 years ago

@unknow-dev hi, have you found a solution? I tried to modify loss function into L1*100, and adjust the lr of level-1, but it did not work. If you find the reason, could you share it with me? Thank you.

@chenquan-hdu Hi, have you found a solution to the problem of image color distortion? If you find the solution, could you share it with me? Thank you!

C-water commented 2 years ago

Hi, I don't know how to solve this problem. Therefore, I used the PyTorch code for training, and the author provides a basic version (pynet-pytorch) that requires modifications to level1 and level2. PyTorch-version code is valid.

JialeHu97 commented 2 years ago

@chenquan-hdu , Thanks for your information.