Trying to reproduce the results from paper

[eugene-goncharov]

Hi,

I'm trying to reproduce the results from paper and the reward never goes above zero.

Could you please let me know if it is possible to reproduce the results with one worker. I'm training on Windows and Chainer is not working on more than one worker.

Many thanks
Eugene

[satoshi-kosugi]

I guess you are using unexpected versions of the libraries.
After installing the following versions of the libraries, please try using multiple workers.

• Python 3.6
• Chainer 7.0
• ChainerRL 0.6

[eugene-goncharov]

I've tried the exact versions of libraries. Using these libraries it is required to do some minor change in the code, because the Chainer API is different.

But anyway, on Windows multiple workers are not working. I'm trying to reproduce the results on Linux. I'll update this issue later today.

[eugene-goncharov]

I've tried on Linux, after 100k steps the reward still below zero, like -5 or something near.

Should it go above zero? How many steps do you do during training?
Is it possible to share the training curves and discriminator weights?

[satoshi-kosugi]

The reward should become larger than at the start of the training but does not need to become larger than zero.

I trained the model for 96000 steps, which is the default value in settings/photo_enhancement.yaml (96000 = 16 processes * 6000 n_update).

To reproduce the paper's results, please try training and testing the model as described in README.md.
That is, to train the model,

python train.py settings/photo_enhancement.yaml logs

and to test the model,

python test.py settings/photo_enhancement.yaml logs --result_dir logs/20200115T223451.986831/96000_finish/test_results --load logs/20200115T223451.986831/96000_finish/

where 20200115T223451.986831 should be replaced with the time when you start the training.