jongchan / distort-and-recover-cvpr18

Code for "Distort-and-Recover: Color Enhancement with Deep Reinforcement Learning", CVPR18

DISTORT-AND-RECOVER-CVPR18

Code for "Distort-and-Recover: Color Enhancement with Deep Reinforcement Learning", CVPR18

Overview

• You can pull&use the docker image from pjc0309/default_setup:latest to run the code. (TensorFlow 0.11.0rc, and other packages such as numpy/scipy)
• Before training,prepare MIT5K train/test images in separate folders (train/raw/, train/target/, test/raw/, test/target/). And edit the path in main.py accordingly.
• run_train.sh starts training.
• Use parse_test.py to parse the test results. (edit the paths accordingly)
• The training speed (iterations per second) should be between 20~40 it/sec. (When trained on i5-6600 and GTX 1080)

Data

• MIT5K Train/Val(RANDOM250) images. Resized to maximum side 500px, JPEG format. (including RANDOM250 list) LINK

distort-and-recover-cvpr18's Issues

where to find ./vgg16_pretrain.npy file

Could you please release the codes for inference?

Dataset link is broken!! :sob:

Please fix! Your version of the dataset is the best I could find!
(I have a backup if needed)

question about 'image_np+0.5'

Hi,

Thanks for sharing the code. May I know why you add 0.5 to image_np before taking any actions?

Regards,
Xiaodong

why action number mismatch

take_action in action.py has 28 types of actions, while in tools/action.py has only 12

will pretrained model be released?

I think in line 577. The op_strs will cause line index out of range.

DISTORT-AND-RECOVER-CVPR18/main.py

Line 577 in 5b1c8dc

image_init.save(os.path.join(test_result_dir, "%s_raw_%s.png" % (fn[i], op_strs[i])))

Hello, what is the version of your tensorflow?

get "duplicated prefix" issues

Hello, thank you for your wonderful work. I have a trouble when I run the test code. I use ./run_test.sh and get duplicated prefix which is so strange that I can not solve it.
Looking forward for your reply!