jiandandan001 / rl-restore Goto Github PK

🚩 Support arbitrary input size. Aug 25
🚩 Add Python3 compatibility. Aug 6
🚩 Training code is ready! Jun 15

• Framework

• Synthetic & real-world results

• Anaconda is highly recommended as you can easily adjust the environment setting.

  pip install opencv-python scipy tqdm h5py
  

• We have tested our code under the following settings:
  

  Python	TensorFlow	CUDA	cuDNN
  2.7	1.3	8.0	5.1
  3.5	1.4	8.0	5.1
  3.6	1.10	9.0	7.0

• Start testing on synthetic dataset

  python main.py --dataset moderate
  

  dataset: choose a test set among mild, moderate and severe

• ❗ Start testing on real-world data (support arbitrary input size)

  python main.py --dataset mine
  

  • You may put your own test images in data/test/mine/

• Dataset

  • All test sets can be downloaded at Google Drive or Baidu Cloud.

  • Replace test_images/ with the downloaded data and play with the whole dataset.

• Naming rules

  • Each saved image name refers to a selected toolchain. Please refer to my second reply in this issue.

• Download training images

  • Download training images (down-sampled DIV2K images) at Google Drive or Baidu Cloud.

  • Move the downloaded file to data/train/ and unzip.

• Generate training data

  • Run data/train/generate_train.m to generate training data in HDF5 format.

  • You may generate multiple .h5 files in data/train/

• Let's train!

  python main.py --is_train True
  

  • When you observe reward_sum is increasing, it indicates training is going well.

  • You can visualize reward increasing by TensorBoard.

  

The DQN algorithm is modified from DQN-tensorflow.

@inproceedings{yu2018crafting,
 author = {Yu, Ke and Dong, Chao and Lin, Liang and Loy, Chen Change},
 title = {Crafting a Toolchain for Image Restoration by Deep Reinforcement Learning},
 booktitle = {Proceedings of IEEE Conference on Computer Vision and Pattern Recognition},
 pages={2443--2452},
 year = {2018} 
}