Pytorch-Learning-to-See-in-the-Dark

Learning to See in the Dark using pytorch 0.4.0

Original tensorflow version

Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun, "Learning to See in the Dark", in CVPR, 2018.
Tensorflow code
Paper

Requirements

• 64 GB RAM
• GTX 1080
• PyTorch 0.4.0
• RawPy 0.10
• Scipy

The program have been tested on Ubuntu 16.04 and Windows 10.

Download Dataset

Download download_dataset.py from the original code and put it in the top level directory of this project and execute: python download_dataset.py

Training

python train_Sony.py

• It will save model and generate result images every 100 epochs.
• The trained models will be saved in saved_model/ and the result images will be saved in result_Sony/.
• The result of the current output is displayed on the right side of the image file with the ground truth counter part shown on the left side.

Testing

Download trained model

You can download the trained pytorch model here and put it in folder saved_model/.

python test_Sony.py

• The trained model is only for .ARW photos taken by Sony cameras.
• Pytorch somehow needs more GPU resources than Tensorflow. Therefore, it is impossible to take in the whole image.
• Testing will only take 1024 * 1024 pixels from the test images.
• This testing script is only for checking the performance of the trained model.
• The result will be saved in test_result_Sony with gt as ground truth images, scale as scaled images, ori as input images, and out as output images.

Todo

I have tried to feed the sliced input images into the model and put the result back to the original size. But there still remains two problems:

1. The edges of the sliced input images are quite obvious in the recovered result images.
2. There is no padding SAME in pytorch. Hence, images with incompetible shape might result in errors.

License

MIT License.

Please contact me you have any questions. twotwolavi@gmail.com