Learning to See in the Dark

Joe Hines and John Zlotek But based entirely on the work found here

Developers

Python Notebooks

Since we did all of our training and testing on Google's Colaboratory, we decided to provided our notebooks in the notebooks/ directory. This will run through the training and validation of a model, and provide access to the outputs of the tests.

This allows you to quickly get access to and play around with the network. However, if you're looking to get more hands on, keep reading.

CLI

We provide all of the necessary files to train, modify, and use this network in the src/ directory of this project. While we do provide our own trained model, we encourage explorations. The following are instructions/requirements for various items.

NOTE: WE ONLY SUPPORT THE SONY DATASET AND MODEL IN THIS VERSION

• Datasets
  • The original paper used a Sony and Fuji dataset, as these cameras have different RAW formats. IF YOU WISH TO USE .DNG FILES, WHICH MOST SMARTPHONES CAN TAKE, USE THE SONY MODEL. IT IS COMPATIBLE WITH .ARW AND .DNG FILES
  • Downloading and unzipping the dataset will take a while as well as roughly ~150gb of disk space, as the dataset is 25GB (Sony)
  • To download (inside the src/ directory):
    • python3 download_datasets.py
• Our trained model
  • We trained a model with 501 epochs, although we recommend 4001 per the paper (but deadlines can be a tricky thing).
  • This model requires a ~100MB of disk space, and is unzipped into src/trained_model/
  • To download (from inside src/)
  • python3 download_our_model.py
• Train the model
  • If you choose to train the model yourself, it is good to be aware of the batch_size variable. A high batch number could provide better results, but will use impractical amounts of memory. Batch sizes less than 30 images can use less than 6GB of RAM, but some experimentation may be required on your system.
  • From within src/
    • python3 train_Sony.py
• Test the model
  • This requires a model to exist in src/trained_model/
  • From inside src/
    • python3 test_Sony.py
• CLI application is run via python correct_img.py
  • Usage:
    • python correct_img.py {source} {optional: destination}
    • Not supplying a destination will create a .png file in the same directory of the source image