==============================
A Pytorch implementation of Perceptual Losses for Real-Time Style Transfer and Super-Resolution along with Instance Normalization: The Missing Ingredient for Fast Stylization.
Created to learn PyTorch and mess around with best practices.
You can run the Dockerfile by downloading the file on its own and using docker build .. Alternatively use the requirements.txt, but the PyTorch version can be 1.0+.
For the dataset I used the MSCOCO 2014 Training images dataset [80K/13GB] (download) - the same dataset used in Johnson et al. Required command line arguments for train.py are cuda, data_dir, and save_model_dir. Required command line arguments for stylize.py are cuda, content_image, output_image, and model. Use train.py --help or stylize.py --help for more details.
Example training command:
python src/train.py --cuda 1 --datasets/train2014 --save_model_dir saved_models/checkpoints --style_image wave_crop.jpg
├── LICENSE
├── requirements.txt <- Use `pip install -r requirements.txt`
├── README.md
├── Dockerfile <- Download this file and use `docker build` to create a Docker image with all dependencies.
├── setup.py <- makes project pip installable (pip install -e .) so src can be imported
└── src <- Source code for use in this project
├── __init__.py <- Makes src a Python module
│
├── train.py <- Run this to train
│
├── stylize.py <- Run this to stylize images given a trained model.
│
├── options <- Files for command line options
│
└── networks <- Code for defining the network structure and loss functions