- An operating system that can run python 3
- Python 3 (downloadable at https://www.python.org/downloads/)
- Enough spare storage space to hold the python packages and the dataset (at least 10 GB recommended)
First, download the repository in any way you would see fit. Then run pip install -r requirements.txt to install the needed python packages.
The dataset is obtained from https://www.animecharactersdatabase.com using the programs in data_gathering/. First, scraper.py downloads images from the site. Then create_dataset.py formats those downloaded images into the format used to train, detecting the faces and then concatenating the face with an image of the face run through the Canny edge detection algorithm.
The image above is an example of an image in the dataset. The first half is the normal face and the second half is the face after being ran through the Canny edge detector.
It is recommended to use an NVIDIA GPU to train the model, however, if one is not available it will automatically train on CPU. To train the model from the dataset, run main.py. If saved models already exist in saved_model_paths/, it will load them up and resume training on them. Each hair color is trained and saved separately. The model is saved every 2000 steps to saved_model_paths/. The following is a link to the final model paths I had after training: https://mega.nz/folder/bJxkmSxI#qhTo7ELXU3xrCCY2qJBWcQ.
The results/logs while training can be seen in the log directory using tensorboard. To see the logs, make sure to have tensorboard installed and go into the project directory and run tensorboard --logdir logs. The logs are made using the validation data. Some results for orange hair after 8000 steps are below.
Condition (what the generator sees):
Generated (what the generator makes):
Real (what the generator was trying to make):
