1. dcgan: Popular DCGAN model with 64x64 output.

2. rel_dcgan: DCGAN model with additional layers for 128x128 output and relativistic discriminator.

3. ss_rel_dcgan: rel_dcgan with self-supervision.

4. ss_rel_gan_improved: ss_rel_dcgan with spectral normalization and less frequent generator training.

5. ss_rel_resgan: ss_rel_gan_improved with residual blocks.

1. Install conda if you don't have it on your machine, ideally by installing Anaconda.

2. Create the conda environment for the project.

conda env create -f env.yml

3. Activate the environment.

conda activate clone-wars

4. Install pre-commit hooks while in the root of this project.

pre-commit install

The dataset contains copyrighted content so we cannot freely distribute it. Please contact us to obtain a private copy of this dataset.

1. Download dataset and place into data/full folder. This folder should contain another folder containing all images as its root.

2. Be sure to activate the clone-wars conda environment as specified in the Getting Started section.

3. Change directory into the model of your choice.

4. Run the training script.

python train.py

1. Be sure you have fully trained a model.

2. Within the directory of the model of your choice (currently only rel_dcgan and ss_rel_dcgan), run the eval script.

python eval.py -n NUMBER_OF_IMAGES -d DEVICE MODEL_FILENAME

If the device flag is not specified, the model defaults to using GPU (if available). Type in "cpu" to use CPU.

MIT