-
dcgan
: Popular DCGAN model with 64x64 output. -
rel_dcgan
: DCGAN model with additional layers for 128x128 output and relativistic discriminator. -
ss_rel_dcgan
:rel_dcgan
with self-supervision. -
ss_rel_gan_improved
:ss_rel_dcgan
with spectral normalization and less frequent generator training. -
ss_rel_resgan
:ss_rel_gan_improved
with residual blocks.
-
Install conda if you don't have it on your machine, ideally by installing Anaconda.
-
Create the conda environment for the project.
conda env create -f env.yml
- Activate the environment.
conda activate clone-wars
- 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.
-
Download dataset and place into
data/full
folder. This folder should contain another folder containing all images as its root. -
Be sure to activate the clone-wars conda environment as specified in the Getting Started section.
-
Change directory into the model of your choice.
-
Run the training script.
python train.py
-
Be sure you have fully trained a model.
-
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.