This repository contains a re-implementation of the CycleGAN architecture found in [1].
Experiment tracking was done using Weights and Biases (wandb
), selected runs are made public here.
Generating a Van Gogh painting from a photograph and converting it back to a photograph.
- Download the zip file containing preconfigured directories and datasets from here.
- Extract the zip file (and all other zip files within) and place contents inside the root directory.
- Set up a virtual environment and install the dependencies using the requirements.txt file by running
pip install -r requirements.txt
- Amend the configurations dictionary as required in the file train.py.
- Train a model by choosing a relevant GPU device with the following command
To run without a GPU use,
CUDA_VISIBLE_DEVICES=[GPU device] python src/train.py
python src/train.py
- Run evaluations by adjusting the configurations dictionary in evaluate.py according to the
required settings and by running,
python src/evaluate.py
- To compute metrics specify the desired dataset in metrics.py file and then run,
Currently metrics can only be computed on a CUDA enabled GPU.
python src/metrics.py
Pre-trained models can be downloaded from here.
- Hyperparameter tuning was not done for the model.