Conditional cartoon faces incorporating a landmark loss in CycleGAN

Report

The Report_FaceToCartoon.ipynb covers the details of the project.

Demo

You find an interactive demo under code/Demo.ipynb.

Code

More details to the code are in code/README.md.

How to train

1. Clone this repo using git clone https://github.com/fs2019-atml/face-to-cartoon.git
2. Get the dataset from google drive https://drive.google.com/open?id=12vU_Dkn13KqsVy5LOYlpoTJGWmzPqPQh
3. Untar the archive to ./code/datasets/ to have all the images under ./code/datasets/faces/{cartoon/*, real/*} (e.g. tar xf faces.tar.gz inside ./code/datasets)
4. Install dependencies (See below)
5. a) Invoke python train.py --gpu_ids=0 to train on cuda. b) Invoke python train.py to wait forever.

Dependencies

We suggest an anaconda environment with the following:

conda install numpy pyyaml mkl mkl-include setuptools cmake cffi typing
conda install pytorch torchvision -c pytorch # add cuda90 if CUDA 9 (or magma-cuda90)
conda install visdom dominate -c conda-forge # install visdom and dominate

If you run in troubles with Cuda try to downgrade pytorch to version 0.41.