This repository contains the code for implementation of the FusionGAN model described in the paper Generating a Fusion Image: One's Identity and Another's Shape
The following are the dependencies required by this repository:
- PyTorch v0.4
- NumPy
- SciPy
- Pickle
- PIL
- Matplotlib
First, download the repository on your local machine by either downloading it or running the following script in the terminal
git clone https://github.com/aarushgupta/FusionGAN.git
Next, go through the instructions mentioned in the Dataset Preparation
section.
As the data is not publically available in the desired form, the frames of the required YouTube videos have been saved at this Google Drive link.
The link contains a compressed train folder which has the following 3 folders:
- class1_cropped
- class2_cropped
- class3_cropped
Download the data from the link and put the folders according to the following directory structure:
/FusionGAN_root_directory/Dataset/train/class1_cropped
/FusionGAN_root_directory/Dataset/train/class2_cropped
/FusionGAN_root_directory/Dataset/train/class3_cropped
The hyperparameters of the model have been preset. To start training of the model, simply run the train.py
file using the following command
python train.py
The code can also be run interactively using the train.ipynb
Jupyter Notebook provided in the repository.
- Train the model and add checkpoints.
- Polish and add code for preparing the dataset.
- Add test script for the model.
- Add keypoint estimation for quantitative evaluation.
- Remove the unused images in the dataset.