GANs in LibTorch.

‼️ According to this thread, C++ as a deployment mechanism will be fully supported, but not recommended to use for Training.

This repository contains a collection of GAN Architectures implemented in Libtorch (C++).
Training is done in Libtorch (C++) but, inference functions will be available in both Python and C++.
Please note that there can be accuracy degragration when loading a model in Python.

Setup

Docker Environment

The repo provides to environments, one for python (pytorch) and one for C++ (libtorch).
Python env will also setup a Jupyter server @ port 8888. To build both enviroments, run the following command.

docker-compose up --build

Compile Training Binary

Create an executable file of the C++ library.

docker exec -it libtorch bash
mkdir build
cd build
cmake ..

make -j

Training

./artium --dataset path/to/dataset --trainer dcgan|cyclegan --width 64|256 --height 64|256 --batch-size 2

DCGAN

A simple DCGAN Implementation is provided here

Example (Random Face Generation)

Results on celebA dataset (todo)

Limitations

• Input size is limited to 64 x 64.

CycleGAN

CycleGAN Implementation is provided here

Example (Style Transfer)

Results on monet2picture dataset (todo)

Limitations

• Input size is limited to 256 x 256.