DCGAN in C++

A implementation of a Deep Convolutional Generative Adversarial Network (DCGAN) written in C++. This project was done towards learning C++ and some of its tools. It trains a Generator and Discriminator based on the MNIST dataset on CPUs. The code was inspired by the PyTorch tutorial on the C++ framework PyTorch C++ Tutorial.

Results

As the GAN trains, the Generator learns to map a 100-dimensional latent noise vector into handwritten digits. Below is the progression of the generated samples across training steps:

Samples from the Generator from Epoch 1 (left) to Epoch 10 (right).

Prerequisites

To build and run this project, your system needs:

• C++ Compiler supporting C++20 (GCC/Clang)
• CMake (>= 3.10)
• LibTorch (The PyTorch C++ API)
• OpenCV (For saving inference tensors as .png files)

Dataset

LibTorch requires the uncompressed, raw MNIST IDX files. Create a data/ directory and run the mnist.py file for dowloading the files.

Build

This project uses CMake to handle linking LibTorch and OpenCV. From the root of the project, run:

mkdir build
cd build
cmake ..
make -j$(nproc)

Usage

After compiling, a bin/dcgan binary is created. It takes a command-line argument to choose between training and inference. To train the DCGAN, use:

./bin/dcgan train

This saves training samples and the model checkpoints in checkpoints/ directory in the root. For inference:

./bin/dcgan inference

Which saves a inference-sample.png in checkpoints/.