DCGAN-CelebA-PyTorch-CPP

DCGAN Implementation (on CelebA dataset) using PyTorch C++ Frontend API (Libtorch)

• Training Code location: src/main.cpp
• Generator and Discriminator Definition: include/network.hpp
• Dataset Class: include/dataset.hpp and src/dataset.cpp
• Tested on Libtorch Version: Stable 1.4.0 (cxx11 ABI) with and without CUDA (10.1), Linux, OpenCV 4.1.0

How is this different from dcgan sample of PyTorch?

1. This loads a custom dataset (which is not in the dataset class of PyTorch) - CelebA.
2. Since some users prefer using Sequential Modules, so this example uses Sequential Module.

Utility Functions (to visualize images & create animation), and architecture is inherited from the PyTorch Example on DCGAN (https://github.com/pytorch/examples/blob/master/cpp/dcgan/).

Please note that this is in no way targeted to achieve a certain accuracy, but only focuses on creating an example template for DCGAN using Libtorch on CelebA Dataset.

Steps to Follow

Note: This project requires OpenCV built from source. Make sure you also have Pillow (to save animation), NumPy, Matplotlib for running files in utils/ folder.

1. Create a build directory: mkdir build/
2. Change the directory path (absolute path) to libtorch in CMakeLists.txt file. Then configure using CMake: cmake ..
3. Create an output directory (in the build directory) to store the results & save checkpoints: mkdir output/
4. Build your project: make
5. Execute: ./bin/example/
6. The saved checkpoints and outputs will be at output/ directory
7. To visualize, go back to the main directory: cd ../
8. Execute: python3 utils/display_samples.py
9. The outputs will be stores in build/output/output_images/ directory
10. If you want to make an animation, run: python3 utils/visualize.py and it will save the animation for you in build/output/output_animation/ directory

Blog

Find more about DCGAN on my blogs here:

1. https://krshrimali.github.io/DCGAN-using-PyTorch-CPP/
2. https://krshrimali.github.io/DCGAN-using-PyTorch-CPP-Part-2/

Results

This is the output from random noise (batch of) images after ~10 epochs of training:

Happy learning!