IntroVAE-PyTorch

Simple implementation of IntroVAE toward MNIST

Abstract

This is the simple implementation of the paper- IntroVAE: Introspective Variational Autoencoders for Photographic Image Synthesis. We only test the idea on MNIST. The LSUM and CelebA example can be found in other implementation.

Usage

• Train the model for 100 epoch:

$ python3 main.py --epochs 100

• Sample for 20 digits images

$ python3 main.py --n 20

Result

The above figure demonstrates the loss curve during IntroVAE training. We only test on the case whose size is 64x64.