// the find
bojone/vae
a simple vae and cvae from keras
A collection of VAE and CVAE implementations in Keras, written by Jianlin Su (bojone) as companion code to his Chinese-language blog posts on kexue.fm. It covers basic VAE, convolutional VAE, conditional VAE, VQ-VAE, and a vMF-VAE variant. This is educational reference code, not a library.
The vMF-VAE and VQ-VAE variants are included alongside the basic implementation, giving you a useful side-by-side comparison of latent space choices. The code is short and self-contained — each file is a standalone experiment you can read in one sitting. The CelebA experiment result is included with a concrete loss number, which is more honest than most tutorial repos. The blog posts (in Chinese) reportedly go deep on the math behind each variant.
Requires Python 2.7 and TensorFlow 1.8/1.13 with Keras 2.2.4 — this stack is years past EOL and will not run on any modern Python without significant porting work. No requirements.txt or setup file of any kind. The README is two sentences and a link; if you can't read Chinese, the companion blog posts are inaccessible. Last touched in 2021, so there's been no migration path offered to TF2 or PyTorch.