This project is simply for fun. I was curious what a VAE would learn from the complete set of Hanzis (about 30k characters). The training set can be easily obtained from *.ttf. The results were not unexpected: VAE learned a space where the only most common ingredients (Bushou, 部首) resided.
However, VAE was unable to reconstruct other ingredients. It was pretty disappointing but expected (based on statistics).
Input a character, the VAE only outputs the
Python 2.7
- Scipy (conda install scipy)
- Matplotlib
- PIL (conda install PIL)
python vae.py