Abstract

The idea of this project is to generate some characters “invented” by Sketch-rnn. The model is trained from ~10k Chinese characters used in Japan. These characters are decoded from random IID Gaussian latent vectors with temperature 0.1. They are orgnized in one paragraph to disguise as real ones.

Results

Files

• FakeHanzi.ipynb - guideline and executable code.
• Checkpoint files are stored in my google drive.

Notes

• Test on Google Colab.

Reference

Sketch-RNN.
KanjiVG Dataset.