hanzi

Generate (fake) Chinese characters using GAN and VAE with tensorflow and keras.

Results

The resulting images are saved to *_samples/.

12MB gif (tanh in generator):

Using tanh in generator is more stable than Relu, and in some case no BN needed.
In the generator, relu is much better than tanh in learning strokes.
Relu is not suitable in the first layer of the generator.
判别器成中filter的数目决定了生成的字是否“残”。

Name		Name	Last commit message	Last commit date
Latest commit History 44 Commits
data		data
infogan_samples		infogan_samples
keras_samples		keras_samples
load-save-keras-model		load-save-keras-model
temp_processed_image/train/1		temp_processed_image/train/1
tensorflow_samples		tensorflow_samples
vae_samples		vae_samples
wgan_samples		wgan_samples
.gitignore		.gitignore
Chinese_inputs.py		Chinese_inputs.py
ChnSentiCorp_htl_all.txt		ChnSentiCorp_htl_all.txt
README.md		README.md
Untitled.ipynb		Untitled.ipynb
Untitled1.ipynb		Untitled1.ipynb
main_keras.py		main_keras.py
main_tensorflow.py		main_tensorflow.py
main_tensorflow_infogan.py		main_tensorflow_infogan.py
main_tensorflow_wgan.py		main_tensorflow_wgan.py
main_vae.py		main_vae.py
process-csv-file.py		process-csv-file.py
save_image.py		save_image.py
simsun.ttc		simsun.ttc