Skip to content

airaria/GAN_hanzi

Repository files navigation

hanzi

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

Results

The resulting images are saved to *_samples/.

Training process

12MB gif (tanh in generator):

Training process

Some issues

  • 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”。

About

Generating (fake) Chinese characters using GAN

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages