search

shekkizh / WassersteinGAN.tensorflow

Tensorflow implementation of Wasserstein GAN - arxiv: https://arxiv.org/abs/1701.07875
MIT License
416 stars 130 forks source link

WGan's result in celeba training-data #8

Closed zhangqianhui closed 7 years ago

zhangqianhui commented 7 years ago

It looks that wgan generator face images have low quality compared to the origin DCGAN.

shekkizh commented 7 years ago

Quality is certainly a relative term. My take from WGAN is that earlier training didn't have a quantifiable measure to understand the learning but with this modified formulation you get an idea of your model. This is actually better than training long hours and hoping your results turn out good or visualizing model every hour or so to check generated results.

Also, note that the results I posted in Readme are from randomly sampled points.

zhangqianhui commented 7 years ago

thanks