search

tatsy / keras-generative

Deep generative networks, coded with Keras.
122 stars 28 forks source link

Noisy Images are produced when using MNIST on ALI #4

Closed xtarx closed 3 years ago

xtarx commented 6 years ago

The generator loss keeps increasing , while the Discriminator loss converges to 0, resulting in Noise like images on MNIST training dataset.

tatsy commented 6 years ago

Hi @xtarx, Actually, I have not tested MNIST dataset on any network model. So, I cannot guarantee that the models properly work for it. Also, the ALI model is often very unstable as far as I tested CelebA dataset. So, I think some parameter tuning is necessary for each dataset. As you can find in the original paper, they use quite different networks for different datasets.