search

hwalsuklee / tensorflow-mnist-VAE

Tensorflow implementation of variational auto-encoder for MNIST
492 stars 182 forks source link

Training on all the digits at once #2

Open udion opened 6 years ago

udion commented 6 years ago

Hi, While training on multiple images of MNIST I was not getting any reasonable output upon generation of images, however when I trained on a single digit (say 3) I was able to generate reasonable looking digits (like 3).

I tried increasing the iteration for training with multiple digits in dataset together, but still I am not getting reasonable outputs like yours, any suggestions?

hwalsuklee commented 6 years ago

Hi, from my experience, I suggest you to use deeper architectures. Generated images are a bit sensitive to network architecture.

DCGAN like architecture would be good start point.

udion commented 6 years ago

I have tried dcgan as well, but it also doesn't when dataset has mixture of all the digits. any suggestions?

lerndeep commented 4 years ago

@udion can you let me know how can I train network just one class and generate one class output?