timsainb/Tensorflow-MultiGPU-VAE-GAN

Optimizing hyperparameters

Opened this issue · 3 comments

RiaanZoetmulder commented

Hey!

I have basically implemented a version of your model myself. I got it to work on MNIST but I also want to make it work on other datasets.

My question is how did you find the optimal hyperparameters on the celebA dataset, did you do a grid search?

regards,

Riaan

timsainb commented

Hi Riaan,

I started with hyperparameters similar to the papers/tutorials I referenced at the bottom of the notebook. Then I wrote a little parameter squashing function for the learning rates for the generator and discriminator. I believe I first wrote it to work with smaller images (16/16 or 32/32?) and then scaled up, changing hyperparameters as needed. But mostly just a few days of playing around with numbers and seeing what worked best.

Best,
Tim

RiaanZoetmulder commented

Ahhh good. I have got some settings now that work.

Have you by any chance gotten good results on cifar10?

timsainb commented

I haven't tried any other datasets other than MNIST. Good to hear that you got yours working!