MNIST experiment - Githubissues

arianhosseini / GMVAE

Gaussian Mixture Variational Autoencoders

1 stars 1 forks source link

MNIST experiment #1

Open arianhosseini opened 6 years ago

arianhosseini commented 6 years ago

make changes to GMVAE.py to get Bernoulli parameters (instead of Gaussian mus and vars) from pygx MLP.
implement getHyperMNIST function for MNIST hyperparameters in hyper.py (e.g. the pygx MLP dimensions, hyperparameter for mode =MNIST/Spiral for later)
implement sampling from Bernoulli (y_samples) in sample function in main.py

slachapelle commented 6 years ago

Detail regarding the sampling of MNIST:

In the spiral data sampling, we sampled "all the way" in the sense that, for each random var in the graph of P, we sample according to a distribution. In particular, we sample from p(y|x) at the very end of the process to get the sample of y.

In MNIST, everything stays the same, but we won't sample from p(y|x) at the end. If we do that, the pixels of the sampled images will be white or black, no in between (since p(y|x) is a bernoulli). We don't want that. Instead, just use the mean of the bernoulli distribution in the images (this way, the images will have grayscale).