• Python 3
• Tensorflow > 1

• 3 models, saae.py, aae.py and gsvae.py.
• utils.py contains all the functions used to build and train models
• index.py contains index information for labels.
• svhn.py creates pickle file for svhn data.

• create a data directory in the parent directory of the code folder.
• The mnist dataset should download automatically to ../data when saae.py, aae.py or gsvae.py are run.
• for svhn download svhn files to ../data and run svhn.py file which creates a pickle of the data (~5gb).
• run any of the 3 models saae.py, aae.py or gsvae.py.

python saae.py 1000 1000 10 10 MSE GAUSS 0.01 0.1 0.1 1.0 0.3 0.1 mnist True 60000 10000 100 100 127 0

units in hidden layer 1 - any +ve int
units in hidden layer 2 - any +ve int
number of y units - any +ve int
number of z units - any +ve int
likelihood / reconstruction - MSE or ABS
posterior / latent - GAUSS or LAP
reconstruction learning rate - float 0-1
regularisation learning rate - float 0-1
semi-supervised learning rate - float 0-1
dropout rate - float 0-1
gumbel-softmax temperature - float 0-1
gaussian noise - float 0-1
dataset - mnist or svhn
use pre-defined labels - True
number training points - between 5000 and 60000
number validation / test points 10000
epochs - any +ve int
number of labels - any +ve int < number training points
random seed - any +ve int
counter - any +ve int

python aae.py 1000 1000 10 10 MSE GAUSS 0.01 0.1 0.1 1.0 mnist True 60000 10000 100 100 127 0

units in hidden layer 1 - any +ve int
units in hidden layer 2 - any +ve int
number of y units - any +ve int
number of z units - any +ve int
likelihood / reconstruction - MSE or ABS
posterior / latent - GAUSS or LAP
reconstruction learning rate - float 0-1
regularisation learning rate - float 0-1
semi-supervised learning rate - float 0-1
dropout rate - float 0-1
dataset - mnist or svhn
use pre-defined labels - True
number training points - between 5000 and 60000
number validation / test points 10000
epochs - any +ve int
number of labels - any +ve int < number training points
random seed - any +ve int
counter - any +ve int

python gsvae.py 1000 1000 10 10 MSE GAUSS 0.01 0.1 1.0 0.3 1.0 mnist True 60000 10000 100 100 127 0

units in hidden layer 1 - any +ve int
units in hidden layer 2 - any +ve int
number of y units - any +ve int
number of z units - any +ve int
likelihood / reconstruction - MSE or ABS
posterior / latent - GAUSS or LAP
reconstruction + regularization learning rate - float 0-1
semi-supervised learning rate - float 0-1
dropout rate - float 0-1
gumbel-softmax temperature - float 0-1
gaussian noise - float 0-1
dataset - mnist or svhn
use pre-defined labels - True
number training points - between 5000 and 60000
number validation / test points 10000
epochs - any +ve int
number of labels - any +ve int < number training points
random seed - any +ve int
counter - any +ve int