Theano implementation of the auto-classifiers-encoders (ACE) from "Towards universal neural nets: Gibbs machines and ACE", Galin Georgiev, http://arxiv.org/abs/1508.06585

---------------------------Time spent on GTX 970 GPU:

-non-generative ACE: 2 sec/epoch for batch size = 10000 -generative ACE: -classification task (2): 8 sec/epoch for batch size = 10000 -density estimation task (1): 15 sec/epoch for batch size = 1000 ---------------------------Generative ACE options:

Density estimator (task=1) or classifier (task=2). Sampling density is either Gaussian (sampling_class=1) or Laplacian (sampling_class=2).

---------------------------Generative ACE output for MNIST:

Every 20 epochs, a 30 x 30 matrix of 900 images is saved:

-the top 300 images are from the so-called "creative" regime (see sub-section 3.2 in paper). -the middle 300 images are the model reconstructions of the first 300 test images in the "non-creative" regime. -the bottom 300 images are the first 300 test images from the original (possibly binarized) data set.

--------------------------Raw data:

Obtain the files below from http://yann.lecun.com/exdb/mnist/ and change datasets_dir in toolbox.py to the desired location:

MNIST: train-images.idx3-ubyte train-labels.idx1-ubyte t10k-images.idx3-ubyte t10k-labels.idx1-ubyte