Code for the paper Disentangling factors of variation in deep representations using adversarial training by Mathieu, Zhao, Ramesh, Sprechmann and LeCun, NIPS 2016.
This code is a cleaned up version of our research code, and does not cover everything that is in the paper (yet). In particular, the code for the NORB dataset not there yet.
This code runs using torch. It is only tested on GPU. The following torch libraries are required: nn, cunn, cudnn, optim, image, nngraph.
The following libraries are optional (but strongly recommended: nnx, display.
We are using preprocessed versions of the dataset. You can download the mnist files here . You need to create a folder data/mnist containing these files.
We provide two model architectures: small and big. By default, the big architecture is used, it should work in all cases. The small architecture is faster to train and should work for MNIST and Sprites.
th train.lua --optimDisc adam
or
th train.lua --modelSize small --optimDisc adam
th train.lua --dataset sprites --nFeatures 14 --optimDisc adam
th train.lua --dataset yaleBExtended --swap2weight 1 --learningRateGen 0.0002 --learningRateDisc 0.0002