Code for the paper "Generative Models and Model Criticism via Optimized Maximum Mean Discrepancy" (arXiv:1611.04488; submitted to ICLR 2017).

• General code for learning kernels for a fixed two-sample test, with Theano, is in two_sample; in particular, a Theano implementation of the estimator code is in two_sample/mmd.py.
• Code for the GAN models, using TensorFlow, is in gan. If you just want a TensorFlow implementation of the estimator code, see gan/mmd.py.
• Code for the efficient permutation test described in Section 3 is in the feature/bigtest branch of Shogun; look under shogun/src/shogun/statistical_testing. We're hoping to get that merged onto mainline shogun soon. An example of using it in the Python API is in two_sample/mmd_test.py; code to reproduce the experiments coming soon.

This code is under a BSD license, but if you use it, please cite the paper.