PyTorch implementation of “Monge-Ampère Flow for Generative Modeling” arXiv:1809.10188

How to run the code

Density estimation of MNIST

python density_estimation.py -dataset MNIST -hdim 1024 -Nsteps 100 -train -cuda 7

Variational free energy of Ising

python variational_free_energy.py -L 16 -fe_exact -2.3159198563359373 -train -cuda 7 -hdim 512 -Nsteps 50 -Batchsize 64 -symmetrize

Plots in the paper

• MNIST NLL

python paper/plot_nll.py -outname nll.pdf 

• Gaussianization MNIST

python density_estimation.py -hdim 1024 -Nsteps 100 -epsilon 0.1 -checkpoint data/learn_mnist/Simple_MLP_hdim1024_Batchsize100_lr0.001_Nsteps100_epsilon0.1/epoch-1.chkp -show 

• Direct sample Ising

python variational_free_energy.py -hdim 512 -Nsteps 50 -checkpoint data/learn_ot/ising_L16_d2_T2.269185314213022_symmetrize_Simple_MLP_hdim512_Batchsize64_lr0.001_delta0.0_Nsteps50_epsilon0.1/epoch-1.chkp -show  -L 16  -symmetrize 

Reference: Exact Ising free energy density at critical temperature on $L\times L$ lattices (For details see Appendix B of the paper)

$L$	periodic	open
4	-2.33604476445	-1.9470001244979966
8	-2.3227349295609376	-2.1909718508291
16	-2.3159198563359373	-2.272901214087426
32	-2.3140498159960936	-2.2993352217736573
64	-2.3135805785878905	-2.3080749864821253