README.md

MDMA

PyTorch implementation of the Marginalizable Density Model Approximator — a density estimator that provides closed-form marginal and conditional densities and enables rapid sampling.

For details, see:

Dar Gilboa, Ari Pakman and Thibault Vatter, Marginalizable Density Models (2021)

Requirements

• python>=3.6
• numpy>=1.20.2
• pytorch>=1.0.0
• pandas>=1.2.3

Optional for visualization and plotting: matplotlib and tensorboardX.

Structure

• mdma/models.py: Implementation of the MDMA class.
• mdma/fit.py: Fitting an MDMA model.
• mdma/utils.py: Various auxiliary functions.
• experiments: Additional code for reproducing the experiments in the paper.

Usage

Below, example commands are given for running experiments.

Download datasets

Download UCI datasets:

bash download_datasets.sh

Toy 3D density estimation

Density estimation on a toy dataset of two spirals, showing the ability of MDMA to compute marginal and conditional densities.

Fit two spirals density using MDMA, and plot marginals and conditionals:

python3 toy_density_estimation.py --dataset spirals

Possible values for dataset are spirals, checkerboard, gaussians.

For a two spiral dataset, the samples and marginal histograms of the data take the following form: Samples from the trained MDMA model and the learned marginal densities evaluated on a grid are indistinguishable: MDMA also provides closed-form expression for all conditional densities:

UCI density estimation

Fit UCI POWER dataset using MDMA:

python3 uci_density_estimation.py --dataset power \
                                  --m 1000 \           # Width of tensor network
                                  --r 3 \              # Width of univariate CDF networks
                                  --l 2 \              # Depth of univariate CDF networks
                                  --batch_size 500 \
                                  --n_epochs 1000 \
                                  --lr 0.01 

Possible values for dataset are power, gas, hepmass, miniboone.

Fit UCI POWER dataset using the non-marginalizable variant nMDMA:

python3 uci_density_estimation.py --dataset power \
                                  --m 1000 \           # Width of tensor network
                                  --r 3 \              # Width of univariate CDF networks
                                  --l 2 \              # Depth of univariate CDF networks
                                  --batch_size 500 \
                                  --n_epochs 1000 \
                                  --lr 0.01 \
                                  --mix_vars 1 \       # Use nMDMA
                                  --n_mix_terms 5\     # Number of diagonals in the mixing matrix T

Density estimation with missing values

Fit UCI POWER dataset with 0.5 probability of missing values per entry using MDMA:

python3 uci_density_estimation.py --dataset gas \
                                  --m 4000 \
                                  --r 5 \
                                  --l 4 \
                                  --batch_size 500 \
                                  --n_epochs 1000 \
                                  --lr 0.01 \
                                  --missing_data_pct 0.5 # proportion of missing values

Density estimation using BNAF on the same dataset after performing MICE imputation:

python3 bnaf_density_estimation.py --dataset gas \
                                   --hidden_dim 320 \
                                   --missing_data_pct 0.5 \
                                   --missing_data_strategy mice

Requires (for imputation):

• miceforest>=2.0.4
• scikit-learn>=0.24.2

Mutual information estimation

Generate data from a multivariate Gaussian, fit the joint density using MDMA and estimate the mutual information between subsets of variables:

python3 mi_estimation.py

Causal discovery

Run the causal discovery experiment, recovering a causal graph from data by testing for conditional independence using MDMA:

python3 causal_discovery.py --dataset "sachs" \
                            --lr .1 \
                            --r 3 \
                            --l 2 \
                            --m 1000 \
                            --batch_size 500 \
                            --patience 100 \
                            --n_epochs 50 \
                            --verbose 1 \
                            --save_checkpoints 0

Requires:

• R>=4.0.5 and R packages
  • pcalg>=2.7-3 (on CRAN)
  • graph>=1.70.0 (on Bioconductor)
  • RBGL>=1.68.0 (on Bioconductor)
  • graph>=1.70.0 (on Bioconductor)

as well as the python packages

• cdt>=0.5.23
• networkx>=2.5.1
• rpy2>=3.4.4

Sampling

The time complexity of sampling from MDMA is logarithmic in the input dimension.

Generate S samples from a trained model:

samples = model.sample(S)

Citation

@misc{gilboa2021marginalizable,
      title={Marginalizable Density Models}, 
      author={Dar Gilboa and Ari Pakman and Thibault Vatter},
      year={2021},
      eprint={2106.04741},
      archivePrefix={arXiv},
      primaryClass={stat.ML}
}