FlowScans

A likelihood estimator for exchangeable data

This is the official repository for the 2020 AAAI article.

General Idea

We utilize (learned) permutaton equivariant flows to transform the data to a space that is better suited to scanning. We then scan over each the first dimension of each point to convert the set into a sequence. Finally, we apply conventional likelihood estimators (such as autoregressive flows).

Results

We compare to several other general, exchangeable likelihood estimators, BRUNO and The Neural Statistician.

DataSet	BRUNO	NS	FlowScan
Synthetic	-2.28	-1.07	0.14
Airplanes	2.71	4.09	4.81
Chairs	0.75	2.02	2.58
ModelNet10	0.49	2.12	3.01
ModelNet10a	1.20	2.82	3.58
Caudate	1.29	4.49	4.87
Thalamus	0.82	2.69	3.12
SpatialMNIST	-5.68	-5.37	-5.26

Use

To use this version of the code you can build the provided Dockerfile or can install the necessary packages. pip install dill tqdm requests matplotlib scipy tensorflow-gpu==1.12.0 should suffice.

To train a model, the data must be provided in a pickle file. On load, the data must be returned as a dictionary of arrays with keys train, valid, and test. Each of these (3D) arrays must be organized as sets, points, and dimensions such that indexing into the first dimension (data['train'][n])) will return a complete set.

The default model is trained by calling

import flowscan.demos.flowscan as fdemo
dataset = 'plane'               # plane sets from ModelNet10
datadir = '/home/me/data/sets'  # the directory in which the data is stored
results = fdemo.main(
    dataset=dataset, datadir=datadir, dims=3, subsample=512, train_iters=40000)

Citation

@inproceedings{bender2020flowscan,
  author    = {Christopher M.~Bender and
               Kevin O'Connor and
               Yang Li and
               Juan Jose Garcia and
               Manzil Zaheer and
               Junier B.~Oliva},
  title     = {Exchangeable Generative Models with Flow Scans},
  publisher = {{AAAI} Press},
  url       = {http://arxiv.org/abs/1902.01967},
  year      = {2020},
}