Research code for the paper "Variational Gibbs inference for statistical estimation from incomplete data".

Overview

Variational Gibbs inference (VGI)

This repository contains the research code for

Simkus, V., Rhodes, B., Gutmann, M. U., 2021. Variational Gibbs inference for statistical model estimation from incomplete data.

The code is shared for reproducibility purposes and is not intended for production use. It should also serve as a reference implementation for anyone wanting to use VGI for model estimation from incomplete data.

Abstract

Statistical models are central to machine learning with broad applicability across a range of downstream tasks. The models are typically controlled by free parameters that are estimated from data by maximum-likelihood estimation. However, when faced with real-world datasets many of the models run into a critical issue: they are formulated in terms of fully-observed data, whereas in practice the datasets are plagued with missing data. The theory of statistical model estimation from incomplete data is conceptually similar to the estimation of latent-variable models, where powerful tools such as variational inference (VI) exist. However, in contrast to standard latent-variable models, parameter estimation with incomplete data often requires estimating exponentially-many conditional distributions of the missing variables, hence making standard VI methods intractable. We address this gap by introducing variational Gibbs inference (VGI), a new general-purpose method to estimate the parameters of statistical models from incomplete data.

VGI demo

We invite the readers of the paper to also see the Jupyter notebook, where we demonstrate VGI on two statistical models and animate the learning process to help better understand the method.

Below is an animation from the notebook of a Gaussian Mixture Model fitted from incomplete data using the VGI algorithm (left), and the variational Gibbs conditional approximations (right) throughout iterations.

demo_vgi_mog_fit.mp4

Dependencies

Install python dependencies from conda and the cdi project package with

conda env create -f environment.yml
conda activate cdi
python setup.py develop

If the dependencies in environment.yml change, update dependencies with

conda env update --file environment.yml

Summary of the repository structure

Data

All data used in the paper are stored in data directory and the corresponding data loaders can be found in cdi/data directory.

Method code

The main code to the various methods used in the paper can be found in cdi/trainers directory.

  • trainer_base.py implements the main data loading and preprocessing code.
  • variational_cdi.py and cdi.py implement the key code for variational Gibbs inference (VGI).
  • mcimp.py implements the code for variational block-Gibbs inference (VBGI) used in the VAE experiments.
  • The other scripts in cdi/trainers implement the comparison methods and variational conditional pre-training.

Statistical models

The code for the statistical (factor analysis, VAEs, and flows) and the variational models are located in cdi/models.

Configuration files

The experiment_configs directory contains the configuration files for all experiments. The config files include all the hyperparameter settings necessary to reproduce our results. The config files are in a json format. They are passed to the main running script as a command-line argument and values in them can be overriden with additional command-line arguments.

Run scripts

train.py is the main code we use to run the experiments, and test.py is the main script to produce analysis results presented in the paper.

Analysis code

The Jupyter notebooks in notebooks directory contain the code which was used to analysis the method and produce figures in the paper. You should also be able to use these notebooks to find the corresponding names of the config files for the experiments in the paper.

Running the code

Before running any code you'll need to activate the cdi conda environment (and make sure you've installed the dependencies)

conda activate cdi

Model fitting

To train a model use the train.py script, for example, to fit a rational-quadratic spline flow on 50% missing MiniBooNE dataset

python train.py --config=experiment_configs/flows_uci/learning_experiments/3/rqcspline_miniboone_chrqsvar_cdi_uncondgauss.json

Any parameters set in the config file can be overriden by passing additionals command-line arguments, e.g.

python train.py --config=experiment_configs/flows_uci/learning_experiments/3/rqcspline_miniboone_chrqsvar_cdi_uncondgauss.json --data.total_miss=0.33

Optional variational model warm-up

Some VGI experiments use variational model "warm-up", which pre-trains the variational model on observed data as probabilistic regressors. The experiment configurations for these runs will have var_pretrained_model set to the name of the pre-trained model. To run the corresponding pre-training script run, e.g.

python train.py --config=experiment_configs/flows_uci/learning_experiments/3/miniboone_chrqsvar_pretraining_uncondgauss.json

Running model evaluation

For model evaluation use test.py with the corresponding test config, e.g.

python test.py --test_config=experiment_configs/flows_uci/eval_loglik/3/rqcspline_miniboone_chrqsvar_cdi_uncondgauss.json

This will store all results in a file that we then analyse in the provided notebook.

For the VAE evaluation, where variational distribution fine-tuning is required for test log-likelihood evaluation use retrain_all_ckpts_on_test_and_run_test.py.

Using this codebase on your own task

While the main purpose of this repository is reproducibility of the research paper and a demonstration of the method, you should be able to adapt the code to fit your statistical models. We would advise you to first see the Jupyter notebook demo. The notebook provides an example of how to implement the target statistical model as well as the variational model of the conditionals, you can find further examples in cdi/models directory. If you intend to use a variational family that is different to ours you will also need to implement the corresponding sampling functions here.

Owner
Vaidotas Šimkus
PhD candidate in Data Science at the University of Edinburgh. Interested in deep generative models, variational inference, and the Bayesian principle.
Vaidotas Šimkus
Docker containers of baseline agents for the Crafter environment

Crafter Baselines This repository contains Docker containers for running various baselines on the Crafter environment. Reward Agents DreamerV2 based o

Danijar Hafner 17 Sep 25, 2022
Reading Group @mila-iqia on Computational Optimal Transport for Machine Learning Applications

Computational Optimal Transport for Machine Learning Reading Group Over the last few years, optimal transport (OT) has quickly become a central topic

Ali Harakeh 11 Aug 26, 2022
Code for the Lovász-Softmax loss (CVPR 2018)

The Lovász-Softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks Maxim Berman, Amal Ranne

Maxim Berman 1.3k Jan 04, 2023
In Search of Probeable Generalization Measures

In Search of Probeable Generalization Measures Exciting News! In Search of Probeable Generalization Measures has been accepted to the International Co

Mahdi S. Hosseini 6 Sep 11, 2022
codes for "Scheduled Sampling Based on Decoding Steps for Neural Machine Translation" (long paper of EMNLP-2022)

Scheduled Sampling Based on Decoding Steps for Neural Machine Translation (EMNLP-2021 main conference) Contents Overview Background Quick to Use Furth

Adaxry 13 Jul 25, 2022
Official implementation of Influence-balanced Loss for Imbalanced Visual Classification in PyTorch.

Official implementation of Influence-balanced Loss for Imbalanced Visual Classification in PyTorch.

Seulki Park 70 Jan 03, 2023
Efficient Sharpness-aware Minimization for Improved Training of Neural Networks

Efficient Sharpness-aware Minimization for Improved Training of Neural Networks Code for “Efficient Sharpness-aware Minimization for Improved Training

Angusdu 32 Oct 18, 2022
Code for "Contextual Non-Local Alignment over Full-Scale Representation for Text-Based Person Search"

Contextual Non-Local Alignment over Full-Scale Representation for Text-Based Person Search This is an implementation for our paper Contextual Non-Loca

Tencent YouTu Research 50 Dec 03, 2022
Human-Pose-and-Motion History

Human Pose and Motion Scientist Approach Eadweard Muybridge, The Galloping Horse Portfolio, 1887 Etienne-Jules Marey, Descent of Inclined Plane, Chron

Daito Manabe 47 Dec 16, 2022
[EMNLP 2020] Keep CALM and Explore: Language Models for Action Generation in Text-based Games

Contextual Action Language Model (CALM) and the ClubFloyd Dataset Code and data for paper Keep CALM and Explore: Language Models for Action Generation

Princeton Natural Language Processing 43 Dec 16, 2022
Instantaneous Motion Generation for Robots and Machines.

Ruckig Instantaneous Motion Generation for Robots and Machines. Ruckig generates trajectories on-the-fly, allowing robots and machines to react instan

Berscheid 374 Dec 23, 2022
Parasite: a tool allowing you to compress and decompress files, to reduce their size

🦠 Parasite 🦠 Parasite is a tool written in Python3 allowing you to "compress" any file, reducing its size. ⭐ Features ⭐ + Fast + Good optimization,

Billy 30 Nov 25, 2022
TransFGU: A Top-down Approach to Fine-Grained Unsupervised Semantic Segmentation

TransFGU: A Top-down Approach to Fine-Grained Unsupervised Semantic Segmentation Zhaoyun Yin, Pichao Wang, Fan Wang, Xianzhe Xu, Hanling Zhang, Hao Li

DamoCV 25 Dec 16, 2022
A new data augmentation method for extreme lighting conditions.

Random Shadows and Highlights This repo has the source code for the paper: Random Shadows and Highlights: A new data augmentation method for extreme l

Osama Mazhar 35 Nov 26, 2022
Code for "Learning Canonical Representations for Scene Graph to Image Generation", Herzig & Bar et al., ECCV2020

Learning Canonical Representations for Scene Graph to Image Generation (ECCV 2020) Roei Herzig*, Amir Bar*, Huijuan Xu, Gal Chechik, Trevor Darrell, A

roei_herzig 24 Jul 07, 2022
Speckle-free Holography with Partially Coherent Light Sources and Camera-in-the-loop Calibration

Speckle-free Holography with Partially Coherent Light Sources and Camera-in-the-loop Calibration Project Page | Paper Yifan Peng*, Suyeon Choi*, Jongh

Stanford Computational Imaging Lab 19 Dec 11, 2022
Code of our paper "Contrastive Object-level Pre-training with Spatial Noise Curriculum Learning"

CCOP Code of our paper Contrastive Object-level Pre-training with Spatial Noise Curriculum Learning Requirement Install OpenSelfSup Install Detectron2

Chenhongyi Yang 21 Dec 13, 2022
Can we visualize a large scientific data set with a surrogate model? We're building a GAN for the Earth's Mantle Convection data set to see if we can!

EarthGAN - Earth Mantle Surrogate Modeling Can a surrogate model of the Earth’s Mantle Convection data set be built such that it can be readily run in

Tim 0 Dec 09, 2021
PyTorch-Geometric Implementation of MarkovGNN: Graph Neural Networks on Markov Diffusion

MarkovGNN This is the official PyTorch-Geometric implementation of MarkovGNN paper under the title "MarkovGNN: Graph Neural Networks on Markov Diffusi

HipGraph: High-Performance Graph Analytics and Learning 6 Sep 23, 2022