VACA: Designing Variational Graph Autoencoders for Interventional and Counterfactual Queries

Last update: Oct 10, 2022

Related tags

Overview

VACA

Code repository for the paper "VACA: Designing Variational Graph Autoencoders for Interventional and Counterfactual Queries (arXiv)". The implementation is based on Pytorch, Pytorch Geometric and Pytorch Lightning. The repository contains the necessary resources to run the experiments of the paper. Follow the instructions below to download the German dataset.

Installation

Create conda environment and activate it:

conda create --name vaca python=3.9 --no-default-packages
conda activate vaca

Option 1: Import the conda environment

conda env create -f environment.yml

Option 2: Commands

conda install pip
pip install torch torchvision torchaudio
pip install pytorch-lightning
pip install -U scikit-learn
pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-1.9.0+cpu.html
pip install matplotlib
pip install seaborn

Note: The German dataset is not contained in this repository. The first time you try to train on the German dataset, you will get an error with instructions on how to download and store it. Please follow the instructions, such that the code runs smoothly.

Datasets

This repository contains 7 different SCMs: - ColliderSCM - MGraphSCM - ChainSCM - TriangleSCM - LoanSCM - AdultSCM - GermanSCM

Additionally, we provide the implementation of the first five SCMs with three different types of structural equations: linear (LIN), non-linear (NLIN) and non-additive (NADD). You can find the implementation of all the datasets inside the folder datasets. To create all datasets at once run python _create_data_toy.py (this is optional since the datasets will be created as needed on the fly).

How to create your custom Toy Datasets

We also provide a function to create custom ToySCM datasets. Here is an example of an SCM with 2 nodes

from datasets.toy import create_toy_dataset
from utils.distributions import *
dataset = create_toy_dataset(root_dir='./my_custom_datasets',
                             name='2graph',
                             eq_type='linear',
                             nodes_to_intervene=['x1'],
                             structural_eq={'x1': lambda u1: u1,
                                            'x2': lambda u2, x1: u2 + x1},
                             noises_distr={'x1': Normal(0,1),
                                           'x2': Normal(0,1)},
                             adj_edges={'x1': ['x2'],
                                        'x2': []},
                             split='train',
                             num_samples=5000,
                             likelihood_names='d_d',
                             lambda_=0.05)

Training

To train a model you need to execute the script main.py. For that, you need to specify three configuration files: - dataset_file: Specifies the dataset and the parameters of the dataset. You can overwrite the dataset parameters -d. - model_file: Specifies the model and the parameters of the model as well as the optimizer. You can overwrite the model parameters with -m and the optimizer parameters with -o. - trainer_file: Specifies the training parameters of the Trainer object from PyTorch Lightning.

For plotting results use --plots 1. For more information, run python main.py --help.

Examples

To train our VACA algorithm on each of the synthetic graphs with linear structural equations (default value in dataset_):

python main.py --dataset_file _params/dataset_adult.yaml --model_file _params/model_vaca.yaml
python main.py --dataset_file _params/dataset_loan.yaml --model_file _params/model_vaca.yaml
python main.py --dataset_file _params/dataset_chain.yaml --model_file _params/model_vaca.yaml
python main.py --dataset_file _params/dataset_collider.yaml --model_file _params/model_vaca.yaml
python main.py --dataset_file _params/dataset_mgraph.yaml --model_file _params/model_vaca.yaml
python main.py --dataset_file _params/dataset_triangle.yaml --model_file _params/model_vaca.yaml

You can also select a different SEM with the -d option and

for linear (LIN) equations -d equations_type=linear,
for non-linear (NLIN) equations -d equations_type=non-linear,
for non-additive (NADD) equation -d equations_type=non-additive.

For example, to train the triangle graph with non linear SEM:

python main.py --dataset_file _params/dataset_triangle.yaml --model_file _params/model_vaca.yaml -d equations_type=non-linear

We can train our VACA algorithm on the German dataset:

python main.py --dataset_file _params/dataset_german.yaml --model_file _params/model_vaca.yaml

To run the CAREFL model:

python main.py --dataset_file _params/dataset_adult.yaml --model_file _params/model_carefl.yaml
python main.py --dataset_file _params/dataset_loan.yaml --model_file _params/model_carefl.yaml
python main.py --dataset_file _params/dataset_chain.yaml --model_file _params/model_carefl.yaml
python main.py --dataset_file _params/dataset_collider.yaml --model_file _params/model_carefl.yaml
python main.py --dataset_file _params/dataset_mgraph.yaml --model_file _params/model_carefl.yaml
python main.py --dataset_file _params/dataset_triangle.yaml --model_file _params/model_carefl.yaml

To run the MultiCVAE model:

python main.py --dataset_file _params/dataset_adult.yaml --model_file _params/model_mcvae.yaml
python main.py --dataset_file _params/dataset_loan.yaml --model_file _params/model_mcvae.yaml
python main.py --dataset_file _params/dataset_chain.yaml --model_file _params/model_mcvae.yaml
python main.py --dataset_file _params/dataset_collider.yaml --model_file _params/model_mcvae.yaml
python main.py --dataset_file _params/dataset_mgraph.yaml --model_file _params/model_mcvae.yaml
python main.py --dataset_file _params/dataset_triangle.yaml --model_file _params/model_mcvae.yaml

How to load a trained model?

To load a trained model:

set the training flag to -i 0.
select configuration file of our training model, i.e. hparams_full.yaml

python main.py --yaml_file=PATH/hparams_full.yaml -i 0

Load a model and train/evaluate counterfactual fairness

Load your model and add the flag --eval_fair. For example:

python main.py --yaml_file=PATH/hparams_full.yaml -i 0 --eval_fair --show_results

TensorBoard visualization

You can track different metrics during (and after) training using TensorBoard. For example, if the root folder of the experiments is exper_test, we can run the following command in a terminal

tensorboard --logdir exper_test/

to display the logs of all experiments contained in such folder. Then, we go to our favourite browser and go to http://localhost:6006/ to visualize all the results.

VACA: Designing Variational Graph Autoencoders for Interventional and Counterfactual Queries

Related tags

Overview

VACA

Installation

Option 1: Import the conda environment

Option 2: Commands

Datasets

How to create your custom Toy Datasets

Training

Examples

How to load a trained model?

Load a model and train/evaluate counterfactual fairness

TensorBoard visualization

Owner

Pablo Sánchez-Martín

Teaching end to end workflow of deep learning

Here is the diagnostic tool for BMVC 2021 paper Diagnosing Errors in Video Relation Detectors.

Official Code for "Non-deep Networks"

Official PyTorch code for the paper: "Point-Based Modeling of Human Clothing" (ICCV 2021)

GarmentNets: Category-Level Pose Estimation for Garments via Canonical Space Shape Completion

Abstractive opinion summarization system (SelSum) and the largest dataset of Amazon product summaries (AmaSum). EMNLP 2021 conference paper.

Control-Robot-Arm-using-PS4-Controller - A Robotic Arm based on Raspberry Pi and Arduino that controlled by PS4 Controller

Implementation of the Remixer Block from the Remixer paper, in Pytorch

Easy to use Python camera interface for NVIDIA Jetson

Prototypical python implementation of the trust-region algorithm presented in Sequential Linearization Method for Bound-Constrained Mathematical Programs with Complementarity Constraints by Larson, Leyffer, Kirches, and Manns.

UMPNet: Universal Manipulation Policy Network for Articulated Objects

Learning Saliency Propagation for Semi-supervised Instance Segmentation

NasirKhusraw - The TSP solved using genetic algorithm and show TSP path overlaid on a map of the Iran provinces & their capitals.

Official PyTorch Implementation of paper EAN: Event Adaptive Network for Efficient Action Recognition

We evaluate our method on different datasets (including ShapeNet, CUB-200-2011, and Pascal3D+) and achieve state-of-the-art results, outperforming all the other supervised and unsupervised methods and 3D representations, all in terms of performance, accuracy, and training time.

The official pytorch implemention of the CVPR paper "Temporal Modulation Network for Controllable Space-Time Video Super-Resolution".

Meta Self-learning for Multi-Source Domain Adaptation： A Benchmark

Official Implementation of "Third Time's the Charm? Image and Video Editing with StyleGAN3" https://arxiv.org/abs/2201.13433

A TensorFlow implementation of DeepMind's WaveNet paper

Causal Imitative Model for Autonomous Driving