GNeuralFlow

This repository includes the supporting code for

Giangiacomo Mercatali, Andre Freitas, Jie Chen. Graph Neural Flows for Unveiling Systemic Interactions Among Irregularly Sampled Time Series. In Advances in Neural Information Processing Systems 37, 2024.

Installation

Install required packages from env.yml. The code was tested with python 3.10 and pytorch 2.1.1.

Training models

The argument --training_scheme indicates what model to train. Choices are:

• lgnf: GNeuralFlow together with DAG learning.
• tgnf: GNeuralFlow with a given DAG (only for synthetic data).
• nfe: Neural flow from Bilos et al. 2021 (serving as a non-graph baseline).

Experiments of synthetic datasets

Use the following commands to train GNeuralFlow models (--training_scheme=lgnf). --data is one of sink, triangle, sawtooth, square. The following example is for sink.

sink=(--experiment=synthetic --data=sink --training_scheme=lgnf --n_ts=20 --epochs=1000 --batch-size=50 --weight-decay=1e-05 --model=flow --flow-layers=4 --time-hidden-dim=8 --hidden-layers=2 --hidden-dim=64 --activation=ReLU --final-activation=Identity)
python -m nfe.train "${sink[@]}" --flow-model=resnet --time-net=TimeFourierBounded --rho=7 --h_tol=1e-10 --h_par=0.19 --rho_max=1e24 --dag_epochs=2
python -m nfe.train "${sink[@]}" --flow-model=coupling --time-net=TimeFourier --rho=8 --h_tol=1e-11 --h_par=0.22 --rho_max=1e27 --dag_epochs=2
python -m nfe.train "${sink[@]}" --flow-model=gru --time-net=TimeFourier --rho=8 --h_tol=1e-11 --h_par=0.22 --rho_max=1e27 --dag_epochs=2

Experiments of latent variable modeling: Smoothing approach

Activity

activity=(--experiment=latent_ode --data activity --training_scheme=lgnf --batch-size 100 --epochs 1000 --lr-decay 0.5 --lr-scheduler-step 20 --weight-decay 0.0001 --extrap 0 --classify 1 --hidden-layers 3 --hidden-dim 100 --rec-dims 30 --latents 20 --gru-units 100 --model flow --flow-layers 2 --time-hidden-dim 8)
python -m nfe.train "${activity[@]}" --flow-model gru --time-net TimeLinear --rho_par=15  --h_par=0.21
python -m nfe.train "${activity[@]}" --flow-model coupling --time-net TimeLinear --rho_par=7  --h_par=0.21
python -m nfe.train "${activity[@]}" --flow-model resnet --time-net TimeTanh --rho_par=7 --h_tol=1e-9  --h_par=0.5

MuJoCo

mujoco=(--experiment=latent_ode --data hopper --training_scheme=lgnf --batch-size 100 --epochs 1000 --lr-decay 0.5 --lr-scheduler-step 20 --weight-decay 0.0001 --extrap 0 --classify 0 --hidden-layers 2 --hidden-dim 100 --rec-dims 100 --gru-units 50 --model flow --flow-layers 2 --time-hidden-dim 8)
python -m nfe.train "${mujoco[@]}" --flow-model gru --time-net TimeLinear --rho_par=10  --latents 20
python -m nfe.train "${mujoco[@]}" --flow-model coupling --time-net TimeLinear --rho_par=7  --latents 20
python -m nfe.train "${mujoco[@]}" --flow-model resnet --time-net TimeTanh --h_par=0.5 --rho_par=15 --latents 15

Physionet

physionet=(--experiment=latent_ode --data physionet --training_scheme=lgnf --dag_epochs=100 --batch-size 100 --epochs 1000 --lr-decay 0.5 --lr-scheduler-step 20 --weight-decay 0.0001 --extrap 0 --classify 1 --hidden-layers 3 --hidden-dim 50 --rec-dims 40 --latents 20 --gru-units 50 --model flow --flow-layers 2 --time-hidden-dim 8)
python -m nfe.train "${physionet[@]}" --flow-model gru --time-net TimeLinear --h_par=0.5 --rho_par=15
python -m nfe.train "${physionet[@]}" --flow-model coupling --time-net TimeLinear --h_par=0.5 --rho_par=10
python -m nfe.train "${physionet[@]}" --flow-model resnet --time-net TimeTanh --rho_par=15

Experiments of latent variable modeling: Filtering approach

Prepare data (MIMIC IV)

1. Download csv datasets. Need to register at physionet.org. Place the data in the folder nfe/experiments/data/physionet.org/files/mimiciv/2.2.
2. Run the scripts in the prep_data folder, in this order: 1. admission, 2. prescription, 3. labevents, 4. inputsevents, 5. outputs.

Train GNeuralFlow

mimic4=(--experiment=gru_ode_bayes --data=mimic4 --training_scheme=lgnf --patience=5 --dag_epochs=1000 --epochs=1000 --batch-size=100 --weight-decay=0.0001 --lr-decay=0.33 --lr-scheduler-step=20 --model=flow --flow-layers=4 --time-net=TimeTanh --time-hidden-dim=8 --hidden-layers=2 --hidden-dim=64 --activation=ReLU --final-activation=Identity)
python -m nfe.train "${mimic4[@]}" --flow-model=gru --rho_par=10 --h_tol=1e-10 --h_par=0.15 --rho_max=1e24 --dag_epochs=1
python -m nfe.train "${mimic4[@]}" --flow-model=coupling --rho_par=10 --h_tol=1e-10 --h_par=0.15 --rho_max=1e24 --dag_epochs=1
python -m nfe.train "${mimic4[@]}" --flow-model=resnet --rho_par=10 --h_tol=1e-10 --h_par=0.15 --rho_max=1e24 --dag_epochs=1

Acknowledgements

Our code was developed based on the following repositories.

• https://github.com/mbilos/neural-flows-experiments
• https://github.com/EnyanDai/GANF