PrimeNet : Pre-Training for Irregular Multivariate Time Series

This is the official PyTorch implementation of the AAAI 2023 paper PrimeNet: Pre-Training for Irregular Multivariate Time Series.

Quick Start

git clone https://github.com/ranakroychowdhury/PrimeNet.git

Datasets

Activity

Run preprocess/download_activity.sh to download and place the data under the proper directory.

Run preprocess/preprocess_activity.py to preprocess the data.

MIMIC-III

Follow the instruction from interp-net to download and preprocess the data.

Appliances Energy

Download the dataset.

Run preprocess/preprocess_ae.py to preprocess the data.

PhysioNet

Run preprocess/preprocess_physionet.py to download and preprocess the data.

The data directory structure should be the same as that in data.zip. Extract data.zip to run experiments with a sample toy dataset.

data/

pretrain/

X_train.pt
X_val.pt

finetune/

X_train.pt
y_train.pt
X_val.pt
y_val.pt
X_test.pt
y_test.pt

Pre-training

Run pretrain.sh to run pretraining experiments on a dataset. The pretrained model is saved under ./models/ and the pretraining results are stored under ./results/. The arguments for pretraining are explained in pretrain.py.

sh pretrain.sh

Fine-tuning and Evaluation

Run finetune.sh to run finetuning experiments on a dataset. The pretrained model saved during the pretraining experiment under ./models/ is used for finetuning. The name of the pretrained model to use is added as an argument in the finetuning command. The finetuning results are stored under ./results/. The arguments for finetuning are explained in finetune.py.

sh finetune.sh

Reference