Temporal Dynamic Embedding for Irregularly Sampled Time Series

This repository is implementation for the paper Temporal Dynamic Embedding for Irregularly Sampled Time Series.

This paper is now under review by International Conference on Machine Learning 2022 (ICML 2022)

Install environments

The source code requires Python 3.7. In detail, all required Python package list is written in requirements.txt

$ pip install -r requirements.txt

Source code

├── main.py                         # implementation of PhysioNet 2012, MIMIC-3
├── main_online.py                  # implementation of PhysioNet 2019
├── data
│   ├── physionet-challenge-2012    # dataset of PhysioNet 2012
│   ├── physionet-challenge-2019    # dataset of PhysioNet 2019
│   └── physionet-mimic-3           # dataset of MIMIC-3
├── logs
│   └── DATA_NAME                   # dataset name
│       ├── checkpoint              # log files for early stop; '.best.h5' is saved model weights for results in paper
│       └── results_learning        # results of learning curve
├── results                         # results of classification performance
│   └── DATA_NAME                   # dataset name
├── data_pipeline.py                # load and preprocess dataset
├── model.py                        # TDE-mean, TDE-attn
└── utils.py                        # several necessary function for implementation

Usage

1. The PhysioNet Challenge 2012

(1) TDE-mean

$ python main.py --dataset physionet2012 --agg_type mean --dropout_rate 0.2 --dim_general 4 --dim_ts 16 --dim_time 32 --dim_instance 16 --dim_hidden 128

(2) TDE-attn

$ python main.py --dataset physionet2012 --agg_type attn --dropout_rate 0.5 --dim_general 2 --dim_ts 16 --dim_attn 8 --num_heads 1 --dim_instance 8 --dim_hidden 256

2. MIMIC-3

(1) TDE-mean

$ python main.py --dataset mimic3 --agg_type mean --dropout_rate 0.2 --dim_general 2 --dim_ts 64 --dim_time 8 --dim_instance 16 --dim_hidden 16

(2) TDE-attn

$ python main.py --dataset mimic3 --agg_type attn --dropout_rate 0.2 --dim_general 2 --dim_ts 128 --dim_attn 16 --num_heads 1 --dim_instance 16 --dim_hidden 64

3. The PhysioNet Challenge 2019

(1) TDE-mean

$ python main_online.py --agg_type mean --dropout_rate 0.2 --dim_general 2 --dim_ts 32 --dim_time 8 --dim_instance 64 --dim_hidden 64

(2) TDE-attn

$ python main_online.py --agg_type attn --dropout_rate 0.2 --dim_general 2 --dim_ts 16 --dim_attn 4 --num_heads 8 --dim_instance 32 --dim_hidden 256

License

MIT License