MMAML-TSR

Multi-modal meta-learning for time series regression

This repo contains the implementation of the paper (*) which adapts MAML (Finn et al., 2017) and MMAML (Vuorio et al., 2019) to Time Series Regression.

The multimodal-meta-learning is based on this official implementation.

Dependencies

• Python 3.7.0
• Pytorch 1.4.0
• Learn2learn 0.1.1

Data

The code can be used on two open datasets that need to be pre-processed before running MAML or MMAML. The data is available on:

• Air Pollution Dataset: PM2.5 Data of Five Chinese Cities Data Set.

• Heart Rate Dataset: PPG-DaLiA Data Set.

Usage

1. Download data and create the following folder structure:

MMAML-TSR/
├── logs
	├── MAML_output/ 
	├── MMAML_output/
	...
├── data
├── code
	├── tools
	├── pre_processing
	├── models
	...

1. Preprocess and generate .pickle file

Change the paths to the raw data in the file pre_processing/ts_dataset.py accordingly. Then run pre_processing/dataset_creation.ipynb to pickle the object with the transformed data. For a new dataset, a loading functionality should be created by taking our datasets as reference. Optionally, you can download the preprocessed data HERE.

2. Run MAML

Assuming that the pickled files are in data/. Training with the default parameters on the Air Pollution Dataset works as:

cd code/
python run_MAML.py

To train on Heart-rate data:

cd code/
python run_MAML.py --dataset HR

3. Run MMAML

Assuming that the pickled files are in data/. Training with the default parameters on the Air Pollution Dataset works as:

cd code
python run_MMAML.py

To train on Heart-rate data:

cd code
python run_MMAML.py --dataset HR

Contact

To ask questions or report issues, please open an issue on the issues tracker.