TS-Rep: Self-supervised time series representation learning from robot sensor data

This is the code for the paper TS-Rep: Self-supervised time series representation learning from robot sensor data presented at the Self-Supervised Learning - Theory and Practice workshop at NeurIPS 2022.

Acknowledgement: This code is developed on top of the T-Loss code github (paper). We thank the authors of T-Loss for making their code available.

Requirements

Install dependencies using the Conda environment:

conda env create -f environment.yml
conda activate ts-rep

Datasets

All pre-processed datasets in .npz can be downloaded with a single command:

bash datasets/download_datasets.sh

Note: Pre-processing scripts for the raw datasets will be uploaded soon.

Usage

Scripts are provided in scripts/ to run the experiments.

For example, to train TS-Rep on the Manipulation dataset run,

bash scripts/manipulation.sh

Dataset type

TS-Rep supports both fixed-length (i.e., padded) and varying length time series. In each script, the fixed-length dataset is the default. To run on a varying-length/ NaN padded dataset, we have to uncomment a few lines. Again, taking the Manipulation dataset as an example, uncomment the following lines in scripts/manipulation.sh:

# dataset_name="manipulation_std_nan_padded"
# dataset_dir="datasets/manipulation/varying"

Tasks

Arguments to run these tasks are provided in the individual scripts; for example in manipulation.sh, we have clustering and anomaly detection tasks.

• Clustering
• Anomaly Detection
• Terrain Classification

Output

Output will be stored in output/%m_%d_%H_%M_%S/ directories, and based on the task, CSV and PNG/ EPS files will be generated.