Description

This repository contains the code for TSA on AutoPilot: Self-tuning Self-supervised Time Series Anomaly Detection.

Data

PhysioNet A-G and Mocap A-B are available here

Training and Testing TSAP

To train TSAP, specify the configuration file in configs/tsap and adapt shell_scripts/train_tsap.sh accordingly. To test TSAP, specify the configuration file in configs/tsap and adapt shell_scripts/test_tsap.sh accordingly.

# Train TSAP on PhysioNet C
sh shell_scripts/train_tsap.sh

Three key parameters in the configuration files are aug_params, a_init, and anom_data_path. aug_params specifies which hyperparameters to learn and which ones to randomize (lvl, loc, len). a_init specifies the initial values of the augmentation hyperparameter $\mathbb{a}$. anom_data_path specifies the path to the anomalous data (i.e. PhysioNet A-G or Mocap A-B).

File Structure

The project is organized as follows:

• /checkpoints: This folder contains saved model checkpoints for the pretrained $f_\mathrm{aug}$. During training of TSAP, checkpoints for $f_\mathrm{det}$ and $\mathbb{a}$ are saved here as well.
• /configs: Configuration files for training TSAP.
• /shell_scripts: Scripts to train and/or test TSAP.
• /src: The source code for the core functionalities of TSAP.
• main.py: The main Python script for running TSAP. --