So Far Yet So Near: Time series data augmentation with exploring non-semantic boundaries based on reinforcement learning (ICASSP2025)

This is the Official PyTorch implementation of our ICASSP2025 paper "So Far Yet So Near: Time series data augmentation with exploring non-semantic boundaries based on reinforcement learning".

Quick Start

Install

# build with python3.10
conda create --name FreqSYNenv python=3.10e
conda activate FreqSYNenv 
pip install -r requirments.txt

Usage

Before usage

1. We provide a data loading API in our code, which can be customized to suit the characteristics of specific datasets.
2. The detailed loading functions are available in Code/Utils/Func.py.

Classifier Training

1. The downstream task classifier should be integrated into our reinforcement learning framework, and must be provided prior to the reinforcement learning training process.
2. We provide a basic training script for MMCNN, a well-known classifier in the EEG signal domain, which can be used by python Code/Train/Classifier_train.py --dataset your_dataset
3. Tips: the path of your dataset should be set in the script before training

Augmentation Agent Training

1. Before training the augmentation agent within the RL framework, the RL training configuration must be set. We provide an example in Config/train_config_2a.json, which should be adjusted to fit specific datasets.

2. For training the RL Agent, you can use the following command:

   python Code/train.py --train_name=your_train_name --train_cfg=/path/to/your/train/config.json
   # for more arguments setting, you can use:
   python Code/train.py --help
   

Data Augment

1. For getting the Augmentated data which generated by the agent, the following command should be used:

   python3 Code/Test/Get_Aug_data.py --train_cfg /same/as/the/config/in/train --agent_path /the/best/agent/saved/in/train
   

2. After getting the augmented data, the classifier should be trained again with augmented data by using:

   # Tips: the augmented data path should be setting before
   python Code/Train/Classifier_train.py --dataset your_dataset --train_mode aug
   

Reference

1. RL-PPO