PATK: Dynamic Semantic Tokenization for Time Series via Elastic Sampling on Physics-aware Perception

Project Introduction

PATK is a Transformer-based time series processing framework focused on time series classification and prediction tasks. The framework supports pre-training and fine-tuning modes, and can handle various time series datasets, including physiological signals, sensor data, and meteorological data.

Main Features

• Time Series Classification: Supports multiple classification datasets such as ECG, EMG, Epilepsy, etc.
• Time Series Prediction: Supports multiple prediction datasets such as ETT, electricity, weather, etc.
• Pre-training and Fine-tuning: Implements a complete workflow for model pre-training and downstream task fine-tuning
• Multiple Model Support: Supports various model architectures including Transformer, Mamba, CNN, RNN, etc.
• Efficient Data Processing: Includes data loading, preprocessing, and augmentation functionality

Project Structure

PATK/
├── data_load/         # Data loading and preprocessing module
├── dataset/           # Dataset directory
│   ├── Cls/           # Classification datasets
├── model/             # Model definitions
├── trainer/           # Training and testing scripts
├── main_cls.py        # Main script
├── contrastive_loss_V2.py # Contrastive loss function

Installation Instructions

Environment Requirements

• Python 3.8+
• PyTorch 1.9+
• CUDA (recommended for GPU acceleration)
• Other dependencies: numpy, einops, matplotlib

Installation Steps

1. Clone the project to your local machine

   git clone <project address>
   cd PATK

2. Install dependencies

   pip install -r requirements.txt

Usage

Classification Task

Pre-training

python main_cls.py --scene pre-train --pre_dataset_name SleepEEG --EPOCH 10 --batch_size 64

Fine-tuning

python main_cls.py --scene fine-tune --pre_dataset_name SleepEEG --tune_dataset_name ECG --EPOCH 10

Main Parameter Description

Parameter	Description	Default Value
--n_layers	Number of encoding layers	4
--n_heads	Number of multi-head attention heads	12
--d_model	Feature dimension	128
--d_ff	Feed-forward network hidden layer dimension	512
--dropout_rate	Dropout rate	0.1
--sampling_num	Sampling times	30
--sampling_length	Sampling length	40
--input_len	Input sequence length	Classification: 178, Prediction: 336
--pred_len	Prediction sequence length	96
--dataset	Task type	Cls or Pred
--pre_dataset_name	Pre-training dataset name	SleepEEG (classification) or ETTh1 (prediction)
--tune_dataset_name	Fine-tuning dataset name	ECG (classification) or ETTh1 (prediction)
--scene	Training scene	pre-train or fine-tune
--EPOCH	Training epochs	10
--batch_size	Batch size	Classification: 64, Prediction: 8
--learning_rate	Learning rate	0.008
--model	Model type	Transformer

Model Architecture

Core Components

1. RoV_time: Time domain random sampling module that generates sampling points through change rate distribution, performs slicing and length scaling on time series
2. SEI_freq: Frequency domain sampling module that generates sampling points through energy distribution, processes frequency domain signals
3. Time_transformer: Transformer-based time series encoder that融合 time domain and frequency domain features
4. Contrastive Learning: Uses contrastive loss function for pre-training to enhance model representation capability

Model Flow

1. Input time series data
2. Perform sampling in time domain and frequency domain through RoV_time and SEI_freq modules respectively
3. Embed and encode the sampling results
4. Extract features through Transformer encoder
5. Fine-tune for classification or prediction tasks

Supported Datasets

Classification Datasets

• ECG: Electrocardiogram data
• EMG: Electromyography data
• Epilepsy: Epileptic seizure data
• FD-A/FD-B: Fault diagnosis data
• Gesture: Gesture recognition data
• HAR: Human activity recognition data
• SleepEEG: Sleep electroencephalogram data

Prediction Datasets

• ETTh1/ETTh2/ETTm1/ETTm2: Electrical transformer temperature data
• electricity: Electricity consumption data
• illness: Disease incidence data
• weather: Meteorological data

Classification Task Example

# Pre-training
python main_cls.py --scene pre-train --pre_dataset_name SleepEEG --EPOCH 10

# Fine-tuning
python main_cls.py --scene fine-tune --pre_dataset_name SleepEEG --tune_dataset_name ECG --EPOCH 10

Notes

1. Ensure datasets are correctly placed in the dataset/ directory
2. Pre-trained models will be saved in the saved_models_pre_train_<dataset_name>/ directory
3. Training logs will be saved in the V2_logs_<dataset_name>/ directory
4. For large datasets, GPU training is recommended
5. You can optimize model performance by adjusting --sampling_num and --sampling_length parameters

Contribution

Contributions are welcome! Please submit Issues and Pull Requests to improve this project.

License

This project is licensed under the MIT License.

Contact

For any questions, please contact the project maintainers.