DoubleAdapt

A refined implementation of the double adaptation framework for time series prediction, based on the paper "Double Adaptation for Time Series Forecasting" (ICML 2023). This repository is a restructured version of the original implementation with improved code organization and usability.

Overview

DoubleAdapt addresses the challenge of distribution shifts in time series forecasting through a novel double adaptation framework that combines:

1. Offline Meta-Learning: Learns generalizable patterns across historical tasks
2. Online Adaptation: Rapidly adapts to new data distributions using meta-learned initialization

Key advantages:

• Handles both gradual and abrupt distribution shifts
• Requires minimal adaptation data
• Maintains model stability while enabling quick adaptation
• Achieves superior performance compared to traditional online learning methods

Features

• Enhanced Implementation:

  • Clean, modular code structure
  • Type hints and comprehensive documentation
  • Improved data preprocessing with robust Z-score normalization
  • Flexible task sampling for time series data

• Core Components:

  • Meta-learning based offline training
  • Efficient online adaptation mechanism
  • Support/query set based task creation
  • Customizable model architectures

Installation

Using UV (Recommended)

uv venv
uv pip install .

Using pip

pip install .

Project Structure

doubleadpt/
├── src/
│   ├── dataset.py      # Rolling task sampler for meta-learning
│   ├── handler.py      # Data preprocessing and normalization
│   ├── model.py        # Neural network architectures
│   └── trainner.py     # Double adaptation implementation
├── main.py             # Training script with configuration
├── pyproject.toml      # Project dependencies
└── README.md

Quick Start

from doubleadpt.src.handler import RobustZScoreNorm
from doubleadpt.src.trainner import DoubleAdaptFramework
from doubleadpt.src.dataset import RollingTaskSampler

# Initialize data preprocessing
normalizer = RobustZScoreNorm()
X_norm = normalizer.fit_transform(X)

# Create task sampler
sampler = RollingTaskSampler(
    X_norm, 
    support_size=60,  # Historical context window
    query_size=20     # Prediction window
)

# Initialize and train framework
framework = DoubleAdaptFramework(
    model=your_model,
    criterion=your_loss_function,
    x_dim=feature_dim,
    device=device
)

# Train offline (meta-learning phase)
framework.offline_training(train_tasks, valid_tasks, max_epochs=100)

# Adapt online (deployment phase)
metrics = framework.online_training(valid_tasks, test_tasks)

Command Line Usage

python main.py --data_path /path/to/your/data.csv --output_dir outputs

Key Arguments

• --data_path: Path to input CSV data file (required)
• --output_dir: Directory for saving outputs (default: 'outputs')
• --support_size: Historical context window size (default: 60)
• --query_size: Prediction window size (default: 20)
• --max_epochs: Maximum training epochs (default: 100)

Input Data Format

Required CSV structure:

• datetime: Temporal index column
• label: Target values column
• Feature columns used for prediction

Citation

If you use this implementation in your research, please cite:

@inproceedings{zhang2023double,
  title={Double Adaptation for Time Series Forecasting},
  author={Zhang, Xiyuan and Wen, Rui and Wang, Zhe and Li, Bo},
  booktitle={International Conference on Machine Learning},
  year={2023}
}

License

MIT License