Skip to content

Samay: Time-series Foundational Models Library

A Unified Interface for Multiple Time-Series Foundational Models

Welcome to Samay

Samay is a comprehensive Python library that provides a unified, easy-to-use interface for training and evaluating state-of-the-art time-series foundational models. Whether you're working on forecasting, classification, anomaly detection, or imputation tasks, Samay simplifies the process of leveraging powerful pre-trained models.

✨ Key Features

  • Unified Interface: Work with multiple foundational models through a consistent API
  • Pre-trained Models: Access state-of-the-art pre-trained models ready for zero-shot forecasting
  • Fine-tuning Support: Easily fine-tune models on your custom datasets
  • Multiple Tasks: Support for forecasting, classification, anomaly detection, and imputation
  • Flexible Data Handling: Built-in dataset classes for common time-series formats
  • Easy Integration: Simple pip installation and minimal code to get started

🚀 Supported Models

Samay currently supports the following foundational models:

Model Paper Strengths
LPTM Large Pre-trained Time Series Models General-purpose forecasting with segmentation
MOMENT MOMENT: A Family of Open Time-series Foundation Models Multi-task learning (forecasting, classification, anomaly detection)
TimesFM A decoder-only foundation model for time-series forecasting Decoder-only architecture by Google Research
Chronos Chronos: Learning the Language of Time Series Language model-based approach
MOIRAI Unified Training of Universal Time Series Forecasting Transformers Universal transformer by Salesforce
TinyTimeMixer TinyTimeMixer: Fast Pre-trained Models for Time Series Lightweight and efficient

📦 Quick Installation

Install Samay directly from GitHub:

pip install git+https://github.com/AdityaLab/Samay.git

🔥 Quick Start

Here's a simple example using LPTM for time-series forecasting:

from samay.model import LPTMModel
from samay.dataset import LPTMDataset

# Configure the model
config = {
    "task_name": "forecasting",
    "forecast_horizon": 192,
    "freeze_encoder": True,
    "freeze_embedder": True,
    "freeze_head": False,
}

# Load the pre-trained model
model = LPTMModel(config)

# Load your dataset
train_dataset = LPTMDataset(
    name="ett",
    datetime_col="date",
    path="./data/ETTh1.csv",
    mode="train",
    horizon=192,
)

# Fine-tune the model
finetuned_model = model.finetune(train_dataset)

# Evaluate on test data
test_dataset = LPTMDataset(
    name="ett",
    datetime_col="date",
    path="./data/ETTh1.csv",
    mode="test",
    horizon=192,
)

avg_loss, trues, preds, histories = model.evaluate(test_dataset)
print(f"Average Loss: {avg_loss}")

📚 What's Next?

🎯 Use Cases

Samay is perfect for:

  • Time-Series Forecasting: Predict future values from historical data
  • Anomaly Detection: Identify unusual patterns in time-series data
  • Classification: Classify time-series sequences into categories
  • Data Imputation: Fill missing values in time-series data
  • Transfer Learning: Leverage pre-trained models for your domain-specific tasks

💡 Why Samay?

Traditional time-series modeling requires extensive expertise and computational resources. Samay democratizes access to state-of-the-art foundational models by:

  1. Providing a unified interface across multiple model architectures
  2. Offering pre-trained models that work out-of-the-box
  3. Enabling easy fine-tuning with minimal code
  4. Supporting multiple tasks with the same infrastructure

🤝 Community and Support

📝 Citation

If you use Samay in your research, please cite:

@inproceedings{
kamarthi2024large,
title={Large Pre-trained time series models for cross-domain Time series analysis tasks},
author={Harshavardhan Kamarthi and B. Aditya Prakash},
booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
year={2024},
url={https://openreview.net/forum?id=vMMzjCr5Zj}
}

📋 System Requirements

  • Python: 3.11-3.13
  • OS: Linux (CPU + GPU), macOS (CPU)
  • GPU: NVIDIA GPUs supported

Platform Support

Windows and Apple Silicon GPU support is planned for future releases.