ViTime: A Visual Intelligence-based Foundation Model for Time Series Forecasting

ViTime is a novel Visual Intelligence-based foundation model for time series forecasting (TSF), leveraging visual data processing paradigms to enhance forecasting capabilities. [Paper]

Key Features

• Visual Intelligence: Utilizes visual data processing to enhance time series analysis and forecasting.
• Real Time Series (RealTS): A novel data synthesis method improving model training and performance.
• Zero-Shot Performance: Achieves state-of-the-art zero-shot forecasting on diverse datasets, often surpassing the best supervised models.

Getting Started

Quick start

Try ViTime via [Colab]!

Prerequisites

• Python 3.9+
• PyTorch

Installation

Clone the repository:

git clone https://github.com/IkeYang/ViTime
cd ViTime

Install the required packages:

pip install -r requirements.txt

Checkpoints

We provide weights for three models in the following links:

1. ViTime_V1.pth: Specialized for look back window = 512

   • Model Weights

2. ViTime_1072.pth: Specialized for look back window = 1072

   • Model Weights

3. ViTime_V2.pth: Flexible input look back window = [256, 1024]

   • Model Weights

Model Inference

To run the inference code, follow these steps:

1. Load the Pretrained Model: Ensure you have the pretrained model file ViTime_V2.pth.

2. Prepare the Inference Script:

import matplotlib.pyplot as plt
from model.model import ViTime
import numpy as np
import torch

deviceNum = 0
torch.cuda.set_device(deviceNum)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
checkpoint = torch.load(r'ViTime_V2.pth', map_location=device)
args = checkpoint['args']
args.device = device
args.flag = 'test'

# Set upscaling parameters
args.upscal = True  # True: max input length = 512, max prediction length = 720
                    # False: max input length = 1024, max prediction length = 1440
model = ViTime(args=args)
model.load_state_dict(checkpoint['model'])
model.to(device)
model.eval()

# Example data
xData=np.sin(np.arange(512)/10)+np.sin(np.arange(512)/5+50)+np.cos(np.arange(512)+50)
args.realInputLength = len(xData)
yp = model.inference(xData)

# Plot results
plt.plot(np.concatenate([xData,yp.flatten()],axis=0),label='Prediction')
plt.plot(xData,label='Input Sequence')
plt.legend()
plt.show()

3. Run the Inference:

python inference.py

This script loads the ViTime model, performs inference on example sinusoidal data, and plots the predicted sequence alongside the input sequence.

Experiments and Results

ViTime has been evaluated on a diverse set of previously unseen forecasting datasets, demonstrating state-of-the-art zero-shot performance. The model often surpasses the best individually trained supervised models, showcasing the potential of visual intelligence in time series forecasting. The proposed ViTime, when fine-tuned with 10% of the training data, outperforms the latest supervised models that were trained on 100% of the data. Furthermore, when finetuned with 100% of the data, ViTime's prediction accuracy greatly surpasses all existing models, further demonstrating its effectiveness.

Contributing

We welcome contributions to enhance ViTime. Please submit pull requests or open issues for any improvements or bug fixes.

License

This project is licensed under the MIT License.