(September 2025) We are happy to announce that the paper has been accepted to NeurIPS2025!

🧪 ImagenFew: Unified Time Series Generation Under Data Scarcity

📄 Read the Paper

This repository accompanies the paper "Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach", which introduces a unified diffusion-based framework for generating high-fidelity time series, especially in few-shot regimes. The model leverages pretraining on a diverse set of datasets, dynamic convolutions for adaptive channel handling, and dataset token conditioning for domain-aware generation.

---

💡 What You Get

• Full pipeline to reproduce the paper: Preprocessed datasets, pretrained checkpoints, and scripts for training, fine-tuning, and evaluating the unified diffusion model under data scarcity.
• Flexible framework for custom models: Easily plug in your own architecture (e.g., Diffusion, GANs, VAEs), run it within the same environment, and benchmark it using standardized protocols and metrics.
• Modular data design: Add new datasets with minimal effort—data loading, preprocessing, and integration are all streamlined for multi-domain support.

---

🛠️ Installation

Clone the repository:

git clone git@github.com:azencot-group/ImagenFew.git
cd ImagenFew

📦 Environment Setup

Using Python virtualenv:

python -m venv .venv
source .venv/bin/activate  # or `./.venv/Scripts/activate` on Windows
pip install -r requirements.txt

Using Conda:

conda create --name ImagenFew python=3.10 pip
conda activate ImagenFew
pip install -r requirements.txt

---

📁 Datasets

Preprocessed datasets are available for both pretraining and fine-tuning:

➡️ Download from: https://drive.google.com/file/d/1EHO1EXBJYg1ohFKxmJjV548qKtbSkVhF/view?usp=drive_link

Unzip the contents into the empty ./data directory.

Pretraining datasets include:

Stocks, Energy, ETTh1, Exchange, MSL, SMAP, PSM, SMD, ECG5000, NonInvasiveFetalECGThorax1, and more.

Fine-tuning/evaluation datasets:

MuJoCo, ETTm1/m2/h2, Sine, Weather, ILI, ECG200, AirQuality, Saugeen River Flow, etc.

---

💾 Pretrained Models

Checkpoints for our pre-trained models are available in the same Google Drive folder: https://drive.google.com/file/d/16MMBjyKT6VH7YwCshp2I5RXx4FQhStiA/view?usp=drive_link

The attached ZIP archive contains four pre-trained ImagenFew models, each using the ./config/pretrain/pretrain.yaml configuration file, with different pretraining sequence lengths.

---

🚀 Getting Started

📌 Configurations

Configuration files are available in the ./configs folder for all datasets and tasks.

Pretraining

python run.py --no_test_model --seq_len <sequence_length> --config <config_file>
# or
./scripts/run_pretrain.sh

Fine-tuning

Percentage-based:

python run.py --subset_p <percent> --model_ckpt <ckpt_path> --config <config_file>

Count-based:

python run.py --subset_n <count> --model_ckpt <ckpt_path> --config <config_file>

Scripts for all datasets:

./scripts/run_finetune_percentage.sh
./scripts/run_finetune_count.sh

---

📊 Visualization

Use PCA, t-SNE and other visualizations:

python run_visualization.py --model_ckpt <ckpt_path> --config <config_file>
# or
./scripts/run_visualization.sh

---

🧩 Plug In Your Own Model

Want to use your own generative architecture? Just follow these steps:

1. Create a folder under ./models/ with your model name.
2. Implement Handler that extends generativeHandler:
   • build_model(): instantiate the model
   • train_iter(): define a single training iteration
   • sample(): generate samples for a given class label
3. Optional: Take advantage of our DDP support for distributed training.
4. In your configuration file, specify the handler path using the format: handler: models.{model_folder}.{handler_file}

---

🔬 Citation

If you use this work, please cite:

@inproceedings{
gonen2025time,
title={Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach},
author={Tal Gonen and Itai Pemper and Ilan Naiman and Nimrod Berman and Omri Azencot},
booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
year={2025},
url={https://openreview.net/forum?id=p324ryBKTc}
}