fahim-sikder/TransFusion

This is the official repository for the paper "TransFusion: Generating Long, High Fidelity Time Series using Diffusion Models with Transformers"

TransFusion: Generating Long, High Fidelity Time Series using Diffusion Models with Transformer

This repo contains the code for TransFusion paper (https://arxiv.org/abs/2307.12667).

First, install the docker container by running the shell build.sh, after that run the environment by using run-shell.sh shell.

chmod +x build.sh
./build.sh
chmod +x run-shell.sh
./run-shell.sh

Training

To train TransFusion, run the following command inside the docker container.

Dataset Option: sine, stock, air, energy

To download the dataset (stock, air and energy), please refer to the supplementary document in this repo. Create a folder called data and paste the datasets there. If you want to change the code of dataset please refer to the data_make.py file.

python3 train.py --dataset_name sine --seq_len 100 --batch_size 256

Generated samples from the runtime can be found in the saved_files directory after running the experiments.

For Long-Sequence Discriminative score metric, we need both original data and synthetic data as torch tensor.

from long_discriminative_score import long_discriminative_score_metrics


long_discriminative_score_metrics(original_data, synthetic_data)

For Long-Sequence Predictive score metric, we need both original data and synthetic data as torch tensor.

from long_predictive_score import long_predictive_score_metrics

long_predictive_score_metrics(original_data, synthetic_data)

Citation

Please refer to our work if you use any parts of the code:

@article{sikder2023transfusion,
  title={Transfusion: Generating long, high fidelity time series using diffusion models with transformers},
  author={Sikder, Md Fahim and Ramachandranpillai, Resmi and Heintz, Fredrik},
  journal={arXiv preprint arXiv:2307.12667},
  year={2023}
}