We introduce the source code of our paper "Generative Time Series Forecasting with Diffusion, Denoise, and Disentanglement" (NeurIPS 2022). If you find this code or any of the ideas in this paper useful, please consider citing:
@inproceedings{li2022generative,
title={Generative Time Series Forecasting with Diffusion, Denoise, and Disentanglement},
author={Yan Li and Xinjiang Lu and Yaqing Wang and Dejing Dou},
booktitle={Advances in Neural Information Processing Systems},
editor={Alice H. Oh and Alekh Agarwal and Danielle Belgrave and Kyunghyun Cho},
year={2022}
}We propose a generative time series forecasting method on top of
VAE equipped with diffusion, denoise, and disentanglement, namely
We implement
Run the following command to install the required dependencies.
pip install -r requirements.txt
Change the $root_path and $data_path to your own data directory and dataset name, respectively. This code only supports .csv file now.
The public datasets used in the paper are listed as follows.
Electricity: https://archive.ics.uci.edu/ml/datasets/ElectricityLoadDiagrams20112014
Traffic: http://pems.dot.ca.gov
Weather: https://www.bgc-jena.mpg.de/wetter/
ETTs: https://github.com/zhouhaoyi/ETDataset
Besides, we adopt a collected WindPower dataset which is available here.
Run the following commands to reproduce the results in our paper,
you need to change the $root_path and $data_path to your own path and data name.
And you need to set the --sequence_length and --prediction_length to train and forcast time series in different lengths.
Traffic
python main.py --root_path $root_path --data_path $data_path \
--input_dim 862 --percentage 0.05 --diff_steps 1000 \
--beta_end 0.1
Electricity
python main.py --root_path $root_path --data_path $data_path \
--input_dim 321 --percentage 0.03 --diff_steps 1000 \
--beta_end 0.1
Weather
python main.py --root_path $root_path --data_path $data_path \
--input_dim 21 --percentage 0.02 --diff_steps 1000 \
--beta_end 0.1
ETTm1
python main.py --root_path $root_path --data_path $data_path \
--input_dim 7 --percentage 0.01 --diff_steps 1000 \
--beta_end 0.1
ETTh1
python main.py --root_path $root_path --data_path $data_path \
--input_dim 7 --percentage 0.05 --diff_steps 1000 \
--beta_end 0.1
Wind
python main.py --root_path $root_path --data_path $data_path \
--input_dim 7 --percentage 0.02 --diff_steps 1000 \
--beta_end 0.1
| Params | Description |
|---|---|
| --sequence_length | The length of input series. |
| --prediction_length | The length of output series, which should be exactly the same as the input length. |
| --input_dim | Number of dimensions of input series. |
| --embedding_dimension | Number of dimensions of the embedding. |
| --percentage | Percentage of the original dataset, it should be set between [0, 1] to decide the percentage of the sliced dataset. |
| Params | Description |
|---|---|
| --diff_steps | The maximum number of the diffusion steps. |
| --beta_schedule | The type of the diffusion schedule, it could be set to ['quad', 'linear', 'const', 'jsd']. |
| --beta_start | The start value of the diffusion schedule. |
| --beta_end | The end value of the diffusion schedule. |
| --scale | The scale parameter of the target diffusion process. |
[NOTE] The implementation of bidirectional vae architecture is built on top of NVAE.
| Params | Description |
|---|---|
| --num_layers | #layers adopted in the RNN model. |
| --num_channels_enc | #channels of the encoder in BVAE. |
| --num_preprocess_blocks | #preprocess blocks. |
| --num_preprocess_cells | #cells in a preprocess block. |
| --groups_per_scale | #total groups. |
| --num_postprocess_blocks | #post-process blocks. |
| --num_channels_dec | #channels of the decoder in BVAE. |
| --num_latent_per_group | #channels of the latent variables in each group. |
Some basic results can be seen as follows. Note that, to obtain better performance, more dedicated parameter tuning is required.
