wmgifford's picture
Update README.md
7fe295d verified
---
tags:
- generated_from_trainer
- time series
- forecasting
- pretrained models
- foundation models
- time series foundation models
- time-series
license: apache-2.0
pipeline_tag: time-series-forecasting
model-index:
- name: patchtst_etth1_forecast
results: []
---
# PatchTST model pre-trained on ETTh1 dataset
<!-- Provide a quick summary of what the model is/does. -->
[`PatchTST`](https://huggingface.co/docs/transformers/model_doc/patchtst) is a transformer-based model for time series modeling tasks, including forecasting, regression, and classification. This repository contains a pre-trained `PatchTST` model encompassing all seven channels of the `ETTh1` dataset.
This particular pre-trained model produces a Mean Squared Error (MSE) of 0.3881 on the `test` split of the `ETTh1` dataset when forecasting 96 hours into the future with a historical data window of 512 hours.
For training and evaluating a `PatchTST` model, you can refer to this [demo notebook](https://github.com/IBM/tsfm/blob/main/notebooks/hfdemo/patch_tst_getting_started.ipynb).
## Model Details
### Model Description
The `PatchTST` model was proposed in A Time Series is Worth [64 Words: Long-term Forecasting with Transformers](https://arxiv.org/abs/2211.14730) by Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, Jayant Kalagnanam.
At a high level the model vectorizes time series into patches of a given size and encodes the resulting sequence of vectors via a Transformer that then outputs the prediction length forecast via an appropriate head.
The model is based on two key components: (i) segmentation of time series into subseries-level patches which are served as input tokens to Transformer; (ii) channel-independence where each channel contains a single univariate time series that shares the same embedding and Transformer weights across all the series. The patching design naturally has three-fold benefit: local semantic information is retained in the embedding; computation and memory usage of the attention maps are quadratically reduced given the same look-back window; and the model can attend longer history. Our channel-independent patch time series Transformer (PatchTST) can improve the long-term forecasting accuracy significantly when compared with that of SOTA Transformer-based models.
In addition, PatchTST has a modular design to seamlessly support masked time series pre-training as well as direct time series forecasting, classification, and regression.
<img src="patchtst_architecture.png" alt="Architecture" width="600" />
### Model Sources
<!-- Provide the basic links for the model. -->
- **Repository:** [PatchTST Hugging Face](https://huggingface.co/docs/transformers/model_doc/patchtst)
- **Paper:** [PatchTST ICLR 2023 paper](https://dl.acm.org/doi/abs/10.1145/3580305.3599533)
- **Demo:** [Get started with PatchTST](https://github.com/IBM/tsfm/blob/main/notebooks/hfdemo/patch_tst_getting_started.ipynb)
## Uses
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
This pre-trained model can be employed for fine-tuning or evaluation using any Electrical Transformer dataset that has the same channels as the `ETTh1` dataset, specifically: `HUFL, HULL, MUFL, MULL, LUFL, LULL, OT`. The model is designed to predict the next 96 hours based on the input values from the preceding 512 hours. It is crucial to normalize the data. For a more comprehensive understanding of data pre-processing, please consult the paper or the demo.
## How to Get Started with the Model
Use the code below to get started with the model.
[Demo](https://github.com/IBM/tsfm/blob/main/notebooks/hfdemo/patch_tst_getting_started.ipynb)
## Training Details
### Training Data
<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
[`ETTh1`/train split](https://github.com/zhouhaoyi/ETDataset/blob/main/ETT-small/ETTh1.csv).
Train/validation/test splits are shown in the [demo](https://github.com/IBM/tsfm/blob/main/notebooks/hfdemo/patch_tst_getting_started.ipynb).
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 5e-05
- train_batch_size: 8
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 10
### Training Results
| Training Loss | Epoch | Step | Validation Loss |
|:-------------:|:-----:|:-----:|:---------------:|
| 0.4306 | 1.0 | 1005 | 0.7268 |
| 0.3641 | 2.0 | 2010 | 0.7456 |
| 0.348 | 3.0 | 3015 | 0.7161 |
| 0.3379 | 4.0 | 4020 | 0.7428 |
| 0.3284 | 5.0 | 5025 | 0.7681 |
| 0.321 | 6.0 | 6030 | 0.7842 |
| 0.314 | 7.0 | 7035 | 0.7991 |
| 0.3088 | 8.0 | 8040 | 0.8021 |
| 0.3053 | 9.0 | 9045 | 0.8199 |
| 0.3019 | 10.0 | 10050 | 0.8173 |
## Evaluation
<!-- This section describes the evaluation protocols and provides the results. -->
### Testing Data
[`ETTh1`/test split](https://github.com/zhouhaoyi/ETDataset/blob/main/ETT-small/ETTh1.csv).
Train/validation/test splits are shown in the [demo](https://github.com/IBM/tsfm/blob/main/notebooks/hfdemo/patch_tst_getting_started.ipynb).
### Metrics
<!-- These are the evaluation metrics being used, ideally with a description of why. -->
Mean Squared Error (MSE).
### Results
It achieves a MSE of 0.3881 on the evaluation dataset.
#### Hardware
1 NVIDIA A100 GPU
#### Framework versions
- Transformers 4.36.0.dev0
- Pytorch 2.0.1
- Datasets 2.14.4
- Tokenizers 0.14.1
## Citation
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
**BibTeX:**
```
@misc{nie2023time,
title={A Time Series is Worth 64 Words: Long-term Forecasting with Transformers},
author={Yuqi Nie and Nam H. Nguyen and Phanwadee Sinthong and Jayant Kalagnanam},
year={2023},
eprint={2211.14730},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
```
**APA:**
```
Nie, Y., Nguyen, N., Sinthong, P., & Kalagnanam, J. (2023). A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. arXiv preprint arXiv:2211.14730.
```