Official implementation code of the paper: "TENT: Tensorized Encoder Transformer for temperature forecasting" (ArXiv link).
(a) Model architectureTensorial. (b) Tensorial self-attention. (c) Tensorial multi-head attention.
The obtained test MAE of the models for USA-Canada dataset averaged over cities (a) and prediction time steps (b).
The comparison between the predictions of TENT model and the real measurements for hourly temperature of the test set of Vancouver.
Attention visualization for Dallas in USA-Canada dataset.
In order to download the data, please email to the following address:
Execute the notebook on colab (Use TPU for TENT): TT_All_models_experiments.ipynb
If you use our data and code, please cite the paper using the following bibtex reference:
@article{bilgin2021tent,
title={TENT: Tensorized Encoder Transformer for Temperature Forecasting},
author={Bilgin, Onur and M{\k{a}}ka, Pawe{\l} and Vergutz, Thomas and Mehrkanoon, Siamak},
journal={arXiv preprint arXiv:2106.14742v2},
year={2021}
}