AA-TransUNet: Attention Augmented TransUNet For Nowcasting Tasks

AA_TransUNet Architecture.

Datasets & Pre-trained Models

If you are interesed in the dataset(precipitation maps & cloud cover dataset) used in this paper，please visit SmaAt-UNet for further details.

For the pre-trained AA_TransUNet models, please contact us: [email protected] or [email protected]

Please put the dataset into "\dataset" directory for training and testing, and put the pre-trained models into "\results\Model_Saved" directory for future loading.

Usages

Required Dependencies:

pip install -r requirements.txt

Training precipitation maps dataset:

python train_precipitation.py

Training cloud cover dataset:

python train_cloud_cover.py

After training or loading pre-trained models, you can evaluate model's performance by:

Evaluating precipitation maps dataset:

python evaluate_precipitation.py

Evaluating cloud cover dataset:

python evaluate_cloud_cover.py

For easier training, we also provide a colab demo:

Colab training demo for cloud cover dataset.

Authors

• Yimin Yang: [email protected]
• Siamak Mehrkanoon: [email protected]

Citation

 @inproceedings{  
    yang2022aa,  
    title={Aa-transunet: Attention augmented transunet for nowcasting tasks},  
    author={Yang, Yimin and Mehrkanoon, Siamak},  
    booktitle={2022 International Joint Conference on Neural Networks (IJCNN)},  
    pages={01--08},  
    year={2022},  
    organization={IEEE}  
 }