Skip to content

Latest commit

 

History

History
49 lines (37 loc) · 3.73 KB

MODEL_ZOO.md

File metadata and controls

49 lines (37 loc) · 3.73 KB

Model Zoo

We provide the experimental results of the MGMA method in SSPL framework for AVP. We implement two MGMA-Nets, including the MGMA-ResNet (RNB, built upon the classic ResNet architecture) and the MGMA-ShuffleNet (SFNB, built upon the lightweight ShuffleNet architecture), used for conducting experiments. We present the performance (RMSE, MAE) and efficiency (Params, FLOPs) metrics of our MGMA-SSPL models in predicting ERA5 variables (T2M, UV10, R and TCC). All the models are trained for 50 epochs, and they can be downloaded via the Google Drive or Baidu Netdisk links.

The Efficiency of ST-SSPL Models

Method Params FLOPs FPS
SFNB-Base 0.37M 2.98G 993
RNB-Base 0.48M 3.66G 931
SFNB-MGMA 0.55M 4.14G 407
RNB-MGMA 0.66M 4.81G 396

The Results of ST-SSPL Models on Temperature (t2m)

Method Variable RMSE MAE Config
SFNB-Base t2m 1.1154 0.7133 configs/weather/t2m_5_625/MGMA_ShuffleV2_NONE.py
RNB-Base t2m 1.1348 0.7339 configs/weather/t2m_5_625/MGMA_Bottleneck_NONE.py
SFNB-MGMA t2m 1.0831 0.6760 configs/weather/t2m_5_625/MGMA_ShuffleV2.py
RNB-MGMA t2m 1.0726 0.6689 configs/weather/t2m_5_625/MGMA_Bottleneck.py

The Results of ST-SSPL Models on Wind Component (uv10)

Method Variable RMSE MAE Config
SFNB-Base uv10 1.3692 0.9430 configs/weather/uv10_5_625/MGMA_ShuffleV2_NONE.py
RNB-Base uv10 1.3606 0.9381 configs/weather/uv10_5_625/MGMA_Bottleneck_NONE.py
SFNB-MGMA uv10 1.2938 0.8660 configs/weather/uv10_5_625/MGMA_ShuffleV2.py
RNB-MGMA uv10 1.2855 0.8600 configs/weather/uv10_5_625/MGMA_Bottleneck.py

The Results of ST-SSPL Models on Humidity (r)

Method Variable RMSE MAE Config
SFNB-Base r 5.8830 4.1028 configs/weather/r_5_625/MGMA_ShuffleV2_NONE.py
RNB-Base r 5.8999 4.1061 configs/weather/r_5_625/MGMA_Bottleneck_NONE.py
SFNB-MGMA r 5.6384 3.8036 configs/weather/r_5_625/MGMA_ShuffleV2.py
RNB-MGMA r 5.6376 3.8242 configs/weather/r_5_625/MGMA_Bottleneck.py

The Results of ST-SSPL Models on Cloud Cover (tcc)

Method Variable RMSE MAE Config
SFNB-Base tcc 0.2250 0.1588 configs/weather/tcc_5_625/MGMA_ShuffleV2_NONE.py
RNB-Base tcc 0.2253 0.1577 configs/weather/tcc_5_625/MGMA_Bottleneck_NONE.py
SFNB-MGMA tcc 0.2150 0.1461 configs/weather/tcc_5_625/MGMA_ShuffleV2.py
RNB-MGMA tcc 0.2150 0.1467 configs/weather/tcc_5_625/MGMA_Bottleneck.py