Yandex Cup 2023 Machine Learning Challenge: Nowcasting Develop AI model for nowcasting.
2nd place out of 89 participants with 166.3084 RMSE (top1 164.9235, top3 166.4652).
Time spent
The approach is based on an UNet model. The inputs to the encoder are 4 previous channel-wise stacked frames with resolution of 252x252 of intensities, events, first 4 reflectivity levels and month. We normalize inputs with min-max scaler. The outputs are intensities for the next 2 hours (12 frames). We directly optimize RMSE using AdamW optimizer and CosineAnnelingLR scheduler. We use vertical flips, rotations as augmentations.
- unet model
- channel-wise stacked 4 previous frames (intensity, events, reflectivity)
- outputs intensity for the next 2 hours (12 frames)
- rmse loss
- random flips and rotations augmentations
- left-right and up-down flips test time augmentations
- Ubuntu 20.04.6 LTS (GNU/Linux 5.4.0-125-generic x86_64)
- Python 3.8.10 (GCC 9.4.0)
- Nvidia driver version: 515.65.01
- CUDA version: 11.8
- NVIDIA Tesla V100-SXM2 32GB
- PyTorch 2.0.1
- Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz
- 64 GB RAM
Download the data from the competition page and unzip into data folder. Then
download pretrained models from yandex disk.
Run the script to produce the submission
sh ./submit.shThe inference takes ~1 minute on 1 GPU V100 32GB.
To train a model from scratch
sh ./run.shIt takes ~39 hours on 1 V100 to train all the models. Actually, the models are trained faster (4-8 epochs), so you may stop earlier and reduce overall time to ~12 hours.