This is the 1st-place solution for:
Google Research - Identify Contrails to Reduce Global Warming
https://www.kaggle.com/competitions/google-research-identify-contrails-reduce-global-warming
Solution post:
The final model is the ensemble of unet1024 model and vit4 model.
- Version 43: 2 fold + 2 fold is the winning solution (Private LB 0.72432)
- Version 47: 5 fold + 5 fold only differ by random fluctuation (Private 0.72420)
y_pred = w1 * y_vit4 + w2 * y_unet1024 > 0.5
See notebook/ for the weights.
The base single-time U-Net model with 256 or 512 input image size in each dimension.
U-Net model with 1024x1024 input image and 512x512 segmentation target. Almost same as the base model, but one less upscale decoder layer.
The input is four panels of 512x512 inputs at t=1,2,3,4, concatenated to one 1024x1024 input image. Target is 512x512 image for t=4.
The inference code and model weights are online on Kaggle:
- https://www.kaggle.com/code/junkoda/contrails-submit
- https://www.kaggle.com/datasets/junkoda/contrails-src
- https://www.kaggle.com/datasets/junkoda/contrails-weights
- entry_points.md: Instructions how to train the model
- SETTINGS.json: Directory configuration for training
- Training the final model would take 30 hours per fold with A100 GPU (40GB RAM)
- Python 3.10.10
- CUDA 11.7 (NVIDIA driver Version: 530.30.02)
- PyTorch 2.0.1
- timm 0.9.2
- segmentation-models-pytorch 0.3.3
- albumentations 1.3.0
- h5py 3.9.0
- Standard packages: numpy, pandas, tqdm
- requirements.txt: List of all python modules, more than sufficient
MIT