In the course of the HIDA DATATHON for Grand Challenges on Climate Change 2020.
Using high level open source frameworks for landscape image segmentation with PyTorch
The configuration
architecture | pyramid attention network (PAN); encoder | se_resnext50_32x4d; batchsize | 8; learning rate | 0.0001; activation | sigmoid; data_size | 256 x 256
Used resources:
PyTorch: https://pytorch.org/
PyTorch Lightning: https://pytorchlightning.ai/
Segmentation models: https://github.com/qubvel/segmentation_models.pytorch
Validation segmentation DICE score: ~ 95%