Scripts for the my 6th place solution with MXNet.
Key features:
- Unet-like network architecture with multiple input branches(A, M, P channels)
- Adaptive crop sampler, based on the performance of the network on different classes, updated each epoch.
Brief description:
b3_data_iter - data generation, augmentation
utils - wkt polygon rasterization, raster polygonization utils
b3_train - network symbol definition, training loop