Kaggle-HuBMAP

Codes for Kaggle competition: HuBMAP - Hacking the Kidney

Instructions

Environment

All codes are based on pytorch and pytorch-lightning, and other common deep learning, image processing packages. Most of imported packages can be easily installed by conda or pip.

Data processing

Training data abd public test data are from Kaggle competition data source, and external data is downloaded from official HuBMAP dataset port.

Visual.ipynb is for visualization of images and corresponding mask, which is not neccessary.
Slicer.ipynb is for slicing WSI into smaller patches. This should be executed before training.

Training phase

This code version only contains EfficinetNet-b3-U-Net training. For other backbones, please refer to timm packages and segmentation_models_pytorch for available U-Net backbones.

Solver.py is training code of training phase without pseudo-labeling.
Simply run python Solver.py.
Trianing config should be set inside Solver.py, including learning rate, batch size, number of epochs, backbone and so on. Most of configs are shown in first several lines of Solver.py
SolverPseudo.py is training code of training phase with pseudo-labeling.

Inferencing phase