I learn a lot by making a "HuBMAP - Hacking the Kidney" kaggle competition https://www.kaggle.com/c/hubmap-kidney-segmentation. The objective of this challenge is segmentation of regions with glomeruli in human kidney tissue images across different tissue preparation pipelines. In this kernel I provide a Pytorch( using GPU )code based on a network UNet. We will write data pipeline for dividing the images into tiles and selection of tiles with tissue. After, we will train and evaluate of UNet model, combining the tile masks back into image level masks, conversion into RLE, and visualize the predicted mask for images tests.
This kernel provides a starter Pytorch-GPU code for inference that performs dividing the images into tiles, selection of tiles with tissue, Train and evaluate of UNet(encode = efficientnet-b3) model, combining the tile masks back into image level masks, conversion into RLE, and visualize the predicted mask for images tests. The inference is pretrained based on models trained in the kernel.