My top 6% approach to the HPA Single Cell Classification Kaggle Competition. The code includes peak response maps, puzzle-cam, and transformer encoders using RoI (cell mask) pooling over the CNN feature maps.
See this page for a more detailed description of my approach.
Overview of the cell-mask-based Transformer model (click on the image for higher resolution):
Overview of the cell-mask-based RoI pooling layer with cell position encoding:
To install the baseline cell segmentation model, run the following:
pip install https://github.com/CellProfiling/HPA-Cell-Segmentation/archive/master.zip
- download public dataset
- exclude images
- multilabel stratify split
- train