Data:

• src | original competition data from kaggle
• Sartorius | prepared data used for training, can be downloaded here
  • LiveCell: (pretrain)
    • added .png masks
    • change image format to png
    • create lightweight coco format .json annotations (following this notebook). This decreases the file size from ~550MB to 230MB for the training annotations.
  • SartoriusCellSeg: (competition)
    • added .png masks
    • copied annotation jsons from Slawek Biels Dataset

Notebooks:

• Detectron: Use a custom trainer with 1cycle policy to (pre-)train for the Sartorius instance segmentation challenge
  • WandB-Train Record training with Weights & Biases
  • CellSeg-Inference Get predictions for test set and create submission; copied from here
• Switch: Using fastai with timm, switch between classification & segmentation (Unet) while sharing a backbone
  • LiveCell-Pretraining
  • CellSeg-Training
  • CellSeg-Inference Get predictions and transform from semantic to instance segmentation, create submission
• LiveCell: Prepare the livecell data for pretraining
  • coco_annotations
  • create_pngMasks
  • tiff_to_png
• Sagemaker: Upload big files from sagemaker to dropbox
  • UploadToDropbox