CellposeModel Eval

How does CP eval work?

• Normalizes (though this has no effect)
• Adds zero padding (256x256 -> 272x272) [also, make it divisible by 16]
• Runs in:
  • _run_tiled(net, imgi, batch_size=8, augment=False, bsize=224, tile_overlap=0.1):
    • [cellpose/core.py]
  • transforms.make_tiles:
    • [cellpose/transforms.py]
    • Creates for 224x224 tiles out of the 272x272 padded img
  • PREDICTS on those 224x224 tiles
  • Transforms.average_tiles
  • Limit back size to 272x272
  • Remove (slice out) the zero padding
  • Returns padded result

Performance

CP eval prediction (with the tiling stuff described above)

CP compute mask prediction (same parameters as CP eval [should be], but without the tiling described above_.

Also, without the 99th percentile normalization being reapplied.

CP TTA compute mask prediction