M181_binary

First stage of the PISCO dual-ViT: living vs not-living. Fine-tuned google/vit-base-patch16-224-in21k on expert-validated crops from cruise M181 (tropical Atlantic, April-May 2022). Test accuracy 97.65%.

Classes (2): living, not-living

Important: this model is coupled to its preprocessing

Trained on crops produced by the 231204 LUCYD deconvolution (lucyd-edof-plankton_231204.pth, the pipeline default) with non-isolated crops (neighbouring particles present). Running it on crops from a different deconvolution model, or on neighbour-isolated crops, is an out-of-domain shift: measured on ATAIR-BSH, swapping the deconvolution alone collapsed the living rate by 83-93% at unchanged confidence. Treat deconvolution + crop isolation + classifier as one versioned unit.

Preprocessing: resize longest edge to 224, centre-pad to 224x224 with white (255), normalise mean=std=0.5. No rotation at inference. Crops must carry no scale bar (or have it removed with utils.strip_scale_bar).

Use in the pipeline

These are the pipeline defaults (--binary-model-dir / --living-model-dir), or from any machine:

process_pisco_profiles.py --binary-model-hf Veit/M181_binary --living-model-hf Veit/M181_multiclass
process_pisco_profiles.py --dualvit-model M181     # resolves both from the Hub when not local