neurolithic/best

ONNX export of the neurolithic lithic-scar segmentation model (UNet++ / EfficientNet-B5, 6-channel input, soft-edge output).

Used directly in-browser by the lithicjs web app (onnxruntime-web). The model is the 2D segmentation network applied to 6 orthographic renders of a PCA-aligned mesh at 512x512; per-view predictions are back-projected and merged on the mesh client-side.

Source checkpoint: learning_curve_100pct_20260305_121700_best.ckpt

Files:

• model_fp32.onnx
• model_fp16.onnx
• config.json — input/inference metadata (channels, resolution, views, etc.)

The exported graph bakes in the per-channel input/output normalization (the checkpoint's real transform stats), so it matches the PyTorch predict path exactly. fp32 is exact; fp16 is ~half the size with a small accuracy trade-off.