Expand description
Phase 4 / V4 — frozen embedding + compact linear classifier head.
Uses the existing CPU-reference embedding (or any VisualModel embed) and a
caller-supplied weight matrix for a closed vocabulary. No Python, no deep
training loop — head can be filled offline or by simple accumulators.
Structs§
- Linear
Head - Compact linear head:
logits[c] = bias[c] + sum_i embed[i] * weight[c * dim + i]. - Linear
Probe Vision - Wraps a base visual model: takes its embedding, runs linear head, emits 0–1 detection.
Functions§
- fit_
two_ class_ centroids - Train a 2-class head by averaging embeddings of two example batches (centroid).