Repository navigation
Inference API: What should the inference API return per image? #4
Answered
by
Ray3R3
D-M-Pearman
asked this question in
Q&A
|
From DMP: Before we build the API: what's the contract for a single inspection? Just a pass/fail and a score, or something richer? Want to make sure the evaluation harness and the eventual frontend can both use it without reshaping. |
Answered by
Ray3R3
Nov 16, 2026
Replies: 1 comment
|
Richer than pass/fail. It returns an
One constraint this creates: |
0 replies
Answer selected by
D-M-Pearman
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Richer than pass/fail. It returns an
InspectionResult(src/vista/schema.py), which is already validated and tested:image_scoreandthresholdare separate fields. Callers can re-threshold (stricter for safety-critical parts) without re-running the model, and evaluation can sweep thresholds for AUROC from stored results.defectsis a list of regions with a bounding box in original-image pixels and a per-region score.labelis optional because PatchCore is unsupervised: it knows where something is wrong, not what it is. A classifier can fill that in later without a schema change.mask_path) rather than inlined, so a result stays small enough to log and store p…