Troubleshooting
What can go wrong
Find the symptom below and check the report before changing settings.
Skipped runs do not count as successful results.
Common problems
Observed symptom
The yolozu command is not found
- Verify
- Confirm which Python environment received the package, then run python3 -m yolozu --help from that same interpreter.
- Next action
- Reinstall into the intended environment; do not switch runtimes silently.
Detailed guidance
Observed symptom
doctor --proof fails
- Verify
- Read the errors and drift_hints in the doctor report. The proof is CPU-only and does not require a model download.
- Next action
- Resolve the reported package, path, or write-permission problem, then rerun into a fresh proof directory.
Detailed guidance
Observed symptom
Predictions validation reports missing or migrated fields
- Verify
- Run strict validation on the wrapped artifact and compare its schema version with the predictions schema.
- Next action
- Fix the producer or use the documented migration path before scoring.
Detailed guidance
Observed symptom
Dataset validation cannot resolve the split
- Verify
- Use the split recorded in the dataset/proof artifact instead of assuming train, val, or val2017.
- Next action
- Correct the dataset manifest or command so image IDs and split identity agree.
Detailed guidance
Observed symptom
Two reports have different metrics
- Verify
- Compare task, dataset/split, prediction source, export settings, protocol identity, class mapping, thresholds, and skip reasons.
- Next action
- Align the differing input or protocol field before attributing the difference to a model or runtime.
Detailed guidance
Observed symptom
A backend row is skipped or conditional
- Verify
- Read the canonical support matrix and the report's explicit status/reason.
- Next action
- Provide the required artifact/runtime or keep the lane skipped; never reinterpret it as successful parity.
Detailed guidance
Observed symptom
A Research run changes the result
- Verify
- Verify that the stable baseline predictions and evaluation report are preserved and linked.
- Next action
- Treat the delta as Research evidence until a fixed protocol, rollback review, and environment qualification support a stronger conclusion.
Detailed guidance
Observed symptom
The visible demo has no overlays
- Verify
- Confirm the demo report and overlays directory under the exact run directory printed by the command.
- Next action
- Rerun the instance-seg demo into a new run directory and inspect the reported output paths.
Detailed guidance