Skip to content

Troubleshooting

General tips

  • Run order matters: Unified Detection → Find Similar → Modify Detections.
  • Prompt hygiene: For Florence-2/OWL, prefer clear, concrete phrases.
  • Memory management: Operator frees VRAM, but splitting stages or lowering batch/points can help with large datasets.
  • Unknown-only safety: Always default to unknown_only=True for bulk edits to preserve verified labels.

Setup & Environment

MongoDB Database Issues

  1. Stop running FiftyOne MongoDB processes: 
    pkill -f 'fiftyone/db/bin/mongod' || true
  2. Remove stale lock and PID files: 
    rm -f ~/.fiftyone/var/lib/mongo/{mongod.lock,WiredTiger.lock} ~/.fiftyone/var/lib/mongo/mongod-*.pid

Plugin Registration Issues

  1. Remove plugin: 
    rm -rf ~/fiftyone/__plugins__/latent_label
  2. Force reinstall: 
    pip install --force-reinstall --extra-index-url https://${REPOSITORY_TOKEN_NAME}:${REPOSITORY_TOKEN_PASSCODE}@repository.latentai.com/repository/pypi/simple latent-label
  3. Restart: 
    latent-label start

Docker Issues

Check service status: 

docker compose ps

View logs: 

docker compose logs -f latent-label

Restart services: 

docker compose restart latent-label

Verify data directory permissions: 

ls -la /path/to/your/images

Unified Detection

Symptom Likely cause Fix
AttributeError: Field 'get' was not selected on this SampleView Operating on a view that doesn’t include unified_detections Refresh or switch to a base dataset; re-run Unified Detection to register bbox_area and unified_detections.
VLM duplicates or overlaps curated labels detection_mode=both with overlapping boxes Adjust overlap_iou_threshold or accept default merge rules (OWL preferred)
Too many tiny or noisy boxes Area thresholds too low Increase min_area_percentage and/or adjust sam_nms_thresh
GPU memory pressure Batch too large or models loaded together Split detection_mode (VLM-only or SAM-only), reduce points_per_batch, or lower batch_size
Florence-2 boxes cover multiple objects Florence-2 sometimes outputs group boxes (confidence = 1.00) Run OWL2 alone instead of picking both

Find Similar

Symptom Likely cause Fix
LDA never runs Not enough labeled examples or classes Ensure ≥ min_classes_for_lda distinct labels with ≥ min_examples_per_class each; seed labels manually first
Clusters not meaningful Dataset has few labeled detections or highly variable objects Add labels via Modify Detections, then re-run FS
UMAP visualization looks empty or distorted Low-quality embeddings or very sparse data Ensure leip_embedding exists and was computed; refresh brain view; adjust n_neighbors if needed
Embeddings don't appear Browser didn't update state Refresh your browser window

Modify Detections

Symptom Likely cause Fix
Labels changed for confirmed detections unknown_only=False Re-run with unknown_only=True
Detection not removed Applied on patch view but not parent sample Use dataset-level target or ensure patch updates propagate
Changes not visible FiftyOne view not refreshed Refresh dataset view or reopen FiftyOne session
Bulk relabel didn’t propagate to LDA New labels not yet seeded After Modify Detections, re-run Find Similar to update clusters and trigger LDA