ActiveLabelingSystem

Formerly known as LabelOps.

Local-first AI-assisted image labeling with active learning, manual box tools, background retraining, and dataset version snapshots.

New features and updates

• New Feature: Freehand region labeling for polygon-style segmentation in manual mode.
• New Feature: Manual toolbox mode switch between bounding boxes and freehand regions, with highlight toggling and point undo.
• New Feature: Active learning strategy selector with Uncertainty, Margin, Diversity, and Balanced modes.
• New Feature: Previous-image navigation and auto-accepted image log from the review bar.
• Update: Modular app structure (src/app, src/core, src/features) for cleaner maintenance.
• Update: Active learning image prioritization when loading folders (unlabeled images first, sorted by uncertainty).
• Update: Retraining policy engine with multi-signal checks (sample count, time, entropy shift, class balance, confidence drift).
• Update: Dataset versioning from UI with metadata, hash integrity, and manifest tracking.
• Update: Label format selection at folder load time: COCO JSON or Plain JSON.
• Update: Safer persistence:
  • autosave per folder (labels_autosave.json)
  • internal state store (.labels_internal.json)
  • atomic file writes for exports.
• Update: Improved manual labeling UX:
  • floating toolbox
  • box and freehand region modes
  • quick class switching (1-9)
  • undo/delete/point-undo shortcuts
  • inline floating action toolbar.
• Update: Better training controls in UI:
  • force retrain
  • live queue/training status
  • shadow model promotion with validation warning flow.

Features

Active learning and triage

• Entropy-aware detection metadata is attached to predictions.
• Folder loading prioritizes unlabeled images for high-value review.
• Replay buffer preserves historical samples for continual learning.

Retraining workflow

• Background training via Ray shadow trainer.
• Training trigger policy requires minimum sample count plus at least one urgency signal.
• Promotion flow supports validation and explicit override confirmation.

Dataset versioning

• Create version snapshots from current labels.
• Stores YOLO-style labels, copied images, metadata, and a dataset hash.
• Keeps lineage and latest pointer in src/datasets/manifest.json.

Label export

• Plain JSON output (labels.json) for direct app consumption.
• COCO JSON output (labels_coco.json) for downstream ML pipelines.

Manual labeling mode

• Draw boxes directly on the canvas.
• Draw freehand regions as polygon segmentations.
• Per-box class assignment with deterministic class colors.
• Save-and-next loop without leaving manual mode.

Project structure

ActiveLabelingSystem/
  images/
  src/
    app/
      window.py
      dialogs.py
      state.py
      actions.py
    core/
      data_manager.py
      entropy.py
      sample_selector.py
      retrain_policy.py
      dataset_versioner.py
      replay_buffer.py
      shadow_trainer.py
      training_orchestrator.py
      model_manager.py
      feedback_validator.py
    features/
      manual.py
      shortcut_manager.py
      shortcut_config.py
      toolbar_manager.py
      toolbar_widget.py
      toolbar_styles.py
    datasets/
    models/
    main.py
    requirements.txt
    run_tests.bat
  README.md
  GUID.md

Installation

git clone https://github.com/sairam-s0/ActiveLabelingSystem.git
cd ActiveLabelingSystem
python -m venv .venv
# Windows
.venv\Scripts\activate
# Linux/macOS
# source .venv/bin/activate

pip install .

Users can install it directly with:

pip install Active-Labeling-System

Run the app

After install, run the bootstrap/setup command first:

als

This command:

• runs run_tests.bat on Windows
• checks Python and runtime dependencies
• installs missing packages automatically
• detects GPU hardware
• installs CUDA-enabled PyTorch automatically for NVIDIA GPUs
• falls back to CPU mode for unsupported GPU setups

After setup completes, start the GUI with:

als --start

How to use

1. Select image folder and output format

• Click Select Folder.
• Choose output format:
  • COCO JSON -> writes labels_coco.json
  • Plain JSON -> writes labels.json
• The app also keeps internal state in .labels_internal.json and autosave in labels_autosave.json inside the selected folder.

2. Select classes

• Click Select Classes.
• Pick one or more classes.
• Add custom classes from the same dialog when needed.

3. Start labeling

• Click START.
• Review detections and use bottom actions:
  • Previous (P)
  • Accept (A)
  • Reject (R)
  • Skip (N)
  • Manual (M)
  • Log

4. Manual mode (box drawing)

• Draw boxes by click-drag on canvas.
• Switch to Freehand Region in the manual toolbox to draw polygon regions.
• In freehand mode, click points to build a region, then double-click, right-click, or click Finish Region to close it.
• Use Highlight On/Off to control shape fill visibility while editing.
• Save boxes and move next with:
  • Space or Enter -> save and next
  • Esc -> exit manual mode
  • Ctrl+Z -> undo last box
  • Backspace -> undo last freehand point
  • Delete -> delete last box
  • 1..9 -> switch class index

5. Monitor active learning and training

• Left panel shows:
  • entropy of current image
  • queue size
  • training progress/status.
• Use Force Retrain if you want to bypass normal policy checks (still requires minimum samples).

6. Version and promote

• Create Version creates a dataset snapshot in src/datasets/v_YYYYMMDD_HHMMSS/.
• List Versions shows stored versions and metadata.
• Promote Shadow promotes trained candidate model to active model.

Output files and artifacts

For each selected image folder:

• labels.json or labels_coco.json (selected format)
• .labels_internal.json (internal metadata store)
• labels_autosave.json (session recovery)

Notes

• If Ray is unavailable, labeling still works; background training features are reduced.
• If class mapping is not available yet, trainer creation waits until first labels are saved.
• Restart app after model promotion for a clean reload of active weights.

Contributing

Please read CONTRIBUTING.md.

License

MIT. See LICENSE.