simple-napari-cci-annotator 0.0.2

Plugin to easly test a yolo model and retrain it

Simple napari annotator

Minimal napari plugin for YOLO bbox detection + quick correction + retraining.

No dataset browser, no extra workflow logic. User provides the image in napari.

UI flow

1. Path to model (.pt) or model folder
2. Load model
3. Predict
4. Optional destination folder for retrained model (browse or type)
5. Edit boxes in napari shapes layer (yolo_bboxes)
6. Add correction
7. Retrain

Starting from scratch

The model input supports either:

• A .pt model file, or
• A folder path.

Folder behavior on Load model:

• If the folder already contains one or more .pt files, the first one is loaded.
• If the folder contains no .pt file, the plugin copies the bundled yolov8n.pt into that folder and loads it.
• Empty model input is invalid and will show an error.

If you do not have a model yet, start from a pretrained YOLOv8 nano checkpoint and use it as your initial file:

• Direct download: https://github.com/ultralytics/assets/releases/latest/download/yolov8n.pt
• Model overview: https://docs.ultralytics.com/models/yolov8/

After downloading, select that yolov8n.pt file in the Model field and continue with the correction/retrain loop.

Assumptions

• Input image is already RGB 8-bit (or compatible with clipping/conversion).
• Single class (0: LABEL) for now.
• Exactly one image layer should be present when using Add correction.

Folder behavior

Given a model path like:

.../my_model/best.pt

The plugin uses .../my_model as root.

Add correction

Each click saves:

• Image to my_model/corrections/<image_name>_<timestamp>.png
• Labels to my_model/corrections/<image_name>_<timestamp>.txt
• Training config to my_model/corrections/training_config.json (created/updated)

Image layer behavior:

• If no image layer exists, Add correction shows an error.
• If more than one image layer exists, Add correction shows an error.
• If exactly one image layer exists, that image layer is used for saving correction image data.

training_config.json defaults:

• image_size: prefilled from the current image size using max(height, width)
• batch: 8
• epochs: 100
• patience: 30

You can edit this file before clicking Retrain.

Label format is YOLO detection:

class x_center y_center width height

normalized to [0, 1].

Retrain

Each click creates:

• <retrain_root>/dataset/
  • images/train, images/val
  • labels/train, labels/val
  • dataset.yaml
• Trains YOLO from those corrections
• Copies best model to:
  • <retrain_root>/best.pt
• Deletes training traces folder after extracting best.pt

<retrain_root> resolution:

• If destination field is set, retraining outputs there.
• Otherwise it defaults to my_model/retrained_<timestamp>.

So the retrained folder keeps a clean dataset + final model, without run artifacts.

Warning: Data labeled will be equally divided between traning and validation (50%).