Set Validation Labels
Training vs Validation Labels
By default, all labels you draw are training labels. These are used to train the model. However, it's important to also set aside some validation labels to monitor model performance during training.
Training Labels: - Used to teach the model patterns and features - The model learns from these labels - Updated during each training iteration
Validation Labels: - Used to evaluate model performance during training - NOT used for learning - kept separate to test generalization - Helps detect overfitting and monitor training progress
Important: Not all frameworks use validation labels. Check your framework documentation or ask on image.sc to see if validation is supported.
How to Set Validation Labels
1️⃣ Select the Label box layer in the layer list
2️⃣ Click the selection tool (hollow arrow) in Napari's left toolbar
3️⃣ Select the label box ROIs that you want to designate as validation
4️⃣ Press the v key to toggle between training and validation
The text on the label box will change to indicate its type (training or validation).
Outputs
In the json file associated with the label you should now see a "name -> value' pair with name "split" and value either "train" or "validation". If you use the labels downstream of Napari-easy-augment-batch-dl then use this name value pair when partitioning into train/validation sets.
Training and validation patches
If you set some labels as validation, the patches generated from those labels will be partitioned into a separate directory (with _validation partitioned on the directory name). See screenshot below...
Notes
Not all frameworks support validation labels yet.
Best Practices
- Reserve 10-20% of your labeled data for validation
- Ensure validation data represents the variety in your full dataset
- Validation labels should be independent from training labels
- Toggle validation labels before running augmentation and training
🔄 Next: Configure Augmentations