Labelbox lets you slice and dice your model predictions, annotations, and metrics. Powerful filtering and sorting make it simpler and more efficient to:
- Analyze the performance of neural networks
- Find low-performing slices of data
- Surface labeling mistakes
- Identify high-impact data to label in order to maximally improve neural network performance
A model run contains a snapshot of annotations, predictions, and metadata
The annotations, predictions and metadata in a model run are versioned. Even if you update these annotations, predictions, and metadata elsewhere in the Labelbox platform, they will remain intact in the model run.
Below are the attributes you can search and filter by in the Model product.
|Find data rows with labels that contain or do not have certain counts of annotations
|Show image assets where X annotation was used at least N times.
|Predictions uploaded to the model run
|Show text assets where the model predicted X at least N times.
|Filter on any data row identifiers (e.g., global key, data row ID, and external ID)
|Show all image assets where the global key does not contain X.
|Find data rows by the ID of associated labels.
|Show all text assets where a label ID contains X.
|Find data rows that belong to a particular dataset.
|Show all image assets uploaded to dataset X.
|Find data rows that are associated with a specific labeling project
|Show all text assets uploaded to project X.
|Model run metrics (auto-generated and custom)
|Show image assets with specific values for precision, recall, intersection over union, etc
|Find data rows that contain a certain metadata field and/or value
|Show text assets that were captured between two datetimes
You can think of creating a filter like constructing a pyramid with layers of logical sequence. Each layer is an AND operation. Within a layer, you can use OR operations. Each filter provides a count of data rows that match the filter. Only non-zero counts of instances of an attribute are available for selection.
Here is a realistic example to help you understand filter construction: An ML engineer is developing an AI model to identify people on an object detection dataset. The engineer learns that the model performs poorly on images containing coastlines. So the engineer queries for images that are at least 200px wide AND belong to the dataset named "SAR dataset (chipped)".
To do so, click on Sort by, then pick the desired metric and decide if the sorting should be in ascending order (Asc) or descending order (Desc).
Updated 3 months ago