Lakera releases one-line Voxel51 integration to bring full-scale model validation to FiftyOne users.

This article was originally posted on our company website. Lakera’s developer platform enables ML teams to ship fail-safe computer vision models.

We love Voxel51 and we know that many of our users do too. That’s why we’re super excited to release a one-line integration with Lakera for existing Voxel51 users.

💡 Already a Voxel51 user? Head over to our integration tutorial here.

one-line Voxel51 integration to bring full-scale model validation to FiftyOne users

This release lets people who already benefit from Voxel51’s insights add full-scale model validation to their workflow with a single line of code.

At the highest level, the workflow is as follows:p

# Step 1: Load your Voxel51 dataset and add predictions.

dataset: fo.Dataset = ...
load_predictions(dataset)

# Step 2: Run MLTest to analyze the model's performance. That's it!

runner = lakera.FiftyOneRunner(
    dataset=dataset,
    path_to_output="/tmp/voxel_results",
)
runner.run()

This is it! With Voxel51 and Lakera’s MLTest, you have full visibility of your model’s performance prior to deployment.

You can find more information on the integration here.

Voxel51 + MLTest to analyze YOLOv8

YOLOv8 is all the hype right now. But is it also more robust than the previous versions? We wanted to find out!

Using Voxel51 and MLTest we analyzed the latest YOLOv8 model on COCO and found quite a number of robustness issues and clusters on which the model performance is much lower than expected:

low performant clusters found by MLTest

We will publish our in-depth analysis of YOLOv8 soon! Stay tuned!