6. Create Stress Test#
After you have created the desired domain, navigate to Stress Tests tab and choose your model, dataset and domain (see the image below) from the drop down menus. You can use the additional parameters menu located on the right side of the Test button to choose the best parameters for your stress tests. Once ready, click on Test.
The platform will first create a baseline on the selected pair model-dataset if it doesn’t exist, which will perform a standard model inference on the selected dataset, providing various metrics and reports at the end. After its completion or if you already have a baseline with the selected pair, the platform will run a stress test to find samples that degrade the performance of your model. When the process is finished, you will find a stress test report that covers standard metrics, as well as manually inspect the generated samples and download them.
The stress test parameters you can change influence the depth and width of the search as follows:
Samples per search- what’s the depth of each individual search. Usually a value of between 25 and 50 can be used depending on the complexity of the problem. Larger depths find issues in harder problems.
Number of searches- how many independent search attempts to perform. In cases where the model has many failure modes, a larger number of searches will be able to find an increasing subset.