Implement Custom Vision Models AI-102 Practice Questions (Page 6)

Questions 26–30

1. 26expert · hard

   A developer is using the Custom Vision Python SDK to train a model. They want to evaluate the model's performance after training. What should they do?

   • AManually test the model with a few images and guess performance
   • BCall the get_iteration_performance method on the training client
   • CUse the Computer Vision SDK to evaluate the model
   • DCall the predict method on the prediction client
   Reveal Solution Discussion Select an answer first
2. 27application · medium

   A team is preparing a dataset to train a custom vision model that distinguishes between healthy and diseased plant leaves. They have 2,000 images. What labeling practice is most important for model accuracy?

   • AEnsure each image is labeled with the correct class and that ambiguous images are reviewed
   • BUse a single label for all images to simplify training
   • CLabel images automatically without human review
   • DLabel only the images that are easiest to classify
   Reveal Solution Discussion Select an answer first
4. 28expert · hard

   A company has a custom vision model that classifies product images into 'approved' and 'rejected'. The model is published as a default iteration. The quality team wants to test a new iteration in production with a small percentage of traffic before fully rolling it out. The application currently calls the prediction endpoint without specifying an iteration name. What is the best way to conduct this A/B test?

   • ADelete the old iteration so the app automatically uses the new one.
   • BChange the default iteration to the new iteration and monitor the results.
   • CPublish the new iteration and update the application to randomly route a small percentage of requests to the new iteration's endpoint.
   • DRetrain the model until the new iteration becomes the default.
   Reveal Solution Discussion Select an answer first
5. 29application · medium

   A manufacturing company wants to automate quality inspection on an assembly line. Each product image may contain multiple defects (scratch, dent, discoloration) at different locations. The system must report which defects are present and where they are located. The team has 5,000 labeled images. What should they build?

   • AAn image classification model with one class per defect type
   • BA multi-label image classification model
   • CAn object detection model with bounding boxes around each defect
   • DA semantic segmentation model with pixel-level masks
   Reveal Solution Discussion Select an answer first
6. 30expert · hard

   A team is training a custom vision model to classify images of different types of vehicles. They have a dataset with 10,000 images per class. After training, the evaluation shows high accuracy on the training set but low accuracy on the test set. What is the most likely cause, and what is the best action?

   • AThe model is not trained long enough; increase the number of training iterations.
   • BThe model is underfitting; add more training data.
   • CThe model is overfitting; add regularization or use a simpler model architecture.
   • DThe test set is too small; use a larger test set.
   Reveal Solution Discussion Select an answer first