Use the Google Cloud console to check your model performance. Analyze testerrors to iteratively improve model quality by fixing data issues.

This tutorial has several pages:

1. Set up your project and environment.

2. Create an image classification dataset, and import images.

3. Train an AutoML image classification model.

4. Evaluate and analyze model performance.

5. Deploy a model to an endpoint, and send a prediction.

6. Clean up your project.

Each page assumes that you have already performed the instructions from theprevious pages of the tutorial.

1. Understand AutoML model evaluation results

After training is completed, your model is automatically evaluated against thetest data split. The corresponding evaluation results are presented by clickingthe model's name from either theModel Registry page or theDatasetpage.

From there, you can find the metrics to measure the model's performance.

You can find a more detailed introduction to different evaluation metrics in theEvaluate, test, and deploy your model section.

2. Analyze test results

If you want to continue improving the model performance, the first step is oftento examine the error cases and investigate the potential causes. Theevaluation page of each class presents detailed test images of the givenclass categorized as false negatives, false positives, and true positives. Thedefinition of each category can be found in theEvaluate, test, and deploy your model section.

For each image under every category, you can further check the predictiondetails by clicking the image and access the detailed analysis results. You willsee theReview similar images panel on the right side of the page, where theclosest samples from the training set are presented with distances measured inthe feature space.

There are two types of data issues that you might want to pay attention:

1. Label inconsistency. If a visually similar sample from the training set hasdifferent labels from the test sample, it's possible that one of them isincorrect, or that the subtle difference requires more data for the model tolearn from,or that the current class labels are simply not accurate enough to describethe given sample.Reviewing similar images can help you get the label information accurate byeither correcting the error cases or excluding the problematic sample fromthe test set. You can conveniently change the label of either the test imageor training images on theReview similar images panel on the same page.

2. Outliers. If a test sample is marked as an outlier, it's possible that thereare no visually similar samples in the training set to help train the model.Reviewing similar images from the training set can help you identify thesesamples and add similar images into the training set to further improve themodel performance on these cases.

What's next

If you're happy with the model performance, follow thenext page of this tutorial to deploy your trainedAutoML model to an endpoint and send an image to the model for prediction.Otherwise, if you make any corrections on the data, train a new model using theTraining an AutoML image classification modeltutorial.