Training is complete, and the image classifier is deployed with its learned weights fixed.
A user submits one image and receives a class label. Separately, the operations team measures accuracy on labeled images.
Which inference step computes the predicted label returned to the user?
Inference uses a trained model to compute a prediction. Training updates the weights. Evaluation compares predictions with known labels to measure performance.
Detailed explanation
Preparing the inputs does not by itself produce a predicted class label.
The trained weights are used to compute a prediction for the image. This step does not update those weights.
Using labeled images to update the weights is part of training the model.
Computing accuracy compares predictions with known labels. Obtaining the predictions used in that comparison can itself require inference.
Check official documentation
Compare the following points with the official documentation.
Read the SageMaker AI model training documentation and the image classifier's input and output specification. With one fictional image, separate the fixed model's prediction from checking whether it matches the known label.
Official references
What to confirm
If both the predicted and known labels are Dog, computing Dog is inference and counting one correct match is evaluation. The model's weights stay fixed in both steps of this example.
Key points
- Training
- Inference
- Evaluation
Notes
- Environment: Paper exercise with AWS documentation and a fictional classification result
- Computing labels for evaluation images is inference. Computing the accuracy of those labels is evaluation.
Foundation review
Prediction and scoring
A fixed classifier returns a label for each image. Evaluation can then compare that label with the known answer to calculate a metric such as accuracy.
Updating weights
Training updates the weights. The deployed model in this scenario is fixed, so no reference label is supplied to update its weights during inference.