Image Classification Evaluation

[abhishekpratapa]

• added support for image classification evaluation in TuriCreate

[kanitw]

Why is the content in the percent empty?

[abhishekpratapa]

So we don't actually have a percent requirement yet in the Confusions Evaluation Table, so I left it unimplemented.

[kanitw]

It looks like this function can be simplified into:

const {sort_by_confusions, sort_direction_confusions} = this.props;
if (sort_by_confusions) {
  if (sort_direction_confusions) {
	return a[sort_by_confusions] < b[sort_by_confusions] ? -1 : 1;
  } else {
    return a[sort_by_confusions] > b[sort_by_confusions] ? -1 : 1;
  }
}

I also wonder if the currently code handles ties consistently. Shouldn't equal return zero?

[abhishekpratapa]

Yaaa, it should prob return zero, however, I'm not sure if it changes the sort that much. Maybe it does?

[znation]

@abhishekpratapa It looks like you're getting a test failure:

self = <turicreate.test.test_logistic_classifier.LogisticRegressionClassifierModelTest_binary_classification_integer_target testMethod=test__list_fields>
msg = "Items in the first set but not the second:\n'training_predictions'"

However, I think the test should simply be updated to accommodate the new field. Seems that training_predictions is added intentionally in this PR, and shouldn't break existing code. What we should consider is:

• What will a serialized model, saved prior to this PR, return when this field is called?
• What will a serialized model, saved after this PR, return when this field is called? (i.e., are we persisting training predictions as part of the model artifact?) If we are persisting to disk, does this massively inflate the model size?
• In either case, if we don't get valid training predictions, is our new evaluation UX broken without it? What should the UX be in that case?