IssueTracker

This will create a feature request which anybody can view and comment on. Please describe your requested enhancement. Good feature requests will solve common problems or enable new use cases.

What you would like to accomplish:

• Customer suggests that there are two properties that should be added to the converted models exported by AutoML Vision as far is image classification models are concerned:

1. The input type of the model should be ImageType instead of MLMultiArrayType. The CoreML documentation speaks to the performance benefits here [1]. In addition to improved performance, usage of ImageType allows iOS/macOS developers to use Apple's Vision API which makes it much easier to run the inferences on CoreML models.

2. The model should be configured as an image classifier, and the labels should be embedded in the classifier configuration. This allows developers to more easily use the inference results, as the classification scores will be directly associated with their respective labels without need of lookup in an accompanying dictionary file.

• [1] https://coremltools.readme.io/docs/image-inputs

How this might work:

• Here is an example code snippet showing how, if developers could access model in Keras SavedModel format, then customer would be able to achieve this behavior.

import coremltools as ct
import tensorflow as tf
labels = 'path/to/dict.txt'
keras_model = tf.keras.models.load_model('path/to/SavedModel')
optimized_coreml_model = ct.convert(keras_model, inputs=[ct.ImageType()], classifier_config=ct.ClassifierConfig(labels))

If applicable, reasons why alternative solutions are not sufficient:

• Unfortunately, there is no option to export in Keras format so this is not possible on customer's end. It's also difficult to modify the existing MLModel [2]. Because of this, customer request that the AutoML Vision product team can change the internal conversion process to include these optimizations, or alternatively provide a TF2 SavedModel export option (the SavedModel export option currently available seems to be based on TF1, and is very difficult to use to convert to MLModel) so that customer can convert it themselves.

• [2] https://github.com/apple/coremltools/issues/1143#issuecomment-805301470

Other information (workarounds you have tried, documentation consulted, etc):