About

This repository contains the source for converting a Tensorflow 2.x SavedModel format into an .mlmodel format supported by CoreML in iOS 13.

Convert a classifier

Prepare the .class_labels file

In order for the model to be considered a classifier, it has to have class labels assigned to the neural network outputs.

The class labels are stored separately from the model in a .class_labels file, where each class label is defined on a new line.

For example, if the model has 3 outputs red, green, blue, the .class_labels file should contain:

red
green
blue

Place the model and class labels in the models directory

The models directory contains the models and class labels that are used for conversion. Note that the SavedModel folder name and the .class_labels file name should match.

For example, for a model named DetectColor, the directory name should be DetectColor and the .class_labels file should be named DetectColor.class_labels.

Convert the model

1. Run the Docker container using make model_name=DetectColor convert where model_name is the name of the model directory.
2. The docker container will be built and run the convert.py script.

The output of that should be out/DetectColor.mlmodel.

License

MIT