Common Pitfalls for iOS Developers Creating CoreML Models

Hello, I'm noppe.
I've been trying to create CoreML models, but things haven't been going smoothly, so I'll share some challenges I faced while getting about halfway through the process.

Don't Create Models as Keras Subclasses

To use coremltools for converting models to CoreML, the Keras structure needs to be a Sequential model or a Functional model.
Therefore, if you build a model using a Keras subclass, you won't be able to convert it to a CoreML model even after training it.

Specify the Filename When You Want to Create an mlpackage

It's not very well-known, but if you include .mlpackage in the filename, it will automatically save it as an mlpackage.

ml_model = ct.convert(model)
ml_model.save("MLModel.mlpackage")

Explicitly Specify ImageType When Input is an Image

If you don't specify ImageType, you'll have to deal with an array called MLMultiArray directly from Swift.

ct.convert(
    model,
    inputs=[
        ct.ImageType()
    ],
)

Name Your Layers

By naming your layers, you can determine the property names used when calling them from Swift.
In particular, layer names like dense_1_input, dense_2_input, etc., change every time you run it, so it's better to define them explicitly.

layer1 = tf.keras.Input(name="input", shape=(82,))
layer2 = tf.keras.layers.Embedding(input_dim=100, output_dim=10, input_length=82)(layer1)
layer3 = tf.keras.layers.LSTM(name="output", units=200, return_sequences=False)(layer2)

Note that if you don't use Input, the name will be the first layer's name plus _input.
I generally recommend adding Input at the beginning.

I'll add more when I think of them.