Possibility to add implementations for unknown layers

[mats-claassen]

Hi, what do you think of the idea of having a parameter in the convert function were the user can pass a function that handles potential unsupported layers?

What I have in mind is something like torch2coreml has with the unknown_layer_converter_fn parameter.

[aseemw]

Just to understand it correctly, the "unknown_layer_converter_fn" allows to insert custom layers in the CoreML graph, right?

Yes, a flag along the lines of the "add_custom_layers" Flag, as in the coremltools Keras converter, would be added.

[mats-claassen]

Actually I think there are two types of unknown layers. First, the one you mention which tackles those layers that CoreML does not support so that they have to be implemented in Swift. The other is for those layers that tf-coreml doesn't support, i.e. does not convert to CoreML. I think both cases would be helpful.

[aseemw]

PR #179 adds the ability to add custom layer proto messages during conversion.