During conversion, Core ML Tools optimizes the model by applying graph transformations, called graph passes, which simplify and canonicalize the representation for a more efficient execution by the Core ML runtime.
For example, dead_code_elimination eliminates unused ops whose outputs do not contribute to final outputs. The const_elimination pass fuses ops with multiple constant inputs into a single constant, saving compute at runtime. The fuse_elementwise_to_batchnorm pass detects combinations of multiplication and add ops, and fuses them to a single
pass_pipeline parameter in
ct.convert, you can control which graph passes to run and the order of the graph passes. You can also specify options for each pass. For an overview, see the MIL Graph Pass Guide in the repo (
For a description of each graph pass, see MIL Graph Passes in the API Reference.