Dala.ML.CoreML (dala v0.0.10)

CoreML integration for iOS.

This module provides an Elixir API for using Apple's CoreML framework on iOS devices. CoreML is optimized by Apple and can use the Neural Engine for hardware-accelerated ML inference.

Prerequisites

iOS device or simulator (CoreML is iOS-only)
CoreML model file (.mlmodel or .mlpackage)

Converting models

You can convert models from other formats to CoreML:

Axon → ONNX → CoreML:

# Train with Axon
model = Axon.input("input", shape: {nil, 784}) |> Axon.dense(10, activation: :softmax)
{init_fn, predict_fn} = Axon.build(model)
params = init_fn.(Nx.template({1, 784}, :f32), %{})

# Export to ONNX (requires ortex or onnx package)
# Then convert ONNX to CoreML using Apple's coremltools (Python)

Use pre-trained CoreML models from Apple or third parties.

Usage

# Load a model
:ok = Dala.ML.CoreML.load_model("/path/to/model.mlmodel", "my_model")

# Check if loaded
true = Dala.ML.CoreML.loaded?("my_model")

# Make prediction
{:ok, result_json} = Dala.ML.CoreML.predict("my_model", %{
  "input": [1.0, 2.0, 3.0]
})

# Unload when done
:ok = Dala.ML.CoreML.unload_model("my_model")

Summary

Functions

load_model(model_path, identifier)

Loads a CoreML model from the given path.

loaded?(identifier)

Checks if a model is loaded.

loaded_models()

Lists all loaded model identifiers.

predict(identifier, inputs)

Makes a prediction using a loaded model.

predict_with_loaded_model(identifier, inputs)

Convenience function to load and predict in one call.

unload_model(identifier)

Unloads a previously loaded model.

Functions

load_model(model_path, identifier)

Loads a CoreML model from the given path.

Parameters

model_path: Path to the .mlmodel or .mlpackage file
identifier: A unique identifier for this model (used in later calls)

Returns

:ok if successful
{:error, reason} if failed

loaded?(identifier)

Checks if a model is loaded.

Parameters

identifier: The model identifier

Returns

true if the model is loaded, false otherwise. Returns false on non-iOS platforms.

loaded_models()

Lists all loaded model identifiers.

Returns

A list of model identifiers, or :none on non-iOS platforms.

predict(identifier, inputs)

Makes a prediction using a loaded model.

Parameters

identifier: The model identifier
inputs: A map of input names to values

Input values can be:

Numbers (floats/integers)
Strings
Lists (converted to MLMultiArray)
Base64-encoded data (for large multi-array inputs)

Returns

{:ok, result_json} where result_json is a JSON string of the outputs. {:error, reason} if the prediction fails. :not_supported on non-iOS platforms.

Example

inputs = %{
  "input1" => 1.0,
  "input2" => [1.0, 2.0, 3.0]
}

case Dala.ML.CoreML.predict("my_model", inputs) do
  {:ok, json} ->
    result = Jason.decode!(json)
    # Use result...
  {:error, reason} ->
    # Handle error...
  :not_supported ->
    # CoreML not available
end

predict_with_loaded_model(identifier, inputs)

Convenience function to load and predict in one call.

Note: Model must be loaded first using load_model/2.

unload_model(identifier)

Unloads a previously loaded model.

Parameters

identifier: The identifier used when loading the model