CODEWITHSUNDEEP
swiftmachine-learningcoreml
Cyril
Cyril

Reputation: 2818

Prepping to update a Core ML Model

Core ML 3 now gives us the ability to perform on-device training. After creating an updatable Core ML Model, we need to perform some function to update it on-device by using the MLUpdateTask function which requires 3 parameters: Model URL, MLBatchProvider and MLModelConfiguration.

Since Core ML 3 was just released, documentation for it is very limited; specifically on how to prepare the training data or an MLBatchProvider

Question: How do you prepare training data or create an MLBatchprovider.

Upvotes: 0

Views: 757

Answers (2)

Seb
Seb

Reputation: 1686

If your Model is Named TestModel a TestModelTrainingInput class should be available.

let singleTrainingData = try TestModelTrainingInput(input: .init[1,2,3], output_true: .init([4,5,6]))
let trainingData = MLArrayBatchProvider(array: [singleTrainingData])

Upvotes: 1

Matthijs Hollemans
Matthijs Hollemans

Reputation: 7892

To provide data to Core ML, you create an MLFeatureProvider object. This returns one or more MLFeatureValue objects, one for each input in your model. Normally the auto-generated class does this behind the scenes.

If you want to work with a batch, you create an MLBatchProvider that has multiple MLFeatureProviders, one for each example.

Making an MLBatchProvider for predictions isn't so hard: just put your MLFeatureProviders in an array and then use MLArrayBatchProvider. Again, the auto-generated class has a helper method for this.

For training you probably want to load the data on-the-fly, do random augmentations, and so on. In that case, you'll want to make a new class that adopts the MLBatchProvider protocol. It should return an MLFeatureProvider for each example. This time the MLFeatureProvider does not just have the MLFeatureValue for the example but also an MLFeatureValue for the target / true label. (The auto-generated class has a helper class for this training feature provider, but not for the training batch provider.)

I haven't actually gotten any of the new training APIs to work yet (they crash on beta 2), so I'm not 100% sure yet how the MLBatchProvider would cycle through an entire epoch of training examples.

Upvotes: 0

Related Questions