Contents

MLState

Handle to the state buffers.

Declaration

class MLState

Overview

A stateful model maintains a state from one prediction to another by storing the information in the state buffers. To use such a model, the client must request the model to create state buffers and get MLState object, which is the handle to those buffers. Then, at the prediction time, pass the MLState object in one of the stateful prediction functions.

// Load a stateful model
let modelAsset = try MLModelAsset(url: modelURL)
let model = try await MLModel.load(asset: modelAsset, configuration: MLModelConfiguration())

// Request a state
let state = model.newState()

// Run predictions
for _ in 0 ..< 42 {
  _ = try await model.prediction(from: inputFeatures, using: state)
}

// Access the state buffer.
state.withMultiArray(for: "accumulator") { stateMultiArray in
  ...
}

The object is a handle to the state buffers. The client shall not read or write the buffers while a prediction is in-flight.

Each stateful prediction that uses the same MLState must be serialized. Otherwise, if two such predictions run concurrently, the behavior is undefined.

Topics

Getting a state buffer

See Also

Model state