SignatureRunner class

A runner for a specific model signature, enabling named tensor access.

Signatures allow models to expose multiple named entry points. This is the foundation for on-device training, where a model typically exposes:

• train: runs one training step (forward pass + backprop + optimizer)
• infer: runs inference / prediction
• get_weights: reads current model weights into Dart buffers
• set_weights: writes weight values from Dart buffers into the model

On-device training example

// Load a model that was exported with training signatures.
final interpreter = await Interpreter.fromAsset('trainable_model.tflite');

// Optionally inspect available signatures:
print(interpreter.signatureKeys); // ['train', 'infer', 'get_weights', 'set_weights']

// --- Training loop ---
final trainRunner = interpreter.getSignatureRunner('train');

for (int epoch = 0; epoch < 10; epoch++) {
  final lossBuffer = Float32List(1);
  trainRunner.run(
    {'x': imageData, 'y': labels},
    {'loss': lossBuffer},
  );
  print('Epoch $epoch loss: ${lossBuffer[0]}');
}
trainRunner.close();

// --- Inference ---
final inferRunner = interpreter.getSignatureRunner('infer');
final predictions = Float32List(10);
inferRunner.run({'x': testImage}, {'output': predictions});
inferRunner.close();

// --- Weight persistence (get/set) ---
final getRunner = interpreter.getSignatureRunner('get_weights');
final w = [[0.0]];
final b = [0.0];
getRunner.run({}, {'w': w, 'b': b});
getRunner.close();

// Restore weights into a fresh interpreter:
final setRunner = freshInterpreter.getSignatureRunner('set_weights');
setRunner.run({'w': w, 'b': b}, {});
setRunner.close();

interpreter.close();

Building a training-capable model (Python)

Export a LiteRT model with train, infer, get_weights, and set_weights signatures and convert with resource variables enabled:

converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
converter.experimental_enable_resource_variables = True
tflite_model = converter.convert()