collectOutputShapes function

Map<int, List<int>> collectOutputShapes(Interpreter itp)

Collects the shape of every output tensor of itp, keyed by output index.

Model wrappers commonly need output shapes at initialization to work out which output carries which result (by element count, trailing dimension, and so on) and to preallocate result buffers. This walks output indices from 0 until getOutputTensor throws, so it works on interpreters that do not expose a reliable output count.

Only shapes are read. Deliberately does not touch Tensor.data, so no buffer views are materialized and quantized outputs are safe to enumerate. Use TensorFloat32Views when you need the tensor buffers themselves.

Interpreter.allocateTensors should have been called first, otherwise the reported shapes may not reflect the allocated tensors.

Implementation

Map<int, List<int>> collectOutputShapes(Interpreter itp) {
  final Map<int, List<int>> shapes = <int, List<int>>{};
  for (int i = 0; ; i++) {
    try {
      shapes[i] = itp.getOutputTensor(i).shape;
    } catch (_) {
      break;
    }
  }
  return shapes;
}