Returns a specification to start this module under a supervisor.

See Supervisor.

@spec predict(
  pid(),
  binary()
  | %StbImage{data: term(), shape: term(), type: term()}
  | %Nx.Tensor{
      data: term(),
      names: term(),
      shape: term(),
      type: term(),
      vectorized_axes: term()
    },
  Keyword.t()
) :: term()

Run the model against an image.

Options

• :timeout - how long to wait for the answer, in milliseconds, or :infinity. Defaults to 30000. Raise it for a large model or a slow board; inference is local and bounded, so waiting is the right answer more often than giving up.

@spec set_label(pid(), String.t() | [String.t()]) :: :ok

Give the classifier its labels, either as a list or as the path to a file holding one label per line.

Results carry an index until this is set; afterwards they carry the label.

@spec set_label_from_associated_file(pid(), String.t()) :: :ok | {:error, String.t()}

Give the classifier its labels from a file the model itself carries.

TFLiteElixir.FlatBufferModel.list_associated_files/1 says what a model has.

@spec start(any(), any()) :: :ignore | {:error, any()} | {:ok, pid()}

Start a classifier for model, a path to a .tflite file or its contents.

Options, all optional: :top_k (1) how many results predict/3 returns, :threshold (0.0) the score below which a result is dropped, :mean (128.0) and :std (128.0) the input normalisation, :jobs (System.schedulers_online/0) the interpreter's thread count, :use_tpu (false) and :tpu ("") to run on a named Edge TPU.