View Source TFLiteElixir.Interpreter.Server (tflite_elixir v1.0.0-rc4)
An interpreter that lives inside a process, so that feeding it, running it and reading the result back is one step that nothing can interleave with.
The direct API is not wrong -- it mirrors TfLite's C API faithfully -- but
nothing in it says that input_tensor/3, invoke/1 and output_tensor/2
have to be treated as one operation. Two processes taking turns badly get each
other's answers: measured on a real model, 147 wrong results in 400 calls,
silently and without a crash.
{:ok, server} = TFLiteElixir.Interpreter.Server.start_link(model_path)
output = TFLiteElixir.Interpreter.Server.predict(server, [input])Concurrent callers are serialised by the process rather than racing inside the
interpreter, so each gets the answer to its own input. TFLiteElixir.Interpreter
stays exactly as it is for callers who would rather serialise access
themselves.
Summary
Functions
Feed, run and read back, as one operation.
Run a function against the interpreter inside the owning process.
Start an interpreter process outside a supervision tree.
Start an interpreter process for a model file, linked to the caller.
Stop the process, and with it the interpreter.
Functions
Feed, run and read back, as one operation.
@spec run(pid(), (reference() -> result), timeout()) :: result | {:error, String.t()} when result: term()
Run a function against the interpreter inside the owning process.
For the sequences predict/3 does not cover -- resizing an input and
reallocating, say, or driving a signature runner. The function runs in the
server process, so nothing else touches the interpreter while it does, and it
should return promptly for the same reason.
TFLiteElixir.Interpreter.Server.run(server, fn interpreter ->
TFLiteElixir.Interpreter.tensors_size(interpreter)
end)Start an interpreter process outside a supervision tree.
Start an interpreter process for a model file, linked to the caller.
Options
:num_threads. Passed to the builder before the interpreter is built, so it reaches the default XNNPACK delegate as well.
@spec stop(pid()) :: :ok
Stop the process, and with it the interpreter.