Ortex is a wrapper around ONNX Runtime (implemented as bindings to ort). Ortex leverages Nx.Serving to easily deploy ONNX models that run concurrently and distributed in a cluster. Ortex also provides a storage-only tensor implementation for ease of use.

ONNX models are a standard machine learning model format that can be exported from most ML libraries like PyTorch and TensorFlow. Ortex allows for easy loading and fast inference of ONNX models using different backends available to ONNX Runtime such as CUDA, TensorRT, Core ML, and ARM Compute Library.

TL;DR:

iex> model = Ortex.load("./models/resnet50.onnx")
#Ortex.Model<
  inputs: [{"input", "Float32", [nil, 3, 224, 224]}]
  outputs: [{"output", "Float32", [nil, 1000]}]>
iex> {output} = Ortex.run(model, Nx.broadcast(0.0, {1, 3, 224, 224}))
iex> output |> Nx.backend_transfer() |> Nx.argmax
#Nx.Tensor<
  s64
  499
>

Inspecting a model shows the expected inputs, outputs, data types, and shapes. Axes with nil represent a dynamic size.

To see more real world examples see the examples folder.

Ortex also implements Nx.Serving behaviour. To use it in your application's supervision tree consult the Nx.Serving docs.

iex> serving = Nx.Serving.new(Ortex.Serving, model)
iex> batch = Nx.Batch.stack([{Nx.broadcast(0.0, {3, 224, 224})}])
iex> {result} = Nx.Serving.run(serving, batch)
iex> result |> Nx.backend_transfer() |> Nx.argmax(axis: 1)
#Nx.Tensor<
  s64[1]
  [499]
>

Ortex can be installed by adding ortex to your list of dependencies in mix.exs:

def deps do
  [
    {:ortex, "~> 0.1.10"}
  ]
end

You will need Rust for compilation to succeed.

Builds that link ONNX Runtime dynamically, which includes the cuda and tensorrt features, fail to compile with Rust 1.100 and later because of the ort version Ortex currently uses. Until ort is upgraded, use Rust 1.99 for these builds, for example by running rustup override set 1.99.1 in your project directory.