Load an Ortex.Model from disk. Optionally pass the execution providers as a list of descending priority and graph optimization level 1-3. Any graph optimization level beyond the range of 1-3 will disable graph optimization.

By default, Ortex only includes some of the supported execution providers of ONNX Runtime. To enable others, first ensure you have downloaded or compiled a version of libonnxruntime that includes them, then set the environment variable ORT_LIB_LOCATION to its location. Then add config :ortex, Ortex.Native, features: [EXECUTION_PROVIDERS] to your config.exs where EXECUTION_PROVIDERS is a list of strings of which execution providers to enable.

examples

Examples

iex> Ortex.load("./models/tinymodel.onnx")
iex> Ortex.load("./models/tinymodel.onnx", [:cuda, :cpu])
iex> Ortex.load("./models/tinymodel.onnx", [:cpu], 0)

Run a forward pass through a model.

This takes a model and tuple of Nx.Tensors, optionally transfers them to the Ortex.Backend if they aren't there already, and runs a forward pass through the model. This will return a tuple of Ortex.Backend tensors, it's up to the user to transfer these back to another backend if additional ops are required.

If there is only one input you can optionally pass a bare tensor rather than a tuple.

examples

Examples

iex> model = Ortex.load("./models/tinymodel.onnx")
iex> {%Nx.Tensor{shape: {1, 10}},
...>  %Nx.Tensor{shape: {1, 10}},
...>  %Nx.Tensor{shape: {1, 10}}} = Ortex.run(
...>    model, {
...>      Nx.broadcast(0, {1, 100}) |> Nx.as_type(:s32),
...>      Nx.broadcast(0, {1, 100}) |> Nx.as_type(:f32)
...>    })