Package 'onnxr'

Find the ONNX Runtime shared library

Description

Searches for the ONNX Runtime shared library in standard locations: the ORT_ROOT environment variable, common system library paths, the per-user install from onnx_install(), the Python onnxruntime package, and pkg-config.

Usage

onnx_find_lib()

Value

Full path to the shared library, or NULL if not found.

Examples

onnx_find_lib()

---

Install ONNX Runtime

Description

Downloads pre-built ONNX Runtime binaries for the current platform and installs them to a per-user data directory. The library is loaded immediately after installation, so there is no need to restart R.

Usage

onnx_install(cuda = NULL)

Arguments

cuda

Whether to install the CUDA-enabled build for GPU acceleration. NULL (default): auto-detect by checking for nvidia-smi on the system PATH. Installs the CUDA build if a GPU is found and the platform is supported (Linux x64 or Windows x64), otherwise falls back to the CPU build. TRUE: force the CUDA build. Errors if the platform is not Linux x64 or Windows x64. FALSE: always install the CPU-only build.

Value

Invisibly, the path to the installation directory.

---

Check whether ONNX Runtime is available

Description

Check whether ONNX Runtime is available

Usage

onnx_is_installed()

Value

TRUE if the ONNX Runtime shared library can be found in any of the standard search locations, FALSE otherwise.

Examples

onnx_is_installed()

---

Check whether ONNX Runtime is loaded

Description

Check whether ONNX Runtime is loaded

Value

TRUE if the ONNX Runtime shared library has been loaded in the current R session, FALSE otherwise.

Examples

onnx_is_loaded()

---

Load an ONNX model

Description

Loads an .onnx model file and creates a model object.

Usage

onnx_model(
  path,
  backend = c("cpu", "coreml", "cuda", "xnnpack", "openvino"),
  cache_dir = tools::R_user_dir("onnxr", "cache"),
  threads = 1L,
  opt_level = 99L
)

Arguments

path	Path to an .onnx model file.
backend	Execution backend. Available options depend on the platform and ORT build: "cpu" — Default, available everywhere. "coreml" — Apple Neural Engine + CPU (macOS/iOS only). "cuda" — NVIDIA GPU (Linux x64 and Windows x64 only). Requires CUDA toolkit and the CUDA-enabled ORT build from onnx_install(cuda = TRUE). "xnnpack" — Optimized CPU kernels (mobile/embedded). Requires an ORT build with XNNPACK support (not provided by onnx_install()). "openvino" — Intel hardware acceleration. Requires OpenVINO installation and ORT build with OpenVINO EP (not provided by onnx_install()).
cache_dir	Optional directory for CoreML model cache. Set to NULL to disable caching.
threads	Number of threads. 0 uses all available; a positive integer sets a fixed thread count.
opt_level	Graph optimization level. 99 (default) enables all optimizations; 1 for basic only; 0 to disable.

Value

An "onnx_model" object (a named list) with model metadata and internal pointers used by onnx_run().

Examples

model_path <- system.file("extdata", "lm_iris.onnx", package = "onnxr")
if (onnx_is_loaded() && nzchar(model_path)) {
    sess <- onnx_model(model_path)
    sess
}

---

Run or predict from an ONNX model

Description

Passes input through the model and returns the results. For models with a single output, returns the output array directly. For models with multiple outputs, returns a named list of arrays.

Usage

onnx_run(model, ..., simplify = FALSE)

Arguments

model	An "onnx_model" object created by onnx_model().
...	Input arrays, either as unnamed arguments (matched to model inputs by position) or as named arguments (matched by name). Each input must be a numeric or integer matrix/array with dimensions matching the model's expected input shape. For single-input models, a single array can be passed directly.
simplify	If TRUE, return the output array directly for single-output models instead of a length-1 named list.

Details

Handles conversion between R's column-major arrays and ONNX's row-major tensors, and between R's numeric types and the model's declared element types (float, double, int32, int64, bool). Note that int64 outputs are cast to doubles, which may lose precision for large integers.

Value

A named list of output arrays, or (if simplify = TRUE and the model has a single output) the output array directly.

Examples

model_path <- system.file("extdata", "lm_iris.onnx", package = "onnxr")
if (onnx_is_loaded() && nzchar(model_path)) {
    sess <- onnx_model(model_path)
    input <- as.matrix(iris[1:5, c("Sepal.Length", "Sepal.Width", "Petal.Length")])
    onnx_run(sess, input)
}

---