--- title: "OnnxModel" sidebar_label: "OnnxModel" description: "The clika_runtime.io OnnxModel class." --- An ONNX model file as an editable graph: open or create one, or build one from a ModelGraph (from_graph), inspect and edit it, save it, or compile() it into a ModelGraph. ## `ir_version` (property) The ONNX IR version. ## `num_initializers` (property) The initializer (graph constant) count. ## `num_nodes` (property) The node count. ## `opset` (property) The ai.onnx opset version. ## `__init__` ```python __init__(self, /, *args, **kwargs) ``` Initialize self. See help(type(self)) for accurate signature. ## `add_initializer` ```python add_initializeradd_initializer(self, value: clika_runtime._core.Tensor) -> str ``` add_initializer(self, value: clika_runtime._core.Tensor) -> str add_initializer(value) -> str Create a graph constant from a tensor's bytes; returns the generated initializer name. ## `add_input` ```python add_inputadd_input(self, name: str, dtype: clika_runtime._core.DataType, shape: collections.abc.Sequence[int]) -> str ``` add_input(self, name: str, dtype: clika_runtime._core.DataType, shape: collections.abc.Sequence[int]) -> str add_input(name, dtype, shape) -> str Register a graph input; any dim <= 0 declares a dynamic dimension. Returns the tensor name. ## `add_node` ```python add_nodeadd_node(self, op_type: str, inputs: collections.abc.Sequence[str], num_outputs: int = 1, attrs: dict = {}, domain: str = '', output_names: collections.abc.Sequence[str] = []) -> list[str] ``` add_node(self, op_type: str, inputs: collections.abc.Sequence[str], num_outputs: int = 1, attrs: dict = {}, domain: str = '', output_names: collections.abc.Sequence[str] = []) -> list[str] add_node(op_type, inputs, num_outputs=1, attrs={}, domain='', output_names=()) -> list[str] Append a node: positional input tensor names in, generated (or caller-named) output names back. Attributes ride a dict of int / float / str / list values. A `domain` other than the standard one must be declared first with add_opset_domain(); a node under a domain the model does not import raises InvalidArgumentError and leaves the model as it was. ## `add_opset_domain` ```python add_opset_domainadd_opset_domain(self, domain: str, version: int) -> None ``` add_opset_domain(self, domain: str, version: int) -> None add_opset_domain(domain, version) -> None Declare an opset import for a non-standard domain (for example 'com.microsoft', 1), before add_node() adds a node under that domain. ## `add_output` ```python add_outputadd_output(self, name: str, dtype: clika_runtime._core.DataType, shape: collections.abc.Sequence[int]) -> str ``` add_output(self, name: str, dtype: clika_runtime._core.DataType, shape: collections.abc.Sequence[int]) -> str add_output(name, dtype, shape) -> str Register a graph output. Returns the tensor name. ## `compile` ```python compilecompile(self, options: object | None = None) -> clika_runtime._core.io.ModelGraph ``` compile(self, options: object | None = None) -> clika_runtime._core.io.ModelGraph compile(options=None) -> ModelGraph Build the executable form: every node parses into its runtime operator (an unsupported op fails here, naming it), weights bind, shapes resolve. `options` (a CompileOptions) pins input shapes and names the weights' device. Returns the ModelGraph as built: optimize() it when wanted, then finalize() it to run. The OnnxModel is unchanged and reusable. ## `create` ```python createcreate(graph_name: str, opset_version: int = 0) -> clika_runtime._core.io.OnnxModel ``` create(graph_name: str, opset_version: int = 0) -> clika_runtime._core.io.OnnxModel create(graph_name, opset_version=0) -> OnnxModel Create an empty model at the given ai.onnx opset (0 = the library default); build it up with add_input / add_initializer / add_node / add_output. ## `from_graph` ```python from_graphfrom_graph(graph: object | None, *, opset_version: int = 0, contrib_ops: bool = True, optimize: bool = True, save_as_external_data: bool | None = None, producer_name: str = '') -> clika_runtime._core.io.OnnxModel ``` from_graph(graph: object | None, *, opset_version: int = 0, contrib_ops: bool = True, optimize: bool = True, save_as_external_data: bool | None = None, producer_name: str = '') -> clika_runtime._core.io.OnnxModel from_graph(graph, *, opset_version=0, contrib_ops=True, optimize=True, save_as_external_data=None, producer_name='') -> OnnxModel Build the ONNX model of `graph`, a ModelGraph, in memory, with nothing written: the model graph.export_onnx(path, ...) would write to its file. Inspect it, edit it, optimize(), save() or compile() it like any other model. It reads the graph's weights only when it is saved or compiled, so it holds no copy of them, and it keeps them alive for as long as it lives. The options are ModelGraph.export_onnx()'s. Raises what export_onnx() raises, and TypeError for a graph that is not a ModelGraph. ## `inputs` ```python inputsinputs(self) -> list ``` inputs(self) -> list inputs() -> list[tuple[str, DataType, tuple[int, ...]]] Graph inputs as (name, dtype, dims) tuples; a dynamic dim reads as -1. ## `merge` ```python mergemerge(self, second: object | None, io_map: collections.abc.Sequence[tuple[str, str]] = []) -> None ``` merge(self, second: object | None, io_map: collections.abc.Sequence[tuple[str, str]] = []) -> None merge(second, io_map=()) -> None Combine `second`, an OnnxModel, into this model. Each io_map entry (this_output, second_input) connects one of this model's outputs to one of second's inputs; the connected outputs stop being outputs and the connected inputs stop being inputs. The combined model's inputs are this model's, then second's unconnected ones, and its outputs second's, then this model's unconnected ones. A name of second's that this model already uses is renamed, and the opset imports merge to the highest version per domain. The call consumes `second` whether it merges or refuses, and leaves it empty. The one exception is second being this model, which is refused before anything moves. Raises InvalidArgumentError when the two cannot combine (an io_map name that names no output or input, second being this model, an empty model), and TypeError for a second that is not an OnnxModel. ## `nodes` ```python nodesnodes(self) -> list[clika_runtime._core.io.OnnxNodeInfo] ``` nodes(self) -> list[clika_runtime._core.io.OnnxNodeInfo] nodes() -> list[OnnxNodeInfo] The graph's nodes in graph order, each with its name, op type, domain and the input and output tensor names; a node's name is what remove_node(name) takes. ## `open` ```python openopen(path: str) -> clika_runtime._core.io.OnnxModel ``` open(path: str) -> clika_runtime._core.io.OnnxModel open(path) -> OnnxModel Open a .onnx file (initializer payloads referenced as external data stay on disk beside it). ## `optimize` ```python optimizeoptimize(self) -> int ``` optimize(self) -> int optimize() -> int Run the graph-rewrite pipeline to a fixed point; returns the pass count. ## `outputs` ```python outputsoutputs(self) -> list ``` outputs(self) -> list outputs() -> list[tuple[str, DataType, tuple[int, ...]]] Graph outputs as (name, dtype, dims) tuples. ## `remove_node` ```python remove_noderemove_node(self, name: str) -> None ``` remove_node(self, name: str) -> None remove_node(name) -> None Remove the node named `name` (an error if absent). ## `save` ```python savesave(self, path: str) -> None ``` save(self, path: str) -> None save(path) -> None Assemble the model and write it to `path`.