ortops.optim#

optimize#

Replaces an operator by another one in the same domain or another one.

Parameters:

onx – proto to modify
op_type – operator to look for
op_domain – domain to look for
new_op_type – new operator name or None for the same name
new_op_domain – new domain name or None the for the same domain
new_opset – new opset for the new domain, if not specified, it is 1 for any opset other than “”
kwargs – modified parameters, set it to None to remove them

Returns:

same type as the input

The function is not recursive yet.

onnx_extended.ortops.optim.optimize.optimize_model(onx: ModelProto, feeds: Dict[str, ndarray], transform: Callable[[ModelProto], ModelProto], session: Callable[[ModelProto], Any], params: Dict[str, List[Any]], baseline: Callable[[ModelProto], Any] | None = None, verbose: bool = False, number: int = 10, repeat: int = 10, warmup: int = 5, n_tries: int = 2, sleep: float = 0.1) → List[Dict[str, str | float]][source]#

Optimizes a model by trying out many possibilities.

Parameters:

onx – ModelProto
feeds – inputs as a dictionary of numpy arrays
transform – function taking a ModelProto and returning a ModelProto based on the values coming from params
session – function which takes a modifed ModelProto and return a session
params – dictionary of values to test { param_name: [ param_values ] }
baseline – function which takes a modifed ModelProto and return a session, identified as the baseline
verbose – use tqdm to show improvment
number – parameter to measure_time
repeat – parameter to measure_time
warmup – parameter to measure_time
n_tries – number of times to measure, if the measurements returns very different results, values for number or repeat should be increased
sleep – time to sleep between two measurements

Returns:

list of results returned by measure_time

See example TreeEnsemble optimization for an example.

onnx_extended.ortops.optim.cpu.get_ort_ext_libs() → List[str][source]#: Returns the list of libraries implementing new simple onnxruntime kernels implemented for the CPUExecutionProvider.

List of implemented kernels

<<<

from onnx_extended.ortops.optim.cpu import documentation

print("\n".join(documentation()))

>>>

It does the sum of two tensors.

Provider

CPUExecutionProvider

Inputs

Outputs

Constraints