microsoft/onnxruntime-extensions
Publicmirrored fromhttps://github.com/microsoft/onnxruntime-extensionsAvailable
onnxruntime_extensions/_ortapi2.py
128lines · modecode
| 1 | # Copyright (c) Microsoft Corporation. All rights reserved. |
| 2 | # Licensed under the MIT License. See License.txt in the project root for |
| 3 | # license information. |
| 4 | ############################################################################### |
| 5 | |
| 6 | import numpy as np |
| 7 | import onnxruntime as _ort |
| 8 | from ._ocos import default_opset_domain, get_library_path # noqa |
| 9 | from ._cuops import * # noqa |
| 10 | |
| 11 | |
| 12 | def get_opset_version_from_ort(): |
| 13 | _ORT_OPSET_SUPPORT_TABLE = { |
| 14 | "1.5": 11, |
| 15 | "1.6": 12, |
| 16 | "1.7": 13, |
| 17 | "1.8": 14, |
| 18 | "1.9": 15 |
| 19 | } |
| 20 | |
| 21 | ort_ver_string = '.'.join(_ort.__version__.split('.')[0:2]) |
| 22 | return _ORT_OPSET_SUPPORT_TABLE.get(ort_ver_string, 11) |
| 23 | |
| 24 | |
| 25 | def make_onnx_model(graph, opset_version=0, extra_domain=default_opset_domain(), extra_opset_version=1): |
| 26 | if opset_version == 0: |
| 27 | opset_version = get_opset_version_from_ort() |
| 28 | fn_mm = onnx.helper.make_model_gen_version if hasattr(onnx.helper, 'make_model_gen_version' |
| 29 | ) else onnx.helper.make_model |
| 30 | model = fn_mm(graph, opset_imports=[ |
| 31 | onnx.helper.make_operatorsetid('ai.onnx', opset_version)]) |
| 32 | model.opset_import.extend([onnx.helper.make_operatorsetid(extra_domain, extra_opset_version)]) |
| 33 | return model |
| 34 | |
| 35 | |
| 36 | class OrtPyFunction: |
| 37 | |
| 38 | @classmethod |
| 39 | def get_ort_session_options(cls): |
| 40 | # ONNXRuntime has an issue to support reusing the SessionOptions object. |
| 41 | # Create a new one every time here |
| 42 | so = _ort.SessionOptions() |
| 43 | so.register_custom_ops_library(get_library_path()) |
| 44 | return so |
| 45 | |
| 46 | def __init__(self): |
| 47 | self._onnx_model = None |
| 48 | self.ort_session = None |
| 49 | self.default_inputs = {} |
| 50 | |
| 51 | def create_from_customop(self, op_type, *args, **kwargs): |
| 52 | graph = SingleOpGraph.build_my_graph(op_type, *args, **kwargs) |
| 53 | self._bind(make_onnx_model(graph)) |
| 54 | return self |
| 55 | |
| 56 | def add_default_input(self, **kwargs): |
| 57 | inputs = { |
| 58 | ky_: val_ if isinstance(val_, (np.ndarray, np.generic)) else \ |
| 59 | np.asarray(list(val_), dtype=np.uint8) for ky_, val_ in kwargs.items() |
| 60 | } |
| 61 | |
| 62 | self.default_inputs.update(inputs) |
| 63 | |
| 64 | @property |
| 65 | def onnx_model(self): |
| 66 | assert self._oxml is not None, "No onnx model attached yet." |
| 67 | return self._oxml |
| 68 | |
| 69 | @property |
| 70 | def input_names(self): |
| 71 | return [vi_.name for vi_ in self.onnx_model.graph.input] |
| 72 | |
| 73 | @property |
| 74 | def output_names(self): |
| 75 | return [vi_.name for vi_ in self.onnx_model.graph.output] |
| 76 | |
| 77 | def _bind(self, oxml): |
| 78 | self.inputs = list(oxml.graph.input) |
| 79 | self.output = list(oxml.graph.output) |
| 80 | self._oxml = oxml |
| 81 | return self |
| 82 | |
| 83 | def _ensure_ort_session(self): |
| 84 | if self.ort_session is None: |
| 85 | sess = _ort.InferenceSession(self.onnx_model.SerializeToString(), self.get_ort_session_options()) |
| 86 | self.ort_session = sess |
| 87 | |
| 88 | return self.ort_session |
| 89 | |
| 90 | @classmethod |
| 91 | def from_customop(cls, op_type, *args, **kwargs): |
| 92 | return cls().create_from_customop(op_type, *args, **kwargs) |
| 93 | |
| 94 | @classmethod |
| 95 | def from_model(cls, path_or_model, *args, **kwargs): |
| 96 | return cls()._bind(onnx.load_model(path_or_model) if isinstance(path_or_model, str) else path_or_model) |
| 97 | |
| 98 | def _argument_map(self, *args, **kwargs): |
| 99 | idx = 0 |
| 100 | feed = {} |
| 101 | for i_ in self.inputs: |
| 102 | if i_.name in self.default_inputs: |
| 103 | feed[i_.name] = self.default_inputs[i_.name] |
| 104 | continue |
| 105 | |
| 106 | x = args[idx] |
| 107 | ts_x = np.array(x) if isinstance(x, (int, float, bool)) else x |
| 108 | # an annoying bug is numpy by default is int32, while pytorch is int64. |
| 109 | # so cast the input here automatically. |
| 110 | feed[i_.name] = \ |
| 111 | ts_x.astype(np.int64) if i_.type.tensor_type.elem_type == onnx_proto.TensorProto.INT64 else ts_x |
| 112 | idx += 1 |
| 113 | |
| 114 | # feed.update(kwargs) |
| 115 | return feed |
| 116 | |
| 117 | def __call__(self, *args, **kwargs): |
| 118 | self._ensure_ort_session() |
| 119 | outputs = self.ort_session.run(None, self._argument_map(*args, **kwargs)) |
| 120 | return outputs[0] if len(outputs) == 1 else tuple(outputs) |
| 121 | |
| 122 | |
| 123 | def optimize_model(model_or_file, output_file): |
| 124 | sess_options = OrtPyFunction.get_ort_session_options() |
| 125 | sess_options.graph_optimization_level = _ort.GraphOptimizationLevel.ORT_ENABLE_BASIC |
| 126 | sess_options.optimized_model_filepath = output_file |
| 127 | _ort.InferenceSession(model_or_file if isinstance(model_or_file, str) |
| 128 | else model_or_file.SerializeToString(), sess_options) |
| 129 | |