|
| 1 | +import functools |
| 2 | +from typing import Dict, List, Optional, Sequence, Tuple, Union |
| 3 | +import onnx |
| 4 | +import torch |
| 5 | +from ..helpers.torch_helper import to_tensor |
| 6 | +from ..torch_onnx.runtime_info import first_used_last_used |
| 7 | +from . import torch_ops |
| 8 | + |
| 9 | + |
| 10 | +@functools.lru_cache |
| 11 | +def get_kernels() -> Dict[Tuple[str, str, int], type[torch_ops.OpRun]]: |
| 12 | + """Retrieves all the available kernels.""" |
| 13 | + res = {} |
| 14 | + for _k, v in torch_ops.__dict__.items(): |
| 15 | + if isinstance(v, type) and issubclass(v, torch_ops.OpRun) and "_" in v.__name__: |
| 16 | + name, version = v.__name__.split("_") |
| 17 | + domain = getattr(v, "domain", "") |
| 18 | + res[domain, name, int(version)] = v |
| 19 | + return res |
| 20 | + |
| 21 | + |
| 22 | +class TorchEvaluator: |
| 23 | + """ |
| 24 | + Torch evaluator for onnx models. |
| 25 | + The model does not stores the original proto it evaluates to avoid |
| 26 | +
|
| 27 | + :param proto: a proto |
| 28 | + :param providers: where to run the model |
| 29 | + :param opsets: needed if proto is a graph |
| 30 | +
|
| 31 | + The class holds the following attributes: |
| 32 | +
|
| 33 | + * `providers`: providers |
| 34 | + * `default_device`: default torch device |
| 35 | + * `constants`: all initializers or constants |
| 36 | + * `kernels`: kernels |
| 37 | + * `runtime_info`: produced by :func:`first_used_last_used |
| 38 | + <onnx_diagnostic.torch_onnx.runtime_info.first_used_last_used>` |
| 39 | + * `last_used`: contains the list of intermediate results, |
| 40 | + to remove after every node execution, |
| 41 | + this avoid the memory to grow too much |
| 42 | +
|
| 43 | + The class is not multithreaded. `runtime_info` gets updated |
| 44 | + by the the class. |
| 45 | + """ |
| 46 | + |
| 47 | + def __init__( |
| 48 | + self, |
| 49 | + proto: Union[onnx.FunctionProto, onnx.GraphProto, onnx.ModelProto], |
| 50 | + providers: Tuple[str, ...] = ("CPUExecutionProvider",), |
| 51 | + opsets: Optional[Dict[str, int]] = None, |
| 52 | + ): |
| 53 | + self.providers = providers |
| 54 | + self.constants: Dict[str, torch.Tensor] = {} |
| 55 | + self.kernels: List[Optional[torch_ops.OpRun]] = [] |
| 56 | + self.CPU = torch.tensor([0]).to("cpu").device |
| 57 | + if "CUDAExecutionProvider" in providers: |
| 58 | + self.CUDA = torch.tensor([0]).to("cuda").device |
| 59 | + self.default_device = self.CUDA |
| 60 | + else: |
| 61 | + self.default_device = self.CPU |
| 62 | + |
| 63 | + if isinstance(proto, onnx.ModelProto): |
| 64 | + assert opsets is None, "proto is a model, opsets must be None in that case" |
| 65 | + assert not proto.graph.sparse_initializer, "sparse_initializer not support yet" |
| 66 | + self.opsets = {d.domain: d.version for d in proto.opset_import} |
| 67 | + self._build_initializers(proto.graph.initializer) |
| 68 | + self._build_initializers(proto.graph.node) |
| 69 | + self._build_kernels(proto.graph.node) |
| 70 | + self.input_names = [i.name for i in proto.graph.input] |
| 71 | + self.output_names = [i.name for i in proto.graph.output] |
| 72 | + elif isinstance(proto, onnx.GraphProto): |
| 73 | + assert opsets, "opsets must be specified if proto is a graph" |
| 74 | + assert not proto.sparse_initializer, "sparse_initializer not support yet" |
| 75 | + self.opsets = opsets |
| 76 | + self._build_initializers(proto) |
| 77 | + self._build_initializers(proto.node) |
| 78 | + self._build_kernels(proto.nodes) |
| 79 | + self.input_names = [i.name for i in proto.input] |
| 80 | + self.output_names = [i.name for i in proto.output] |
| 81 | + elif isinstance(proto, onnx.FunctionProto): |
| 82 | + assert opsets is None, "proto is a model, opsets must be None in that case" |
| 83 | + self.opsets = {d.domain: d.version for d in proto.opset_import} |
| 84 | + self._build_initializers(proto.node) |
| 85 | + self._build_kernels(proto.node) |
| 86 | + self.input_names = list(proto.input) |
| 87 | + self.output_names = list(proto.output) |
| 88 | + else: |
| 89 | + raise TypeError(f"Unexpected type {type(proto)} for proto") |
| 90 | + |
| 91 | + self.runtime_info = first_used_last_used(proto, constant_as_initializer=True) |
| 92 | + self.last_used: List[List[str]] = [[] for _ in self.kernels] |
| 93 | + for name, info in self.runtime_info.items(): |
| 94 | + assert isinstance(info.last_used, int), f"Missing field last_used in {info!r}" |
| 95 | + if not info.is_output and not info.is_initializer: |
| 96 | + self.last_used[info.last_used].append(name) |
| 97 | + |
| 98 | + @property |
| 99 | + def on_cuda(self) -> bool: |
| 100 | + return self.default_device == self.CUDA |
| 101 | + |
| 102 | + def _build_initializers(self, inits: Sequence[Union[onnx.NodeProto, onnx.TensorProto]]): |
| 103 | + for init in inits: |
| 104 | + if isinstance(init, onnx.TensorProto): |
| 105 | + self.constants[init.name] = to_tensor(init).to(self.default_device) |
| 106 | + elif ( |
| 107 | + isinstance(init, onnx.NodeProto) |
| 108 | + and init.op_type == "Constant" |
| 109 | + and init.domain == "" |
| 110 | + ): |
| 111 | + value = None |
| 112 | + for att in init.attribute: |
| 113 | + if att.name == "value": |
| 114 | + value = to_tensor(att.t).to(self.default_device) |
| 115 | + assert value is not None, f"No attribute value in node {init}" |
| 116 | + self.constants[init.output[0]] = value |
| 117 | + |
| 118 | + def _build_kernels(self, nodes: Sequence[onnx.NodeProto]): |
| 119 | + kernels = get_kernels() |
| 120 | + self.kernels.clear() |
| 121 | + for node in nodes: |
| 122 | + if node.op_type == "Constant" and node.domain == "": |
| 123 | + # Treated as a constant. |
| 124 | + self.kernels.append(None) |
| 125 | + continue |
| 126 | + opset = self.opsets[node.domain] |
| 127 | + key = node.domain, node.op_type, opset |
| 128 | + while key not in kernels: |
| 129 | + opset -= 1 |
| 130 | + key = node.domain, node.op_type, opset |
| 131 | + assert ( |
| 132 | + key in kernels |
| 133 | + ), f"Missing kernel for node type {node.op_type!r} from domain {node.domain!r}" |
| 134 | + self.kernels.append(kernels[key](node, opset)) |
| 135 | + |
| 136 | + def run( |
| 137 | + self, outputs: Optional[List[str]], feeds: Dict[str, torch.Tensor] |
| 138 | + ) -> List[torch.Tensor]: |
| 139 | + """ |
| 140 | + Runs the ONNX model. |
| 141 | +
|
| 142 | + :param outputs: outputs required: |
| 143 | + :param feeds: inputs |
| 144 | + :return: output tensors. |
| 145 | + """ |
| 146 | + if outputs is None: |
| 147 | + outputs = self.output_names |
| 148 | + |
| 149 | + # sets constants |
| 150 | + for k, v in self.constants.items(): |
| 151 | + r = self.runtime_info[k] |
| 152 | + if not r.has_value: |
| 153 | + r.set_value(v.to(self.CUDA) if r.is_shape and self.on_cuda else v) |
| 154 | + |
| 155 | + # inputs |
| 156 | + for k, v in feeds.items(): |
| 157 | + r = self.runtime_info[k] |
| 158 | + r.set_value(v.to(self.CUDA) if r.is_shape and self.on_cuda else v) |
| 159 | + |
| 160 | + # node execution |
| 161 | + for it, kernel in enumerate(self.kernels): |
| 162 | + if kernel is not None: |
| 163 | + # kernel execution |
| 164 | + inputs = [(self.runtime_info[i].value if i else None) for i in kernel.input] |
| 165 | + res = kernel.run(*inputs) |
| 166 | + if isinstance(res, tuple): |
| 167 | + for name, t in zip(kernel.output, res): |
| 168 | + self.runtime_info[name].set_value(t) |
| 169 | + else: |
| 170 | + self.runtime_info[kernel.output[0]].set_value(res) |
| 171 | + |
| 172 | + # free intermediate results |
| 173 | + for name in self.last_used[it]: |
| 174 | + self.runtime_info[name].clean_value() |
| 175 | + |
| 176 | + # outputs |
| 177 | + res = [self.runtime_info[o].value for o in outputs] |
| 178 | + |
| 179 | + # clean previous execution |
| 180 | + for k in feeds: |
| 181 | + self.runtime_info[k].clean_value() |
| 182 | + for o in outputs: |
| 183 | + self.runtime_info[o].clean_value() |
| 184 | + |
| 185 | + return res |
0 commit comments