kin*_*eld 7 python onnx tensorflow2.0
我正在尝试使用 onnx 将detr 模型转换为张量流。torch.onnx.export我使用opset_version=12转换了模型。（它生成一个detr.onnx文件）
然后我尝试使用此示例将 onnx 文件转换为张量流模型。我添加了onnx.check_model一行以确保模型正确加载。
import math
from PIL import Image
import requests
import matplotlib.pyplot as plt
import torch
from torch import nn
from torchvision.models import resnet50
import onnx
from onnx_tf.backend import prepare
import torchvision.transforms as T
torch.set_grad_enabled(False)
model = torch.hub.load('facebookresearch/detr', 'detr_resnet50', pretrained=True)
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
im = Image.open(requests.get(url, stream=True).raw)
transform = T.Compose([
T.Resize(800),
T.ToTensor(),
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
img = transform(im).unsqueeze(0)
torch.onnx.export(model, img, 'detr.onnx', opset_version = 12)
onnx_model = onnx.load('./detr.onnx')
result = onnx.checker.check_model(onnx_model)
tf_rep = prepare(onnx_model)
tf_rep.export_graph('./model.pb')
此代码在到达行时引发异常 tf_rep.export_graph('./model.pb')。
onnx 版本 = 1.13.0 ，火炬版本 = 1.13.0+cu117 ，onnx_tf = 1.10.0
异常消息：
KeyError Traceback (most recent call last)
Cell In[19], line 26
23 result = onnx.checker.check_model(onnx_model)
25 tf_rep = prepare(onnx_model)
---> 26 tf_rep.export_graph('./model.pb')
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\onnx_tf\backend_rep.py:143, in TensorflowRep.export_graph(self, path)
129 """Export backend representation to a Tensorflow proto file.
130
131 This function obtains the graph proto corresponding to the ONNX
(...)
137 :returns: none.
138 """
139 self.tf_module.is_export = True
140 tf.saved_model.save(
141 self.tf_module,
142 path,
--> 143 signatures=self.tf_module.__call__.get_concrete_function(
144 **self.signatures))
145 self.tf_module.is_export = False
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\eager\def_function.py:1239, in Function.get_concrete_function(self, *args, **kwargs)
1237 def get_concrete_function(self, *args, **kwargs):
1238 # Implements GenericFunction.get_concrete_function.
-> 1239 concrete = self._get_concrete_function_garbage_collected(*args, **kwargs)
1240 concrete._garbage_collector.release() # pylint: disable=protected-access
1241 return concrete
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\eager\def_function.py:1219, in Function._get_concrete_function_garbage_collected(self, *args, **kwargs)
1217 if self._stateful_fn is None:
1218 initializers = []
-> 1219 self._initialize(args, kwargs, add_initializers_to=initializers)
1220 self._initialize_uninitialized_variables(initializers)
1222 if self._created_variables:
1223 # In this case we have created variables on the first call, so we run the
1224 # defunned version which is guaranteed to never create variables.
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\eager\def_function.py:785, in Function._initialize(self, args, kwds, add_initializers_to)
782 self._lifted_initializer_graph = lifted_initializer_graph
783 self._graph_deleter = FunctionDeleter(self._lifted_initializer_graph)
784 self._concrete_stateful_fn = (
--> 785 self._stateful_fn._get_concrete_function_internal_garbage_collected( # pylint: disable=protected-access
786 *args, **kwds))
788 def invalid_creator_scope(*unused_args, **unused_kwds):
789 """Disables variable creation."""
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\eager\function.py:2523, in Function._get_concrete_function_internal_garbage_collected(self, *args, **kwargs)
2521 args, kwargs = None, None
2522 with self._lock:
-> 2523 graph_function, _ = self._maybe_define_function(args, kwargs)
2524 return graph_function
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\eager\function.py:2760, in Function._maybe_define_function(self, args, kwargs)
2758 # Only get placeholders for arguments, not captures
2759 args, kwargs = placeholder_dict["args"]
-> 2760 graph_function = self._create_graph_function(args, kwargs)
2762 graph_capture_container = graph_function.graph._capture_func_lib # pylint: disable=protected-access
2763 # Maintain the list of all captures
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\eager\function.py:2670, in Function._create_graph_function(self, args, kwargs)
2665 missing_arg_names = [
2666 "%s_%d" % (arg, i) for i, arg in enumerate(missing_arg_names)
2667 ]
2668 arg_names = base_arg_names + missing_arg_names
2669 graph_function = ConcreteFunction(
-> 2670 func_graph_module.func_graph_from_py_func(
2671 self._name,
2672 self._python_function,
2673 args,
2674 kwargs,
2675 self.input_signature,
2676 autograph=self._autograph,
2677 autograph_options=self._autograph_options,
2678 arg_names=arg_names,
2679 capture_by_value=self._capture_by_value),
2680 self._function_attributes,
2681 spec=self.function_spec,
2682 # Tell the ConcreteFunction to clean up its graph once it goes out of
2683 # scope. This is not the default behavior since it gets used in some
2684 # places (like Keras) where the FuncGraph lives longer than the
2685 # ConcreteFunction.
2686 shared_func_graph=False)
2687 return graph_function
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\framework\func_graph.py:1247, in func_graph_from_py_func(name, python_func, args, kwargs, signature, func_graph, autograph, autograph_options, add_control_dependencies, arg_names, op_return_value, collections, capture_by_value, acd_record_initial_resource_uses)
1244 else:
1245 _, original_func = tf_decorator.unwrap(python_func)
-> 1247 func_outputs = python_func(*func_args, **func_kwargs)
1249 # invariant: `func_outputs` contains only Tensors, CompositeTensors,
1250 # TensorArrays and `None`s.
1251 func_outputs = nest.map_structure(
1252 convert, func_outputs, expand_composites=True)
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\eager\def_function.py:677, in Function._defun_with_scope.<locals>.wrapped_fn(*args, **kwds)
673 with default_graph._variable_creator_scope(scope, priority=50): # pylint: disable=protected-access
674 # __wrapped__ allows AutoGraph to swap in a converted function. We give
675 # the function a weak reference to itself to avoid a reference cycle.
676 with OptionalXlaContext(compile_with_xla):
--> 677 out = weak_wrapped_fn().__wrapped__(*args, **kwds)
678 return out
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\eager\function.py:3317, in class_method_to_instance_method.<locals>.bound_method_wrapper(*args, **kwargs)
3312 return wrapped_fn(weak_instance(), *args, **kwargs)
3314 # If __wrapped__ was replaced, then it is always an unbound function.
3315 # However, the replacer is still responsible for attaching self properly.
3316 # TODO(mdan): Is it possible to do it here instead?
-> 3317 return wrapped_fn(*args, **kwargs)
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\framework\func_graph.py:1233, in func_graph_from_py_func.<locals>.autograph_handler(*args, **kwargs)
1231 except Exception as e: # pylint:disable=broad-except
1232 if hasattr(e, "ag_error_metadata"):
-> 1233 raise e.ag_error_metadata.to_exception(e)
1234 else:
1235 raise
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\framework\func_graph.py:1222, in func_graph_from_py_func.<locals>.autograph_handler(*args, **kwargs)
1220 # TODO(mdan): Push this block higher in tf.function's call stack.
1221 try:
-> 1222 return autograph.converted_call(
1223 original_func,
1224 args,
1225 kwargs,
1226 options=autograph.ConversionOptions(
1227 recursive=True,
1228 optional_features=autograph_options,
1229 user_requested=True,
1230 ))
1231 except Exception as e: # pylint:disable=broad-except
1232 if hasattr(e, "ag_error_metadata"):
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\impl\api.py:439, in converted_call(f, args, kwargs, caller_fn_scope, options)
437 try:
438 if kwargs is not None:
--> 439 result = converted_f(*effective_args, **kwargs)
440 else:
441 result = converted_f(*effective_args)
File ~\AppData\Local\Temp\__autograph_generated_fileq0h7j9t_.py:30, in outer_factory.<locals>.inner_factory.<locals>.tf____call__(self, **kwargs)
28 node = ag__.Undefined('node')
29 onnx_node = ag__.Undefined('onnx_node')
---> 30 ag__.for_stmt(ag__.ld(self).graph_def.node, None, loop_body, get_state, set_state, (), {'iterate_names': 'node'})
31 outputs = ag__.converted_call(ag__.ld(dict), (), None, fscope)
33 def get_state_4():
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\operators\control_flow.py:463, in for_stmt(iter_, extra_test, body, get_state, set_state, symbol_names, opts)
459 _tf_distributed_iterable_for_stmt(
460 iter_, extra_test, body, get_state, set_state, symbol_names, opts)
462 else:
--> 463 _py_for_stmt(iter_, extra_test, body, None, None)
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\operators\control_flow.py:512, in _py_for_stmt(***failed resolving arguments***)
510 else:
511 for target in iter_:
--> 512 body(target)
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\operators\control_flow.py:478, in _py_for_stmt.<locals>.protected_body(protected_iter)
477 def protected_body(protected_iter):
--> 478 original_body(protected_iter)
479 after_iteration()
480 before_iteration()
File ~\AppData\Local\Temp\__autograph_generated_fileq0h7j9t_.py:23, in outer_factory.<locals>.inner_factory.<locals>.tf____call__.<locals>.loop_body(itr)
21 node = itr
22 onnx_node = ag__.converted_call(ag__.ld(OnnxNode), (ag__.ld(node),), None, fscope)
---> 23 output_ops = ag__.converted_call(ag__.ld(self).backend._onnx_node_to_tensorflow_op, (ag__.ld(onnx_node), ag__.ld(tensor_dict), ag__.ld(self).handlers), dict(opset=ag__.ld(self).opset, strict=ag__.ld(self).strict), fscope)
24 curr_node_output_map = ag__.converted_call(ag__.ld(dict), (ag__.converted_call(ag__.ld(zip), (ag__.ld(onnx_node).outputs, ag__.ld(output_ops)), None, fscope),), None, fscope)
25 ag__.converted_call(ag__.ld(tensor_dict).update, (ag__.ld(curr_node_output_map),), None, fscope)
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\impl\api.py:439, in converted_call(f, args, kwargs, caller_fn_scope, options)
437 try:
438 if kwargs is not None:
--> 439 result = converted_f(*effective_args, **kwargs)
440 else:
441 result = converted_f(*effective_args)
File ~\AppData\Local\Temp\__autograph_generated_filetsq4l59p.py:62, in outer_factory.<locals>.inner_factory.<locals>.tf___onnx_node_to_tensorflow_op(cls, node, tensor_dict, handlers, opset, strict)
60 pass
61 handler = ag__.Undefined('handler')
---> 62 ag__.if_stmt(ag__.ld(handlers), if_body_1, else_body_1, get_state_1, set_state_1, ('do_return', 'retval_'), 2)
64 def get_state_2():
65 return ()
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\operators\control_flow.py:1363, in if_stmt(cond, body, orelse, get_state, set_state, symbol_names, nouts)
1361 _tf_if_stmt(cond, body, orelse, get_state, set_state, symbol_names, nouts)
1362 else:
-> 1363 _py_if_stmt(cond, body, orelse)
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\operators\control_flow.py:1416, in _py_if_stmt(cond, body, orelse)
1414 def _py_if_stmt(cond, body, orelse):
1415 """Overload of if_stmt that executes a Python if statement."""
-> 1416 return body() if cond else orelse()
File ~\AppData\Local\Temp\__autograph_generated_filetsq4l59p.py:56, in outer_factory.<locals>.inner_factory.<locals>.tf___onnx_node_to_tensorflow_op.<locals>.if_body_1()
54 nonlocal retval_, do_return
55 pass
---> 56 ag__.if_stmt(ag__.ld(handler), if_body, else_body, get_state, set_state, ('do_return', 'retval_'), 2)
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\operators\control_flow.py:1363, in if_stmt(cond, body, orelse, get_state, set_state, symbol_names, nouts)
1361 _tf_if_stmt(cond, body, orelse, get_state, set_state, symbol_names, nouts)
1362 else:
-> 1363 _py_if_stmt(cond, body, orelse)
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\operators\control_flow.py:1416, in _py_if_stmt(cond, body, orelse)
1414 def _py_if_stmt(cond, body, orelse):
1415 """Overload of if_stmt that executes a Python if statement."""
-> 1416 return body() if cond else orelse()
File ~\AppData\Local\Temp\__autograph_generated_filetsq4l59p.py:48, in outer_factory.<locals>.inner_factory.<locals>.tf___onnx_node_to_tensorflow_op.<locals>.if_body_1.<locals>.if_body()
46 try:
47 do_return = True
---> 48 retval_ = ag__.converted_call(ag__.ld(handler).handle, (ag__.ld(node),), dict(tensor_dict=ag__.ld(tensor_dict), strict=ag__.ld(strict)), fscope)
49 except:
50 do_return = False
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\impl\api.py:439, in converted_call(f, args, kwargs, caller_fn_scope, options)
437 try:
438 if kwargs is not None:
--> 439 result = converted_f(*effective_args, **kwargs)
440 else:
441 result = converted_f(*effective_args)
File ~\AppData\Local\Temp\__autograph_generated_filec7_esoft.py:41, in outer_factory.<locals>.inner_factory.<locals>.tf__handle(cls, node, **kwargs)
39 nonlocal retval_, do_return
40 raise ag__.converted_call(ag__.ld(BackendIsNotSupposedToImplementIt), (ag__.converted_call('{} version {} is not implemented.'.format, (ag__.ld(node).op_type, ag__.ld(cls).SINCE_VERSION), None, fscope),), None, fscope)
---> 41 ag__.if_stmt(ag__.ld(ver_handle), if_body, else_body, get_state, set_state, ('do_return', 'retval_'), 2)
42 return fscope.ret(retval_, do_return)
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\operators\control_flow.py:1363, in if_stmt(cond, body, orelse, get_state, set_state, symbol_names, nouts)
1361 _tf_if_stmt(cond, body, orelse, get_state, set_state, symbol_names, nouts)
1362 else:
-> 1363 _py_if_stmt(cond, body, orelse)
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\operators\control_flow.py:1416, in _py_if_stmt(cond, body, orelse)
1414 def _py_if_stmt(cond, body, orelse):
1415 """Overload of if_stmt that executes a Python if statement."""
-> 1416 return body() if cond else orelse()
File ~\AppData\Local\Temp\__autograph_generated_filec7_esoft.py:33, in outer_factory.<locals>.inner_factory.<locals>.tf__handle.<locals>.if_body()
31 try:
32 do_return = True
---> 33 retval_ = ag__.converted_call(ag__.ld(ver_handle), (ag__.ld(node),), dict(**ag__.ld(kwargs)), fscope)
34 except:
35 do_return = False
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\impl\api.py:439, in converted_call(f, args, kwargs, caller_fn_scope, options)
437 try:
438 if kwargs is not None:
--> 439 result = converted_f(*effective_args, **kwargs)
440 else:
441 result = converted_f(*effective_args)
File ~\AppData\Local\Temp\__autograph_generated_filevddqx9qt.py:12, in outer_factory.<locals>.inner_factory.<locals>.tf__version(cls, node, **kwargs)
10 try:
11 do_return = True
---> 12 retval_ = ag__.converted_call(ag__.ld(cls)._common, (ag__.ld(node),), dict(**ag__.ld(kwargs)), fscope)
13 except:
14 do_return = False
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\impl\api.py:439, in converted_call(f, args, kwargs, caller_fn_scope, options)
437 try:
438 if kwargs is not None:
--> 439 result = converted_f(*effective_args, **kwargs)
440 else:
441 result = converted_f(*effective_args)
File ~\AppData\Local\Temp\__autograph_generated_filedezd6jrz.py:122, in outer_factory.<locals>.inner_factory.<locals>.tf___common(cls, node, **kwargs)
120 paddings = ag__.Undefined('paddings')
121 constant_values = ag__.Undefined('constant_values')
--> 122 ag__.if_stmt(ag__.ld(cls).SINCE_VERSION < 11, if_body_1, else_body_1, get_state_1, set_state_1, ('constant_values', 'paddings'), 2)
123 cond = ag__.converted_call(ag__.ld(tf).cond, (ag__.converted_call(ag__.ld(check_positive), (ag__.ld(paddings),), None, fscope), ag__.autograph_artifact(lambda : ag__.converted_call(ag__.ld(process_pos_pads), (ag__.ld(x), ag__.ld(paddings), ag__.ld(constant_values)), None, fscope)), ag__.autograph_artifact(lambda : ag__.converted_call(ag__.ld(process_neg_pads), (ag__.ld(x), ag__.ld(paddings), ag__.ld(constant_values)), None, fscope))), None, fscope)
124 try:
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\operators\control_flow.py:1363, in if_stmt(cond, body, orelse, get_state, set_state, symbol_names, nouts)
1361 _tf_if_stmt(cond, body, orelse, get_state, set_state, symbol_names, nouts)
1362 else:
-> 1363 _py_if_stmt(cond, body, orelse)
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\operators\control_flow.py:1416, in _py_if_stmt(cond, body, orelse)
1414 def _py_if_stmt(cond, body, orelse):
1415 """Overload of if_stmt that executes a Python if statement."""
-> 1416 return body() if cond else orelse()
File ~\AppData\Local\Temp\__autograph_generated_filedezd6jrz.py:119, in outer_factory.<locals>.inner_factory.<locals>.tf___common.<locals>.else_body_1()
117 nonlocal paddings, constant_values
118 paddings = ag__.ld(tensor_dict)[ag__.ld(node).inputs[1]]
--> 119 constant_values = ag__.if_exp(ag__.converted_call(ag__.ld(len), (ag__.ld(node).inputs,), None, fscope) == 3, lambda : ag__.ld(tensor_dict)[ag__.ld(node).inputs[2]], lambda : 0, 'ag__.converted_call(len, (node.inputs,), None, fscope) == 3')
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\operators\conditional_expressions.py:27, in if_exp(cond, if_true, if_false, expr_repr)
25 return _tf_if_exp(cond, if_true, if_false, expr_repr)
26 else:
---> 27 return _py_if_exp(cond, if_true, if_false)
File ~\AppData\Local\Programs\Python\Python39\lib\site-packages\tensorflow\python\autograph\o
您面临的问题是由于在模型源使用动态填充而不是静态填充形状造成的。当您在导出期间降低 onnx opset 版本时，会暴露此问题。
import warnings
warnings.filterwarnings("ignore")
#import onnxruntime
import math
from PIL import Image
import requests
import matplotlib.pyplot as plt
import torch
from torch import nn
from torchvision.models import resnet50
import torchvision.transforms as T
import onnx
from onnx_tf.backend import prepare
#from onnxsim import simplify
torch.set_grad_enabled(False)
model = torch.hub.load('facebookresearch/detr', 'detr_resnet50', pretrained=True)
url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
im = Image.open(requests.get(url, stream=True).raw)
transform = T.Compose([
T.Resize(800),
T.ToTensor(),
T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
img = transform(im).unsqueeze(0)
model.eval()
torch.onnx.export(model, img, 'detr.onnx', opset_version = 10)
onnx_model = onnx.load('./detr.onnx')
#onnx_model, _ = simplify(model)
result = onnx.checker.check_model(onnx_model)
tf_rep = prepare(onnx_model)
tf_rep.export_graph('./model.pb')
它将抛出以下输出：
SymbolicValueError Traceback (most recent call last)
c:\Anaconda3\envs\workenv\lib\site-packages\torch\onnx\symbolic_opset9.py in _convert_padding_node(input)
1821 try:
-> 1822 padding = [
1823 symbolic_helper._get_const(v, "i", "padding") for v in input_list
c:\Anaconda3\envs\workenv\lib\site-packages\torch\onnx\symbolic_opset9.py in <listcomp>(.0)
1822 padding = [
-> 1823 symbolic_helper._get_const(v, "i", "padding") for v in input_list
1824 ]
c:\Anaconda3\envs\workenv\lib\site-packages\torch\onnx\symbolic_helper.py in _get_const(value, desc, arg_name)
169 if not _is_constant(value):
--> 170 raise errors.SymbolicValueError(
171 f"ONNX symbolic expected a constant value of the '{arg_name}' argument, "
SymbolicValueError: ONNX symbolic expected a constant value of the 'padding' argument, got '509 defined in (%509 : Long(requires_grad=0, device=cpu) = onnx::Sub(%max_size_i, %498), scope: models.detr.DETR:: # C:\Users\Anurag/.cache\torch\hub\facebookresearch_detr_main\util\misc.py:349:0
)' [Caused by the value '509 defined in (%509 : Long(requires_grad=0, device=cpu) = onnx::Sub(%max_size_i, %498), scope: models.detr.DETR:: # C:\Users\Anurag/.cache\torch\hub\facebookresearch_detr_main\util\misc.py:349:0
)' (type 'Tensor') in the TorchScript graph. The containing node has kind 'onnx::Sub'.]
(node defined in C:\Users\Anurag/.cache\torch\hub\facebookresearch_detr_main\util\misc.py(349): <listcomp>
C:\Users\Anurag/.cache\torch\hub\facebookresearch_detr_main\util\misc.py(349): _onnx_nested_tensor_from_tensor_list
C:\Users\Anurag/.cache\torch\hub\facebookresearch_detr_main\util\misc.py(313): nested_tensor_from_tensor_list
C:\Users\Anurag/.cache\torch\hub\facebookresearch_detr_main\models\detr.py(60): forward
c:\Anaconda3\envs\workenv\lib\site-packages\torch\nn\modules\module.py(1182): _slow_forward
...
#5: 507 defined in (%507 : Long(requires_grad=0, device=cpu) = onnx::Sub(%478, %466), scope: models.detr.DETR:: # C:\Users\Anurag/.cache\torch\hub\facebookresearch_detr_main\util\misc.py:349:0
) (type 'Tensor')
Outputs:
#0: 510 defined in (%510 : int[] = prim::ListConstruct(%459, %509, %459, %508, %459, %507), scope: models.detr.DETR::
) (type 'List[int]')
我的建议是从其他来源获取模型。
作为参考，请查看：ONNX 符号期望填充参数的常量值
| 归档时间： |
|
| 查看次数： |
2052 次 |
| 最近记录： |