onnx文件的转换以及pt文件的修改

将训练后的pt文件转换为onnx文件

pip install onnx coremltools onnx-simplifier

导出onnx文件

pip install netron
python
import netron
netron.start('xxx.onnx')

配置Python中的onnx环境

pip install onnxruntime
pip install onnx

Python中一些基础的onnx操作

• onnx模型通常只能拿到最后输出节点的输出数据，若想拿到中间节点的输出数据，需要我们自己添加相应的输出节点信息；首先需要构建指定的节点（层名称、数据类型、维度信息）；然后再通过insert的方式将节点插入到模型中。

获取onnx的输出层

import onnx
# 加载模型
model = onnx.load('onnx_model.onnx')
# 检查模型格式是否完整及正确
onnx.checker.check_model(model)
# 获取输出层，包含层名称、维度信息
output = self.model.graph.output
print(output)

获取节点的输出数据

import onnx
from onnx import helper
# 加载模型
model = onnx.load('onnx_model.onnx')
# 创建中间节点：层名称、数据类型、维度信息
prob_info =  helper.make_tensor_value_info('layer1',onnx.TensorProto.FLOAT, [1, 3, 320, 280])
# 将构建完成的中间节点插入到模型中
model.graph.output.insert(0, prob_info)
# 保存新的模型
onnx.save(model, 'onnx_model_new.onnx')

# 扩展：
# 删除指定的节点方法： item为需要删除的节点
# model.graph.output.remove(item)

获取查询节点的位置

def find_node(model,name):  # 返回节点的序号
    for node_id,node in enumerate(model.graph.node):
        if node.name == name:
            return node_id

删除节点

def del_node(model,name):
    if type(name) == list:
        for id_ in name:
            del_node(model,id_)
    else:
        id_ = find_node(model,name)
        if id_ == None :
            print("输入名字错误，请核对信息")
        else:
            tem_node = model.graph.node[id_]
            model.graph.node.remove(tem_node)

onnx前向推理InferenceSession的使用

import onnxruntime

# 创建一个InferenceSession的实例，并将模型的地址传递给该实例
sess = onnxruntime.InferenceSession('onnxmodel.onnx')
# 调用实例sess的润方法进行推理
outputs = sess.run(output_layers_name, {input_layers_name: x})


class InferenceSession(Session):
    """
    This is the main class used to run a model.
    """
    def __init__(self, path_or_bytes, sess_options=None, providers=[]):
        """
        :param path_or_bytes: filename or serialized model in a byte string
        :param sess_options: session options
        :param providers: providers to use for session. If empty, will use
            all available providers.
        """
        self._path_or_bytes = path_or_bytes
        self._sess_options = sess_options
        self._load_model(providers)
        self._enable_fallback = True
        Session.__init__(self, self._sess)

    def _load_model(self, providers=[]):
        if isinstance(self._path_or_bytes, str):
            self._sess = C.InferenceSession(
                self._sess_options if self._sess_options else C.get_default_session_options(), self._path_or_bytes,
                True)
        elif isinstance(self._path_or_bytes, bytes):
            self._sess = C.InferenceSession(
                self._sess_options if self._sess_options else C.get_default_session_options(), self._path_or_bytes,
                False)
        # elif isinstance(self._path_or_bytes, tuple):
        # to remove, hidden trick
        #   self._sess.load_model_no_init(self._path_or_bytes[0], providers)
        else:
            raise TypeError("Unable to load from type '{0}'".format(type(self._path_or_bytes)))

        self._sess.load_model(providers)

        self._sess_options = self._sess.session_options
        self._inputs_meta = self._sess.inputs_meta
        self._outputs_meta = self._sess.outputs_meta
        self._overridable_initializers = self._sess.overridable_initializers
        self._model_meta = self._sess.model_meta
        self._providers = self._sess.get_providers()

        # Tensorrt can fall back to CUDA. All others fall back to CPU.
        if 'TensorrtExecutionProvider' in C.get_available_providers():
            self._fallback_providers = ['CUDAExecutionProvider', 'CPUExecutionProvider']
        else:
            self._fallback_providers = ['CPUExecutionProvider']

• 在_load_model函数，可以发现在load模型的时候是通过C.InferenceSession，并且将相关的操作也委托给该类。
• 从导入语句from onnxruntime.capi import _pybind_state as C可知其实就是一个c++实现的Python接口,其源码在onnxruntime\onnxruntime\python\onnxruntime_pybind_state.cc中。

模型推理run，源码分析

    def run(self, output_names, input_feed, run_options=None):
        """
        Compute the predictions.

        :param output_names: name of the outputs
        :param input_feed: dictionary ``{ input_name: input_value }``
        :param run_options: See :class:`onnxruntime.RunOptions`.

        ::

            sess.run([output_name], {input_name: x})
        """
        num_required_inputs = len(self._inputs_meta)
        num_inputs = len(input_feed)
        # the graph may have optional inputs used to override initializers. allow for that.
        if num_inputs < num_required_inputs:
            raise ValueError("Model requires {} inputs. Input Feed contains {}".format(num_required_inputs, num_inputs))
        if not output_names:
            output_names = [output.name for output in self._outputs_meta]
        try:
            return self._sess.run(output_names, input_feed, run_options)
        except C.EPFail as err:
            if self._enable_fallback:
                print("EP Error: {} using {}".format(str(err), self._providers))
                print("Falling back to {} and retrying.".format(self._fallback_providers))
                self.set_providers(self._fallback_providers)
                # Fallback only once.
                self.disable_fallback()
                return self._sess.run(output_names, input_feed, run_options)
            else:
                raise

• 以下内容为个人修改onnx文件所用，不是通用版本，请勿使用

import onnx
from onnx import helper
# 删除节点的参考
# https://blog.csdn.net/DazeAJD/article/details/126055822  
# https://blog.csdn.net/qq_16792139/article/details/128720432?spm=1001.2101.3001.6650.4&utm_medium=distribute.pc_relevant.none-task-blog-2%7Edefault%7EYuanLiJiHua%7EPosition-4-128720432-blog-117163212.pc_relevant_3mothn_strategy_recovery&depth_1-utm_source=distribute.pc_relevant.none-task-blog-2%7Edefault%7EYuanLiJiHua%7EPosition-4-128720432-blog-117163212.pc_relevant_3mothn_strategy_recovery&utm_relevant_index=5
name = ["Concat_628","Unsqueeze_627","Unsqueeze_626","Unsqueeze_625",] #需要删除节点的名字
# 删除节点的示例
# prob_info = helper.make_tensor_value_info("layer1",onnx.TensorProto.FLOAT,[1,3,416,416])
#model.graph.output.insert(0,prob_info)  # 插入输出端
# del_1 = model.graph.output[0]
# model.graph.output.remove(del_1)
names_concat_3 = ["/model.24/Concat","/model.24/Concat_1","/model.24/Concat_2"]
name_concat_1 = ["model.24/Concat_3"]

def del_node(model,name):
    if type(name) == list:
        for id_ in name:
            del_node(model,id_)
    else:
        id_ = find_node(model,name)
        if id_ == None :
            print("输入名字错误，请核对信息")

        elif id_ > 0 and id_ + 1 < (len(model.graph.node) - 1):
            # pre,past = [model.graph.node[id-1].name,model.graph.node[id+1].name]
            # 首先将上下节点位置的输入输出对接起来
            # model.graph.node[id_ + 1].input = model.graph.node[id-1].output

            # 删除节点
            tem_node = model.graph.node[id_]
            model.graph.node.remove(tem_node)
        else:
            tem_node = model.graph.node[id_]
            model.graph.node.remove(tem_node)


def find_node(model,name):  # 返回节点的序号
    id_list = []
    tem_list = []
    for node_id,node in enumerate(model.graph.node):
        if node.name == name:
            return node_id




if __name__=="__main__":


    model = onnx.load(r"C:\yolo\1\yolov5s_new_2.onnx")

    del_node(model,"/model.24/Concat_3")
    # 创建节点使用  onnx.helper.make_node("type_name",name = "your node name",other_property)
    # 插入新的节点请使用 model.graph.node.insert(find_node(model,"your delet node last input name") + 1, new_node)
    onnx.save(model,r"C:\yolo\1\yolov5s_new.onnx")

"""
ben_1 = onnx.helper.make_node("Concat",name = "/model.24.Concat_ben_1",axis = 4,inputs =["/model.24/Mul_1_output_0","/model.24/Mul_3_output_0"],outputs =["ben_1"] )
ben_2 = onnx.helper.make_node("Concat",name = "/model.24.Concat_ben_2",axis = 4,inputs =["/model.24/Split_output_2","ben_1"],outputs =["ben_2"] )
ben_3 = onnx.helper.make_node("Concat",name = "/model.24.Concat_ben_3",axis = 4,inputs = ["ben_1","ben_2"],outputs = ["output_0"])
"""