YOLOv8模型pytorch格式转为onnx格式

一、YOLOv8的Pytorch网络结构

model DetectionModel((model): Sequential((0): Conv((conv): Conv2d(3, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))(act): SiLU(inplace=True))(1): Conv((conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))(act): SiLU(inplace=True))(2): C2f((cv1): Conv((conv): Conv2d(128, 128, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(320, 128, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(m): ModuleList((0-2): 3 x Bottleneck((cv1): Conv((conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True)))))(3): Conv((conv): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))(act): SiLU(inplace=True))(4): C2f((cv1): Conv((conv): Conv2d(256, 256, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(m): ModuleList((0-5): 6 x Bottleneck((cv1): Conv((conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True)))))(5): Conv((conv): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))(act): SiLU(inplace=True))(6): C2f((cv1): Conv((conv): Conv2d(512, 512, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(2048, 512, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(m): ModuleList((0-5): 6 x Bottleneck((cv1): Conv((conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True)))))(7): Conv((conv): Conv2d(512, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))(act): SiLU(inplace=True))(8): C2f((cv1): Conv((conv): Conv2d(512, 512, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(1280, 512, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(m): ModuleList((0-2): 3 x Bottleneck((cv1): Conv((conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True)))))(9): SPPF((cv1): Conv((conv): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(m): MaxPool2d(kernel_size=5, stride=1, padding=2, dilation=1, ceil_mode=False))(10): Upsample(scale_factor=2.0, mode='nearest')(11): Concat()(12): C2f((cv1): Conv((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(1280, 512, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(m): ModuleList((0-2): 3 x Bottleneck((cv1): Conv((conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True)))))(13): Upsample(scale_factor=2.0, mode='nearest')(14): Concat()(15): C2f((cv1): Conv((conv): Conv2d(768, 256, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(640, 256, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(m): ModuleList((0-2): 3 x Bottleneck((cv1): Conv((conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True)))))(16): Conv((conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))(act): SiLU(inplace=True))(17): Concat()(18): C2f((cv1): Conv((conv): Conv2d(768, 512, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(1280, 512, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(m): ModuleList((0-2): 3 x Bottleneck((cv1): Conv((conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True)))))(19): Conv((conv): Conv2d(512, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))(act): SiLU(inplace=True))(20): Concat()(21): C2f((cv1): Conv((conv): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(1280, 512, kernel_size=(1, 1), stride=(1, 1))(act): SiLU(inplace=True))(m): ModuleList((0-2): 3 x Bottleneck((cv1): Conv((conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(cv2): Conv((conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True)))))(22): PostDetect((cv2): ModuleList((0): Sequential((0): Conv((conv): Conv2d(256, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(1): Conv((conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(2): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1)))(1-2): 2 x Sequential((0): Conv((conv): Conv2d(512, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(1): Conv((conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(2): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1))))(cv3): ModuleList((0): Sequential((0): Conv((conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(1): Conv((conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(2): Conv2d(256, 35, kernel_size=(1, 1), stride=(1, 1)))(1-2): 2 x Sequential((0): Conv((conv): Conv2d(512, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(1): Conv((conv): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))(act): SiLU(inplace=True))(2): Conv2d(256, 35, kernel_size=(1, 1), stride=(1, 1))))(dfl): DFL((conv): Conv2d(16, 1, kernel_size=(1, 1), stride=(1, 1), bias=False))))
)

yolov8网络从1-21层与pt文件相对应是BackBone和Neck模块，22层是Head模块。

二、转ONNX步骤

 2.1 yolov8官方

"""
代码解释
pt模型转为onnx格式
"""
import os
from ultralytics import YOLO
model = YOLO("weights/best.pt")
success = model.export(format="onnx")print("导出成功！")

将pytorch转为onnx后，pytorch支持的一系列计算就会转为onnx所支持的算子，若没有相对应的就会使用其他方式进行替换（比如多个计算替换其单个）。比较常见是conv和SiLU合并成一个Conv模块进行。

其中，1*4*8400表示每张图片预测 8400 个候选框，每个框有 4 个参数边界框坐标 (x,y,w,h)。 1*35*8400类同，1和4800代表意义相同，35是类别属性包含了其置信度概率值。

最后两个输出Concat操作，得到1*39*8400。最后根据这个结果去进行后续操作。

2.2 自定义转换

所谓的自定义转换其实是在转onnx时，对1*39*8400多加了一系列自定义操作例如NMS等。

2.2.1 加载权重并优化结构

YOLOv8 = YOLO(args.weights) #替换为自己的权重
model = YOLOv8.model.fuse().eval()

2.2.2 后处理检测模块

def gen_anchors(feats: Tensor,strides: Tensor,grid_cell_offset: float = 0.5) -> Tuple[Tensor, Tensor]:"""生成锚点，并计算每个锚点的步幅。参数：feats (Tensor): 特征图，通常来自不同的网络层。strides (Tensor): 每个特征图的步幅（stride）。grid_cell_offset (float): 网格单元的偏移量，默认为0.5。返回：Tuple[Tensor, Tensor]: 锚点的坐标和对应的步幅张量。"""anchor_points, stride_tensor = [], []assert feats is not None  # 确保输入的特征图不为空dtype, device = feats[0].dtype, feats[0].device  # 获取特征图的数据类型和设备# 遍历每个特征图，计算锚点for i, stride in enumerate(strides):_, _, h, w = feats[i].shape  # 获取特征图的高（h）和宽（w）sx = torch.arange(end=w, device=device,dtype=dtype) + grid_cell_offset  # 计算 x 轴上的锚点位置sy = torch.arange(end=h, device=device,dtype=dtype) + grid_cell_offset  # 计算 y 轴上的锚点位置sy, sx = torch.meshgrid(sy, sx)  # 生成网格坐标anchor_points.append(torch.stack((sx, sy), -1).view(-1, 2))  # 将 x 和 y 组合成坐标点stride_tensor.append(torch.full((h * w, 1), stride, dtype=dtype, device=device))  # 生成步幅张量return torch.cat(anchor_points), torch.cat(stride_tensor)  # 返回合并后的锚点和步幅class customize_NMS(torch.autograd.Function):"""继承torch.autograd.Function用于TensorRT的非极大值抑制（NMS）自定义函数。"""@staticmethoddef forward(ctx: Graph,boxes: Tensor,scores: Tensor,iou_threshold: float = 0.65,score_threshold: float = 0.25,max_output_boxes: int = 100,background_class: int = -1,box_coding: int = 0,plugin_version: str = '1',score_activation: int = 0) -> Tuple[Tensor, Tensor, Tensor, Tensor]:"""正向计算NMS输出，模拟真实的TensorRT NMS过程。参数：boxes (Tensor): 预测的边界框。scores (Tensor): 预测框的置信度分数。其他参数同样为NMS的超参数。返回：Tuple[Tensor, Tensor, Tensor, Tensor]: 包含检测框数量、框坐标、置信度分数和类别标签。"""batch_size, num_boxes, num_classes = scores.shape  # 获取批量大小、框数量和类别数num_dets = torch.randint(0,max_output_boxes, (batch_size, 1),dtype=torch.int32)  # 随机生成检测框数量（仅为模拟）boxes = torch.randn(batch_size, max_output_boxes, 4)  # 随机生成预测框scores = torch.randn(batch_size, max_output_boxes)  # 随机生成分数labels = torch.randint(0,num_classes, (batch_size, max_output_boxes),dtype=torch.int32)  # 随机生成类别标签return num_dets, boxes, scores, labels  # 返回模拟的结果@staticmethoddef symbolic(g,boxes: Value,scores: Value,iou_threshold: float = 0.45,score_threshold: float = 0.25,max_output_boxes: int = 100,background_class: int = -1,box_coding: int = 0,score_activation: int = 0,plugin_version: str = '1') -> Tuple[Value, Value, Value, Value]:"""计算图的符号函数，供TensorRT使用。参数：g: 计算图对象boxes (Value), scores (Value): 传入的边界框和得分其他参数是用于配置NMS的参数。返回：经过NMS处理的检测框、得分、类别标签及检测框数量。"""out = g.op('TRT::EfficientNMS_TRT',boxes,scores,iou_threshold_f=iou_threshold,score_threshold_f=score_threshold,max_output_boxes_i=max_output_boxes,background_class_i=background_class,box_coding_i=box_coding,plugin_version_s=plugin_version,score_activation_i=score_activation,outputs=4)  # 使用TensorRT的EfficientNMS插件nums_dets, boxes, scores, classes = out  # 获取输出的检测框数量、框坐标、得分和类别return nums_dets, boxes, scores, classes  # 返回结果class Post_process_Detect(nn.Module):"""用于后处理的检测模块，执行检测后的非极大值抑制（NMS）。"""export = Trueshape = Nonedynamic = Falseiou_thres = 0.65  # 默认的IoU阈值conf_thres = 0.25  # 默认的置信度阈值topk = 100  # 输出的最大检测框数量def __init__(self, *args, **kwargs):super().__init__()def forward(self, x):"""执行后处理操作，提取预测框、置信度和类别。参数：x (Tensor): 输入的特征图。返回：Tuple[Tensor, Tensor, Tensor]: 预测框、置信度和类别。"""shape = x[0].shape  # 获取输入的形状b, res, b_reg_num = shape[0], [], self.reg_max * 4# b为特征列表第一个元素的批量大小，表示处理的样本数量，# res声明一个空列表存储处理过的特征图# b_reg_num为回归框的数量#遍历特征层（self.nl表示特征层数），将每一层的框预测和分类预测拼接。for i in range(self.nl):res.append(torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1))  # 特征拼接# 调用# make_anchors# 生成锚点和步幅，用于还原边界框的绝对坐标。if self.dynamic or self.shape != shape:self.anchors, self.strides = (x.transpose(0, 1) for x in gen_anchors(x, self.stride, 0.5))  # 生成锚点和步幅self.shape = shape  # 更新输入的形状x = [i.view(b, self.no, -1) for i in res]  # 调整特征图形状y = torch.cat(x, 2)  # 拼接所有特征图boxes, scores = y[:, :b_reg_num, ...], y[:, b_reg_num:, ...].sigmoid()  # 提取框和分数boxes = boxes.view(b, 4, self.reg_max, -1).permute(0, 1, 3, 2)  # 变换框的形状boxes = boxes.softmax(-1) @ torch.arange(self.reg_max).to(boxes)  # 对框进行softmax处理boxes0, boxes1 = -boxes[:, :2, ...], boxes[:, 2:, ...]  # 分离框的不同部分boxes = self.anchors.repeat(b, 2, 1) + torch.cat([boxes0, boxes1], 1)  # 合并框坐标boxes = boxes * self.strides  # 乘以步幅return customize_NMS.apply(boxes.transpose(1, 2), scores.transpose(1, 2),self.iou_thres, self.conf_thres, self.topk)  # 执行NMSdef optim(module: nn.Module):setattr(module, '__class__', Post_process_Detect)for item in model.modules():optim(item)item.to(args.device) #输入cpu或者gpu的卡号

自定义这里是在yolo官方得到的1*4*8400和1*35*8400进行矩阵转换2<->3，最后引入EfficientNMS_TRT插件后处理，可以有效加速NMS处理。

2.2.3  EfficientNMS_TRT插件

EfficientNMS_TRT 是 TensorRT 中的一个高效非极大值抑制 (NMS) 插件，用于快速过滤检测框。它通过优化的 CUDA 实现来执行 NMS 操作，特别适合于深度学习推理阶段中目标检测任务的后处理。支持在一个批次中对多个图像同时执行 NMS。

输出结果为num_dets， detection_boxes， detection_scores， detection_classes ，分别代表经过 NMS 筛选后保留的边界框数，每张图片保留的检测框的坐标，每张图片中保留下来的检测框的分数（由高到低），每个保留下来的边界框的类别索引。

三、结语 

仅供学习使用！！！