[人工智能]Yolov5 TensorRT推理加速(c++版)

Yolov5 不做赘述，目前目标检测里使用非常多的模型，效果和速度兼顾，性能强悍，配合TensorRT推理加速，在工业界可以说是非常流行的组合。

废话不多说，直接开整，以下使用的Tensor RT部署推理路线为：Pytorch-> ONNX -> TensorRT。

pytorch导出到onnx模型，可以非常方便，并且支持dynamic维度，配合netron工具，可以查看模型的网络结构，而TensorRT对ONNX的支持也非常完整，所以选择这一套流程，可以非常轻松的完成TensorRT的部署，同时，tensor RT提供官方的nms插件，使得推理代码可以免去编写nms的部分，极大提高效率。

GitHub仓库地址： gentlemanarch/yolov5-tensorrt

环境准备

1. 系统：Ubuntu20.04LTS系统，或者tensorRT官方docker镜像：nvcr.io/nvidia/tensorrt:21.05-py3（推荐）

2. TensorRT: 本篇使用的TensorRT7.2.3

3. Yolov5: 截止2021.09.11的develop分支代码

4. gcc: 9.3.0

5. torch: 1.8.2

6. onnx:1.10.1

7. onnx-simplifier:0.3.6

pytorch模型训练

clone Yolov5的官方代码，按照教程训练得到pt权重文件。

github地址：

https://github.com/ultralytics/yolov5

Torch -> onnx

在导出到onnx之前，为了方便后续添加nms插件，需要对torch的模型输出做一些修改.

models/yolo.py:

将这部分代码

 ? ? ? ? ? ? ? ?

                if self.inplace:
                    y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
                    y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
                else:  # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953
                    xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
                    wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i].view(1, self.na, 1, 1, 2)  # wh
                    y = torch.cat((xy, wh, y[..., 4:]), -1)
                z.append(y.view(bs, -1, self.no))

替换为：

                if self.inplace:
                    xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
                    wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i].view(1, self.na, 1, 1, 2).expand(bs, self.na, 1, 1, 2)  # wh
                    rest = y[..., 4:]
                    yy = torch.cat((xy, wh, rest), -1)
                    z.append(yy.view(bs, -1, self.no))
                else:
                    xy = (y[..., 0:2] * 2. - 0.5 + self.grid[i]) * self.stride[i]  # xy
                    wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i].view(1, self.na, 1, 1, 2)  # wh
                    y = torch.cat((xy, wh, y[..., 4:]), -1)

同时，这句:

return x if self.training else (torch.cat(z, 1), x)

替换为：

return x if self.training else torch.cat(z, 1)

最后如图所示：

models/export.py

改写后，官方的export.py已不适用，使用以下export代码：

'''
export yolov5 .pt model to onnx model

Usage:
    python models/export.py --weights yolov5s.pt --img-size 640 \
    --batch-size 1 --device 0 --include onnx --inplace --dynamic \
    --simplify --opset-version 11 --img test_img/1.jpg
'''

import argparse
import sys
import time
from pathlib import Path

sys.path.append(Path(__file__).parent.parent.absolute().__str__())  # to run '$ python *.py' files in subdirectories

import torch
import torch.nn as nn
from torch.utils.mobile_optimizer import optimize_for_mobile

import models
from models.experimental import attempt_load
from utils.activations import Hardswish, SiLU
from utils.general import colorstr, check_img_size, check_requirements, file_size, set_logging
from utils.torch_utils import select_device


def xywh2xyxy(x):
    center = x[:, :, :2]
    wh = x[:, :, 2:] / 2.
    return torch.cat([center - wh, center + wh], -1)


class Yolov5(nn.Module):
    def __init__(self, opt):
        super().__init__()
        self.model = self.init_model(opt)
    
    def init_model(self, opt):
        # load PyTorch model
        model = attempt_load(opt.weights)

        for k, m in model.named_modules():
            m._non_persistent_buffers_set = set()  # pytorch 1.6 compatbility
            if isinstance(m, models.common.Conv):
                if isinstance(m.act, nn.Hardswish):
                    m.act = Hardswish()
                elif isinstance(m.act, nn.SiLU):
                    m.act = SiLU()
            elif isinstance(m, models.yolo.Detect):
                m.inplace = opt.inplace
                m.onnx_dynamic = opt.dynamic
        return model

    def forward(self, x):
        output = self.model(x)
        output = self.post_processing(output)
        return output

    def post_processing(self, x):
        bs, nb_box, infos = x.shape

        boxes_input = xywh2xyxy(x[..., :4]).reshape(bs, nb_box, 1, 4)
        scores_input = x[..., 5:] * x[..., 4:5]
        return [boxes_input, scores_input]

def remove_initializer_from_input(model):
    if model.ir_version < 4:
        print(
            'Model with ir_version below 4'
        )
        return
    inputs = model.graph.input
    name_to_input = {}
    for input in inputs:
        name_to_input[input.name] = input
    
    for initializer in model.graph.initializer:
        if initializer.name in name_to_input:
            inputs.remove(name_to_input[initializer.name])
    
    return model

def to_numpy(tensor):
    return tensor.detach().cpu().numpy() if tensor.requires_grad else tensor.cpu().numpy()

def preprocess_image(image_raw, INPUT_W=640, INPUT_H=640):
    h, w, c = image_raw.shape
    image = image_raw.copy()
    r_w = INPUT_W / w
    r_h = INPUT_H / h
    if r_h > r_w:
        tw = INPUT_W
        th = int(r_w * h)
        tx1 = tx2 = 0
        ty1 = int((INPUT_H - th) / 2)
        ty2 = INPUT_H - th - ty1
    else:
        tw = int(r_h * w)
        th = INPUT_H
        tx1 = int((INPUT_W - tw) / 2)
        tx2 = INPUT_W - tw - tx1
        ty1 = ty2 = 0
    image = cv2.resize(image, (tw, th))
    image = cv2.copyMakeBorder(
        image, ty1, ty2, tx1, tx2, cv2.BORDER_CONSTANT, (128, 128, 128)
    )
    return image


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--weights', type=str, default='./yolov5s.pt', help='weights path')
    parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='image size')
    parser.add_argument('--batch-size', type=int, default=1, help='batch size')
    parser.add_argument('--device', default='cpu', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
    parser.add_argument('--include', nargs='+', default=['torchscript', 'onnx', 'coreml'], help='include format')
    parser.add_argument('--half', action='store_true', help='FP16 half-precision export')
    parser.add_argument('--inplace', action='store_true', help='set Yolov5 Detect() inplace=True')
    parser.add_argument('--train', action='store_true', help='model.train() mode')
    parser.add_argument('--optimize', action='store_true', help='optimize TorchScript for mobile') # TorchScript-only
    parser.add_argument('--dynamic', action='store_true', help='dynamic ONNX axes')
    parser.add_argument('--simplify', action='store_true', help='simplify ONNX model')
    parser.add_argument('--opset-version', type=int, default=11, help='ONNX opset version')
    parser.add_argument('--img', type=str, default='', help='test image path')
    opt = parser.parse_args()
    opt.img_size *= 2 if len(opt.img_size) == 1 else 1
    opt.img_size = (640, 640)
    opt.include = [x.lower() for x in opt.include]
    print(opt)
    set_logging()
    t = time.time()


    device = select_device(opt.device)


    import cv2
    image_path = opt.img
    image = cv2.imread(image_path)

    from utils.datasets import letterbox
    frame = preprocess_image(image, opt.img_size[0], opt.img_size[1])
    print('frame.shape: ', frame.shape)

    img = torch.from_numpy(frame).float().unsqueeze(0)
    img = img.permute(0, 3, 1, 2)
    img = img[:, [2, 1, 0]] / 255.

    assert not (opt.device.lower() == 'cpu' and opt.half), '--half only compatible with GPU export, i.e. use --device 0'
    model = Yolov5(opt)

    if opt.half:
        img, model = img.half(), model.half()
    model.train() if opt.train else model.eval()

    for k, m in model.named_modules():
        m._non_persistent_buffers_set = set()  # pytorch 1.6 compatbility
        if isinstance(m, models.common.Conv):
            if isinstance(m.act, nn.Hardswish):
                m.act = Hardswish()
            elif isinstance(m.act, nn.SiLU):
                m.act = SiLU()
        elif isinstance(m, models.yolo.Detect):
            m.inplace = opt.inplace
            m.onnx_dynamic = opt.dynamic

    for _ in range(2):
        y = model(img)
    print(f"\n{colorstr('PyTorch:')} starting from {opt.weights} ({file_size(opt.weights):.1f} MB)")

    if 'onnx' in opt.include:
        prefix = colorstr('ONNX:')
        try:
            import onnx

            print(f'{prefix} starting export with onnx {onnx.__version__}...')
            f = opt.weights.replace('.pt', 'fix.onnx') if not opt.dynamic else opt.weights.replace('.pt', '_dynamic.onnx')
            dynamic_axes = {'input': {0: 'batch'},
                            'boxes': {0: 'batch'},
                            'confs': {0: 'batch'}}

            torch.onnx.export(model, img, f, verbose=True, opset_version=opt.opset_version,
                              training=torch.onnx.TrainingMode.TRAINING if opt.train else torch.onnx.TrainingMode.EVAL,
                              do_constant_folding=True, export_params=True, operator_export_type=torch.onnx.OperatorExportTypes.ONNX,
                              input_names=['input'],
                              output_names=['boxes', 'confs'],
                              dynamic_axes=dynamic_axes if opt.dynamic else None)
            model_onnx = onnx.load(f)
            onnx.checker.check_model(model_onnx)

            import onnxoptimizer
            print("Beginning ONNX model path optimization")
            all_passes = onnxoptimizer.get_available_passes()
            passes = ["extract_constant_to initializer", "eliminate_unused_initializer", "fuse_bn_into_conv"]
            assert all(p in all_passes for p in passes)
            model_onnx = onnoptimizer.optimize(model_onnx, passes)
            print("Completed ONNX model path optimization")

            if opt.simplify:
                try:
                    check_requirements(['onnx-simplifier'])
                    import onnxsim

                    print(f'{prefix} simplifying with onnx-simplifier {onnxsim.__version__}...')
                    model_onnx, check = onnxsim.simplify(
                        model_onnx,
                        dynamic_input_shape=opt.dynamic,
                        input_shapes={'input': list(img.shape)} if opt.dynamic else None
                    )
                    assert check, 'assert check failed'
                    model_onnx = remove_initializer_from_input(model_onnx)
                    onnx.save(model_onnx, f)
                except Exception as e:
                    print(f'{prefix} simplifier failure: {e}')
                print(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)')

        except Exception as e:
            print(f'{prefix} export failure: {e}')
            print(e)

    print(f'\nExport complete ({time.time() - t:.2f}s). Visualize with https://github.com/lutzroeder/netron.')

    import onnxruntime
    import numpy as np
    ort_session = onnxruntime.InferenceSession(f)
    ort_inputs = {ort_session.get_inputs()[0].name: to_numpy(img)}
    onnx_out = ort_session.run(None, ort_inputs)
    torch_out = model(img)

    import ipdb
    ipdb.set_trace()
    np.testing.assert_allclose(to_numpy(torch_out[0]), onnx_out[0], rtol=1e-03, atol=1e-05)
    print("Exported model has been tested with ONNXRuntime, and the result looks good!")
    np.testing.assert_allclose(to_numpy(torch_out[1]), onnx_out[1], rtol=1e-03, atol=1e-05)
    print("Exported model has been tested with ONNXRuntime, and the result looks good!")

在yolov5项目根目录中，使用以下命令导出onnx模型：

python models/export.py --weights yolov5s.pt --img-size 640 \
    --batch-size 1 --device 0 --include onnx --inplace --dynamic \
    --simplify --opset-version 11 --img test_img/1.jpg

其中参数：

• weights指定pytorch权重路径

• img-size指定图片输入尺寸，以上程序中，固定了输入维度为640，640，所以这个参数并不起作用，可以修改代码中的opt.img_size = (640, 640)部分

• --dynamic，允许输入维度可变，以上提供的代码中，只有batchsize维度可变，如果需要height和width都可变，可将dynamic_axes修改如下：

  dynamic_axes = {'input': {0: 'batch', 2: 'height', 3: 'width'},
                  'boxes': {0: 'batch'},
                  'confs': {0: 'batch'}}

• --simplify简化onnx模型，去掉梯度、优化器等推理中不需要的部分

• opset-version, 算法版本，目前11支持比较完善

• img, 指定一张测试使用的图片

执行完导出命令后，会在pt权重文件对应的目录下，得到一个onnx模型。

onnx->TensorRT & TensorRT inference

编译C++代码

clone yolov5_trt代码到本地，

cd yolov5_trt
mkdir build
cd build
cmake ..
make -j 10

完成编译代码

自定义yaml配置文件

进入到data目录中，新建一个自己的数据目录，copy person目录中的yaml配置文件到自己目录中，修改其中内容：

# 以下所有与路径相关的配置文件的根目录， 左斜杠结尾
path: "../data/person/"
# 模型文件路径，trt若不存在，会使用onnx生成对应trt
model:
  onnx: "person.onnx"
  tensorrt: "person.trt"
# 输入测试图片，与结果保存路径
image:
  input: "person.jpg"
  output: "person_result.jpg"
# 输出类别，names文件路径
args:
  n_classes: 1
  names: "person.names"
  # yolov5输入图片尺寸
  channels: 3
  height: 640
  width: 640
  # 如果需要dynamic支持，需要配置下面：三个尺寸，最小，opt，最大，开启ifdynamic为1，否则为0
  ifdynamic: 1
  min: [1, 3, 640, 640]
  opt: [1, 3, 640, 640]
  max: [1, 3, 640, 640]
# 运行参数， demo运行测试，若trt不存在，会使用onnx默认生成fp32模型
# fp16表示使用fp16生成模型，int8同理， 取值{0, 1}
build:
  demo: 1
  fp16: 1
  int8: 0
  # 工作空间大小，单位M，暂时不支持修改
  workspace: 4
# nms 插件配置
nms:
  topK: 512
  keepTopK: 100
?
  # 以下nms相关参数，暂时不支持修改，使用下面列出的默认配置
  clipBoxes: 0
  iouThreshold: 0.25
  scoreThreshold: 0.45
  isNormalized: false
  output: ["num_detections", "nmsed_boxes", "nmsed_scores", "nmsed_classes"]

yaml中注释足够详细，简单说下用法，以需要检测person为例：

1. 在data中，新建person目录（随便命名），复制yaml到person目录， 其中，yaml中的 path: "../data/person/" 路径中person为自己命名的路径名字；

2. 为检测标签新建一个names文件，命名为person.names, 并修改yaml中args.names项为对应名字

3. onnx文件同上

4. 其他设置参见yaml中的注释

运行代码

进入到build文件夹中，执行：

./yolov5_trt -c ../data/person/person.yaml

最后参数为指定yaml配置文件的路径，即可

检测结果如下：