libtorch+yolov5

前言

这是本人第一篇博客，只是对近期学习工作的一些总结。主要是利用libtorch对pytorch训练的模型进行部署，之前也是成功使用pyinstaller将整个python项目进行打包成exe，但是不满足对方的需求才使用的libtorch。

环境配置

vs2019+opencv4.5+libtorch1.7.1：
1.vs2019下载：链接: link
2. opencv官网：链接: link
3. Libtorch下载：链接: link推荐下载release版本（版本需要与训练模型pytorch版本符合，cuda版本需要相符）

环境搭建参考：

https://blog.csdn.net/weixin_44936889/article/details/111186818
https://blog.csdn.net/zzz_zzz12138/article/details/109138805
https://blog.csdn.net/wenghd22/article/details/112512231
vs2019和opencv配置过程参考链接；
https://blog.csdn.net/sophies671207/article/details/89854368

给出我的libtorch配置

给出我的libtorch配置过程：新建空项目 新建main.cpp文件
1、新建项目->属性->VC++目录->包含目录

2新建项目->属性->VC++目录->库目录

3新建项目->属性->C/C++目录->常规->附加包含目录

4新建项目->属性->C/C++目录->常规->SDL检查 ：否

5新建项目->属性->连接器->输入->附加依赖项：写入以下
E:\opencv\build\x64\vc15\lib\opencv_world450.lib
c10.lib
asmjit.lib
c10_cuda.lib
caffe2_detectron_ops_gpu.lib
caffe2_module_test_dynamic.lib
caffe2_nvrtc.lib
clog.lib
cpuinfo.lib
dnnl.lib
fbgemm.lib
libprotobuf.lib
libprotobuf-lite.lib
libprotoc.lib
mkldnn.lib
torch.lib
torch_cuda.lib
torch_cpu.lib
kernel32.lib
user32.lib
gdi32.lib
winspool.lib
comdlg32.lib
advapi32.lib
shell32.lib
ole32.lib
oleaut32.lib
uuid.lib
odbc32.lib
odbccp32.lib

6新建项目->属性->连接器->命令行：输入/INCLUDE:?warp_size@cuda@at@@YAHXZ

7新建项目->属性->C/C++目录->语言->符合模式 ：否

配置好了以上环境，打包好的文件夹如下图：
权重文件：官方权重导出固定尺度模型即可 直接运行main.cpp即可。采用的samples文件夹下的图片进行测试。
main.cpp代码 十六行使用GPU时需注意

#include 
#include 
#include 
#include 
#include 
#include 

std::vector<torch::Tensor> non_max_suppression(torch::Tensor preds, float score_thresh = 0.01, float iou_thresh = 0.35)
{
     
    std::vector<torch::Tensor> output;
    for (size_t i = 0; i < preds.sizes()[0]; ++i)
    {
     
        torch::Tensor pred = preds.select(0, i);

        //GPU推理结果为cuda数据类型，nms之前要转成cpu，否则会报错
        pred = pred.to(at::kCPU); //增加到函数里pred = pred.to(at::kCPU); 注意preds的数据类型，转成cpu进行后处理。

        // Filter by scores
        torch::Tensor scores = pred.select(1, 4) * std::get<0>(torch::max(pred.slice(1, 5, pred.sizes()[1]), 1));
        pred = torch::index_select(pred, 0, torch::nonzero(scores > score_thresh).select(1, 0));
        if (pred.sizes()[0] == 0) continue;

        // (center_x, center_y, w, h) to (left, top, right, bottom)
        pred.select(1, 0) = pred.select(1, 0) - pred.select(1, 2) / 2;
        pred.select(1, 1) = pred.select(1, 1) - pred.select(1, 3) / 2;
        pred.select(1, 2) = pred.select(1, 0) + pred.select(1, 2);
        pred.select(1, 3) = pred.select(1, 1) + pred.select(1, 3);

        // Computing scores and classes
        std::tuple<torch::Tensor, torch::Tensor> max_tuple = torch::max(pred.slice(1, 5, pred.sizes()[1]), 1);
        pred.select(1, 4) = pred.select(1, 4) * std::get<0>(max_tuple);
        pred.select(1, 5) = std::get<1>(max_tuple);

        torch::Tensor  dets = pred.slice(1, 0, 6);

        torch::Tensor keep = torch::empty({
      dets.sizes()[0] });
        torch::Tensor areas = (dets.select(1, 3) - dets.select(1, 1)) * (dets.select(1, 2) - dets.select(1, 0));
        std::tuple<torch::Tensor, torch::Tensor> indexes_tuple = torch::sort(dets.select(1, 4), 0, 1);
        torch::Tensor v = std::get<0>(indexes_tuple);
        torch::Tensor indexes = std::get<1>(indexes_tuple);
        int count = 0;
        while (indexes.sizes()[0] > 0)
        {
     
            keep[count] = (indexes[0].item().toInt());
            count += 1;

            // Computing overlaps
            torch::Tensor lefts = torch::empty(indexes.sizes()[0] - 1);
            torch::Tensor tops = torch::empty(indexes.sizes()[0] - 1);
            torch::Tensor rights = torch::empty(indexes.sizes()[0] - 1);
            torch::Tensor bottoms = torch::empty(indexes.sizes()[0] - 1);
            torch::Tensor widths = torch::empty(indexes.sizes()[0] - 1);
            torch::Tensor heights = torch::empty(indexes.sizes()[0] - 1);
            for (size_t i = 0; i < indexes.sizes()[0] - 1; ++i)
            {
     
                lefts[i] = std::max(dets[indexes[0]][0].item().toFloat(), dets[indexes[i + 1]][0].item().toFloat());
                tops[i] = std::max(dets[indexes[0]][1].item().toFloat(), dets[indexes[i + 1]][1].item().toFloat());
                rights[i] = std::min(dets[indexes[0]][2].item().toFloat(), dets[indexes[i + 1]][2].item().toFloat());
                bottoms[i] = std::min(dets[indexes[0]][3].item().toFloat(), dets[indexes[i + 1]][3].item().toFloat());
                widths[i] = std::max(float(0), rights[i].item().toFloat() - lefts[i].item().toFloat());
                heights[i] = std::max(float(0), bottoms[i].item().toFloat() - tops[i].item().toFloat());
            }
            torch::Tensor overlaps = widths * heights;

            // FIlter by IOUs
            torch::Tensor ious = overlaps / (areas.select(0, indexes[0].item().toInt()) + torch::index_select(areas, 0, indexes.slice(0, 1, indexes.sizes()[0])) - overlaps);
            indexes = torch::index_select(indexes, 0, torch::nonzero(ious <= iou_thresh).select(1, 0) + 1);
        }
        keep = keep.toType(torch::kInt64);
        output.push_back(torch::index_select(dets, 0, keep.slice(0, 0, count)));
    }
    return output;
}

#include  
#include 
#include 
//int main(int argc, const char* argv[]) {
     
//    std::cout << "cuda::is_available():" << torch::cuda::is_available() << std::endl;
//    torch::DeviceType device_type = at::kCPU; // 定义设备类型
//    if (torch::cuda::is_available())
//        device_type = at::kCUDA;
//}


int main(int argc, char* argv[])
{
     
    std::cout << "cuda::is_available():" << torch::cuda::is_available() << std::endl;
    torch::DeviceType device_type = at::kCPU; // 定义设备类型
    if (torch::cuda::is_available())
        device_type = at::kCUDA;
    // Loading  Module
    torch::jit::script::Module module = torch::jit::load("yolov5x.torchscript.pt");//best.torchscript3.pt//yolov5x.torchscript.pt
    module.to(device_type); // 模型加载至GPU

    std::vector<std::string> classnames;
    std::ifstream f("coco.names");
    std::string name = "";
    while (std::getline(f, name))
    {
     
        classnames.push_back(name);
    }
    if (argc < 2)
    {
     
        std::cout << "Please run with test video." << std::endl;
        return -1;
    }
    std::string video = argv[1];
    cv::VideoCapture cap = cv::VideoCapture(video);
    // cap.set(cv::CAP_PROP_FRAME_WIDTH, 1920);
    // cap.set(cv::CAP_PROP_FRAME_HEIGHT, 1080);
    cv::Mat frame, img;
    cap.read(frame);
    int width = frame.size().width;
    int height = frame.size().height;
    int count = 0;
    while (cap.isOpened())
    {
     
        count++;
        clock_t start = clock();
        cap.read(frame);
        if (frame.empty())
        {
     
            std::cout << "Read frame failed!" << std::endl;
            break;
        }

        // Preparing input tensor
        cv::resize(frame, img, cv::Size(640, 640));
        // cv::cvtColor(img, img, cv::COLOR_BGR2RGB);
        // torch::Tensor imgTensor = torch::from_blob(img.data, {img.rows, img.cols,3},torch::kByte);
        // imgTensor = imgTensor.permute({2,0,1});
        // imgTensor = imgTensor.toType(torch::kFloat);
        // imgTensor = imgTensor.div(255);
        // imgTensor = imgTensor.unsqueeze(0);
        // imgTensor = imgTensor.to(device_type);
        cv::cvtColor(img, img, cv::COLOR_BGR2RGB);  // BGR -> RGB
        img.convertTo(img, CV_32FC3, 1.0f / 255.0f);  // normalization 1/255
        auto imgTensor = torch::from_blob(img.data, {
      1, img.rows, img.cols, img.channels() }).to(device_type);
        imgTensor = imgTensor.permute({
      0, 3, 1, 2 }).contiguous();  // BHWC -> BCHW (Batch, Channel, Height, Width)
        std::vector<torch::jit::IValue> inputs;
        inputs.emplace_back(imgTensor);
        // preds: [?, 15120, 9]
        torch::jit::IValue output = module.forward(inputs);
        auto preds = output.toTuple()->elements()[0].toTensor();
        // torch::Tensor preds = module.forward({ imgTensor }).toTensor();
        std::vector<torch::Tensor> dets = non_max_suppression(preds, 0.35, 0.5);
        if (dets.size() > 0)
        {
     
            // Visualize result
            for (size_t i = 0; i < dets[0].sizes()[0]; ++i)
            {
     
                float left = dets[0][i][0].item().toFloat() * frame.cols / 640;
                float top = dets[0][i][1].item().toFloat() * frame.rows / 640;
                float right = dets[0][i][2].item().toFloat() * frame.cols / 640;
                float bottom = dets[0][i][3].item().toFloat() * frame.rows / 640;
                float score = dets[0][i][4].item().toFloat();
                int classID = dets[0][i][5].item().toInt();

                cv::rectangle(frame, cv::Rect(left, top, (right - left), (bottom - top)), cv::Scalar(0, 255, 0), 2);

                cv::putText(frame,
                    classnames[classID] + ": " + cv::format("%.2f", score),
                    cv::Point(left, top),
                    cv::FONT_HERSHEY_SIMPLEX, (right - left) / 200, cv::Scalar(0, 255, 0), 2);
            }
        }
        // std::cout << "-[INFO] Frame:" <<  std::to_string(count) << " FPS: " + std::to_string(float(1e7 / (clock() - start))) << std::endl;
        std::cout << "-[INFO] Frame:" << std::to_string(count) << std::endl;
        // cv::putText(frame, "FPS: " + std::to_string(int(1e7 / (clock() - start))),
        //     cv::Point(50, 50),
        //     cv::FONT_HERSHEY_SIMPLEX, 1, cv::Scalar(0, 255, 0), 2);
        cv::imshow("", frame);
        // cv::imwrite("../images/"+cv::format("%06d", count)+".jpg", frame);
        cv::resize(frame, frame, cv::Size(width, height));
        if (cv::waitKey(1) == 27) break;
    }
    cap.release();
    return 0;
}

GPU模型 导出 export代码

注意修改导出模型尺度。

"""Exports a YOLOv5 *.pt model to ONNX and TorchScript formats

Usage:
    $ export PYTHONPATH="$PWD" && python models/export.py --weights ./weights/yolov5s.pt --img 640 --batch 1
"""

import argparse
import sys
import time

sys.path.append('./')  # to run '$ python *.py' files in subdirectories

import torch
import torch.nn as nn

import models
from models.experimental import attempt_load
from utils.activations import Hardswish, SiLU
from utils.general import set_logging, check_img_size

if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--weights', type=str, default='E:\\yolov5-master\\runs\\train\\exp\weights\\best.pt', help='weights path')  # from yolov5/models/
    parser.add_argument('--img-size', nargs='+', type=int, default=[352, 640], help='image size')  # height, width
    parser.add_argument('--batch-size', type=int, default=1, help='batch size')
    opt = parser.parse_args()
    opt.img_size *= 2 if len(opt.img_size) == 1 else 1  # expand
    print(opt)
    set_logging()
    t = time.time()

    # Load PyTorch model
    model = attempt_load(opt.weights, map_location=torch.device('cuda'))  # load FP32 model
    labels = model.names

    # Checks
    gs = int(max(model.stride))  # grid size (max stride)
    opt.img_size = [check_img_size(x, gs) for x in opt.img_size]  # verify img_size are gs-multiples

    # Input
    img = torch.zeros(opt.batch_size, 3, *opt.img_size).to(device='cuda')
    # image size(1,3,320,192) iDetection

    # Update model
    for k, m in model.named_modules():
        m._non_persistent_buffers_set = set()  # pytorch 1.6.0 compatibility
        if isinstance(m, models.common.Conv):  # assign export-friendly activations
            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.forward = m.forward_export  # assign forward (optional)
    #model.model[-1].export = True  # set Detect() layer export=True
    model.model[-1].export = False
    y = model(img)  # dry run

    # TorchScript export
    try:
        print('\nStarting TorchScript export with torch %s...' % torch.__version__)
        f = opt.weights.replace('.pt', '.torchscript.pt')  # filename
        ts = torch.jit.trace(model, img)
        ts.save(f)
        print('TorchScript export success, saved as %s' % f)
    except Exception as e:
        print('TorchScript export failure: %s' % e)

    # ONNX export
    try:
        import onnx

        print('\nStarting ONNX export with onnx %s...' % onnx.__version__)
        f = opt.weights.replace('.pt', '.onnx')  # filename
        torch.onnx.export(model, img, f, verbose=False, opset_version=12, input_names=['images'],
                          output_names=['classes', 'boxes'] if y is None else ['output'])

        # Checks
        onnx_model = onnx.load(f)  # load onnx model
        onnx.checker.check_model(onnx_model)  # check onnx model
        # print(onnx.helper.printable_graph(onnx_model.graph))  # print a human readable model
        print('ONNX export success, saved as %s' % f)
    except Exception as e:
        print('ONNX export failure: %s' % e)

    # CoreML export
    try:
        import coremltools as ct

        print('\nStarting CoreML export with coremltools %s...' % ct.__version__)
        # convert model from torchscript and apply pixel scaling as per detect.py
        model = ct.convert(ts, inputs=[ct.ImageType(name='image', shape=img.shape, scale=1 / 255.0, bias=[0, 0, 0])])
        f = opt.weights.replace('.pt', '.mlmodel')  # filename
        model.save(f)
        print('CoreML export success, saved as %s' % f)
    except Exception as e:
        print('CoreML export failure: %s' % e)

    # Finish
    print('\nExport complete (%.2fs). Visualize with https://github.com/lutzroeder/netron.' % (time.time() - t))

效果展示

测试结果：

视频测试：

2021-01-20 14-31-49

libtorch+yolov5

另一个视频链接: link

结束

自己训练的模型效果就不展示了，感谢大家观看，请多多关注，共同学习进步！

###问题 关于forward耗时很长的问题 其实早在我批处理多张图片的时候发现了 如下图：前两张很慢 后面正常。自己查过也在GitHub咨询过 原因可能是libtorch1.7.1版本存在warm up的问题。