实战YOLO V5推演（TensorRT版本 C++）

发布时间：2023-03-31 文章分类：电脑基础 投稿人：樱花 字号： 默认 | 大 | 超大 打印

兜兜转转还是来到了C++，欠的债该还的得还，因此本篇文章试图从C++来撬动tensorrt 完成转换模型和推理，而不是借助python库

文章目录

  • 前言
  • 一、Tensorrt前置
      • 1.1 Tensorrt基本概念
      • 1.2 异步推演说明
  • 二、实战YOLO V5推演
    • 1.TensorRT模型转换
    • 所以要替换模型了，明天见！
    • 分割线,mmp今天阳性了,明天请了天假,今明两天搞定它!!!!
    • 3. 再来一遍
  • 总结

前言

希望本篇文章结束后，我能回答两个问题：

  1. python不好用么，为什么还要转C++； yolov5的export.py 已经实现在python端的模型转化了啊？
    不需要c++的可以看这个《【深度学习】目标检测 yolov5模型量化安装教程以及转ONXX，torchscript，engine和速度比较一栏表》
  2. C++的学习曲线那么陡峭，性能在服务器端不如go，所以为什么不是go？
    go 在服务器端性能牛，但是在终端设备不咋样，所以从这个角度，还是C++

一、Tensorrt前置

首先需要有个环境，搭建环境的指引《【深度学习】基于深度学习的linux服务器，需要搭建哪些服务，一步步搭建深度学习的环境，cuda，pytorch，opencv，ftp服务， nfs服务 docker等等》
文章很长，是深度学习的最大化安装，基于进入Tensorrt的操作，至少要安装显卡驱动，cuda，cudann，tensorrt，gpu的torch环境，并且他们之间的版本要匹配。老实说，新手直接绕道《【深度学习】目标检测 yolov5模型量化安装教程以及转ONXX，torchscript，engine和速度比较一栏表》
读到后面你会发现,我在模型转换的时候,踩了yolov的各种坑,你会看到我用yolov5的 tag7.0 tag6.1 tag6.2 都有问题.
尚不清楚是pytorch的版本问题,还是tensorrt的版本问题,我再量化那篇文章用的是pytorch1.8 没遇到问题.

1.1 Tensorrt基本概念

本章节你将了解TensorRT中的以下几个关键概念：Runtime、CudaEngine、ExecutionContext、与Binding。

在使用ONNX Runtime的Python API执行模型推演大致分四步：

  1. 初始化：使用onnx格式存储的模型文件初始化 InferenceSession 。

  2. 预处理：将输入数据做预处理。

  3. 执行推演：调用 run 方法执行模型推演，获得模型输出。

  4. 后处理：执行后处理逻辑。
    ‌
    TensorRT的推演与其类似，只是要更加复杂一些：

  5. 模型准备：TensorRT使用私有模型格式，其它格式的模型文件如ONNX需要先完成转换才可以被使用。该转换过程可以通过调 用TensorRT的C++或者Python API来在线实现，也可以通过 trtexec 这个工具离线完成。

  6. 初始化：包括初始化运行时 Runtime ，使用 Runtime 加载模型文件形成 CudaEngine ，再基于 CudaEngine 创建 ExecutionContext ；这三者组合起来可以大致等同为 ONNX Runtime中的 InferenceSession 。此外还需要按照模型的输入与输出规格分别申请CPU一侧的内存与GPU一侧的显存，指向这些存储区域的指针集合称为 Binding 。

  7. 将输入数据做预处理，然后将其从CPU侧存储（内存）拷贝到GPU侧存储（显存）。

  8. 调用 ExecutionContext 的enqueueV2 方法异步执行模型推演，然后调用 cudaStreamSynchronize 方法同步结果，获得模型输出。

  9. 将输出结果从GPU侧存储（显存）拷贝到CPU侧存储（内存），然后执行后处理逻辑。

1.2 异步推演说明

TensorRT支持同步与异步推演，这里以异步推演作为说明实例。

与ONNX Runtime Python API的推演流程相比，TensorRT C++ API推演明显有着更加底层、更加细粒度的数据与流程控制。这些正是使用C++作为开发语言与Python相比所带来的主要区别。

接下来，你将通过YOLO v5这个实例来掌握TensorRT的实际使用过程。本章节所涉及到的概念都会出现在代码中，希望到时候你看到它们时不会感觉太陌生。

二、实战YOLO V5推演

本章节你将借助YOLO v5模型推演的完整代码掌握TensorRT C++ 推演API，以及将ONNX格式神经网络模型转换为TensorRT专用格式的具体方法。

YOLO v5是一款在物体检测领域颇为流行的神经网络模型。在这里我们不会深入探究其原理与设计，而将直接使用GitHub上的一款开源实现所生成的预训练模型。在此基础上，本章节展示了以下内容。

  1. 使用TensorRT的 trtexec 工具将ONNX模型文件转换为TensorRT的Engine格式。
  2. 使用CMake构建TensorRT工程。
  3. 使用OpenCV实现输入图像读取与存储。
  4. YOLO v5模型预处理与后处理。
  5. 使用TensorRT实现YOLO v5模型推演。
    让我们开始吧！

1.TensorRT模型转换

首先下载YOLO v5预训练好的模型权重文件，其地址为：

https://github.com/ultralytics/yolov5/releases/download/v6.1/yolov5s.pt

然后遵循官方教程，导出ONNX格式模型：
代码如下（示例）：
在写一次吧，安装

pip install -r requirements.txt coremltools onnx onnx-simplifier onnxruntime-gpu openvino-dev tensorflow  # GPU

其实只需要onnx相关的，其它的有tf使用的，apple使用的opencv使用的，用不到 ，我这里偷个懒，安装少的

pip install -r requirements.txt onnx onnxruntime-gpu # GPU

上边这一步完成后，你应当获得一个ONNX格式模型文件 yolov5s.onnx 。
转换一版试一下：

 python export.py --weights weights/yolov5s.pt -- --include onnx engine

转化报错，原因是:https://onnxruntime.ai/docs/execution-providers/CUDA-ExecutionProvider.html,默认安装的1.13，匹配11.6的吗我的是11.4，换回来重新运行
实战YOLO V5推演（TensorRT版本 C++）

接下来，使用如下命令将其转换为TensoRT格式模型文件。
还是报错

[8] Assertion failed: scales.is_weights() && "Resize scales must be an initializer!"
TensorRT: export failure ❌ 4.3s: failed to load ONNX file: weights/yolov5s.onnx

解决问题指引：
https://blog.csdn.net/zaf0516/article/details/122932651

snap install netron
pip install onnx-simplifier
pip install onnx_graphsurgeon --index-url https://pypi.ngc.nvidia.com

还是报错：
指引是改模型，我偷个懒，退一下版本 yolov5 tag6.2搞下来试一试，现在用的是最新的用于分割的tag7.0

实战YOLO V5推演（TensorRT版本 C++）
都是相同的错误，先不改了，直接转onnx是可以的

python export.py --weights weights/yolov5s.pt --device 0 --include onnx

切換到6.1 上面報錯，解決指引
https://blog.csdn.net/Thebest_jack/article/details/124723687
然后转换engine

trtexec --onnx=weights/yolov5s.onnx --saveEngine=yolov5s.engine

问题出来了，不管用7.0 还是 6.1， 6.2 都报错


[12/07/2022-20:49:09] [E] [TRT] ModelImporter.cpp:776: --- End node ---
[12/07/2022-20:49:09] [E] [TRT] ModelImporter.cpp:779: ERROR: builtin_op_importers.cpp:3609 In function importResize:
[8] Assertion failed: scales.is_weights() && "Resize scales must be an initializer!"
[12/07/2022-20:49:09] [E] Failed to parse onnx file
[12/07/2022-20:49:09] [I] Finish parsing network model
[12/07/2022-20:49:09] [E] Parsing model failed
[12/07/2022-20:49:09] [E] Failed to create engine from model.
[12/07/2022-20:49:09] [E] Engine set up failed
&&&& FAILED TensorRT.trtexec [TensorRT v8205] # trtexec --onnx=weights/yolov5s.onnx --saveEngine=yolov5s.engine

python 命令和trtexec命令都报错

 trtexec --onnx=weights/yolov5s.onnx --saveEngine=yolov5s.engine
[12/07/2022-20:53:43] [E] [TRT] ModelImporter.cpp:776: --- End node ---
[12/07/2022-20:53:43] [E] [TRT] ModelImporter.cpp:779: ERROR: builtin_op_importers.cpp:3609 In function importResize:
[8] Assertion failed: scales.is_weights() && "Resize scales must be an initializer!"
[12/07/2022-20:53:43] [E] Failed to parse onnx file
[12/07/2022-20:53:43] [I] Finish parsing network model
[12/07/2022-20:53:43] [E] Parsing model failed
[12/07/2022-20:53:43] [E] Failed to create engine from model.
[12/07/2022-20:53:43] [E] Engine set up failed
&&&& FAILED TensorRT.trtexec [TensorRT v8205] # trtexec --onnx=weights/yolov5s.onnx --saveEngine=yolov5s.engine

python export.py --weights weights/yolov5s.pt --device 0 --include onnx engine

所以要替换模型了，明天见！

分割线,mmp今天阳性了,明天请了天假,今明两天搞定它!!!

restart 还是用yolov5 tag7.0
环境:
torch 1.12.1
torchaudio 0.12.1
torchvision 0.13.1
NVIDIA-SMI 470.63.01 Driver Version: 470.63.01 CUDA Version: 11.4
cuda:11.4
cudnn: cudann:cudnn-11.4-linux-x64-v8.2.4.15.tgz 8.2.4
tensorrt: 8.2.5.1

我会以这个环境为起点,试图debug,让他在yolov5的项目包python export.py 转onnx 和 engine都成功!
错误再现

(base) [root@localhost road_water_yolov5-7.0]# python export.py --weights weights/yolov5s.pt --device 0 --include onnx engine
export: data=data/coco128.yaml, weights=['weights/yolov5s.pt'], imgsz=[640, 640], batch_size=1, device=0, half=False, inplace=False, keras=False, optimize=False, int8=False, dynamic=False, simplify=False, opset=12, verbose=False, workspace=4, nms=False, agnostic_nms=False, topk_per_class=100, topk_all=100, iou_thres=0.45, conf_thres=0.25, include=['onnx', 'engine']
Unknown option: -C
usage: git [--version] [--help] [-c name=value]
           [--exec-path[=<path>]] [--html-path] [--man-path] [--info-path]
           [-p|--paginate|--no-pager] [--no-replace-objects] [--bare]
           [--git-dir=<path>] [--work-tree=<path>] [--namespace=<name>]
           <command> [<args>]
YOLOv5 🚀 2022-12-2 Python-3.9.12 torch-1.12.1 CUDA:0 (NVIDIA GeForce RTX 3080 Ti, 12054MiB)
Fusing layers...
YOLOv5s summary: 213 layers, 7225885 parameters, 0 gradients, 16.4 GFLOPs
PyTorch: starting from weights/yolov5s.pt with output shape (1, 25200, 85) (14.1 MB)
ONNX: starting export with onnx 1.12.0...
ONNX: export success ✅ 2.7s, saved as weights/yolov5s.onnx (28.0 MB)
TensorRT: starting export with TensorRT 8.2.5.1...
[12/08/2022-18:42:33] [TRT] [I] [MemUsageChange] Init CUDA: CPU +466, GPU +0, now: CPU 3433, GPU 1917 (MiB)
[12/08/2022-18:42:34] [TRT] [I] [MemUsageSnapshot] Begin constructing builder kernel library: CPU 3433 MiB, GPU 1917 MiB
[12/08/2022-18:42:34] [TRT] [I] [MemUsageSnapshot] End constructing builder kernel library: CPU 3586 MiB, GPU 1961 MiB
[12/08/2022-18:42:34] [TRT] [I] ----------------------------------------------------------------
[12/08/2022-18:42:34] [TRT] [I] Input filename:   weights/yolov5s.onnx
[12/08/2022-18:42:34] [TRT] [I] ONNX IR version:  0.0.7
[12/08/2022-18:42:34] [TRT] [I] Opset version:    12
[12/08/2022-18:42:34] [TRT] [I] Producer name:    pytorch
[12/08/2022-18:42:34] [TRT] [I] Producer version: 1.12.1
[12/08/2022-18:42:34] [TRT] [I] Domain:
[12/08/2022-18:42:34] [TRT] [I] Model version:    0
[12/08/2022-18:42:34] [TRT] [I] Doc string:
[12/08/2022-18:42:34] [TRT] [I] ----------------------------------------------------------------
[12/08/2022-18:42:34] [TRT] [W] onnx2trt_utils.cpp:366: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32.
[12/08/2022-18:42:34] [TRT] [E] ModelImporter.cpp:773: While parsing node number 141 [Resize -> "onnx::Concat_271"]:
[12/08/2022-18:42:34] [TRT] [E] ModelImporter.cpp:774: --- Begin node ---
[12/08/2022-18:42:34] [TRT] [E] ModelImporter.cpp:775: input: "onnx::Resize_266"
input: "onnx::Resize_270"
input: "onnx::Resize_445"
output: "onnx::Concat_271"
name: "Resize_141"
op_type: "Resize"
attribute {
  name: "coordinate_transformation_mode"
  s: "asymmetric"
  type: STRING
}
attribute {
  name: "cubic_coeff_a"
  f: -0.75
  type: FLOAT
}
attribute {
  name: "mode"
  s: "nearest"
  type: STRING
}
attribute {
  name: "nearest_mode"
  s: "floor"
  type: STRING
}
[12/08/2022-18:42:34] [TRT] [E] ModelImporter.cpp:776: --- End node ---
[12/08/2022-18:42:34] [TRT] [E] ModelImporter.cpp:779: ERROR: builtin_op_importers.cpp:3609 In function importResize:
[8] Assertion failed: scales.is_weights() && "Resize scales must be an initializer!"
TensorRT: export failure ❌ 4.1s: failed to load ONNX file: weights/yolov5s.onnx
ONNX: starting export with onnx 1.12.0...
ONNX: export success ✅ 2.6s, saved as weights/yolov5s.onnx (28.0 MB)
Export complete (11.8s)
Results saved to /root/workplace/road_water_yolov5-7.0/weights
Detect:          python detect.py --weights weights/yolov5s.onnx
Validate:        python val.py --weights weights/yolov5s.onnx
PyTorch Hub:     model = torch.hub.load('ultralytics/yolov5', 'custom', 'weights/yolov5s.onnx')
Visualize:       https://netron.app

onnx 转换成功了,.engine没有
用c++的方式转换:

(base) [root@localhost road_water_yolov5-7.0]#  trtexec --onnx=weights/yolov5s.onnx --saveEngine=yolov5s.engine
&&&& RUNNING TensorRT.trtexec [TensorRT v8205] # trtexec --onnx=weights/yolov5s.onnx --saveEngine=yolov5s.engine
[12/08/2022-18:48:42] [I] === Model Options ===
[12/08/2022-18:48:42] [I] Format: ONNX
[12/08/2022-18:48:42] [I] Model: weights/yolov5s.onnx
[12/08/2022-18:48:42] [I] Output:
[12/08/2022-18:48:42] [I] === Build Options ===
[12/08/2022-18:48:42] [I] Max batch: explicit batch
[12/08/2022-18:48:42] [I] Workspace: 16 MiB
[12/08/2022-18:48:42] [I] minTiming: 1
[12/08/2022-18:48:42] [I] avgTiming: 8
[12/08/2022-18:48:42] [I] Precision: FP32
[12/08/2022-18:48:42] [I] Calibration:
[12/08/2022-18:48:42] [I] Refit: Disabled
[12/08/2022-18:48:42] [I] Sparsity: Disabled
[12/08/2022-18:48:42] [I] Safe mode: Disabled
[12/08/2022-18:48:42] [I] DirectIO mode: Disabled
[12/08/2022-18:48:42] [I] Restricted mode: Disabled
[12/08/2022-18:48:42] [I] Save engine: yolov5s.engine
[12/08/2022-18:48:42] [I] Load engine:
[12/08/2022-18:48:42] [I] Profiling verbosity: 0
[12/08/2022-18:48:42] [I] Tactic sources: Using default tactic sources
[12/08/2022-18:48:42] [I] timingCacheMode: local
[12/08/2022-18:48:42] [I] timingCacheFile:
[12/08/2022-18:48:42] [I] Input(s)s format: fp32:CHW
[12/08/2022-18:48:42] [I] Output(s)s format: fp32:CHW
[12/08/2022-18:48:42] [I] Input build shapes: model
[12/08/2022-18:48:42] [I] Input calibration shapes: model
[12/08/2022-18:48:42] [I] === System Options ===
[12/08/2022-18:48:42] [I] Device: 0
[12/08/2022-18:48:42] [I] DLACore:
[12/08/2022-18:48:42] [I] Plugins:
[12/08/2022-18:48:42] [I] === Inference Options ===
[12/08/2022-18:48:42] [I] Batch: Explicit
[12/08/2022-18:48:42] [I] Input inference shapes: model
[12/08/2022-18:48:42] [I] Iterations: 10
[12/08/2022-18:48:42] [I] Duration: 3s (+ 200ms warm up)
[12/08/2022-18:48:42] [I] Sleep time: 0ms
[12/08/2022-18:48:42] [I] Idle time: 0ms
[12/08/2022-18:48:42] [I] Streams: 1
[12/08/2022-18:48:42] [I] ExposeDMA: Disabled
[12/08/2022-18:48:42] [I] Data transfers: Enabled
[12/08/2022-18:48:42] [I] Spin-wait: Disabled
[12/08/2022-18:48:42] [I] Multithreading: Disabled
[12/08/2022-18:48:42] [I] CUDA Graph: Disabled
[12/08/2022-18:48:42] [I] Separate profiling: Disabled
[12/08/2022-18:48:42] [I] Time Deserialize: Disabled
[12/08/2022-18:48:42] [I] Time Refit: Disabled
[12/08/2022-18:48:42] [I] Skip inference: Disabled
[12/08/2022-18:48:42] [I] Inputs:
[12/08/2022-18:48:42] [I] === Reporting Options ===
[12/08/2022-18:48:42] [I] Verbose: Disabled
[12/08/2022-18:48:42] [I] Averages: 10 inferences
[12/08/2022-18:48:42] [I] Percentile: 99
[12/08/2022-18:48:42] [I] Dump refittable layers:Disabled
[12/08/2022-18:48:42] [I] Dump output: Disabled
[12/08/2022-18:48:42] [I] Profile: Disabled
[12/08/2022-18:48:42] [I] Export timing to JSON file:
[12/08/2022-18:48:42] [I] Export output to JSON file:
[12/08/2022-18:48:42] [I] Export profile to JSON file:
[12/08/2022-18:48:42] [I]
[12/08/2022-18:48:42] [I] === Device Information ===
[12/08/2022-18:48:42] [I] Selected Device: NVIDIA GeForce RTX 3080 Ti
[12/08/2022-18:48:42] [I] Compute Capability: 8.6
[12/08/2022-18:48:42] [I] SMs: 80
[12/08/2022-18:48:42] [I] Compute Clock Rate: 1.665 GHz
[12/08/2022-18:48:42] [I] Device Global Memory: 12053 MiB
[12/08/2022-18:48:42] [I] Shared Memory per SM: 100 KiB
[12/08/2022-18:48:42] [I] Memory Bus Width: 384 bits (ECC disabled)
[12/08/2022-18:48:42] [I] Memory Clock Rate: 9.501 GHz
[12/08/2022-18:48:42] [I]
[12/08/2022-18:48:42] [I] TensorRT version: 8.2.5
[12/08/2022-18:48:43] [I] [TRT] [MemUsageChange] Init CUDA: CPU +474, GPU +0, now: CPU 486, GPU 491 (MiB)
[12/08/2022-18:48:43] [I] [TRT] [MemUsageSnapshot] Begin constructing builder kernel library: CPU 486 MiB, GPU 491 MiB
[12/08/2022-18:48:43] [I] [TRT] [MemUsageSnapshot] End constructing builder kernel library: CPU 640 MiB, GPU 533 MiB
[12/08/2022-18:48:43] [I] Start parsing network model
[12/08/2022-18:48:43] [I] [TRT] ----------------------------------------------------------------
[12/08/2022-18:48:43] [I] [TRT] Input filename:   weights/yolov5s.onnx
[12/08/2022-18:48:43] [I] [TRT] ONNX IR version:  0.0.7
[12/08/2022-18:48:43] [I] [TRT] Opset version:    12
[12/08/2022-18:48:43] [I] [TRT] Producer name:    pytorch
[12/08/2022-18:48:43] [I] [TRT] Producer version: 1.12.1
[12/08/2022-18:48:43] [I] [TRT] Domain:
[12/08/2022-18:48:43] [I] [TRT] Model version:    0
[12/08/2022-18:48:43] [I] [TRT] Doc string:
[12/08/2022-18:48:43] [I] [TRT] ----------------------------------------------------------------
[12/08/2022-18:48:43] [W] [TRT] onnx2trt_utils.cpp:366: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32.
[12/08/2022-18:48:43] [E] [TRT] ModelImporter.cpp:773: While parsing node number 141 [Resize -> "onnx::Concat_271"]:
[12/08/2022-18:48:43] [E] [TRT] ModelImporter.cpp:774: --- Begin node ---
[12/08/2022-18:48:43] [E] [TRT] ModelImporter.cpp:775: input: "onnx::Resize_266"
input: "onnx::Resize_270"
input: "onnx::Resize_445"
output: "onnx::Concat_271"
name: "Resize_141"
op_type: "Resize"
attribute {
  name: "coordinate_transformation_mode"
  s: "asymmetric"
  type: STRING
}
attribute {
  name: "cubic_coeff_a"
  f: -0.75
  type: FLOAT
}
attribute {
  name: "mode"
  s: "nearest"
  type: STRING
}
attribute {
  name: "nearest_mode"
  s: "floor"
  type: STRING
}
[12/08/2022-18:48:43] [E] [TRT] ModelImporter.cpp:776: --- End node ---
[12/08/2022-18:48:43] [E] [TRT] ModelImporter.cpp:779: ERROR: builtin_op_importers.cpp:3609 In function importResize:
[8] Assertion failed: scales.is_weights() && "Resize scales must be an initializer!"
[12/08/2022-18:48:43] [E] Failed to parse onnx file
[12/08/2022-18:48:43] [I] Finish parsing network model
[12/08/2022-18:48:43] [E] Parsing model failed
[12/08/2022-18:48:43] [E] Failed to create engine from model.
[12/08/2022-18:48:43] [E] Engine set up failed
&&&& FAILED TensorRT.trtexec [TensorRT v8205] # trtexec --onnx=weights/yolov5s.onnx --saveEngine=yolov5s.engine

问题再一次明晰:
python脚本转onnx能成功,但无论python还是C++由onnx转engine都失败!!!
而且从export.py的代码分析,python代码,也是调用的C++的

实战YOLO V5推演（TensorRT版本 C++）
[8] Assertion failed: scales.is_weights() && “Resize scales must be an initializer!”
报错是这个,
1,目前不确认是onnx不能供转化engine使用,还是engine的转换代码有问题
2,不确认是pytorch的版本问题,还是yolov的版本问题,还是tensorrt的版本问题
我们假设:onnx没问题,是转换.engine有问题, 有onnx.load 的几个文件

(base) [root@localhost python3.9]# grep -r onnx.load /root/miniconda3/pkgs
/root/miniconda3/pkgs/pytorch-1.12.1-py3.9_cuda11.3_cudnn8.3.2_0/lib/python3.9/site-packages/torch/utils/tensorboard/_onnx_graph.py:    m = onnx.load(fname)
/root/miniconda3/pkgs/pytorch-1.12.1-py3.9_cuda11.3_cudnn8.3.2_0/lib/python3.9/site-packages/caffe2/python/onnx/backend.py:                model = onnx.load(f);
/root/miniconda3/pkgs/pytorch-1.12.1-py3.9_cuda11.3_cudnn8.3.2_0/lib/python3.9/site-packages/caffe2/python/trt/test_trt.py:        model_def = onnx.load(os.path.join(model_dir, 'model.onnx'))
/root/miniconda3/pkgs/pytorch-1.12.1-py3.9_cuda11.3_cudnn8.3.2_0/lib/python3.9/site-packages/caffe2/contrib/aten/docs/sample.py:graph = onnx.load(f.name)

用一个图片来推理,验证假设,mmp todesk连不到公司了 睡觉!!!

在家用wls2 的ubuntu系统试试,基础情况
torch 1.13.0+cu116
torchaudio 0.13.0+cu116
torchvision 0.14.0+cu116

 nvidia-smi -l
Fri Dec  9 03:52:56 2022
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 510.73.01    Driver Version: 512.78       CUDA Version: 11.6     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  NVIDIA GeForce ...  On   | 00000000:01:00.0 Off |                  N/A |
| N/A   51C    P0    25W /  N/A |      0MiB /  8192MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
(base) justin@DESKTOP-NIK28BI:~/software$ nvcc -V
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Tue_Mar__8_18:18:20_PST_2022
Cuda compilation tools, release 11.6, V11.6.124
Build cuda_11.6.r11.6/compiler.31057947_0
cuda: cuda_11.6.2_510.47.03_linux.run
cudann:  cudnn-linux-x86_64-8.5.0.96_cuda11-archive.tar.xz
# tensorrt
version="8.5.1.7"
arch=$(uname -m)
cuda="cuda-11.6"
cudnn="cudnn8.5"
tar -xzvf TensorRT-${version}.Linux.${arch}-gnu.${cuda}.${cudnn}.tar.gz

用的是yolov5 最新的包,应该是tag7.0
执行转换命令

python export.py --weights weights/yolov5s.pt --device 0 --include onnx engine

都转换成功了!!! 问题明晰,可能是torch版本的问题

/usr/local/TensorRT-8.5.1.7/bin/trtexec --onnx=weights/yolov5s.onnx --saveEngine=weights/yolov5s.engine

用c++ 也能转出来.

--------------------------------------------------------------------------------------------------------------分割线-------------------------------------------------------------------------------------------------------------------------------

3. 再来一遍

2023.1.13
cudnn的版本在这里
/usr/local/cuda-11.1/include/cudnn_version.h
环境：cuda 11.1

#ifndef CUDNN_VERSION_H_
#define CUDNN_VERSION_H_
#define CUDNN_MAJOR 8
#define CUDNN_MINOR 0
#define CUDNN_PATCHLEVEL 5

cudnn：8.05
tenssort：TensorRT v8501
yolov5tag7
转化命令

 trtexec --onnx=yolov5s-seg.onnx --saveEngine=yolov5s-seg.engine

trtexec找不到的话，再解压包的这里/usr/local/TensorRT-8.5.1.7/bin/trtexec
我是把他mv到/usr/local 下了 你可以export 或者绝对路径使用
最后输出这个，说明成功了。
实战YOLO V5推演（TensorRT版本 C++）
能转化成功，但是安装opencv的C++版本装不上，换了2台server都装不上
···········································································
2023.1.18
在docker 上搞
升级cmake指引:https://www.lhl.zone/Linux/33.html
ref:https://blog.csdn.net/qq_41043389/article/details/127754384
opencv C++的安装指引:https://docs.opencv.org/4.5.5/d7/d9f/tutorial_linux_install.html
opencv 的下载：https://opencv.org/releases/
我安装的是4.5.5, 安装成功了。哈哈~
cudnn的版本在这里
/usr/local/cuda-11.1/include/cudnn_version.h
环境：cuda 11.1

#ifndef CUDNN_VERSION_H_
#define CUDNN_VERSION_H_
#define CUDNN_MAJOR 8
#define CUDNN_MINOR 0
#define CUDNN_PATCHLEVEL 5

cudnn：8.05
tenssort：TensorRT v8501
yolov5tag7
转化命令

 trtexec --onnx=yolov5s-seg.onnx --saveEngine=yolov5s-seg.engine

trtexec找不到的话，再解压包的这里/usr/local/TensorRT-8.5.1.7/bin/trtexec
我是把他mv到/usr/local 下了 你可以export 或者绝对路径使用
最后输出这个，说明成功了。
······································
走这个blog指引的时候 走不通，找不到cuda环境。哎~。
https://blog.csdn.net/qq_41043389/article/details/127754384

报错也先记录一下:

(py38) [root@localhost build]# cmake ..
-- The C compiler identification is GNU 4.8.5
-- The CXX compiler identification is GNU 4.8.5
-- Check for working C compiler: /usr/bin/cc
-- Check for working C compiler: /usr/bin/cc -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Detecting C compile features
-- Detecting C compile features - done
-- Check for working CXX compiler: /usr/bin/c++
-- Check for working CXX compiler: /usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Found OpenCV: /usr/local (found version "4.5.5")
-- Found CUDA: /usr/local/cuda (found version "11.1")
--  libraries: /usr/local/cuda/lib64/libcudart.so
--  include path: /usr/local/cuda/include
CMake Error: Could not find cmake module file: CMakeDetermineCUDACompiler.cmake
CMake Error: Error required internal CMake variable not set, cmake may be not be built correctly.
Missing variable is:
CMAKE_CUDA_COMPILER_ENV_VAR
CMake Error: Could not find cmake module file: /app/Yolov5-instance-seg-tensorrt-main/build/CMakeFiles/3.7.2/CMakeCUDACompiler.cmake
CMake Error: Could not find cmake module file: CMakeCUDAInformation.cmake
CMake Error: Could not find cmake module file: CMakeTestCUDACompiler.cmake
-- Configuring incomplete, errors occurred!
See also "/app/Yolov5-instance-seg-tensorrt-main/build/CMakeFiles/CMakeOutput.log".

用docker安装首先要安装cmake，接着安装opencv2，还要安装gxx
我把history也贴出来

    1  ls
    2  python -V
    3  pip list
    4  pip uninstall python
    5  pip uninstall python
    6  which pip
    7  conda uninstall python
    8  ls
    9  python -V
   10  pip list
   11  pip uninstall python
   12  pip uninstall python
   13  which pip
   14  conda uninstall python
   15  which pip
   16  cat ~/.condarc
   17  nvcc -V
   18  python
   19  conda create -n py38 python=3.8
   20  conda activate py38
   21  conda install pytorch1.8.0 torchvision0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge
   22  ls
   23  cd app
   24  ls
   25  pip install -r requirements.txt
   26   conda install pytorch1.8.0 torchvision0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge
   27  conda install pytorch1.8.0 torchvision0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge
   28  conda activate py38
   29  import tensorrt
   30  python
   31  import torch
   32  python
   33  echo  $LD_LIBRARY_PATH
   34  find / -name libnvinfer.so.8
   35  vim ~/.bashrc
   36  fg
   37  source  ~/.bashrc
   38  vim ~/.bashrc
   39  vim /etc/profile
   40  source /etc/profile
   41  source ~/.bashrc
   42  conda env list
   43  ls
   44  cd app/
   45  ls
   46  cd weights/
   47  ls
   48  which trtexec
   49  /software/TensorRT-8.5.1.7/bin/trtexec --onnx=yolov5m6.onnx --saveEngine=yolov5m6.engine
   50  export LD_LIBRARY_PATH=/software/TensorRT-8.5.1.7/targets/x86_64-linux-gnu/lib:$LD_LIBRARY_PATH
   51  /software/TensorRT-8.5.1.7/bin/trtexec --onnx=yolov5m6.onnx --saveEngine=yolov5m6.engine
   52  程cd 。。
   53  cd。 ..
   54  cd ..
   55  ls
   56  cd cpp_project/
   57  g++ -verson
   58  g++ -V
   59  g++ --help
   60  g++ --version
   61  ls
   62  g++ yolov5_trt.cpp
   63  which yum
   64  yum install glog-devel
   65  yum install glog-devel
   66  yum install gflags-devel glog-devel
   67  cd /software/
   68   wget https://google-glog.googlecode.com/files/glog-0.3.3.tar.gz
   69  ls
   70  ls -alh
   71  sudo chmod -R 0777 glog-0.3.3.tar.gz
   72   chmod -R 0777 glog-0.3.3.tar.gz
   73  tar glog-0.3.3.tar.gz
   74  tar zxvf glog-0.3.3.tar.gz
   75  cd glog-0.3.3
   76  ./configure
   77  make && make install
   78  cd /app/
   79  cd cpp_project/
   80  history
   81  g++ yolov5_trt.cpp
   82  cd /software/
   83  ls
   84  unzip opencv-4.5.5.zip
   85  yum install unzip
   86  unzip opencv-4.5.5.zip
   87  mkdir -p build && cd build
   88  cmake ../opencv-4.5.5
   89  cd ..
   90  ;s
   91  ;s
   92  ls
   93  tar -zxvf cmake-3.25.1.tar.gz
   94  cd cmake-3.25.1
   95  ls -alh
   96  ./bootstrap
   97  yum install libssl-dev
   98  yum install openssl
   99  yum install openssl-devel
  100  ./bootstrap
  101  make
  102  make install
  103  cd -
  104  ls
  105  cd /app/
  106  ;s
  107  ls
  108  cd cpp_project/
  109  ls
  110  history | grep gcc
  111  history | grep g++
  112  g++ yolov5_trt.cpp
  113  cd /software/
  114  ls
  115  cd build/
  116  cmake ../opencv-4.5.5
  117  cmake --build .
  118  make -j4
  119  ls bin
  120  ls lib
  121  ls OpenCVConfig*.cmake
  122  ls
  123  make install
  124  程cd ▒/app
  125  cd /app/
  126  ls
  127  cd cpp_project/
  128  ls
  129  history | grep g++
  130  ▒g++ yolov5_trt.cpp
  131  g++ yolov5_trt.cpp
  132  cd ..
  133  ls
  134  cd ..
  135  cd software/
  136  ls
  137  unzip opencv-2.4.13.6.zip
  138  cd opencv-2.4.13.6
  139  mkdir release
  140  cd release/
  141  cmake -D CMAKE_BUILD_TYPE=RELEASE -D CMAKE_INSTALL_PREFIX=/usr/local ..
  142  cmake -D CMAKE_BUILD_TYPE=RELEASE -DCMAKE_INSTALL_PREFIX=/usr/local ..
  143  cmake -DCMAKE_BUILD_TYPE=RELEASE -DCMAKE_INSTALL_PREFIX=/usr/local ..
  144  cd ..
  145  cd ..
  146  ls
  147  cd opencv-2.4.13.6
  148  ls
  149  rm -rf release/
  150  mkdir release
  151  cd release
  152  cmake -DCMAKE_BUILD_TYPE=RELEASE -DCMAKE_INSTALL_PREFIX=/usr/local ..
  153  make
  154  cd /app/
  155  g++ yolov5_trt.cpp
  156  cd cpp_project/
  157  g++ yolov5_trt.cpp
  158  g++ yolov5_trt.cpp
  159  cd  /
  160  ls
  161  cd app/
  162  ls
  163  cd ..
  164  ls
  165  cd /app/
  166  ls
  167  cd ..
  168  ls
  169  ps
  170  ls
  171  cd opt/
  172  ls
  173  cd ..
  174  cd app/
  175  ls
  176  cd ..
  177  ls
  178  ls
  179  cd software/
  180  ls
  181  cd opencv-2.4.13.6
  182  ls
  183  cat CMakeLists.txt
  184  cd ..
  185  ls
  186  rm -rf opencv-2.4.13.6
  187  unzip opencv-2.4.13.6.zip
  188  rm -rf build/
  189  cd opencv-2.4.13.6
  190  mkdir build && cd build
  191  cmake ..
  192  make -j4
  193  ls
  194  cd ..
  195  ls
  196  vim CMakeLists.txt
  197  grep -rn "compute"
  198  grep -rn "compute_20"
  199  rm -rf build/
  200  grep -rn "compute_20"
  201  mkdir build && cd build
  202  cd ..
  203  ls
  204  vim CMakeLists.txt
  205  mkdir build && cd build
  206  rm -rf build/
  207  mkdir build && cd build
  208  cmake .. -DCMAKE_BUILD_TYPE=Release -DCUDA_nppi_LIBRARY=true -DWITH_CUDA=OFF -DBUILD_TIFF=ON
  209  make -j4
  210  make install
  211  opencv_vision
  212  cd ..
  213  ls
  214  cd ..
  215  ls
  216  cd ..
  217  cd app/
  218  cd cpp_project/
  219  history | grep g++
  220  ▒g++ yolov5_trt.cpp
  221   g++ yolov5_trt.cpp
  222  find /  NvInfer.h
  223  echo PATH
  224  echo $PATH
  225  ls
  226  cd /software/
  227  ls
  228  cd TensorRT-8.5.1.7
  229  ls
  230  pwd
  231  cd ..
  232  ls
  233  cd ..
  234  ls
  235  cd app/
  236  ls
  237  cd cpp_project/
  238   g++ yolov5_trt.cpp
  239   g++ yolov5_trt.cpp
  240  find -name NvInfer.h
  241  cd /
  242  find -name NvInfer.h
  243  cd app/cpp_project/
  244   g++ yolov5_trt.cpp
  245  find / -name cuda_runtime_api.h
  246   g++ yolov5_trt.cpp
  247   g++ yolov5_trt.cpp
  248  mv /software/TensorRT-8.5.1.7 /usr/local/
  249  cp /usr/local/TensorRT-8.5.1.7/ /software/
  250  cp -r /usr/local/TensorRT-8.5.1.7 /software/TensorRT-8.5.1.7
  251  export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/TensorRT-8.5.1.7/lib
  252  g++ yolov5_trt.cpp
  253  g++ yolov5_trt.cpp
  254  mkdir build && cd build
  255  cmake ..
  256  make
  257  ls
  258  cd ..
  259  ls
  260  cd build
  261  ls
  262  程cd ..
  263  ls
  264  cd ..
  265  ls
  266  g++  yolov5_trt.cpp
  267  find / NvInfer.h
  268  find / -name  NvInfer.h
  269  g++  yolov5_trt.cpp
  270  g++  yolov5_trt.cpp
  271  vim /usr/local/TensorRT-8.5.1.7/include/NvInferRuntimeCommon.h:19:30
  272  vim /usr/local/TensorRT-8.5.1.7/include/NvInferRuntimeCommon.h 19
  273  g++  yolov5_trt.cpp
  274  find / -name  cuda_runtime_api.h
  275  vim /usr/local/TensorRT-8.5.1.7/include/NvInferRuntimeCommon.h 19
  276  g++  yolov5_trt.cpp
  277  find / -name  NvOnnxParser.h
  278  g++  yolov5_trt.cpp
  279  find / -name  cuda_runtime_api.h
  280  g++  yolov5_trt.cpp
  281  g++  yolov5_trt.cpp
  282  g++  yolov5_trt.cpp
  283  clear
  284  g++  yolov5_trt.cpp
  285  history

总结

没完成，但情况开始清晰：
1.环境冲突的问题已解决
2.主要是c++不灵光所致
c++不灵光包括cmakelist不会写，只能挨家挨户的找指引，其实应该自己写出一个c++版本的前后处理来掉yolov5.engine权重！！！加油
结论清晰了，补充c++ 和继续尝试去吧~！！！
20230130

标签 , ,