初学TX2之jetson-inference安装、分割模型推理测试记录

官方配置参考：https://github.com/dusty-nv/jetson-inference/blob/master/docs/building-repo-2.md

一、基于TX2的jetson-inference安装：

1.安装依赖

$sudo apt-get install git cmake

2.clone the jetson-inference repo

$ git clone https://github.com/dusty-nv/jetson-inference
$ cd jetson-inference
$ git submodule update --init

3.Configuring with CMake

1.注释掉CMakePreBuild.sh中下载模型、解压模型的带以下关键词的代码，不去下载那些模型文件

#sudo apt-get install -y libopencv-calib3d-dev libopencv-dev 
#wget
#mv
#tar -xzvf


2.cmake安装
$ mkdir build
$ cd build
$ cmake ../

##cmake 错误总结##
=========================================================================================
error 1.如果上面cmake出错如下提示(jetson nano开发板不会出现这个问题），则使用下面solve解决方案1的命令行进行cmake：
error1:cmake ../
CMake Error at /usr/share/cmake-3.5/Modules/FindPackageHandleStandardArgs.cmake:148 (message):
  Could NOT find CUDA (missing: CUDA_CUDART_LIBRARY) (found suitable exact
  version "9.0")
Call Stack (most recent call first):
  /usr/share/cmake-3.5/Modules/FindPackageHandleStandardArgs.cmake:388 (_FPHSA_FAILURE_MESSAGE)
  /usr/share/cmake-3.5/Modules/FindCUDA.cmake:949 (find_package_handle_standard_args)
  /usr/local/share/OpenCV/OpenCVConfig.cmake:86 (find_package)
  /usr/local/share/OpenCV/OpenCVConfig.cmake:105 (find_host_package)
  trt-console/CMakeLists.txt:4 (find_package)

CMake Error: The following variables are used in this project, but they are set to NOTFOUND.
Please set them or make sure they are set and tested correctly in the CMake files:
CUDA_CUDART_LIBRARY (ADVANCED)
    linked by target "jetson-inference" in directory /home/nvidia/jetson-inference
    linked by target "imagenet-console" in directory /home/nvidia/jetson-inference/imagenet-console
    linked by target "imagenet-camera" in directory /home/nvidia/jetson-inference/imagenet-camera
    linked by target "detectnet-console" in directory /home/nvidia/jetson-inference/detectnet-console
    linked by target "detectnet-camera" in directory /home/nvidia/jetson-inference/detectnet-camera
    linked by target "segnet-console" in directory /home/nvidia/jetson-inference/segnet-console
    linked by target "segnet-camera" in directory /home/nvidia/jetson-inference/segnet-camera
    linked by target "trt-bench" in directory /home/nvidia/jetson-inference/trt-bench

-- Configuring incomplete, errors occurred!
See also "/home/nvidia/jetson-inference/build/CMakeFiles/CMakeOutput.log".

解决方案1（solve）：
$cmake -DCUDA_CUDART_LIBRARY=/usr/local/cuda/lib64/libcudart.so ../
========================================================================================
error 2.不能兼容opencv 4.0.0版本
解决方案2（solve）：
修改工程中jetson-inference/tools/trt-console/CMakeLists.txt
改为：find_package(OpenCV 4.0.0 COMPONENTS core calib3d REQUIRED)
========================================================================================
error 3.没有支持python3.5版本的接口
解决方案3（solve）：
3.1 修改工程中jetson-inference/python/CMakeLists.txt 第8行
改为：set(PYTHON_BINDING_VERSIONS 3.5 3.6 3.7)

3.2 修改工程中jetson-inference/utils/python/CMakeLists.txt 第8行
set(PYTHON_BINDING_VERSIONS 3.5 3.6 3.7)
========================================================================================

4.Compiling the Project

$ cd jetson-inference/build			# omit if pwd is already /build from above
$ make
$ sudo make install

二、基于TX2的jetson-inference分割模型推理：

1.训练好的模型文件的文件夹放置的位置：

放到xxx/jetson-inference/build/aarch/bin/

比如我模型文件的文件夹：

/home/ljm/jetson-inference/build/aarch64/bin/20190625-152849-e5e9_epoch_16.0

2.关于模型网络

在FCN-Alexnet里面重要的网络层，TensorRT已经支持它们了，所以在deploy.prototxt 的末尾处，要保留deconv 和 crop层：

所以看网络层的最后一层知道输出节点的名称为：score

3.分割模型推理测试:命令行如下

$cd ~/jetson-inference/build/aarch/bin
$NET=20190625-152849-e5e9_epoch_16.0
$./segnet-console 11.png output_11.png 
--prototxt=$NET/deploy.prototxt  
--model=$NET/snapshot_iter_54680.caffemodel 
--labels=$NET/label_names.txt 
--colors=$NET/color_map.txt 
--input_blob=data 
--output_blob=score

4.推理测试显示日志：

nvidia@tegra-ubuntu:~/jetson-inference/build/aarch64/bin$ ./segnet-console 
11.png output_11.png 
--prototxt=$NET/deploy.prototxt  
--model=$NET/snapshot_iter_54680.caffemodel 
--labels=$NET/label_names.txt 
--colors=$NET/color_map.txt 
--input_blob=data 
--output_blob=score 
segnet-console
  args (9):  0 [./segnet-console]  
1 [11.png]  
2 [output_11.png]  
3 [--prototxt=20190625-152849-e5e9_epoch_16.0/deploy.prototxt]  
4 [--model=20190625-152849-e5e9_epoch_16.0/snapshot_iter_54680.caffemodel]  
5 [--labels=20190625-152849-e5e9_epoch_16.0/label_names.txt]  
6 [--colors=20190625-152849-e5e9_epoch_16.0/color_map.txt]  
7 [--input_blob=data]  
8 [--output_blob=score]  


segNet -- loading segmentation network model from:
       -- prototxt:   20190625-152849-e5e9_epoch_16.0/deploy.prototxt
       -- model:      20190625-152849-e5e9_epoch_16.0/snapshot_iter_54680.caffemodel
       -- labels:     20190625-152849-e5e9_epoch_16.0/label_names.txt
       -- colors:     20190625-152849-e5e9_epoch_16.0/color_map.txt
       -- input_blob  'data'
       -- output_blob 'score'
       -- batch_size  2

       -- precision  2

[TRT]  TensorRT version 4.0.2
[TRT]  detected model format - caffe  (extension '.caffemodel')
[TRT]  desired precision specified for GPU: FP32
[TRT]  native precisions detected for GPU:  FP32, FP16
[TRT]  attempting to open engine cache file 20190625-152849-e5e9_epoch_16.0/snapshot_iter_54680.caffemodel.2.1.GPU.FP32.engine
[TRT]  loading network profile from engine cache... 20190625-152849-e5e9_epoch_16.0/snapshot_iter_54680.caffemodel.2.1.GPU.FP32.engine
[TRT]  device GPU, 20190625-152849-e5e9_epoch_16.0/snapshot_iter_54680.caffemodel loaded
[TRT]  device GPU, CUDA engine context initialized with 2 bindings
[TRT]  binding -- index   0
               -- name    'data'
               -- type    FP32
               -- in/out  INPUT
               -- # dims  3
               -- dim #0  3 (CHANNEL)
               -- dim #1  480 (SPATIAL)
               -- dim #2  640 (SPATIAL)
[TRT]  binding -- index   1
               -- name    'score'
               -- type    FP32
               -- in/out  OUTPUT
               -- # dims  3
               -- dim #0  21 (CHANNEL)
               -- dim #1  480 (SPATIAL)
               -- dim #2  640 (SPATIAL)
[TRT]  binding to input 0 data  binding index:  0
[TRT]  binding to input 0 data  dims (b=2 c=3 h=480 w=640) size=7372800
[cuda]  cudaAllocMapped 7372800 bytes, CPU 0x101540000 GPU 0x101540000
[TRT]  binding to output 0 score  binding index:  1
[TRT]  binding to output 0 score  dims (b=2 c=21 h=480 w=640) size=51609600
[cuda]  cudaAllocMapped 51609600 bytes, CPU 0x101c50000 GPU 0x101c50000
device GPU, 20190625-152849-e5e9_epoch_16.0/snapshot_iter_54680.caffemodel initialized.
[cuda]  cudaAllocMapped 336 bytes, CPU 0x101340200 GPU 0x101340200
[TRT]  segNet outputs -- s_w 640  s_h 480  s_c 21
[cuda]  cudaAllocMapped 307200 bytes, CPU 0x104d90000 GPU 0x104d90000
segNet -- class 00  color 0 0 0 255
segNet -- class 01  color 0 255 255 255
segNet -- loaded 2 class colors
'egNet -- class 00  label '_background_
'egNet -- class 01  label 'water
segNet -- loaded 2 class labels
loaded image  11.png  (640 x 480)  4915200 bytes
[cuda]  cudaAllocMapped 4915200 bytes, CPU 0x104f90000 GPU 0x104f90000
[cuda]  cudaAllocMapped 4915200 bytes, CPU 0x105440000 GPU 0x105440000
segnet-console:  beginning processing (1561975363278)
[TRT]  layer shift - 1.396928 ms
[TRT]  layer conv1 + relu1 - 76.159103 ms
[TRT]  layer pool1 - 1.299744 ms
[TRT]  layer norm1 - 0.325120 ms
[TRT]  layer conv2 + relu2 - 21.611776 ms
[TRT]  layer pool2 - 0.905824 ms
[TRT]  layer norm2 - 0.253536 ms
[TRT]  layer conv3 + relu3 - 13.983904 ms
[TRT]  layer conv4 + relu4 - 10.975136 ms
[TRT]  layer conv5 + relu5 - 6.627584 ms
[TRT]  layer pool5 - 0.246560 ms
[TRT]  layer fc6 + relu6 - 72.884415 ms
[TRT]  layer fc7 + relu7 - 25.698463 ms
[TRT]  layer score_fr - 0.800960 ms
[TRT]  layer upscore - 1590.975830 ms
[TRT]  layer score - 1.977600 ms
[TRT]  layer network time - 1826.122437 ms
segnet-console:  finished processing (1561975366445)
segnet-console:  completed saving 'output_11.png'

shutting down...

5.测试结果：

6.推理精度改FP16为FP32:

6.1 查看有哪些精度可以选择：查看tensorNet.h文件，第67行

enum precisionType
{
	TYPE_DISABLED = 0,	/**< Unknown, unspecified, or disabled type */
	TYPE_FASTEST,		/**< The fastest detected precision should be use (i.e. try INT8, then FP16, then FP32) */
	TYPE_FP32,		/**< 32-bit floating-point precision (FP32) */
	TYPE_FP16,		/**< 16-bit floating-point half precision (FP16) */
	TYPE_INT8,		/**< 8-bit integer precision (INT8) */
	NUM_PRECISIONS		/**< Number of precision types defined */
};

6.2 修改segNet.h文件中两个地方：改为precisionType precision=TYPE_FP32

static segNet* Create( NetworkType networkType=FCN_ALEXNET_CITYSCAPES_SD, uint32_t maxBatchSize=2,
         precisionType precision=TYPE_FP32, deviceType device=DEVICE_GPU, bool allowGPUFallback=true );  //TYPE_FASTEST改为TYPE_FP32

/**
* Load a new network instance
* @param prototxt_path File path to the deployable network prototxt
* @param model_path File path to the caffemodel
* @param class_labels File path to list of class name labels
* @param class_colors File path to list of class colors
* @param input Name of the input layer blob. @see SEGNET_DEFAULT_INPUT
* @param output Name of the output layer blob. @see SEGNET_DEFAULT_OUTPUT
* @param maxBatchSize The maximum batch size that the network will support and be optimized for.
*/
static segNet* Create( const char* prototxt_path, const char* model_path, 
		       const char* class_labels, const char* class_colors=NULL,
		       const char* input = SEGNET_DEFAULT_INPUT, 
		       const char* output = SEGNET_DEFAULT_OUTPUT,
	               uint32_t maxBatchSize=2, precisionType precision=TYPE_FP32,  // TYPE_FASTEST改为TYPE_FP32
		       deviceType device=DEVICE_GPU, bool allowGPUFallback=true );