Jetson平台使用Inter 神经网络计算棒(Neural Compute Stick, NCS)
工作方式
NCS工作方式分为两种,一种是在主机上将训练好的模型生成NCS可执行graph文件,该文件用于推理过程;另一种是在树莓派、Jetson TX2等便携式计算机上加速推理过程。
主机端使用NCS
安装
将神经计算棒插入主机中,在终端执行以下命令:
git clone https://github.com/movidius/ncsdk
cd ncsdk
make install
make install的作用如下:
-
检查安装Tensorflow;
-
检查安装Caffe(SSD-caffe);
-
编译安装ncsdk(不包含inference模块,只包含mvNCCompile相关模块,用来将Caffe或Tensorflow模型转成NCS graph的)
之后执行:
make example
程序顺利运行不报错的话,就说明已经安装成功了。
使用
将训练好的模型生成NCS可以执行的graph文件,在终端执行以下命令:
mvNCCompile network.prototxt -w network.caffemodel -s MaxNumberOfShaves -in InputNodeName -on OutputNodeName-is InputWidth InputHeight -o OutputGraphFilename
network.prototxt:.prototxt文件的路径
-w network.caffemode:模型文件的路径
-s MaxNumberOfShaves:1, 2, 4, 8, 12。默认为12
-in InputNodeName:选择指定一个特定的输入图层(它将匹配prototxt文件中的名称,可选项)
-on OutputNodeName:默认情况下网络是通过输出张量进行处理的,这个选项允许用户在网络中选择一个替代端点(可选项)
-is InputWidth InputHeight:输入尺寸,需要与网络匹配
-o OutputGraphFilename:生成的graph文件存储路径
在Jetson TX2上安装NCS
在TX2上只完成推理(Inference)过程,所以只需安装API-only模式即可,将NCS插入到TX2上。
安装依赖
sudo apt-get install -y libusb-1.0-0-dev libprotobuf-dev libleveldb-dev libsnappy-dev libopencv-dev libhdf5-serial-dev protobuf-compiler libatlas-base-dev git automake byacc lsb-release cmake libgflags-dev libgoogle-glog-dev liblmdb-dev swig3.0 graphviz libxslt-dev libxml2-dev gfortran python3-dev python-pip python3-pip python3-setuptools python3-markdown python3-pillow python3-yaml python3-pygraphviz python3-h5py python3-nose python3-lxml python3-matplotlib python3-numpy python3-protobuf python3-dateutil python3-skimage python3-scipy python3-six python3-networkx python3-tk
下载源码
mkdir ~/workspace
cd ~/workspace
git clone https://github.com/movidius/ncsdk
编译和安装NCSDK API框架
cd ~/workspace/ncsdk/api/src
make
sudo make install
测试
cd ~/workspace
git clone https://github.com/movidius/ncappzoo
cd ncappzoo/apps/hello_ncs_py
python3 hello_ncs.py
出现以下结果:
Hello NCS! Device opened normally.
Goodbye NCS! Device closed normally.
NCS device working.
API-only模式安装成功。
TX2利用NCS加速推理使用
参数预定义:
- GRAPH_PATH:graph文件路径;
- IMAGE_PATH:要分类的图片的路径;
- IMAGE_DIM:由选择的神经网络定义的图像尺寸;例:GoogLeNet uses 224x224 pixels, AlexNet uses 227x227 pixels
- IMAGE_STDDEV:由选择的神经网络定义的标准差(标度值);例:GoogLeNet uses no scaling factor, InceptionV3 uses 128 (stddev = 1/128)
- IMAGE_MEAN:平均减法是深度学习中常用的一种技术,用于对数据进行中心处理。例:ILSVRC dataset, the mean is B = 102 Green = 117 Red = 123
使用NCS做图像分类的5个步骤:
从mvnc库中引入mvncapi模块
import mvnc.mvncapi as mvnc
01
将NCS插入应用处理器(Ubuntu笔记本电脑/台式机)USB端口时,它将自身列为USB设备。通过调用API来查找枚举的NCS设备:
# Look for enumerated Intel Movidius NCS device(s); quit program if none found.
devices = mvnc.EnumerateDevices()
if len( devices ) == 0:
print( ‘No devices found‘ )
quit()
如果插入了多个NCS,还需要选择一个NCS并打开:
# Get a handle to the first enumerated device and open it
device = mvnc.Device( devices[0] )
device.OpenDevice()
02
加载graph文件到NCS
# Read the graph file into a buffer
with open( GRAPH_PATH, mode=‘rb‘ ) as f:
blob = f.read()
# Load the graph buffer into the NCS
graph = device.AllocateGraph( blob )
03
将图像加载到Intel Movidius NCS上以运行推理
图像预处理:
1.调整图像大小/裁剪图像以匹配预先训练的网络定义的尺寸。例:GoogLeNet uses 224x224 pixels, AlexNet uses 227x227 pixels.
2.每个通道的平均值(蓝色,绿色和红色)从整个数据集中减去。这是深度学习中常用的一种技术,可以集中数据。
3.将图像转换为半精度浮点数(fp16)数组(NCS输入数据格式为fp16),并使用LoadTensor函数调用将图像加载到NCS上。skimage库可以在一行代码中完成此操作。
# Read & resize image (Image size is defined during training)
img = print_img = skimage.io.imread( IMAGES_PATH )
img = skimage.transform.resize( img, IMAGE_DIM, preserve_range=True )
# Convert RGB to BGR [skimage reads image in RGB, but Caffe uses BGR]
img = img[:, :, ::-1]
# Mean subtraction & scaling [A common technique used to center the data]
img = img.astype( numpy.float32 )
img = ( img - IMAGE_MEAN ) * IMAGE_STDDEV
# Load the image as a half-precision floating point array
graph.LoadTensor( img.astype( numpy.float16 ), ‘user object‘ )
04
从NCS读取并打印推理结果
# Get the results from NCS
output, userobj = graph.GetResult()
# Print the results
print(‘\n------- predictions --------‘)
labels = numpy.loadtxt( LABELS_FILE_PATH, str, delimiter = ‘\t‘ )
order = output.argsort()[::-1][:6]
for i in range( 0, 5 ):
print (‘prediction ‘ + str(i) + ‘ is ‘ + labels[order[i]])
# Display the image on which inference was performed
skimage.io.imshow( IMAGES_PATH )
skimage.io.show( )
05
卸载图形并关闭设备
为了避免内存泄漏和/或分段错误,我们应该关闭所有打开的文件或资源并释放所有使用的内存。
graph.DeallocateGraph()
device.CloseDevice()
运行Demo
Demo采用Adrian Rosebrock博客Real-time object detection on the Raspberry Pi with the Movidius NCS - PyImageSearch里的程序(https://www.pyimagesearch.com/2018/02/19/real-time-object-detection-on-the-raspberry-pi-with-the-movidius-ncs/),这个程序基于Mobilenet-ssd模型对视频流做实时检测,如图2所示,Demo采用USB摄像头读取实时视频。图3为检测结果。
参考
https://github.com/movidius/ncsdk
https://movidius.github.io/blog/ncs-apps-on-rpi/