1.Ubuntu虚拟机上安装NC SDK
cd /home/shine/Downloads/ mkdir NC_SDK git clone https://github.com/movidius/ncsdk.git make install
通过运行测试例程判断是否正确安装
make examples
2.激活USB设备
在启动ubuntu前，请不要插入movidius，等ubuntu启动以后，再插入（知乎用户经验，笔者测试不影响）
3.测试工程
cd /home/shine/Downloads/NC_SDK/ncsdk/examples/apps/hello_ncs_cpp/ make run cd /home/shine/Downloads/NC_SDK/ncsdk/examples/apps/hello_ncs_py/ make run
正常结果显示
Hello NCS! Device opened normally.
Goodbye NCS! Device Closed normally.
NCS device working.
4.训练工程
ncappzoo中提供了大量的工程样例提供分析，为开发者的模型选择提供了极大的便利，在选择模型的时候需要综合权衡训练样本的类型、大小以及部署后的运行速度。
cd /home/shine/Downloads/NC_SDK/ncsdk/ git clone https://github.com/ashwinvijayakumar/ncappzoo git checkout dogsvscats
以猫和狗的分类任务为例
- 数据集的准备（在百度网盘中共享测试数据集和训练数据集）
https://pan.baidu.com/s/1mtXYfB61Czkadjrgs4RXzw
https://pan.baidu.com/s/1ZD4Hocgk4bMcl8tQGkTHcQ
cd ncappzoo/apps/dogsvscats mkdir data mv /home/shine/Downloads/test1.zip ~/Downloads/ncappzoo/apps/dogsvscats/data/ mv /home/shine/Downloads/train.zip ~/Downloads/ncappzoo/apps/dogsvscats/data/ cd ncappzoo/apps/dogsvscats make
上述操作主要执行
Image pre-processing - resizing , cropping , histogram equalization (图像预处理) Shuffling the images (图像打乱) Splitting the images into training and validation (图像分割为训练集和测试集) Creating an lmdb database of these images (格式转换) Computing image mean -a common deep learning technique used to normalize data (计算图像均值)
- 模型对比
①比较模型的差异
export CAFFE_PATH=~/Downloads/caffe-master diff -u $CAFFE_PATH/models/bvlc_googlenet bvlc_googlenet/org
- 数据训练
①下载caffe预训练模型，使用本地CPU或GPU进行训练，CAFFE_PATH需要替换为本地安装目录
$CAFFE_PATH/scripts/download_model_binary.py $CAFFE_PATH/models/bvlc_googlenet
$CAFFE_PATH/build/tools/caffe train --solver bvlc_googlenet/org/solver.prototxt --weights $CAFFE_PATH/models/bvlc_googlenet/bvlc_googlenet.caffemodel 2>&1 | tee bvlc_googlenet/org/train.log
#错误1：Cannot use GPU in CPU-only Caffe: check mode cd ~/Downloads/ncappzoo/apps/dogsvscats/bvlc_googlenet/org vim solver.prototxt 将其中的 solver_mode: GPU改为 solver_mode: CPU 或者将caffe重新编译成GPU模式 #错误2：Check failed: error == cudaSuccess (2 vs. 0) out of memory 由于博主使用的是GTX 650Ti 显存只有979Mb，执行GPU运算的时候出现显存不足的现象
②使用Intel AI Cloud 加速训练
如上文所述，在本地训练数据是一个巨大的运算量，常常需要几周或几个月，因此使用Intel提供的云服务器可以极大缩短训练的时间
在terminal中使用如下语句登陆到AI Cloud 服务器
ssh colfax mkdir dogsvscats
登陆成功后即显示
######################################################################## # Welcome to Intel AI DevCloud! ########################################################################
将训练dogsvscats工程所需的数据集及shell命令上传到服务器（请根据实际目录进行调整,若上传速度较慢请尝试云服务器wget直接下载开放数据集）
scp /home/shine/Downloads/ncappzoo/apps/dogsvscats/data/train.zip colfax:/home/u14673/ncappzoo/apps/dogsvscats/data/
scp /home/shine/Downloads/ncappzoo/apps/dogsvscats/data/test1.zip .zip colfax:/home/u14673/ncappzoo/apps/dogsvscats/data/
将对应的shell文件和Makefile上传到服务器用于训练数据预处理（请根据实际目录进行调整）
scp /home/shine/Downloads/ncappzoo/apps/dogsvscats/Makefile colfax:/home/u14673/ncappzoo/apps/dogsvscats/ scp /home/shine/Downloads/ncappzoo/apps/dogsvscats/create-labels.py colfax:/home/u14673/ncappzoo/apps/dogsvscats/ scp /home/shine/Downloads/ncappzoo/apps/dogsvscats/create-lmdb.sh colfax:/home/u14673/ncappzoo/apps/dogsvscats/
使用Makefile进行预处理，由于Makefile中deps含有sudo apt-get -y install unzip和sudo pip install pyyaml，且sudo apt-get在AI Cloud中无法运行
vi Makefile
将deps更改为
@echo "Installing dependencies..." # sudo apt-get -y install unzip # sudo pip install pyyaml
:wq!保存后退出，创建任务用于数据预处理
vi data_process.sh
在打开的界面中输入（请根据实际目录进行调整）
echo "Start Data Process" cd /home/u14673/ncappzoo/apps/dogsvscats/ make all echo "Data Process Finished"
:wq!保存后退出，提交任务开始数据预处理
qsub data_process.sh
使用qstat可以查看任务完成的情况，完成后会在当前目录中生成对应的日志文件
将训练所需的prototxt及预训练模型上传至AI Cloud
scp -r /home/shine/Downloads/ncappzoo/apps/dogsvscats/bvlc_googlenet colfax:/home/u14673/ncappzoo/apps/dogsvscats/
scp /home/shine/Downloads/caffe/models/bvlc_googlenet/bvlc_googlenet.caffemodel colfax:/home/u14673/ncappzoo/apps/dogsvscats/
创建任务用于数据训练
vi data_train.sh
在打开的界面中输入如下语句（请根据实际目录进行调整）
cd /home/u14673/ncappzoo/apps/dogsvscats/ echo 'Start Trainning'
# >&表示所有的标准输出和标准错误输出都将被重定向 /glob/intel-python/python3/bin/caffe train --solver bvlc_googlenet/org/solver.prototxt --weights /home/u14673/ncappzoo/apps/dogsvscats/bvlc_googlenet.caffemodel 2>&1 | tee bvlc_googlenet/org/train.log
关于caffe train命令的定义，标准的范例如下
caffe train \ --solver=solver_1st.prototxt \ --weights=VGG/VGG_ILSVRC_16_layers.caffemodel \ --gpu=0,1,2,3 2>&1 | tee log_1st.txt
其中--solver为必要的参数，配置solver文件
如果从头开始训练模型，则无需配置--weights
如果从快照中恢复，则需要配置--snapshot--weights
如果是fine-tuning，则需要配置
:wq!保存后退出，提交任务开始训练，训练完成后在当前目录可以看到日志文件
qsub data_train.sh
查看日志，日志保存在 bvlc_googlenet/org 目录，使用如下命令将数据拷贝到本地
scp colfax:/home/u14673/ncappzoo/apps/dogsvscats/bvlc_googlenet/org/train.log ./
使用caffe自带的工具绘制（位于caffe/tools/extra目录）训练数据，caffe中支持很多种曲线绘制，通过指定不同的类型参数即可，具体参数如下
Notes: 1. Supporting multiple logs. 2. Log file name must end with the lower-cased ".log". Supported chart types: 0: Test accuracy vs. Iters 1: Test accuracy vs. Seconds 2: Test loss vs. Iters 3: Test loss vs. Seconds 4: Train learning rate vs. Iters 5: Train learning rate vs. Seconds 6: Train loss vs. Iters 7: Train loss vs. Seconds
解析日志并生成Test accuracy vs. Seconds曲线(实际应该为Test Loss，参考https://www.cnblogs.com/WaitingForU/p/9130327.html的解析)
cd ~/Downloads/ncappzoo/apps/dogsvscats/bvlc_googlenet/org cp -r /home/shine/Downloads/caffe/tools/extra ~/Downloads/ncappzoo/apps/dogsvscats/bvlc_googlenet/org mv train.log ./extra/ ./plot_training_log.py.example 0 save.png ./train.log
Test Loss Vs Seconds
Train Loss Vs Seconds
从上述两张图来看，似乎训练过程并未收敛，对于这一问题，原作者并未给出原因，而是建议去掉--weights重新进行训练
/glob/intel-python/python3/bin/caffe train --solver bvlc_googlenet/org/solver.prototxt 2>&1 | tee bvlc_googlenet/org/train_withoutweights.log
Test Loss Vs Iters
Test Accuracy Vs Iters
将训练后的模型拷贝到本地
scp colfax:/home/u14673/ncappzoo/apps/dogsvscats/bvlc_googlenet/org/bvlc_googlenet_iter_40000.caffemodel /home/shine/Downloads/ncappzoo/apps/dogsvscats/bvlc_googlenet/org/cd ~/Downloads/ncappzoo/apps/dogsvscats/bvlc_googlenet/org/
本地机器（需要安装NCSDK）查看网络分析，大致可以得到如下的图形，显示了各层连接的带宽和运行时间
mvNCProfile -s 12 deploy.prototxt -w bvlc_googlenet_iter_40000.caffemodel
firefox output_report.html
- 模型调优
作者对比了dogsvscats例程中改进的网络和GoogLenet原始网络，通过Caffe自带的Python工具分别绘制对应网络拓扑
cd ~/Downloads/ncappzoo/apps/dogsvscats/bvlc_googlenet/org python ~/Downloads/caffe/python/draw_net.py train_val.prototxt train_val_plot.png
eog train_val_plot.png
cd ~/Downloads/ncappzoo/apps/dogsvscats/bvlc_googlenet/custom python ~/Downloads/caffe/python/draw_net.py train_val.prototxt train_val_plot.png
eog train_val_plot.png
使用python-caffe自带的工具draw_net.py时可能会遇到如下错误
#错误1 ImportError: No module named google.protobuf (没有安装python-protobuf)
sudo apt-get install python-protobuf
#错误２ ImportError: No module named _caffe (caffe源码编译的时候没有编译pycaffe)
cd ~/Downloads/caffe/
sudo make pycaffe
#错误3 ImportError: No module named skimage.io (没有安装python-skimage)
sudo apt-get install python-skimage
#错误4 ImportError: No module named pydot (没有安装python-pydot)
sudo apt install python-pydot python-pydot-ng graphviz
- 模型部署
根据最新训练的结果，生成graph文件
cd ~/workspace/ncappzoo/apps/dogsvscats/bvlc_googlenet/org (由于前面训练过程未能收敛，使用该模型预测时会出现Ｎａｎ的结果)
mvNCCompile -s 12 deploy.prototxt -w bvlc_googlenet_iter_40000.caffemodel -o dogsvscats-org.graph
cd ~/workspace/ncappzoo/apps/dogsvscats/bvlc_googlenet/custom （定制优化后的网络）
mvNCCompile -s 12 deploy.prototxt -w bvlc_googlenet_iter_40000.caffemodel -o dogsvscats-org.graph
- 模型测试
修改ncappzoo/apps/image-classifier.py，原文件如下
# User modifiable input parameters NCAPPZOO_PATH = '../..' GRAPH_PATH = NCAPPZOO_PATH + '/caffe/GoogLeNet/graph' IMAGE_PATH = NCAPPZOO_PATH + '/data/images/cat.jpg' CATEGORIES_PATH = NCAPPZOO_PATH + '/data/ilsvrc12/synset_words.txt' IMAGE_MEAN = numpy.float16( [104.00698793, 116.66876762, 122.67891434] ) IMAGE_STDDEV = ( 1 ) IMAGE_DIM = ( 224, 224 )
修改后的文件如下
NCAPPZOO_PATH = '../..' GRAPH_PATH = NCAPPZOO_PATH +'/apps/dogsvscats/bvlc_googlenet/custom/dogsvscats-org.graph' IMAGE_PATH = NCAPPZOO_PATH +'/apps/dogsvscats/data/test1/173.jpg' CATEGORIES_PATH = NCAPPZOO_PATH +'/apps/dogsvscats/data/categories.txt' IMAGE_MEAN = numpy.float16( [106.202, 115.912, 124.449] ) IMAGE_STDDEV = ( 1 ) IMAGE_DIM = ( 224, 224 )
使用生成的graph测试准确率(注意是python3，使用python image-classifier.py时会报错，具体原因待查明)
cd ~/Downloads/ncappzoo/apps/image-classifier
python3 image-classifier.py
得到结果如下
------- predictions --------
Prediction for : dog with 100.0% confidence in 89.67 ms
Ｐrediction for : cat with 0.0% confidence in 89.67 ms
那么关于本步骤部署，系统具体作了哪些事情呢，深入查看image-classifier.py我们可以得知
# ---- Step 1: Open the enumerated device and get a handle to it ------------- # 枚举Movidius神经元计算棒 # ---- Step 2: Load a graph file onto the NCS device ------------------------- # 加载graph文件 # ---- Step 3: Offload image onto the NCS to run inference ------------------- # 加载image文件 # ---- Step 4: Read & print inference results from the NCS ------------------- # 读取并打印运算结果 # ---- Step 5: Unload the graph and close the device ------------------------- # 关闭神经元计算棒
- 模型调优
- 参考文献：
1.https://movidius.github.io/blog/deploying-custom-caffe-models
2.https://communities.intel.com/community/tech/intel-ai-academy