1.Ubuntu虛擬機上安裝NC SDK
cd /home/shine/Downloads/ mkdir NC_SDK git clone https://github.com/movidius/ncsdk.git make install
通過運行測試例程判斷是否正確安裝
make examples
2.激活USB設備
在啟動ubuntu前，請不要插入movidius，等ubuntu啟動以后，再插入（知乎用戶經驗，筆者測試不影響）
3.測試工程
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
正常結果顯示
Hello NCS! Device opened normally.
Goodbye NCS! Device Closed normally.
NCS device working.
4.訓練工程
ncappzoo中提供了大量的工程樣例提供分析，為開發者的模型選擇提供了極大的便利，在選擇模型的時候需要綜合權衡訓練樣本的類型、大小以及部署后的運行速度。
cd /home/shine/Downloads/NC_SDK/ncsdk/ git clone https://github.com/ashwinvijayakumar/ncappzoo git checkout dogsvscats
以貓和狗的分類任務為例
- 數據集的准備（在百度網盤中共享測試數據集和訓練數據集）
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
上述操作主要執行
Image pre-processing - resizing , cropping , histogram equalization (圖像預處理) Shuffling the images (圖像打亂) Splitting the images into training and validation (圖像分割為訓練集和測試集) Creating an lmdb database of these images (格式轉換) Computing image mean -a common deep learning technique used to normalize data (計算圖像均值)
- 模型對比
①比較模型的差異
export CAFFE_PATH=~/Downloads/caffe-master diff -u $CAFFE_PATH/models/bvlc_googlenet bvlc_googlenet/org
- 數據訓練
①下載caffe預訓練模型，使用本地CPU或GPU進行訓練，CAFFE_PATH需要替換為本地安裝目錄
$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
#錯誤1：Cannot use GPU in CPU-only Caffe: check mode cd ~/Downloads/ncappzoo/apps/dogsvscats/bvlc_googlenet/org vim solver.prototxt 將其中的 solver_mode: GPU改為 solver_mode: CPU 或者將caffe重新編譯成GPU模式 #錯誤2：Check failed: error == cudaSuccess (2 vs. 0) out of memory 由於博主使用的是GTX 650Ti 顯存只有979Mb，執行GPU運算的時候出現顯存不足的現象
②使用Intel AI Cloud 加速訓練
如上文所述，在本地訓練數據是一個巨大的運算量，常常需要幾周或幾個月，因此使用Intel提供的雲服務器可以極大縮短訓練的時間
在terminal中使用如下語句登陸到AI Cloud 服務器
ssh colfax mkdir dogsvscats
登陸成功后即顯示
######################################################################## # Welcome to Intel AI DevCloud! ########################################################################
將訓練dogsvscats工程所需的數據集及shell命令上傳到服務器（請根據實際目錄進行調整,若上傳速度較慢請嘗試雲服務器wget直接下載開放數據集）
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/
將對應的shell文件和Makefile上傳到服務器用於訓練數據預處理（請根據實際目錄進行調整）
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進行預處理，由於Makefile中deps含有sudo apt-get -y install unzip和sudo pip install pyyaml，且sudo apt-get在AI Cloud中無法運行
vi Makefile
將deps更改為
@echo "Installing dependencies..." # sudo apt-get -y install unzip # sudo pip install pyyaml
:wq!保存后退出，創建任務用於數據預處理
vi data_process.sh
在打開的界面中輸入（請根據實際目錄進行調整）
echo "Start Data Process" cd /home/u14673/ncappzoo/apps/dogsvscats/ make all echo "Data Process Finished"
:wq!保存后退出，提交任務開始數據預處理
qsub data_process.sh
使用qstat可以查看任務完成的情況，完成后會在當前目錄中生成對應的日志文件
將訓練所需的prototxt及預訓練模型上傳至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/
創建任務用於數據訓練
vi data_train.sh
在打開的界面中輸入如下語句（請根據實際目錄進行調整）
cd /home/u14673/ncappzoo/apps/dogsvscats/ echo 'Start Trainning'
# >&表示所有的標准輸出和標准錯誤輸出都將被重定向 /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
關於caffe train命令的定義，標准的范例如下
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為必要的參數，配置solver文件
如果從頭開始訓練模型，則無需配置--weights
如果從快照中恢復，則需要配置--snapshot--weights
如果是fine-tuning，則需要配置
:wq!保存后退出，提交任務開始訓練，訓練完成后在當前目錄可以看到日志文件
qsub data_train.sh
查看日志，日志保存在 bvlc_googlenet/org 目錄，使用如下命令將數據拷貝到本地
scp colfax:/home/u14673/ncappzoo/apps/dogsvscats/bvlc_googlenet/org/train.log ./
使用caffe自帶的工具繪制（位於caffe/tools/extra目錄）訓練數據，caffe中支持很多種曲線繪制，通過指定不同的類型參數即可，具體參數如下
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
解析日志並生成Test accuracy vs. Seconds曲線(實際應該為Test Loss，參考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
從上述兩張圖來看，似乎訓練過程並未收斂，對於這一問題，原作者並未給出原因，而是建議去掉--weights重新進行訓練
/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
將訓練后的模型拷貝到本地
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/
本地機器（需要安裝NCSDK）查看網絡分析，大致可以得到如下的圖形，顯示了各層連接的帶寬和運行時間
mvNCProfile -s 12 deploy.prototxt -w bvlc_googlenet_iter_40000.caffemodel
firefox output_report.html
- 模型調優
作者對比了dogsvscats例程中改進的網絡和GoogLenet原始網絡，通過Caffe自帶的Python工具分別繪制對應網絡拓撲
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自帶的工具draw_net.py時可能會遇到如下錯誤
#錯誤1 ImportError: No module named google.protobuf (沒有安裝python-protobuf)
sudo apt-get install python-protobuf
#錯誤２ ImportError: No module named _caffe (caffe源碼編譯的時候沒有編譯pycaffe)
cd ~/Downloads/caffe/
sudo make pycaffe
#錯誤3 ImportError: No module named skimage.io (沒有安裝python-skimage)
sudo apt-get install python-skimage
#錯誤4 ImportError: No module named pydot (沒有安裝python-pydot)
sudo apt install python-pydot python-pydot-ng graphviz
- 模型部署
根據最新訓練的結果，生成graph文件
cd ~/workspace/ncappzoo/apps/dogsvscats/bvlc_googlenet/org (由於前面訓練過程未能收斂，使用該模型預測時會出現Ｎａｎ的結果)
mvNCCompile -s 12 deploy.prototxt -w bvlc_googlenet_iter_40000.caffemodel -o dogsvscats-org.graph
cd ~/workspace/ncappzoo/apps/dogsvscats/bvlc_googlenet/custom （定制優化后的網絡）
mvNCCompile -s 12 deploy.prototxt -w bvlc_googlenet_iter_40000.caffemodel -o dogsvscats-org.graph
- 模型測試
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測試准確率(注意是python3，使用python image-classifier.py時會報錯，具體原因待查明)
cd ~/Downloads/ncappzoo/apps/image-classifier
python3 image-classifier.py
得到結果如下
------- predictions -------- Prediction for : dog with 100.0% confidence in 89.67 ms Ｐrediction for : cat with 0.0% confidence in 89.67 ms
那么關於本步驟部署，系統具體作了哪些事情呢，深入查看image-classifier.py我們可以得知
# ---- Step 1: Open the enumerated device and get a handle to it ------------- # 枚舉Movidius神經元計算棒 # ---- Step 2: Load a graph file onto the NCS device ------------------------- # 加載graph文件 # ---- Step 3: Offload image onto the NCS to run inference ------------------- # 加載image文件 # ---- Step 4: Read & print inference results from the NCS ------------------- # 讀取並打印運算結果 # ---- Step 5: Unload the graph and close the device ------------------------- # 關閉神經元計算棒
- 模型調優
- 參考文獻：
1.https://movidius.github.io/blog/deploying-custom-caffe-models
2.https://communities.intel.com/community/tech/intel-ai-academy