Caffe 深度學習框架上手教程

Caffe (CNN, deep learning) 介紹

Caffe -----------Convolution Architecture For Feature Embedding (Extraction)

1. Caffe 是什麽東東？
   • CNN (Deep Learning) 工具箱
   • C++ 語言架構
   • CPU 和GPU 無縫交換
   • Python 和matlab的封裝
   • 但是，Decaf只是CPU 版本。

2. 為什麽要用Caffe？

   • 運算速度快。簡單 友好的架構 用到的一些庫：
   • Google Logging library (Glog): 一個C++語言的應用級日誌記錄框架，提供了C++風格的流操作和各種助手宏.
   • lebeldb(數據存儲)： 是一個google實現的非常高效的kv數據庫，單進程操作。
   • CBLAS library（CPU版本的矩陣操作）
   • CUBLAS library (GPU 版本的矩陣操作)

3. Caffe 架構

1. 預處理圖像的leveldb構建
   輸入：一批圖像和label （2和3）
   輸出：leveldb （4）
   指令裏包含如下信息：
   • conver_imageset （構建leveldb的可運行程序）
   • train/ （此目錄放處理的jpg或者其他格式的圖像)
   • label.txt (圖像文件名及其label信息)
   • 輸出的leveldb文件夾的名字
   • CPU/GPU (指定是在cpu上還是在gpu上運行code)

2. CNN網絡配置文件

   • Imagenet_solver.prototxt （包含全局參數的配置的文件）
   • Imagenet.prototxt （包含訓練網絡的配置的文件）
   • Imagenet_val.prototxt （包含測試網絡的配置文件）

Caffe深度學習之圖像分類模型AlexNet解讀

在imagenet上的圖像分類challenge上Alex提出的alexnet網絡結構模型贏得了2012屆的冠軍。要研究CNN類型DL網絡模型在圖像分類上的應用，就逃不開研究alexnet，這是CNN在圖像分類上的經典模型（DL火起來之後）。

在DL開源實現caffe的model樣例中，它也給出了alexnet的復現，具體網絡配置文件如下train_val.prototxt452

接下來本文將一步步對該網絡配置結構中各個層進行詳細的解讀（訓練階段）：

1. conv1階段DFD（data flow diagram）：
   

2. conv2階段DFD（data flow diagram）：
   

3. conv3階段DFD（data flow diagram）：

   SouthEast2455x274

4. conv4階段DFD（data flow diagram）：
   

5. conv5階段DFD（data flow diagram）：
   

6. fc6階段DFD（data flow diagram）：
   

7. fc7階段DFD（data flow diagram）：
   

8. fc8階段DFD（data flow diagram）：
   

各種layer的operation更多解釋可以參考Caffe Layer Catalogue356

從計算該模型的數據流過程中，該模型參數大概5kw+。

caffe的輸出中也有包含這塊的內容日誌，詳情如下：

I0721 10:38:15.326920  4692 net.cpp:125] Top shape: 256 3 227 227 (39574272)
I0721 10:38:15.326971  4692 net.cpp:125] Top shape: 256 1 1 1 (256)
I0721 10:38:15.326982  4692 net.cpp:156] data does not need backward computation.
I0721 10:38:15.327003  4692 net.cpp:74] Creating Layer conv1
I0721 10:38:15.327011  4692 net.cpp:84] conv1 <- data
I0721 10:38:15.327033  4692 net.cpp:110] conv1 -> conv1
I0721 10:38:16.721956  4692 net.cpp:125] Top shape: 256 96 55 55 (74342400)
I0721 10:38:16.722030  4692 net.cpp:151] conv1 needs backward computation.
I0721 10:38:16.722059  4692 net.cpp:74] Creating Layer relu1
I0721 10:38:16.722070  4692 net.cpp:84] relu1 <- conv1
I0721 10:38:16.722082  4692 net.cpp:98] relu1 -> conv1 (in-place)
I0721 10:38:16.722096  4692 net.cpp:125] Top shape: 256 96 55 55 (74342400)
I0721 10:38:16.722105  4692 net.cpp:151] relu1 needs backward computation.
I0721 10:38:16.722116  4692 net.cpp:74] Creating Layer pool1
I0721 10:38:16.722125  4692 net.cpp:84] pool1 <- conv1
I0721 10:38:16.722133  4692 net.cpp:110] pool1 -> pool1
I0721 10:38:16.722167  4692 net.cpp:125] Top shape: 256 96 27 27 (17915904)
I0721 10:38:16.722187  4692 net.cpp:151] pool1 needs backward computation.
I0721 10:38:16.722205  4692 net.cpp:74] Creating Layer norm1
I0721 10:38:16.722221  4692 net.cpp:84] norm1 <- pool1
I0721 10:38:16.722234  4692 net.cpp:110] norm1 -> norm1
I0721 10:38:16.722251  4692 net.cpp:125] Top shape: 256 96 27 27 (17915904)
I0721 10:38:16.722260  4692 net.cpp:151] norm1 needs backward computation.
I0721 10:38:16.722272  4692 net.cpp:74] Creating Layer conv2
I0721 10:38:16.722280  4692 net.cpp:84] conv2 <- norm1
I0721 10:38:16.722290  4692 net.cpp:110] conv2 -> conv2
I0721 10:38:16.725225  4692 net.cpp:125] Top shape: 256 256 27 27 (47775744)
I0721 10:38:16.725242  4692 net.cpp:151] conv2 needs backward computation.
I0721 10:38:16.725253  4692 net.cpp:74] Creating Layer relu2
I0721 10:38:16.725261  4692 net.cpp:84] relu2 <- conv2
I0721 10:38:16.725270  4692 net.cpp:98] relu2 -> conv2 (in-place)
I0721 10:38:16.725280  4692 net.cpp:125] Top shape: 256 256 27 27 (47775744)
I0721 10:38:16.725288  4692 net.cpp:151] relu2 needs backward computation.
I0721 10:38:16.725298  4692 net.cpp:74] Creating Layer pool2
I0721 10:38:16.725307  4692 net.cpp:84] pool2 <- conv2
I0721 10:38:16.725317  4692 net.cpp:110] pool2 -> pool2
I0721 10:38:16.725329  4692 net.cpp:125] Top shape: 256 256 13 13 (11075584)
I0721 10:38:16.725338  4692 net.cpp:151] pool2 needs backward computation.
I0721 10:38:16.725358  4692 net.cpp:74] Creating Layer norm2
I0721 10:38:16.725368  4692 net.cpp:84] norm2 <- pool2
I0721 10:38:16.725378  4692 net.cpp:110] norm2 -> norm2
I0721 10:38:16.725389  4692 net.cpp:125] Top shape: 256 256 13 13 (11075584)
I0721 10:38:16.725399  4692 net.cpp:151] norm2 needs backward computation.
I0721 10:38:16.725409  4692 net.cpp:74] Creating Layer conv3
I0721 10:38:16.725419  4692 net.cpp:84] conv3 <- norm2
I0721 10:38:16.725427  4692 net.cpp:110] conv3 -> conv3
I0721 10:38:16.735193  4692 net.cpp:125] Top shape: 256 384 13 13 (16613376)
I0721 10:38:16.735213  4692 net.cpp:151] conv3 needs backward computation.
I0721 10:38:16.735224  4692 net.cpp:74] Creating Layer relu3
I0721 10:38:16.735234  4692 net.cpp:84] relu3 <- conv3
I0721 10:38:16.735242  4692 net.cpp:98] relu3 -> conv3 (in-place)
I0721 10:38:16.735250  4692 net.cpp:125] Top shape: 256 384 13 13 (16613376)
I0721 10:38:16.735258  4692 net.cpp:151] relu3 needs backward computation.
I0721 10:38:16.735302  4692 net.cpp:74] Creating Layer conv4
I0721 10:38:16.735312  4692 net.cpp:84] conv4 <- conv3
I0721 10:38:16.735321  4692 net.cpp:110] conv4 -> conv4
I0721 10:38:16.743952  4692 net.cpp:125] Top shape: 256 384 13 13 (16613376)
I0721 10:38:16.743988  4692 net.cpp:151] conv4 needs backward computation.
I0721 10:38:16.744000  4692 net.cpp:74] Creating Layer relu4
I0721 10:38:16.744010  4692 net.cpp:84] relu4 <- conv4
I0721 10:38:16.744020  4692 net.cpp:98] relu4 -> conv4 (in-place)
I0721 10:38:16.744030  4692 net.cpp:125] Top shape: 256 384 13 13 (16613376)
I0721 10:38:16.744038  4692 net.cpp:151] relu4 needs backward computation.
I0721 10:38:16.744050  4692 net.cpp:74] Creating Layer conv5
I0721 10:38:16.744057  4692 net.cpp:84] conv5 <- conv4
I0721 10:38:16.744067  4692 net.cpp:110] conv5 -> conv5
I0721 10:38:16.748935  4692 net.cpp:125] Top shape: 256 256 13 13 (11075584)
I0721 10:38:16.748955  4692 net.cpp:151] conv5 needs backward computation.
I0721 10:38:16.748965  4692 net.cpp:74] Creating Layer relu5
I0721 10:38:16.748975  4692 net.cpp:84] relu5 <- conv5
I0721 10:38:16.748983  4692 net.cpp:98] relu5 -> conv5 (in-place)
I0721 10:38:16.748998  4692 net.cpp:125] Top shape: 256 256 13 13 (11075584)
I0721 10:38:16.749011  4692 net.cpp:151] relu5 needs backward computation.
I0721 10:38:16.749022  4692 net.cpp:74] Creating Layer pool5
I0721 10:38:16.749030  4692 net.cpp:84] pool5 <- conv5
I0721 10:38:16.749039  4692 net.cpp:110] pool5 -> pool5
I0721 10:38:16.749050  4692 net.cpp:125] Top shape: 256 256 6 6 (2359296)
I0721 10:38:16.749058  4692 net.cpp:151] pool5 needs backward computation.
I0721 10:38:16.749074  4692 net.cpp:74] Creating Layer fc6
I0721 10:38:16.749083  4692 net.cpp:84] fc6 <- pool5
I0721 10:38:16.749091  4692 net.cpp:110] fc6 -> fc6
I0721 10:38:17.160079  4692 net.cpp:125] Top shape: 256 4096 1 1 (1048576)
I0721 10:38:17.160148  4692 net.cpp:151] fc6 needs backward computation.
I0721 10:38:17.160166  4692 net.cpp:74] Creating Layer relu6
I0721 10:38:17.160177  4692 net.cpp:84] relu6 <- fc6
I0721 10:38:17.160190  4692 net.cpp:98] relu6 -> fc6 (in-place)
I0721 10:38:17.160202  4692 net.cpp:125] Top shape: 256 4096 1 1 (1048576)
I0721 10:38:17.160212  4692 net.cpp:151] relu6 needs backward computation.
I0721 10:38:17.160222  4692 net.cpp:74] Creating Layer drop6
I0721 10:38:17.160230  4692 net.cpp:84] drop6 <- fc6
I0721 10:38:17.160238  4692 net.cpp:98] drop6 -> fc6 (in-place)
I0721 10:38:17.160258  4692 net.cpp:125] Top shape: 256 4096 1 1 (1048576)
I0721 10:38:17.160265  4692 net.cpp:151] drop6 needs backward computation.
I0721 10:38:17.160277  4692 net.cpp:74] Creating Layer fc7
I0721 10:38:17.160286  4692 net.cpp:84] fc7 <- fc6
I0721 10:38:17.160295  4692 net.cpp:110] fc7 -> fc7
I0721 10:38:17.342094  4692 net.cpp:125] Top shape: 256 4096 1 1 (1048576)
I0721 10:38:17.342157  4692 net.cpp:151] fc7 needs backward computation.
I0721 10:38:17.342175  4692 net.cpp:74] Creating Layer relu7
I0721 10:38:17.342185  4692 net.cpp:84] relu7 <- fc7
I0721 10:38:17.342198  4692 net.cpp:98] relu7 -> fc7 (in-place)
I0721 10:38:17.342208  4692 net.cpp:125] Top shape: 256 4096 1 1 (1048576)
I0721 10:38:17.342217  4692 net.cpp:151] relu7 needs backward computation.
I0721 10:38:17.342228  4692 net.cpp:74] Creating Layer drop7
I0721 10:38:17.342236  4692 net.cpp:84] drop7 <- fc7
I0721 10:38:17.342245  4692 net.cpp:98] drop7 -> fc7 (in-place)
I0721 10:38:17.342254  4692 net.cpp:125] Top shape: 256 4096 1 1 (1048576)
I0721 10:38:17.342262  4692 net.cpp:151] drop7 needs backward computation.
I0721 10:38:17.342274  4692 net.cpp:74] Creating Layer fc8
I0721 10:38:17.342283  4692 net.cpp:84] fc8 <- fc7
I0721 10:38:17.342291  4692 net.cpp:110] fc8 -> fc8
I0721 10:38:17.343199  4692 net.cpp:125] Top shape: 256 22 1 1 (5632)
I0721 10:38:17.343214  4692 net.cpp:151] fc8 needs backward computation.
I0721 10:38:17.343231  4692 net.cpp:74] Creating Layer loss
I0721 10:38:17.343240  4692 net.cpp:84] loss <- fc8
I0721 10:38:17.343250  4692 net.cpp:84] loss <- label
I0721 10:38:17.343264  4692 net.cpp:151] loss needs backward computation.
I0721 10:38:17.343305  4692 net.cpp:173] Collecting Learning Rate and Weight Decay.
I0721 10:38:17.343327  4692 net.cpp:166] Network initialization done.
I0721 10:38:17.343335  4692 net.cpp:167] Memory required for Data 1073760256

CIFAR-10在caffe上進行訓練與學習

使用數據庫：CIFAR-10

60000張 32X32 彩色圖像 10類，50000張訓練，10000張測試

準備

在終端運行以下指令：

cd $CAFFE_ROOT/data/cifar10
./get_cifar10.sh
cd $CAFFE_ROOT/examples/cifar10
./create_cifar10.sh

其中CAFFE_ROOT是caffe-master在你機子的地址

運行之後，將會在examples中出現數據庫文件./cifar10-leveldb和數據庫圖像均值二進制文件./mean.binaryproto

模型

該CNN由卷積層，POOLing層，非線性變換層，在頂端的局部對比歸一化線性分類器組成。該模型的定義在CAFFE_ROOT/examples/cifar10 directory’s cifar10_quick_train.prototxt中，可以進行修改。其實後綴為prototxt很多都是用來修改配置的。

訓練和測試

訓練這個模型非常簡單，當我們寫好參數設置的文件cifar10_quick_solver.prototxt和定義的文件cifar10_quick_train.prototxt和cifar10_quick_test.prototxt後，運行train_quick.sh或者在終端輸入下面的命令：

cd $CAFFE_ROOT/examples/cifar10
./train_quick.sh

即可，train_quick.sh是一個簡單的腳本，會把執行的信息顯示出來，培訓的工具是train_net.bin,cifar10_quick_solver.prototxt作為參數。

然後出現類似以下的信息：這是搭建模型的相關信息

I0317 21:52:48.945710 2008298256 net.cpp:74] Creating Layer conv1
I0317 21:52:48.945716 2008298256 net.cpp:84] conv1 <- data
I0317 21:52:48.945725 2008298256 net.cpp:110] conv1 -> conv1
I0317 21:52:49.298691 2008298256 net.cpp:125] Top shape: 100 32 32 32 (3276800)
I0317 21:52:49.298719 2008298256 net.cpp:151] conv1 needs backward computation.

接著：

0317 21:52:49.309370 2008298256 net.cpp:166] Network initialization done.
I0317 21:52:49.309376 2008298256 net.cpp:167] Memory required for Data 23790808
I0317 21:52:49.309422 2008298256 solver.cpp:36] Solver scaffolding done.
I0317 21:52:49.309447 2008298256 solver.cpp:47] Solving CIFAR10_quick_train

之後，訓練開始

I0317 21:53:12.179772 2008298256 solver.cpp:208] Iteration 100, lr = 0.001
I0317 21:53:12.185698 2008298256 solver.cpp:65] Iteration 100, loss = 1.73643
...
I0317 21:54:41.150030 2008298256 solver.cpp:87] Iteration 500, Testing net
I0317 21:54:47.129461 2008298256 solver.cpp:114] Test score #0: 0.5504
I0317 21:54:47.129500 2008298256 solver.cpp:114] Test score #1: 1.27805

其中每100次叠代次數顯示一次訓練時lr(learningrate),和loss（訓練損失函數），每500次測試一次，輸出score 0（準確率）和score 1（測試損失函數）

當5000次叠代之後，正確率約為75%，模型的參數存儲在二進制protobuf格式在cifar10_quick_iter_5000

然後，這個模型就可以用來運行在新數據上了。

其他

另外，更改cifar*solver.prototxt文件可以使用CPU訓練，

# solver mode: CPU or GPU
solver_mode: CPU

可以看看CPU和GPU訓練的差別。

主要資料來源：caffe官網教程

Caffe 深度學習框架上手教程