zhuwei3014的个人博客分享 http://blog.sciencenet.cn/u/zhuwei3014

博文

Caffe 实例测试一: MNIST

已有 27696 次阅读 2014-11-13 19:19 |个人分类:Caffe|系统分类:科研笔记|关键词:学者| Learning, deep, caffe, MNIST

参考:http://caffe.berkeleyvision.org/gathered/examples/mnist.html


原笔记在印象笔记中,复制过来格式就乱了,原笔记地址:

http://app.yinxiang.com/l/AARGaZ2rlLdEn7ikhvPcGA7hd7_EnB2OlIo/

 

caffe安装目录:home/user/caffe-master
cd caffe-master
download MNIST dataset
注意:新版caffe都需要从根目录上执行,不然可能会遇到这个错误:
./create_mnist.sh: 16: ./create_mnist.sh: build/examples/mnist/convert_mnist_data.bin: not found
./data/mnist/get_mnist.sh

生成mnist-train-leveldb/ 和 mnist-test-leveldb/文件夹,这里包含了LDB格式的数据集
./examples/mnist/create_mnist.sh

LeNet:

caffe中用的模型结构是非常著名的手写体识别模型LeNet,唯一的区别是把其中的sigmoid激活函数换成了ReLU,整个结构中包含两个convolution layer、两个pooling layer和两个fully connected layer。结构定义在$caffe-master/examples/mnist/lenet_train_test.prototxt中。
定义MNIST Network:

该结构定义在lenet_train_test.prototxt中,需要对google protobuf有一定了解并且看过Caffe中protobuf的定义,其定义在$caffe-master/src/caffe/proto/caffe.proto。
protobuf简介:
protobuf是google公司的一个开源项目,主要功能是把某种数据结构的信息以某种格式保存及传递,类似微软的XML,但是效率较高。目前提供C++、java和python的API。
protobuf简介:http://blog.163.com/jiang_tao_2010/blog/static/12112689020114305013458/
使用实例       :http://www.ibm.com/developerworks/cn/linux/l-cn-gpb/
编写LeNet的protobuf:
1、首先给定一个名字
name: "LeNet"
2、数据层
从生成的lmdb文件中读取MNIST数据,定义数据层如下:
layers {  name: "mnist"  type: DATA  data_param {    source: "mnist_train_lmdb"    backend: LMDB    batch_size: 64              scale: 0.00390625     //归一化到[0,1)  }  top: "data"  top: "label"}
3、卷积层
layers {  name: "conv1"  type: CONVOLUTION  blobs_lr: 1.         // weight learning rate, same as learning rate    blobs_lr: 2.         // bias learning rate, twice as learning rate  convolution_param {    num_output: 20     // 20 channels    kernelsize: 5    stride: 1    weight_filler {      type: "xavier"   // xavier algorithm, automatically initialize weight filler    }    bias_filler {      type: "constant" // initialize bias filler as constant, default value 0    }  }  bottom: "data"  top: "conv1" }
4、pooling层
layers {  name: "pool1"  type: POOLING  pooling_param {    kernel_size: 2    stride: 2    pool: MAX  }  bottom: "conv1"  top: "pool1"}
5、全连接层1
layers {  name: "ip1"  type: INNER_PRODUCT      // caffe calls it an innerprodect layer  blobs_lr: 1.  blobs_lr: 2.  inner_product_param {    num_output: 500        // 500 outputs    weight_filler {      type: "xavier"    }    bias_filler {      type: "constant"    }  }  bottom: "pool2"  top: "ip1"}
6、ReLU层
layers {  name: "relu1"  type: RELU  bottom: "ip1"  top: "ip1"}layers {7、全连接层2
layers {  name: "ip2"  type: INNER_PRODUCT  blobs_lr: 1.  blobs_lr: 2.  inner_product_param {    num_output: 10              // 10 outputs    weight_filler {      type: "xavier"    }    bias_filler {      type: "constant"    }  }  bottom: "ip1"  top: "ip2"}
8、loss层(softmax_loss layer)
这一层实现了softmax和多项式回归误差计算。
layers {  name: "loss"  type: SOFTMAX_LOSS  bottom: "ip2"  bottom: "label"}
9、additional notes
还可以在网络定义中添加规则,如下:
layers {  // ...layer definition...  include: { phase: TRAIN }}
另外还有一个accuracy layer,用来在测试阶段每100次迭代报告模型准确率。
更多的配置和规则可以在$caffe-master/src/caffe/proto/caffe.proto中查看。
可以在$caffe-master/examples/mnist/lenet_solver.prototxt中查看查看protobuf的解释:
# The train/test net protocol buffer definitionnet: "examples/mnist/lenet_train_test.prototxt"# test_iter specifies how many forward passes the test should carry out.# In the case of MNIST, we have test batch size 100 and 100 test iterations,# covering the full 10,000 testing images.test_iter: 100# Carry out testing every 500 training iterations.test_interval: 500# The base learning rate, momentum and the weight decay of the network.base_lr: 0.01momentum: 0.9weight_decay: 0.0005# The learning rate policylr_policy: "inv"gamma: 0.0001power: 0.75# Display every 100 iterationsdisplay: 100# The maximum number of iterationsmax_iter: 10000# snapshot intermediate resultssnapshot: 5000snapshot_prefix: "examples/mnist/lenet"# solver mode: CPU or GPUsolver_mode: GPU
接下来就是训练和测试LeNet模型:
cd $caffe-master./examples/mnist/train_lenet.sh
出现这样的结果:
I1203 net.cpp:66] Creating Layer conv1I1203 net.cpp:76] conv1 <- dataI1203 net.cpp:101] conv1 -> conv1I1203 net.cpp:116] Top shape: 20 24 24I1203 net.cpp:127] conv1 needs backward computation.    // 显示连接信息和输出信息
I1203 net.cpp:142] Network initialization done.I1203 solver.cpp:36] Solver scaffolding done.I1203 solver.cpp:44] Solving LeNet
I1203 solver.cpp:204] Iteration 100, lr = 0.00992565     // learning rate printed every 100 iterationsI1203 solver.cpp:66] Iteration 100, loss = 0.26044       // loss function ...I1203 solver.cpp:84] Testing netI1203 solver.cpp:111] Test score #0: 0.9785              // test accuracy every 1000 iterations  I1203 solver.cpp:111] Test score #1: 0.0606671           // test loss function
I1203 solver.cpp:84] Testing netI1203 solver.cpp:111] Test score #0: 0.9897I1203 solver.cpp:111] Test score #1: 0.0324599I1203 solver.cpp:126] Snapshotting to lenet_iter_10000I1203 solver.cpp:133] Snapshotting solver state to lenet_iter_10000.solverstateI1203 solver.cpp:78] Optimization Done.
最终训练完的模型存储为一个二进制的protobuf文件:
lenet_iter_10000
Caffe默认的训练是在GPU上的,但是如果想改为CPU上,只需要在lenet_solver.prototxt中修改一行:
# solver mode: CPU or GPUsolver_mode: CPU




https://m.sciencenet.cn/blog-1583812-843207.html

上一篇:Ubuntu+CUDA6.5+Caffe安装配置汇总
下一篇:DIY Deep Learning for Vision- a Hands-On Tutorial with Caffe

0

该博文允许注册用户评论 请点击登录 评论 (1 个评论)

数据加载中...

Archiver|手机版|科学网 ( 京ICP备07017567号-12 )

GMT+8, 2024-6-3 23:49

Powered by ScienceNet.cn

Copyright © 2007- 中国科学报社

返回顶部