caffe运行手写体数字识别例程
下载MNIST数据集
假设我们处在根目录(之后的文件夹操作我们都假定自己处于根目录下)，进入data/mnist，可以看到一个get_mnist.sh的脚本
#!/usr/bin/env sh
# This scripts downloads the mnist data and unzips it.
DIR="$( cd "$(dirname "$0")" ; pwd -P )"
cd "$DIR"
echo "Downloading..."
for fname in train-images-idx3-ubyte train-labels-idx1-ubyte t10k-images-idx3-ubyte t10k-labels-idx1-ubyte
do
if [ ! -e $fname ]; then
wget --no-check-certificate http://yann.lecun.com/exdb/mnist/${fname}.gz
gunzip ${fname}.gz
fi
done
可见该脚本实现从下载地址下载mnist数据集，bash get_mnist.sh执行该脚本，会将该数据集下载到当前目录下，该数据集的格式可以在相关的网站查询得到。
TRAINING SET LABEL FILE (train-labels-idx1-ubyte):
| [offset] | [type] | [value] | [description] |
|---|---|---|---|
| 0000 | 32 bit integer | 0x00000801(2049) | magic number (MSB first) |
| 0004 | 32 bit integer | 60000 | number of items |
| 0008 | unsigned byte | ?? | label |
| 0009 | unsigned byte | ?? | label |
| ........ | |||
| xxxx | unsigned byte | ?? | label |
TRAINING SET IMAGE FILE (train-images-idx3-ubyte):
| [offset] | [type] | [value] | [description] |
|---|---|---|---|
| 0000 | 32 bit integer | 0x00000803(2051) | magic number |
| 0004 | 32 bit integer | 60000 | number of images |
| 0008 | 32 bit integer | 28 | number of rows |
| 0012 | 32 bit integer | 28 | number of columns |
| 0016 | unsigned byte | ?? | pixel |
| 0017 | unsigned byte | ?? | pixel |
| ........ | |||
| xxxx | unsigned byte | ?? | pixel |
测试集格式同样类似
转换格式为LMDB
caffe对mnist样例写好了转换格式的代码，将原二进制格式转化为LMDB，在caffe根目录下执行./examples/mnist/create_mnist.sh，会发现在examples/mnist/mnist_train_lmdb和examples/mnist/mnist_test_lmdb两个目录，每个目录下都有两个文件：data.lmdb和lock.lmdb，由名称就可以知道一个为LMDB格式的训练集，一个为LMDB格式的测试集。
可以分析以下脚本了解是如何转换的。用vim打开create_mnist.sh
#!/usr/bin/env sh
# This script converts the mnist data into lmdb/leveldb format,
# depending on the value assigned to $BACKEND.
set -e
EXAMPLE=examples/mnist
DATA=data/mnist
BUILD=build/examples/mnist
BACKEND="lmdb"
echo "Creating ${BACKEND}..."
# 将原本examples/mnist中有关train和test的lmdb或leveldb格式的文件删除，重新转换
rm -rf $EXAMPLE/mnist_train_${BACKEND}
rm -rf $EXAMPLE/mnist_test_${BACKEND}
# 使用已经写好的转换工具进行类型转换
$BUILD/convert_mnist_data.bin $DATA/train-images-idx3-ubyte \
$DATA/train-labels-idx1-ubyte $EXAMPLE/mnist_train_${BACKEND} --backend=${BACKEND}
$BUILD/convert_mnist_data.bin $DATA/t10k-images-idx3-ubyte \
$DATA/t10k-labels-idx1-ubyte $EXAMPLE/mnist_test_${BACKEND} --backend=${BACKEND}
echo "Done."
由上面的脚本可知调用了build/examples/mnist下的convert_mnist_data.bin的编译成功的二进制工具，之前我们已经知道了make编译后的文件都存放在build目录下，所以我们根据目录的提示可以知道源码应在examples/mnist中，其源码应为convert_mnist_data.cpp
// This script converts the MNIST dataset to a lmdb (default) or
// leveldb (--backend=leveldb) format used by caffe to load data.
// Usage:
// convert_mnist_data [FLAGS] input_image_file input_label_file
// output_db_file
// The MNIST dataset could be downloaded at
// http://yann.lecun.com/exdb/mnist/
#include <gflags/gflags.h>
#include <glog/logging.h>
#include <google/protobuf/text_format.h>
#if defined(USE_LEVELDB) && defined(USE_LMDB)
#include <leveldb/db.h>
#include <leveldb/write_batch.h>
#include <lmdb.h>
#endif
#include <stdint.h>
#include <sys/stat.h>
#include <fstream> // NOLINT(readability/streams)
#include <string>
#include "boost/scoped_ptr.hpp"
#include "caffe/proto/caffe.pb.h"
#include "caffe/util/db.hpp"
#include "caffe/util/format.hpp"
#if defined(USE_LEVELDB) && defined(USE_LMDB)
using namespace caffe; // NOLINT(build/namespaces)
using boost::scoped_ptr;
using std::string;
// GFALGS工具定义命令行选项，默认值为lmdb，定义方式为: --backend=lmdb
DEFINE_string(backend, "lmdb", "The backend for storing the result");
// 大小端转换。MNIST原始数据文件中32为的整型值为大端存储，C/C++变量为小端存储，因此需要加入转换机制，MNIST文件中的说明为魔数
uint32_t swap_endian(uint32_t val) {
val = ((val << 8) & 0xFF00FF00) | ((val >> 8) & 0xFF00FF);
return (val << 16) | (val >> 16);
}
void convert_dataset(const char* image_filename, const char* label_filename,
const char* db_path, const string& db_backend) {
// Open files
// 用C++输入文件流以二进制方式打开文件
std::ifstream image_file(image_filename, std::ios::in | std::ios::binary);
std::ifstream label_file(label_filename, std::ios::in | std::ios::binary);
// CHECK宏来自于google的glog库，类似于标准库中的assert宏
// 给定条件不满时终止程序，条件不满时输出后面程序
CHECK(image_file) << "Unable to open file " << image_filename;
CHECK(label_file) << "Unable to open file " << label_filename;
// Read the magic and the meta data
uint32_t magic;
uint32_t num_items;
uint32_t num_labels;
uint32_t rows;
uint32_t cols;
// read()是以字符的类型读取，而MNIST的魔数由32位整型组成，并且使用大端存储
// 应读4个字符共32位，read()将4个字符读取到magic地址的缓存中
// 必要时进行强制类型转换，并将其转换为小端存储。总条目数，行数，列数同样如此
image_file.read(reinterpret_cast<char*>(&magic), 4);
magic = swap_endian(magic);
// 将大端存储转换为小端存储后
// 使用同样来自于Glog的CHECK_EQ来判断其是否与2051相等
// 标签文件则要判断其是否与2049相等
CHECK_EQ(magic, 2051) << "Incorrect image file magic.";
label_file.read(reinterpret_cast<char*>(&magic), 4);
magic = swap_endian(magic);
CHECK_EQ(magic, 2049) << "Incorrect label file magic.";
image_file.read(reinterpret_cast<char*>(&num_items), 4);
num_items = swap_endian(num_items);
label_file.read(reinterpret_cast<char*>(&num_labels), 4);
num_labels = swap_endian(num_labels);
CHECK_EQ(num_items, num_labels);
image_file.read(reinterpret_cast<char*>(&rows), 4);
rows = swap_endian(rows);
image_file.read(reinterpret_cast<char*>(&cols), 4);
cols = swap_endian(cols);
scoped_ptr<db::DB> db(db::GetDB(db_backend));
db->Open(db_path, db::NEW);
scoped_ptr<db::Transaction> txn(db->NewTransaction());
// Storing to db
// 将数据存储到db中
char label;
char* pixels = new char[rows * cols];
int count = 0;
string value;
// Datum是caffe.proto定义的message，这里用来存储图片
// 由于image_file是采用unsigned byte类型，Datum中也有相似的bytes类型
// 因此这里采用data来进行存取，Datum也支持float类型的数据
// C++里没有byte类型，但是有与之相似的char类型，用其保存byte
Datum datum;
datum.set_channels(1);
datum.set_height(rows);
datum.set_width(cols);
LOG(INFO) << "A total of " << num_items << " items.";
LOG(INFO) << "Rows: " << rows << " Cols: " << cols;
for (int item_id = 0; item_id < num_items; ++item_id) {
image_file.read(pixels, rows * cols);
label_file.read(&label, 1);
datum.set_data(pixels, rows*cols);
datum.set_label(label);
string key_str = caffe::format_int(item_id, 8);
datum.SerializeToString(&value);
txn->Put(key_str, value);
if (++count % 1000 == 0) {
txn->Commit();
}
}
// write the last batch
if (count % 1000 != 0) {
txn->Commit();
}
LOG(INFO) << "Processed " << count << " files.";
delete[] pixels;
db->Close();
}
int main(int argc, char** argv) {
#ifndef GFLAGS_GFLAGS_H_
namespace gflags = google;
#endif
FLAGS_alsologtostderr = 1;
gflags::SetUsageMessage("This script converts the MNIST dataset to\n"
"the lmdb/leveldb format used by Caffe to load data.\n"
"Usage:\n"
" convert_mnist_data [FLAGS] input_image_file input_label_file "
"output_db_file\n"
"The MNIST dataset could be downloaded at\n"
" http://yann.lecun.com/exdb/mnist/\n"
"You should gunzip them after downloading,"
"or directly use data/mnist/get_mnist.sh\n");
gflags::ParseCommandLineFlags(&argc, &argv, true);
// FLAGS_backend在前面通过DEFINE_string定义，是字符串类型
const string& db_backend = FLAGS_backend;
if (argc != 4) {
gflags::ShowUsageWithFlagsRestrict(argv[0],
"examples/mnist/convert_mnist_data");
} else {
google::InitGoogleLogging(argv[0]);
convert_dataset(argv[1], argv[2], argv[3], db_backend);
}
return 0;
}
#else
int main(int argc, char** argv) {
LOG(FATAL) << "This example requires LevelDB and LMDB; " <<
"compile with USE_LEVELDB and USE_LMDB.";
}
#endif // USE_LEVELDB and USE_LMDB
数据类型多种多样，不可能用一套代码实现所有类型输入数据的读取。转化成统一格式可以简化数据读取层的实现，另一方面，使用LMDB可以提高磁盘的IO利用率。
接下来我们用LeNet-5模型来进行训练，原版的LeNet-5模型稍有不同，比如将激活函数由Sigmoid改为ReLU，该模型存放在examples/mnist/lenet_train_test.prototxt中。使用python/draw_net.py可以查看网络结构。
训练超参数
caffe也为该例程提供了训练的超参数和脚本，查看examples/mnist/train_lenet.sh脚本。
#!/usr/bin/env sh
set -e
./build/tools/caffe train --solver=examples/mnist/lenet_solver.prototxt $@
该脚本使用了build/tools/caffe.bin的二进制程序(caffe是指向caffe.bin的)，并且使用examples/mnist/lenet_solver.prototxt作为求解器。
# The train/test net protocol buffer definition
net: "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: 0.01
momentum: 0.9
weight_decay: 0.0005
# The learning rate policy|学习速率的衰减策略
lr_policy: "inv"
gamma: 0.0001
power: 0.75
# Display every 100 iterations|每经过100次迭代，在屏幕上打印一次运行log
display: 100
# The maximum number of iterations
max_iter: 10000
# snapshot intermediate results|每5000次迭代打印一次快照
snapshot: 5000
snapshot_prefix: "examples/mnist/lenet"
# solver mode: CPU or GPU|Caffe求解为CPU模式。可以改为GPU模式
solver_mode: GPU
执行脚本后会在examples/mnist下生成四个文件，因为每5000次迭代会生成一次快照。具体caffe的运作机制可以在tools/caffe.cpp源代码中查看。
lenet_iter_10000.caffemodel
lenet_iter_10000.solverstate
lenet_iter_5000.caffemodel
lenet_iter_5000.solverstate
在测试中我们选择的模型依旧为lenet_train_test.prototxt，它既包括训练的模型，也包括测试的模型，权重选择迭代10000次后的模型lenet_iter_10000.caffemodel，选择100次迭代为一个batch，100个batch刚好可以测试完10000张图片，并且使用指定的gpu进行测试。
./build/tools/caffe test -model examples/mnist/lenet_train_test.prototxt -weights examples/mnist/lenet_iter_10000.caffemodel -iterations 100 -gpu 0
效果如下所示，正确率达到了99.07%。
I0116 17:55:27.196736 13017 caffe.cpp:309] Loss: 0.0284669
I0116 17:55:27.196743 13017 caffe.cpp:321] accuracy = 0.9907
I0116 17:55:27.196750 13017 caffe.cpp:321] loss = 0.0284669 (* 1 = 0.0284669 loss)
浙公网安备 33010602011771号