mxnet是最近流行的机器学习框架之一，使用起来体验不错，不过平常都是用python接口写程序，本文介绍如何在linux下从源码编译mxnet并使用其C++接口编程。
本文所使用的环境是ubunt14.04，g++4.8，如果是其他类unix发行版（fedora，mac os）同理。
目标
- 编译出libmxnet.a，libmxnet.so，本文只编译cpu版
- 链接共享库，调用C++接口
构建依赖
最小构建条件：
- 最新支持C++11的编译器
- blas库（比如libblas，atlas，openblas），opencv库
可选条件：
- CUDA Toolkit>=7.0（nvidia GPU， compute capability>=2.0）
- cudnn>=3，加速GPU computation
编译步骤
安装依赖：
sudo apt-get update sudo apt-get install -y build-essential git libatlas-base-dev libopencv-dev
如果需要使用GPU版，还需要安装cudnn，cuda
修改配置：
config.mk修改
必选，生成op.h文件
USE_CPP_PACKAGE=1
可选，cuda相关
USE_CUDA = 1 USE_CUDA_PATH = /usr/local/cuda # 根据实际目录 USE_CUDNN = 1
构建mxnet：
git clone --recursive https://github.com/dmlc/mxnet cd mxnet; cp make/config.mk . make -j4
编译完成后，会在mxnet根目录的lib目录生成libmxnet.a和libmxnet.so
注意:
- 一定要使用git --recursive方式下载代码，因为很多目录里面代码是递归下载的，如果直接在github网页点击下载会导致代码补全，缺少很多库。
- 如果报internal compiler error: Killed (program cc1plus)这个错误，是因为内存不足，解决方法：stackoverflow
编写程序
在前面的链接库（静态和动态）编译好后，建立一个cpp工程，引入mxnet相关头文件和共享库，编写测试
这里跑一个机器学习的经典例子：mnist手写数据识别，需要预先下载好mnist数据集（地址：mnist）
工程结构：
mxnet_cpp_test项目目录
.
├── CMakeLists.txt
└── src
└── main.cpp
CMakeLists.txt
project(mxnet_cpp_test)
cmake_minimum_required(VERSION 2.8)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -g -std=c++11 -W")
include_directories(
/home/user/codetest/mxnet/include
/home/user/codetest/mxnet/dmlc-core/include
/home/user/codetest/mxnet/nnvm/include
/home/user/codetest/mxnet/cpp-package/include
)
aux_source_directory(./src DIR_SRCS)
link_directories(/home/user/codetest/mxnet/lib)
add_executable(mxnet_cpp_test ${DIR_SRCS})
target_link_libraries(mxnet_cpp_test mxnet)
main.cpp
#include <chrono>
#include "mxnet-cpp/MxNetCpp.h"
using namespace std;
using namespace mxnet::cpp;
Symbol mlp(const vector<int> &layers)
{
auto x = Symbol::Variable("X");
auto label = Symbol::Variable("label");
vector<Symbol> weights(layers.size());
vector<Symbol> biases(layers.size());
vector<Symbol> outputs(layers.size());
for (size_t i = 0; i < layers.size(); ++i)
{
weights[i] = Symbol::Variable("w" + to_string(i));
biases[i] = Symbol::Variable("b" + to_string(i));
Symbol fc = FullyConnected(
i == 0 ? x : outputs[i - 1], // data
weights[i],
biases[i],
layers[i]);
outputs[i] = i == layers.size() - 1 ? fc : Activation(fc, ActivationActType::kRelu);
}
return SoftmaxOutput(outputs.back(), label);
}
int main(int argc, char** argv)
{
const int image_size = 28;
const vector<int> layers{ 128, 64, 10 };
const int batch_size = 100;
const int max_epoch = 10;
const float learning_rate = 0.1;
const float weight_decay = 1e-2;
auto train_iter = MXDataIter("MNISTIter")
.SetParam("image", "./mnist_data/train-images.idx3-ubyte")
.SetParam("label", "./mnist_data/train-labels.idx1-ubyte")
.SetParam("batch_size", batch_size)
.SetParam("flat", 1)
.CreateDataIter();
auto val_iter = MXDataIter("MNISTIter")
.SetParam("image", "./mnist_data/t10k-images.idx3-ubyte")
.SetParam("label", "./mnist_data/t10k-labels.idx1-ubyte")
.SetParam("batch_size", batch_size)
.SetParam("flat", 1)
.CreateDataIter();
auto net = mlp(layers);
Context ctx = Context::cpu(); // Use CPU for training
//Context ctx = Context::gpu();
std::map<string, NDArray> args;
args["X"] = NDArray(Shape(batch_size, image_size*image_size), ctx);
args["label"] = NDArray(Shape(batch_size), ctx);
// Let MXNet infer shapes other parameters such as weights
net.InferArgsMap(ctx, &args, args);
// Initialize all parameters with uniform distribution U(-0.01, 0.01)
auto initializer = Uniform(0.01);
for (auto& arg : args)
{
// arg.first is parameter name, and arg.second is the value
initializer(arg.first, &arg.second);
}
// Create sgd optimizer
Optimizer* opt = OptimizerRegistry::Find("sgd");
opt->SetParam("rescale_grad", 1.0 / batch_size)
->SetParam("lr", learning_rate)
->SetParam("wd", weight_decay);
// Create executor by binding parameters to the model
auto *exec = net.SimpleBind(ctx, args);
auto arg_names = net.ListArguments();
// Start training
for (int iter = 0; iter < max_epoch; ++iter)
{
int samples = 0;
train_iter.Reset();
auto tic = chrono::system_clock::now();
while (train_iter.Next())
{
samples += batch_size;
auto data_batch = train_iter.GetDataBatch();
// Set data and label
data_batch.data.CopyTo(&args["X"]);
data_batch.label.CopyTo(&args["label"]);
// Compute gradients
exec->Forward(true);
exec->Backward();
// Update parameters
for (size_t i = 0; i < arg_names.size(); ++i)
{
if (arg_names[i] == "X" || arg_names[i] == "label") continue;
opt->Update(i, exec->arg_arrays[i], exec->grad_arrays[i]);
}
}
auto toc = chrono::system_clock::now();
Accuracy acc;
val_iter.Reset();
while (val_iter.Next())
{
auto data_batch = val_iter.GetDataBatch();
data_batch.data.CopyTo(&args["X"]);
data_batch.label.CopyTo(&args["label"]);
// Forward pass is enough as no gradient is needed when evaluating
exec->Forward(false);
acc.Update(data_batch.label, exec->outputs[0]);
}
float duration = chrono::duration_cast<chrono::milliseconds>(toc - tic).count() / 1000.0;
LG << "Epoch: " << iter << " " << samples / duration << " samples/sec Accuracy: " << acc.Get();
}
delete exec;
MXNotifyShutdown();
return 0;
}
使用cmake构建makefile，编译时会链接libmxnet.a（或者libmxnet.so），在cmake的生成目录下可以看到libmxnet.so被拷贝进去了，在编译生成的目录中需要将mnist数据拷贝进去，供程序load进来进行训练
编译生成目录结构：
. ├── libmxnet.so ├── Makefile ├── mnist_data │ ├── t10k-images.idx3-ubyte │ ├── t10k-labels.idx1-ubyte │ ├── train-images.idx3-ubyte │ └── train-labels.idx1-ubyte ├── mxnet_cpp_test
运行结果：
MNISTIter: load 60000 images, shuffle=1, shape=(100,784) MNISTIter: load 10000 images, shuffle=1, shape=(100,784) Epoch: 0 25178.3 samples/sec Accuracy: 0.1135 Epoch: 1 24340.8 samples/sec Accuracy: 0.536 Epoch: 2 25031.3 samples/sec Accuracy: 0.8278 Epoch: 3 25466.9 samples/sec Accuracy: 0.8729 Epoch: 4 25370 samples/sec Accuracy: 0.9042 Epoch: 5 23819 samples/sec Accuracy: 0.9159 Epoch: 6 24067.4 samples/sec Accuracy: 0.9229 Epoch: 7 26513.5 samples/sec Accuracy: 0.9287 Epoch: 8 25575.4 samples/sec Accuracy: 0.9335 Epoch: 9 25619.1 samples/sec Accuracy: 0.9371可以感受到用C++直接写mxnet训练程序运行速度非常快