大多数情况下,mxnet都使用python接口进行深度学习程序的编写,方便快捷,但是有的时候,需要把机器学习训练和识别的程序部署到生产版的程序中去,比如游戏或者云服务,此时采用C++等高级语言去编写才能提高性能,本文介绍了如何在windows系统下从源码编译mxnet,安装python版的包,并使用C++原生接口创建示例程序。
目标
- 编译出libmxnet.lib和libmxnet.dll的gpu版本
- 从源码安装mxnet python包
- 构建mxnet C++示例程序
环境
- windows10
- vs2015
- cmake3.7.2
- Miniconda2(python2.7.14)
- CUDA8.0
- mxnet1.2
- opencv3.4.1
- OpenBLAS-v0.2.19-Win64-int32
- cudnn-8.0-windows10-x64-v7.1(如果编译cpu版本的mxnet,则此项不需要)
步骤
下载源码
最好用git下载,递归地下载所有依赖的子repo,源码的根目录为mxnet
git clone --recursive https://github.com/dmlc/mxnet
依赖库
在此之前确保cmake和python已经正常安装,并且添加到环境变量,然后再下载第三方依赖库
- 下载安装cuda,确保机器是英伟达显卡,且支持cuda,地址:https://developer.nvidia.com/cuda-toolkit
- 下载安装opencv预编译版,地址:https://sourceforge.net/projects/opencvlibrary/files/opencv-win/3.4.1/opencv-3.4.1-vc14_vc15.exe/download
- 下载openblas预编译版,地址:https://sourceforge.net/projects/openblas/files/v0.2.19/
- 下载cudnn预编译版,注意与cuda版本对应,地址:https://developer.nvidia.com/compute/machine-learning/cudnn/secure/v7.0.5/prod/8.0_20171129/cudnn-8.0-windows10-x64-v7
cmake配置
打开cmake-gui,配置源码目录和生成目录,编译器选择vs2015 win64
配置第三方依赖库
configure和generate
注意:在这一步中不要勾选cpppackage选项,不然会编译不过,因为这选项是为了生成op.h这个头文件的,但是由于官网文档的bug,目前无法通过配置cmake编译自动生成,后面会讲手动生成
编译vs工程
打开mxnet.sln,配置成release x64模式,编译整个solution
编译完成后会在对应文件夹生成mxnet的lib和dll
此时整个过程成功了一半
安装mxnet的python包
有了libmxnet.dll就可以同源码安装python版的mxnet包了
不过,前提是需要集齐所有依赖到的其他dll,如图所示,将这些dll全部拷贝到mxnet/python/mxnet目录下
tip: 关于dll的来源
- opencv,openblas,cudnn相关dll都是从这几个库的目录里拷过来的
- libgcc_s_seh-1.dll和libwinpthread-1.dll是从mingw相关的库目录里拷过来的,git,qt等这些目录都有
- libgfortran-3.dll和libquadmath_64-0.dll是从adda(https://github.com/adda-team/adda/releases)这个库里拷过来的,注意改名
然后,在mxnet/python目录下使用命令行安装mxnet的python包
python setup.py install
安装过程中,python会自动把对应的dll考到安装目录,正常安装完成后,在python中就可以 import mxnet 了
生成C++依赖头文件 op.h
为了能够使用C++原生接口,这一步是很关键的一步,目的是生成mxnet C++程序依赖的op.h文件
如果直接指向编译mxnet并使用C++,前面的python包安装可以不做
在mxnet/cpp-package/scripts目录,将所有依赖到的dll拷贝进来
在此目录运行命令行
python OpWrapperGenerator.py libmxnet.dll
正常情况下就可以在mxnet/cpp-package/include/mxnet-cpp目录下生成op.h了
如果这个过程中出现一些error,多半是dll文件缺失或者版本不对,很好解决
构建C++示例程序
建立cpp工程,这里使用经典的mnist手写数字识别训练示例(请提前下载好mnist数据,地址:mnist),启用GPU支持
选择release x64模式
配置include和lib目录以及附加依赖项
include目录包括:
- D:\mxnet\include
- D:\mxnet\dmlc-core\include
- D:\mxnet\nnvm\include
- D:\mxnet\cpp-package\include
lib目录:
- D:\mxnet\build_x64\Release
附加依赖项:
- libmxnet.lib
代码 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);
// start traning
cout << "==== mlp training begin ====" << endl;
auto start_time = chrono::system_clock::now();
Context ctx = Context::gpu(); // Use GPU for training
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 of 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);
std::unique_ptr<LRScheduler> lr_sch(new FactorScheduler(5000, 0.1));
opt->SetLRScheduler(std::move(lr_sch));
// Create executor by binding parameters to the model
auto *exec = net.SimpleBind(ctx, args);
auto arg_names = net.ListArguments();
// Create metrics
Accuracy train_acc, val_acc;
// Start training
for (int iter = 0; iter < max_epoch; ++iter)
{
int samples = 0;
train_iter.Reset();
train_acc.Reset();
auto tic = chrono::system_clock::now();
while (train_iter.Next())
{
samples += batch_size;
auto data_batch = train_iter.GetDataBatch();
// Data provided by DataIter are stored in memory, should be copied to GPU first.
data_batch.data.CopyTo(&args["X"]);
data_batch.label.CopyTo(&args["label"]);
// CopyTo is imperative, need to wait for it to complete.
NDArray::WaitAll();
// 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]);
}
// Update metric
train_acc.Update(data_batch.label, exec->outputs[0]);
}
// one epoch of training is finished
auto toc = chrono::system_clock::now();
float duration = chrono::duration_cast<chrono::milliseconds>(toc - tic).count() / 1000.0;
LG << "Epoch[" << iter << "] " << samples / duration \
<< " samples/sec " << "Train-Accuracy=" << train_acc.Get();;
val_iter.Reset();
val_acc.Reset();
while (val_iter.Next())
{
auto data_batch = val_iter.GetDataBatch();
data_batch.data.CopyTo(&args["X"]);
data_batch.label.CopyTo(&args["label"]);
NDArray::WaitAll();
// Only forward pass is enough as no gradient is needed when evaluating
exec->Forward(false);
val_acc.Update(data_batch.label, exec->outputs[0]);
}
LG << "Epoch[" << iter << "] Val-Accuracy=" << val_acc.Get();
}
// end training
auto end_time = chrono::system_clock::now();
float total_duration = chrono::duration_cast<chrono::milliseconds>(end_time - start_time).count() / 1000.0;
cout << "total duration: " << total_duration << " s" << endl;
cout << "==== mlp training end ====" << endl;
//delete exec;
MXNotifyShutdown();
getchar(); // wait here
return 0;
}
编译生成目录
- 预先把mnist数据拷进去,维持相对目录结构
- 在执行目录也要把所有依赖的dll拷贝进来
运行结果
在官方的example里面有mlp的cpu和gpu两个版本,有兴趣的话可以跑起来做一个对比
其实,在某些数据量小的情况下,gpu版本并不明显比cpu版本消耗的训练时间少
至此,大功告成