Cranium
1.0.0
matrix.h
中第 7 行的注释即可启用 BLAS sgemm
函数以进行快速矩阵乘法。 由于 Cranium 仅包含头文件,因此只需将src
目录复制到您的项目中,然后#include "src/cranium.h"
即可开始使用它。
它唯一需要的编译器依赖项来自
标头,因此使用-lm
进行编译。
如果您使用 CBLAS,您还需要使用-lcblas
进行编译,并通过-I
包含特定机器的 BLAS 实现所在的路径。常见的有 OpenBLAS 和 ATLAS。
它已经过测试,可以与任何级别的 gcc 优化完美配合,因此请随意使用它们。
#include "cranium.h"
/*
This basic example program is the skeleton of a classification problem.
The training data should be in matrix form, where each row is a data point, and
each column is a feature.
The training classes should be in matrix form, where the ith row corresponds to
the ith training example, and each column is a 1 if it is of that class, and
0 otherwise. Each example may only be of 1 class.
*/
// create training data and target values (data collection not shown)
int rows , features , classes ;
float * * training ;
float * * classes ;
// create datasets to hold the data
DataSet * trainingData = createDataSet ( rows , features , training );
DataSet * trainingClasses = createDataSet ( rows , classes , classes );
// create network with 2 input neurons, 1 hidden layer with sigmoid
// activation function and 5 neurons, and 2 output neurons with softmax
// activation function
srand ( time ( NULL ));
size_t hiddenSize [] = { 5 };
Activation hiddenActivation [] = { sigmoid };
Network * net = createNetwork ( 2 , 1 , hiddenSize , hiddenActivation , 2 , softmax );
// train network with cross-entropy loss using Mini-Batch SGD
ParameterSet params ;
params . network = net ;
params . data = trainingData ;
params . classes = trainingClasses ;
params . lossFunction = CROSS_ENTROPY_LOSS ;
params . batchSize = 20 ;
params . learningRate = .01 ;
params . searchTime = 5000 ;
params . regularizationStrength = .001 ;
params . momentumFactor = .9 ;
params . maxIters = 10000 ;
params . shuffle = 1 ;
params . verbose = 1 ;
optimize ( params );
// test accuracy of network after training
printf ( "Accuracy is %fn" , accuracy ( net , trainingData , trainingClasses ));
// get network's predictions on input data after training
forwardPass ( net , trainingData );
int * predictions = predict ( net );
free ( predictions );
// save network to a file
saveNetwork ( net , "network" );
// free network and data
destroyNetwork ( net );
destroyDataSet ( trainingData );
destroyDataSet ( trainingClasses );
// load previous network from file
Network * previousNet = readNetwork ( "network" );
destroyNetwork ( previousNet );
要运行测试,请查看tests
文件夹。
Makefile
包含用于运行每批单元测试或同时运行所有单元测试的命令。
如果您想添加任何功能或发现错误,请随时发送拉取请求。
检查问题选项卡以了解一些可能要做的事情。