Skip to content

Latest commit

 

History

History
25 lines (20 loc) · 1.09 KB

readme-cn.md

File metadata and controls

25 lines (20 loc) · 1.09 KB
Raw

JavaCNN

一个卷积神经网络的java实现. 仿Matlab toolbox(https://github.com/rasmusbergpalm/DeepLearnToolbox )实现的，同时进行了部分改进，使得卷积核和采样块可以为矩形而不仅仅是正方形。更多细节，请查看http://www.cnblogs.com/fengfenggirl/p/cnn_implement.html

创建一个卷积神经网络

LayerBuilder builder = new LayerBuilder();
builder.addLayer(Layer.buildInputLayer(new Size(28, 28)));
builder.addLayer(Layer.buildConvLayer(6, new Size(5, 5)));
builder.addLayer(Layer.buildSampLayer(new Size(2, 2)));
builder.addLayer(Layer.buildConvLayer(12, new Size(5, 5)));
builder.addLayer(Layer.buildSampLayer(new Size(2, 2)));
builder.addLayer(Layer.buildOutputLayer(10));
CNN cnn = new CNN(builder, 50);

在 MNIST 数据集上测试

String fileName = "data/train.format";
Dataset dataset = Dataset.load(fileName, ",", 784);
cnn.train(dataset, 100);
Dataset testset = Dataset.load("data/test.format", ",", -1);
cnn.predict(testset, "data/test.predict");

迭代100次，四核CPU大约需要运行一个小时后，正确率97.8%

##Lisence MIT