(1) 运行create data,其中包括下载cifar10,并转换为hdf5格式(详见百度百科:http://baike.baidu.com/view/771949.htm#4_2):
cifar10的数据简介见:http://www.cs.toronto.edu/~kriz/cifar.html
cd data
python create_cifar10.py
(2) 训练模型:
cd examples
python cifar10_cnn.py
关键在于网络结构的设置:convolution layer >> pooling >> convolution layer >> pooling >> convolution layer >> pooling >> fully connected
训练算法:SGD with momentum
(3)训练结果:只使用CPU,共耗时约179min,一个epoch约耗时为8~9min,max epoch=20