SVHN 数据集

一直都是直接把test集直接作为测试集的，结果今天才发现错了。

是 training data without any data augmentation, and a validation set with 6,000 images is split from the training set. 

6000的来源是: select 400 samples per class from the training set and 200samples per class from the extra set. The remaining 598,388 images are used for training

SVHN consists of 73, 257 digits for training, 26, 032 digitsfor testing, and 53, 1131 extra training samples on 32 × 32 color images.

598388= 73257+531131-6000

使用了这种方法测试集的论文有：

1.Maxout Networks

2.Deep Networks with Stochastic Depth   https://github.com/yueatsprograms/Stochastic_Depth.  torch 7实现。

3.Deeply-Supervised Nets   https://github.com/s9xie/DSN   caffe实现，但没有关于SVHN的部分

4.Network in network  https://github.com/BVLC/caffe/wiki/Model-Zoo#network-in-network-model 没有SVHN部分

https://github.com/hangyao/street_view_house_numbers 这个是SVHN在tensorflow上的实现，是ipython版本的。