svhn数据集处理

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

使用上面链接里的方法，在ipython中改成三个集合，分别是trainset，validset和testset

然后再保存成mat格式，再用之前找到的mat转lmdb的代码执行，就ok了。

里面有两个需要注意的，一是数据X不需要做维度的变化，二是标签y是n*1的uint8格式，没有改成n*1的double型

没有做local contrast normalization.

根据新的trainset重新计算了均值。

把ipython中的源码贴下：extract_trainset_and_validset.ipython

from __future__ import print_function
import matplotlib.pyplot as plt
import numpy as np
import os
import sys
import tarfile
#import tensorflow as tf
from IPython.display import display, Image
from scipy import ndimage
from sklearn.linear_model import SGDClassifier
from sklearn.metrics import classification_report, confusion_matrix
from six.moves.urllib.request import urlretrieve
from six.moves import cPickle as pickle

import scipy.io
train_data = scipy.io.loadmat('train_32x32.mat', variable_names='X').get('X')
train_labels = scipy.io.loadmat('train_32x32.mat', variable_names='y').get('y')
test_data = scipy.io.loadmat('test_32x32.mat', variable_names='X').get('X')
test_labels = scipy.io.loadmat('test_32x32.mat', variable_names='y').get('y')
extra_data = scipy.io.loadmat('extra_32x32.mat', variable_names='X').get('X')
extra_labels = scipy.io.loadmat('extra_32x32.mat', variable_names='y').get('y')


print(train_data.shape, train_labels.shape)
print(test_data.shape, test_labels.shape)
print(extra_data.shape, extra_labels.shape)

train_labels[train_labels == 10] = 0
test_labels[test_labels == 10] = 0
extra_labels[extra_labels == 10] = 0

import random


random.seed()


n_labels = 10
valid_index = []
valid_index2 = []
train_index = []
train_index2 = []
for i in np.arange(n_labels):
    valid_index.extend(np.where(train_labels[:,0] == (i))[0][:400].tolist())
    train_index.extend(np.where(train_labels[:,0] == (i))[0][400:].tolist())
    valid_index2.extend(np.where(extra_labels[:,0] == (i))[0][:200].tolist())
    train_index2.extend(np.where(extra_labels[:,0] == (i))[0][200:].tolist())


random.shuffle(valid_index)
random.shuffle(train_index)
random.shuffle(valid_index2)
random.shuffle(train_index2)


#valid_data = np.concatenate((extra_data[:,:,:,valid_index2], train_data[:,:,:,valid_index]), axis=3).transpose((3,0,1,2))
valid_data = np.concatenate((extra_data[:,:,:,valid_index2], train_data[:,:,:,valid_index]), axis=3)
#valid_labels = np.concatenate((extra_labels[valid_index2,:], train_labels[valid_index,:]), axis=0)[:,0]
valid_labels = np.concatenate((extra_labels[valid_index2,:], train_labels[valid_index,:]), axis=0)
#train_data_t = np.concatenate((extra_data[:,:,:,train_index2], train_data[:,:,:,train_index]), axis=3).transpose((3,0,1,2))
train_data_t = np.concatenate((extra_data[:,:,:,train_index2], train_data[:,:,:,train_index]), axis=3)
#train_labels_t = np.concatenate((extra_labels[train_index2,:], train_labels[train_index,:]), axis=0)[:,0]
train_labels_t = np.concatenate((extra_labels[train_index2,:], train_labels[train_index,:]), axis=0)
#test_data = test_data.transpose((3,0,1,2))
test_data = test_data
test_labels = test_labels


print(train_data_t.shape, train_labels_t.shape)
print(test_data.shape, test_labels.shape)
print(valid_data.shape, valid_labels.shape)

scipy.io.savemat('train_32x32_1.mat',{'X': train_data_t,'y': train_labels_t})

scipy.io.savemat('valid_32x32.mat',{'X': valid_data,'y': valid_labels})

scipy.io.savemat('test_32x32_1.mat',{'X': test_data,'y': test_labels})