TFlearn——（2）SVHN

1，数据集简介

　　SVHN（Street View House Number）Dateset 来源于谷歌街景门牌号码，原生的数据集1也就是官网的 Format 1 是一些原始的未经处理的彩色图片，如下图所示（不含有蓝色的边框），下载的数据集含有 PNG 的图像和 digitStruct.mat  的文件，其中包含了边框的位置信息，这个数据集每张图片上有好几个数字，适用于 OCR 相关方向。

　　这里采用 Format2, Format2 将这些数字裁剪成32x32的大小，如图所示，并且数据是 .mat 文件。

    

2，数据处理

　　数据集含有两个变量 X 代表图像， 训练集 X 的 shape 是  （32，32，3，73257） 也就是(width, height, channels, samples),  tensorflow 的张量需要 （samples, width, height, channels），所以需要转换一下，由于直接调用 cifar 10 的网络模型，数据只需要先做个归一化，所有像素除于255就 OK，另外原始数据 0 的标签是 10，这里要转化成 0，并提供 one_hot 编码。

#!/usr/bin/env python2
# -*- coding: utf-8 -*-
"""
Created on Thu Jan 19 09:55:36 2017

@author: cheers
"""

import scipy.io as sio
import matplotlib.pyplot as plt
import numpy as np

image_size = 32
num_labels = 10

def display_data():
    print 'loading Matlab data...'
    train = sio.loadmat('train_32x32.mat')
    data=train['X']
    label=train['y']
    for i in range(10):
        plt.subplot(2,5,i+1)
        plt.title(label[i][0])
        plt.imshow(data[...,i])
        plt.axis('off')
    plt.show()

def load_data(one_hot = False):
    
    train = sio.loadmat('train_32x32.mat')
    test = sio.loadmat('test_32x32.mat')
    
    train_data=train['X']
    train_label=train['y']
    test_data=test['X']
    test_label=test['y']
    
    
    train_data = np.swapaxes(train_data, 0, 3)
    train_data = np.swapaxes(train_data, 2, 3)
    train_data = np.swapaxes(train_data, 1, 2)
    test_data = np.swapaxes(test_data, 0, 3)
    test_data = np.swapaxes(test_data, 2, 3)
    test_data = np.swapaxes(test_data, 1, 2)
    
    test_data = test_data / 255.
    train_data =train_data / 255.
    
    for i in range(train_label.shape[0]):
         if train_label[i][0] == 10:
             train_label[i][0] = 0
                        
    for i in range(test_label.shape[0]):
         if test_label[i][0] == 10:
             test_label[i][0] = 0

    if one_hot:
        train_label = (np.arange(num_labels) == train_label[:,]).astype(np.float32)
        test_label = (np.arange(num_labels) == test_label[:,]).astype(np.float32)

    return train_data,train_label, test_data,test_label

if __name__ == '__main__':
    load_data(one_hot = True)
    display_data()

3，TFearn 训练

注意 ImagePreprocessing 对数据做了 0 均值化。网络结构也比较简单，直接调用 TFlearn 的 cifar10 例子。

from __future__ import division, print_function, absolute_import

import tflearn
from tflearn.data_utils import shuffle, to_categorical
from tflearn.layers.core import input_data, dropout, fully_connected
from tflearn.layers.conv import conv_2d, max_pool_2d
from tflearn.layers.estimator import regression
from tflearn.data_preprocessing import ImagePreprocessing
from tflearn.data_augmentation import ImageAugmentation

# Data loading and preprocessing
import svhn_data as SVHN
X, Y, X_test, Y_test = SVHN.load_data(one_hot = True)
X, Y = shuffle(X, Y)

# Real-time data preprocessing
img_prep = ImagePreprocessing()
img_prep.add_featurewise_zero_center()
img_prep.add_featurewise_stdnorm()


# Convolutional network building
network = input_data(shape=[None, 32, 32, 3],
                     data_preprocessing=img_prep)
network = conv_2d(network, 32, 3, activation='relu')
network = max_pool_2d(network, 2)
network = conv_2d(network, 64, 3, activation='relu')
network = conv_2d(network, 64, 3, activation='relu')
network = max_pool_2d(network, 2)
network = fully_connected(network, 512, activation='relu')
network = dropout(network, 0.5)
network = fully_connected(network, 10, activation='softmax')
network = regression(network, optimizer='adam',
                     loss='categorical_crossentropy',
                     learning_rate=0.001)

# Train using classifier
model = tflearn.DNN(network, tensorboard_verbose=0)
model.fit(X, Y, n_epoch=15, shuffle=True, validation_set=(X_test, Y_test),
          show_metric=True, batch_size=96, run_id='svhn_cnn')

训练结果：

Training Step: 11452  | total loss: 0.68217 | time: 7.973s
| Adam | epoch: 015 | loss: 0.68217 - acc: 0.9329 -- iter: 72576/73257
Training Step: 11453  | total loss: 0.62980 | time: 7.983s
| Adam | epoch: 015 | loss: 0.62980 - acc: 0.9354 -- iter: 72672/73257
Training Step: 11454  | total loss: 0.58649 | time: 7.994s
| Adam | epoch: 015 | loss: 0.58649 - acc: 0.9356 -- iter: 72768/73257
Training Step: 11455  | total loss: 0.53254 | time: 8.005s
| Adam | epoch: 015 | loss: 0.53254 - acc: 0.9421 -- iter: 72864/73257
Training Step: 11456  | total loss: 0.49179 | time: 8.016s
| Adam | epoch: 015 | loss: 0.49179 - acc: 0.9416 -- iter: 72960/73257
Training Step: 11457  | total loss: 0.45679 | time: 8.027s
| Adam | epoch: 015 | loss: 0.45679 - acc: 0.9433 -- iter: 73056/73257
Training Step: 11458  | total loss: 0.42026 | time: 8.038s
| Adam | epoch: 015 | loss: 0.42026 - acc: 0.9469 -- iter: 73152/73257
Training Step: 11459  | total loss: 0.38929 | time: 8.049s
| Adam | epoch: 015 | loss: 0.38929 - acc: 0.9491 -- iter: 73248/73257
Training Step: 11460  | total loss: 0.35542 | time: 9.928s
| Adam | epoch: 015 | loss: 0.35542 - acc: 0.9542 | val_loss: 0.40315 - val_acc: 0.9085 -- iter: 73257/73257