keras入门(三) VGG 10折交叉验证实现cifar10的分类

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Sun Sep 30 17:12:12 2018


这是用keras搭建的vgg16网络
这是很经典的cnn,在图像和时间序列分析方面有很多的应用
@author: lg
"""
#################

import keras
from keras import regularizers
from keras.datasets import cifar10
from keras.models import Sequential
from keras.layers import Conv2D, MaxPooling2D, Dense, Dropout, Flatten, BatchNormalization
from keras.optimizers import SGD
import os
import argparse
import random
import numpy as np
from scipy.misc import imread, imresize, imsave
import pickle

from sklearn.model_selection import StratifiedKFold

parser = argparse.ArgumentParser()
parser.add_argument('--train_dir', default='./train/')
parser.add_argument('--test_dir', default='./test/')
parser.add_argument('--log_dir', default='./')
parser.add_argument('--batch_size', default=16)
parser.add_argument('--gpu', type=int, default=0)
args = parser.parse_args()

os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu)
type_list = ['cat', 'dog']


def vgg_16_net():
    model = Sequential()
    model.add(Conv2D(64, (3, 3), input_shape=(32, 32, 3), padding='same', activation='relu', name='conv1_block'))
    model.add(Conv2D(64, (3, 3), activation='relu', padding='same', name='conv2_block'))
    model.add(BatchNormalization())
    model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
    model.add(Conv2D(128, (3, 3), activation='relu', padding='same', name='conv3_block'))
    model.add(Conv2D(128, (3, 3), activation='relu', padding='same', name='conv4_block'))
    model.add(BatchNormalization())
    model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
    model.add(Conv2D(256, (3, 3), activation='relu', padding='same', name='conv5_block'))
    model.add(Conv2D(256, (3, 3), activation='relu', padding='same', name='conv6_block'))
    model.add(Conv2D(256, (1, 1), activation='relu', padding='same', name='conv7_block'))
    model.add(BatchNormalization())
    model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
    model.add(Conv2D(512, (3, 3), activation='relu', padding='same', name='conv8_block'))
    model.add(Conv2D(512, (3, 3), activation='relu', padding='same', name='conv9_block'))
    model.add(Conv2D(512, (1, 1), activation='relu', padding='same', name='conv10_block'))
    model.add(BatchNormalization())
    model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
    model.add(Conv2D(512, (3, 3), activation='relu', padding='same', name='conv11_block'))
    model.add(Conv2D(512, (3, 3), activation='relu', padding='same', name='conv12_block'))
    model.add(Conv2D(512, (1, 1), activation='relu', padding='same', name='conv13_block'))
    model.add(BatchNormalization())
    model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))
    model.add(Flatten())
    model.add(Dense(2048, activation='relu'))
    model.add(Dropout(0.5))
    model.add(Dense(4096, activation='relu'))
    model.add(Dropout(0.5))
    model.add(Dense(10, activation='softmax'))
    #model.add(Dense(1, activation='sigmoid'))
    return model


def prepare_data():
    file_dict1 = unpickle('F:/cifar10/cifar-10-batches-py/data_batch_1')
    label = file_dict1[b'labels']
    image = file_dict1[b'data']
    print(type(image))
    file_dict2 = unpickle('F:/cifar10/cifar-10-batches-py/data_batch_2')
    label = label + file_dict2[b'labels']
    image = np.vstack((image, file_dict2[b'data']))
    file_dict3 = unpickle('F:/cifar10/cifar-10-batches-py/data_batch_3')
    label = label + file_dict3[b'labels']
    image = np.vstack((image, file_dict3[b'data']))
    file_dict4 = unpickle('F:/cifar10/cifar-10-batches-py/data_batch_4')
    label = label + file_dict4[b'labels']
    image = np.vstack((image, file_dict4[b'data']))
    file_dict5 = unpickle('F:/cifar10/cifar-10-batches-py/data_batch_5')
    label = label + file_dict5[b'labels']
    image = np.vstack((image, file_dict5[b'data']))
    image = np.reshape(image/255, (-1, 32, 32, 3))
    label = keras.utils.to_categorical(label, 10)
    #seed = 7
    #np.random.seed(seed)
    #train_data, test_data, train_label, test_label = train_test_split(image, label, test_size=0.2, random_state=0)
    #train_num = int(len(label) * 0.8 )
    #train_data, train_label, test_data, test_label = image[0:train_num], label[0:train_num], image[train_num:], label[train_num:]
    # (X_train, y_train), (X_test, y_test) = cifar10.load_data()
    # train_data = np.reshape(X_train/255, (-1, 32, 32, 3))
    # train_label = keras.utils.to_categorical(y_train, 10)
    # test_data = np.reshape(X_test/255, (-1, 32, 32, 3))
    # test_label = keras.utils.to_categorical(y_test, 10)
    # return train_data, train_label, test_data, test_label
    return image, label

def unpickle(file):
    with open(file, 'rb') as fo:
        file_dict = pickle.load(fo, encoding='bytes')
    return file_dict


def train():
    kfold = StratifiedKFold(n_splits=10, shuffle=True, random_state=7)
    data, label = prepare_data()
    index = 1
    for train, test in kfold.split(data, label.argmax(1)):
        model = vgg_16_net()
        sgd = SGD(lr=0.001, decay=1e-8, momentum=0.9, nesterov=True)
        model.compile(optimizer=sgd, loss='categorical_crossentropy', metrics=['accuracy'])
        model.fit(data[train], label[train], validation_data=(data[test], label[test]), epochs=20, batch_size=32, shuffle=True)
        #train_data, train_label = prepare_data()
        #model.fit(train_data, train_label, batch_size=64, epochs=20, shuffle=True, validation_split=0.2)
        model.save_weights(args.log_dir + 'model_' + str(index) + '.h5')
        index = index + 1



if __name__ == '__main__':
    try:
        train()
        #model.load_weights(args.log_dir + '/model.h5')
        #predict(model)
    except Exception  as err:
        print(err)

最开始没有使用交叉验证，但是测试集与验证集的准确率一直维持在50%~60%，基本属于盲猜系列。原因大概是数据量太多，进行随机划分时，测试数据的分类不是很均匀。所以采用了交叉验证的方式，最终测试集与训练集的准确率能够达到99%，应该是有点过拟合了，结果还是非常满意的。