keras数据增强用于mnist

 keras的数据增强，有篇博客写的不错博客地址

 keras关于数据增强的介绍keras数据增强

from keras.preprocessing.image import ImageDataGenerator, array_to_img, img_to_array, load_img

datagen = ImageDataGenerator(featurewise_center=True,featurewise_std_normalization=True,rotation_range=20,width_shift_range=0.2,height_shift_range=0.2,horizontal_flip=True)

img = load_img('cat.jpeg')  # 这是一个PIL图像
x = img_to_array(img)  # 把PIL图像转换成一个numpy数组，形状为(3, 150, 150)
x = x.reshape((1,) + x.shape)  # 这是一个numpy数组，形状为 (1, 3, 150, 150)
#计算依赖于数据的变换所需要的统计信息(均值方差等),只有使用featurewise_center，featurewise_std_normalization或zca_whitening时需要此函数。
#datagen.fit(x)
# 下面是生产图片的代码
# 生产的所有图片保存在 `data1/` 目录下
i = 0
for batch in datagen.flow(x, batch_size=1,save_to_dir='data1', save_prefix='cat', save_format='jpeg'):
    i += 1
    if i > 50:
        break  # 否则生成器会退出循环

当没有加datagen.fit()这个函数时，会报这个警告：

当加上datagen.fit()这句话时（即把上面代码中#datagen.fit()的#符号去掉），就不会报上面的警告

结果：

keras实现mnist：

import os
import numpy as np
import tensorflow

from keras.models import *
from keras.layers import Input, merge, Conv2D, MaxPooling2D, UpSampling2D, Dropout, Cropping2D,Dense,Reshape
from keras.optimizers import *
from keras.callbacks import ModelCheckpoint, LearningRateScheduler
from keras import backend as keras
import keras
from keras.preprocessing.image import ImageDataGenerator
from keras.datasets import mnist



import scipy.io as sio
class myUnet(object):

    def get_unet(self):
        inputs = Input((self.img_rows, self.img_cols,1))
        conv1 =Conv2D(32, 3, activation='relu', padding='same', kernel_initializer='he_normal')(inputs)
        conv2 = Conv2D(64, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv1)
        conv3=MaxPooling2D(pool_size=(2, 2))(conv2)
        conv4=keras.layers.core.Dropout(0.25, noise_shape=None, seed=None)(conv3)
        conv5=keras.layers.core.Flatten()(conv4)
        conv6 = Dense(128)(conv5)
        conv7 = keras.layers.core.Dropout(0.5, noise_shape=None, seed=None)(conv6)
        conv8=Dense(10, activation='softmax')(conv7)

        model = Model(input = inputs, output = conv8)
        model.compile(optimizer=Adam(lr=1e-4), loss='categorical_crossentropy', metrics=['accuracy'])

        return model


    def train(self):

        print("loading data")
        (x_train, y_train), (x_test, y_test) = mnist.load_data()
        img_rows=28;
        img_cols=28
        num_classes=10

        if K.image_data_format() == 'channels_first':
            x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
            x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
            # 单通道灰度图像,channel=1
            input_shape = (1, img_rows, img_cols)
        else:
            x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
            x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
            input_shape = (img_rows, img_cols, 1)
        # 数据转为float32型
        x_train = x_train.astype('float32')
        x_test = x_test.astype('float32')
        print('y_test:', x_train.shape)
        print('y_test:', y_test.shape)
        # 先通过np生成一个1000*1维的其值为0-9的矩阵，然后再通过```keras.utils.to_categorical```方法获取成一个1000*10维的二元矩阵。比如当那一行的数为5时生成的新的一行为[0,0,0,0,0,1,0,0,0,0]即第6个数为1，其它为0
        y_train = keras.utils.to_categorical(y_train, num_classes)
        y_test = keras.utils.to_categorical(y_test, num_classes)
        print('y_test:',y_test.shape)
        print("loading data done")


        #datagen = ImageDataGenerator(featurewise_center=True, featurewise_std_normalization=True, rotation_range=20,width_shift_range=0.2, height_shift_range=0.2, horizontal_flip=True)

        model = self.get_unet()
        model_checkpoint = ModelCheckpoint('unet.hdf5', monitor='loss',verbose=1, save_best_only=True)
        print('Fitting model...')
        #model.fit_generator(datagen.flow(x_train, y_train, batch_size=12),steps_per_epoch=x_train.shape[0], epochs=10, verbose=1)
        model.fit(x_train, y_train, batch_size=100, nb_epoch=10, verbose=1,validation_split=0.1, shuffle=True, callbacks=[model_checkpoint])
        score = model.evaluate(x_test, y_test, verbose=0)
        print('Test loss:', score[0])
        print('Test accuracy:', score[1])

        model.save('a_rcnn1.h5')

if __name__ == '__main__':
    myunet = myUnet()
    myunet.train()

其中(x_train, y_train), (x_test, y_test) = mnist.load_data()的数据会下载到下面路径

• path：如果你在本机上已有此数据集（位于'~/.keras/datasets/'+path），则载入。否则数据将下载到该目录下
• 关于keras的mnist数据集合的介绍

在我的电脑上的路径为：D:\Users\kmin\.keras

其中.keras下还有个文件：keras.json 内容：

{
    "floatx": "float32",
    "epsilon": 1e-07,
    "image_data_format": "channels_last",
    "backend": "tensorflow"
}

如果将tensorflow改为theano那么backend就会变为theano

我的训练结果为：

keras采用数据增强：

import os
import numpy as np
import tensorflow

from keras.models import *
from keras.layers import Input, merge, Conv2D, MaxPooling2D, UpSampling2D, Dropout, Cropping2D,Dense,Reshape
from keras.optimizers import *
from keras.callbacks import ModelCheckpoint, LearningRateScheduler
from keras import backend as keras
import keras
from keras.preprocessing.image import ImageDataGenerator
from keras.datasets import mnist



import scipy.io as sio
class myUnet(object):

    def get_unet(self):
        inputs = Input((self.img_rows, self.img_cols,1))
        conv1 =Conv2D(32, 3, activation='relu', padding='same', kernel_initializer='he_normal')(inputs)
        conv2 = Conv2D(64, 3, activation='relu', padding='same', kernel_initializer='he_normal')(conv1)
        conv3=MaxPooling2D(pool_size=(2, 2))(conv2)
        conv4=keras.layers.core.Dropout(0.25, noise_shape=None, seed=None)(conv3)
        conv5=keras.layers.core.Flatten()(conv4)
        conv6 = Dense(128)(conv5)
        conv7 = keras.layers.core.Dropout(0.5, noise_shape=None, seed=None)(conv6)
        conv8=Dense(10, activation='softmax')(conv7)

        model = Model(input = inputs, output = conv8)
        model.compile(optimizer=Adam(lr=1e-4), loss='categorical_crossentropy', metrics=['accuracy'])

        return model


    def train(self):

        print("loading data")
        (x_train, y_train), (x_test, y_test) = mnist.load_data()
        img_rows=28;
        img_cols=28
        num_classes=10

        if K.image_data_format() == 'channels_first':
            x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
            x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
            # 单通道灰度图像,channel=1
            input_shape = (1, img_rows, img_cols)
        else:
            x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
            x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
            input_shape = (img_rows, img_cols, 1)
        # 数据转为float32型
        x_train = x_train.astype('float32')
        x_test = x_test.astype('float32')
        print('y_test:', x_train.shape)
        print('y_test:', y_test.shape)
        # 先通过np生成一个1000*1维的其值为0-9的矩阵，然后再通过```keras.utils.to_categorical```方法获取成一个1000*10维的二元矩阵。比如当那一行的数为5时生成的新的一行为[0,0,0,0,0,1,0,0,0,0]即第5个数为1，其它为0
        y_train = keras.utils.to_categorical(y_train, num_classes)
        y_test = keras.utils.to_categorical(y_test, num_classes)
        print('y_test:',y_test.shape)
        print("loading data done")

        #具体说明看官方文档
        datagen = ImageDataGenerator(featurewise_center=True, featurewise_std_normalization=True, rotation_range=20,width_shift_range=0.2, height_shift_range=0.2, horizontal_flip=True)

        model = self.get_unet()
        model_checkpoint = ModelCheckpoint('unet.hdf5', monitor='loss',verbose=1, save_best_only=True)
        print('Fitting model...')
        #batch_size就是一张图片要生成多少张图片，这里为12，steps_per_epoch就是一次epochs要取多少张图片，这里是训练图片的个数
        model.fit_generator(datagen.flow(x_train, y_train, batch_size=12),steps_per_epoch=x_train.shape[0], epochs=10, verbose=1)
        #model.fit(x_train, y_train, batch_size=100, nb_epoch=10, verbose=1,validation_split=0.1, shuffle=True, callbacks=[model_checkpoint])
        score = model.evaluate(x_test, y_test, verbose=0)
        print('Test loss:', score[0])
        print('Test accuracy:', score[1])

        model.save('a_rcnn1.h5')

if __name__ == '__main__':
    myunet = myUnet()
    myunet.train()

结果好像很差，这里就不贴出来了，可能是因为数据增强 用了 featurewise_center ， featurewise_std_normalization 或 zca_whitening

操作导致数据不好，具体原因，有待参考。