ImageDataGenerator(featurewise_center=False, #布爾值。將輸入數(shù)據(jù)的均值設(shè)置為 0，逐特征進(jìn)行
 samplewise_center=False, #布爾值。將每個(gè)樣本的均值設(shè)置為 0
 featurewise_std_normalization=False, #布爾值。將輸入除以數(shù)據(jù)標(biāo)準(zhǔn)差，逐特征進(jìn)行
 samplewise_std_normalization=False, #布爾值。將每個(gè)輸入除以其標(biāo)準(zhǔn)差
 zca_whitening=False, #是否進(jìn)行ZAC白化
 zca_epsilon=1e-06, #ZCA 白化的 epsilon 值
 rotation_range=0, #整數(shù)。隨機(jī)旋轉(zhuǎn)的度數(shù)范圍
 width_shift_range=0.0, 
 height_shift_range=0.0, 
 brightness_range=None, 
 shear_range=0.0, #浮點(diǎn)數(shù)。剪切強(qiáng)度（以弧度逆時(shí)針方向剪切角度）
 zoom_range=0.0, #浮點(diǎn)數(shù) 或 [lower, upper]。隨機(jī)縮放范圍。如果是浮點(diǎn)數(shù)，[lower, upper] = [1-zoom_range, 1+zoom_range]。
 channel_shift_range=0.0, #浮點(diǎn)數(shù)。隨機(jī)通道轉(zhuǎn)換的范圍
 fill_mode='nearest', #輸入邊界以外的點(diǎn)的模式填充
 cval=0.0, #當(dāng) fill_mode = "constant",邊界點(diǎn)的填充值
 horizontal_flip=False, #隨機(jī)水平翻轉(zhuǎn)
 vertical_flip=False, #隨機(jī)垂直翻
 rescale=None, #默認(rèn)為 None。如果是 None 或 0，不進(jìn)行縮放，否則將數(shù)據(jù)乘以所提供的值（在應(yīng)用任何其他轉(zhuǎn)換之前）
 preprocessing_function=None, #應(yīng)用于每個(gè)輸入的函數(shù)。這個(gè)函數(shù)會(huì)在任何其他改變之前運(yùn)行。這個(gè)函數(shù)需要一個(gè)參數(shù)：一張圖像（秩為 3 的 Numpy 張量），并且應(yīng)該輸出一個(gè)同尺寸的 Numpy 張量。
 data_format=None, #圖像數(shù)據(jù)格式，{"channels_first", "channels_last"} 之一
 validation_split=0.0, 
 dtype=None) #生成數(shù)組使用的數(shù)據(jù)類型

def flow_from_directory(self,
 directory, #子文件夾所在的目錄
 target_size=(256, 256), #輸出的圖片的尺寸
 color_mode='rgb', #單通道還是三通道
 classes=None, #類別,有多少個(gè)子文件夾就有多少個(gè)類別,填寫的是子文件夾的名稱
 class_mode='categorical', #通常默認(rèn),表示標(biāo)簽采用one-hot形式,
 batch_size=32, 
 shuffle=True, #是否隨機(jī)打亂順序
 seed=None,
 save_to_dir=None, #把圖片保存,輸入的是路徑
 save_prefix='', #圖像前綴名,
 save_format='png', #圖像后綴名
 follow_links=False,
 subset=None,
 interpolation='nearest')

from tensorflow.keras.preprocessing.image import ImageDataGenerator #我是直接裝tensorflow,然后使用里面的keras的,
 
#實(shí)例化對(duì)象datagen
datagen=ImageDataGenerator() 
 
#讀訓(xùn)練集圖片
train_generator = datagen.flow_from_directory(
 '/home/hky/folder/kaggle/DataGenerator/train',
 classes=['cat','dog'],
 target_size=(227, 227),
 class_mode='categorical',
 batch_size=batch_size)
 
#讀驗(yàn)證集圖片
validation_generator = datagen.flow_from_directory(
 '/home/hky/folder/kaggle/DataGenerator/validation',
 classes=['cat','dog'],
 target_size=(227, 227),
 class_mode='categorical',
 batch_size=batch_size)
 
'''開始訓(xùn)練'''
#steps_per_epoch是為了判斷是否完成了一個(gè)epoch,這里我訓(xùn)練集有20000張圖片,然后batch_size=16,所以是10000/16
#同樣,validation_steps=2496/16是因?yàn)槲业尿?yàn)證集有2496張圖片
model.fit_generator(generator=train_generator,steps_per_epoch=20000/16,epochs=10,validation_data=validation_generator,validation_steps=2496/16)

import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.preprocessing import image
from tensorflow.keras.callbacks import EarlyStopping
from tensorflow.keras import optimizers
from tensorflow.keras.preprocessing.image import ImageDataGenerator
import numpy as np
import cv2
import yaml
from tensorflow.keras.models import model_from_yaml
 
batch_size = 16
 
''' 搭建模型'''
l=tf.keras.layers
model=Sequential()
 
#第一層卷積和池化
model.add(l.Conv2D(filters=96,kernel_size=(11,11),strides=(4,4),padding='valid',input_shape=(227,227,3),activation='relu'))
model.add(l.BatchNormalization())
model.add(l.MaxPooling2D(pool_size=(3,3),strides=(2,2),padding='valid'))
 
#第二層卷積和池化
model.add(l.Conv2D(256,(5,5),(1,1),padding='same',activation='relu'))
model.add(l.BatchNormalization())
model.add(l.MaxPooling2D((3,3),(2,2),padding='valid'))
 
#第三層卷積
model.add(l.Conv2D(384,(3,3),(1,1),'same',activation='relu'))
 
#第四層卷積
model.add(l.Conv2D(384,(3,3),(1,1),'same',activation='relu'))
 
#第五層卷積和池化
model.add(l.Conv2D(256,(3,3),(1,1),'same',activation='relu'))
model.add(l.MaxPooling2D((3,3),(2,2),'valid'))
 
#全連接層
model.add(l.Flatten())
model.add(l.Dense(4096,activation='relu'))
model.add(l.Dropout(0.5))
 
model.add(l.Dense(4096,activation='relu'))
model.add(l.Dropout(0.5))
 
model.add(l.Dense(1000,activation='relu'))
model.add(l.Dropout(0.5))
 
#輸出層
model.add(l.Dense(2,activation='softmax'))
model.compile(optimizer='sgd',loss='categorical_crossentropy',metrics=['accuracy'])
 
'''導(dǎo)入圖片數(shù)據(jù)'''
#利用ImageDataGenerator生成一個(gè)batch一個(gè)batch的數(shù)據(jù)
 
datagen=ImageDataGenerator(samplewise_center=True,rescale=1.0/255) #samplewise_center:使輸入數(shù)據(jù)的每個(gè)樣本均值為0,rescale:歸一化
train_generator = datagen.flow_from_directory(
 '/home/hky/folder/kaggle/DataGenerator/train',
 classes=['cat','dog'],
 target_size=(227, 227),
 class_mode='categorical',
 batch_size=batch_size)
 
validation_generator = datagen.flow_from_directory(
 '/home/hky/folder/kaggle/DataGenerator/validation',
 classes=['cat','dog'],
 target_size=(227, 227),
 class_mode='categorical',
 batch_size=batch_size)
 
'''開始訓(xùn)練'''
model.fit_generator(generator=train_generator,steps_per_epoch=20000/16,epochs=10,validation_data=validation_generator,validation_steps=2496/16)
 
yaml_string = model.to_yaml() # 保存模型結(jié)構(gòu)到y(tǒng)aml文件
open('./model_architecture.yaml', 'w').write(yaml_string)
model.save_weights('./AlexNet_model.h5') #保存模型參數(shù)
 
'''導(dǎo)入模型'''
#model = model_from_yaml(open('./model_architecture.yaml').read())
#model.load_weights('./AlexNet_model.h5')
 
'''隨便輸入一張圖片測(cè)試一下'''
imgs=[]
img=cv2.imread('/home/hky/folder/kaggle/test/120.jpg')
img=cv2.resize(img,(227,227))
imgs.append(img)
a=np.array(imgs)
 
result=model.predict(a)
idx=np.argmax(result)
 
if idx==0:
 print('the image is cat\n')
else:
 print('the image is dog\n')
 
cv2.imshow("image",img)
cv2.waitKey(0)