keras实现多种分类网络的方式

Keras应该是最简单的一种深度学习框架了，入门非常的简单.

简单记录一下keras实现多种分类网络：如AlexNet、Vgg、ResNet

采用kaggle猫狗大战的数据作为数据集.

由于AlexNet采用的是LRN标准化,Keras没有内置函数实现,这里用batchNormalization代替

收件建立一个model.py的文件,里面存放着alexnet,vgg两种模型,直接导入就可以了

#coding=utf-8
from keras.models import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten
from keras.layers import Conv2D, MaxPooling2D, ZeroPadding2D, BatchNormalization
from keras.layers import *
from keras.layers.advanced_activations import LeakyReLU,PReLU
from keras.models import Model

def keras_batchnormalization_relu(layer):
BN = BatchNormalization()(layer)
ac = PReLU()(BN)
return ac

def AlexNet(resize=227, classes=2):
model = Sequential()
# 第一段
model.add(Conv2D(filters=96, kernel_size=(11, 11),
     strides=(4, 4), padding='valid',
     input_shape=(resize, resize, 3),
     activation='relu'))
model.add(BatchNormalization())
model.add(MaxPooling2D(pool_size=(3, 3),
      strides=(2, 2),
      padding='valid'))
# 第二段
model.add(Conv2D(filters=256, kernel_size=(5, 5),
     strides=(1, 1), padding='same',
     activation='relu'))
model.add(BatchNormalization())
model.add(MaxPooling2D(pool_size=(3, 3),
      strides=(2, 2),
      padding='valid'))
# 第三段
model.add(Conv2D(filters=384, kernel_size=(3, 3),
     strides=(1, 1), padding='same',
     activation='relu'))
model.add(Conv2D(filters=384, kernel_size=(3, 3),
     strides=(1, 1), padding='same',
     activation='relu'))
model.add(Conv2D(filters=256, kernel_size=(3, 3),
     strides=(1, 1), padding='same',
     activation='relu'))
model.add(MaxPooling2D(pool_size=(3, 3),
      strides=(2, 2), padding='valid'))
# 第四段
model.add(Flatten())
model.add(Dense(4096, activation='relu'))
model.add(Dropout(0.5))

model.add(Dense(4096, activation='relu'))
model.add(Dropout(0.5))

model.add(Dense(1000, activation='relu'))
model.add(Dropout(0.5))

# Output Layer
model.add(Dense(classes,activation='softmax'))
# model.add(Activation('softmax'))

return model

def AlexNet2(inputs, classes=2, prob=0.5):
'''
自己写的函数,尝试keras另外一种写法
:param inputs: 输入
:param classes: 类别的个数
:param prob: dropout的概率
:return: 模型
'''
# Conv2D(32, (3, 3), dilation_rate=(2, 2), padding='same')(inputs)
print "input shape:", inputs.shape

conv1 = Conv2D(filters=96, kernel_size=(11, 11), strides=(4, 4), padding='valid')(inputs)
conv1 = keras_batchnormalization_relu(conv1)
print "conv1 shape:", conv1.shape
pool1 = MaxPool2D(pool_size=(3, 3), strides=(2, 2))(conv1)
print "pool1 shape:", pool1.shape

conv2 = Conv2D(filters=256, kernel_size=(5, 5), padding='same')(pool1)
conv2 = keras_batchnormalization_relu(conv2)
print "conv2 shape:", conv2.shape
pool2 = MaxPool2D(pool_size=(3, 3), strides=(2, 2))(conv2)
print "pool2 shape:", pool2.shape

conv3 = Conv2D(filters=384, kernel_size=(3, 3), padding='same')(pool2)
conv3 = PReLU()(conv3)
print "conv3 shape:", conv3.shape

conv4 = Conv2D(filters=384, kernel_size=(3, 3), padding='same')(conv3)
conv4 = PReLU()(conv4)
print "conv4 shape:", conv4

conv5 = Conv2D(filters=256, kernel_size=(3, 3), padding='same')(conv4)
conv5 = PReLU()(conv5)
print "conv5 shape:", conv5

pool3 = MaxPool2D(pool_size=(3, 3), strides=(2, 2))(conv5)
print "pool3 shape:", pool3.shape

dense1 = Flatten()(pool3)
dense1 = Dense(4096, activation='relu')(dense1)
print "dense2 shape:", dense1
dense1 = Dropout(prob)(dense1)
# print "dense1 shape:", dense1

dense2 = Dense(4096, activation='relu')(dense1)
print "dense2 shape:", dense2
dense2 = Dropout(prob)(dense2)
# print "dense2 shape:", dense2

predict= Dense(classes, activation='softmax')(dense2)

model = Model(inputs=inputs, outputs=predict)
return model

def vgg13(resize=224, classes=2, prob=0.5):
model = Sequential()
model.add(Conv2D(64, (3, 3), strides=(1, 1), input_shape=(resize, resize, 3), padding='same', activation='relu',
     kernel_initializer='uniform'))
model.add(Conv2D(64, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(128, (3, 2), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(Conv2D(128, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(256, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(Conv2D(256, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(512, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(Conv2D(512, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(512, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(Conv2D(512, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Flatten())
model.add(Dense(4096, activation='relu'))
model.add(Dropout(prob))
model.add(Dense(4096, activation='relu'))
model.add(Dropout(prob))
model.add(Dense(classes, activation='softmax'))
return model

def vgg16(resize=224, classes=2, prob=0.5):
model = Sequential()
model.add(Conv2D(64, (3, 3), strides=(1, 1), input_shape=(resize, resize, 3), padding='same', activation='relu',
     kernel_initializer='uniform'))
model.add(Conv2D(64, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(128, (3, 2), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(Conv2D(128, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(256, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(Conv2D(256, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(Conv2D(256, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(512, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(Conv2D(512, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(Conv2D(512, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(512, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(Conv2D(512, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(Conv2D(512, (3, 3), strides=(1, 1), padding='same', activation='relu', kernel_initializer='uniform'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Flatten())
model.add(Dense(4096, activation='relu'))
model.add(Dropout(prob))
model.add(Dense(4096, activation='relu'))
model.add(Dropout(prob))
model.add(Dense(classes, activation='softmax'))
return model

然后建立一个train.py文件，用于读取数据和训练数据的.

#coding=utf-8
import keras
import cv2
import os
import numpy as np
import model
import modelResNet
import tensorflow as tf
from keras.layers import Input, Dense
from keras.preprocessing.image import ImageDataGenerator

resize = 224
batch_size = 128
path = "/home/hjxu/PycharmProjects/01_cats_vs_dogs/data"

trainDirectory = '/home/hjxu/PycharmProjects/01_cats_vs_dogs/data/train/'
def load_data():
imgs = os.listdir(path + "/train/")
num = len(imgs)
train_data = np.empty((5000, resize, resize, 3), dtype="int32")
train_label = np.empty((5000, ), dtype="int32")
test_data = np.empty((5000, resize, resize, 3), dtype="int32")
test_label = np.empty((5000, ), dtype="int32")
for i in range(5000):
 if i ％ 2:
  train_data[i] = cv2.resize(cv2.imread(path + '/train/' + 'dog.' + str(i) + '.jpg'), (resize, resize))
  train_label[i] = 1
 else:
  train_data[i] = cv2.resize(cv2.imread(path + '/train/' + 'cat.' + str(i) + '.jpg'), (resize, resize))
  train_label[i] = 0
for i in range(5000, 10000):
 if i ％ 2:
  test_data[i-5000] = cv2.resize(cv2.imread(path + '/train/' + 'dog.' + str(i) + '.jpg'), (resize, resize))
  test_label[i-5000] = 1
 else:
  test_data[i-5000] = cv2.resize(cv2.imread(path + '/train/' + 'cat.' + str(i) + '.jpg'), (resize, resize))
  test_label[i-5000] = 0
return train_data, train_label, test_data, test_label

def main():

train_data, train_label, test_data, test_label = load_data()
train_data, test_data = train_data.astype('float32'), test_data.astype('float32')
train_data, test_data = train_data/255, test_data/255

train_label = keras.utils.to_categorical(train_label, 2)
'''
 #one_hot转码,如果使用 categorical_crossentropy,就需要用到to_categorical函数完成转码
'''
test_label = keras.utils.to_categorical(test_label, 2)

inputs = Input(shape=(224, 224, 3))

modelAlex = model.AlexNet2(inputs, classes=2)
'''
导入模型
'''
modelAlex.compile(loss='categorical_crossentropy',
    optimizer='sgd',
    metrics=['accuracy'])
'''
def compile(self, optimizer, loss, metrics=None, loss_weights=None,
    sample_weight_mode=None, **kwargs):
 optimizer:优化器,为预定义优化器名或优化器对象,参考优化器
 loss: 损失函数,为预定义损失函数名或者一个目标函数
 metrics:列表,包含评估模型在训练和测试时的性能指标,典型用法是 metrics=['accuracy']
 sample_weight_mode:如果需要按时间步为样本赋值,需要将改制设置为"temoral"
 如果想用自定义的性能评估函数:如下
  def mean_pred(y_true, y_pred):
  return k.mean(y_pred)
 model.compile(loss = 'binary_crossentropy', metrics=['accuracy', mean_pred],...)
 损失函数同理,再看 keras内置支持的损失函数有
  mean_squared_error
 mean_absolute_error
 mean_absolute_percentage_error
 mean_squared_logarithmic_error
 squared_hinge
 hinge
 categorical_hinge
 logcosh
 categorical_crossentropy
 sparse_categorical_crossentropy
 binary_crossentropy
 kullback_leibler_divergence
 poisson
 cosine_proximity
'''
modelAlex.summary()
'''
# 打印模型信息
'''
modelAlex.fit(train_data, train_label,
   batch_size=batch_size,
   epochs=50,
   validation_split=0.2,
   shuffle=True)
'''
def fit(self, x=None,   # x:输入数据
  y=None,     # y:标签 Numpy array
  batch_size=32,   # batch_size:训练时,一个batch的样本会被计算一次梯度下降
  epochs=1,    # epochs: 训练的轮数,每个epoch会把训练集循环一遍
  verbose=1,    # 日志显示:0表示不在标准输入输出流输出,1表示输出进度条,2表示每个epoch输出
  callbacks=None,   # 回调函数
  validation_split=0.,  # 0-1的浮点数,用来指定训练集一定比例作为验证集,验证集不参与训练
  validation_data=None, # (x,y)的tuple,是指定的验证集
  shuffle=True,   # 如果是"batch",则是用来处理HDF5数据的特殊情况,将在batch内部将数据打乱
  class_weight=None,  # 字典,将不同的类别映射为不同的权值,用来在训练过程中调整损失函数的
  sample_weight=None,  # 权值的numpy array,用于训练的时候调整损失函数
  initial_epoch=0,   # 该参数用于从指定的epoch开始训练,继续之前的训练
  **kwargs):
返回:返回一个History的对象,其中History.history损失函数和其他指标的数值随epoch变化的情况
'''
scores = modelAlex.evaluate(train_data, train_label, verbose=1)
print(scores)

scores = modelAlex.evaluate(test_data, test_label, verbose=1)
print(scores)
modelAlex.save('my_model_weights2.h5')

def main2():
train_datagen = ImageDataGenerator(rescale=1. / 255,
         shear_range=0.2,
         zoom_range=0.2,
         horizontal_flip=True)
test_datagen = ImageDataGenerator(rescale=1. / 255)
train_generator = train_datagen.flow_from_directory(trainDirectory,
             target_size=(224, 224),
             batch_size=32,
             class_mode='binary')

validation_generator = test_datagen.flow_from_directory(trainDirectory,
              target_size=(224, 224),
              batch_size=32,
              class_mode='binary')

inputs = Input(shape=(224, 224, 3))
# modelAlex = model.AlexNet2(inputs, classes=2)
modelAlex = model.vgg13(resize=224, classes=2, prob=0.5)
# modelAlex = modelResNet.ResNet50(shape=224, classes=2)
modelAlex.compile(loss='sparse_categorical_crossentropy',
     optimizer='sgd',
     metrics=['accuracy'])
modelAlex.summary()

modelAlex.fit_generator(train_generator,
     steps_per_epoch=1000,
     epochs=60,
     validation_data=validation_generator,
     validation_steps=200)

modelAlex.save('model32.hdf5')
#
if __name__ == "__main__":
'''
如果数据是按照猫狗大战的数据,都在同一个文件夹下,使用main()函数
如果数据按照猫和狗分成两类,则使用main2()函数
'''
main2()

得到模型后该怎么测试一张图像呢？

建立一个testOneImg.py脚本，代码如下

#coding=utf-8
from keras.preprocessing.image import load_img#load_image作用是载入图片
from keras.preprocessing.image import img_to_array
from keras.applications.vgg16 import preprocess_input
from keras.applications.vgg16 import decode_predictions
import numpy as np
import cv2
import model
from keras.models import Sequential

pats = '/home/hjxu/tf_study/catVsDogsWithKeras/my_model_weights.h5'
modelAlex = model.AlexNet(resize=224, classes=2)
# AlexModel = model.AlexNet(weightPath='/home/hjxu/tf_study/catVsDogsWithKeras/my_model_weights.h5')

modelAlex.load_weights(pats)
#
img = cv2.imread('/home/hjxu/tf_study/catVsDogsWithKeras/111.jpg')
img = cv2.resize(img, (224, 224))
x = img_to_array(img/255) # 三维（224，224，3）

x = np.expand_dims(x, axis=0) # 四维（1，224，224，3）#因为keras要求的维度是这样的，所以要增加一个维度
# x = preprocess_input(x) # 预处理
print(x.shape)
y_pred = modelAlex.predict(x) # 预测概率 t1 = time.time() print("测试图：", decode_predictions(y_pred)) # 输出五个最高概率(类名, 语义概念, 预测概率)
print y_pred

不得不说，Keras真心简单方便。

补充知识：keras中的函数式API——残差连接+权重共享的理解

1、残差连接

# coding: utf-8
"""残差连接 residual connection：
 是一种常见的类图网络结构，解决了所有大规模深度学习的两个共性问题：
  1、梯度消失
  2、表示瓶颈
 （甚至，向任何>10层的神经网络添加残差连接，都可能会有帮助）

残差连接：让前面某层的输出作为后面某层的输入，从而在序列网络中有效地创造一条捷径。
      """
from keras import layers

x = ...
y = layers.Conv2D(128, 3, activation='relu', padding='same')(x)
y = layers.Conv2D(128, 3, activation='relu', padding='same')(y)
y = layers.Conv2D(128, 3, activation='relu', padding='same')(y)

y = layers.add([y, x]) # 将原始x与输出特征相加

# ---------------------如果特征图尺寸不同，采用线性残差连接-------------------
x = ...
y = layers.Conv2D(128, 3, activation='relu', padding='same')(x)
y = layers.Conv2D(128, 3, activation='relu', padding='same')(y)
y = layers.MaxPooling2D(2, strides=2)(y)

residual = layers.Conv2D(128, 1, strides=2, padding='same')(x) # 使用1*1的卷积，将原始张量线性下采样为y具有相同的形状

y = layers.add([y, residual]) # 将原始x与输出特征相加

2、权重共享

即多次调用同一个实例

# coding: utf-8
"""函数式子API：权重共享
 能够重复的使用同一个实例，这样相当于重复使用一个层的权重，不需要重新编写"""
from keras import layers
from keras import Input
from keras.models import Model

lstm = layers.LSTM(32) # 实例化一个LSTM层,后面被调用很多次

# ------------------------左边分支--------------------------------
left_input = Input(shape=(None, 128))
left_output = lstm(left_input) # 调用lstm实例

# ------------------------右分支---------------------------------
right_input = Input(shape=(None, 128))
right_output = lstm(right_input) # 调用lstm实例

# ------------------------将层进行连接合并------------------------
merged = layers.concatenate([left_output, right_output], axis=-1)

# -----------------------在上面构建一个分类器---------------------
predictions = layers.Dense(1, activation='sigmoid')(merged)

# -------------------------构建模型，并拟合训练-----------------------------------
model = Model([left_input, right_input], predictions)
model.fit([left_data, right_data], targets)

来源：https://blog.csdn.net/hjxu2016/article/details/83504756