model = keras.models.Sequential([
 #卷積層1
 keras.layers.Conv2D(32,kernel_size=5,strides=1,padding="same",data_format="channels_last",activation=tf.nn.relu,kernel_regularizer=keras.regularizers.l2(0.01)),
 #池化層1
 keras.layers.MaxPool2D(pool_size=2,strides=2,padding="same"),
 #卷積層2
 keras.layers.Conv2D(64,kernel_size=5,strides=1,padding="same",data_format="channels_last",activation=tf.nn.relu),
 #池化層2
 keras.layers.MaxPool2D(pool_size=2,strides=2,padding="same"),
 #數(shù)據(jù)整理
 keras.layers.Flatten(),
 #1024個(gè)，全連接層
 keras.layers.Dense(1024,activation=tf.nn.relu),
 #100個(gè)，全連接層
 keras.layers.Dense(100,activation=tf.nn.softmax)
 ])

import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
 
from tensorflow.python.keras.datasets import cifar100
from tensorflow.python import keras
import tensorflow as tf
 
class CNNMnist(object):
 
 model = keras.models.Sequential([
 #卷積層1
 keras.layers.Conv2D(32,kernel_size=5,strides=1,padding="same",data_format="channels_last",activation=tf.nn.relu,kernel_regularizer=keras.regularizers.l2(0.01)),
 #池化層1
 keras.layers.MaxPool2D(pool_size=2,strides=2,padding="same"),
 #卷積層2
 keras.layers.Conv2D(64,kernel_size=5,strides=1,padding="same",data_format="channels_last",activation=tf.nn.relu),
 #池化層2
 keras.layers.MaxPool2D(pool_size=2,strides=2,padding="same"),
 #數(shù)據(jù)整理
 keras.layers.Flatten(),
 #1024個(gè)，全連接層
 keras.layers.Dense(1024,activation=tf.nn.relu),
 #100個(gè)，全連接層
 keras.layers.Dense(100,activation=tf.nn.softmax)
 ])
 
 def __init__(self):
 (self.x_train,self.y_train),(self.x_test,self.y_test) = cifar100.load_data()
 
 self.x_train = self.x_train/255.0
 self.x_test = self.x_test/255.0
 
 
 def compile(self):
 CNNMnist.model.compile(optimizer=keras.optimizers.Adam(),loss=keras.losses.sparse_categorical_crossentropy,metrics=["accuracy"])
 
 def fit(self):
 CNNMnist.model.fit(self.x_train,self.y_train,epochs=1,batch_size=32)
 
 def evaluate(self):
 test_loss,test_acc = CNNMnist.model.evaluate(self.x_test,self.y_test)
 print(test_loss,test_acc)
 
if __name__ == '__main__':
 cnn = CNNMnist()
 print(CNNMnist.model.summary())
 cnn.compile()
 cnn.fit()

#含有L1正則化的損失函數(shù)：
loss = tf.reduce_mean(tf.square(y_ - y)) + tf.contrib.layers.l1_regularizer(λ)(w)

#含有L2正則化的損失函數(shù)：
loss = tf.reduce_mean(tf.square(y_ - y)) + tf.contrib.layers.l2_regularizer(λ)(w)

import tensorflow as tf
weights = tf.constant([[1.0, -2.0], [-3.0, 4.0]])
with tf.Session() as sess:
 #計(jì)算結(jié)果為5.0
 print(sess.run(tf.contrib.layers.l1_regularizer(0.5)(weights)))
 #計(jì)算結(jié)果為15 * 1/2 = 7.5，L2正則化乘以1/2可以方便求導(dǎo)
 print(sess.run(tf.contrib.layers.l2_regularizer(0.5)(weights)))

import tensorflow as tf

#獲取一層神經(jīng)網(wǎng)絡(luò)邊上的權(quán)重，并將這個(gè)權(quán)重的L2正則化損失加入名稱為losses的集合中
def get_weight(shape, r):
 #生成一個(gè)變量
 var = tf.Variable(tf.random_normal(shape, stddev=1, seed=1), dtype=tf.float32)
 '''add_to_collection函數(shù)將這個(gè)新生成變量的L2正則化損失項(xiàng)加入集合
 這個(gè)函數(shù)的第一個(gè)參數(shù)losses是集合的名字，第二個(gè)參數(shù)是要加入這個(gè)集合的內(nèi)容'''
 tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(r)(var))
 return var

x = tf.placeholder(tf.float32, shape=(None, 2))
y_ = tf.placeholder(tf.float32, shape=(None, 1))

#定義了每一層網(wǎng)絡(luò)中節(jié)點(diǎn)的個(gè)數(shù)
layer_dimension = [2, 10, 10, 10, 1]
#神經(jīng)網(wǎng)絡(luò)的層數(shù)
n_layers = len(layer_dimension)

#這個(gè)變量維護(hù)前向傳播時(shí)最深層的節(jié)點(diǎn)，開始的時(shí)候就是輸入層
cur_layer = x
#in_dimension為當(dāng)前層的節(jié)點(diǎn)個(gè)數(shù)
in_dimension = layer_dimension[0]

#通過(guò)一個(gè)循環(huán)來(lái)生成5層全連接的神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)
for i in range(1, n_layers):
 #out_dimension為下一層的節(jié)點(diǎn)個(gè)數(shù)
 out_dimension = layer_dimension[i]
 #生成當(dāng)前層中權(quán)重的變量，并將這個(gè)變量的L2正則化損失加入losses集合
 weight = get_weight([in_dimension, out_dimension], 0.001)
 bias = tf.Variable(tf.fill([1, out_dimension], 0.1))
 #使用ReLU激活函數(shù)
 cur_layer = tf.nn.relu(tf.matmul(cur_layer, weight) + bias)
 #進(jìn)入下一層之前將下一層的節(jié)點(diǎn)個(gè)數(shù)更新為當(dāng)前層節(jié)點(diǎn)個(gè)數(shù)
 in_dimension = out_dimension

'''在定義神經(jīng)網(wǎng)絡(luò)前向傳播的同時(shí)已經(jīng)將所有的L2正則化損失加入了losses集合
這里只需要計(jì)算刻畫模型在訓(xùn)練數(shù)據(jù)上表現(xiàn)的損矣函數(shù)。'''
mse_loss = tf.reduce_mean(tf.square(y_ - cur_layer))

#將均方誤差損失函數(shù)加入損失集合
tf.add_to_collection('losses', mse_loss)

'''get_collection返回一個(gè)列表，這個(gè)列表包含所有這個(gè)集合中的元素
在這個(gè)樣例中這些元素就是損失函數(shù)的不同部分，將它們加起來(lái)就可以得到最終的損失函數(shù)。'''
loss = tf.add_n(tf.get_collection('losses'))