这几天学习的内容有点多,光看定义和一些理论知识还是太抽象了,结合代码看上手比较容易。建议理论部分就去看吴恩达老师的深度学习课程,然后再看北大的tensorflow课程,先理论后实际。
目录
1. 学习率 learning rate
1.1定义
学习率:决定了参数每次更新的幅度
其中:为更新后的参数,
为当前参数,learning rate为学习率,
为损失函数的梯度
eg: , 则
= 2w+2
参数w0初始化为5,学习率为0.2,则
w1 = 5-0.2*(2*5+2)=2.6
w2 = 2.6-0.2*(2*2.6+2)=1.16
学习率大了,振荡不收敛;小了,收敛速度慢
指数衰减学习率:
其中:learning_rate_base为学习率初始值,learning_rate_decay为衰减率,
LEARNING_RATE_BASE = 0.1 #初始学习率
LEARNING_RATE_DECAY = 0.99 #学习率衰减率
LEARNING_RATE_STEP =1 #喂了多少轮batchsize后,更新一次学习率,
#一般设置喂样本总数/batchsize
#运行了几轮batchsize的计数器,初值为0,设置为不被训练
global_step = tf.Variable(0,trainable = False)
#定义指数下降学习率
learning_rate = tf.train.exponential_decay(
LEARNING_RATE_BASE,
global_step,
LEARNING_RATE_STEP,
LEARNING_RATE_DECAY,
staircase= True)
1.2示例
#coding:utf-8
#设置损失函数loss = (w+1)^2 令w初值为10
import tensorflow as tf
LEARNING_RATE_BASE = 0.1 #初始学习率
LEARNING_RATE_DECAY = 0.99 #学习率衰减率
LEARNING_RATE_STEP =1 #喂了多少轮batchsize后,更新一次学习率,
#一般设置喂样本总数/batchsize
#运行了几轮batchsize的计数器,初值为0,设置为不被训练
global_step = tf.Variable(0,trainable = False)
#定义指数下降学习率
learning_rate = tf.train.exponential_decay(
LEARNING_RATE_BASE,
global_step,
LEARNING_RATE_STEP,
LEARNING_RATE_DECAY,
staircase= True)
#定义待优化参数,初值为10
w = tf.Variable(tf.constant(5, dtype = tf.float32))
#定义损失函数loss
loss = tf.square(w+1)
#定义反向传播方法
train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step = global_step)
#生成会话,训练40轮
with tf.Session() as sess:
init_opt = tf.global_variables_initializer()
sess.run(init_opt)
for i in range(40):
sess.run(train_step)
learning_rate_val = sess.run(learning_rate)
global_step_val = sess.run(global_step)
w_val = sess.run(w)
loss_val = sess.run(loss)
print("after %s steps: global_step is %f, w is %f, learning rate is %f, loss is %f"
% (i,global_step_val, w_val,learning_rate_val, loss_val))
2.滑动平均
2.1 定义
可以参考链接 https://blog.csdn.net/tiao_god/article/details/104555827
滑动平均(影子):记录了每个参数一段时间内过往值的平均,增加了模型的泛化能力。针对所有的参数(w和b)。给参数加了影子,参数变化,影子缓慢追随。
新的影子 = 衰减率*当前影子 +(1-衰减率)*参数
其中:影子初值=参数初值
衰减率 = min (moving_average_decay, )
eg: 设置moving_average_decay 为0.99,w1为0,global_step为0,w1的滑动平均值为0
当w1更新为1时,滑动平均值=min(0.99,1/10)*0 +(1-min(0.99,1/10))*1 =0.9
当w1更新为10,global_step为100时,滑动平均值=min(0.99,101/110)*0.9 +(1-min(0.99,101/110))*10
global_step = tf.Variable(0, trainable = False)
MOVING_AVERAGE_DECAY = 0.99
ema = tf.train.ExponentialMovingAverage (MOVING_AVERAGE_DECAY, global_step)
#每运行此句,所有待优化的参数求滑动平均
ema_op = ema.apply(tf.trainable_variable())
#将训练过程和滑动平均绑定在一起,形成一个结点
with tf.control_dependencies([train_step, ema_op]):
train_op = tf.no_op(name ="train")
#查看某参数的滑动平均值
ema.average(参数名)
2.2示例
#coding:utf-8
import tensorflow as tf
w1 =tf.Variable(0,dtype = tf.float32)
global_step = tf.Variable(0, trainable = False)
MOVING_AVERAGE_DECAY = 0.99
ema = tf.train.ExponentialMovingAverage (MOVING_AVERAGE_DECAY, global_step)
#每运行此句,所有待优化的参数求滑动平均
ema_op = ema.apply(tf.trainable_variable())
with tf.Session() as sess:
init_op = tf.global_variables_initializer()
sess.run(init_op)
print(sess.run([w1, ema.average(w1)]))
#模拟经历100轮后,w1的值为10
sess.run(tf.assign(global_step,100))
sess.run(tf.assign(w1,10))
sess.run(ema_op)
print(sess.run([w1, ema.average(w1)]))
#每次sess.run会更新一次w1的huadong平均值
sess.run(ema_op)
print(sess.run([w1, ema.average(w1)]))
3.正则化
3.1定义
正则化在损失函函数中引入模型复杂度指标,利用给w加权值,弱化了训练数据的噪声(一般不正则化b)
loss = loss(y与y')+ regularizer *loss(w)
loss(y与y'):模型中所有参数的损失函数
regularizer:用超参数给出参数w在总loss的比例
w:需要正则化的参数
def get_weight (shape, regularizer):
w = tf.Variable(tf.random_normal(shape), dtype = tf.float32)
tf.add_to_collection('losses',tf.contrib.layers.l2_regularizer(regularizer)(w))
return w
3.2 复杂的示例
#coding:utf-8
#0. 导入模块,生成数据集
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
BATCH_SIZE = 30
seed =2
#基于seed产生随机数
rdm = np.random.RandomState (seed)
#随机数返回300行2列的矩阵,表示300组坐标点
X = rdm.randn(300,2)
#标记 ,作为输入数据集的标签(正确答案) 若<2则标记为1 ,否则为0
Y_ = [int (x0*x0+x1*x1<2) for (x0, x1) in X]
#1赋值为红色,其余赋值为蓝色
Y_c = [['red' if y else 'blue'] for y in Y_]
#对数据集X和标签Y进行shape整理,第一个元素为-1表示,随第二个元素计算得到,第二个元素表示列数
X = np.vstack(X).reshape(-1,2) #n行2列
Y_ = np.vstack(Y_).reshape(-1,1)#n行1列
print(X)
print(Y_)
print(Y_c)
#plt.scatter画出数据集X各行中0列元素和1列元素
plt.scatter(X[:,0], X[:,1] , c= np.squeeze (Y_c))
plt.show()
#1.定义神经网络的输入、参数和输出,定义前向传播过程
def get_weight (shape, regularizer):
w = tf.Variable(tf.random_normal(shape), dtype = tf.float32)
tf.add_to_collection('losses',tf.contrib.layers.l2_regularizer(regularizer)(w))
return w
def get_bias(shape):
b = tf.Variable(tf.constant(0.01, shape = shape))
return b
x = tf.placeholder (tf.float32, shape = (None, 2))
y_ = tf.placeholder (tf.float32, shape = (None, 1))
#输入的x为n*2的维度, 所以w1的维度为2 ,有11个节点,所以后面的维度是11
w1 = get_weight ([2,11], 0.01)
b1 = get_bias([11])
y1 = tf.nn.relu(tf.matmul(x,w1) +b1) #得到的y1维度是n*11
#最后输出的是n*1维的y,所以 w2的维度是11*1
w2 = get_weight ([11,1], 0.01)
b2 = get_bias([1])
y = tf.matmul(y1,w2) +b2 #输出层不过激活函数
#定义损失函数
loss_mse = tf.reduce_mean(tf.square(y-y_))
loss_total = loss_mse + tf.add_n(tf.get_collection('losses'))
#定义反向传播方法:不含正则化
train_step = tf.train.AdamOptimizer(0.0001).minimize(loss_mse)
#会话执行
with tf.Session() as sess:
#初始化所有变量
init_op = tf.global_variables_initializer()
sess.run(init_op)
STEPS = 40000
for i in range(STEPS):
start = (i*BATCH_SIZE) % 300
end = start + BATCH_SIZE
sess.run(train_step, feed_dict = {x:X[start:end], y_:Y_[start:end]})
#打印数据
if i%2000==0:
loss_mse_v = sess.run(loss_mse,feed_dict = {x:X, y_:Y_})
print('after' +str(i) + 'steps, loss is' + str(loss_mse_v))
#------------------------------生成测试数据,放入训练好的模型--------------------------------------------#
#xx在-3和3之间,步长为0.01
xx,yy = np.mgrid[-3:3:.01, -3:3:.01]
#将xx和yy拉直,并合并成一个2列矩阵,得到一个网格
grid = np.c_[xx.ravel(), yy.ravel()]
#将网格坐标点喂入神经网络,probs为输出
probs = sess.run(y, feed_dict= {x:grid})
#probs 的shape调整成xx的样子
probs = probs.reshape(xx.shape)
print("w1"+ str(sess.run(w1)))
print("b1"+str(sess.run(b1)))
print("w2"+str(sess.run(w2)))
print("b2"+str(sess.run(b2)))
plt.scatter(X[:,0], X[:,1], c= np.squeeze(Y_c))
plt.contour(xx,yy,probs,levels = [0.5])
plt.show()
#定义反向传播方法:含正则化
train_step = tf.train.AdamOptimizer(0.0001).minimize(loss_total)
#会话执行
with tf.Session() as sess:
init_op = tf.global_variables_initializer()
sess.run(init_op)
STEPS = 40000
for i in range(STEPS):
start = (i*BATCH_SIZE) % 300
end = start + BATCH_SIZE
sess.run(train_step, feed_dict = {x:X[start:end], y_:Y_[start:end]})
if i%2000==0:
loss_mse_v = sess.run(loss_total,feed_dict = {x:X, y_:Y_})
print('after' +str(i) + 'steps, loss is' +str(loss_mse_v))
#xx在-3和3之间,步长为0.01
xx,yy = np.mgrid[-3:3:0.01, -3:3:0.01]
#将xx和yy拉直,并合并成一个2列矩阵,得到一个网格
grid = np.c_[xx.ravel(), yy.ravel()]
#将网格坐标点喂入神经网络,probs为输出
probs = sess.run(y, feed_dict= {x:grid})
#probs 的shape调整成xx的样子
probs = probs.reshape(xx.shape)
print("w1"+ str(sess.run(w1)))
print("b1"+str(sess.run(b1)))
print("w2"+str(sess.run(w2)))
print("b2"+str(sess.run(b2)))
plt.scatter(X[:,0], X[:,1], c= np.squeeze(Y_c))
plt.contour(xx,yy,probs,levels = [0.5])
plt.show()
#执行结果(我把次数40000改成了4000,),不正则化VS正则化