文章目录
深度学习与深层神经网络
深度学习的精确定义为“一类通过多层非线性变换对高复杂性数据建模算法的合集”。
激活函数去线性化:
ReLU:
f
(
x
)
=
m
a
x
(
x
,
0
)
f(x)=max(x,0)
f(x)=max(x,0)
sigmoid:
f
(
x
)
=
1
1
+
e
−
x
f(x)=\frac{1}{1+e^{-x}}
f(x)=1+e−x1
tanh:
f
(
x
)
=
1
+
e
−
2
x
1
+
e
−
2
x
f(x)=\frac{1+e^{-2x}}{1+e^{-2x}}
f(x)=1+e−2x1+e−2x
多层网络解决异或运算
损失函数的定义
分类
交叉熵:刻画了两个概率分布之间的距离,它是分类问题中使用比较广的一种损失函数。需要概率来计算。它刻画的是
通过概率分布q来表达概率分布p的困难程度
softmax回归本身可以作为一个学习算法来优化分类结果,但在TensorFlow 中,softmax回归的参数被去掉了,它只是一层额外的处理层,将神经网络的输出变成一个概率分布。
使用了Softmax回归之后的交叉熵:
cross_entropy= tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y, logits=y)
回归
均方误差(MSE, mean squared error)
mse = tf.reduce_mean(tf.square(y_-y))
自定义损失函数
TensorFlow支持自定义损失函数
import tensorflow as tf
from numpy.random import RandomState
batch_size = 8
x = tf.placeholder(tf.float32, shape=(None, 2), name="x-input")
y_ = tf.placeholder(tf.float32, shape=(None, 1), name="y-input")
w1 = tf.Variable(tf.random_normal([2, 1], stddev=1, seed=1))
y = tf.matmul(x, w1)
# 自定义损失函数
loss_less = 10
loss_more = 1
loss = tf.reduce_sum(tf.where(tf.greater(y, y_), (y-y_)*loss_more, (y_-y)*loss_less))
train_step = tf.train.AdamOptimizer(0.001).minimize(loss)
rdm = RandomState(1)
dataset_size = 128
X = rdm.rand(dataset_size, 2)
Y = [[x1 + x2 + rdm.rand()/10.0 - 0.05] for (x1, x2) in X]
with tf.Session() as sess:
init_op = tf.global_variables_initializer()
sess.run(init_op)
STEPS = 5000 # 训练轮数
for i in range(STEPS):
# 每次选取batch_size=8个样本进行训练。
start = (i * batch_size) % dataset_size
end = min(start + batch_size, dataset_size)
sess.run(train_step, feed_dict={x: X[start:end], y_: Y[start:end]})
print(sess.run(w1))
[[1.019347 ]
[1.0428089]]
神经网络优化算法
《Learning representations by back-propagating errors》
梯度下降算法:不一定能达到全局最优,计算时间太长。
随机梯度下降算法:每次随机挑选一个
综合:每次选取一小部分数据"batch"
神经网络的进一步优化
学习率的设置
指数衰减法
# decayed_learning_rate = learning_rate * decay_rate ^ (global_step / decay_steps)
# decayed_learning_rate 为每一轮优化时使用的学习率
# learning_rate 为事先设定的初始学习率
# decay_rate 为衰减系数
# decay_steps 为衰减速度
#====================通过expinential_decay函数生成学习率
#learning_rate= tf.train.exponential_decay(0.1, global_step, 100, 0.96, staircase=True)
#训练100轮后学习率乘以0.96.
过拟合问题
所谓过拟合,指的是当一个模型过为复杂之后,它可以很好地“记忆”每一个训练数据中随机噪音的部分而忘记了要去“学习”训练数据中通用的趋势。
正则化:正则化的思想就是在损失函数中加入刻画模型复杂程度的指标
损失函数
J
(
θ
)
J(\theta)
J(θ),那么在优化时不是直接优化
J
(
θ
)
J(\theta)
J(θ),而是优化
J
(
θ
)
+
λ
R
(
w
)
J(\theta)+\lambda R(w)
J(θ)+λR(w).其中
R
(
w
)
R(w)
R(w)刻画的是模型的复杂程度,而
λ
\lambda
λ表示模型复杂损失在总损失中的比例。基本的思想都是希望通过限制权重的大小,使得模型不能任意拟合训练数据中的随机噪音。
L1正则化:更稀疏
R
(
w
)
=
∣
∣
w
∣
∣
1
=
∑
i
∣
w
i
∣
R(w)=||w||_{1}=\sum_{i}|w_{i}|
R(w)=∣∣w∣∣1=∑i∣wi∣
L2正则化:
R
(
w
)
=
∣
∣
w
∣
∣
2
2
=
∑
i
∣
w
i
2
∣
R(w)=||w||_{2}^{2}=\sum_{i}|w_{i}^{2}|
R(w)=∣∣w∣∣22=∑i∣wi2∣
# 简单的带L2正则化的损失函数定义
# w = tf.Variable(tf.random_normal([2, 1], stddev=1, seed=1))
# y = tf.matmul(x, w)
# loss = tf.reduce_mean(tf.square(y_ - y)) + \ # 均方误差损失函数
# tf.contrib.layers.l2_regularizer(lambda)(w) # 正则化
import tensorflow as tf
weights = tf.constant([[1.0, -2.0], [-3.0, 4.0]])
with tf.Session() as sess:
print(sess.run(tf.contrib.layers.l1_regularizer(.5)(weights)))
print(sess.run(tf.contrib.layers.l2_regularizer(.5)(weights)))
5.0
7.5
通过集合计算一个5层神经网络带L2正则化的损失函数的计算方法
import tensorflow as tf
# 获取一层神经网络边上的权重,并将这个权重的L2正则化损失加入名称为“losses”的集合中
def get_weight(shape, lambda1):
var = tf.Variable(tf.random_normal(shape), dtype=tf.float32)
tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(lambda1)(var))
return var
x = tf.placeholder(tf.float32, shape=(None, 2))
y_ = tf.placeholder(tf.float32, shape=(None, 1))
batch_size = 8
# 定义每一层网络中节点的个数
layer_dimension = [2, 10, 10, 10, 1]
# 神经网络的层数
n_layers = len(layer_dimension)
# 这个变量维护前向传播时最深层的节点,开始的时候就是输入层
cur_layer = x
# 当前层的节点个数
in_dimension = layer_dimension[0]
# 循环生成网络结构
for i in range(1, n_layers):
out_dimension = layer_dimension[i] # 下一层的节点个数
# 当前层中权重的变量,并将这个变量的L2正则化损失加入计算图上的集合
weight = get_weight([in_dimension, out_dimension], 0.001)
bias = tf.Variable(tf.constant(0.1, shape=[out_dimension]))
# 使用ReLU激活函数
cur_layer = tf.nn.relu(tf.matmul(cur_layer, weight) + bias)
# 进入下一层之前将下一层的节点个数更新为当前层节点个数
in_dimension = layer_dimension[i]
# 在定义神经网络前向传播的同时已经将所有的L2正则化损失加入了图上的集合,
# 这里只需要计算刻画模型在训练数据上表现的损失函数。
mse_loss = tf.reduce_mean(tf.square(y_ - cur_layer))
# 将均方误差损失函数加入损失集合。
tf.add_to_collection("losses", mse_loss)
# get_collection返回一个列表,这个列表是所有这个集合中的元素。在这个样例中,
# 这些元素就是损失函数的不同部分,将它们加起来就可以得到最终的损失函数。
loss = tf.add_n(tf.get_collection("losses"))
例子
1. 生成模拟数据集。
import tensorflow as tf
import matplotlib.pyplot as plt
import numpy as np
data = []
label0 = []
np.random.seed(0)
# 以原点为圆心,半径为1的圆把散点划分成红蓝两部分,并加入随机噪音。
for i in range(150):
x1 = np.random.uniform(-1, 1)
x2 = np.random.uniform(0, 2)
if x1**2 + x2**2 <= 1:
data.append([np.random.normal(x1, 0.1),np.random.normal(x2, 0.1)])
label0.append(0)
else:
data.append([np.random.normal(x1, 0.1), np.random.normal(x2, 0.1)])
label0.append(1)
data = np.hstack(data).reshape(-1, 2)
plt.scatter(data[:, 0], data[:, 1], c=label0,
cmap="RdBu", vmin=-0.2, vmax=1.2, edgecolor="white")
plt.show()
label = np.hstack(label0).reshape(-1, 1)
2. 定义一个获取权重,并自动加入正则项到损失的函数。
def get_weight(shape, lambda1):
var = tf.Variable(tf.random_normal(shape), dtype=tf.float32)
tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(lambda1)(var))
return var
3. 定义神经网络。
x = tf.placeholder(tf.float32, shape=(None, 2))
y_ = tf.placeholder(tf.float32, shape=(None, 1))
sample_size = len(data)
# 每层节点的个数
layer_dimension = [2, 10, 5, 3, 1]
n_layers = len(layer_dimension)
cur_layer = x
in_dimension = layer_dimension[0]
# 循环生成网络结构
for i in range(1, n_layers):
out_dimension = layer_dimension[i]
weight = get_weight([in_dimension, out_dimension], 0.003)
bias = tf.Variable(tf.constant(0.1, shape=[out_dimension]))
cur_layer = tf.nn.relu(tf.matmul(cur_layer, weight) + bias)
in_dimension = layer_dimension[i]
y= cur_layer
# 损失函数的定义。
mse_loss = tf.reduce_sum(tf.pow(y_ - y, 2)) / sample_size
tf.add_to_collection('losses', mse_loss)
loss = tf.add_n(tf.get_collection('losses'))
4. 训练不带正则项的损失函数mse_loss。
# 定义训练的目标函数mse_loss,训练次数及训练模型
train_op = tf.train.AdamOptimizer(0.001).minimize(mse_loss)
TRAINING_STEPS = 40000
with tf.Session() as sess:
tf.global_variables_initializer().run()
for i in range(TRAINING_STEPS):
sess.run(train_op, feed_dict={x: data, y_: label})
if i % 2000 == 0:
print("After %d steps, mse_loss: %f" % (i,sess.run(mse_loss, feed_dict={x: data, y_: label})))
# 画出训练后的分割曲线
xx, yy = np.mgrid[-1.2:1.2:.01, -0.2:2.2:.01]
grid = np.c_[xx.ravel(), yy.ravel()]
probs = sess.run(y, feed_dict={x:grid})
probs = probs.reshape(xx.shape)
plt.scatter(data[:,0], data[:,1], c=label0,
cmap="RdBu", vmin=-.2, vmax=1.2, edgecolor="white")
plt.contour(xx, yy, probs, levels=[.5], cmap="Greys", vmin=0, vmax=.1)
plt.show()
After 0 steps, mse_loss: 0.597972
After 2000 steps, mse_loss: 0.024146
After 4000 steps, mse_loss: 0.021657
After 6000 steps, mse_loss: 0.015733
After 8000 steps, mse_loss: 0.010555
After 10000 steps, mse_loss: 0.007910
After 12000 steps, mse_loss: 0.007517
After 14000 steps, mse_loss: 0.006668
After 16000 steps, mse_loss: 0.006667
After 18000 steps, mse_loss: 0.006667
After 20000 steps, mse_loss: 0.006667
After 22000 steps, mse_loss: 0.006667
After 24000 steps, mse_loss: 0.006667
After 26000 steps, mse_loss: 0.006667
After 28000 steps, mse_loss: 0.006667
After 30000 steps, mse_loss: 0.006667
After 32000 steps, mse_loss: 0.006667
After 34000 steps, mse_loss: 0.006667
After 36000 steps, mse_loss: 0.006667
After 38000 steps, mse_loss: 0.006667
5. 训练带正则项的损失函数loss。
# 定义训练的目标函数loss,训练次数及训练模型
train_op = tf.train.AdamOptimizer(0.001).minimize(loss)
TRAINING_STEPS = 40000
with tf.Session() as sess:
tf.global_variables_initializer().run()
for i in range(TRAINING_STEPS):
sess.run(train_op, feed_dict={x: data, y_: label})
if i % 2000 == 0:
print("After %d steps, loss: %f" % (i, sess.run(loss, feed_dict={x: data, y_: label})))
# 画出训练后的分割曲线
xx, yy = np.mgrid[-1:1:.01, 0:2:.01]
grid = np.c_[xx.ravel(), yy.ravel()]
probs = sess.run(y, feed_dict={x:grid})
probs = probs.reshape(xx.shape)
plt.scatter(data[:,0], data[:,1], c=label0,
cmap="RdBu", vmin=-0.2, vmax=1.2, edgecolor="white")
plt.contour(xx, yy, probs, levels=[0.5], cmap="Greys", vmin=0, vmax=.1)
plt.show()
After 0 steps, loss: 0.640029
After 2000 steps, loss: 0.048152
After 4000 steps, loss: 0.039120
After 6000 steps, loss: 0.038880
After 8000 steps, loss: 0.038837
After 10000 steps, loss: 0.038654
After 12000 steps, loss: 0.038642
After 14000 steps, loss: 0.038642
After 16000 steps, loss: 0.038645
After 18000 steps, loss: 0.038643
After 20000 steps, loss: 0.038642
After 22000 steps, loss: 0.038644
After 24000 steps, loss: 0.038644
After 26000 steps, loss: 0.038642
After 28000 steps, loss: 0.038642
After 30000 steps, loss: 0.038643
After 32000 steps, loss: 0.038644
After 34000 steps, loss: 0.038643
After 36000 steps, loss: 0.038644
After 38000 steps, loss: 0.038643
滑动平均模型
参考链接:https://blog.csdn.net/Sengo_GWU/article/details/80519250
在未知数据上更加健壮
tf.train.ExponentialMovingAverage来实现滑动平均模型
import tensorflow as tf
# 定义一个变量用于计算滑动平均,这个变量的初始值为0.
v1 = tf.Variable(0, dtype=tf.float32)
step = tf.Variable(0, trainable=False) # 模拟神经网络中迭代的轮数,可以用于动态控制衰减率
# 定义一个滑动平均的类,初始化时给定了衰减率为0.99和控制衰减率的变量step
ema = tf.train.ExponentialMovingAverage(0.99, step)
# 定义一个更新变量滑动平均的操作,每次执行这个操作时这个列表中的变量都会被更新
maintain_averages_op = ema.apply([v1])
with tf.Session() as sess:
# 初始化
init_op = tf.global_variables_initializer()
sess.run(init_op)
print(sess.run([v1, ema.average(v1)]))
# 更新变量v1的取值
sess.run(tf.assign(v1, 5))
# 更新v1的滑动平均值,衰减率为min{0.99,(1+step)/(10+step)=0.1}=0.1
# v1的滑动平均会被更新为0.1*0+0.9*5=0.45
sess.run(maintain_averages_op)
print(sess.run([v1, ema.average(v1)]) )
# 更新step和v1的取值
sess.run(tf.assign(step, 10000))
sess.run(tf.assign(v1, 10))
sess.run(maintain_averages_op)
print(sess.run([v1, ema.average(v1)]))
# 更新一次v1的滑动平均值
sess.run(maintain_averages_op)
print(sess.run([v1, ema.average(v1)]))
[0.0, 0.0]
[5.0, 4.5]
[10.0, 4.555]
[10.0, 4.60945]