Dropout 解决 overfitting

声明

来源于莫烦Python：Dropout 解决 overfitting

---

往期回顾

过拟合 (Overfitting)及其解决方法

---

栗子

   图中黑色曲线是正常模型，绿色曲线就是overfitting模型。尽管绿色曲线很精确的区分了所有的训练数据，但是并没有描述数据的整体特征，对新测试数据的适应性较差。

---

   第三条曲线存在overfitting问题，尽管它经过了所有的训练点，但是不能很好的反应数据的趋势，预测能力严重不足。 TensorFlow提供了强大的dropout方法来解决overfitting问题。

---

代码

import tensorflow as tf
from sklearn.datasets import load_digits
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import LabelBinarizer

# load data
digits = load_digits()
X = digits.data
y = digits.target
y = LabelBinarizer().fit_transform(y)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=.3)


def add_layer(inputs, in_size, out_size, layer_name, activation_function=None, ):
    # add one more layer and return the output of this layer
    Weights = tf.Variable(tf.random_normal([in_size, out_size]))
    biases = tf.Variable(tf.zeros([1, out_size]) + 0.1, )
    Wx_plus_b = tf.matmul(inputs, Weights) + biases
    # here to dropout
    Wx_plus_b = tf.nn.dropout(Wx_plus_b, keep_prob)
    if activation_function is None:
        outputs = Wx_plus_b
    else:
        outputs = activation_function(Wx_plus_b, )
    tf.summary.histogram(layer_name + '/outputs', outputs)
    return outputs


# define placeholder for inputs to network
keep_prob = tf.placeholder(tf.float32)
xs = tf.placeholder(tf.float32, [None, 64])  # 8x8
ys = tf.placeholder(tf.float32, [None, 10])

# add output layer
l1 = add_layer(xs, 64, 50, 'l1', activation_function=tf.nn.tanh)
prediction = add_layer(l1, 50, 10, 'l2', activation_function=tf.nn.softmax)

# the loss between prediction and real data
cross_entropy = tf.reduce_mean(-tf.reduce_sum(ys * tf.log(prediction),
                                              reduction_indices=[1]))  # loss
tf.summary.scalar('loss', cross_entropy)
train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)

sess = tf.Session()
merged = tf.summary.merge_all()
# summary writer goes in here
train_writer = tf.summary.FileWriter("logs/train", sess.graph)
test_writer = tf.summary.FileWriter("logs/test", sess.graph)

init = tf.global_variables_initializer()
sess.run(init)

for i in range(500):
    # here to determine the keeping probability
    sess.run(train_step, feed_dict={xs: X_train, ys: y_train, keep_prob: 0.5})
    if i % 50 == 0:
        # record loss
        train_result = sess.run(merged, feed_dict={xs: X_train, ys: y_train, keep_prob: 1})
        test_result = sess.run(merged, feed_dict={xs: X_test, ys: y_test, keep_prob: 1})
        train_writer.add_summary(train_result, i)
        test_writer.add_summary(test_result, i)

---

代码注释

建立 dropout 层

---

import tensorflow as tf
from sklearn.datasets import load_digits
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import LabelBinarizer

---

keep_prob = tf.placeholder(tf.float32)
...
...
Wx_plus_b = tf.nn.dropout(Wx_plus_b, keep_prob)

---

   这里的keep_prob是保留概率，即我们要保留的结果所占比例，它作为一个placeholder，在run时传入， 当keep_prob=1的时候，相当于100%保留，也就是dropout没有起作用。 下面我们分析一下程序结构，首先准备数据

digits = load_digits()
X = digits.data
y = digits.target
y = LabelBinarizer().fit_transform(y)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=.3)

其中X_train是训练数据, X_test是测试数据。 然后添加隐含层和输出层

# add output layer
l1 = add_layer(xs, 64, 50, 'l1', activation_function=tf.nn.tanh)
prediction = add_layer(l1, 50, 10, 'l2', activation_function=tf.nn.softmax)

loss函数（即最优化目标函数）选用交叉熵函数。交叉熵用来衡量预测值和真实值的相似程度，如果完全相同，交叉熵就等于零。

cross_entropy = tf.reduce_mean(-tf.reduce_sum(ys * tf.log(prediction),
                               reduction_indices=[1]))  # loss

train方法（最优化算法）采用梯度下降法。

train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)

训练

最后开始train，总共训练500次。

sess.run(train_step, feed_dict={xs: X_train, ys: y_train, keep_prob: 0.5})
#sess.run(train_step, feed_dict={xs: X_train, ys: y_train, keep_prob: 1})

可视化结果

训练中keep_prob=1时，就可以暴露出overfitting问题。keep_prob=0.5时，dropout就发挥了作用。 我们可以两种参数分别运行程序，对比一下结果。

当keep_prob=1时，模型对训练数据的适应性优于测试数据，存在overfitting，输出如下： 红线是 train 的误差, 蓝线是 test 的误差.

当keep_prob=0.5时效果好了很多，输出如下：