TensorFlow激励函数

激励函数

  常见的激励函数有：线性激励函数、阈值或阶跃激励函数、S形激励函数、双曲正切激励函数和高斯激励函数等。

  TensorFlow三种常用激励 relu sigmoid tanh 本次利用relu激励进行离散点拟合

  激励函数（activation function）是一群空间魔法师，扭曲翻转特征空间，在其中寻找线性的边界。

激励函数可以引入非线性因素，解决线性模型所不能解决的问题

卷积神经网络中激励函数应用

完整代码实现

# 激励函数
# 三种激励 relu sigmoid tanh
# 多层神经网络时 激励选择需慎重 会导致梯度消失或梯度爆炸
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt


def add_layer(inputs, inSize, outSize, activationFunction=None):
    # 输入层
    with tf.name_scope('Inputs'):
        weights = tf.Variable(tf.random_normal([inSize, outSize]))  # 矩阵
        biases = tf.Variable(tf.zeros([1, outSize]) + 0.1)
        plusResult = tf.matmul(tf.cast(inputs, tf.float32), weights) + biases
        if activationFunction is None:
            outputs = plusResult
        else:
            outputs = activationFunction(plusResult)
        return outputs


xData = np.linspace(-1, 1, 300)[:, np.newaxis]
noise = np.random.normal(0, 0.05, xData.shape)
yData = np.square(xData) - 0.5 + noise

# 输入层
with tf.name_scope('Inputs'):
    xs = tf.placeholder(tf.float32, [None, 1], name='xInput')
    ys = tf.placeholder(tf.float32, [None, 1], name='yInput')

# relu激励
layer1 = add_layer(xData, 1, 10, activationFunction=tf.nn.relu)
predition = add_layer(layer1, 10, 1, activationFunction=None)

loss = tf.square(ys - predition)
add = tf.reduce_sum(loss, reduction_indices=[1])
averloss = tf.reduce_mean(add)

trainStep = tf.train.GradientDescentOptimizer(0.1).minimize(averloss)  # 训练趋向

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

fig = plt.figure()
ax = fig.add_subplot(1, 1, 1)
ax.scatter(xData, yData)
plt.ion()
plt.show()

for i in range(1000):
    sess.run(trainStep, feed_dict={xs: xData, ys: yData})
    if i % 50 == 0:
        try:
            ax.lines.remove(lines[0])
        except Exception:
            pass
        predictionValue = sess.run(predition, feed_dict={xs: xData})
        lines = ax.plot(xData, predictionValue, 'r-', lw=5)
        plt.pause(0.1)

注意：如果用Pycharm的同学出不来效果，请顺次点击File->Settings->Tools->Python Scientific->Show plots in tool window选项的勾打没即可

最终效果（红线为拟合结果）