tensorflow 曲线拟合
Python代码:
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
# from tensorflow.examples.tutorials.mnist import input_data
# creating data
mu,sigma=0, 0.1
data_size = 300
x_data = np.linspace(-1, 1,data_size)[:, np.newaxis]
# noise = np.random.normal(0,0.05, x_data.shape)
y_data = np.sign(x_data)
# mnist data
# mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
# x_data, y_data = mnist.train.next_batch(300)
# input layer
xs = tf.placeholder(tf.float32, [None, 1])
ys = tf.placeholder(tf.float32, [None, 1])
# layer function
def layer(data_in, size, func = None):
w = tf.Variable(tf.random_normal(size))
b = tf.Variable(tf.zeros([1, size[1]]))
z = tf.matmul(data_in, w) + b
if(func):
data_out = func(z)
else:
data_out = z
return data_out
# hidden layer
output1 = layer(xs, [1, 10], tf.nn.relu)
output2 = layer(output1, [10, 20], tf.nn.softmax)
output3 = layer(output2, [20, 20], tf.nn.relu)
output4 = layer(output3, [20, 10], tf.nn.softmax)
output5 = layer(output4, [10, 10], tf.nn.relu)
# output layer
out = layer(output5, [10, 1])
# loss function
# loss = tf.reduce_sum(ys * tf.log(out))
loss = tf.reduce_mean(tf.reduce_sum(tf.square((out - ys))))
# trainning method
# train = tf.train.GradientDescentOptimizer(0.1).minimize(loss)
train = tf.train.AdamOptimizer().minimize(loss)
# init all variables
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
# print loss value for every 50 times loop
print_step = 50
# loop less than 50 * 1000
loop_max_count = 1000
while True:
print_step -= 1
_,loss_value = sess.run([train,loss],feed_dict={xs:x_data,ys:y_data})
if(print_step == 0):
print(loss_value)
print_step = 50
loop_max_count -= 1
if(loss_value < .00001 or loop_max_count <= 0):
break
# print loop times and show the output
print("loop_count = ", (1000 - loop_max_count) * 50)
y_out = sess.run(out, feed_dict={xs:x_data})
plt.plot(x_data, y_out, label="out")
plt.plot(x_data, y_data, label="in")
plt.show()
可以用来看看不同数目的隐含层和不同的激活函数对曲线拟合的训练性能和训练结果有何影响。