文章作者：Tyan
博客：noahsnail.com ?|? CSDN ?|? 簡書

本文主要是介紹利用tensorflow創(chuàng)建一個簡單的神經(jīng)網(wǎng)絡(luò)并進(jìn)行訓(xùn)練。

#!/usr/bin/env python
# _*_ coding: utf-8 _*_

import tensorflow as tf
import numpy as np

# 創(chuàng)建一個神經(jīng)網(wǎng)絡(luò)層
def add_layer(input, in_size, out_size, activation_function = None):
    """
    :param input:
        神經(jīng)網(wǎng)絡(luò)層的輸入
    :param in_zize:
        輸入數(shù)據(jù)的大小
    :param out_size:
        輸出數(shù)據(jù)的大小
    :param activation_function:
        神經(jīng)網(wǎng)絡(luò)激活函數(shù)，默認(rèn)沒有
    """

    # 定義神經(jīng)網(wǎng)絡(luò)的初始化權(quán)重
    Weights = tf.Variable(tf.random_normal([in_size, out_size]))
    # 定義神經(jīng)網(wǎng)絡(luò)的偏置
    biases = tf.Variable(tf.zeros([1, out_size]) + 0.1)
    # 計算w*x+b
    W_mul_x_plus_b = tf.matmul(input, Weights) + biases
    # 根據(jù)是否有激活函數(shù)進(jìn)行處理
    if activation_function is None:
        output = W_mul_x_plus_b
    else:
        output = activation_function(W_mul_x_plus_b)

    return output

# 創(chuàng)建一個具有輸入層、隱藏層、輸出層的三層神經(jīng)網(wǎng)絡(luò)，神經(jīng)元個數(shù)分別為1,10,1
# 創(chuàng)建只有一個特征的輸入數(shù)據(jù)，數(shù)據(jù)數(shù)目為300，輸入層
x_data = np.linspace(-1, 1, 300)[:, np.newaxis]
# 創(chuàng)建數(shù)據(jù)中的噪聲
noise = np.random.normal(0, 0.05, x_data.shape)
# 創(chuàng)建輸入數(shù)據(jù)對應(yīng)的輸出
y_data = np.square(x_data) + 1 + noise

# 定義輸入數(shù)據(jù)，None是樣本數(shù)目，表示多少輸入數(shù)據(jù)都行，1是輸入數(shù)據(jù)的特征數(shù)目
xs = tf.placeholder(tf.float32, [None, 1])
# 定義輸出數(shù)據(jù)，與xs同理
ys = tf.placeholder(tf.float32, [None, 1])

# 定義一個隱藏層
hidden_layer = add_layer(xs, 1, 10, activation_function = tf.nn.relu)
# 定義輸出層
prediction = add_layer(hidden_layer, 10, 1, activation_function = None)

# 求解神經(jīng)網(wǎng)絡(luò)參數(shù)
# 定義損失函數(shù)
loss = tf.reduce_mean(tf.reduce_sum(tf.square(ys - prediction), reduction_indices = [1]))
# 定義訓(xùn)練過程
train_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss)
# 變量初始化
init = tf.global_variables_initializer()
# 定義Session
sess = tf.Session()
# 執(zhí)行初始化工作
sess.run(init)
# 進(jìn)行訓(xùn)練
for i in range(1000):
    # 執(zhí)行訓(xùn)練，并傳入數(shù)據(jù)
    sess.run(train_step, feed_dict = {xs: x_data, ys: y_data})
    if i % 100 == 0:
        print sess.run(loss, feed_dict = {xs: x_data, ys: y_data})
# 關(guān)閉Session
sess.close()

執(zhí)行結(jié)果如下：

1.06731
0.0111914
0.00651229
0.00530187
0.00472237
0.00429948
0.00399815
0.00377548
0.00359714
0.00345819