小白学Tensorflow之多层神经网络

在本博客中&＃xff0c;我们将利用Tensorflow来构建一个多层神经网络。因为本博客是为了学习目的&＃xff0c;所以我们就来构建一个四层神经网络&＃xff0c;即一个输入层&＃xff0c;两个隐藏层&＃xff0c;一个输出层。第一&＃xff0c;我们需要定义层与层之间的转移矩阵

# 定义输入层到第一个隐藏层之间的连接矩阵
w_layer_1 &＃61; init_weights([784, 625])# 定义第一个隐藏层到第二个隐藏层之间的连接矩阵
w_layer_2 &＃61; init_weights([625, 625])# 定义第二个隐藏层到输出层之间的连接矩阵
w_layer_3 &＃61; init_weights([625, 10])

第二&＃xff0c;构建模型。在此模型中&＃xff0c;我们加入了dropout参数&＃xff0c;该参数是为了防止过拟合。也就是说&＃xff0c;如果在某一层中使用了dropout参数&＃xff0c;那么该层只有一部分神经元放电。比如&＃xff0c;dropout &＃61; 0.8&＃xff0c;那么只有80%的神经元是出于放电状态的&＃xff0c;其他都是关闭状态。

def model(X, w_layer_1, w_layer_2, w_layer_3, p_keep_input, p_keep_hidden): X &＃61; tf.nn.dropout(X, p_keep_input) hidden_1 &＃61; tf.nn.relu(tf.matmul(X, w_layer_1)) hidden_1 &＃61; tf.nn.dropout(hidden_1, p_keep_hidden) hidden_2 &＃61; tf.nn.relu(tf.matmul(hidden_1, w_layer_2)) hidden_2 &＃61; tf.nn.dropout(hidden_2, p_keep_hidden) return tf.matmul(hidden_2, w_layer_3)

完整代码&＃xff0c;如下&＃xff1a;

#!/usr/bin/env python
# -*- coding: utf-8 -*-import numpy as np
import tensorflow as tf
import input_datadef init_weights(shape):return tf.Variable(tf.random_normal(shape, stddev &＃61; 0.01))def model(X, w_layer_1, w_layer_2, w_layer_3, p_keep_input, p_keep_hidden):X &＃61; tf.nn.dropout(X, p_keep_input)hidden_1 &＃61; tf.nn.relu(tf.matmul(X, w_layer_1))hidden_1 &＃61; tf.nn.dropout(hidden_1, p_keep_hidden)hidden_2 &＃61; tf.nn.relu(tf.matmul(hidden_1, w_layer_2))hidden_2 &＃61; tf.nn.dropout(hidden_2, p_keep_hidden)return tf.matmul(hidden_2, w_layer_3)# 导入数据
mnist &＃61; input_data.read_data_sets("MNIST_data/", one_hot&＃61;True)
trX, trY, teX, teY &＃61; mnist.train.images, mnist.train.labels, mnist.test.images, mnist.test.labelsX &＃61; tf.placeholder("float", [None, 784])
Y &＃61; tf.placeholder("float", [None, 10])# 在该模型中我们一共有4层&＃xff0c;一个输入层&＃xff0c;两个隐藏层&＃xff0c;一个输出层
# 定义输入层到第一个隐藏层之间的连接矩阵
w_layer_1 &＃61; init_weights([784, 625])# 定义第一个隐藏层到第二个隐藏层之间的连接矩阵
w_layer_2 &＃61; init_weights([625, 625])# 定义第二个隐藏层到输出层之间的连接矩阵
w_layer_3 &＃61; init_weights([625, 10])# dropout 系数
# 定义有多少有效的神经元将作为输入神经元&＃xff0c;比如 p_keep_intput &＃61; 0.8&＃xff0c;那么只有80%的神经元将作为输入
p_keep_input &＃61; tf.placeholder("float")# 定义有多少的有效神经元将在隐藏层被激活
p_keep_hidden &＃61; tf.placeholder("float")# 构建模型
py_x &＃61; model(X, w_layer_1, w_layer_2, w_layer_3, p_keep_input, p_keep_hidden)cost &＃61; tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(py_x, Y))
train_op &＃61; tf.train.RMSPropOptimizer(0.001, 0.9).minimize(cost)
predict_op &＃61; tf.argmax(py_x, 1)with tf.Session() as sess:init &＃61; tf.initialize_all_variables()sess.run(init)for i in xrange(100):for start, end in zip(range(0, len(trX), 128), range(128, len(trX), 128)):sess.run(train_op, feed_dict &＃61; {X: trX[start:end], Y: trY[start:end],p_keep_input: 0.8, p_keep_hidden: 0.5})print i, np.mean(np.argmax(teY, axis &＃61; 1) &＃61;&＃61; sess.run(predict_op, feed_dict &＃61; {X: teX, Y: teY, p_keep_input: 1.0, p_keep_hidden: 1.0}))