tensorflow训练权重不更新_TensorFlow学习Program1——15.权重和偏置更新要点、单层感知机、反向传播算法BPN...

# tensorflow实现单层感知机&＃xff0c;是一个单层神经网络。
# 它使用阈值激活函数&＃xff0c;只能解决线性可分的问题&＃xff0c;不能使用TensorFlow优化器来更新权重
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt# from Activation_Function_Threshold import thresholddef threshold(x):# tf.less返回两个张量各元素比较&＃xff08;x# hyper parameters
learn_rate &＃61; 0.4
min_err &＃61; 1e-3 # minmum accepted error
max_epochs &＃61; 100 # maximum epochs# 2.指定训练数据。在这个例子中&＃xff0c;取三个输入神经元&＃xff08;A&＃xff0c;B&＃xff0c;C&＃xff09;并训练它学习逻辑 AB&＃43;BC&＃xff1a;
# traning data Y&＃61;AB&＃43;BC, sum of two linear functions
# 疑问&＃xff0c;AB算是线性函数&＃xff1f;线性X线性&＃61;线性&＃xff1f; 好像多项式都算线性&＃xff1f;
T, F &＃61; 1.0, 0.0
X_in &＃61; [[T, T, T, T],[T, T, F, T],[T, F, T, T],[T, F, F, T],[F, T, T, T],[F, T, F, T],[F, F, T, T],[F, F, F, T]]
Y &＃61; [[T],[T],[F],[F],[T],[F],[F],[F]]
# 3.定义要用到的变量和用于计算更新的计算图&＃xff0c;最后执行计算图
W &＃61; tf.Variable(tf.random_normal([4, 1], stddev&＃61;2, seed&＃61;0))
h &＃61; tf.matmul(X_in, W)
Y_hat &＃61; threshold(h)
error &＃61; Y - Y_hat
mean_error &＃61; tf.reduce_mean(tf.square(error))
# 用mean_err作为判断标准&＃xff0c;但不用来修订dW
dW &＃61; learn_rate * tf.matmul(X_in, error, transpose_a&＃61;True)
# shape(dW)&＃61;[4,1]&＃61;倒置[8,4]*[8,1]&＃61;[4,8]*[8,1]
train &＃61; tf.assign(W, W &＃43; dW)
init &＃61; tf.global_variables_initializer()
err &＃61; 1
epoch &＃61; 0
with tf.Session() as sess:sess.run(init)while err > min_err and epoch

# MNIST数据集是分类问题的标准。本例用TensorFlow实现反向传播算法
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
mnist &＃61; input_data.read_data_sets(&＃39;A:/software/Anaconda3/Lib/site-packages/tensorflow/examples/tutorials/mnist/MNIST_data&＃39;, one_hot&＃61;True)# 1. 定义超参数和其他常量。
# 每个手写数字的尺寸是 28×28&＃61;784 像素。数据集被分为 10 类&＃xff0c;以 0 到 9 之间的数字表示。这两点是固定的。
# 学习率、最大迭代周期数、每次批量训练的批量大小以及隐藏层中的神经元数量都是超参数。
# 可以通过调整这些超参数&＃xff0c;看看它们是如何影响网络表现的&＃xff1a;
n_input &＃61; 784
n_classes &＃61; 10
max_epochs &＃61; 1000
learn_rate &＃61; 0.5
batch_size &＃61; 10
seed &＃61; 0
n_hidden &＃61; 30 # Number of neurons in the hidden layer# 2. 定义Sigmoid函数s(x)的导数来进行权重更新,其导数为s(x)*(1-s(x))
def sigmaprime(x):return tf.multiply(tf.sigmoid(x), tf.subtract(tf.constant(1.0), tf.sigmoid(x)))# 3. 为训练数据创建占位符
x_in &＃61; tf.placeholder(tf.float32, [None, n_input])
y &＃61; tf.placeholder(tf.float32, [None, n_classes])# 4. 创建模型
def multilayer_perceptron(x, weights, biases):# Hidder layer with RELU activationh_layer_1 &＃61; tf.add(tf.matmul(x, weights[&＃39;h1&＃39;]), biases[&＃39;h1&＃39;])out_layer_1 &＃61; tf.sigmoid(h_layer_1)# output layer with linear activationh_out &＃61; tf.add(tf.matmul(out_layer_1, weights[&＃39;out&＃39;]), biases[&＃39;out&＃39;])return tf.sigmoid(h_out), h_out, out_layer_1, h_layer_1# 5. 定义权重和偏置变量
weights &＃61; {&＃39;h1&＃39;: tf.Variable(tf.random_normal([n_input, n_hidden], seed&＃61;seed)),&＃39;out&＃39;: tf.Variable(tf.random_normal([n_hidden, n_classes], seed&＃61;seed))}
biases &＃61; {&＃39;h1&＃39;: tf.Variable(tf.random_normal([1, n_hidden], seed&＃61;seed)),&＃39;out&＃39;: tf.Variable(tf.random_normal([1, n_classes], seed&＃61;seed))}# 6. 为正向传播、误差、梯度和更新计算创建计算图&＃xff1a;
# Forward pass
y_hat, h_2, o_1, h_1 &＃61; multilayer_perceptron(x_in, weights, biases)
# error
err &＃61; y_hat - y
# backward pass
delta_2 &＃61; tf.multiply(err, sigmaprime(h_2))
delta_w_2 &＃61; tf.matmul(tf.transpose(o_1), delta_2)
wtd_error &＃61; tf.matmul(delta_2, tf.transpose(weights[&＃39;out&＃39;]))
delta_1 &＃61; tf.multiply(wtd_error, sigmaprime(h_1))
delta_w_1 &＃61; tf.matmul(tf.transpose(x_in), delta_1)eta &＃61; tf.constant(learn_rate)
# 7.更新权重update weights
# 注意为什么这里是 W&＃61;W-dW,而不是W&＃61;W&＃43;dW。因为当err&＃61;y_hat-y,就是W&＃61;W-dW&＃xff0c;当err&＃61;y-y_hat,则W&＃61;W&＃43;dW
step &＃61; [tf.assign(weights[&＃39;h1&＃39;], tf.subtract(weights[&＃39;h1&＃39;], tf.multiply(eta, delta_w_1))),tf.assign(biases[&＃39;h1&＃39;], tf.subtract(biases[&＃39;h1&＃39;], tf.multiply(eta, tf.reduce_mean(delta_1, axis&＃61;[0])))),tf.assign(weights[&＃39;out&＃39;], tf.subtract(weights[&＃39;out&＃39;], tf.multiply(eta, delta_w_2))),tf.assign(biases[&＃39;out&＃39;], tf.subtract(biases[&＃39;out&＃39;], tf.multiply(eta, tf.reduce_mean(delta_2, axis&＃61;[0])))),]
# 8.定义计算精度accuracy和初始化
acc_mat &＃61; tf.equal(tf.argmax(y_hat, 1), tf.argmax(y, 1))
accuracy &＃61; tf.reduce_sum(tf.cast(acc_mat, tf.float32))
init &＃61; tf.global_variables_initializer()# 9.执行
with tf.Session() as sess:sess.run(init)for epoch in range(max_epochs):batch_xs, batch_ys &＃61; mnist.train.next_batch(batch_size)sess.run(step, feed_dict&＃61;{x_in: batch_xs, y: batch_ys})if epoch % 100 &＃61;&＃61; 0:acc_test &＃61; sess.run(accuracy, feed_dict&＃61;{x_in: mnist.test.images, y: mnist.test.labels})acc_train &＃61; sess.run(accuracy, feed_dict&＃61;{x_in: mnist.train.images, y: mnist.train.labels})# 训练集和测试集样本大小分别为55000和10000&＃xff0c;这里所求精度即向满分55000和10000靠齐&＃xff0c;除以550和100就可获得相对百分数# 也有写600的&＃xff0c;认为训练集为60000&＃xff0c;但是我验证了&＃xff0c;是train&＃61;55000, validation&＃61;5000print(&＃39;Epoch:{0} accuracy train%:{1} accuracy test%: {2}&＃39;.format(epoch, acc_train / 550, (acc_test / 100)))