最简单反向传播神经网络TensorFlow实现

#ecoding=utf-8
import numpy as np
import tensorflow as tf

#准备训练数据集
trainX = np.random.random(size=[100,4])
trainY = np.array([[int(sum(y)<2)] for y in trainX])

#定义一次输入网络样本个数
BATCH_SIZE = 8

#定义输入
x_in = tf.placeholder(tf.float32, shape=(None, trainX.shape[1]))
y_in = tf.placeholder(tf.float32, shape=(None,1))

#定义前向传播（网络结构）
w1 = tf.Variable(tf.random_normal(shape=[trainX.shape[1], 5] , stddev=1, mean=0 ))
w2 = tf.Variable(tf.random_normal(shape=[5, 1], stddev=1, mean=0))

nn1 = tf.matmul(x_in, w1)
y_pre = tf.matmul(nn1, w2)

#定义反向传播（迭代学习网络参数）
train_loss = tf.reduce_mean(tf.square(y_pre - y_in))
train_step = tf.train.GradientDescentOptimizer(learning_rate=0.002).minimize(loss=train_loss)

#训练网络
with tf.Session() as sess:
    init_op = tf.global_variables_initializer()
    sess.run(init_op)
    print('迭代前网络参数:','\nw1:\n',sess.run(w1), '\nw2:\n', sess.run(w2))

    for i in range(1000):
        start = i*BATCH_SIZE% trainX.shape[0]
        end = start + BATCH_SIZE
        sess.run(train_step, feed_dict={x_in:trainX[start:end], y_in:trainY[start:end]})
        if i % 100 == 0:
            print('train_loss:', sess.run(train_loss, feed_dict={x_in:trainX, y_in:trainY}))
    # print('ypre:', sess.run(y_pre , feed_dict={x_in:trainX}))