定义单隐藏层前馈网络模型的训练样本X和Y、定义输入x，输出y，隐藏层参数分别定义为w1和b1，隐藏层的激活函数选取ReLU；输出层参数为w2，b2，输出层激活函数选取sigmoid。定义深度前馈网络模型输出out，定义损失函数为均方差损失函数loss，并且使用Adam算法的Optimizer

deep_feedforward_networks_03.py

import tensorflow as tf

# 输入训练数据，这里是python的list, 也可以定义为numpy的ndarray
x_data = [[1., 0.], [0., 1.], [0., 0.], [1., 1.]]
x = tf.placeholder(tf.float32, shape=[None, 2])  # 定义占位符，占位符在运行图的时候必须feed数据
y_data = [[1], [1], [0], [0]]  # 训练数据的标签，注意维度
y = tf.placeholder(tf.float32, shape=[None, 1])
# 定义variables，在运行图的过程中会被按照优化目标改变和保存
weights = {'w1': tf.Variable(tf.random_normal([2, 16])),
           'w2': tf.Variable(tf.random_normal([16, 1]))}
bias = {'b1': tf.Variable(tf.zeros([1])),
        'b2': tf.Variable(tf.zeros([1]))}  # b1,b2初始为0，正态化初始也可


# 定义神经网络计算图
def nn(x, weights, bias):
    d1 = tf.matmul(x, weights['w1']) + bias['b1']
    d1 = tf.nn.relu(d1)
    d2 = tf.matmul(d1, weights['w2']) + bias['b2']
    d2 = tf.nn.sigmoid(d2)
    return d2


pred = nn(x, weights, bias)  # 预测值
cost = tf.reduce_mean(tf.square(y - pred))  # 损失函数
learning_rate = 0.01  # 学习率取0.01
# 定义tf.train用来训练
# train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)  ## max_step: 20000, loss: 0.002638
train_step = tf.train.AdamOptimizer(learning_rate).minimize(cost)  ## max_step: 2000, loss: 0.000014
init = tf.global_variables_initializer()  # 初始化参数，图运行的一开始必须初始化所有变量
# 运行图
with tf.Session() as sess:
    sess.run(init)
    max_step = 500
    for i in range(max_step + 1):
        sess.run(train_step, feed_dict={x: x_data, y: y_data})
        loss = sess.run(cost, feed_dict={x: x_data, y: y_data})
        if i % 100 == 0:
            print('step: ' + str(i) + '  loss:' + "{:.6f}".format(loss))  # + '    accuracy:' + "{:.6f}".format(acc))
            print(sess.run(pred, feed_dict={x: x_data}))
    print('end')