import tensorflow as tf
from numpy.random import RandomState
tf.compat.v1.disable_eager_execution()
se = tf.compat.v1.Session()
def print_tf(s):
print(se.run(s))
def demo1():
batch_size = 8
# 定义训练数据batch的大小
w1 = tf.Variable(tf.compat.v1.random_normal([2, 3], stddev=1, seed=1))
w2 = tf.Variable(tf.compat.v1.random_normal([3, 1], stddev=1, seed=1))
# 分别定义一二层和二三层之间的网络参数,标准差为1,随机产生的数保持一致
x = tf.compat.v1.placeholder(tf.float32, shape=(None, 2), name='x-input')
y_ = tf.compat.v1.placeholder(tf.float32, shape=(None, 1), name='y-input')
# 输入为两个维度,即两个特征,输出为一个标签,声明数据类型float32,None即一个batch大小
# y_是真实的标签
# x为(1,2),w1为(2,3),故a为(1,3),y为(1,1)
a = tf.matmul(x, w1)
y = tf.matmul(a, w2)
# 定义神经网络前向传播过程
# clip_by_value函数将y压缩在1e-10到1之间, tf.compat.v1.log计算自然对数, reduce_mean计算均值
cross_entropy = -tf.reduce_mean(y_ * tf.compat.v1.log(tf.clip_by_value(y, 1e-10, 1.0)))
train_step = tf.compat.v1.train.AdamOptimizer(0.001).minimize(cross_entropy)
# 定义损失函数和反向传播算法
rdm = RandomState(1)
dataset_size = 128
# 产生128组数据
X = rdm.rand(dataset_size, 2)
Y = [[int(x1 + x2 < 1)] for (x1, x2) in X]
# 将所有x1+x2<1的样本视为正样本,表示为1;其余为0
# 创建会话来运行TensorFlow程序
with tf.compat.v1.Session() as sess:
init_op = tf.compat.v1.global_variables_initializer()
# 初始化变量
sess.run(init_op)
print(sess.run(w1))
print(sess.run(w2))
# 打印出训练网络之前网络参数的值
STEPS = 5000
# 设置训练的轮数
for i in range(STEPS):
start = (i * batch_size) % dataset_size
end = min(start + batch_size, dataset_size)
# 每次选取batch_size个样本进行训练
sess.run(train_step, feed_dict={x: X[start:end], y_: Y[start:end]})
# 通过选取的样本训练神经网络并更新参数
if i % 1000 == 0:
total_cross_entropy = sess.run(cross_entropy, feed_dict={x: X, y_: Y})
print("After %d training step(s),cross entropy on all data is %g" % (i, total_cross_entropy))
# 每隔一段时间计算在所有数据上的交叉熵并输出,随着训练的进行,交叉熵逐渐变小
print(sess.run(w1))
print(sess.run(w2))
if __name__ == '__main__':
demo1()