tensorlfow自定义损失函数

tf支持自定义损失函数，使得神经网络优化更加贴近时间问题的需求。

import tensorflow as tf
import numpy
from numpy.random.mtrand import rand

w1 = tf.Variable(tf.random_normal([2, 1], stddev=1, seed=1))

# 两个输入结点
x = tf.placeholder(tf.float32, shape=(None, 2), name="x-input")
# 回归问题一般只有一个结点
y_ = tf.placeholder(tf.float32, shape=(None, 1), name='y-input')
# 单层前向传播，简单的加权和
y = tf.matmul(x, w1)


# 自定义损失函数
loss_more = 1
loss_less = 10
loss = tf.reduce_sum(tf.where(tf.greater(y, y_),
                     (y - y_) * loss_more,
                     (y_ - y) * loss_less))


learning_rate = 0.001

train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)

# 通过随机数生成一个模拟数据集
rdm = numpy.random.RandomState(1)
dataset_size = 128
X = rdm.rand(dataset_size, 2)

# x1+x2 + 噪声
Y = [[int(x1+x2) + (rand()-rand())] for (x1, x2) in X]

# 创建会话
with tf.Session() as sess:
    init_op = tf.global_variables_initializer()
    sess.run(init_op)

    print(sess.run(w1))

    # 设定训练轮数
    STEP = 5000
    # 设定batch大小
    batch_size = 8
    for i in range(STEP):
        start = (i * batch_size) % dataset_size
        end = min(start + batch_size, dataset_size)

        # 选取样本训练网络并更新参数
        sess.run(train_step,
                 feed_dict={x: X[start:end], y_: Y[start:end]})

        # 每隔一段时间输出交叉熵
        if i % 1000 == 0:
            total_cross_entropy = sess.run(
                loss, feed_dict={x: X, y_: Y})
            print("After %d traning step(s), cross entropy on all data is %g" %(i, total_cross_entropy))
    # 打印神经网络中的参数
    print(sess.run(w1))