TensorFlow总结

第一 基础

1. 定义变量

#定义维度为[2,3], 平均值为·1， 标准差为1，类型为float32，名称为w1的服从正态分布的变量

w1 = tf.Variable(tf.random_normal(2, 3), stddev=1, seed=1, dtype=tf.float32, name='w1')

#定义维度为[2,3]，平均值为1，标准差为1，类型为float32，名称为w1的服从正态分布的变量，偏离平均值超过两个标准差则重新生成

w1 = tf.Variable(tf.truncated_normal(2, 3), stddev=1, seed=1, dtype=tf.float32, name='w1')

2.定义常量

#定义值为[[1,2], [4,5]]， 类型为float32， 名称为x的常量

x = tf.constant([[1, 2], [4, 5]], dtype=tf.float32, name='x')

3.定义占位符

#定义类型为float32，维度为[None, 2], 名称为x的占位符

x = tf.placeholder(tf.float32, shape=(None, 2), name='x')

4.开启回话

init_op = tf.initialize_all_variables()

with tf.Session() as sess:

　　sess.run(init_op)

第二  优化

1.非线性与激活函数

INPUT_NODE = 2

OUTPUT_NODE = 1

LAYER1_NODE = 3

x = tf.placeholder(tf.float32, shape=(None, INPUT_NODE), name='x-input')

_y = tf.placeholder(tf.float32, shape=(None, OUTPUT_NODE), name='y-output')

weight1 = tf.Variable(tf.truncated_normal(INPUT_NODE, LAYER1_NODE), dtype=tf.float32, name='weight1')

biases1 = tf.Variable(tf.constant(0.1, shape=LAYER1_NODE), dtype=tf.float32, name='biases1')

weight2 = tf.Variable(tf.truncated_normal(LAYER1_NODE, OUTPUT_NODE), dtype=tf.float32, name='weight2')

biases2 = tf.Variable(tf.constant(0.1, shape=OUTPUT_NODE), dtype=tf.float32, name='biases2')

layer1 = tf.nn.relu(tf.matmul(x, weight1) + biases1)

y = tf.matmul(layer1, weight2) + biases2

2.交叉熵损失函数

（1）不封装写法

cross_entropy = -tf.reduce_mean(_y * tf.log(tf.clip_by_value(y, 1e-10, 1.0)))

train_step = tf.train.AdamOptimizer(0.001).minimize(cross_entropy)

（2）封装写法

cross_entropy = tf.nn.softmax_cross_entropy_with_logits(y, _y)

train_step = tf.train.AdamOptimizer(0.001).minimize(cross_entropy)

3.均方误差损失函数

mse = tf.reduce_mean(tf.square(_y - y))

train_step = tf.train.AdamOptimizer(0.001).minimize(mse)

4.自定义损失函数

LOSS_MORE = 10

LOSS_LESS = 1

loss = tf.reduce_sum(tf.select(tf.greater(y, _y), (y- _y) * LOSS_MORE, (_y - y) * LOSS_LESS))

train_step = tf.train.AdamOptimizer(0.001).minimize(loss)

5.指数衰减学习率

LEARNING_RATE_BASE = 0.1

LEARNING_RATE_DECAY = 0.99

ALL_SIZE = 10000

BATCH_SIZE = 10

global_step = tf.Variable(0, trainable=False)

learning_rate = tf.train.exponential_decay(LEARNING_RATE_BASE, global_step, ALL_SIZE/BATCH_SIZE, LEARNING_RATE_DECAY)

train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step)

6.过拟合问题与正则化

REGULARIZATION_RATE = 0.0001

regularizer = tf.contrib.layers.l2_regularizer(REGULARIZATION_RATE)

regularization = regularizer(weight1) + regularizer(weight2)

loss = cross_entropy + regularization

7.滑动平均模型

MOVING_AVERAGE_DECAY = 0.99

ema = tf.traim.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, global_step)

ema_op = ema.apply(tf.trainable_variables())