『TensorFlow』批处理类

当前位置:

首页 > 编程开发 > 批处理教程 >

『TensorFlow』批处理类

回到顶部

基础知识

下面有莫凡的对于批处理的解释:

fc_mean,fc_var = tf.nn.moments(

    Wx_plus_b,

    axes=[0],

    # 想要 normalize 的维度, [0] 代表 batch 维度

    # 如果是图像数据, 可以传入 [0, 1, 2], 相当于求[batch, height, width] 的均值/方差, 注意不要加入 channel 维度

)

scale = tf.Variable(tf.ones([out_size]))

shift = tf.Variable(tf.zeros([out_size]))

epsilon = 0.001

Wx_plus_b = tf.nn.batch_normalization(Wx_plus_b,fc_mean,fc_var,shift,scale,epsilon)

# 上面那一步, 在做如下事情:

# Wx_plus_b = (Wx_plus_b - fc_mean) / tf.sqrt(fc_var + 0.001)

# Wx_plus_b = Wx_plus_b * scale + shift

回到顶部

tf.contrib.layers.batch_norm：封装好的批处理类

class batch_norm():

    '''batch normalization层'''

    def __init__(self, epsilon=1e-5,

                 momentum=0.9, name='batch_norm'):

        '''

        初始化

        :param epsilon:    防零极小值

        :param momentum:   滑动平均参数

        :param name:       节点名称

        '''

        with tf.variable_scope(name):

            self.epsilon = epsilon

            self.momentum = momentum

            self.name = name

    def __call__(self, x, train=True):

        # 一个封装了的会在内部调用batch_normalization进行正则化的高级接口

        return tf.contrib.layers.batch_norm(x,

                                            decay=self.momentum,        # 滑动平均参数

                                            updates_collections=None,

                                            epsilon=self.epsilon,

                                            scale=True,

                                            is_training=train,          # 影响滑动平均

                                            scope=self.name)

1.

Note: when training, the moving_mean and moving_variance need to be updated.
    By default the update ops are placed in `tf.GraphKeys.UPDATE_OPS`, so they
    need to be added as a dependency to the `train_op`. For example:

    ```python
      update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
      with tf.control_dependencies(update_ops):
        train_op = optimizer.minimize(loss)
    ```

    One can set updates_collections=None to force the updates in place, but that
    can have a speed penalty, especially in distributed settings.

2.

is_training: Whether or not the layer is in training mode. In training mode
        it would accumulate the statistics of the moments into `moving_mean` and
        `moving_variance` using an exponential moving average with the given
        `decay`. When it is not in training mode then it would use the values of
        the `moving_mean` and the `moving_variance`.

回到顶部

tf.nn.batch_normalization：原始接口封装使用

实际上tf.contrib.layers.batch_norm对于tf.nn.moments和tf.nn.batch_normalization进行了一次封装，这个类又进行了一次封装（主要是制订了一部分默认参数），实际操作时可以仅仅使用tf.contrib.layers.batch_norm函数，它已经足够方便了。

添加了滑动平均处理之后，也就是不使用封装，直接使用tf.nn.moments和tf.nn.batch_normalization实现的batch_norm函数：

def batch_norm(x,beta,gamma,phase_train,scope='bn',decay=0.9,eps=1e-5):

    with tf.variable_scope(scope):

        # beta = tf.get_variable(name='beta', shape=[n_out], initializer=tf.constant_initializer(0.0), trainable=True)

        # gamma = tf.get_variable(name='gamma', shape=[n_out],

        #                         initializer=tf.random_normal_initializer(1.0, stddev), trainable=True)

        batch_mean,batch_var = tf.nn.moments(x,[0,1,2],name='moments')

        ema = tf.train.ExponentialMovingAverage(decay=decay)

        def mean_var_with_update():

            ema_apply_op = ema.apply([batch_mean,batch_var])

            with tf.control_dependencies([ema_apply_op]):

                return tf.identity(batch_mean),tf.identity(batch_var)

                # identity之后会把Variable转换为Tensor并入图中，

                # 否则由于Variable是独立于Session的，不会被图控制control_dependencies限制

        mean,var = tf.cond(phase_train,

                           mean_var_with_update,

                           lambda: (ema.average(batch_mean),ema.average(batch_var)))

　　     normed = tf.nn.batch_normalization(x, mean, var, beta, gamma, eps)

    return normed

另一种将滑动平均展开了的方式，

def batch_norm(x, size, training, decay=0.999):

    beta = tf.Variable(tf.zeros([size]), name='beta')

    scale = tf.Variable(tf.ones([size]), name='scale')

    pop_mean = tf.Variable(tf.zeros([size]))

    pop_var = tf.Variable(tf.ones([size]))

    epsilon = 1e-3

    batch_mean, batch_var = tf.nn.moments(x, [0, 1, 2])

    train_mean = tf.assign(pop_mean, pop_mean * decay + batch_mean * (1 - decay))

    train_var = tf.assign(pop_var, pop_var * decay + batch_var * (1 - decay))

    def batch_statistics():

        with tf.control_dependencies([train_mean, train_var]):

            return tf.nn.batch_normalization(x, batch_mean, batch_var, beta, scale, epsilon, name='batch_norm')

    def population_statistics():

        return tf.nn.batch_normalization(x, pop_mean, pop_var, beta, scale, epsilon, name='batch_norm')

return tf.cond(training, batch_statistics, population_statistics)

注， tf.cond：流程控制，参数一True，则执行参数二的函数，否则执行参数三函数。

出处：https://www.cnblogs.com/hellcat/p/7380022.html

栏目列表