Batch Normalization Tensorflow代码

    2015年深度学习领域，非常值得学习的一篇文献：《Batch Normalization: Accelerating Deep Network Training by  Reducing Internal Covariate Shift》，这个算法目前已经被大量的应用，最新的文献算法很多都会引用这个算法，进行网络训练。     

      在训练网络的时会将输入减去均值，还有些人甚至会对输入做白化等操作，目的是为了加快训练。白化的方式有好几种，常用的有PCA白化：即对数据进行PCA操作之后，在进行方差归一化。这样数据基本满足0均值、单位方差、弱相关性。作者首先考虑，对每一层数据都使用白化操作，但分析认为这是不可取的。因为白化需要计算协方差矩阵、求逆等操作，计算量很大，此外，反向传播时，白化操作不一定可导。于是，作者采用下面的Normalization方法。整个BN的算法如下：

其中的m是batchsize，因此BN是作用在Layer层面的归一化操作，这种操作可以和每个神经元绑定，一个Layer的每个神经元都对应着各自的gamma和beta，用于重构输入数据的分布。这两个参数和W和b一样需要学习出来的。最终的实现中beta和b的功能冗余，故去掉一个。关于BN在CNN中的使用在Group Normalization这篇文章中有直观描述：

算法的TensorFlow实现代码：

    def add_layer(inputs, in_size, out_size, activation_function=None, norm=False):
        # weights and biases (bad initialization for this case)
        Weights = tf.Variable(tf.random_normal([in_size, out_size], mean=0., stddev=1.))
        biases = tf.Variable(tf.zeros([1, out_size]) + 0.1)

        # fully connected product
        Wx_plus_b = tf.matmul(inputs, Weights) + biases

        # normalize fully connected product对每层输出做批正则化
        if norm:
            # Batch Normalize
            fc_mean, fc_var = tf.nn.moments(
                Wx_plus_b,
                axes=[0],   # the dimension you wanna normalize, here [0] for batch
                            # for image, you wanna do [0, 1, 2] for [batch, height, width] but not channel
            )
            scale = tf.Variable(tf.ones([out_size])) #需要训练出来的放缩尺度伽马
            shift = tf.Variable(tf.zeros([out_size]))#需要寻来拿出来的平移变换贝塔
            epsilon = 0.001                          #防止方差为零

            # apply moving average for mean and var when train on batch
            ema = tf.train.ExponentialMovingAverage(decay=0.5)
            def mean_var_with_update():
                ema_apply_op = ema.apply([fc_mean, fc_var])
                with tf.control_dependencies([ema_apply_op]):
                    return tf.identity(fc_mean), tf.identity(fc_var)
            mean, var = mean_var_with_update()

            Wx_plus_b = tf.nn.batch_normalization(Wx_plus_b, mean, var, shift, scale, epsilon)
            # similar with this two steps:
            # Wx_plus_b = (Wx_plus_b - fc_mean) / tf.sqrt(fc_var + epsilon)
            # Wx_plus_b = Wx_plus_b * scale + shift

        # activation
        if activation_function is None:
            outputs = Wx_plus_b
        else:
            outputs = activation_function(Wx_plus_b)

        return outputs

    fix_seed(1)

    if norm:   
        # BN for the first input对输入前数据做批正则化
        fc_mean, fc_var = tf.nn.moments(
            xs,     #输入数据
            axes=[0],
        )
        scale = tf.Variable(tf.ones([1]))
        shift = tf.Variable(tf.zeros([1]))
        epsilon = 0.001
        # apply moving average for mean and var when train on batch
        ema = tf.train.ExponentialMovingAverage(decay=0.5)
        def mean_var_with_update():
            ema_apply_op = ema.apply([fc_mean, fc_var])
            with tf.control_dependencies([ema_apply_op]):
                return tf.identity(fc_mean), tf.identity(fc_var)
        mean, var = mean_var_with_update()
        xs = tf.nn.batch_normalization(xs, mean, var, shift, scale, epsilon)

    # record inputs for every layer
    layers_inputs = [xs]

    # build hidden layers
    for l_n in range(N_LAYERS):   #循环添加神经网络层
        layer_input = layers_inputs[l_n]
        in_size = layers_inputs[l_n].get_shape()[1].value

        output = add_layer(
            layer_input,    # input
            in_size,        # input size
            N_HIDDEN_UNITS, # output size
            ACTIVATION,     # activation function
            norm,           # normalize before activation
        )
        layers_inputs.append(output)    # add output for next run

    # build output layer
    prediction = add_layer(layers_inputs[-1], 30, 1, activation_function=None)

    cost = tf.reduce_mean(tf.reduce_sum(tf.square(ys - prediction), reduction_indices=[1]))
    train_op = tf.train.GradientDescentOptimizer(0.001).minimize(cost)
    return [train_op, cost, layers_inputs]

关于Group Normalization的Tensorflow实现比BN更加简洁：

def GroupNorm(x, gamma, beta, G, eps=1e.5):
    # x: input features with shape [N,C,H,W]
    # gamma, beta: scale and offset, with shape [1,C,1,1]
    # G: number of groups for GN
    N, C, H, W = x.shape
    x = tf.reshape(x, [N, G, C // G, H, W])
    mean, var = tf.nn.moments(x, [2, 3, 4], keep dims=True)
    x = (x . mean) / tf.sqrt(var + eps)
    x = tf.reshape(x, [N, C, H, W])
    return x gamma + beta