BatchNorm算法详解

1 BatchNorm原理

BatchNorm通过对输入的每个mini-batch的数据进行标准化，使得网络的输入分布更加稳定。

在训练过程中，每轮迭代网络层的输入数据分布变化很大的话，使得数据抖动很大，导致权重变化也会很大，网络很难收敛。而batch norm会将数据归一化，减少不同batch间数据的抖动情况，从而提高训练速度加快收敛。

BatchNorm计算流程

输入： 设一个mini-batch为 B = { x 1... m } \mathcal{B}=\{x_{1...m}\} B={x1...m​}，$\gamma ,\beta$为可学习的参数

首先计算$\mathcal{B}$的均值：
$${\mu }_{\mathcal{B}}\leftarrow \frac{1}{m}\sum _{i=1}^m{x}_i$$
然后计算$\mathcal{B}$的方差：
$${\sigma }_{\mathcal{B}}^2\leftarrow \frac{1}{m}\sum _{i=1}^m(x_i-{\mu }_{\mathcal{B}}{)}^2$$
归一化数据：
$$\widehat{x_i}\leftarrow \frac{x_i-{\mu }_{\mathcal{B}}}{\sqrt{{\sigma }_{\mathcal{B}}^2+ϵ}}$$
其中，$ϵ$的作用是防止方差为0导致出错，$ϵ$的值为1e-5。

最后，对归一化的数据进行缩放（scale）和平移（shift）
$$y_i\leftarrow \gamma \widehat{x_i}+\beta$$
其中，$\gamma ,\beta$是通过训练学习到的。

2 BatchNorm代码实现

def batchnorm_forward(x, gamma, beta, bn_param):"""Forward pass for batch normalization.During training the sample mean and (uncorrected) sample variance arecomputed from minibatch statistics and used to normalize the incoming data.During training we also keep an exponentially decaying running mean of themean and variance of each feature, and these averages are used to normalizedata at test-time.At each timestep we update the running averages for mean and variance usingan exponential decay based on the momentum parameter:running_mean = momentum * running_mean + (1 - momentum) * sample_meanrunning_var = momentum * running_var + (1 - momentum) * sample_varInput:- x: Data of shape (N, D)- gamma: Scale parameter of shape (D,)- beta: Shift paremeter of shape (D,)- bn_param: Dictionary with the following keys:- mode: 'train' or 'test'; required- eps: Constant for numeric stability- momentum: Constant for running mean / variance.- running_mean: Array of shape (D,) giving running mean of features- running_var Array of shape (D,) giving running variance of featuresReturns a tuple of:- out: of shape (N, D)- cache: A tuple of values needed in the backward pass"""mode = bn_param['mode']eps = bn_param.get('eps', 1e-5)momentum = bn_param.get('momentum', 0.9)N, D = x.shaperunning_mean = bn_param.get('running_mean', np.zeros(D, dtype=x.dtype))running_var = bn_param.get('running_var', np.ones(D, dtype=x.dtype))if mode == 'train':sample_mean = x.mean(axis=0)sample_var = x.var(axis=0)running_mean = momentum * running_mean + (1 - momentum) * sample_meanrunning_var = momentum * running_var + (1 - momentum) * sample_varstd = np.sqrt(sample_var + eps)x_centered = x - sample_meanx_norm = x_centered / stdout = gamma * x_norm + betacache = (x_norm, x_centered, std, gamma)elif mode == 'test':x_norm = (x - running_mean) / np.sqrt(running_var + eps)out = gamma * x_norm + betaelse:raise ValueError('Invalid forward batchnorm mode "%s"' % mode)# Store the updated running means back into bn_parambn_param['running_mean'] = running_meanbn_param['running_var'] = running_varreturn out, cache

3 为什么要做滑动平均

我们一开始训练不可能获得整个训练集的均值和方差，

就算我们在训练前，把整个训练集做一次完全的forward，拿到了均值和方差，但是在模型参数变化后，均值和方差也会随之变化。所以我们要通过滑动平均的方法来获取整个训练集的均值和方差。

4 BN中的滑动平均

训练过程中的每一个batch都会进行一次滑动平均的计算：

初始值，moving_mean = 0，moving_var = 1，相当于标准正态分布。理论上初始化为任意值。momentum = 0.9

moving_mean -= (moving_mean - batch_mean) * (1 - momentum)
moving_var -= (moving_var - batch_var) * (1 - momentum)