pytorch BN函数说明

pytorch中的BN继承自：

class _BatchNorm(Module):

这个类的代码:

class _BatchNorm(Module):

    def __init__(self, num_features, eps=1e-5, momentum=0.1, affine=True,
                 track_running_stats=True):
        super(_BatchNorm, self).__init__()
        self.num_features = num_features
        self.eps = eps
        self.momentum = momentum
        self.affine = affine
        self.track_running_stats = track_running_stats
        if self.affine:
            self.weight = Parameter(torch.Tensor(num_features))
            self.bias = Parameter(torch.Tensor(num_features))
        else:
            self.register_parameter('weight', None)
            self.register_parameter('bias', None)
        if self.track_running_stats:
            self.register_buffer('running_mean', torch.zeros(num_features))
            self.register_buffer('running_var', torch.ones(num_features))
        else:
            self.register_parameter('running_mean', None)
            self.register_parameter('running_var', None)
        self.reset_parameters()

    def reset_parameters(self):
        if self.track_running_stats:
            self.running_mean.zero_()
            self.running_var.fill_(1)
        if self.affine:
            self.weight.data.uniform_()
            self.bias.data.zero_()

    def forward(self, input):
        self._check_input_dim(input)

        return F.batch_norm(
            input, self.running_mean, self.running_var, self.weight, self.bias,
            self.training or not self.track_running_stats, self.momentum, self.eps)

一维BN，pytorch源码注释：

class BatchNorm1d(_BatchNorm):
    r"""Applies Batch Normalization over a 2D or 3D input (a mini-batch of 1D
    inputs with optional additional channel dimension) as described in the paper
    `Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift`_ .

    .. math::

        y = \frac{x - \mathrm{E}[x]}{\sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta

    The mean and standard-deviation are calculated per-dimension over
    the mini-batches and :math:`\gamma` and :math:`\beta` are learnable parameter vectors
    of size `C` (where `C` is the input size).

    By default, during training this layer keeps running estimates of its
    computed mean and variance, which are then used for normalization during
    evaluation. The running estimates are kept with a default :attr:`momentum`
    of 0.1.

    If :attr:`track_running_stats` is set to ``False``, this layer then does not
    keep running estimates, and batch statistics are instead used during
    evaluation time as well.

    .. note::
        This :attr:`momentum` argument is different from one used in optimizer
        classes and the conventional notion of momentum. Mathematically, the
        update rule for running statistics here is
        :math:`\hat{x}_\text{new} = (1 - \text{momentum}) \times \hat{x} + \text{momemtum} \times x_t`,
        where :math:`\hat{x}` is the estimated statistic and :math:`x_t` is the
        new observed value.

    Because the Batch Normalization is done over the `C` dimension, computing statistics
    on `(N, L)` slices, it's common terminology to call this Temporal Batch Normalization.

    Args:
        num_features: :math:`C` from an expected input of size
            :math:`(N, C, L)` or :math:`L` from input of size :math:`(N, L)`
        eps: a value added to the denominator for numerical stability.
            Default: 1e-5
        momentum: the value used for the running_mean and running_var
            computation. Default: 0.1
        affine: a boolean value that when set to ``True``, this module has
            learnable affine parameters. Default: ``True``
        track_running_stats: a boolean value that when set to ``True``, this
            module tracks the running mean and variance, and when set to ``False``,
            this module does not track such statistics and always uses batch
            statistics in both training and eval modes. Default: ``True``

    Shape:
        - Input: :math:`(N, C)` or :math:`(N, C, L)`
        - Output: :math:`(N, C)` or :math:`(N, C, L)` (same shape as input)

    Examples::

        >>> # With Learnable Parameters
        >>> m = nn.BatchNorm1d(100)
        >>> # Without Learnable Parameters
        >>> m = nn.BatchNorm1d(100, affine=False)
        >>> input = torch.randn(20, 100)
        >>> output = m(input)

    .. _`Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift`:
        https://arxiv.org/abs/1502.03167
    """

    def _check_input_dim(self, input):
        if input.dim() != 2 and input.dim() != 3:
            raise ValueError('expected 2D or 3D input (got {}D input)'
                             .format(input.dim()))

参数解释:

    If :attr:`track_running_stats` is set to ``False``, this layer then does not
    keep running estimates, and batch statistics are instead used during
    evaluation time as well.

1 . 属性trac_running_stats默认是True,即使用训练时的均值和方差的指数加权平均来估计真实的均值与方差。

若设置为False，验证时，使用每个batch计算出的均值和方差。

若单个样本Predict时，该如何处理？

  :attr:`momentum`: By default, during training this layer keeps running estimates of its
    computed mean and variance, which are then used for normalization during
    evaluation. The running estimates are kept with a default
    of 0.1

2. 属性momentum。在训练时，一般使用指数加权平均来估算正是样本的均值和方差。若momentum取0.1，真实的均值和方差的估计值 约等于 最后10个batch的均值和方差的平均。

32. 属性affine。默认是True,即使用BN时加入可训练的参数gamma、beta。

一维举例:

batch20， 每个样本时100维，则此时共有100对 需要学习的gamma 和 beta

>>> # With Learnable Parameters
>>> m = nn.BatchNorm1d(100)
>>> # Without Learnable Parameters
>>> m = nn.BatchNorm1d(100, affine=False)
>>> input = torch.randn(20, 100)
>>> output = m(input)

二维BN源码注释

class BatchNorm2d(_BatchNorm):
    r"""Applies Batch Normalization over a 4D input (a mini-batch of 2D inputs
    with additional channel dimension) as described in the paper
    `Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift`_ .

    .. math::

        y = \frac{x - \mathrm{E}[x]}{ \sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta

    The mean and standard-deviation are calculated per-dimension over
    the mini-batches and :math:`\gamma` and :math:`\beta` are learnable parameter vectors
    of size `C` (where `C` is the input size).

    By default, during training this layer keeps running estimates of its
    computed mean and variance, which are then used for normalization during
    evaluation. The running estimates are kept with a default :attr:`momentum`
    of 0.1.

    If :attr:`track_running_stats` is set to ``False``, this layer then does not
    keep running estimates, and batch statistics are instead used during
    evaluation time as well.

    .. note::
        This :attr:`momentum` argument is different from one used in optimizer
        classes and the conventional notion of momentum. Mathematically, the
        update rule for running statistics here is
        :math:`\hat{x}_\text{new} = (1 - \text{momentum}) \times \hat{x} + \text{momemtum} \times x_t`,
        where :math:`\hat{x}` is the estimated statistic and :math:`x_t` is the
        new observed value.

    Because the Batch Normalization is done over the `C` dimension, computing statistics
    on `(N, H, W)` slices, it's common terminology to call this Spatial Batch Normalization.

    Args:
        num_features: :math:`C` from an expected input of size
            :math:`(N, C, H, W)`
        eps: a value added to the denominator for numerical stability.
            Default: 1e-5
        momentum: the value used for the running_mean and running_var
            computation. Default: 0.1
        affine: a boolean value that when set to ``True``, this module has
            learnable affine parameters. Default: ``True``
        track_running_stats: a boolean value that when set to ``True``, this
            module tracks the running mean and variance, and when set to ``False``,
            this module does not track such statistics and always uses batch
            statistics in both training and eval modes. Default: ``True``

    Shape:
        - Input: :math:`(N, C, H, W)`
        - Output: :math:`(N, C, H, W)` (same shape as input)

    Examples::

        >>> # With Learnable Parameters
        >>> m = nn.BatchNorm2d(100)
        >>> # Without Learnable Parameters
        >>> m = nn.BatchNorm2d(100, affine=False)
        >>> input = torch.randn(20, 100, 35, 45)
        >>> output = m(input)

    .. _`Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift`:
        https://arxiv.org/abs/1502.03167
    """

    def _check_input_dim(self, input):
        if input.dim() != 4:
            raise ValueError('expected 4D input (got {}D input)'
                             .format(input.dim()))

从2维BN的例子中可以看出，若输入的batch是20， 每张图图片的feature map是100， 每个feature map的维度是35×35.则：为节省参数（正常情况下应该是100×35×35对gamma、beta），二维BN共有100对gamma、beta。

还有3维BN，但是我不知道具体应用场景是什么，我能猜可能是时间序列图片，比如给定一个时间序列的图片，描述这一时间段图片里的内容。

class BatchNorm3d(_BatchNorm):
    r"""Applies Batch Normalization over a 5D input (a mini-batch of 3D inputs
    with additional channel dimension) as described in the paper
    `Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift`_ .

    .. math::

        y = \frac{x - \mathrm{E}[x]}{ \sqrt{\mathrm{Var}[x] + \epsilon}} * \gamma + \beta

    The mean and standard-deviation are calculated per-dimension over
    the mini-batches and :math:`\gamma` and :math:`\beta` are learnable parameter vectors
    of size `C` (where `C` is the input size).

    By default, during training this layer keeps running estimates of its
    computed mean and variance, which are then used for normalization during
    evaluation. The running estimates are kept with a default :attr:`momentum`
    of 0.1.

    If :attr:`track_running_stats` is set to ``False``, this layer then does not
    keep running estimates, and batch statistics are instead used during
    evaluation time as well.

    .. note::
        This :attr:`momentum` argument is different from one used in optimizer
        classes and the conventional notion of momentum. Mathematically, the
        update rule for running statistics here is
        :math:`\hat{x}_\text{new} = (1 - \text{momentum}) \times \hat{x} + \text{momemtum} \times x_t`,
        where :math:`\hat{x}` is the estimated statistic and :math:`x_t` is the
        new observed value.

    Because the Batch Normalization is done over the `C` dimension, computing statistics
    on `(N, D, H, W)` slices, it's common terminology to call this Volumetric Batch Normalization
    or Spatio-temporal Batch Normalization.

    Args:
        num_features: :math:`C` from an expected input of size
            :math:`(N, C, D, H, W)`
        eps: a value added to the denominator for numerical stability.
            Default: 1e-5
        momentum: the value used for the running_mean and running_var
            computation. Default: 0.1
        affine: a boolean value that when set to ``True``, this module has
            learnable affine parameters. Default: ``True``
        track_running_stats: a boolean value that when set to ``True``, this
            module tracks the running mean and variance, and when set to ``False``,
            this module does not track such statistics and always uses batch
            statistics in both training and eval modes. Default: ``True``

    Shape:
        - Input: :math:`(N, C, D, H, W)`
        - Output: :math:`(N, C, D, H, W)` (same shape as input)

    Examples::

        >>> # With Learnable Parameters
        >>> m = nn.BatchNorm3d(100)
        >>> # Without Learnable Parameters
        >>> m = nn.BatchNorm3d(100, affine=False)
        >>> input = torch.randn(20, 100, 35, 45, 10)
        >>> output = m(input)

    .. _`Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift`:
        https://arxiv.org/abs/1502.03167
    """

    def _check_input_dim(self, input):
        if input.dim() != 5:
            raise ValueError('expected 5D input (got {}D input)'
                             .format(input.dim()))