pytorch关于卷积操作的初始化方式(kaiming_uniform_详解)
作者：两只蜡笔的小新
这篇文章主要介绍了pytorch关于卷积操作的初始化方式(kaiming_uniform_详解),具有很好的参考价值,希望对大家有所帮助,如有错误或未考虑完全的地方,望不吝赐教
摘要：
最近写了一篇文章，reviewers给了几个意见，其中之一就是：不同配置下的网络初始化条件是否相同，是怎样初始化的？
之前竟然没有关注过这个问题，应该是torch默认情况下会初始化卷积核参数，这里详细讲解一下torch卷积操作的初始化过程。
1. pytorch中的卷积运算分类
在pycharm的IDE中，按住ctrl+鼠标点击torch.nn.Conv2d可以进入torch的内部卷积运算的源码(conv.py)
搭建网络经常使用到的模块
如下图所示：
class _ConvNd(Module): class Conv1d(_ConvNd): class Conv2d(_ConvNd): class Conv3d(_ConvNd): class _ConvTransposeNd(_ConvNd): class ConvTranspose1d(_ConvTransposeNd): class ConvTranspose2d(_ConvTransposeNd): class ConvTranspose3d(_ConvTransposeNd):
可以看到：常用的卷积的父类均是
class _ConvNd(Module):
并且点开 class Conv2d(_ConvNd): 并没有发现参数初始化的具体方法，
如下图所示
所以猜想卷积初始化参数的方法应该在父类 _ConvNd(Module):
2. pytorch中的卷积操作的父类
下面是父类 _ConvNd 的源码，其中初始化参数的 方法是
def reset_parameters(self) -> None:
class _ConvNd(Module):
__constants__ = ['stride', 'padding', 'dilation', 'groups',
'padding_mode', 'output_padding', 'in_channels',
'out_channels', 'kernel_size']
__annotations__ = {'bias': Optional[torch.Tensor]}
def _conv_forward(self, input: Tensor, weight: Tensor, bias: Optional[Tensor]) -> Tensor:
...
_in_channels: int
out_channels: int
kernel_size: Tuple[int, ...]
stride: Tuple[int, ...]
padding: Tuple[int, ...]
dilation: Tuple[int, ...]
transposed: bool
output_padding: Tuple[int, ...]
groups: int
padding_mode: str
weight: Tensor
bias: Optional[Tensor]
def __init__(self,
in_channels: int,
out_channels: int,
kernel_size: Tuple[int, ...],
stride: Tuple[int, ...],
padding: Tuple[int, ...],
dilation: Tuple[int, ...],
transposed: bool,
output_padding: Tuple[int, ...],
groups: int,
bias: bool,
padding_mode: str) -> None:
super(_ConvNd, self).__init__()
if in_channels % groups != 0:
raise ValueError('in_channels must be divisible by groups')
if out_channels % groups != 0:
raise ValueError('out_channels must be divisible by groups')
valid_padding_modes = {'zeros', 'reflect', 'replicate', 'circular'}
if padding_mode not in valid_padding_modes:
raise ValueError("padding_mode must be one of {}, but got padding_mode='{}'".format(
valid_padding_modes, padding_mode))
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.padding = padding
self.dilation = dilation
self.transposed = transposed
self.output_padding = output_padding
self.groups = groups
self.padding_mode = padding_mode
# `_reversed_padding_repeated_twice` is the padding to be passed to
# `F.pad` if needed (e.g., for non-zero padding types that are
# implemented as two ops: padding + conv). `F.pad` accepts paddings in
# reverse order than the dimension.
self._reversed_padding_repeated_twice = _reverse_repeat_tuple(self.padding, 2)
if transposed:
self.weight = Parameter(torch.Tensor(
in_channels, out_channels // groups, *kernel_size))
else:
self.weight = Parameter(torch.Tensor(
out_channels, in_channels // groups, *kernel_size))
if bias:
self.bias = Parameter(torch.Tensor(out_channels))
else:
self.register_parameter('bias', None)
self.reset_parameters()
def reset_parameters(self) -> None:
init.kaiming_uniform_(self.weight, a=math.sqrt(5))
if self.bias is not None:
fan_in, _ = init._calculate_fan_in_and_fan_out(self.weight)
bound = 1 / math.sqrt(fan_in)
init.uniform_(self.bias, -bound, bound)
def extra_repr(self):
s = ('{in_channels}, {out_channels}, kernel_size={kernel_size}'
', stride={stride}')
if self.padding != (0,) * len(self.padding):
s += ', padding={padding}'
if self.dilation != (1,) * len(self.dilation):
s += ', dilation={dilation}'
if self.output_padding != (0,) * len(self.output_padding):
s += ', output_padding={output_padding}'
if self.groups != 1:
s += ', groups={groups}'
if self.bias is None:
s += ', bias=False'
if self.padding_mode != 'zeros':
s += ', padding_mode={padding_mode}'
return s.format(**self.__dict__)
def __setstate__(self, state):
super(_ConvNd, self).__setstate__(state)
if not hasattr(self, 'padding_mode'):
self.padding_mode = 'zeros'3. def reset_parameters(self) -> None
卷积操作的默认的初始化方式：
def reset_parameters(self) -> None:
init.kaiming_uniform_(self.weight, a=math.sqrt(5))
if self.bias is not None:
fan_in, _ = init._calculate_fan_in_and_fan_out(self.weight)
bound = 1 / math.sqrt(fan_in)
init.uniform_(self.bias, -bound, bound)该类中的参数的初始化方式是： Kaiming 初始化
由我国计算机视觉领域专家何凯明提出了针对于relu的初始化方法，pytorch默认使用kaiming正态分布初始化卷积层参数。
Fills the input Tensor with values according to the method described in Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification - He, K. et al. (2015),
using a uniform distribution.
The resulting tensor will have values sampled from U( − bound, bound) where bound = gain × √((3)/( fan_mode))Also known as He initialization.
3.1 卷积核部分的参数初始化：
init.kaiming_uniform_(self.weight, a=math.sqrt(5))
关于init.kaiming_uniform_这个函数，源码如下：
def kaiming_uniform_(tensor, a=0, mode='fan_in', nonlinearity='leaky_relu'):
r"""Fills the input `Tensor` with values according to the method
described in `Delving deep into rectifiers: Surpassing human-level
performance on ImageNet classification` - He, K. et al. (2015), using a
uniform distribution. The resulting tensor will have values sampled from
:math:`\mathcal{U}(-\text{bound}, \text{bound})` where
.. math::
\text{bound} = \text{gain} \times \sqrt{\frac{3}{\text{fan\_mode}}}
Also known as He initialization.
Args:
tensor: an n-dimensional `torch.Tensor`
a: the negative slope of the rectifier used after this layer (only
used with ``'leaky_relu'``)
mode: either ``'fan_in'`` (default) or ``'fan_out'``. Choosing ``'fan_in'``
preserves the magnitude of the variance of the weights in the
forward pass. Choosing ``'fan_out'`` preserves the magnitudes in the
backwards pass.
nonlinearity: the non-linear function (`nn.functional` name),
recommended to use only with ``'relu'`` or ``'leaky_relu'`` (default).
Examples:
>>> w = torch.empty(3, 5)
>>> nn.init.kaiming_uniform_(w, mode='fan_in', nonlinearity='relu')
"""
fan = _calculate_correct_fan(tensor, mode)
gain = calculate_gain(nonlinearity, a)
std = gain / math.sqrt(fan)
bound = math.sqrt(3.0) * std # Calculate uniform bounds from standard deviation
with torch.no_grad():
return tensor.uniform_(-bound, bound)torch中卷积核默认的初始化的详细参数为：
init.kaiming_uniform_(self.weight, a=math.sqrt(5),mode='fan_in', nonlinearity='leaky_relu'))
关于 init.kaiming_uniform_中所使用的其他函数 ，如下不做进一步的分析，不过还是简单介绍一下。
_calculate_correct_fan(tensor, mode) # 用于计算计算当前网络层的fan_in（输入神经元个数）或 fan_out（输出神经元个数的），取决于 mode 的值 'fan_in' 'fan_out' calculate_gain：# 对于给定的非线性函数，返回推荐的增益值，其实就是一个数，从下面图中的列表中选出对应的值
- _calculate_correct_fan：在这里 model = fan_in， 计算 的是 当前网络层的fan_in（输入神经元个数）
- calculate_gain： 在这里 nonlinearity='leaky_relu'，param = a = math.sqrt(5) 得到的值就是：（negative_slope = param = math.sqrt(5)）
gan = math.sqrt(2.0 / (1 + negative_slope ** 2))
前文讲到，
The resulting tensor will have values sampled from U( − bound, bound) where bound = gain × √((3)/( fan_mode))
所以上面的一通计算得到了bound
下面的 uniform_(from=0, to=1) → Tensor， 将tensor用从均匀分布中抽样得到的值填充。
3.2 bias部分的初始化
这里不做详细介绍了，相信认真看了 weights部分的初始化过程，这部分自然会明白。
if self.bias is not None:
fan_in, _ = init._calculate_fan_in_and_fan_out(self.weight)
bound = 1 / math.sqrt(fan_in)
init.uniform_(self.bias, -bound, bound)附加的：
init._calculate_fan_in_and_fan_out(self.weight)
函数来计算当前网络层的fan_in（输入神经元个数）和fan_out（输出神经元个数的）
总结
以上为个人经验，希望能给大家一个参考，也希望大家多多支持脚本之家。