pytorch中卷积操作的初始化⽅法（kaiming_uniform_详解）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 _ConvTranspoNd(_ConvNd):
class ConvTranspo1d(_ConvTranspoNd):
class ConvTranspo2d(_ConvTranspoNd):
class ConvTranspo3d(_ConvTranspoNd):
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):
考研线代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 ret_parameters(lf) -> 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(lf, input: Tensor, weight: Tensor, bias: Optional[Tensor]) -> Tensor: ...
_in_channels: int
out_channels: int
kernel_size: Tuple[int, ...]
电脑网速慢
stride: Tuple[int, ...]
padding: Tuple[int, ...]
dilation: Tuple[int, ...]
transpod: bool
output_padding: Tuple[int, ...]
groups: int
padding_mode: str
weight: Tensor
bias: Optional[Tensor]
def __init__(lf,
in_channels: int,
in_channels: int,
out_channels: int,
kernel_size: Tuple[int, ...],
stride: Tuple[int, ...],
padding: Tuple[int, ...],
dilation: Tuple[int, ...],
transpod: bool,
output_padding: Tuple[int, ...],
groups: int,
bias: bool,
padding_mode: str) -> None:
super(_ConvNd, lf).__init__()
if in_channels % groups != 0:
rai ValueError('in_channels must be divisible by groups')
让爱飞扬
if out_channels % groups != 0:
rai ValueError('out_channels must be divisible by groups')
valid_padding_modes = {'zeros', 'reflect', 'replicate', 'circular'}
if padding_mode not in valid_padding_modes:
rai ValueError("padding_mode must be one of {}, but got padding_mode='{}'".format( valid_padding_modes, padding_mode))
lf.in_channels = in_channels
lf.out_channels = out_channels
lf.kernel_size = kernel_size
lf.stride = stride
lf.padding = padding
lf.dilation = dilation
lf.output_padding = output_padding
lf.padding_mode = padding_mode
# `_reverd_padding_repeated_twice` is the padding to be pasd 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
# rever order than the dimension.
lf._reverd_padding_repeated_twice = _rever_repeat_tuple(lf.padding, 2)
if transpod:
lf.weight = Parameter(torch.Tensor(
in_channels, out_channels // groups, *kernel_size))
el:
lf.weight = Parameter(torch.Tensor(
out_channels, in_channels // groups, *kernel_size))
if bias:
lf.bias = Parameter(torch.Tensor(out_channels))
el:
<_parameters()
def ret_parameters(lf) -> None:
init.kaiming_uniform_(lf.weight, a=math.sqrt(5))
if lf.bias is not None:
fan_in, _ = init._calculate_fan_in_and_fan_out(lf.weight)
bound = 1 / math.sqrt(fan_in)
init.uniform_(lf.bias, -bound, bound)
def extra_repr(lf):
s = ('{in_channels}, {out_channels}, kernel_size={kernel_size}'
', stride={stride}')
if lf.padding != (0,) * len(lf.padding):
s += ', padding={padding}'
if lf.dilation != (1,) * len(lf.dilation):
s += ', dilation={dilation}'
if lf.output_padding != (0,) * len(lf.output_padding):
s += ', output_padding={output_padding}'
ups != 1:
s += ', groups={groups}'
if lf.bias is None:
if lf.bias is None:
s += ', bias=Fal'
if lf.padding_mode != 'zeros':
s += ', padding_mode={padding_mode}'
return s.format(**lf.__dict__)
def __tstate__(lf, state):
super(_ConvNd, lf).__tstate__(state)
if not hasattr(lf, 'padding_mode'):
lf.padding_mode = 'zeros'
3. def ret_parameters(lf) -> None:
卷积操作的默认的初始化⽅式：小米面
def ret_parameters(lf) -> None:
init.kaiming_uniform_(lf.weight, a=math.sqrt(5))
if lf.bias is not None:
fan_in, _ = init._calculate_fan_in_and_fan_out(lf.weight)
bound = 1 / math.sqrt(fan_in)
init.uniform_(lf.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))
李蔚语
Also known as He initialization.
3.1 卷积核部分的参数初始化：
init.kaiming_uniform_(lf.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:汪国真经典诗文
tensor: an n-dimensional `torch.Tensor`
手柄震动测试a: the negative slope of the rectifier ud after this layer (only
ud with ``'leaky_relu'``)
mode: either ``'fan_in'`` (default) or ``'fan_out'``. Choosing ``'fan_in'``
prerves the magnitude of the variance of the weights in the
forward pass. Choosing ``'fan_out'`` prerves the magnitudes in the
backwards pass.
nonlinearity: the non-linear function (`nn.functional` name),
recommended to u only with ``'relu'`` or ``'leaky_relu'`` (default).
Examples:
>>> w = pty(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
如何赚钱_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_(lf.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_o 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）
