目录
前言
在学习transformer时，遇到过非常频繁的nn.Linear()函数，这里对nn.Linear进行一个详解。
参考：https://pytorch.org/docs/stable/_modules/torch/nn/modules/linear.html
1. nn.Linear的原理:
从名称就可以看出来，nn.Linear表示的是线性变换，原型就是初级数学里学到的线性函数：y=kx+b
不过在深度学习中，变量都是多维张量，乘法就是矩阵乘法，加法就是矩阵加法，因此nn.Linear()运行的真正的计算就是：
output = weight @ input + bias
@: 在python中代表矩阵乘法
input: 表示输入的Tensor，可以有多个维度
weights: 表示可学习的权重，shape=(output_feature,in_feature)
bias: 表示科学习的偏置，shape=(output_feature)
in_feature： nn.Linear 初始化的第一个参数，即输入Tensor最后一维的通道数
out_feature： nn.Linear 初始化的第二个参数，即返回Tensor最后一维的通道数
output: 表示输入的Tensor，可以有多个维度
2. nn.Linear的使用:
常用头文件：import torch.nn as nn
nn.Linear()的初始化：
nn.Linear(in_feature,out_feature,bias)
in_feature： int型, 在forward中输入Tensor最后一维的通道数
out_feature： int型, 在forward中输出Tensor最后一维的通道数
bias: bool型， Linear线性变换中是否添加bias偏置
nn.Linear()的执行：（即执行forward函数）
out=nn.Linear(input)
input: 表示输入的Tensor，可以有多个维度
output: 表示输入的Tensor，可以有多个维度
举例：
2维 Tensor
m = nn.Linear(20, 40) input = torch.randn(128, 20) output = m(input) print(output.size()) # [(128，40])
4维 Tensor:
m = nn.Linear(128, 64) input = torch.randn(512, 3,128,128) output = m(input) print(output.size()) # [(512, 3,128,64))
3. nn.Linear的源码定义:
import math
import torch
import torch.nn as nn
from torch import Tensor
from torch.nn.parameter import Parameter, UninitializedParameter
from torch.nn import functional as F
from torch.nn import init
# from .lazy import LazyModuleMixin
class myLinear(nn.Module):
r"""Applies a linear transformation to the incoming data: :math:`y = xA^T + b`
This module supports :ref:`TensorFloat32<tf32_on_ampere>`.
Args:
in_features: size of each input sample
out_features: size of each output sample
bias: If set to ``False``, the layer will not learn an additive bias.
Default: ``True``
Shape:
- Input: :math:`(*, H_{in})` where :math:`*` means any number of
dimensions including none and :math:`H_{in} = \text{in\_features}`.
- Output: :math:`(*, H_{out})` where all but the last dimension
are the same shape as the input and :math:`H_{out} = \text{out\_features}`.
Attributes:
weight: the learnable weights of the module of shape
:math:`(\text{out\_features}, \text{in\_features})`. The values are
initialized from :math:`\mathcal{U}(-\sqrt{k}, \sqrt{k})`, where
:math:`k = \frac{1}{\text{in\_features}}`
bias: the learnable bias of the module of shape :math:`(\text{out\_features})`.
If :attr:`bias` is ``True``, the values are initialized from
:math:`\mathcal{U}(-\sqrt{k}, \sqrt{k})` where
:math:`k = \frac{1}{\text{in\_features}}`
Examples::
>>> m = nn.Linear(20, 30)
>>> input = torch.randn(128, 20)
>>> output = m(input)
>>> print(output.size())
torch.Size([128, 30])
"""
__constants__ = ['in_features', 'out_features']
in_features: int
out_features: int
weight: Tensor
def __init__(self, in_features: int, out_features: int, bias: bool = True,
device=None, dtype=None) -> None:
factory_kwargs = {'device': device, 'dtype': dtype}
super(myLinear, self).__init__()
self.in_features = in_features
self.out_features = out_features
self.weight = Parameter(torch.empty((out_features, in_features), **factory_kwargs))
if bias:
self.bias = Parameter(torch.empty(out_features, **factory_kwargs))
else:
self.register_parameter('bias', None)
self.reset_parameters()
def reset_parameters(self) -> None:
# Setting a=sqrt(5) in kaiming_uniform is the same as initializing with
# uniform(-1/sqrt(in_features), 1/sqrt(in_features)). For details, see
# https://github.com/pytorch/pytorch/issues/57109
print("333")
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) if fan_in > 0 else 0
init.uniform_(self.bias, -bound, bound)
def forward(self, input: Tensor) -> Tensor:
print("111")
print("self.weight.shape=(", )
return F.linear(input, self.weight, self.bias)
def extra_repr(self) -> str:
print("www")
return 'in_features={}, out_features={}, bias={}'.format(
self.in_features, self.out_features, self.bias is not None
)
# m = myLinear(20, 40)
# input = torch.randn(128, 40, 20)
# output = m(input)
# print(output.size())
m = myLinear(128, 64)
input = torch.randn(512, 3,128,128)
output = m(input)
print(output.size()) # [(512, 3,128,64))
4. nn.Linear的官方源码:
import math
import torch
from torch import Tensor
from torch.nn.parameter import Parameter, UninitializedParameter
from .. import functional as F
from .. import init
from .module import Module
from .lazy import LazyModuleMixin
class Identity(Module):
r"""A placeholder identity operator that is argument-insensitive.
Args:
args: any argument (unused)
kwargs: any keyword argument (unused)
Shape:
- Input: :math:`(*)`, where :math:`*` means any number of dimensions.
- Output: :math:`(*)`, same shape as the input.
Examples::
>>> m = nn.Identity(54, unused_argument1=0.1, unused_argument2=False)
>>> input = torch.randn(128, 20)
>>> output = m(input)
>>> print(output.size())
torch.Size([128, 20])
"""
def __init__(self, *args, **kwargs):
super(Identity, self).__init__()
def forward(self, input: Tensor) -> Tensor:
return input
class Linear(Module):
r"""Applies a linear transformation to the incoming data: :math:`y = xA^T + b`
This module supports :ref:`TensorFloat32<tf32_on_ampere>`.
Args:
in_features: size of each input sample
out_features: size of each output sample
bias: If set to ``False``, the layer will not learn an additive bias.
Default: ``True``
Shape:
- Input: :math:`(*, H_{in})` where :math:`*` means any number of
dimensions including none and :math:`H_{in} = \text{in\_features}`.
- Output: :math:`(*, H_{out})` where all but the last dimension
are the same shape as the input and :math:`H_{out} = \text{out\_features}`.
Attributes:
weight: the learnable weights of the module of shape
:math:`(\text{out\_features}, \text{in\_features})`. The values are
initialized from :math:`\mathcal{U}(-\sqrt{k}, \sqrt{k})`, where
:math:`k = \frac{1}{\text{in\_features}}`
bias: the learnable bias of the module of shape :math:`(\text{out\_features})`.
If :attr:`bias` is ``True``, the values are initialized from
:math:`\mathcal{U}(-\sqrt{k}, \sqrt{k})` where
:math:`k = \frac{1}{\text{in\_features}}`
Examples::
>>> m = nn.Linear(20, 30)
>>> input = torch.randn(128, 20)
>>> output = m(input)
>>> print(output.size())
torch.Size([128, 30])
"""
__constants__ = ['in_features', 'out_features']
in_features: int
out_features: int
weight: Tensor
def __init__(self, in_features: int, out_features: int, bias: bool = True,
device=None, dtype=None) -> None:
factory_kwargs = {'device': device, 'dtype': dtype}
super(Linear, self).__init__()
self.in_features = in_features
self.out_features = out_features
self.weight = Parameter(torch.empty((out_features, in_features), **factory_kwargs))
if bias:
self.bias = Parameter(torch.empty(out_features, **factory_kwargs))
else:
self.register_parameter('bias', None)
self.reset_parameters()
def reset_parameters(self) -> None:
# Setting a=sqrt(5) in kaiming_uniform is the same as initializing with
# uniform(-1/sqrt(in_features), 1/sqrt(in_features)). For details, see
# https://github.com/pytorch/pytorch/issues/57109
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) if fan_in > 0 else 0
init.uniform_(self.bias, -bound, bound)
def forward(self, input: Tensor) -> Tensor:
return F.linear(input, self.weight, self.bias)
def extra_repr(self) -> str:
return 'in_features={}, out_features={}, bias={}'.format(
self.in_features, self.out_features, self.bias is not None
)
# This class exists solely to avoid triggering an obscure error when scripting
# an improperly quantized attention layer. See this issue for details:
# https://github.com/pytorch/pytorch/issues/58969
# TODO: fail fast on quantization API usage error, then remove this class
# and replace uses of it with plain Linear
class NonDynamicallyQuantizableLinear(Linear):
def __init__(self, in_features: int, out_features: int, bias: bool = True,
device=None, dtype=None) -> None:
super().__init__(in_features, out_features, bias=bias,
device=device, dtype=dtype)
[docs]class Bilinear(Module):
r"""Applies a bilinear transformation to the incoming data:
:math:`y = x_1^T A x_2 + b`
Args:
in1_features: size of each first input sample
in2_features: size of each second input sample
out_features: size of each output sample
bias: If set to False, the layer will not learn an additive bias.
Default: ``True``
Shape:
- Input1: :math:`(*, H_{in1})` where :math:`H_{in1}=\text{in1\_features}` and
:math:`*` means any number of additional dimensions including none. All but the last dimension
of the inputs should be the same.
- Input2: :math:`(*, H_{in2})` where :math:`H_{in2}=\text{in2\_features}`.
- Output: :math:`(*, H_{out})` where :math:`H_{out}=\text{out\_features}`
and all but the last dimension are the same shape as the input.
Attributes:
weight: the learnable weights of the module of shape
:math:`(\text{out\_features}, \text{in1\_features}, \text{in2\_features})`.
The values are initialized from :math:`\mathcal{U}(-\sqrt{k}, \sqrt{k})`, where
:math:`k = \frac{1}{\text{in1\_features}}`
bias: the learnable bias of the module of shape :math:`(\text{out\_features})`.
If :attr:`bias` is ``True``, the values are initialized from
:math:`\mathcal{U}(-\sqrt{k}, \sqrt{k})`, where
:math:`k = \frac{1}{\text{in1\_features}}`
Examples::
>>> m = nn.Bilinear(20, 30, 40)
>>> input1 = torch.randn(128, 20)
>>> input2 = torch.randn(128, 30)
>>> output = m(input1, input2)
>>> print(output.size())
torch.Size([128, 40])
"""
__constants__ = ['in1_features', 'in2_features', 'out_features']
in1_features: int
in2_features: int
out_features: int
weight: Tensor
def __init__(self, in1_features: int, in2_features: int, out_features: int, bias: bool = True,
device=None, dtype=None) -> None:
factory_kwargs = {'device': device, 'dtype': dtype}
super(Bilinear, self).__init__()
self.in1_features = in1_features
self.in2_features = in2_features
self.out_features = out_features
self.weight = Parameter(torch.empty((out_features, in1_features, in2_features), **factory_kwargs))
if bias:
self.bias = Parameter(torch.empty(out_features, **factory_kwargs))
else:
self.register_parameter('bias', None)
self.reset_parameters()
def reset_parameters(self) -> None:
bound = 1 / math.sqrt(self.weight.size(1))
init.uniform_(self.weight, -bound, bound)
if self.bias is not None:
init.uniform_(self.bias, -bound, bound)
def forward(self, input1: Tensor, input2: Tensor) -> Tensor:
return F.bilinear(input1, input2, self.weight, self.bias)
def extra_repr(self) -> str:
return 'in1_features={}, in2_features={}, out_features={}, bias={}'.format(
self.in1_features, self.in2_features, self.out_features, self.bias is not None
)
class LazyLinear(LazyModuleMixin, Linear):
r"""A :class:`torch.nn.Linear` module where `in_features` is inferred.
In this module, the `weight` and `bias` are of :class:`torch.nn.UninitializedParameter`
class. They will be initialized after the first call to ``forward`` is done and the
module will become a regular :class:`torch.nn.Linear` module. The ``in_features`` argument
of the :class:`Linear` is inferred from the ``input.shape[-1]``.
Check the :class:`torch.nn.modules.lazy.LazyModuleMixin` for further documentation
on lazy modules and their limitations.
Args:
out_features: size of each output sample
bias: If set to ``False``, the layer will not learn an additive bias.
Default: ``True``
Attributes:
weight: the learnable weights of the module of shape
:math:`(\text{out\_features}, \text{in\_features})`. The values are
initialized from :math:`\mathcal{U}(-\sqrt{k}, \sqrt{k})`, where
:math:`k = \frac{1}{\text{in\_features}}`
bias: the learnable bias of the module of shape :math:`(\text{out\_features})`.
If :attr:`bias` is ``True``, the values are initialized from
:math:`\mathcal{U}(-\sqrt{k}, \sqrt{k})` where
:math:`k = \frac{1}{\text{in\_features}}`
"""
cls_to_become = Linear # type: ignore[assignment]
weight: UninitializedParameter
bias: UninitializedParameter # type: ignore[assignment]
def __init__(self, out_features: int, bias: bool = True,
device=None, dtype=None) -> None:
factory_kwargs = {'device': device, 'dtype': dtype}
# bias is hardcoded to False to avoid creating tensor
# that will soon be overwritten.
super().__init__(0, 0, False)
self.weight = UninitializedParameter(**factory_kwargs)
self.out_features = out_features
if bias:
self.bias = UninitializedParameter(**factory_kwargs)
def reset_parameters(self) -> None:
if not self.has_uninitialized_params() and self.in_features != 0:
super().reset_parameters()
def initialize_parameters(self, input) -> None: # type: ignore[override]
if self.has_uninitialized_params():
with torch.no_grad():
self.in_features = input.shape[-1]
self.weight.materialize((self.out_features, self.in_features))
if self.bias is not None:
self.bias.materialize((self.out_features,))
self.reset_parameters()
# TODO: PartialLinear - maybe in sparse?
补充：许多细节需要声明
1）nn.Linear是一个类，使用时进行类的实例化
2）实例化的时候，nn.Linear需要输入两个参数，in_features为上一层神经元的个数，out_features为这一层的神经元个数
3）不需要定义w和b。所有nn.Module的子类，形如nn.XXX的层，都会在实例化的同时随机生成w和b的初始值。所以实例化之后，我们就可以调用属性weight和bias来查看生成的w和b。其中w是必然会生成的，b是我们可以控制是否要生成的。在nn.Linear类中，有参数bias，默认 bias = True。如果我们希望不拟合常量b，在实例化时将参数bias设置为False即可。
4）由于w和b是随机生成的，所以同样的代码多次运行后的结果是不一致的。如果我们希望控制随机性，则可以使用torch中的random类。如：torch.random.manual_seed(420) #人为设置随机数种子
5）由于不需要定义常量b，因此在特征张量中，不需要留出与常数项相乘的那一列，只需要输入特征张量。
6）输入层只有一层，并且输入层的结构（神经元的个数）由输入的特征张量X决定，因此在PyTorch中构筑神经网络时，不需要定义输入层。
7）实例化之后，将特征张量输入到实例化后的类中。
