https://pytorch.org/docs/stable/generated/torch.nn.Linear.html?highlight=linear#torch.nn.Linear

功能

对输入nn.Linear - 图1进行线性变换nn.Linear - 图2

源码

  1. [torch/nn/modules/linear.py]
  2. class Linear(Module):
  3.     r"""Applies a linear transformation to the incoming data: :math:`y = xA^T + b`
  5.     Args:
  6.         in_features: size of each input sample
  7.         out_features: size of each output sample
  8.         bias: If set to False, the layer will not learn an additive bias.
  9.             Default: True
  11.     Examples::
  13.         >>> m = nn.Linear(20, 30)
  14.         >>> input = torch.randn(128, 20)
  15.         >>> output = m(input)
  16.         >>> print(output.size())
  17.         torch.Size([128, 30])
  18.     """
  19.     # https://discuss.pytorch.org/t/why-do-we-use-constants-or-final/70331
  20.     __constants__ = ['in_features', 'out_features']
  21.     in_features: int
  22.     out_features: int
  23.     weight: Tensor
  25.     def __init__(self, in_features: int, out_features: int, bias: bool = True) -> None:
  26.         super(Linear, self).__init__()
  27.         self.in_features = in_features
  28.         self.out_features = out_features
  29.         self.weight = Parameter(torch.Tensor(out_features, in_features))
  30.         if bias:
  31.             self.bias = Parameter(torch.Tensor(out_features))
  32.         else:
  33.             self.register_parameter('bias', None)
  34.         self.reset_parameters()
  36.     def reset_parameters(self) -> None:
  37.         # 初始化 Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification
  38.         init.kaiming_uniform_(self.weight, a=math.sqrt(5))
  39.         if self.bias is not None:
  40.             fan_in, _ = init._calculate_fan_in_and_fan_out(self.weight)
  41.             bound = 1 / math.sqrt(fan_in)
  42.             init.uniform_(self.bias, -bound, bound)
  44.     def forward(self, input: Tensor) -> Tensor:
  45.         return F.linear(input, self.weight, self.bias)
  47.     def extra_repr(self) -> str:
  48.         return 'in_features={}, out_features={}, bias={}'.format(
  49.             self.in_features, self.out_features, self.bias is not None
  50.         )