Module

定义 Class

从 nn.Module 中继承搭建神经网络模块。在 __init__() 中完成网络的搭建并在 forward() 中完成运算。

class NeuralNetwork(nn.Module):
    def __init__(self):
        super(NeuralNetwork, self).__init__()
        self.flatten = nn.Flatten()
        self.linear_relu_stack = nn.Sequential(
            nn.Linear(28*28, 512),
            nn.ReLU(),
            nn.Linear(512, 512),
            nn.ReLU(),
            nn.Linear(512, 10),
        )

    def forward(self, x):
        x = self.flatten(x)
        logits = self.linear_relu_stack(x)
        return logits

可视化

可以通过 print(model) 查看搭建的网络结构。上述示例打印结果为：

NeuralNetwork(
  (flatten): Flatten(start_dim=1, end_dim=-1)
  (linear_relu_stack): Sequential(
    (0): Linear(in_features=784, out_features=512, bias=True)
    (1): ReLU()
    (2): Linear(in_features=512, out_features=512, bias=True)
    (3): ReLU()
    (4): Linear(in_features=512, out_features=10, bias=True)
  )
)

组件

容器

API	说明	其它
nn.Module	基类
nn.Sequential	序列容器	nn.Sequential(nn.Linear(1, 2), nn.ReLU())

卷积

API	说明	其它
nn.Conv1d	1D 卷积	nn.Conv1d(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True, padding_mode='zeros', device=None, dtype=None)
nn.Conv2d	2D 卷积	nn.Conv2d(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True, padding_mode='zeros', device=None, dtype=None)
nn.ConvTranspose2d	上采样用，2D Transpose 卷积	nn.ConvTranspose2d(in_channels, out_channels, kernel_size, stride=1, padding=0, output_padding=0, groups=1, bias=True, dilation=1, padding_mode='zeros', device=None, dtype=None)

池化

API	说明	其它
nn.MaxPool1d	1D 最大池化	nn.MaxPool1d(kernel_size, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False)
nn.MaxPool2d	2D 最大池化	nn.MaxPool2d(kernel_size, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False)
nn.AvgPool1d	1D 平均池化	nn.AvgPool1d(kernel_size, stride=None, padding=0, ceil_mode=False, count_include_pad=True
nn.AvgPool2d	2D 平均池化	nn.AvgPool2d(kernel_size, stride=None, padding=0, ceil_mode=False, count_include_pad=True
nn.AdaptiveAvgPool2d	2D 自适应平均池化	nn.AdaptiveAvgPool2d(output_size)

填充

API	说明	其它
nn.ReflectionPad2d	对称填充	nn.ReflectionPad2d(padding)
nn.ReplicationPad2d	复制填充	nn.ReplicationPad2d(padding)
nn.ZeroPad2d	零填充	nn.ZeroPad2d(padding)
nn.ConstantPad2d	常量填充	nn.ConstantPad2d(padding, value)

激活函数

API	说明	其它
nn.ReLU	ReLU	nn.ReLU(inplace-False)
nn.LeakyReLU	LeakyReLU	nn.LeakyReLU(negative_slope=0.01, inplace=False)
nn.Sigmoid	Sigmoid	nn.Sigmoid()
nn.Tanh	Tanh	nn.Tanh()

正则化层

API	说明	其它
nn.BatchNorm1d	Batch Norm 1D	nn.BatchNorm1d(num_features, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True, device=None, dtype=None)
nn.BatchNorm2d	Batch Norm 2D	nn.BatchNorm2d(num_features, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True, device=None, dtype=None)
nn.LayerNorm	Layer Norm	nn.LayerNorm(normalized_shape, eps=1e-05, elementwise_affine=True, device=None, dtype=None)

线性层

API	说明	其它
nn.Linear	线性层	nn.Linear(in_features, out_features, bias=True, device=None, dtype=None)

dropout 层

API	说明	其它
nn.Dropout	以伯努利分布随机丢弃	nn.Dropout(p=0.5, inplace=False)
nn.Dropout1d	随机丢弃整个通道（1D特征图）	nn.Dropout1d(p=0.5, inplace=False)
nn.Dropout2d	随机丢弃整个通道（2D特征图）	nn.Dropout2d(p=0.5, inplace=False)

损失函数

API	说明	其它
nn.MSELoss	MSE，接受输入 $x$ 和目标 $y$，reduction 可选 none, mean, sum	nn.MSELoss(size_average=None, reduce=None, reduction='mean')
nn.CrossEntropyLoss	交叉熵，接受 (minibatch, C) 或 (minibatch, C, d1, ..., dk) ，minibatch 可以是 raw	nn.CrossEntropyLoss(weight=None, size_average=None, ignore_index=- 100, reduce=None, reduction='mean', label_smoothing=0.0)
nn.BCELoss	BCE，比较输入和输出的概率	nn.BCELoss(weight=None, size_average=None, reduce=None, reduction='mean')
nn.BCEWithLogitsLoss	BCE + Sigmoid	nn.BCEWithLogitsLoss(weight=None, size_average=None, reduce=None, reduction='mean', pos_weight=None)

---

最后更新: 2022年9月2日