PyTorch卷积网络如何动态设置全连接层in_features？

在调整CNN超参数（如卷积核大小、通道数、dropout率）时，硬编码全连接层的in_features会非常繁琐，以下是几种无需硬编码的解决方案：

方法1：提前计算输入维度 #

根据输入序列长度、卷积和池化的参数，手动计算经过所有卷积池化操作后的扁平化维度，直接在初始化阶段确定全连接层的输入特征数。

import torch

class Conv_v1(torch.nn.Module):
    def __init__(self, input_seq_len, in_channels=4, out_channels1=3, kernel_size1=17, pool_kernel_size=5, dropout_rate=0.5, num_classes=2):
        super(Conv_v1, self).__init__()
        self.conv1 = torch.nn.Conv1d(in_channels=in_channels, out_channels=out_channels1, kernel_size=kernel_size1)
        self.activation = torch.nn.ReLU()
        self.maxpool = torch.nn.MaxPool1d(kernel_size=pool_kernel_size)
        self.dropout = torch.nn.Dropout(p=dropout_rate)
        
        # 计算卷积池化后的扁平化维度
        conv_out_len = input_seq_len - kernel_size1 + 1
        pool_out_len = conv_out_len // pool_kernel_size  # 若池化后有剩余可根据需求调整计算方式
        fc_in_features = out_channels1 * pool_out_len
        
        self.fc = torch.nn.Linear(in_features=fc_in_features, out_features=num_classes)

    def forward(self, x):
        x = x.permute(0, 2, 1)
        x = self.conv1(x)
        x = self.activation(x)
        x = self.maxpool(x)
        x = x.view(x.size(0), -1)
        x = self.dropout(x)
        x = self.fc(x)
        return x

使用时只需传入输入序列的长度，超参数调整时自动适配全连接层的输入维度。

方法2：使用PyTorch的LazyLinear模块 #

PyTorch 1.8.0及以上版本提供了LazyLinear模块，无需提前指定in_features，第一次前向传播时会自动根据输入维度初始化权重，彻底省去手动计算的步骤。

import torch

class Conv_v2(torch.nn.Module):
    def __init__(self, out_channels1=3, kernel_size1=17, pool_kernel_size=5, dropout_rate=0.5, num_classes=2):
        super(Conv_v2, self).__init__()
        self.conv1 = torch.nn.Conv1d(in_channels=4, out_channels=out_channels1, kernel_size=kernel_size1)
        self.activation = torch.nn.ReLU()
        self.maxpool = torch.nn.MaxPool1d(kernel_size=pool_kernel_size)
        self.dropout = torch.nn.Dropout(p=dropout_rate)
        
        # 使用LazyLinear，无需指定in_features
        self.fc = torch.nn.LazyLinear(out_features=num_classes)

    def forward(self, x):
        x = x.permute(0, 2, 1)
        x = self.conv1(x)
        x = self.activation(x)
        x = self.maxpool(x)
        x = x.view(x.size(0), -1)
        x = self.dropout(x)
        x = self.fc(x)
        return x

该模块会在第一次前向传播后固定权重维度，后续不会重新初始化，保证模型正常学习。

方法3：第一次前向传播时动态初始化全连接层 #

如果不想依赖Lazy模块，可以在第一次前向传播时计算输入维度，创建全连接层并将其参数注册到模型中，确保仅初始化一次。

import torch

class Conv_v3(torch.nn.Module):
    def __init__(self, out_channels1, kernel_size1, out_channels2, kernel_size2, dropout_rate, num_classes=2):
        super(Conv_v3, self).__init__()
        self.conv1 = torch.nn.Conv1d(in_channels=4, out_channels=out_channels1, kernel_size=kernel_size1)
        self.conv2 = torch.nn.Conv1d(in_channels=out_channels1, out_channels=out_channels2, kernel_size=kernel_size2)
        self.dropout = torch.nn.Dropout(p=dropout_rate)
        self.fc = None  # 先初始化为None
        self.num_classes = num_classes

    def forward(self, x):
        x = x.permute(0, 2, 1)
        x = torch.nn.functional.max_pool1d(torch.tanh(self.conv1(x)), 2)
        x = torch.nn.functional.max_pool1d(torch.tanh(self.conv2(x)), 2)
        x = x.view(x.size(0), -1)
        x = self.dropout(x)
        
        # 第一次前向时初始化fc层并注册参数
        if self.fc is None:
            self.fc = torch.nn.Linear(in_features=x.size(1), out_features=self.num_classes)
            self.add_module('fc', self.fc)
        
        out = self.fc(x)
        return out

通过add_module将全连接层注册到模型后，优化器就能正常更新其权重，避免了重复初始化的问题。

内容的提问来源于stack exchange，提问作者Marina Thalassini