基于论文实现InceptionEEGNet时torch.cat尺寸匹配报错求助

错误核心原因 #

torch.cat拼接时，除通道维度（dim=1）外的时间维度尺寸不匹配：

• x1、x2经过padding='same'的卷积后，时间维度保持为128
• x3通过AvgPool2d((1,2))、x4通过stride=(1,2)将时间维度压缩为64
  两者维度冲突导致拼接失败。同时还有隐藏问题：拼接后通道数为64×4=256，但batchnorm2d_elu_2中的BatchNorm2d设置为64，会引发后续通道不匹配错误。

修复步骤 #

1. 统一Inception分支的时间维度 #

给x1、x2的第二个卷积层添加stride=(1,2)，让它们的时间维度和x3、x4一致（压缩为64）：

self.inception_block_1 = nn.Sequential(
    nn.Conv2d(256,64,(1,1),padding='same'),
    nn.Conv2d(64,64,(1,7),padding='same', stride=(1,2)),  # 添加stride压缩时间维度
)
self.inception_block_2 = nn.Sequential(
    nn.Conv2d(256,64,(1,1),padding='same'),
    nn.Conv2d(64,64,(1,9),padding='same', stride=(1,2)),  # 添加stride压缩时间维度
)

2. 修正BatchNorm的通道数 #

将batchnorm2d_elu_2中的BatchNorm2d改为256，匹配拼接后的总通道数：

self.batchnorm2d_elu_2 = nn.Sequential(
    nn.BatchNorm2d(256),  # 从64改为256
    nn.ELU(),
    nn.Dropout2d(p=0.5)
)

3. 优化自适应池化层（可选） #

原代码用F.adaptive_avg_pool3d处理4D张量不合理，改为F.adaptive_avg_pool2d更贴合输入维度：

x = F.adaptive_avg_pool2d(x, (1,3))

完整修复代码 #

import torch
from torch import nn
import torch.nn.functional as F

class DepthwiseConv2d(torch.nn.Conv2d):
    def __init__(self,
                 in_channels,
                 depth_multiplier=1,
                 kernel_size=3,
                 stride=1,
                 padding=0,
                 dilation=1,
                 bias=True,
                 padding_mode='zeros'
                 ):
        out_channels = in_channels * depth_multiplier
        super().__init__(
            in_channels=in_channels,
            out_channels=out_channels,
            kernel_size=kernel_size,
            stride=stride,
            padding=padding,
            dilation=dilation,
            groups=in_channels,
            bias=bias,
            padding_mode=padding_mode
        )

class InceptionEEGNet(nn.Module):
    def __init__(self, bathsize): # input = (1,22,256)
        super().__init__()
        self.bathsize = bathsize
        self.conv2d_1 = nn.Conv2d(1,64,(1,16),padding='same')
        self.batchnorm2d_1 = nn.BatchNorm2d(64)
        self.conv2d_2 = DepthwiseConv2d(64,depth_multiplier=4,kernel_size=(22,1),padding='valid')
        self.batchnorm2d_elu_1 = nn.Sequential(
            nn.BatchNorm2d(256),
            nn.ELU()
        )
        self.averagepooling = nn.Sequential(
            nn.AvgPool2d((1,2)),
            nn.Dropout2d(p=0.5)
        )
        self.inception_block_1 = nn.Sequential(
            nn.Conv2d(256,64,(1,1),padding='same'),
            nn.Conv2d(64,64,(1,7),padding='same', stride=(1,2)),
        )
        self.inception_block_2 = nn.Sequential(
            nn.Conv2d(256,64,(1,1),padding='same'),
            nn.Conv2d(64,64,(1,9),padding='same', stride=(1,2)),
        )
        self.inception_block_3 = nn.Sequential(
            nn.AvgPool2d((1,2)),
            nn.Conv2d(256,64,(1,1),padding='same'),
        )
        self.inception_block_4 = nn.Conv2d(256,64,(1,1),stride=(1,2))
        self.batchnorm2d_elu_2 = nn.Sequential(
            nn.BatchNorm2d(256),
            nn.ELU(),
            nn.Dropout2d(p=0.5)
        )
        self.conv2d_3 = nn.Conv2d(256,256,(1,5),padding='same')

    def forward(self,x):
        x = self.conv2d_1(x)
        x = self.batchnorm2d_1(x)
        x = self.conv2d_2(x)
        x = self.batchnorm2d_elu_1(x)
        x = self.averagepooling(x)
        x1 = self.inception_block_1(x)
        x2 = self.inception_block_2(x)
        x3 = self.inception_block_3(x)
        x4 = self.inception_block_4(x)
        x = torch.cat((x1, x2, x3, x4), 1)
        x = self.batchnorm2d_elu_2(x)
        x = self.conv2d_3(x)
        x = self.batchnorm2d_elu_2(x)
        x = F.adaptive_avg_pool2d(x, (1,3))
        x = x.squeeze()
        return x

net = InceptionEEGNet(10)
x = torch.rand(10,1,22,256)
print(net(x).shape)  # 输出应为torch.Size([10, 3])

验证说明 #

修复后四个分支的输出形状均为(10,64,1,64)，拼接后得到(10,256,1,64)，后续层可正常处理，最终输出符合预期。

内容的提问来源于stack exchange，提问作者assa