PyTorch中conv2d()参数类型不匹配错误排查求助

错误原因分析 #

1. Block类初始化参数顺序错误
   Block类的__init__方法定义参数顺序为(in_channel, out_channel, stride=1, downsample=None)，但在ResNet的layers方法中调用Block时，传参顺序写成了block(self.in_channels, out_channels, downsample, stride)，将downsample和stride的位置搞反。这导致Block初始化时，stride参数被赋值为downsample（一个Sequential模块或None），而Conv2d的stride参数要求是int或tuple类型，最终在forward阶段调用conv2d时触发参数类型不匹配错误。

2. Block类中的拼写与逻辑错误

   • BatchNorm实例拼写错误：self.batcnorm2应为self.batchnorm2；
   • forward中重复调用同一BatchNorm：第二次归一化错误使用self.batchnorm1，而非正确的self.batchnorm2；
   • 残差边下采样对象错误：对经过卷积的x而非原始identity进行下采样，导致维度不匹配。

3. ResNet类中的多处错误

   • forward中大写X应为小写x；
   • 全连接层调用错误：不存在linear_layer1属性，且调用写法错误；
   • 通道数数值错误：layer2的out_channels写为124，不符合ResNet标准的128；
   • AdaptiveAvgPool2d输出未展平，直接传入全连接层会导致维度错误。

解决方案 #

1. 修正Block类的调用参数顺序 #

在ResNet的layers方法中，调整Block的传参顺序，匹配其__init__定义：

# 原错误代码
# layers.append(block(self.in_channels,out_channels,downsample,stride))
# 修改后
layers.append(block(self.in_channels, out_channels, stride, downsample))

2. 修复Block类的拼写与逻辑错误 #

class blocks(nn.Module):
  def __init__(self, in_channel, out_channel, stride=1, downsample=None):
    super(blocks,self).__init__()
    self.conv1=nn.Conv2d(in_channel,out_channel,kernel_size=3,stride=stride,padding=1)
    self.batchnorm1=nn.BatchNorm2d(out_channel)
    self.conv2=nn.Conv2d(out_channel,out_channel,kernel_size=3,padding=1,stride=1)
    # 修正拼写错误
    self.batchnorm2=nn.BatchNorm2d(out_channel)
    self.relu=nn.ReLU()
    self.identity_downsample=downsample
  def forward(self,x):
    identity=x
    x=self.conv1(x)
    x=self.batchnorm1(x)
    x=self.relu(x)
    x=self.conv2(x)
    # 修正BatchNorm调用对象
    x=self.batchnorm2(x)
    if self.identity_downsample is not None:
      # 修正下采样对象为原始identity
      identity=self.identity_downsample(identity)
    x=x+identity
    x=self.relu(x)
    return x

3. 修复ResNet类的错误 #

class ResNet(nn.Module):
  def __init__(self,block,layers,num_classes=136):
    super(ResNet,self).__init__()
    self.in_channels = 64
    self.conv1=nn.Conv2d(3,self.in_channels,kernel_size=7,padding=1)
    self.norm1=nn.BatchNorm2d(self.in_channels)
    self.relu=nn.ReLU()
    self.layer1=self.layers(block,64,layers[0])
    # 修正通道数为128
    self.layer2=self.layers(block,128,layers[1],stride=2)
    self.layer3=self.layers(block,256,layers[2],stride=2)
    self.layer4=self.layers(block,512,layers[3],stride=2)
    self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
    self.linear_classifier1=nn.Linear(512,256)
    self.linear_classifier2=nn.Linear(256,num_classes)
  def layers(self,block,out_channels,no_of_blocks,stride=1):
    downsample=None
   
    if stride!=1 or self.in_channels!=out_channels:
      downsample=nn.Sequential(
          nn.Conv2d(self.in_channels,out_channels,kernel_size=1,stride=stride),
          nn.BatchNorm2d(out_channels)
      )
    layers=[]
    # 已修正参数顺序
    layers.append(block(self.in_channels,out_channels,stride,downsample))
    self.in_channels=out_channels
    for i in range(1,no_of_blocks):
      layers.append(block(self.in_channels,out_channels))
    
    return nn.Sequential(*layers)

  def forward(self,x):
    x=self.conv1(x)
    x=self.norm1(x)
    x=ff.max_pool2d(x,kernel_size=3,stride=2,padding=1)
    x=self.relu(x)
    x=self.layer1(x)
    x=self.layer2(x)
    x=self.layer3(x)
    x=self.layer4(x)   
    # 修正大写X为小写x
    x=self.avgpool(x)
    # 展平张量，适配全连接层输入
    x = x.view(x.size(0), -1)
    # 修正全连接层调用
    x=self.linear_classifier1(x)
    x=self.relu(x)
    x=ff.dropout(x,0.4)
    x=self.linear_classifier2(x)
    x=ff.sigmoid(x)         
    return x

4. 补充必要导入 #

确保代码开头导入torch.nn.functional：

import torch.nn.functional as ff

内容的提问来源于stack exchange，提问作者Sachin