ResNet-18替换BatchNorm为InstanceNorm时遇空间维度不足报错

架构老林

2026-8-22

问题

我原本有一个运行正常的ResNet-18模型，为了把归一化层从BatchNorm替换成InstanceNorm，按以下方式修改了所有BatchNorm2d层：

resnet18.bn1 = nn.InstanceNorm2d(64, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)

resnet18.layer1[0].bn1 = nn.InstanceNorm2d(64, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer1[0].bn2 = nn.InstanceNorm2d(64, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer1[1].bn1 = nn.InstanceNorm2d(64, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer1[1].bn2 = nn.InstanceNorm2d(64, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)

resnet18.layer2[0].bn1 = nn.InstanceNorm2d(128, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer2[0].bn2 = nn.InstanceNorm2d(128, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer2[1].bn1 = nn.InstanceNorm2d(128, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer2[1].bn2 = nn.InstanceNorm2d(128, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer2[0].downsample[1] = nn.InstanceNorm2d(128, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)

resnet18.layer3[0].bn1 = nn.InstanceNorm2d(256, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer3[0].bn2 = nn.InstanceNorm2d(256, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer3[1].bn1 = nn.InstanceNorm2d(256, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer3[1].bn2 = nn.InstanceNorm2d(256, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer3[0].downsample[1] = nn.InstanceNorm2d(256, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)

resnet18.layer4[0].bn1 = nn.InstanceNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer4[0].bn2 = nn.InstanceNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer4[1].bn1 = nn.InstanceNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer4[1].bn2 = nn.InstanceNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer4[0].downsample[1] = nn.InstanceNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)

resnet18.fc = nn.Linear(in_features=512, out_features=10, bias=True)

所有num_features参数和原BatchNorm2d一致，只替换了归一化层类型。修改后的模型结构如下：

ResNet(
  (conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
  (bn1): InstanceNorm2d(64, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
  (relu): ReLU(inplace=True)
  (maxpool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
  (layer1): Sequential(
    (0): BasicBlock(
      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): InstanceNorm2d(64, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): InstanceNorm2d(64, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
    )
    (1): BasicBlock(
      (conv1): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): InstanceNorm2d(64, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): InstanceNorm2d(64, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
    )
  )
  (layer2): Sequential(
    (0): BasicBlock(
      (conv1): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (bn1): InstanceNorm2d(128, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): InstanceNorm2d(128, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
      (downsample): Sequential(
        (0): Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)
        (1): InstanceNorm2d(128, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
      )
    )
    (1): BasicBlock(
      (conv1): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): InstanceNorm2d(128, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): InstanceNorm2d(128, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
    )
  )
  (layer3): Sequential(
    (0): BasicBlock(
      (conv1): Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (bn1): InstanceNorm2d(256, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): InstanceNorm2d(256, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
      (downsample): Sequential(
        (0): Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)
        (1): InstanceNorm2d(256, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
      )
    )
    (1): BasicBlock(
      (conv1): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): InstanceNorm2d(256, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): InstanceNorm2d(256, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
    )
  )
  (layer4): Sequential(
    (0): BasicBlock(
      (conv1): Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (bn1): InstanceNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): InstanceNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
      (downsample): Sequential(
        (0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)
        (1): InstanceNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
      )
    )
    (1): BasicBlock(
      (conv1): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn1): InstanceNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): InstanceNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
    )
  )
  (avgpool): AdaptiveAvgPool2d(output_size=(1, 1))
  (fc): Linear(in_features=512, out_features=10, bias=True)
)

训练时出现错误：Instance Norm: ValueError: Expected more than 1 spatial element when training, got input size torch.Size([128, 512, 1, 1])，请问该如何解决？

解决方案

错误核心原因

InstanceNorm在训练模式下，需要对每个样本的空间维度（H×W）计算均值和方差。当空间维度缩小到1×1时，单个元素无法计算有效统计量（方差会趋近于0，触发除以0的异常）。错误出现在layer4最后一个BasicBlock的bn2层，输入空间维度已变为1×1。

具体解决方法

方法1：替换layer4的归一化层为BatchNorm2d

这是最稳妥的方案，因为ResNet-18的layer4输出空间维度天然是1×1，不适合InstanceNorm。修改代码示例：

# 替换layer4所有归一化层
resnet18.layer4[0].bn1 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer4[0].bn2 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer4[1].bn1 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer4[1].bn2 = nn.BatchNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)
resnet18.layer4[0].downsample[1] = nn.BatchNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=True)

方法2：禁用InstanceNorm的运行统计跟踪

将layer4中InstanceNorm的track_running_stats设为False，训练时直接使用当前batch的空间维度统计量。不过这种情况下，1×1的InstanceNorm本质退化为简单的缩放平移操作，效果不如BatchNorm：

# 修改layer4的bn层参数
resnet18.layer4[0].bn2 = nn.InstanceNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=False)
resnet18.layer4[1].bn1 = nn.InstanceNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=False)
resnet18.layer4[1].bn2 = nn.InstanceNorm2d(512, eps=1e-05, momentum=0.9, affine=True, track_running_stats=False)
# 其他layer4的bn层同理修改

方法3：调整网络结构避免空间维度变为1×1

比如修改layer4的卷积步长、移除部分池化层，让输出空间维度大于1×1。但这会改变ResNet的原始结构，需要重新调整全连接层输入维度，成本较高，仅适合有特殊需求的场景。

内容的提问来源于stack exchange，提问作者CasellaJr