PyTorch——解决报错“RuntimeError: running_mean should contain *** elements not ***”

1 问题描述

在使用PyTorch编程的时候，经常遇到一种报错就是：“RuntimeError: running_mean should contain *** elements not ***”；

这次我具体的报错信息是：

  File "/home/songyuc/software/python/anaconda/anaconda3/envs/conda-general/lib/python3.7/site-packages/torch/nn/modules/module.py", line 541, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/songyuc/software/python/anaconda/anaconda3/envs/conda-general/lib/python3.7/site-packages/torch/nn/modules/container.py", line 92, in forward
    input = module(input)
  File "/home/songyuc/software/python/anaconda/anaconda3/envs/conda-general/lib/python3.7/site-packages/torch/nn/modules/module.py", line 541, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/songyuc/software/python/anaconda/anaconda3/envs/conda-general/lib/python3.7/site-packages/torch/nn/modules/batchnorm.py", line 81, in forward
    exponential_average_factor, self.eps)
  File "/home/songyuc/software/python/anaconda/anaconda3/envs/conda-general/lib/python3.7/site-packages/torch/nn/functional.py", line 1670, in batch_norm
    training, momentum, eps, torch.backends.cudnn.enabled
RuntimeError: running_mean should contain 192 elements not 768

从最后一行的报错信息，可以看到：进行求均值元素的总数应该是192而不是768；

2 解决方案

我们可以继续看看上一条提示信息：“File "/home/songyuc/software/python/anaconda/anaconda3/envs/conda-general/lib/python3.7/site-packages/torch/nn/functional.py", line 1670, in batch_norm”

有一个值得注意的信息是batch_norm，而我们的模型中也刚好使用了BN的操作，所以应该是BN的设置出现了问题，

我们回到代码定位的部分进行查看，需要查看的是BN初始化设置的代码，然后看到了下面的代码：

modules = [nn.Sequential(
    nn.Conv2d(in_channels, OUT_CHANNELS, 1, groups=1, bias=False),
    nn.BatchNorm2d(in_channels),
    nn.ReLU()),

我们可以看到，果然，BatchNorm2d的输入通道数与前一层Conv2d的输出通道数不一致，而这里的OUT_CHANNELS=192，in_channels=192，所以造成了这种维度的不一致，所以才会报错；

所以，我们需要根据自己模型的设计，将BN层与Conv层的输出维度保持一致。