BatchNormalization.lua:87: got 32-feature tensor, expected 0

[emredog]

Hello,

I have a model that makes use of the nn.SpatialBatchNormalization, and I got the error in the title.

I want to make inference on a single image (3x224x224). I've tried both "unsqueezed" (1x3x224x224) and the original, both lead to the same error.

The relevant part of the network is as follows:

  (1): nn.SpatialConvolution(3 -> 32, 3x3, 1,1, 1,1) without bias
  (2): nn.SpatialBatchNormalization (4D) (0)

I've added the following prints to the checkInput function that throws the error:

function BN:checkInputDim(input)
   local iDim = input:dim()
   assert(iDim == self.nDim or
              (iDim == self.nDim - 1 and self.train == false), string.format(
      'only mini-batch supported (%dD tensor), got %dD tensor instead',
      self.nDim, iDim))
   local featDim = (iDim == self.nDim - 1) and 1 or 2
   print("train mode?", self.train)
   print("featDim", featDim)
   print("runningMean:nElement", self.running_mean:nElement())
   print("input size:", input:size())
   assert(input:size(featDim) == self.running_mean:nElement(), string.format(
      'got %d-feature tensor, expected %d',
      input:size(featDim), self.running_mean:nElement()))
end

this yields the following output for the unsqueezed image:

train mode?	false	
featDim	2	
runningMean:nElement	0	
input size:	   1
  32
 224
 224
[torch.LongStorage of size 4]

Any help or lead to what to try next is much appreciated, thanks!

[emredog]

I believe this is caused by the lack of the running_mean being null (0 element). Another person encountered probably the same error as me: gaobb/DLDL-v2#4