Unexpected output shape value for BatchNorm1dToQuantScaleBias

[green-cabbage]

Hey, I have been trying to implement brevitas' version of nn.BatchNorm1d(), and I receive this error:
ValueError: Using a target size (torch.Size([64, 1])) that is different to the input size (torch.Size([1, 30, 64, 1])) is deprecated. Please ensure they have the same size.

I have backtracked the error to the point where I believe this is the cause. When I run this:

input_dim = 30
batch_dim = 64
x = torch.rand(64,30)
quant_batchnorm = qnn.BatchNorm1dToQuantScaleBias(
                  num_features=input_dim,
                  return_quant_tensor=True
)  
print(quant_batchnorm(x).shape)

the code prints: torch.Size([1, 30, 64, 30]) when I would expect it to return something like torch.Size([64,30]). Do you guys know why the quant layer returns in that specific shape of [1, input_dim, batch_dim, input_dim] ? Is this behavior intentional? If so why?
Thanks in advance

[green-cabbage]

Issue solved:

There was two small changes I made which made it work.

First, the default parameter runtime_shape=(1, -1, 1) for class BatchNorm1dToQuantScaleBias should actually be runtime_shape=(1, -1) . At least that's the shape that worked for me, and makes sense in the context of torch.nn.BatchNorm1d

Second, I was initially getting runtime_shape=(1, -1, 1, 1) because that is class ScaleBias's default value, and class BatchNorm1dToQuantScaleBias wasn't passing down runtime_shapeparameter properly.
I solved this by changing from ScaleBias.__init__(self, num_features, bias) to ScaleBias.__init__(self, num_features, bias, runtime_shape=runtime_shape) when initializing class QuantScaleBias

If you would like, I can make a pull request with these changes.

[Giuseppe5]

Hi,

Thanks for looking into this!
Please feel free to open a PR on the dev branch with the fix.

[MikeHatchi]

Hi @green-cabbage many thanks, I've applied the solution.
I have not noticed normalization actually. In my case, for eg.: among the input data, I had 1.32.... and after BN1dScaleBias 1.39 au lieu de instead of 0.something.

Hi @Giuseppe5 , I was wondering why BN1dScaleBias operation is a missing topic as BN1d is a very largely used in non-quantize context? Is it because as quantization is about low precision, then it is not supposed to use BN logic?