expanding 1x1xn global average pooling output to the size of the input map HxWxn

Question

I am trying to recreate the model in this paper : BiSeNetV2 There is one module called Context Embedding Block, in which Global Average Pooling is used to embed global information into the feature maps using a residual connection. In the figure provided in the paper, after global pooling and a 1x1 convolution, it mentions a broadcast operation that I traced back to this idea from another paper.

Now, I find myself in front of a relatively easy thing, which is taking a vector of 1x1 feature maps and emulate a feature map of HxW size by repeating the 1x1 element over its entirety. My question is: does it make sense to use Bilinear Upsampling to do so, or will such operation produce a different output from what I have in mind? And moreover, if I were to use Bilinear Upsampling, how should I choose the scale factor in order to accomodate for different input sizes?

Answer 1

I have found a workaround if anyone is interested. In pytorch in order to perform a global pooling operation one can use:

x = torch.randn(16, 128, 128)
out = F.adaptive_max_pool2d(x.unsqueeze(0), output_size=1)

as it was suggested here

In my case I am actually going to avoid adaptive pooling since it is not supported by onnx and I intend to run inference using onnx runtime later. So I am going to use torch.mean() in the following way:

x = torch.randn(1, 16, 128, 128)
out = torch.mean(x, dim=(2, 3), keepdim=True)

This way I obtain a 1x16x1x1 tensor.

In the end the answer to my original question was rather simple, since if I sum the tensor "out" and the input tensor "x", every element of each 128x128 feature map belonging to x will be summed to the one value of the out tensor corresponding to the index of that map, resulting in a tensor of the same size as x.

x = torch.randn(1, 16, 128, 128)
out = torch.mean(x, dim=(2, 3), keepdim=True)
out = out+x