I don’t think it is explicitly stated which variant of densent with regards to bottleneck+compression torchvision.models implements.
AFAIK the DenseNet paper defines three variants: DenseNet, DensNet-B, and DenseNet-BC.
With DenseNet-B the input of a ConvLayer is downsampled to a feature map size of 4 * ‘growth rate’.
In DenseNet-BC the transition layers downsample the feature map size by a factor of theta < 1 (often 0.5)
I looked at the implementation and it seems to me that torchvision implements the DenseNet-BC, is that correct?