RegNet - Number of channels in and out for each layer in a bottleneck

[polars05]

In the RegNet paper (https://arxiv.org/pdf/2003.13678.pdf), Fig 4b shows that (in the downsampling block):

• the number of channels into the first conv1x1 == the number of channels out of the first conv1x1
• the change in the number of channels happens at the conv3x3 layer

However, in timm, it seems that the change in number of channels happens at the conv1x1 instead: https://github.com/rwightman/pytorch-image-models/blob/715519a5eff9046e40958b4c222e0e96f75014e9/timm/models/regnet.py#L143

Is there a reason for implementing it this way? Thanks!