Questions about the transition layers between stages in the codes

[prophett0527]

Hello,
I have some questions about the transition layers between stages in PoseHighResolutionNet class.

The transition layers contains conv3x3 layers with stride=1 and stride=2.
The codes seem similar with the fusion layers in HighResolutionModule class.
Fusion layers have conv1x1 layers for upsampling, but transition layers do not have conv1x1.
What is the difference between these two type layers?
What happens if there are no transition layers between stages?

I couldn't find any explanations about transition layers in the original paper.
Could you please kindly explain the transition layers here?
Thank you!

[aFewThings]

Hi, have you found your answer?