Questions about the Knowledge Transfer block

[Cyrano67]

I know you might be busy right now and will update the code of "Knowledge Transfe" part soon, but I'm just curious and want to know about your design. The question is, are the latent space representations of $b_{past}$ & $b_{both}$ from the same network? Or do you train two different networks (maybe stereo_matching_net) to get $b_{past}$ and $b_{both}$ for distillation? Have you tried these two different ways and if so, what is the difference? I'll appreciate it if you could answer my question.