HaoDot
commented
3 years ago hey zzwei1!
when I first get attach to this paper, the training style confuses me a lot. However, after looking through for a few times, the following pic attracts me ,which shows how the loss function works.
As u have mentioned above, i take it right
such a video reconstruction network, you have N reconstructions that SPADE_E2VID outputed , while you only calculate the loss between the Nth reconstruction (the last reconstruction in total N reconstructions) and the Nth ground truth
Nice work! However, I'm a bit confused about the "many to one " training style. Does it mean that, in such a video reconstruction network, you have N reconstructions that SPADE_E2VID outpued , while you only calculate the loss between the Nth reconstruction (the last reconstruction in total N reconstructions) and the Nth ground truth? Does your "many to one" training stye means many inputs, "one" loss calculation instead of many inputs, "one" reconstrution, since you still generate N constructions? What confuses me is that, I thought your "many to one" means many inputs and only one reconstructed frame of the video. If so, we can't reconstruct a video by one model since one model only generate one frame.