Questions about parallel SPPE? about alphapose

[Fang-Haoshu]

To calculate the loss between the center-located ground truth poses and the output, we use the same loss as the main SPPE.
The center-located ground truth pose image is cropped from the original image using ground truth bounding box.

[Sakuralala]

So what if I discard the parallel SPPE and just use the main branch?Should I add a new loss or just train the net as before(with stn but without parallel SPPE)? @Fang-Haoshu

[Fang-Haoshu]

you can train it as before, but a question is why you want to do that..

[Sakuralala]

emmm...poor gpu memory?

[Fang-Haoshu]

Oh I see, then you may try using a smaller batch size and enlarge the iter_size. Refer to the opt.lua

[Sakuralala]

Thanks for replying!