I wonder if overlapping the patches would improve accuracy further as a way to kind of anti alias the data learned / inferred. In other words, if position 0 is 0,0 - 16,16 and position 1 is 16,0 - 32,16 instead we used 12,0-28,16 for position 1 where it overlaps 4 pixels of the previous position. You’d have more patches / it would be more expensive compute wise, but it might dealias any artificial aliasing that the patches create during both training and inference.
Logged into my personal account for this one! I'm a lead author on a paper that explored exactly. It does enable faster training and smaller model sizes. For reference, you can get 80% accuracy on CIFAR-10 in ~30 minutes of CPU (not using crazy optimizations). There are open questions about scaling but at the time we did not have access to big compute (really still don't) and our goals were focused on addressing the original ViT's claims of data constraints and necessities of pretraining for smaller datasets (spoiler, augmentation + overlapping patches plays a huge role). Basically we wanted to make a network that allowed people to train transformers from scratch for their data projects because pretrained models aren't always the best solutions or practical.
I also want to give maybe some better information about ViTs in general. Lucas Beyer is a good source and has some lectures as well as Hila Chefer and Sayak Paul's tutorials. Also, just follow Ross Wightman, the man is a beast
Thanks for all the good work and all the pointers! Awesome stuff. Let me know if you would want to join us live on a Friday and go over some of your newer work or any recent papers you find interesting. Feel free to reach out at hello@oxen.ai if so :)
I very cursorily skimmed your paper but I didn’t spot where it discusses overlapping the patches. Is it the section about using the hybrid model with a convolutional step which de facto accomplishes it (maybe?) instead of overlapping patches?
Yeah so I can get how that might be confusing. Sometimes code is clearer. So in the vanilla transformer you do a patch and then embed operation, right? A quick way to do that is actually with non-overlapping convolutions. Your strides are the same size as your kernel sizes. Look closely at Figure 2 (you can also see a visual representation in Figure 1 but I'll admit there is some artistic liberty there because we wanted to stress the combined patch and embed operation. Those are real outputs though. But basically yeah, change the stride so you overlap. Those create patches, then you embed. So we don't really call it a hybrid the same way you may call a 1x1 cov a channel wise linear (which is the same as permuting linear then permute).
