Simo Ryu

Ah, indeed. Maybe @pedrogengo can help us here

Nice work as always!!

Ooooo so alpha = 4 is the default? This is going to make my head hurt 🫠 So in the paper they had high alpha values, but it seemed to...

What if we set alpha = 1 as default, and scale up lora instead?

Hmm I want to talk about this bit more deeper. Its just that I"m not really understanding why we need to introduce scale term here. I think that the previous...

But its just my thought, and I think this will change a lot of code all at once, (with readme and some of your analysis as well) so im happy...

I think whatever works fine if it is just backward compatible. Overall, I just want to keep the notion of "full training == $\alpha = 1$" and "base model ==...

Basically because I've built many test scripts with those notion, and lot of other implementations are on that idea as well, so I think it would surprise a lot of...

So maybe keeping alpha intact, setting default scale = 1, and exposing it the advanced users might work? + during saving, fallback scale into A? $$ W' = W +...

Yes, so because they are going to be blurred anyways, I've used just plane face detection module. You can use any sort of pipeline you wish, but It is going...