MohammadrezaBanaei / LoRA-XS

LoRA-XS: Low-Rank Adaptation with Extremely Small Number of Parameters
https://arxiv.org/abs/2405.17604
14 stars 2 forks source link

Shape mismatch in paper? #1

Open ro1406 opened 4 weeks ago

ro1406 commented 4 weeks ago

Hi, thanks for your work!

I was just going through the paper, and noticed a possible shape mismatch in the formulas, and wanted to clarify. In Equations 1 & 2, the product shown is BAx and BRAx. However, based on the formulation using SVD, i got the shapes of B, R, and A as follows:

Hence, computing BRAx would produce a possible shape mismatch (BRA -> rxn X rxr X mxr shapes dont match).

Is this a typo and the original formula was meant to be ARBx?

Thanks and best of luck

MohammadrezaBanaei commented 2 weeks ago

Hi! Thank you for your interest in our work! We have indeed noticed a typo in the paper's formula, and we will correct it in the paper soon. Below is the version consistent with our code and experiments:

Thanks again for catching this!