Code for FoCa

[alexarmbr]

Hello - thank you so much for the amazing work designing feature caching algorithms. I am wondering, do you have any plans to release the code for FoCa? There are some ambiguities in the methods section of the paper, I have tried implementing it several ways and cannot reproduce the result from the paper, saying that it allows you to approximate more steps than TaylorSeer without losing quality. If I interpret it one way it reduces to TaylorSeer(m=1) and if I interpret it another way the results look bad.

[Tammytcl]

Thank you very much for your interest in our work. We plan to make the code publicly available once the paper has been accepted.

[Fanghaipeng]

I tried to derive the scheme based on the BDF2+Heun formula proposed in FoCa, using the Δ notation instead of F. I found that, under the assumption \Delta(x_{t+1}) = \Delta(x_t), the method appears to reduce to the first-order TaylorSeer implemented in the code. In particular, when the last denoising step is t, this seems to completely lose the effect of using BDF2+Heun. Could you help me identify where my derivation is wrong?

[Fanghaipeng]

@alexarmbr Could you help me point out the mistakes in my derivation?