Ziyao Li
Ziyao Li
Thank you for the detailed feedback. With regard to you questions: 1. chain orders do not matter. 2. currently there is no off-the-shelf method to predict the 3 chain prediction....
Sounds like a solid plan for me. But I would suggest you to use small learning rates, or mix your data with other pdb data to avoid overfitting. | |...
You’ll have to process them into feature.pkl data. Example data are in example_data folder. | | 李子尧 博士研究生 | | 北京大学大数据科学研究中心 ***@***.*** 北京市海淀区颐和园路5号 | ---- Replied Message ---- | From...
From your result I believe the model try to form a pseudo homodimer between A and B while ignoring Fv. I would suggest you to try AC and BC, and...
see the updates of datasets in #28.
关于KAN的主要讨论可以分为以下几类： 1. KAN可以拟合任意函数么？答案应该是正确的，且提出 KAN 的核心意义在于用 KAT bridging 一维和高维情况。而在1D情况下，无论你是用 Taylor、 fourier、 spline、还是 RBF 都没太大关系，怎么快、怎么准就怎么来。BTW， RBF最有意思的一点是它的基函数甚至和b-spline基本没差，所以实现上，特别是数值与优化上看，是最接近原版 KAN 的。你提出的所有工程化问题在RBF（[fastkan](https://github.com/ZiyaoLi/fast-kan/tree/master/fastkan)）都基本不存在。 2. KAN能抗过拟合么？这个问题是focus在机器学习的人最感兴趣的问题。当某个函数值的observation是含噪音的时候，KAN是否倾向于捕捉一个较为光滑的解来刻画其中单变量的变换？这个问题更empirical且很可能没有答案，因为泛化性这个问题本身就很难定义。我尝试过在信噪比较低的数据上用KAN，效果并不理想。 3. 关于Grid的选取，看上去似乎和MLP的宽度非常一致。越高的num grid其实就是越高的单元素变换的dof，毕竟单层KAN的参数量为 d_in * d_out * (n_grid + order)，越细的grid越能捕捉变换里面高频的部分，这在高非线性的函数拟合里可能很有用，但对于general 的 machine...
Now FastKAN is 3+ times faster (fwd) compared with efficient_kan. Believe you all want to try this. https://github.com/ZiyaoLi/fast-kan
> Very interesting, what is your benchmark and validation performance? Results are shown in the provided repo. I tested the forward speed of my FastKANLayer and efficient KAN's KANLinear, and...
> I think yes, FastKAN-like implementations that use RBF approximations are the Fastest. I am aware of 3 implementations so far, that are all more or less extremly similar: >...
> I inspected your github and adjusted my lr=1e-3, weight_decay=1e-5, gamma=0.85, with yours, @1ssb. > > So for the results I got after 15 epochs: > > -with FastKAN([28 *...