Sanyam Kapoor

This has been kind of troubling me and glad that you pointed this out. If you make the modifications, the system seems to collapse to almost the mode something like...

Comparing the kernel function and gradients to the paper's code, here's some sanity checks. ```python X_init = (3 * torch.randn(n, *gauss.event_shape)).to(device) X = X_init.clone().requires_grad_(True) kxy_pt = K(X, X.detach()) dxkxy_pt =...

> the minus sign in author's code is due to the derivation that is hardcoded by the author Ah apologies, I misspoke. > What happens when you compare dxkxy against...

UPDATE: I don't particular see the difference b/w the TF2 code you referenced and my implementation except for the gradient sign. It is weird that it works on all the...

Hey @shwangtangjun ! Just noticed that you've taken much effort to figure this out. Thanks for this!

I've ported this for the static site generator Hugo (https://github.com/activatedgeek/distillpub) and wanted to make this option for places like an about page. For now, I'm using CSS `visibility: hidden;` to...

/assign @yixinshi @jlewi