Gaussian smoothing is wrong

[ad8e]

It needs to be a convolution with a Gaussian kernel.
The source repository https://github.com/petoem/taira doesn't do it correctly because it skips the first and last points.
Even if the source repository were correct, you'd have to scale up the kernel as the width (sigma) increased. A decent threshold is 3 * sigma. Currently, it's fixed at kernel size 2*2+1 = 5:

clearml-web/src/app/webapp-common/shared/single-graph/single-graph.utils.ts

Line 33 in 946d6ec

return arr.length > 5 ? smoothing.smoothen(arr, smoothing.ALGORITHMS.GAUSSIAN, 2, Math.round(weight / 100 * 5), false) : arr;

This small kernel causes a big difference at Gaussian smoothing 100, which should be almost flat. But there's still a lot of local variation:

That's because a big Gaussian width with a small kernel size is just a small rectangular window.

It is nontrivial to do performant Gaussian smoothing; a FFT is needed for larger kernels. You can try to find a library that does this; the term is "windowed convolution". If you try to use Taira for this, it will be very slow (and still wrong). One library I found is https://github.com/mljs/convolution, but I have never used it.

With Gaussian smoothing, you should set sigma = "Smoothing Value", rather than what it attempts now, which is Math.round(weight / 100 * 5).

EDIT: actually, since your points can be spaced unevenly, your task may be even harder. The conventional windowed convolution algorithms only work with evenly spaced points. But your existing solution has the same problem, so it won't be a downgrade.

[nirallegro]

@ad8e Thanks for the input. We'll look into it.