numdifftools usage in lmfit

[eendebakpt]

The numdifftools package is used in lmfit for calculation of the covariance matrix. The package is giving various deprecation warnings on our platform. I created a PR to address one of the warnings, but the numdifftools author is not responding.
I anyone from the lmfit community in contact with the numdifftools author? Or are there alternative options we could use in lmfit?

[reneeotten]

I added the numdifftools calculation of the covariance matrix to lmfit, but have no contact with the author. It appears that the NumPy deprecation warnings were introduced in version 1.22.0 which means there is still a fair amount of time for the numdifftools author to resolve this before it becomes a real issue.

If the warnings really bother you, you can for sure just suppress them - I don't recall the exact syntax but Google will know ;) I have no idea if there are alternatives, but have certainly no intentions to change anything in lmfit until the numdifftools package doesn't work anymore because of NumPy failures. But again, presumably numdifftools will have resolved it by then.

[newville]

@eendebakpt @reneeotten Yeah, I see those deprecations too when running our tests. I agree that a) they are really deprecations that numdifftools should resolve - and thanks very much for that PR, and b) we don't really want to have to re-tool just because releases of numdifftools is a little slower to fix deprecations than we might like.

And, maybe related: we should probably plan to push 1.0.4 relatively soon to avoid similar views on slow progress!

I don't know that author myself, but would be willing to write to them if we thought that would be helpful.

[peendebak]

The warnings have been addressed in pbrod/numdifftools#63

[newville]

@eendebakpt Thanks! Since these are being addressed upstream, don't think there is any need to do anything more about the warnings on our end.

[neoncloud]

Speaking of numdifftools alternatives, has anyone tried an automatic differential engine? Torch may be too heavy, but there are some lightweight autograd engines that work well with numpy, such as tinygrad and Autograd.

[reneeotten]

please don't hijack old topics for a new discussion, instead open a new one ;) What would be the benefit of changing to any of them compared to the use of numdifftools? The only thing we want from the package is a Hessian matrix.