Feature wishlist

[itsdfish]

Hello,

I realize you are leave. So don't feel obligated to respond. I would like to make a list of features that would be nice to add to the package.

• SetTransformer: I believe this has an advantage over DeepSet for certain models

• TimeSeriesTransformer and/or TimeSeriesNetwork: works for time series models.

• Support for Enzyme I tried using Enzyme.Duplicated with the inputs to DeepSet, but I received an error about using it with nested networks

• Support for using posterior distributions with PiecewiseEstimator if it has a significant advantage over treating sample size as a random variable

• Support for model comparison

I would like to help out where I can. From my naive perspective, it seems like adding support for model comparison is the easiest. I might be able to make some progress if you would be willing to provide some feedback on the design. I'll look at the code over the next couple of days and sketch out a plan.

[msainsburydale]

Hello,

Many thanks for making this list, and for offering to contribute to the package.

Regarding the different neural networks (SetTransformer, TimeSeriesTransformer, TimeSeriesNetwork): the package doesn't enforce the use of any particular neural network, so any Flux model can be used. (Many of the documentation examples use DeepSet, but this is mainly for "historical" reasons.) So, it would just be a matter of finding a suitable Flux implementation of these neural networks, and then possibly adding some examples in the documentation examples. I don't think this would be too difficult and it would be a great benefit to the package.

I have never used Enzyme but agree that adding support for it would be good.

I think that PiecewiseEstimator with PosteriorEstimator is lower priority for now, but we can keep it in mind if treating the sample size as a random variable is not leading to the results we would expect.

Support for model comparison would be very helpful. In addition to the link you shared, there is also this paper that does model comparison/selection. That paper uses this package, so I might be able to ask the authors if they'd be willing to share their code so that the general methodology can be added to the package.

Cheers,
Matt

[itsdfish]

That sounds good. Thanks for the opportunity to contribute and for your willingness to provide guidance where needed. My goal this weekend to do background reading was unfortunately derailed by the flu. I should be able to make some progress next weekend after I have recovered. I will add the André et al. (2025) paper to my list.

Update

I gave the André et al. (2025) paper a cursory read and now see that the code used NeuralEstimators.jl. Indeed, having access to the code used for training would be a great addition to the package. I did find the trained networks, but not the code for training.

[msainsburydale]

Thanks, and hope you're on the mend soon. Just FYI, I've added a to-do list here.