Summarizing network / learned summary statistics

[cranmer]

From the API documentation, it seems like the package is restricted to simulators that produce output x with a constant shape. I wondered about simulators that produce output of varying size and shape. Currently, it seems this would possible if one introduces a function that wraps the simulator to produce fixed length summary statistics. This was not immediately clear to me from the API documentation or the home page in the documentation.

I do see that the Hodgkin-Huxley example does exactly this with domain-motivated summary statistics (as opposed to learned summary statistics).

However, the software paper mentions

Moreover, if dimensionality reduction of the simulator output is desired, sbi can use a trainable summarizing network to extract relevant features from raw simulator output and spare the user manual feature engineering.

I can see how this wold work with a pre-trained network that produces summary statistics, but my reading of that sentence with the word "trainable" is that SBI can also train the summarizing network. I haven't found this yet in the API documentation. If it is possible, an example would be great. Either way it would be nice to clarify this in the documentation and/or reword the sentence in the software paper.

[dfm]

Linking to openjournals/joss-reviews#2505 for posterity

[michaeldeistler]

Good point, thanks for raising this!

It is indeed possible to use a trainable network by passing an embedding_net to e.g. posterior_nn here. The embedding_net can be any trainable torch.nn.Module. It is also possible to pass such a network for SNRE.

This argument got lost from the the documentation page since we started using **kwargs here at some point. We will fix this in #299. The PR will also add a bit more explanation on the embedding_net.

However, there are not yet any pre-configured nets yet and no examples yet. Both of these things are on our hot desk, and are also documented in #279.

[cranmer]

👋 I see that #299 is merged. I think the API documentation online is still the same. Once that API documentation is updated and exposes embedding_net I will approve the JOSS submission. The example can come later and is not needed for JOSS paper.

[cranmer]

@michaeldeistler just want to make sure you see note above.

[michaeldeistler]

Thanks, saw it. We'll make a new PyPI release soon (hopefully tomorrow) and then update the docs :)