Changing num_sims in TSNPE loop

[danielmk]

I am currently using the TSNPE loop from here, and I am considering to start with a large number of n_sims in the first round and then go to a lower number or iteratively decrease it. The TSNPE paper says regarding the convergent proof: "In this equation, we assumed that all rounds contain equally many simulations, but the proof can easily be extended to rounds with different numbers of simulations by adding weights to the above sum."

I assume this means that changing n_sims could be done in theory. Does the current sbi implementation take care of the weighting or must n_sims be equal between rounds?

[michaeldeistler]

Hi there,

Thanks for reaching out! I am an author on the paper so I think I can clarify.

TL;DR: Changing n_sims can be done in theory and in practice. It does not require any modifications to the code.

Long version: I can see how this sentence is a bit confusing, so let me rephrase it. What the sentence meant to say is: The algorithm also converges if rounds have different numbers of simulations because the proof holds also if one considers weighted sums of the proposal densities.

In other words, one does not have to add any explicit weights, ever. The reweighing is implicit in having different numbers of simulations per round and the proof still holds for these (implicitly) reweighed proposal distributions.

Does this make it clear?

Michael

[danielmk]

Thank you, that clarifies it.