Default SNPE Variant

[ali-akhavan89]

Hi,

I'm trying to understand how the default SNPE method is determined. Specifically, when I initialize an instance with inference = SNPE(prior=prior), which variant (A, B, or C) is used by default? In general, does it depend on the model we define for posterior_nn, such as model="maf”?

Thank you very much in advance,
Ali

[michaeldeistler]

Hi there,

the default SNPE method is SNPE_C. Note that, when doing single-round (amortized) inference (i.e., you are not doing multiround inference), then all SNPE methods are identical.

Michael

[ali-akhavan89]

Thank you so much! I have two follow-up questions:

1. Why are the SNPE methods identical in amortized inference, and how do they differ in multi-round scenarios?

2. Could you please point out the hyperparameters used in the inference network?

[michaeldeistler]

1. All SNPE methods use the maximum likelihood loss function in the first round, see here. In the follow-up rounds, they train with simulations generated from a proposal different from the prior, e.g., the current posterior estimate conditioned on your x_o. This results in a different posterior in the end, the proposal posterior. To correct for this mismatch, there are different techniques implemented in the different SNPE methods, e.g., by a post-hoc correction (SNPE-A) or by using a different loss functions (SNPE B and C).

2. You can find them here