AB Testing with elfi, bernoulli+lognorm about elfi HOT 4 CLOSED

I'll propably just refurbish the whole model to a more usable and robust syntax, I'll be in touch once I get something worthwhile going :D

from elfi.

Hi,

Without complete code I'm unable to reproduce this. Most crucially your handling of batch_size remains unclear.

However, based on the error message the problem is that the data argument given to legacy_updated is 1D, but you're trying to use it as 2D. Assuming data is an elfi.Prior, this is the default behaviour, and data.shape[0] = batch_size. You can use the size keyword, e.g. elfi.Prior('lognorm', 2, 10, size=3), to change this.

Also, if you intend to use var et al. as elfi.Summary, note that ELFI uses the first dimension for internal batches, so you probably should replace axis=0 with axis=1 everywhere.

That said, all this assumes that you make use of ELFI's internal batching. It is certainly possible to circumvent this (by essentially forcing batch_size to 1), but I do not recommended such. :)

from elfi.

Thx for the help, I got the simulator node working but ran to a trouble with the distance node.
I'm not really using batch_size atm since I just want to get the basic concept working first.
I think that this is due to summaries being 1d arrays instead of 2d:
In executing node 'd': all the input array dimensions except for the concatenation axis must match exactly.
That being said I think that I can get this working (most of the issues I'm facing are due to STAN and c++ and I feel like I'm going to a tree with my ass up when doing this in python...)

from elfi.

ELFI always uses batches, and batch_size is always the length of the first dimension (unless you hard-code stuff otherwise, which I do not recommend). This is true even if you use batch_size=1, in which case the first dimension of all arrays has length 1.

The summaries are typically 1d arrays of length batch_size (but can be 2d as well).

from elfi.