Is it possible to store arbitrary node output? about elfi HOT 4 CLOSED

Closing this due to inactivity.

from elfi.

Hi!

Have you given pool for the inference method (see the attached example)?
If you're dealing with "accepted" simulations, then you're doing rejection sampling or smc? You can look at the results-object which contains the accepted samples for each of the nodes. e.g. result.outputs['t1']. If you want all the generated samples, you can now access the batches in pool.stores['t1'].

import elfi
from elfi.examples import ma2

# load the model from elfi.examples
model = ma2.get_model()

# Set up output pool for the desired nodes
pool = elfi.OutputPool(['MA2', 'S1', 'S2', 't1', 't2'])
rej = elfi.Rejection(model['d'], batch_size=10, pool=pool)
result = rej.sample(n_samples=100, quantile=0.2)
print("Accepted samples of node \'t1\' = \n" + str(result.outputs['t1']))
print("Generated samples from node \'t1\' = \n" + str(pool.stores['t1']))

from elfi.

Thanks for the quick response @hpesonen !

Yes I do hand the pool = elfi.OutputPool instance to the inference method (elfi.Rejection(d, pool=pool)). The MA2 example also works fine: I can access both the result.outputs and the pool.stores. Although the results.outputs does not contain the node outputs from MA2, S1 and S2 - just from d, t1 and t2.

In my own model result.outputs only contains the prior samples as well as the output of my distance function - but not of either the summary nor the simulator function. My simulator function output is a set of tuples (graph edges) that get compared to the observation(s) using a custom Summary statistic function (Jaccard index), which outputs a float. Could the set be a dtype that can't get stored?

from elfi.

Could you include a small reproducible example of the issue? You could give keyword output_names (list) to Rejection to get the node outputs from the ELFI graph. However, unfortunately this feature currently supports only numpy arrays and won't help if your simulator data are sets.

You should be able to access all the samples you have generated via OutputPool even if the node output is e.g. a list of sets of size batch_size. And if you use Rejection, you could manually choose and store only samples with the lowest discrepancy.

from elfi.