Divergent flows on subsets of photometric redshift data

[bscot]

Description

Training pzflow on combined photometric redshift catalogs produces divergent flows on subset catalogs. For example, a catalog consisting of filter magnitudes for Euclid + LSST has (nearly) monotonically decreasing loss while training only on p(z | Euclid filters) diverges.

Screenshots

Output of training:

Catalog definitions for unsuccessful training:

And superset which is successful:

to reproduce:

Definition of the flow ensembles: