TODO.md

Incomplete implementations/bigger jobs

containers

cosmetic/consistency/speed

min/max bounds enforced on derived parameters (more generally, "bounds" as well as priors)

make portalocker/numba/dill a requirement?

version attribute should be in all components not just theory (samplers can have versions) [done for samplers; missing: likelihoods]

In the docs "Bases" (and UML diagram) not hyperlinked correctly (not sure how to fix)

dump log info along with each chain file if saving to file (currently in stdout)

PolyChord: check overhead

PolyChord:

update like/theory dict types to allow for instances (remove log.warning if all OK)

skip measure_and_set_speeds if only one component

maybe re-dump model info with measured speeds (need to allow saving of speeds from helper theories like camb.transfers)

Add other sample job scripts; add auto-configure database based on cluster names

Enhancements/Refactorings

Way to have parameters with different speeds within the same component without splitting into separate theories or sub-HelpTheories, or specify a sub-blocking for a component's parameters

Support "parameterization" option of theory .yaml to specify parameter yaml variants?/generalize !defaults

Let classes do all defaults combining; allow separate like instantiation + use equivalent to loading in cobaya

check_conflicts theory method or similar (so likelihoods can raise error when used in combination with other variant likelihoods using non-independent data)

If non-linear lensing on, model the non-linear correction via limber for faster semi-slow parameters

minimize:

• unbounded parameters with flat prior (this would make it safe to rotate the unbounded ones in minimize) [JT: not very much in favour, since that would break a bunch of other stuff. Maybe let's explore an alternative solution? e.g. auto-extend uniform priors.]

mcmc:

• finish removing .checkpoint in favour of updated.yaml and .progress
• For learning checks, X should perhaps ideally also depend slightly on the speed of the cycles, e.g. if either check becomes slow compared to a fast cycle.
• Update output thin factor once chains get over a given size, so that asymptotically the memory size of the chains doesn't grow indefinitely (and convergence checking time also doesn't grow correspondingly), just the thinning factor increases.
• more clever learning of covmat when only a few parameters missing: update only the row/columns of missing params, shrinkage estimator etc.

CLASS: make it non-agnostic

Test installed only (would need more clever pytest marking?)

auto-covmats:

• separate parameter matching into slow ones and fast ones, and prefer missing some fast parameters than missing slow ones.
• refactor to more be general and don't hard code Planck etc in main source (e.g. so external likelihood distributions can provide own covmat databases)

doc, install, model may be better documented generally rather than only in cosmo sections.

parameterization: there should be no need for "drop" if there are no agnostic components.

Dependencies system

• Maybe remove distinction between input parameters and requirements, so that calculate/logp takes both of them, which would be prepared by check_cache_and_compute. This would simplify the code a bit (in particular the part about input parameters that can be requirements, e.g. YHe) and makes all likelihood automatically callable outside a Model feeding requirements by hand. Problem: to prepare requirements we need arguments (e.g. units, ell_factor for Cl's) which are not passed to must_compute.
• Provider: it should be possible to save retrieving methods at initialisation so that everything (params, results, methods) can be retrieved with Provider.get(**args). Maybe it is interesting?