Rules for units in modeling

gistfile1.txt

Current definitions:
 
input_units_strict

# If set to True, values that are passed in compatible units will be converted to
# the exact units specified in input_units.

input_units_equivalencies

# This can be set to a dictionary that maps the input names to a list of equivalencies, for example:

input_units_allow_dimensionless

# If set to True, values that are plain scalars or Numpy arrays can be passed
# to evaluate even if input_units specifies that the input should have units. 
# It is up to the evaluate() to then decide how to handle these dimensionless values.

==========
Rules for single models
============

- If one parameter is a Quantity (Q), then all should be

g = Gaussian1D(10 * u.Jy, 2 * u.m, 1 )

In [41]: g
Out[41]: <Gaussian1D(amplitude=10.0 Jy, mean=2.0 m, stddev=1.0)>

In [42]: g= Gaussian1D(10, 2 * u.m, 1 * u.m)

In [43]: g(2*u.m)
Out[43]: <Quantity 10.0>

===========================
- If models are defined with Q, inputs also must be Q

for Scale:
scl = Scale(2*u.one)

10*u.m *5*u.one
Out[44]: <Quantity 50.0 m>

=======================
input_units_allow_dimensionless should not be settable

==================
input_units_strict should not be settable

====================
Should input_units and return_units be settable?

input_units can be settable
return_units should not be settable.

================
when creating a model from a function with @custom_model decorator, input_units are not
defined. Function annotations should be used 


===========
Rules for Compound models
====================

If one of the submodels is instantiated with units then all submodels should be initialized with Q

==========================
- If initialized with Q, inputs shuld be Q.
==============
The rest of the rules are the same as for single models.

@nden , what would be useful for return_units for me is to actually allow it to be settable. For example, a spectrum evaluates to Jy but I want FLAM. xref astropy/astropy#6505 . However, if that is too much pain, then I guess I just have to deal with such things on my end.

The rest of the rules sounds reasonable. If for whatever reason, I don't want one of my parameters to have units, I will simply use the model as unitless and then re-apply units outside of models, like the old days.