Scope separator for netcdf

[ferrine]

resolves #5657

Breaking changes

Use / separator for model scopes

Before #5607	After
with pm.Model("sub") as model: b = pm.Normal("var") model.named_vars.keys() # dict_keys(['sub_var'])	with pm.Model("sub") as model: b = pm.Normal("var") model.named_vars.keys() # dict_keys(['sub|var'])

Fix an error for nesting

Before #5607	After
with pm.Model("sub") as model: b = pm.Normal("var") with pm.Model("sub") as model: b = pm.Normal("var") # Error	with pm.Model("sub") as model: b = pm.Normal("var") with pm.Model("sub") as sub_model: b = pm.Normal("var") model.named_vars.keys() # dict_keys(['sub|var', 'sub|sub|var'])

It is also possible to save to netcdf

with pm.Model("scope") as model:
    b = pm.Normal("var")
    trace = pm.sample()
az.to_netcdf(trace, "trace.nc")
trace1 = az.from_netcdf("trace.nc")
assert "scope|var" in trace1.posterior

[codecov]

Codecov Report

Merging #5663 (82df819) into main (fa015e3) will decrease coverage by 3.14%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main    #5663      +/-   ##
==========================================
- Coverage   86.86%   83.72%   -3.15%     
==========================================
  Files          75       75              
  Lines       13715    13722       +7     
==========================================
- Hits        11914    11489     -425     
- Misses       1801     2233     +432     

Impacted Files	Coverage Δ
pymc/model.py	85.61% <100.00%> (+<0.01%)	⬆️
pymc/distributions/transforms.py	54.12% <0.00%> (-45.88%)	⬇️
pymc/distributions/dist_math.py	57.71% <0.00%> (-29.72%)	⬇️
pymc/distributions/logprob.py	68.80% <0.00%> (-26.61%)	⬇️
pymc/data.py	58.82% <0.00%> (-24.79%)	⬇️
pymc/tuning/scaling.py	35.18% <0.00%> (-24.08%)	⬇️
pymc/distributions/continuous.py	87.87% <0.00%> (-10.31%)	⬇️
pymc/aesaraf.py	79.39% <0.00%> (-9.05%)	⬇️
pymc/distributions/discrete.py	91.48% <0.00%> (-8.25%)	⬇️
pymc/distributions/multivariate.py	86.61% <0.00%> (-5.43%)	⬇️
... and 6 more

[ricardoV94]

It's not as nice :(

[ferrine]

Well, | seems to be the least ugly otherwise. I'll wait for others to approve or propose a better separator. Possible candidates are:

::

scope::variable
scope::sub::variable

:

scope:variable
scope:sub:variable

#

scope#variable
scope#sub#variable

[ricardoV94]

Those actually look a bit better. What about > ?

My problem with the | is that it makes me thing of or too much

[ferrine]

Those actually look a bit better. What about > ?

My problem with the | is that it makes me thing of or too much

scope>variable
scope>sub>variable

It compares well to the rest, what do you think of ::, so far one of the best candidates for me?

[ricardoV94]

Which one do you prefer?

[ferrine]

::

scope::variable
scope::sub::variable

[ferrine]

it is at least pytest like

[michaelosthege]

All separators | / \ . > → :: will require to access with idata.posterior["sub|var"] instead of idata.posterior.sub_var.
That's manageable I guess.

But are there consequences for doing ArviZ things with such variable names? Will it break?

Note that if we'd (finally) introduce a pm.rcParams (#4657), the choice of separator would be a prime candidate.

[ferrine]

All separators | / \ . > → :: will require to access with idata.posterior["sub|var"] instead of idata.posterior.sub_var.

Yes, but with _ separator you can't tell if it is two words or a scope and a variable name

[ferrine]

But are there consequences for doing ArviZ things with such variable names? Will it break?

I'm using it with arviz just fine, there were no troubles so far

[ferrine]

I've merged this for now to enable saving to disk. Agree that with #4657 it makes sense to specify the separator in rcParams.