negate variable names

[aloctavodia]

This allow to pass a variable to var_names and get all the variables except the one that was passed.

Thus if data is a model with variables names ['mu', 'theta', 'tau']

az.plot_joint(data, var_names=['~theta']);

returns

One caveat is that variables names are then no allowed to starts with ~.

[canyon289]

Can I like this feature more than once? This is a great idea

[ahartikainen]

Good idea. I'm assuming "~a" is not valid name?

[aloctavodia]

Thanks @canyon289 and @ahartikainen. Right names like "~a" are not valid.

[ahartikainen]

How can we handle this situation?

az.from_dict({"mu" : np.random.randn(2,30,3), "~mu" : np.random.randn(2,30)})

[ahartikainen]

exclude_vars = [i[1:] for i in var_names if i.startswith("~") and i not in all_vars]

[aloctavodia]

This will give the wrong result

[aloctavodia]

ups, sorry I misread that!

[aloctavodia]

OK, with this change negation will work as expected except that if a variable name starts with ~ like ~mu when passing ~mu to var_names it will understand that you want the variable ~mu and not to negate a variable named mu. Passing ~~mu will be understood as negating the variable ~mu. Is this OK?

[ahartikainen]

Yes, that sounds logical.

The is the corner case where we have mu and ~mu ... Maybe it should throw a warning if there are both parameters: par and ~par?

[ahartikainen]

But... how do you negate mu in that situation?

[canyon289]

you're saying if a model contains both mu and ~mu how would you negative one but not the other?

[ahartikainen]

How do you negate mu. I'm not saying we need to handle it, but warn users

[canyon289]

For now I think that's reasonable. Scan the var_names in InferenceData and see if any are leading with a tilda. If so provide warning negation behavior might be weird.

If people submit tickets about warning and want a "better" we can then decide to solve further?

[aloctavodia]

oops, I missed this conversation, I will add the warning.

[ahartikainen]

LGTM