Evaluate joint density about crosscat

[vkmvkmvkmvkm]

Those tests are probably buggy and should be updated.
ᐧ

On Sat, Oct 10, 2015 at 9:44 PM, F Saad [email protected] wrote:

We need to expose density in the interface. My first idea was to change
simple_predictive_probability which takes a list of disjoint univariate
queries and interpret the input instead as joint. However it appears there
are several tests that invoke simple_predictive_probability which such a
change might break.

The second best option is to create joint_predictive_probability function
in the interface (and invoke this one through bayeslite).

—
Reply to this email directly or view it on GitHub
#62.

[axch]

Just looked at simple_predictive_probability in crosscat version 0.1.35. A few facts appear to hold, which are jointly rather confusing:

• As @fsaad noted, the interface to simple_predictive_probability itself maps over the query argument.
• I can't see any performance advantage to mapping over the query argument: all the work is inside the loop anyway, and nothing is cached except by the global cache on create_cluster_model_from_X_L that I introduced recently.
• simple_predictive_probability_multistate, which is what Bayeslite's crosscat metamodel calls, relies on the answer being a scalar (specifically, a numpy array of size 1, which can be converted to a floating point scalar by float -- unlike numpy arrays of other sizes).

In light of this, and @vkmvkmvkmvkm 's comment above, I suggest redefining simple_predictive_probability to return the joint probability of all the query cells (as a number, not a numpy array) and adjusting the tests to assume that (most of them seem to be crash-only tests anyway?)

I also explicitly suggest dropping the automatic mapping idea from this interface: if the client wants multiple independent queries, let them call the function many times.

[riastradh-probcomp]

@fsaad, have you confirmed that any tests actually break by doing this?

Preference for renaming it joint_predictive_probability; that will help catch and force review of old users too.

[fsaad]

@riastradh-probcomp for tests that rely on the function in its current form, please see

https://github.com/probcomp/crosscat/blob/master/crosscat/tests/test_pred_prob_and_density.py#L118-L124

https://github.com/probcomp/crosscat/blob/master/crosscat/tests/quality_tests/test_component_model_quality.py#L131

https://github.com/probcomp/crosscat/blob/master/crosscat/tests/quality_test_utils.py#L160-L164

https://github.com/probcomp/crosscat/blob/master/crosscat/cython_code/test_pred_prob_and_density.py#L82-L88

[axch]

Decision reached in conversation: simultaneously with changing the semantics, rename to predictive_probability.

[axch]

Fixed by 373b884 .