continuous variable child with discrete variable parent

[borjasor]

First of all congratulate the team for 'pgmpy', it is a great tool!

Hopefully someone can guide me on how to implement this case.
I define a model with two nodes:

model = BayesianNetwork([('A', 'B')])

'A' is a discrete variable with 4 estates, its CPD is defined by:
cpd_A = TabularCPD(variable='A', variable_card=4, values=[[0.25], [0.25], [0.25], [0.25]], state_names={'A': ['one', 'two', 'three', 'four']})

'B' is a continuous variable with a different 'PDF' for each evidence of 'A';
one_factor = ContinuousFactor(['B'], pdf_one)
two_factor = ContinuousFactor(['B'], pdf_two)
three_factor = ContinuousFactor(['B'], pdf_three)
four_factor = ContinuousFactor(['B'], pdf_four)

The question is how can I assign these PDFs of B to the BayesianNetwork 'model'?

Thank you for your help.

[ankurankan]

@borjasor Unfortunately, continuous and hybrid (continuous + discrete) factors are not fully supported yet.

[borjasor]

Thank you @ankurankan , I assume the only way to proceed with pgmpy (in my case) would be to discretize the continuous variable. However, there seems to be quite a few issues when trying to do this:

1. The examples in 'discretize.py' seems to refer to a nonexisting class 'ContinuousNode'
2. Assuming that 'ContinuousNode' = ContinuousFactor, when using 'RoundingDiscretizer' or 'UnbiasedDiscretizer' there are issues as both make use of 'cdf' = cumulative density function which there does not seem to be way to define in 'ContinuousFactor'.
3. When using class 'BaseDiscretizer', currently is defined in a way that uses 'ABCMeta' class incorrectly as it is a 'meta class', not a 'base class', so you will need to change the class definition to:
   class BaseDiscretizer(metaclass=ABCMeta):

Even when fixing the above, when using 'discretize' on 'BaseDiscretizer' we get an error:
TypeError: Can't instantiate abstract class BaseDiscretizer with abstract method get_discrete_values

[ankurankan]

@borjasor Yes, the discretizer methods need to be fixed. The continuous variable implementation (still incomplete) has changed quite a bit since the methods were written. I would just avoid using that for now. I just added a new function pgmpy.utils.discretize (https://github.com/pgmpy/pgmpy/blob/dev/pgmpy/utils/utils.py#L119), which is a simple wrapper around pandas.cut methods that you can use for discretization.

[borjasor]

Thank you again @ankurankan , I will look into this.