DBNInference ValueError: Self loops are not allowed

[576774572]

Hi!
when I try to construct model with DBN like this:

from pgmpy.models import DynamicBayesianNetwork as DBN
from pgmpy.factors.discrete import TabularCPD
from pgmpy.inference import DBNInference
dbn = DBN()
dbn.add_edges_from([(('a', 0),('b', 0)),(('a', 0),('a', 1)), (('a', 0),('b', 1))])
a_0=TabularCPD(('a', 0), 2, [[0.5],[0.5]])
b_0=TabularCPD(('b', 0), 2, [[0.3,0.7],[0.7,0.3]],[('a',0)],[2])
a_1=TabularCPD(('a', 1), 2, [[0.3,0.7],[0.7,0.3]],[('a',0)],[2])
dbn.add_cpds(a_0, b_0, a_1)
dbn_inf = DBNInference(dbn)

But when I execute the code I get this error:

File "C:\users\a5767\appdata\local\programs\python\python36\lib\site-packages\pgmpy\models\untitled9.py", line 25, in <module> dbn_inf = DBNInference(dbn)
File "c:\users\a5767\appdata\local\programs\python\python36\lib\site-packages\pgmpy\inference\dbn_inference.py", line 70, in __init__ start_markov_model.add_edges_from(combinations_slice_0[0])
File "c:\users\a5767\appdata\local\programs\python\python36\lib\site-packages\pgmpy\base\UndirectedGraph.py", line 252, in add_edges_from self.add_edge(edge[0], edge[1])
File "c:\users\a5767\appdata\local\programs\python\python36\lib\site-packages\pgmpy\models\MarkovModel.py", line 105, in add_edge raise ValueError("Self loops are not allowed")
ValueError: Self loops are not allowed

Thank you very much for someone who can help me with this error.

[FelipeGiro]

I think I got the problem.

The issue is related to the function get_interference_nodes of DynamicBayesianNetwork model.

If we run dbn.get_inter_edges() you get [(a_0, a_1), (a_0, b_1)].Which is in, fact, correct. Although, when using dbn.get_interface_nodes(0) and dbn.get_interface_nodes(0), we get [a_0, a_0] and [a_1, a_1], and then the program accuses wrongly a self loop.

Going in deep to get_interference_nodes return value:

[DynamicNode(edge[0][0], time_slice) for edge in self.get_inter_edges()]

It gets the first letter of the first node of a edge and attach the input time_slice. This can work well when the edges are "parallels" (for example [(D_0, D_1), (I_0, I_1)], resulting in [D_0, I_0] and [D_1, I_1]), but not when you have the interception starting from the same node.

One possible solution is to set the inputs of DynamicNode directly from edge.

[FelipeGiro]

@576774572 Keep in mind that perhaps your model is not completely correct, despite the bug. I would say that it would be more correct like this:

dbn = DBN()
dbn.add_edges_from([(('a', 0),('a', 1)), (('a', 0),('b', 1)), (('a', 1),('b', 1)), (('a', 0),('b', 0))])
a_0=TabularCPD(('a', 0), 2, [[0.5],[0.5]])
a_1=TabularCPD(('a', 1), 2, [[0.3,0.7],[0.7,0.3]],[('a',0)],[2])
b_0=TabularCPD(('b', 0), 2, [[0.3,0.7],[0.7,0.3]],[('a',0)],[2])
b_1=TabularCPD(('b', 1), 2, [[0.9,0.5,0.9,0.7],[0.1,0.5,0.1,0.3]],
               [('a',0), ('a',1)],[2,2])

dbn.add_cpds(a_0, a_1, b_0, b_1)

Because first, all 4 nodes in all time slices are represented by a CPD; second, all 4 edges are represented; and third, the CDP of b_1 receive two edges, from a_0 and a_1, having a "longer" table.

[ankurankan]

This doesn't seem to be fixed yet. Another example:

import numpy as np
from scipy.special import softmax
from pgmpy.models import DynamicBayesianNetwork
from pgmpy.inference import DBNInference
from pgmpy.factors.discrete import TabularCPD

edges=[
( ('A', 0) , ('A', 1) ),
( ('A', 0) , ('B', 1) ),
( ('B', 0) , ('A', 1) ),
( ('B', 0) , ('B', 1) ),
( ('L', 0) , ('A', 0) ),
( ('L', 0) , ('B', 0) ),
( ('L', 0) , ('L', 1) ),
( ('L', 1) , ('A', 1) ),
( ('L', 1) , ('B', 1) ),
]

bnet=DynamicBayesianNetwork(edges)#, latents=set('L'))

bnet_cpds = []
bnet_cpds.append(
    TabularCPD(('A',0), 2, softmax(np.ones((2,2)), axis=0), evidence=[('L', 0)], evidence_card=[2]))
bnet_cpds.append(
    TabularCPD(('B',0), 2, softmax(np.ones((2,2)), axis=0), evidence=[('L', 0)], evidence_card=[2]))
bnet_cpds.append(
    TabularCPD(('L',0), 2, softmax(np.ones((2,1)), axis=0)))
bnet_cpds.append(
    TabularCPD(('A', 1), 2, softmax(np.ones((2,8)), axis=0), evidence=[('A', 0), ('B', 0), ('L', 1)], evidence_card=[2,2,2]))
bnet_cpds.append(
    TabularCPD(('B', 1), 2, softmax(np.ones((2,8)), axis=0), evidence=[('A', 0), ('B', 0), ('L', 1)], evidence_card=[2,2,2]))
bnet_cpds.append(
    TabularCPD(('L',1), 2, softmax(np.ones((2,2)), axis=0), evidence=[('L', 0)], evidence_card=[2]))

bnet.add_cpds(*bnet_cpds)

ie=DBNInference(bnet)
# Error: Self loops are not allowed

[FelipeGiro]

Ok, my tests were not so good to check the networks.

I think the issue is linked to this part, specially with combinations_slice_0 and combinations_slice_1.

pgmpy/pgmpy/inference/dbn_inference.py

Lines 67 to 73 in ec7e5b4

	combinations_slice_0 = tee(combinations(self.interface_nodes_0, 2), 2)
	combinations_slice_1 = combinations(self.interface_nodes_1, 2)
	start_markov_model.add_edges_from(combinations_slice_0[0])
	one_and_half_markov_model.add_edges_from(
	chain(combinations_slice_0[1], combinations_slice_1)
	)

@ankurankan You can assign me to this issue to remember me to solve it in the following weeks, if you want.