-
-
Save rtbs-dev/1349864212b25cce91dbe5e336d794b4 to your computer and use it in GitHub Desktop.
Added example of programmatically instantiating the PyMC3 random variable objects using NetworkX dicts. In theory, one could now "loop-over" an existing network and build up a pymc3 model to do inference.
This is really cool. Thanks!
How do you make a query ? For example P(C3 |C2 = 12)
I've got the following error message when I tried the Networkx with PyMC3:
TypeError Traceback (most recent call last)
in
8 BN.node['D2']['dist_obj'] = pm.Categorical('D2',p=BN.node['D2']['prob'])
9
---> 10 BN.node['C1']['dist_obj'] = pm.Normal('C1', mu=10+4*gpm(BN,'C1'),
11 tau = (1./2)*2)
12 BN.node['C2']['dist_obj'] = pm.Normal('C2', mu=6+2(gpm(BN,'C2'))**2,
in gpm(BN, node, num)
1 # utily to quickly access the parent nodes' pymc3 objects
2 def gpm(BN, node, num=0):
----> 3 return BN.node[BN.predecessors(node)[num]]['dist_obj']
4
5 with pm.Model() as mod2:
TypeError: 'dict_keyiterator' object is not subscriptable
I'm currently using networkx 2.2, pymc3 3.5, and python 3.5 on OS X
I got the following error message:
Traceback (most recent call last):
File "./pmml_test.py", line 126, in
BN.node['C1']['dist_obj'] = pm.Normal('C1', mu=10+4*gpm(BN,'C1'),tau = (1./2)**2)
File "./pmml_test.py", line 118, in gpm
return BN.node[BN.predecessors(node)[num]]['dist_obj']
TypeError: 'dictionary-keyiterator' object has no attribute 'getitem'
I'm currently using networkx 2.2, pymc3 3.5, and python 2.7 on OS X. Can you help me?
I got the following error message:
Traceback (most recent call last):
File "./pmml_test.py", line 126, in
BN.node['C1']['dist_obj'] = pm.Normal('C1', mu=10+4*gpm(BN,'C1'),tau = (1./2)**2)
File "./pmml_test.py", line 118, in gpm
return BN.node[BN.predecessors(node)[num]]['dist_obj']
TypeError: 'dictionary-keyiterator' object has no attribute 'getitem'I'm currently using networkx 2.2, pymc3 3.5, and python 2.7 on OS X. Can you help me?
Here, he is using network 1.x. Here is the Migration guide :https://networkx.github.io/documentation/stable/release/migration_guide_from_1.x_to_2.0.html
In :
"# utily to quickly access the parent nodes' pymc3 objects
def gpm(BN, node, num=0):
return BN.node[BN.predecessors(node)[num]]['dist_obj'] "
you have to change BN.predecessors(node) to list(BN.predecessors(node)).
That is why you might get that keyiterator error.
Edit: aslo, you might want to change BN.node to BN.nodes
Hope it helps!
I've got the following error message when I tried the Networkx with PyMC3:
TypeError Traceback (most recent call last)
in
8 BN.node['D2']['dist_obj'] = pm.Categorical('D2',p=BN.node['D2']['prob'])
9
---> 10 BN.node['C1']['dist_obj'] = pm.Normal('C1', mu=10+4gpm(BN,'C1'),
11 tau = (1./2)2) 12 BN.node['C2']['dist_obj'] = pm.Normal('C2', mu=6+2(gpm(BN,'C2'))**2,in gpm(BN, node, num)
1 # utily to quickly access the parent nodes' pymc3 objects
2 def gpm(BN, node, num=0):
----> 3 return BN.node[BN.predecessors(node)[num]]['dist_obj']
4
5 with pm.Model() as mod2:TypeError: 'dict_keyiterator' object is not subscriptable
I'm currently using networkx 2.2, pymc3 3.5, and python 3.5 on OS X
Here, he is using network 1.x. Here is the Migration guide :https://networkx.github.io/documentation/stable/release/migration_guide_from_1.x_to_2.0.html
In :
"# utily to quickly access the parent nodes' pymc3 objects
def gpm(BN, node, num=0):
return BN.node[BN.predecessors(node)[num]]['dist_obj'] "
you have to change BN.predecessors(node) to list(BN.predecessors(node)).
That is why you might get that keyiterator error.
Edit: aslo, you might want to change BN.node to BN.nodes
Hope it helps!
How do you make a query such as: P(C1_a<C1<C1_b | C3_a<C3<C3_b, D4 = 1)?
I am getting the following error:
Traceback (most recent call last):
File "", line 30, in
BN.edge['C3']['D4']['bin']], observed=BN.node['D4']['observe'])
File "C:\Users\Andrea\Anaconda3\lib\site-packages\theano\tensor\var.py", line 544, in getitem
theano.tensor.subtensor.Subtensor.convert(arg)
File "C:\Users\Andrea\Anaconda3\lib\site-packages\theano\tensor\subtensor.py", line 354, in convert
raise TypeError("Expected an integer")
TypeError: Expected an integer
Any help, please? Thanks!
I get the same error with pymc3 3.6 and python3.
Traceback (most recent call last):
File "pmml_test.py", line 147, in
[gpm(BN,'D4', num=1),BN.adj['C3']['D4']['bin']], observed=BN.node['D4']['observe'])
File "/Users/fangwq/Library/Python/3.6/lib/python/site-packages/theano/tensor/var.py", line 544, in getitem
theano.tensor.subtensor.Subtensor.convert(arg)
File "/Users/fangwq/Library/Python/3.6/lib/python/site-packages/theano/tensor/subtensor.py", line 354, in convert
raise TypeError("Expected an integer")
TypeError: Expected an integer
Any help, please? Thanks!
Thanks to all above!
I have about four hours experience in pymc3, so the above comments are much appreciated.
Using networkx 2.3 and pyMC3 3.7, cpu on Windows.
Fixed the dict_keyiterator problem.
32 BN.edge['C3']['D4']['bin'] = T.switch(T.lt(gpm(BN,'D4'),9), 0,
---> 33 T.switch(T.gt(gpm(BN,'D4'), 9) & T.lt(gpm(BN,'D4'),11), 1, 2))
AttributeError: 'DiGraph' object has no attribute 'edge'
When print(BN.__dict__) I see no edge attribute.
The edges seem to be embedded in _adj, _pred and _succ.
BN = nx.DiGraph()
BN.add_node('D1', dtype='Discrete', prob=d1_prob, pos=(2, 4))
BN.add_node('D2', dtype='Discrete', prob=d2_prob, pos=(4, 4))
BN.add_node('C1', dtype='Continuous', mu=c1_mu, sd=c1_sd, pos=(1,3))
BN.add_edge('D1', 'C1')
print(BN.__dict__)
-----------------------------------------------
{
'graph_attr_dict_factory': <class 'dict'>, 'node_dict_factory': <class 'dict'>,
'node_attr_dict_factory': <class 'dict'>, 'adjlist_outer_dict_factory': <class 'dict'>,
'adjlist_inner_dict_factory': <class 'dict'>, 'edge_attr_dict_factory': <class 'dict'>,
'graph': {},
'_node': {
'D1': {'dtype': 'Discrete', 'prob': array([0.3, 0.7]), 'pos': (2, 4)},
'D2': {'dtype': 'Discrete', 'prob': array([0.6, 0.3, 0.1]), 'pos': (4, 4)},
'C1': {'dtype': 'Continuous', 'mu': array([10, 14]), 'sd': array([2, 2]), 'pos': (1, 3)}
},
'_adj': {'D1': {'C1': {}}, 'D2': {}, 'C1': {}},
'_pred': {'D1': {}, 'D2': {}, 'C1': {'D1': {}}},
'_succ': {'D1': {'C1': {}}, 'D2': {}, 'C1': {}}
}
D1 is a predecessor for C1.
C1 is a successor for D1.
But C1 is also adjacent to D1?
My bigger issue is what to do with the error on line 32/33 above.
I have no idea how to handle this. Any help will be appreciated!
Wow, I had no idea there were comments here! Thanks to all for checking out the example. For those of you with issues/ideas (especially networkx v2 problems), check out the repository that automates a lot of the underlying mechanical/boilerplate code in this example. This notebook was really just a proof-of-concept for that repository.
I haven't gone back to the code for a while, but if there's interest I would be happy to update things to newer pymc3 features. Please see if your above issues persist in the more mature repository. Thanks!
@NelisW certainly, and like I said, feel free to open an issue if something works unexpectedly on that repo (as I said, it's been a while).
If you are having issues with BN.edges['C3']['D4']['bin'], substitute for BN.edges['C3','D4']['bin'].
As described in the migration guide mentioned before
G.edge have been removed in favor of using G.nodes[n] and G.edges[u, v]
Do you have any example for dynamic BN simulated on PYMC3?
Updated to use tabular look-up for pre-defined categorical composition nodes. Removed need for
theano.tensor.switch()in most cases, using a shared numpy slice on the table instead. Runs much faster.Still uses
T.switch()for splitting the continuous node C3