cbp.graph

cbp.graph.base_graph

class cbp.graph.base_graph.BaseGraph(silent=True, epsilon=1, coef_policy=<function bp_policy>, config=BaseConfig())

Bases: object

add_factornode(factornode)

add one factor node to the graph Do the following tasks

• add node to the recorders

• set connections

• set parent relation

Parameters

factornode (FactorNode) – one factor node

Returns

name of factor node

Return type

str

add_varnode(node)

add one ~cbp.node.VarNode to this graph, idx follow the increasing order

Parameters

node (VarNode) – one VarNode

Returns

name of varnode

Return type

str

bake()

cal_bethe(margin)

calculate bethe energy

Parameters

margin (dict) – node_name : margin

Returns

KL divergence between expoert joint dist and p_graph

Return type

float

copy_bp_initialization(another_graph)

copy message setup from the another graph has same structure

Parameters

another_graph (BaseGraph) – another graph which close to the optimal point

delete_node(name_str)

delete node from graph, needs to check following

• judge type, Factor or Variable

• delete from the various recorders

• clear connections

Parameters

name_str ([type]) – [description]

Raises

RuntimeError – [description]

exact_marginal()

export_convergence_marginals()

export the marginal for variable nodes and factor nodes

Returns

{node.key:node.marginal}

Return type

dict

export_marginals()

export the marginal for variable nodes

Returns

{node.key:node.marginal}

Return type

dict

export_sinkhorn()

first_belief_propagation()

get_node(name_str)

init_node_recorder()

plot(png_name='file.png')

plot the graph through graphviz

• red: constrained variable node

• blue: free variable node

• green: factor

Parameters

png_name (str, optional) – name of figure, defaults to ‘file.png’

Raises

ValueError – [description]

pmf()

output the probability mass matrix through brutal-force methods

Returns

joint probability mass matrix

Return type

ndarray

set_node(node_name, potential=None, isconstrained=None)

change node property 1. check whether or not in recorder 2. change potential easily[this may a duplicate function] 3. change isconstrained if possible, delete from recorder

Parameters

• node_name ([type]) – [description]

• potential ([type], optional) – [description], defaults to None

• isconstrained ([type], optional) – [description], defaults to None

Raises

RuntimeError – [description]

sinkhorn(max_iter=5000000, tolerance=1e-05)

sinkhorn_update(tilde_c)

tree_bp()

run classical belief propagation on a tree graph, only need forward and backward

• add attr: is_send_forward: begin send forward false, after forward

before backward true, after backward false

Raises

RuntimeError – Only works for the tree graph, loopy graph does

not work, root node not decided

cbp.graph.coef_policy

cbp.graph.coef_policy.avg_policy(node, node_map)

cbp.graph.coef_policy.bp_policy(node, node_map)

cbp.graph.coef_policy.crazy_policy(node, node_map)

cbp.graph.coef_policy.factor_policy(node, node_map)

cbp.graph.graph_model

class cbp.graph.graph_model.GraphModel(silent=True, epsilon=1, coef_policy=<function bp_policy>, config=BaseConfig())

Bases: cbp.graph.base_graph.BaseGraph

bake()

engine_loop(engine_fun, max_iter=5000000, tolerance=0.01, error_fun=None, isoutput=False)

its_next_looplink()

itsbp()

run sinkhorn or iterative scaling inference

Returns

[description]

Return type

[type]

itsbp_outer_loop()

norm_product_bp(max_iter=5000000, tolerance=1e-05, error_fun=None)

parallel_message(run_constrained=True)

run_bp()

run_cnp()

cbp.graph.graph_utils

cbp.graph.graph_utils.cal_marginal_from_tensor(prob_tensor, varnode_list)

cbp.graph.graph_utils.find_link(node_a, node_b)

cbp.graph.graph_utils.get_node2root(node)

cbp.graph.graph_utils.itsbp_inner_loop(loop_link, is_silent)