def inference_subgraph(self, var_id):
'''推理子图
输入: 用于实体识别则var_id是一个变量id,用于情感分类， 则var_id是k个变量的集合
输出：推理出的此隐变量的概率，此概率需写回variables
'''
learn = 1000
ns = numbskull.NumbSkull(n_inference_epoch=10,
n_learning_epoch=learn,
quiet=True,
learn_non_evidence=True,
stepsize=0.0001,
burn_in=10,
decay=0.001**(1.0 / learn),
regularization=1,
reg_param=0.01)
weight, variable, factor, fmap, domain_mask, edges_num, var_map, feature_map_weight = self.construct_subgraph(
var_id)
subgraph = weight, variable, factor, fmap, domain_mask, edges_num
ns.loadFactorGraph(*subgraph)
# 因子图参数学习
ns.learning()
logging.info("subgraph learning finished")
# 因子图推理
# 参数学习完成后将weight的isfixed属性置为true
for w in ns.factorGraphs[0].weight:
w["isFixed"] = True
ns.learn_non_evidence = False
ns.inference()
logging.info("subgraph inference finished")
# 写回概率到self.variables
if type(var_id) == set() or type(var_id) == list():
for id in var_id:
self.variables[var_id]['probability'] = ns.factorGraphs[
0].marginals[var_map[var_id]]
elif type(var_id) == int:
self.variables[var_id]['probability'] = ns.factorGraphs[
0].marginals[var_map[var_id]]
else:
print('参数错误，var_id应为int或者set和list')
logging.info("inferenced probability recored")
fmp_index += 1
factor_index += 1
# 将定义好的因子图保存为factorGraph.pkl文件，以备后期验证使用
#factorGraph = weight, variable, factor, fmap, domain_mask, edges
# with open('factorGraph_original.pkl', 'wb') as f: #原始的图
# pickle.dump(factorGraph, f)
return weight, variable, factor, fmap, domain_mask, edges
#1.已知所有参数，推理所有变量
learn = 1000
ns = numbskull.NumbSkull(n_inference_epoch=10,
n_learning_epoch=learn,
quiet=True,
learn_non_evidence=True,
stepsize=0.0001,
burn_in=10,
decay=0.001**(1.0 / learn),
regularization=1,
reg_param=0.01)
fg = create_fg()
ns.loadFactorGraph(*fg)
ns.inference()
print("First Inference Finished!")
for i in range(1000):
if ns.factorGraphs[0].marginals[
i] > 0.5: #推理出来的变量是概率值,保存在factorGraphs[0].marginals，现在写回到fg
fg[1][i]['initialValue'] = 1
else:
fg[1][i]['initialValue'] = 0
fmap = np.zeros(edges, FactorToVar)
domain_mask = np.zeros(variables, np.bool)
weight[0]["isFixed"] = True
weight[0]["initialValue"] = 1
for i in range(variables):
variable[i]["isEvidence"] = 0
variable[i]["initialValue"] = 0
variable[i]["dataType"] = 0
variable[i]["cardinality"] = 2
factor[0]["factorFunction"] = value
factor[0]["weightId"] = 0
factor[0]["featureValue"] = 1
factor[0]["arity"] = variables
factor[0]["ftv_offset"] = 0
for i in range(variables):
fmap[i]["vid"] = i
ns = numbskull.NumbSkull(n_inference_epoch=100,
n_learning_epoch=100,
quiet=True)
ns.loadFactorGraph(weight, variable, factor, fmap, domain_mask, edges)
ns.learning()
ns.inference()
print(ns.factorGraphs[0].count)
def init_numbskull(self, argv):
"""TODO."""
self.args = self.parse_args(argv)
self.ns = numbskull.NumbSkull(**vars(self.args))
def prep_local_numbskull(self):
"""TODO."""
self.ns = numbskull.NumbSkull(**vars(self.args))
factor[factor_index]["featureValue"] = 1
factor[factor_index]["arity"] = 2
factor[factor_index]["ftv_offset"] = copy * (1 + 2 * n) + 1 + 2 * i
fmap_index = copy * (1 + 2 * n) + 1 + 2 * i
fmap[fmap_index]["vid"] = copy * (1 + n) # y
fmap[fmap_index]["vid"] = copy * (1 + n) + i + 1 # labeling func i
return weight, variable, factor, fmap, domain_mask, edges
learn = 100
ns = numbskull.NumbSkull(n_inference_epoch=100,
n_learning_epoch=learn,
quiet=True,
learn_non_evidence=True,
stepsize=0.01,
burn_in=100,
decay=0.001**(1.0 / learn),
reg_param=0.15)
prior = 0
accuracy = [1, 0.5]
abstain = [0, 0, 0]
copies = 10
fg = create_fg(prior, accuracy, abstain, copies)
print("weight")
print(fg[0])
print()
print("variable")
print(fg[1])
