def _predict_tree(self, session, model, sent, debug = False):
""" Generate a greedy decoding parsing of a sent object with list of [word, tag]
Return: dep parse tree, dep-label is in its id
"""
target_num = len(self.rev_arc_labels)
#print ("target number is %d" % target_num)
config = Configuration(sent) # create a parsing transition arc-standard system configuration, see (chen and manning.2014)
while not config.is_terminal():
features = transition_system.get_features(config, self.feature_tpl) # 1-D list
X = np.array([features]) # convert 1-D list to ndarray size [1, dim]
Y = np.array([[0] * target_num])
# generate greedy prediction of the next_arc_id
fetches = [model.loss, model.logit] # fetch out prediction logit
feed_dict = {}
feed_dict[model.X] = X
feed_dict[model.Y] = Y # dummy input y: 1D list of shape [, target_num]
_, logit = session.run(fetches, feed_dict) # not running eval_op, just for prediction
pred_next_arc_id = int(np.argmax(logit)) # prediction of next arc_idx of 1 of (2*Nl +1)
pred_next_arc = self.rev_arc_labels[pred_next_arc_id] # 5-> L(2) idx -> L(label_id)
config.step(pred_next_arc) # Configuration Take One Step
if (debug):
print ("DEBUG: next arc idx is %d and next arc is %s" % (pred_next_arc_id, pred_next_arc))
# When config is terminal, return the final dependence trees object
dep_tree = config.tree
return dep_tree
def run_beam_search(session, model, eval_op, ):
#initialize a new empty beam with size K
bs = BeamSearch(K)
#Get the features of one sent and its gold_sequence
features = [] # list of size T
gold_sequence = [] # list of size T, int, Action Sequence
#Compare with golden sequence
config = Configuration(sent)
step = 0
max_step = 100
while step < max_step:
step += 1
all_hpys = []
# top K hyps in current step in Beam
for h in bs.hyps:
config = Configuration(sent) # new empty config
config.step(h.tokens) # repeat the path in hyps
feat = transition_system.get_features(config)
top_k = bs.search(session, model, feat) # topK of current hpys
all_hpys.append(top_k)
# all_hpys: top_k K*K hyps
cur_hyps = get_topk(all_hpys) # current step top K
# check if golden path is in cur_hyps
if (contain()):
continue
else:
# run update function
model.update(X, Y)
# end of the beam search, update network
model.update(X, Y)
def predict(session, model, sent):
""" Generate a greedy decoding parsing of a sent object with list of [word, tag]
"""
config = Configuration(sent)
while not config.is_terminal():
features = transition_system.get_features(config) # 1-D list
# generate greedy prediction of the next_arc_id
fetches = [model.loss, model.logit] # fetch out prediction logit
feed_dict = {}
feed_dict[model.X] = features
feed_dict[model.Y] = [0] * model.target_num # dummy input y: 1D list of shape [, target_num]
_, logit = session.run(fetches, feed_dict) # not running eval_op, just for prediction
pred_next_arc_id = int(np.argmax(logits)) # prediction of next arc_idx of 1 of (2*Nl +1)
pred_next_arc = []
config.step(pred_next_arc) # Configuration Take One Step
print ("next arc idx is %d and next arc is %s" % (pred_next_arc_id, pred_next_arc))
next_arc = get_next_arc(config, tree) # number
next_arc_id = arc_labels[next_arc]