src

exp2/w2v_transE/tester.py

@@ -1,3 +1,4 @@
+from pickle import FALSE, TRUE
 import numpy as np
 import codecs
 import operator
@@ -56,12 +57,25 @@ def __init__(self,entity_dict,relation_dict,test_triple,train_triple,isFit = Tru
  self.hits5 = 0
  self.mean_rank = 0
 
- def rank(self, output_path):
- hits = 0
- step = 1
- f = open(output_path + 'tail_predict.txt', 'w')
+ def rank(self, output_path, re_continue, re_hits):
+ if re_continue:
+ f = open(output_path + 'tail_predict.txt', 'r')
+ step = len(f.readlines())
+ hits = re_hits
+ f.close()
+ f = open(output_path + 'tail_predict.txt', 'a')
+ past = step
+ else:
+ hits = 0
+ step = 0
+ f = open(output_path + 'tail_predict.txt', 'w')
+ past = 0
+
 
  for triple in self.test_triple:
+ if past > 0:
+ past -= 1
+ continue
  rank_tail_dict = {}
 
  for entity in self.entity_dict.keys():
@@ -100,11 +114,11 @@ def rank(self, output_path):
  first_hit = False
 
  step += 1
- if step % 10 == 0:
+ if step % 20 == 0:
  print("step ", step, ", hits ",hits, ', rate: ',hits/step)
  print()
 
- self.hits5 = hits / (2*len(self.test_triple))
+ self.hits5 = hits / len(self.test_triple)
 
 
 if __name__ == '__main__':
@@ -116,7 +130,7 @@ def rank(self, output_path):
 
 
  test = Test(entity_dict,relation_dict,test_triple,train_triple,isFit=True)
- test.rank('./output/')
+ test.rank('./output/',re_continue=False, re_hits=0)
  print("entity hits@5: ", test.hits5)
 
 

exp2/w2v_transE/transE.py

@@ -41,7 +41,7 @@ def distanceL1(h,r,t):
  return np.sum(np.fabs(h+r-t))
 
 class TransE:
- def __init__(self,entity_set, relation_set, triple_list, ent_rel_sim_path, use_w2v = False, nbatch = 400,
+ def __init__(self,entity_set, relation_set, triple_list, ent_rel_sim_path, use_w2v = False,mu = 1, nbatch = 400,
  embedding_dim=100, learning_rate=0.01, margin=1,L1=True):
  self.embedding_dim = embedding_dim
  self.learning_rate = learning_rate
@@ -50,6 +50,7 @@ def __init__(self,entity_set, relation_set, triple_list, ent_rel_sim_path, use_w
  self.relation = relation_set
  self.triple_list = triple_list
  self.L1=L1
+ self.mu = mu
 
  self.nbatch = nbatch
  self.use_w2v = use_w2v
@@ -176,8 +177,10 @@ def update_embeddings(self, Tbatch):
  t_corrupt = self.entity[corrupted_triple[1]]
 
  if self.use_w2v:
- w2v_correct = self.ent_rel_sim[int(triple[0])][int(triple[2])] * self.ent_rel_sim[int(triple[1])][int(triple[2])]
- w2v_corrupt = self.ent_rel_sim[int(corrupted_triple[0])][int(triple[2])] * self.ent_rel_sim[int(corrupted_triple[1])][int(triple[2])]
+ w2v_correct = self.ent_rel_sim[int(triple[0])][int(triple[2])] \
+ * self.ent_rel_sim[int(triple[1])][int(triple[2])]
+ w2v_corrupt = self.ent_rel_sim[int(corrupted_triple[0])][int(triple[2])] \
+ * self.ent_rel_sim[int(corrupted_triple[1])][int(triple[2])]
  w2v_revise = w2v_correct - w2v_corrupt
 
  if self.L1:
@@ -240,7 +243,7 @@ def update_embeddings(self, Tbatch):
  self.relation = copy_relation
 
  def hinge_loss(self,dist_correct,dist_corrupt, w2v_revise):
- return max(0,dist_correct-dist_corrupt+self.margin + w2v_revise)
+ return max(0,dist_correct-dist_corrupt+self.margin + self.mu*w2v_revise)
 
 
 if __name__=='__main__':
@@ -249,7 +252,7 @@ def hinge_loss(self,dist_correct,dist_corrupt, w2v_revise):
  print("load file...")
  print("Complete load. entity : %d , relation : %d , triple : %d" % (len(entity_set),len(relation_set),len(triple_list)))
 
- transE = TransE(entity_set, relation_set, triple_list, './output/ent_rel_sim.npy', use_w2v = False, nbatch=100, embedding_dim=200, learning_rate=0.01, margin=5,L1=True)
- # transE.emb_initialize()
- transE.reload('./temp/')
- transE.train(epochs=0, temp_path='./temp/')
+ transE = TransE(entity_set, relation_set, triple_list, './output/ent_rel_sim.npy', use_w2v = True, mu=12,nbatch=100, embedding_dim=200, learning_rate=0.001, margin=4.0,L1=True)
+ transE.emb_initialize()
+ # transE.reload('./temp/')
+ transE.train(epochs=100, temp_path='./temp/')