-
Notifications
You must be signed in to change notification settings - Fork 14
/
trainer.py
1873 lines (1588 loc) · 105 KB
/
trainer.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
from __future__ import absolute_import
from __future__ import division
from tqdm import tqdm
import json
import time
import os
import logging
import numpy as np
import tensorflow as tf
# Export PYTHONPATH so that 'rl_code' folder can be regarded as a package
import sys
sys.path.append('/data/base2/Bio-Relation-Extract/BioRE-master/rl+PCNN/Joint')
from rl_code.model.agent import Agent, AgentTarget
from rl_code.options import read_options
from rl_code.model.environment import env
import codecs
from collections import defaultdict
import gc
import resource
import sys
from rl_code.model.baseline import ReactiveBaseline
from rl_code.model.nell_eval import nell_eval
from scipy.misc import logsumexp as lse
logger = logging.getLogger()
logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)
import random
import numpy as np
import pandas as pd
import pickle
class priorityDictionary(dict):
def __init__(self):
self.__heap = []
dict.__init__(self)
def smallest(self):
'''Find smallest item after removing deleted items from heap.'''
if len(self) == 0:
print("smallest of empty priorityDictionary")
heap = self.__heap
while heap[0][1] not in self or self[heap[0][1]] != heap[0][0]:
lastItem = heap.pop()
insertionPoint = 0
while 1:
smallChild = 2 * insertionPoint + 1
if smallChild + 1 < len(heap) and \
heap[smallChild] > heap[smallChild + 1]:
smallChild += 1
if smallChild >= len(heap) or lastItem <= heap[smallChild]:
heap[insertionPoint] = lastItem
break
heap[insertionPoint] = heap[smallChild]
insertionPoint = smallChild
return heap[0][1]
def __iter__(self):
'''Create destructive sorted iterator of priorityDictionary.'''
def iterfn():
while len(self) > 0:
x = self.smallest()
yield x
del self[x]
return iterfn()
def __setitem__(self, key, val):
dict.__setitem__(self, key, val)
heap = self.__heap
if len(heap) > 2 * len(self):
self.__heap = [(v, k) for k, v in self.iteritems()]
self.__heap.sort() # builtin sort likely faster than O(n) heapify
else:
newPair = (val, key)
insertionPoint = len(heap)
heap.append(None)
while insertionPoint > 0 and \
newPair < heap[(insertionPoint - 1) // 2]:
heap[insertionPoint] = heap[(insertionPoint - 1) // 2]
insertionPoint = (insertionPoint - 1) // 2
heap[insertionPoint] = newPair
def setdefault(self, key, val):
'''Reimplement setdefault to call our customized __setitem__.'''
if key not in self:
self[key] = val
return self[key]
def Dijkstra(G, start, end=None):
D = {} # dictionary of final distances
P = {} # dictionary of predecessors
Q = priorityDictionary() # est.dist. of non-final vert.
Q[start] = 0
for v in Q:
D[v] = Q[v]
if v == end: break
if v not in G: # this node is not a head node in the graph
continue
for w in G[v]:
vwLength = D[v] + G[v][w]
if w in D:
if vwLength < D[w]:
print("Dijkstra: found better path to already-final vertex")
elif w not in Q or vwLength < Q[w]:
Q[w] = vwLength
P[w] = v
return (D, P)
def shortestPath(G, start, end):
"""
Find a single shortest path from the given start vertex
to the given end vertex.
The input has the same conventions as Dijkstra().
The output is a list of the vertices in order along
the shortest path.
"""
D, P = Dijkstra(G, start, end)
if end in D:
distance = D[end]
Path = []
while 1:
Path.append(end)
if end == start: break
if end not in P:
break
end = P[end]
Path.reverse()
return distance, Path
else:
return 'nan', []
def constructGraph(train, edge_train):
Graph = {}
print("Constructing graph from training set...")
for index, row in tqdm(train.iterrows()):
if row['e1'] not in Graph:
Graph[row['e1']] = {}
if row['e1'] == row['e2']:
pass
else:
Graph[row['e1']][row['e2']] = 1
else:
if row['e2'] not in Graph[row['e1']]:
if row['e1'] == row['e2']:
pass
else:
Graph[row['e1']][row['e2']] = 1
print("Constructing graph from added edges training set...")
for index, row in tqdm(edge_train.iterrows()):
if row['e1'] not in Graph:
Graph[row['e1']] = {}
if row['e1'] == row['e2']:
pass
else:
Graph[row['e1']][row['e2']] = 0.5
else:
if row['e2'] not in Graph[row['e1']]:
if row['e1'] == row['e2']:
pass
else:
Graph[row['e1']][row['e2']] = 0.5
return Graph
class Memory(object):
def __init__(self, memory_size=4000):
self.memory = {}
self.memory_size = memory_size
def insert(self, key, extend_value):
# TODO: replace old memory with new memory if it's full
if key not in self.memory:
self.memory[key] = []
self.memory[key].extend(extend_value)
random.shuffle(self.memory[key])
self.memory[key] = self.memory[key][:self.memory_size]
def sample(self, batch_size, keys):
# TODO: if memory size < batch_size, deal with it!
try:
indices = list(range(len(self.memory[keys[0]])))
random.shuffle(indices)
indices = indices[:batch_size]
res = [np.array(self.memory[key])[indices] for key in keys]
return res
except:
print("Wrong Input!")
class Trainer(object):
def __init__(self, params):
# transfer parameters to self
for key, val in params.items(): setattr(self, key, val);
self.params = params
self.save_path = None
self.train_environment = env(params, 'train')
self.dev_test_environment = env(params, 'dev')
self.test_test_environment = env(params, 'test') # test set: path_test.txt
self.test_environment = self.dev_test_environment
self.rev_relation_vocab = self.train_environment.grapher.rev_relation_vocab
self.rev_entity_vocab = self.train_environment.grapher.rev_entity_vocab
self.max_hits_at_10 = 0
self.ePAD = self.entity_vocab['PAD']
self.rPAD = self.relation_vocab['PAD']
# optimize
self.baseline = ReactiveBaseline(l=self.Lambda)
self.optimizer = tf.train.AdamOptimizer(self.learning_rate)
def sample_PCNN(self, framework):
indices = np.random.randint(len(self.pcnn_experience.memory['triples']), size=self.params['pcnn_batch_size'])
sampled = np.array(self.pcnn_experience.memory['triples'])[indices]
# return pcnn index for e1, e2, rn
# relation_map = framework.test_data_loader.rel2id
# entity_map = framework.test_data_loader.word2id
#
# def function(row):
# return [entity_map[row[0]], entity_map[row[1]], relation_map[row[2]]]
#
# sampled = np.apply_along_axis(function, 1, sampled)
return sampled
def sample_GFAW(self, ratio):
"""
:param batch_size: batch size in GFAW
:return: feed_dict list for every step in this path
"""
# TODO: 1. sample 数据
indices_pos = np.random.randint(len(self.pos_experience.memory['entity_path']), size=round((1-ratio)*self.num_rollouts*self.batch_size))
indices_neg = np.random.randint(len(self.neg_experience.memory['entity_path']), size=round(ratio*self.num_rollouts*self.batch_size))
path_rewards_pos = np.array(self.pos_experience.memory['path_rewards'])[indices_pos] # np.repeat( , self.num_rollouts)
path_rewards_neg = np.array(self.neg_experience.memory['path_rewards'])[indices_neg] # np.repeat( , self.num_rollouts)
path_rewards = np.concatenate((path_rewards_pos, path_rewards_neg), axis=0)
state_rewards_pos = np.array(self.pos_experience.memory['state_rewards'])[indices_pos] # np.repeat( , self.num_rollouts, axis=0)
state_rewards_neg = np.array(self.neg_experience.memory['state_rewards'])[indices_neg] # np.repeat( , self.num_rollouts, axis=0)
state_rewards = np.concatenate((state_rewards_pos, state_rewards_neg), axis=0)
# TODO: 2. 放入 feed_dict
path_length = len(self.pos_experience.memory['relation_path'][0])
feed_dict = [{} for _ in range(path_length)]
feed_dict[0][self.first_state_of_test] = False
query_relation_pos = np.array(self.pos_experience.memory['query_relation'])[indices_pos] # np.repeat( , self.num_rollouts)
query_relation_neg = np.array(self.neg_experience.memory['query_relation'])[indices_neg] # np.repeat( , self.num_rollouts)
query_relation = np.concatenate((query_relation_pos, query_relation_neg), axis=0)
feed_dict[0][self.query_relation] = query_relation.reshape(self.batch_size * self.num_rollouts)
feed_dict[0][self.range_arr] = np.arange(self.batch_size * self.num_rollouts)
for i in range(path_length):
entities_pos = np.array(self.pos_experience.memory['entity_path'])[indices_pos][:, i]
entities_neg = np.array(self.neg_experience.memory['entity_path'])[indices_neg][:, i]
entities = np.concatenate((entities_pos, entities_neg), axis=0)
entities = np.array([ety_index if ety_index < len(self.train_environment.grapher.array_store) else
self.train_environment.grapher.entity_vocab['UNK'] for ety_index in entities])
feed_dict[i][self.input_path[i]] = np.zeros(self.batch_size * self.num_rollouts) # placebo
feed_dict[i][self.candidate_relation_sequence[i]] = self.train_environment.grapher.array_store[entities][:, :, 1] # np.repeat( , self.num_rollouts, axis=0) # TODO 2.2
feed_dict[i][self.candidate_entity_sequence[i]] = self.train_environment.grapher.array_store[entities][:, :, 0] # np.repeat( , self.num_rollouts, axis=0) # TODO 2.3
feed_dict[i][self.entity_sequence[i]] = entities # np.repeat( , self.num_rollouts) # TODO 2.4
return feed_dict, path_rewards, state_rewards
def sample_pos_GFAW(self):
"""
:param batch_size: batch size in GFAW
:return: feed_dict list for every step in this path
"""
# TODO: 1. sample 数据
indices = np.random.randint(len(self.pos_experience.memory['entity_path']), size=self.batch_size * self.num_rollouts)
path_rewards = np.array(self.pos_experience.memory['path_rewards'])[indices] # np.repeat( , self.num_rollouts)
state_rewards = np.array(self.pos_experience.memory['state_rewards'])[indices] # np.repeat( , self.num_rollouts, axis=0)
# TODO: 2. 放入 feed_dict
path_length = len(self.pos_experience.memory['relation_path'][0])
feed_dict = [{} for _ in range(path_length)]
feed_dict[0][self.first_state_of_test] = False
feed_dict[0][self.query_relation] = np.array(self.pos_experience.memory['query_relation'])[indices].reshape(self.batch_size * self.num_rollouts) # np.repeat( ,self.num_rollouts) # TODO: 2.1 to update
feed_dict[0][self.range_arr] = np.arange(self.batch_size * self.num_rollouts)
for i in range(path_length):
entities = np.array(self.pos_experience.memory['entity_path'])[indices][:, i]
entities = np.array([ety_index if ety_index < len(self.train_environment.grapher.array_store) else
self.train_environment.grapher.entity_vocab['UNK'] for ety_index in entities])
feed_dict[i][self.input_path[i]] = np.zeros(self.batch_size * self.num_rollouts) # placebo
feed_dict[i][self.candidate_relation_sequence[i]] = self.train_environment.grapher.array_store[entities][:, :, 1] # np.repeat( , self.num_rollouts, axis=0) # TODO 2.2
feed_dict[i][self.candidate_entity_sequence[i]] = self.train_environment.grapher.array_store[entities][:, :, 0] # np.repeat( , self.num_rollouts, axis=0) # TODO 2.3
feed_dict[i][self.entity_sequence[i]] = entities # np.repeat( , self.num_rollouts) # TODO 2.4
return feed_dict, path_rewards, state_rewards
def target_model_setup(self):
self.agent_target = AgentTarget(self.params)
self.initialize_target()
self.initialize_weight_update()
def GFAW_Q_initialize(self, path_length):
self.Q_target = tf.placeholder(tf.float32, [None, self.max_num_actions], name='Q_target')
self.Q_eval = self.per_example_logits[i]
self.Q_loss = tf.reduce_mean(tf.squared_difference(self.Q_target, self.Q_eval))
def calc_reinforce_loss(self):
loss = tf.stack(self.per_example_loss, axis=1) # [B, T]
self.tf_baseline = self.baseline.get_baseline_value()
# self.pp = tf.Print(self.tf_baseline)
# multiply with rewards
final_reward = self.cum_discounted_reward - self.tf_baseline
# reward_std = tf.sqrt(tf.reduce_mean(tf.square(final_reward))) + 1e-5 # constant addded for numerical stability
reward_mean, reward_var = tf.nn.moments(final_reward, axes=[0, 1])
# Constant added for numerical stability
reward_std = tf.sqrt(reward_var) + 1e-6
final_reward = tf.div(final_reward - reward_mean, reward_std)
loss = tf.multiply(loss, final_reward) # [B, T]
self.loss_before_reg = loss
total_loss = tf.reduce_mean(loss) - self.decaying_beta * self.entropy_reg_loss(self.per_example_logits) # scalar
return total_loss
def entropy_reg_loss(self, all_logits):
all_logits = tf.stack(all_logits, axis=2) # [B, MAX_NUM_ACTIONS, T]
entropy_policy = - tf.reduce_mean(tf.reduce_sum(tf.multiply(tf.exp(all_logits), all_logits), axis=1)) # scalar
return entropy_policy
def initialize(self, restore=None, sess=None):
with tf.device("/gpu:0"):
self.agent = Agent(self.params)
logger.info("Creating TF graph...")
self.candidate_relation_sequence = []
self.candidate_entity_sequence = []
self.input_path = []
self.first_state_of_test = tf.placeholder(tf.bool, name="is_first_state_of_test")
self.query_relation = tf.placeholder(tf.int32, [None], name="query_relation")
self.range_arr = tf.placeholder(tf.int32, shape=[None, ])
self.global_step = tf.Variable(0, trainable=False)
self.decaying_beta = tf.train.exponential_decay(self.beta, self.global_step,
200, 0.90, staircase=False)
self.entity_sequence = []
# to feed in the discounted reward tensor
self.cum_discounted_reward = tf.placeholder(tf.float32, [None, self.path_length],
name="cumulative_discounted_reward")
for t in range(self.path_length):
next_possible_relations = tf.placeholder(tf.int32, [None, self.max_num_actions],
name="next_relations_{}".format(t))
next_possible_entities = tf.placeholder(tf.int32, [None, self.max_num_actions],
name="next_entities_{}".format(t))
input_label_relation = tf.placeholder(tf.int32, [None], name="input_label_relation_{}".format(t))
start_entities = tf.placeholder(tf.int32, [None, ])
self.input_path.append(input_label_relation)
self.candidate_relation_sequence.append(next_possible_relations)
self.candidate_entity_sequence.append(next_possible_entities)
self.entity_sequence.append(start_entities)
self.loss_before_reg = tf.constant(0.0)
self.per_example_loss, self.per_example_logits, self.action_idx = self.agent(
self.candidate_relation_sequence,
self.candidate_entity_sequence, self.entity_sequence,
self.input_path,
self.query_relation, self.range_arr, self.first_state_of_test, self.path_length)
self.loss_op = self.calc_reinforce_loss()
# mark trainable_variables
self.trainable_variables = tf.trainable_variables()
# backprop
self.train_op = self.bp(self.loss_op)
# Building the test graph
self.prev_state = tf.placeholder(tf.float32, self.agent.get_mem_shape(), name="memory_of_agent")
self.prev_relation = tf.placeholder(tf.int32, [None, ], name="previous_relation")
self.query_embedding = tf.nn.embedding_lookup(self.agent.relation_lookup_table, self.query_relation) # [B, 2D]
layer_state = tf.unstack(self.prev_state, self.LSTM_layers)
formated_state = [tf.unstack(s, 2) for s in layer_state]
self.next_relations = tf.placeholder(tf.int32, shape=[None, self.max_num_actions])
self.next_entities = tf.placeholder(tf.int32, shape=[None, self.max_num_actions])
self.current_entities = tf.placeholder(tf.int32, shape=[None,])
with tf.variable_scope("policy_steps_unroll") as scope:
scope.reuse_variables()
self.test_loss, test_state, self.test_logits, self.test_action_idx, self.chosen_relation = self.agent.step(
self.next_relations, self.next_entities, formated_state, self.prev_relation, self.query_embedding,
self.current_entities, self.input_path[0], self.range_arr, self.first_state_of_test)
self.test_state = tf.stack(test_state)
logger.info('TF Graph creation done..')
self.model_saver = tf.train.Saver(max_to_keep=2)
# return the variable initializer Op.
if not restore:
return tf.global_variables_initializer()
else:
return self.model_saver.restore(sess, restore)
def initialize_target(self):
""" Initialize the target model, which is a fixed model as the same as self.agent
We proposed it for Fixed Target Update
"""
logger.info("Creating TF graph...")
self.candidate_relation_sequence_target = []
self.candidate_entity_sequence_target = []
self.input_path_target = []
self.first_state_of_test_target = tf.placeholder(tf.bool, name="is_first_state_of_test_target")
self.query_relation_target = tf.placeholder(tf.int32, [None], name="query_relation_target")
self.range_arr_target = tf.placeholder(tf.int32, shape=[None, ])
self.global_step_target = tf.Variable(0, trainable=False)
self.decaying_beta_target = tf.train.exponential_decay(self.beta, self.global_step_target,
200, 0.90, staircase=False)
self.entity_sequence_target = []
# to feed in the discounted reward tensor
self.cum_discounted_reward_target = tf.placeholder(tf.float32, [None, self.path_length],
name="cumulative_discounted_reward_target")
for t in range(self.path_length):
next_possible_relations_target = tf.placeholder(tf.int32, [None, self.max_num_actions],
name="next_relations_{}_target".format(t))
next_possible_entities_target = tf.placeholder(tf.int32, [None, self.max_num_actions],
name="next_entities_{}_target".format(t))
input_label_relation_target = tf.placeholder(tf.int32, [None], name="input_label_relation_{}_target".format(t))
start_entities_target = tf.placeholder(tf.int32, [None, ])
self.input_path_target.append(input_label_relation_target)
self.candidate_relation_sequence_target.append(next_possible_relations_target)
self.candidate_entity_sequence_target.append(next_possible_entities_target)
self.entity_sequence_target.append(start_entities_target)
self.loss_before_reg_target = tf.constant(0.0)
self.per_example_loss_target, self.per_example_logits_target, self.action_idx_target = self.agent_target(
self.candidate_relation_sequence_target,
self.candidate_entity_sequence_target, self.entity_sequence_target,
self.input_path_target,
self.query_relation_target, self.range_arr_target, self.first_state_of_test_target, self.path_length)
# Building the test graph
self.prev_state_target = tf.placeholder(tf.float32, self.agent_target.get_mem_shape(), name="memory_of_agent")
self.prev_relation_target = tf.placeholder(tf.int32, [None, ], name="previous_relation")
self.query_embedding_target = tf.nn.embedding_lookup(self.agent_target.relation_lookup_table, self.query_relation_target) # [B, 2D]
layer_state_target = tf.unstack(self.prev_state_target, self.LSTM_layers)
formated_state_target = [tf.unstack(s, 2) for s in layer_state_target]
self.next_relations_target = tf.placeholder(tf.int32, shape=[None, self.max_num_actions])
self.next_entities_target = tf.placeholder(tf.int32, shape=[None, self.max_num_actions])
self.current_entities_target = tf.placeholder(tf.int32, shape=[None, ])
with tf.variable_scope("policy_steps_unroll_target") as scope:
scope.reuse_variables()
self.test_loss_target, test_state_target, self.test_logits_target, self.test_action_idx_target, self.chosen_relation_target = self.agent_target.step(
self.next_relations_target, self.next_entities_target, formated_state_target, self.prev_relation_target, self.query_embedding_target,
self.current_entities_target, self.input_path_target[0], self.range_arr_target, self.first_state_of_test_target)
self.test_state_target = tf.stack(test_state_target)
return tf.global_variables_initializer()
def initialize_weight_update(self):
""" call this function after initializing all parameters for self.agent and self.agent_target
Update self.agent_target using variables in self.agent
"""
variables = tf.trainable_variables()
e_params = sorted(self.trainable_variables, key=lambda x: x.name)
t_params = sorted(list(set(variables) - set(self.trainable_variables)), key=lambda x: x.name)
self.replace_target_op = [tf.assign(t, e) for t, e in zip(t_params, e_params)] # extend the replace operations
def gpu_io_setup_target(self):
# create fetches for partial_run_setup
fetches = self.per_example_loss_target + self.action_idx_target + self.per_example_logits_target
feeds = [self.first_state_of_test_target] + self.candidate_relation_sequence_target+ self.candidate_entity_sequence_target + self.input_path_target + \
[self.query_relation_target] + [self.cum_discounted_reward_target] + [self.range_arr_target] + self.entity_sequence_target
feed_dict = [{} for _ in range(self.path_length)]
feed_dict[0][self.first_state_of_test_target] = False
feed_dict[0][self.query_relation_target] = None
feed_dict[0][self.range_arr_target] = np.arange(self.batch_size*self.num_rollouts)
for i in range(self.path_length):
feed_dict[i][self.input_path_target[i]] = np.zeros(self.batch_size * self.num_rollouts) # placebo
feed_dict[i][self.candidate_relation_sequence_target[i]] = None
feed_dict[i][self.candidate_entity_sequence_target[i]] = None
feed_dict[i][self.entity_sequence_target[i]] = None
return fetches, feeds, feed_dict
def initialize_pretrained_embeddings(self, sess):
if self.pretrained_embeddings_action != '':
embeddings = np.loadtxt(open(self.pretrained_embeddings_action))
_ = sess.run((self.agent.relation_embedding_init),
feed_dict={self.agent.action_embedding_placeholder: embeddings})
if self.pretrained_embeddings_entity != '':
embeddings = np.loadtxt(open(self.pretrained_embeddings_entity))
_ = sess.run((self.agent.entity_embedding_init),
feed_dict={self.agent.entity_embedding_placeholder: embeddings})
def bp(self, cost):
self.baseline.update(tf.reduce_mean(self.cum_discounted_reward))
tvars = self.trainable_variables
grads = tf.gradients(cost, tvars)
grads, _ = tf.clip_by_global_norm(grads, self.grad_clip_norm)
train_op = self.optimizer.apply_gradients(zip(grads, tvars))
with tf.control_dependencies([train_op]): # see https://github.com/tensorflow/tensorflow/issues/1899
self.dummy = tf.constant(0)
return train_op
def calc_cum_discounted_reward(self, rewards):
"""
calculates the cumulative discounted reward.
:param rewards:
:param T:
:param gamma:
:return:
"""
running_add = np.zeros([rewards.shape[0]]) # [B]
cum_disc_reward = np.zeros([rewards.shape[0], self.path_length]) # [B, T]
cum_disc_reward[:,
self.path_length - 1] = rewards # set the last time step to the reward received at the last state
for t in reversed(range(self.path_length)):
running_add = self.gamma * running_add + cum_disc_reward[:, t]
cum_disc_reward[:, t] = running_add
return cum_disc_reward
def gpu_io_setup(self):
# create fetches for partial_run_setup
fetches = self.per_example_loss + self.action_idx + [self.loss_op] + self.per_example_logits + [self.dummy]
feeds=[self.first_state_of_test] + self.candidate_relation_sequence+ self.candidate_entity_sequence + self.input_path + \
[self.query_relation] + [self.cum_discounted_reward] + [self.range_arr] + self.entity_sequence
feed_dict = [{} for _ in range(self.path_length)]
feed_dict[0][self.first_state_of_test] = False
feed_dict[0][self.query_relation] = None
feed_dict[0][self.range_arr] = np.arange(self.batch_size*self.num_rollouts)
for i in range(self.path_length):
feed_dict[i][self.input_path[i]] = np.zeros(self.batch_size * self.num_rollouts) # placebo
feed_dict[i][self.candidate_relation_sequence[i]] = None
feed_dict[i][self.candidate_entity_sequence[i]] = None
feed_dict[i][self.entity_sequence[i]] = None
return fetches, feeds, feed_dict
def gpu_io_setup_test(self):
# create fetches for partial_run_setup
fetches = self.per_example_loss + self.action_idx + self.per_example_logits
feeds = [self.first_state_of_test] + self.candidate_relation_sequence+ self.candidate_entity_sequence + self.input_path + \
[self.query_relation] + [self.range_arr] + self.entity_sequence
feed_dict = [{} for _ in range(self.path_length)]
feed_dict[0][self.first_state_of_test] = False
feed_dict[0][self.query_relation] = None
feed_dict[0][self.range_arr] = np.arange(self.batch_size*self.num_rollouts)
for i in range(self.path_length):
feed_dict[i][self.input_path[i]] = np.zeros(self.batch_size * self.num_rollouts) # placebo
feed_dict[i][self.candidate_relation_sequence[i]] = None
feed_dict[i][self.candidate_entity_sequence[i]] = None
feed_dict[i][self.entity_sequence[i]] = None
return fetches, feeds, feed_dict
def train(self, sess):
# import pdb
# pdb.set_trace()
fetches, feeds, feed_dict = self.gpu_io_setup()
train_loss = 0.0
start_time = time.time()
self.batch_counter = 0
for episode in self.train_environment.get_episodes():
self.batch_counter += 1
h = sess.partial_run_setup(fetches=fetches, feeds=feeds)
feed_dict[0][self.query_relation] = episode.get_query_relation()
# get initial state
state = episode.get_state()
# for each time step
loss_before_regularization = []
logits = []
for i in range(self.path_length):
feed_dict[i][self.candidate_relation_sequence[i]] = state['next_relations']
feed_dict[i][self.candidate_entity_sequence[i]] = state['next_entities']
feed_dict[i][self.entity_sequence[i]] = state['current_entities']
per_example_loss, per_example_logits, idx = sess.partial_run(h, [self.per_example_loss[i], self.per_example_logits[i], self.action_idx[i]],
feed_dict=feed_dict[i])
loss_before_regularization.append(per_example_loss)
logits.append(per_example_logits)
# action = np.squeeze(action, axis=1) # [B,]
state = episode(idx)
loss_before_regularization = np.stack(loss_before_regularization, axis=1)
# get the final reward from the environment
rewards = episode.get_reward()
# computed cumulative discounted reward
cum_discounted_reward = self.calc_cum_discounted_reward(rewards) # [B, T]
# backprop
batch_total_loss, _ = sess.partial_run(h, [self.loss_op, self.dummy],
feed_dict={self.cum_discounted_reward: cum_discounted_reward})
# print statistics
train_loss = 0.98 * train_loss + 0.02 * batch_total_loss
avg_reward = np.mean(rewards)
# now reshape the reward to [orig_batch_size, num_rollouts], I want to calculate for how many of the
# entity pair, atleast one of the path get to the right answer
reward_reshape = np.reshape(rewards, (self.batch_size, self.num_rollouts)) # [orig_batch, num_rollouts]
reward_reshape = np.sum(reward_reshape, axis=1) # [orig_batch]
reward_reshape = (reward_reshape > 0)
num_ep_correct = np.sum(reward_reshape)
if np.isnan(train_loss):
raise ArithmeticError("Error in computing loss")
logger.info("batch_counter: {0:4d}, num_hits: {1:7.4f}, avg. reward per batch {2:7.4f}, "
"num_ep_correct {3:4d}, avg_ep_correct {4:7.4f}, train loss {5:7.4f}".
format(self.batch_counter, np.sum(rewards), avg_reward, num_ep_correct,
(num_ep_correct / self.batch_size),
train_loss))
if self.batch_counter%self.eval_every == 0:
with open(self.output_dir + '/scores.txt', 'a') as score_file:
score_file.write("Score for iteration " + str(self.batch_counter) + "\n")
if not os.path.exists(self.path_logger_file + "/" + str(self.batch_counter)):
os.mkdir(self.path_logger_file + "/" + str(self.batch_counter))
self.path_logger_file_ = self.path_logger_file + "/" + str(self.batch_counter) + "/paths"
self.test(sess, beam=True, print_paths=False)
logger.info('Memory usage: %s (kb)' % resource.getrusage(resource.RUSAGE_SELF).ru_maxrss)
gc.collect()
if self.batch_counter >= self.total_iterations:
break
def get_pcnn_predictions(self, framework, state, id_entities_dict):
entpair_list, result_list = framework.predict(state['current_entities'], id_entities_dict)
pcnn_edge_idx = []
next_relations, next_entities, pcnn_confidence_recorder = [], [], []
_, max_length = state['next_relations'].shape
# map from PCNN to GFAW
self.train_environment.batcher.relation_vocab['NA'] = self.train_environment.batcher.relation_vocab['UNK'] # TODO: delete it
relation_map = {framework.test_data_loader.rel2id[k]:self.train_environment.batcher.relation_vocab[k] for k in framework.test_data_loader.rel2id}
entity_map = self.train_environment.batcher.entity_vocab
for i, (entpairs, results) in enumerate(zip(entpair_list, result_list)):
current_ety_idx = state['current_entities'][i]
# Add edge from GFAW
if len(results) == 0:
pcnn_edge_idx.append(self.max_num_actions)
next_rlt = self.train_environment.grapher.array_store[current_ety_idx][:, 1]
next_ety = self.train_environment.grapher.array_store[current_ety_idx][:, 0]
next_relations.append(next_rlt)
next_entities.append(next_ety)
# Add edge from PCNN
else:
length = sum(np.array(entpairs) != "None#None")
pcnn_rlt = [relation_map[j] for j in np.argmax(results, axis=1)[:length]]
pcnn_ety = [entity_map[j.split("#")[-1]] if j.split("#")[-1] in entity_map else entity_map['UNK'] for j in entpairs[:length]]
next_rlt = self.train_environment.grapher.array_store[current_ety_idx][:, 1]
next_ety = self.train_environment.grapher.array_store[current_ety_idx][:, 0]
for edge_idx in range(len(pcnn_rlt)):
next_rlt[-(edge_idx+1)] = pcnn_rlt[edge_idx]
next_ety[-(edge_idx+1)] = pcnn_ety[edge_idx]
# if predicted idx > pcnn edge idx, it's pcnn edge
pcnn_edge_idx.append(self.max_num_actions-len(pcnn_rlt)-1)
next_relations.append(next_rlt)
next_entities.append(next_ety)
next_relations = np.array(next_relations)
next_entities = np.array(next_entities)
return next_relations, next_entities, pcnn_edge_idx
def load_dicts(self):
print("Using bfs to store positive pcnn sequence in replay memory...")
if os.path.exists('./_processed_data/'+self.params['gfaw_dataset']+'-pre_train.pkl'):
print("Dicts already Stored!")
dir = './_processed_data/'
pre_train = pickle.load(open(os.path.join(dir, self.params['gfaw_dataset']+'-pre_train.pkl'), 'rb'))
Graph = pickle.load(open(os.path.join(dir, self.params['gfaw_dataset']+'-Graph.pkl'), 'rb'))
entpair2rlt = pickle.load(open(os.path.join(dir, self.params['gfaw_dataset']+'-entpair2rlt.pkl'), 'rb'))
edge_entpair2rlt = pickle.load(open(os.path.join(dir, self.params['gfaw_dataset']+'-edge_entpair2rlt.pkl'), 'rb'))
else:
dir = "../../data/"
load_dir = './_processed_data/'
pre_dataset = os.path.join(dir, self.params['gfaw_dataset'])
after_dataset = os.path.join(dir, self.params['gfaw_dataset'][:-3]+'0.0')
pre_train = pd.read_csv(os.path.join(pre_dataset, "train.txt"), sep='\t', names=['e1', 'r', 'e2'])
after_train = pd.read_csv(os.path.join(after_dataset, "train.txt"), sep='\t', names=['e1', 'r', 'e2'])
edges = after_train.append(pre_train).drop_duplicates(keep=False)
print("Storing entity pairs to relations dict...")
edge_entpair2rlt = {(row['e1'] + '#' + row['e2']): row['r'] for (index, row) in tqdm(edges.iterrows())}
entpair2rlt = {(row['e1'] + '#' + row['e2']): row['r'] for (index, row) in tqdm(after_train.iterrows())}
Graph = constructGraph(pre_train, edges)
print("Saving Dicts for future use...")
pickle.dump(pre_train, open(os.path.join(load_dir, self.params['gfaw_dataset']+'-pre_train.pkl'), 'wb'))
pickle.dump(Graph, open(os.path.join(load_dir, self.params['gfaw_dataset']+'-Graph.pkl'), 'wb'))
pickle.dump(entpair2rlt, open(os.path.join(load_dir, self.params['gfaw_dataset']+'-entpair2rlt.pkl'), 'wb'))
pickle.dump(edge_entpair2rlt, open(os.path.join(load_dir, self.params['gfaw_dataset']+'-edge_entpair2rlt.pkl'), 'wb'))
del after_train, edges
return pre_train, Graph, entpair2rlt, edge_entpair2rlt
def use_bfs(self):
pre_train, Graph, entpair2rlt, edge_entpair2rlt = self.load_dicts()
count = 0
# dir = "../../data/"
# pre_dataset = os.path.join(dir, "GFAW-cutoff-0.5-PCNN-1.0")
# test = pd.read_csv(os.path.join(pre_dataset, "test.txt"), sep='\t', names=['e1', 'r', 'e2'])
for index, row in tqdm(pre_train.sample(n=100).iterrows()): # , random_state=self.params['random_seed']
if row['e1'] == row['e2']:
pass
else:
del Graph[row['e1']][row['e2']]
distance, path = shortestPath(Graph, row['e1'], row['e2'])
Graph[row['e1']][row['e2']] = 1
## ===================== for pcnn edge:
if len(path) and distance<(len(path)-1) and (len(path)-1) == self.params['path_length']:
# exist path and exist pcnn added edge
flag = 0
for e in path:
if e not in self.train_environment.grapher.entity_vocab:
print(e, " is not in vocab!")
flag = 1
if not flag:
count += 1
entity_path = [self.train_environment.grapher.entity_vocab[e] for e in path]
relation_path_ = [entpair2rlt[path[i] + '#' + path[i + 1]] for i in range(len(path) - 1)]
relation_path = [self.train_environment.grapher.relation_vocab[r] for r in relation_path_]
pcnn_edge = [1 if path[i] + '#' + path[i + 1] in edge_entpair2rlt else 0 for i in
range(len(path) - 1)]
self.pos_experience.insert('entity_path', np.array([entity_path]))
self.pos_experience.insert('relation_path', np.array([relation_path]))
self.pos_experience.insert('path_rewards', np.array([1]))
self.pos_experience.insert('state_rewards', np.array([[1] * self.params['path_length']]))
self.pos_experience.insert('query_relation',
np.array([[self.train_environment.grapher.relation_vocab[row['r']]]]))
self.pos_experience.insert('pcnn_edge', np.array([pcnn_edge]))
for index, is_pcnn in enumerate(pcnn_edge):
if is_pcnn:
self.pcnn_experience.insert('triples',
[[path[index], path[index + 1], relation_path_[index]]])
del distance, path
## ============================ for gfaw edge:
# if len(path) and (len(path) - 1) == self.params['path_length']:
# # exist path and exist pcnn added edge
# flag = 0
# for e in path:
# if e not in self.train_environment.grapher.entity_vocab:
# print(e, " is not in vocab!")
# flag = 1
# if not flag:
# count += 1
# entity_path = [self.train_environment.grapher.entity_vocab[e] for e in path]
# relation_path_ = [entpair2rlt[path[i]+'#'+path[i+1]] for i in range(len(path)-1)]
# relation_path = [self.train_environment.grapher.relation_vocab[r] for r in relation_path_]
# pcnn_edge = [1 if path[i]+'#'+path[i+1] in edge_entpair2rlt else 0 for i in range(len(path)-1)]
# self.pos_experience.insert('entity_path', np.array([entity_path]))
# self.pos_experience.insert('relation_path', np.array([relation_path]))
# self.pos_experience.insert('path_rewards', np.array([1]))
# self.pos_experience.insert('state_rewards', np.array([[1] * self.params['path_length']]))
# self.pos_experience.insert('query_relation', np.array([[self.train_environment.grapher.relation_vocab[row['r']]]]))
# self.pos_experience.insert('pcnn_edge', np.array([pcnn_edge]))
#
# for index, is_pcnn in enumerate(pcnn_edge):
# if is_pcnn:
# self.pcnn_experience.insert('triples', [[path[index], path[index+1], relation_path_[index]]])
# del distance, path
print("Done! There are ", str(count), " positive samples!")
del pre_train, edge_entpair2rlt, entpair2rlt, Graph
def store_bfs(self):
self.pos_experience = Memory(memory_size=50000)
self.pcnn_experience = Memory(memory_size=50000)
pre_train, Graph, entpair2rlt, edge_entpair2rlt = self.load_dicts()
count = 0
for index, row in pre_train.sample(frac=1).iterrows(): # , random_state=self.params['random_seed']
print(count, end='\r')
if row['e1'] == row['e2']:
pass
else:
del Graph[row['e1']][row['e2']]
distance, path = shortestPath(Graph, row['e1'], row['e2'])
Graph[row['e1']][row['e2']] = 1
## ===================== for pcnn edge:
if len(path) and distance < (len(path) - 1) and (len(path) - 1) == self.params['path_length']:
# exist path and exist pcnn added edge
flag = 0
for e in path:
if e not in self.train_environment.grapher.entity_vocab:
print(e, " is not in vocab!")
flag = 1
if not flag:
count += 1
entity_path = [self.train_environment.grapher.entity_vocab[e] for e in path]
relation_path_ = [entpair2rlt[path[i] + '#' + path[i + 1]] for i in range(len(path) - 1)]
relation_path = [self.train_environment.grapher.relation_vocab[r] for r in relation_path_]
pcnn_edge = [1 if path[i] + '#' + path[i + 1] in edge_entpair2rlt else 0 for i in
range(len(path) - 1)]
self.pos_experience.insert('entity_path', np.array([entity_path]))
self.pos_experience.insert('relation_path', np.array([relation_path]))
self.pos_experience.insert('path_rewards', np.array([1]))
self.pos_experience.insert('state_rewards', np.array([[1] * self.params['path_length']]))
self.pos_experience.insert('query_relation',
np.array([[self.train_environment.grapher.relation_vocab[row['r']]]]))
self.pos_experience.insert('pcnn_edge', np.array([pcnn_edge]))
for index, is_pcnn in enumerate(pcnn_edge):
if is_pcnn:
self.pcnn_experience.insert('triples', [[path[index], path[index + 1], relation_path_[index]]])
if 'entity_path' in self.pos_experience.memory and len(self.pos_experience.memory['entity_path']) == self.pos_experience.memory_size:
print("Memory Full!")
break
pickle.dump(self.pos_experience, open(os.path.join('./_processed_data/', self.params['gfaw_dataset'] + '-pre_bfs.pkl'), 'wb'))
pickle.dump(self.pcnn_experience, open(os.path.join('./_processed_data/', self.params['gfaw_dataset'] + '-pre_pcnn.pkl'), 'wb'))
def one_path(self, one_line):
triples = one_line.strip().split(';')
query = eval(triples[0])
path_triples = triples[2:-1]
relation_path = [eval(triple)[1] for triple in path_triples]
entity_path = [eval(triple)[0] for triple in path_triples] + [query[2]]
entity_path.extend([0] * (self.path_length-len(path_triples))) # add padding
relation_path.extend([0] * (self.path_length-len(path_triples))) # add padding
self.pos_experience.insert('entity_path', np.array([entity_path]))
self.pos_experience.insert('relation_path', np.array([relation_path]))
self.pos_experience.insert('path_rewards', np.array([1]))
self.pos_experience.insert('state_rewards', np.array([[1] * self.path_length]))
self.pos_experience.insert('query_relation', np.array([[query[1]]]))
# self.pos_experience.insert('pcnn_edge', np.array([pcnn_edge]))
def use_bfs_new(self):
for _ in tqdm(range(self.batch_size * self.num_rollouts)):
self.one_path(self.bfs_path[random.randint(0, len(self.bfs_path) - 1)])
def train_joint_withoutRM(self, sess, framework):
# import pdb
# pdb.set_trace()
fetches, feeds, feed_dict = self.gpu_io_setup()
# fetches_test, feeds_test, feed_dict_test = self.gpu_io_setup_test()
train_loss = 0.0
start_time = time.time()
self.batch_counter = 0
id_entities_dict = {self.train_environment.batcher.entity_vocab[k]: k
for k in self.train_environment.batcher.entity_vocab}
for episode in self.train_environment.get_episodes():
self.batch_counter += 1
h = sess.partial_run_setup(fetches=fetches, feeds=feeds)
feed_dict[0][self.query_relation] = episode.get_query_relation()
# get initial state
state = episode.get_state()
# for each time step
loss_before_regularization = []
logits = []
for i in range(self.path_length):
next_relations, next_entities, pcnn_edge_idx = self.get_pcnn_predictions(framework, state,
id_entities_dict)
# Switch to GFAW
with sess.as_default():
with sess.graph.as_default():
# TODO: Adapt GFAW feed_dict according to the PCNN prediction
feed_dict[i][self.candidate_relation_sequence[i]] = next_relations
feed_dict[i][self.candidate_entity_sequence[i]] = next_entities
feed_dict[i][self.entity_sequence[i]] = state['current_entities'] # [batch_size*num_rollouts, ]
# GFAW predict next action
# TODO: sess.partial_run
per_example_loss, per_example_logits, idx = sess.partial_run(h, [self.per_example_loss[i],
self.per_example_logits[i],
self.action_idx[i]],
feed_dict=feed_dict[i])
loss_before_regularization.append(per_example_loss)
logits.append(per_example_logits)
# action = np.squeeze(action, axis=1) # [B,]
# GFAW return next state
# episode.state['next_entities'] = np.array(next_entities)
# state = episode(idx) # __call__(self, action) return state
state['current_entities'] = np.array(next_entities)[np.arange(self.batch_size * self.num_rollouts), idx]
loss_before_regularization = np.stack(loss_before_regularization, axis=1)
# get the final reward from the environment
# rewards = episode.get_reward()
reward = (state['current_entities'] == episode.end_entities)
condlist = [reward == True, reward == False]
choicelist = [episode.positive_reward, episode.negative_reward]
rewards = np.select(condlist, choicelist) # [B,]
# computed cumulative discounted reward
cum_discounted_reward = self.calc_cum_discounted_reward(rewards) # [B, T]
# backprop
batch_total_loss, _ = sess.partial_run(h, [self.loss_op, self.dummy],
feed_dict={self.cum_discounted_reward: cum_discounted_reward})
# print statistics
train_loss = 0.98 * train_loss + 0.02 * batch_total_loss
avg_reward = np.mean(rewards)
# now reshape the reward to [orig_batch_size, num_rollouts], I want to calculate for how many of the
# entity pair, atleast one of the path get to the right answer
reward_reshape = np.reshape(rewards, (self.batch_size, self.num_rollouts)) # [orig_batch, num_rollouts]
reward_reshape = np.sum(reward_reshape, axis=1) # [orig_batch]
reward_reshape = (reward_reshape > 0)
num_ep_correct = np.sum(reward_reshape)
if np.isnan(train_loss):
raise ArithmeticError("Error in computing loss")
logger.info("batch_counter: {0:4d}, num_hits: {1:7.4f}, avg. reward per batch {2:7.4f}, "
"num_ep_correct {3:4d}, avg_ep_correct {4:7.4f}, train loss {5:7.4f}".
format(self.batch_counter, np.sum(rewards), avg_reward, num_ep_correct,
(num_ep_correct / self.batch_size),
train_loss))
if self.batch_counter%self.eval_every == 0:
with open(self.output_dir + '/scores.txt', 'a') as score_file:
score_file.write("Score for iteration " + str(self.batch_counter) + "\n")
if not os.path.exists(self.path_logger_file + "/" + str(self.batch_counter)):
os.mkdir(self.path_logger_file + "/" + str(self.batch_counter))
self.path_logger_file_ = self.path_logger_file + "/" + str(self.batch_counter) + "/paths"
self.test(sess, beam=True, print_paths=False)
logger.info('Memory usage: %s (kb)' % resource.getrusage(resource.RUSAGE_SELF).ru_maxrss)
gc.collect()
if self.batch_counter >= self.total_iterations:
break
def train_joint(self, sess, framework):
# import pdb
# pdb.set_trace()
fetches, feeds, feed_dict = self.gpu_io_setup()
fetches_test, feeds_test, feed_dict_test = self.gpu_io_setup_test()
# setup target model
# self.target_model_setup()
# sess.run(self.replace_target_op)
# if self.params['bfs_iteration']:
# with open(os.path.join(self.params['data_input_dir'], 'path_8r.txt'), 'r') as f:
# self.bfs_path = f.readlines()
train_loss = 0.0
pcnn_pos_edge_appearance = {}
self.batch_counter = 0
self.pos_experience = Memory()
self.neg_experience = Memory()