forked from salesforce/MultiHopKG
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathexperiments.py
765 lines (700 loc) · 34.8 KB
/
experiments.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
#!/usr/bin/env python3
"""
Copyright (c) 2018, salesforce.com, inc.
All rights reserved.
SPDX-License-Identifier: BSD-3-Clause
For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause
Experiment Portal.
"""
import copy
import itertools
import numpy as np
import os, sys
import random
import torch
from src.parse_args import parser
from src.parse_args import args
import src.data_utils as data_utils
import src.eval
from src.hyperparameter_range import hp_range
from src.knowledge_graph import KnowledgeGraph
from src.emb.fact_network import ComplEx, ConvE, DistMult
from src.emb.fact_network import get_conve_kg_state_dict, get_complex_kg_state_dict, get_distmult_kg_state_dict
from src.emb.emb import EmbeddingBasedMethod
from src.rl.graph_search.pn import GraphSearchPolicy
from src.rl.graph_search.pg import PolicyGradient
from src.rl.graph_search.rs_pg import RewardShapingPolicyGradient
from src.utils.ops import flatten
torch.cuda.set_device(args.gpu)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
def process_data():
data_dir = args.data_dir
raw_kb_path = os.path.join(data_dir, 'raw.kb')
train_path = data_utils.get_train_path(args)
dev_path = os.path.join(data_dir, 'dev.triples')
test_path = os.path.join(data_dir, 'test.triples')
data_utils.prepare_kb_envrioment(raw_kb_path, train_path, dev_path, test_path, args.test, args.add_reverse_relations)
def initialize_model_directory(args, random_seed=None):
# add model parameter info to model directory
model_root_dir = args.model_root_dir
dataset = os.path.basename(os.path.normpath(args.data_dir))
reverse_edge_tag = '-RV' if args.add_reversed_training_edges else ''
entire_graph_tag = '-EG' if args.train_entire_graph else ''
if args.xavier_initialization:
initialization_tag = '-xavier'
elif args.uniform_entity_initialization:
initialization_tag = '-uniform'
else:
initialization_tag = ''
# Hyperparameter signature
if args.model in ['rule']:
hyperparam_sig = '{}-{}-{}-{}-{}-{}-{}-{}-{}-{}'.format(
args.baseline,
args.entity_dim,
args.relation_dim,
args.history_num_layers,
args.learning_rate,
args.emb_dropout_rate,
args.ff_dropout_rate,
args.action_dropout_rate,
args.bandwidth,
args.beta
)
elif args.model.startswith('point'):
if args.baseline == 'avg_reward':
print('* Policy Gradient Baseline: average reward')
elif args.baseline == 'avg_reward_normalized':
print('* Policy Gradient Baseline: average reward baseline plus normalization')
else:
print('* Policy Gradient Baseline: None')
if args.action_dropout_anneal_interval < 1000:
hyperparam_sig = '{}-{}-{}-{}-{}-{}-{}-{}-{}-{}-{}-{}'.format(
args.baseline,
args.entity_dim,
args.relation_dim,
args.history_num_layers,
args.learning_rate,
args.emb_dropout_rate,
args.ff_dropout_rate,
args.action_dropout_rate,
args.action_dropout_anneal_factor,
args.action_dropout_anneal_interval,
args.bandwidth,
args.beta
)
if args.mu != 1.0:
hyperparam_sig += '-{}'.format(args.mu)
else:
hyperparam_sig = '{}-{}-{}-{}-{}-{}-{}-{}-{}-{}'.format(
args.baseline,
args.entity_dim,
args.relation_dim,
args.history_num_layers,
args.learning_rate,
args.emb_dropout_rate,
args.ff_dropout_rate,
args.action_dropout_rate,
args.bandwidth,
args.beta
)
if args.reward_shaping_threshold > 0:
hyperparam_sig += '-{}'.format(args.reward_shaping_threshold)
elif args.model == 'distmult':
hyperparam_sig = '{}-{}-{}-{}-{}'.format(
args.entity_dim,
args.relation_dim,
args.learning_rate,
args.emb_dropout_rate,
args.label_smoothing_epsilon
)
elif args.model == 'complex':
hyperparam_sig = '{}-{}-{}-{}-{}'.format(
args.entity_dim,
args.relation_dim,
args.learning_rate,
args.emb_dropout_rate,
args.label_smoothing_epsilon
)
elif args.model in ['conve', 'hypere', 'triplee']:
hyperparam_sig = '{}-{}-{}-{}-{}-{}-{}-{}-{}'.format(
args.entity_dim,
args.relation_dim,
args.learning_rate,
args.num_out_channels,
args.kernel_size,
args.emb_dropout_rate,
args.hidden_dropout_rate,
args.feat_dropout_rate,
args.label_smoothing_epsilon
)
else:
raise NotImplementedError
model_sub_dir = '{}-{}{}{}{}-{}'.format(
dataset,
args.model,
reverse_edge_tag,
entire_graph_tag,
initialization_tag,
hyperparam_sig
)
if args.model == 'set':
model_sub_dir += '-{}'.format(args.beam_size)
model_sub_dir += '-{}'.format(args.num_paths_per_entity)
if args.relation_only:
model_sub_dir += '-ro'
elif args.relation_only_in_path:
model_sub_dir += '-rpo'
elif args.type_only:
model_sub_dir += '-to'
if args.test:
model_sub_dir += '-test'
if random_seed:
model_sub_dir += '.{}'.format(random_seed)
model_dir = os.path.join(model_root_dir, model_sub_dir)
if not os.path.exists(model_dir):
os.makedirs(model_dir)
print('Model directory created: {}'.format(model_dir))
else:
print('Model directory exists: {}'.format(model_dir))
args.model_dir = model_dir
def construct_model(args):
"""
Construct NN graph.
"""
kg = KnowledgeGraph(args)
if args.model.endswith('.gc'):
kg.load_fuzzy_facts()
if args.model in ['point', 'point.gc']:
pn = GraphSearchPolicy(args)
lf = PolicyGradient(args, kg, pn)
elif args.model.startswith('point.rs'):
pn = GraphSearchPolicy(args)
fn_model = args.model.split('.')[2]
fn_args = copy.deepcopy(args)
fn_args.model = fn_model
fn_args.relation_only = False
if fn_model == 'complex':
fn = ComplEx(fn_args)
fn_kg = KnowledgeGraph(fn_args)
elif fn_model == 'distmult':
fn = DistMult(fn_args)
fn_kg = KnowledgeGraph(fn_args)
elif fn_model == 'conve':
fn = ConvE(fn_args, kg.num_entities)
fn_kg = KnowledgeGraph(fn_args)
lf = RewardShapingPolicyGradient(args, kg, pn, fn_kg, fn)
elif args.model == 'complex':
fn = ComplEx(args)
lf = EmbeddingBasedMethod(args, kg, fn)
elif args.model == 'distmult':
fn = DistMult(args)
lf = EmbeddingBasedMethod(args, kg, fn)
elif args.model == 'conve':
fn = ConvE(args, kg.num_entities)
lf = EmbeddingBasedMethod(args, kg, fn)
else:
raise NotImplementedError
return lf
def train(lf):
train_path = data_utils.get_train_path(args)
dev_path = os.path.join(args.data_dir, 'dev.triples')
entity_index_path = os.path.join(args.data_dir, 'entity2id.txt')
relation_index_path = os.path.join(args.data_dir, 'relation2id.txt')
train_data = data_utils.load_triples(
train_path, entity_index_path, relation_index_path, group_examples_by_query=args.group_examples_by_query,
add_reverse_relations=args.add_reversed_training_edges)
if 'NELL' in args.data_dir:
adj_list_path = os.path.join(args.data_dir, 'adj_list.pkl')
seen_entities = data_utils.load_seen_entities(adj_list_path, entity_index_path)
else:
seen_entities = set()
dev_data = data_utils.load_triples(dev_path, entity_index_path, relation_index_path, seen_entities=seen_entities)
if args.checkpoint_path is not None:
lf.load_checkpoint(args.checkpoint_path)
lf.run_train(train_data, dev_data)
def inference(lf):
lf.batch_size = args.dev_batch_size
lf.eval()
if args.model == 'hypere':
conve_kg_state_dict = get_conve_kg_state_dict(torch.load(args.conve_state_dict_path))
lf.kg.load_state_dict(conve_kg_state_dict)
secondary_kg_state_dict = get_complex_kg_state_dict(torch.load(args.complex_state_dict_path))
lf.secondary_kg.load_state_dict(secondary_kg_state_dict)
elif args.model == 'triplee':
conve_kg_state_dict = get_conve_kg_state_dict(torch.load(args.conve_state_dict_path))
lf.kg.load_state_dict(conve_kg_state_dict)
complex_kg_state_dict = get_complex_kg_state_dict(torch.load(args.complex_state_dict_path))
lf.secondary_kg.load_state_dict(complex_kg_state_dict)
distmult_kg_state_dict = get_distmult_kg_state_dict(torch.load(args.distmult_state_dict_path))
lf.tertiary_kg.load_state_dict(distmult_kg_state_dict)
else:
lf.load_checkpoint(get_checkpoint_path(args))
entity_index_path = os.path.join(args.data_dir, 'entity2id.txt')
relation_index_path = os.path.join(args.data_dir, 'relation2id.txt')
if 'NELL' in args.data_dir:
adj_list_path = os.path.join(args.data_dir, 'adj_list.pkl')
seen_entities = data_utils.load_seen_entities(adj_list_path, entity_index_path)
else:
seen_entities = set()
eval_metrics = {
'dev': {},
'test': {}
}
if args.compute_map:
relation_sets = [
'concept:athletehomestadium',
'concept:athleteplaysforteam',
'concept:athleteplaysinleague',
'concept:athleteplayssport',
'concept:organizationheadquarteredincity',
'concept:organizationhiredperson',
'concept:personborninlocation',
'concept:teamplayssport',
'concept:worksfor'
]
mps = []
for r in relation_sets:
print('* relation: {}'.format(r))
test_path = os.path.join(args.data_dir, 'tasks', r, 'test.pairs')
test_data, labels = data_utils.load_triples_with_label(
test_path, r, entity_index_path, relation_index_path, seen_entities=seen_entities)
pred_scores = lf.forward(test_data, verbose=False)
mp = src.eval.link_MAP(test_data, pred_scores, labels, lf.kg.all_objects, verbose=True)
mps.append(mp)
map_ = np.mean(mps)
print('Overall MAP = {}'.format(map_))
eval_metrics['test']['avg_map'] = map
elif args.eval_by_relation_type:
dev_path = os.path.join(args.data_dir, 'dev.triples')
dev_data = data_utils.load_triples(dev_path, entity_index_path, relation_index_path, seen_entities=seen_entities)
pred_scores = lf.forward(dev_data, verbose=False)
to_m_rels, to_1_rels, _ = data_utils.get_relations_by_type(args.data_dir, relation_index_path)
relation_by_types = (to_m_rels, to_1_rels)
print('Dev set evaluation by relation type (partial graph)')
src.eval.hits_and_ranks_by_relation_type(
dev_data, pred_scores, lf.kg.dev_objects, relation_by_types, verbose=True)
print('Dev set evaluation by relation type (full graph)')
src.eval.hits_and_ranks_by_relation_type(
dev_data, pred_scores, lf.kg.all_objects, relation_by_types, verbose=True)
elif args.eval_by_seen_queries:
dev_path = os.path.join(args.data_dir, 'dev.triples')
dev_data = data_utils.load_triples(dev_path, entity_index_path, relation_index_path, seen_entities=seen_entities)
pred_scores = lf.forward(dev_data, verbose=False)
seen_queries = data_utils.get_seen_queries(args.data_dir, entity_index_path, relation_index_path)
print('Dev set evaluation by seen queries (partial graph)')
src.eval.hits_and_ranks_by_seen_queries(
dev_data, pred_scores, lf.kg.dev_objects, seen_queries, verbose=True)
print('Dev set evaluation by seen queries (full graph)')
src.eval.hits_and_ranks_by_seen_queries(
dev_data, pred_scores, lf.kg.all_objects, seen_queries, verbose=True)
else:
dev_path = os.path.join(args.data_dir, 'dev.triples')
test_path = os.path.join(args.data_dir, 'test.triples')
dev_data = data_utils.load_triples(
dev_path, entity_index_path, relation_index_path, seen_entities=seen_entities, verbose=False)
test_data = data_utils.load_triples(
test_path, entity_index_path, relation_index_path, seen_entities=seen_entities, verbose=False)
print('Dev set performance:')
pred_scores = lf.forward(dev_data, verbose=args.save_beam_search_paths)
dev_metrics = src.eval.hits_and_ranks(dev_data, pred_scores, lf.kg.dev_objects, verbose=True)
eval_metrics['dev'] = {}
eval_metrics['dev']['hits_at_1'] = dev_metrics[0]
eval_metrics['dev']['hits_at_3'] = dev_metrics[1]
eval_metrics['dev']['hits_at_5'] = dev_metrics[2]
eval_metrics['dev']['hits_at_10'] = dev_metrics[3]
eval_metrics['dev']['mrr'] = dev_metrics[4]
src.eval.hits_and_ranks(dev_data, pred_scores, lf.kg.all_objects, verbose=True)
print('Test set performance:')
pred_scores = lf.forward(test_data, verbose=False)
test_metrics = src.eval.hits_and_ranks(test_data, pred_scores, lf.kg.all_objects, verbose=True)
eval_metrics['test']['hits_at_1'] = test_metrics[0]
eval_metrics['test']['hits_at_3'] = test_metrics[1]
eval_metrics['test']['hits_at_5'] = test_metrics[2]
eval_metrics['test']['hits_at_10'] = test_metrics[3]
eval_metrics['test']['mrr'] = test_metrics[4]
return eval_metrics
def run_ablation_studies(args):
"""
Run the ablation study experiments reported in the paper.
"""
def set_up_lf_for_inference(args):
initialize_model_directory(args)
lf = construct_model(args)
lf.cuda()
lf.batch_size = args.dev_batch_size
lf.load_checkpoint(get_checkpoint_path(args))
lf.eval()
return lf
def rel_change(metrics, ab_system, kg_portion):
ab_system_metrics = metrics[ab_system][kg_portion]
base_metrics = metrics['ours'][kg_portion]
return int(np.round((ab_system_metrics - base_metrics) / base_metrics * 100))
entity_index_path = os.path.join(args.data_dir, 'entity2id.txt')
relation_index_path = os.path.join(args.data_dir, 'relation2id.txt')
if 'NELL' in args.data_dir:
adj_list_path = os.path.join(args.data_dir, 'adj_list.pkl')
seen_entities = data_utils.load_seen_entities(adj_list_path, entity_index_path)
else:
seen_entities = set()
dataset = os.path.basename(args.data_dir)
dev_path = os.path.join(args.data_dir, 'dev.triples')
dev_data = data_utils.load_triples(
dev_path, entity_index_path, relation_index_path, seen_entities=seen_entities, verbose=False)
to_m_rels, to_1_rels, (to_m_ratio, to_1_ratio) = data_utils.get_relations_by_type(args.data_dir, relation_index_path)
relation_by_types = (to_m_rels, to_1_rels)
to_m_ratio *= 100
to_1_ratio *= 100
seen_queries, (seen_ratio, unseen_ratio) = data_utils.get_seen_queries(args.data_dir, entity_index_path, relation_index_path)
seen_ratio *= 100
unseen_ratio *= 100
systems = ['ours', '-ad', '-rs']
mrrs, to_m_mrrs, to_1_mrrs, seen_mrrs, unseen_mrrs = {}, {}, {}, {}, {}
for system in systems:
print('** Evaluating {} system **'.format(system))
if system == '-ad':
args.action_dropout_rate = 0.0
if dataset == 'umls':
# adjust dropout hyperparameters
args.emb_dropout_rate = 0.3
args.ff_dropout_rate = 0.1
elif system == '-rs':
config_path = os.path.join('configs', '{}.sh'.format(dataset.lower()))
args = parser.parse_args()
args = data_utils.load_configs(args, config_path)
lf = set_up_lf_for_inference(args)
pred_scores = lf.forward(dev_data, verbose=False)
_, _, _, _, mrr = src.eval.hits_and_ranks(dev_data, pred_scores, lf.kg.dev_objects, verbose=True)
if to_1_ratio == 0:
to_m_mrr = mrr
to_1_mrr = -1
else:
to_m_mrr, to_1_mrr = src.eval.hits_and_ranks_by_relation_type(
dev_data, pred_scores, lf.kg.dev_objects, relation_by_types, verbose=True)
seen_mrr, unseen_mrr = src.eval.hits_and_ranks_by_seen_queries(
dev_data, pred_scores, lf.kg.dev_objects, seen_queries, verbose=True)
mrrs[system] = {'': mrr * 100}
to_m_mrrs[system] = {'': to_m_mrr * 100}
to_1_mrrs[system] = {'': to_1_mrr * 100}
seen_mrrs[system] = {'': seen_mrr * 100}
unseen_mrrs[system] = {'': unseen_mrr * 100}
_, _, _, _, mrr_full_kg = src.eval.hits_and_ranks(dev_data, pred_scores, lf.kg.all_objects, verbose=True)
if to_1_ratio == 0:
to_m_mrr_full_kg = mrr_full_kg
to_1_mrr_full_kg = -1
else:
to_m_mrr_full_kg, to_1_mrr_full_kg = src.eval.hits_and_ranks_by_relation_type(
dev_data, pred_scores, lf.kg.all_objects, relation_by_types, verbose=True)
seen_mrr_full_kg, unseen_mrr_full_kg = src.eval.hits_and_ranks_by_seen_queries(
dev_data, pred_scores, lf.kg.all_objects, seen_queries, verbose=True)
mrrs[system]['full_kg'] = mrr_full_kg * 100
to_m_mrrs[system]['full_kg'] = to_m_mrr_full_kg * 100
to_1_mrrs[system]['full_kg'] = to_1_mrr_full_kg * 100
seen_mrrs[system]['full_kg'] = seen_mrr_full_kg * 100
unseen_mrrs[system]['full_kg'] = unseen_mrr_full_kg * 100
# overall system comparison (table 3)
print('Partial graph evaluation')
print('--------------------------')
print('Overall system performance')
print('Ours(ConvE)\t-RS\t-AD')
print('{:.1f}\t{:.1f}\t{:.1f}'.format(mrrs['ours'][''], mrrs['-rs'][''], mrrs['-ad']['']))
print('--------------------------')
# performance w.r.t. relation types (table 4, 6)
print('Performance w.r.t. relation types')
print('\tTo-many\t\t\t\tTo-one\t\t')
print('%\tOurs\t-RS\t-AD\t%\tOurs\t-RS\t-AD')
print('{:.1f}\t{:.1f}\t{:.1f} ({:d})\t{:.1f} ({:d})\t{:.1f}\t{:.1f}\t{:.1f} ({:d})\t{:.1f} ({:d})'.format(
to_m_ratio, to_m_mrrs['ours'][''], to_m_mrrs['-rs'][''], rel_change(to_m_mrrs, '-rs', ''), to_m_mrrs['-ad'][''], rel_change(to_m_mrrs, '-ad', ''),
to_1_ratio, to_1_mrrs['ours'][''], to_1_mrrs['-rs'][''], rel_change(to_1_mrrs, '-rs', ''), to_1_mrrs['-ad'][''], rel_change(to_1_mrrs, '-ad', '')))
print('--------------------------')
# performance w.r.t. seen queries (table 5, 7)
print('Performance w.r.t. seen/unseen queries')
print('\tSeen\t\t\t\tUnseen\t\t')
print('%\tOurs\t-RS\t-AD\t%\tOurs\t-RS\t-AD')
print('{:.1f}\t{:.1f}\t{:.1f} ({:d})\t{:.1f} ({:d})\t{:.1f}\t{:.1f}\t{:.1f} ({:d})\t{:.1f} ({:d})'.format(
seen_ratio, seen_mrrs['ours'][''], seen_mrrs['-rs'][''], rel_change(seen_mrrs, '-rs', ''), seen_mrrs['-ad'][''], rel_change(seen_mrrs, '-ad', ''),
unseen_ratio, unseen_mrrs['ours'][''], unseen_mrrs['-rs'][''], rel_change(unseen_mrrs, '-rs', ''), unseen_mrrs['-ad'][''], rel_change(unseen_mrrs, '-ad', '')))
print()
print('Full graph evaluation')
print('--------------------------')
print('Overall system performance')
print('Ours(ConvE)\t-RS\t-AD')
print('{:.1f}\t{:.1f}\t{:.1f}'.format(mrrs['ours']['full_kg'], mrrs['-rs']['full_kg'], mrrs['-ad']['full_kg']))
print('--------------------------')
print('Performance w.r.t. relation types')
print('\tTo-many\t\t\t\tTo-one\t\t')
print('%\tOurs\t-RS\t-AD\t%\tOurs\t-RS\t-AD')
print('{:.1f}\t{:.1f}\t{:.1f} ({:d})\t{:.1f} ({:d})\t{:.1f}\t{:.1f}\t{:.1f} ({:d})\t{:.1f} ({:d})'.format(
to_m_ratio, to_m_mrrs['ours']['full_kg'], to_m_mrrs['-rs']['full_kg'], rel_change(to_m_mrrs, '-rs', 'full_kg'), to_m_mrrs['-ad']['full_kg'], rel_change(to_m_mrrs, '-ad', 'full_kg'),
to_1_ratio, to_1_mrrs['ours']['full_kg'], to_1_mrrs['-rs']['full_kg'], rel_change(to_1_mrrs, '-rs', 'full_kg'), to_1_mrrs['-ad']['full_kg'], rel_change(to_1_mrrs, '-ad', 'full_kg')))
print('--------------------------')
print('Performance w.r.t. seen/unseen queries')
print('\tSeen\t\t\t\tUnseen\t\t')
print('%\tOurs\t-RS\t-AD\t%\tOurs\t-RS\t-AD')
print('{:.1f}\t{:.1f}\t{:.1f} ({:d})\t{:.1f} ({:d})\t{:.1f}\t{:.1f}\t{:.1f} ({:d})\t{:.1f} ({:d})'.format(
seen_ratio, seen_mrrs['ours']['full_kg'], seen_mrrs['-rs']['full_kg'], rel_change(seen_mrrs, '-rs', 'full_kg'), seen_mrrs['-ad']['full_kg'], rel_change(seen_mrrs, '-ad', 'full_kg'),
unseen_ratio, unseen_mrrs['ours']['full_kg'], unseen_mrrs['-rs']['full_kg'], rel_change(unseen_mrrs, '-rs', 'full_kg'), unseen_mrrs['-ad']['full_kg'], rel_change(unseen_mrrs, '-ad', 'full_kg')))
def export_to_embedding_projector(lf):
lf.load_checkpoint(get_checkpoint_path(args))
lf.export_to_embedding_projector()
def export_reward_shaping_parameters(lf):
lf.load_checkpoint(get_checkpoint_path(args))
lf.export_reward_shaping_parameters()
def export_fuzzy_facts(lf):
lf.load_checkpoint(get_checkpoint_path(args))
lf.export_fuzzy_facts()
def export_error_cases(lf):
lf.load_checkpoint(get_checkpoint_path(args))
lf.batch_size = args.dev_batch_size
lf.eval()
entity_index_path = os.path.join(args.data_dir, 'entity2id.txt')
relation_index_path = os.path.join(args.data_dir, 'relation2id.txt')
dev_path = os.path.join(args.data_dir, 'dev.triples')
dev_data = data_utils.load_triples(dev_path, entity_index_path, relation_index_path)
lf.load_checkpoint(get_checkpoint_path(args))
print('Dev set performance:')
pred_scores = lf.forward(dev_data, verbose=False)
src.eval.hits_and_ranks(dev_data, pred_scores, lf.kg.dev_objects, verbose=True)
src.eval.export_error_cases(dev_data, pred_scores, lf.kg.dev_objects, os.path.join(lf.model_dir, 'error_cases.pkl'))
def compute_fact_scores(lf):
data_dir = args.data_dir
train_path = os.path.join(data_dir, 'train.triples')
dev_path = os.path.join(data_dir, 'dev.triples')
test_path = os.path.join(data_dir, 'test.triples')
entity_index_path = os.path.join(args.data_dir, 'entity2id.txt')
relation_index_path = os.path.join(args.data_dir, 'relation2id.txt')
train_data = data_utils.load_triples(train_path, entity_index_path, relation_index_path)
dev_data = data_utils.load_triples(dev_path, entity_index_path, relation_index_path)
test_data = data_utils.load_triples(test_path, entity_index_path, relation_index_path)
lf.eval()
lf.load_checkpoint(get_checkpoint_path(args))
train_scores = lf.forward_fact(train_data)
dev_scores = lf.forward_fact(dev_data)
test_scores = lf.forward_fact(test_data)
print('Train set average fact score: {}'.format(float(train_scores.mean())))
print('Dev set average fact score: {}'.format(float(dev_scores.mean())))
print('Test set average fact score: {}'.format(float(test_scores.mean())))
def get_checkpoint_path(args):
if not args.checkpoint_path:
return os.path.join(args.model_dir, 'model_best.tar')
else:
return args.checkpoint_path
def load_configs(config_path):
with open(config_path) as f:
print('loading configuration file {}'.format(config_path))
for line in f:
if not '=' in line:
continue
arg_name, arg_value = line.strip().split('=')
if arg_value.startswith('"') and arg_value.endswith('"'):
arg_value = arg_value[1:-1]
if hasattr(args, arg_name):
print('{} = {}'.format(arg_name, arg_value))
arg_value2 = getattr(args, arg_name)
if type(arg_value2) is str:
setattr(args, arg_name, arg_value)
elif type(arg_value2) is bool:
if arg_value == 'True':
setattr(args, arg_name, True)
elif arg_value == 'False':
setattr(args, arg_name, False)
else:
raise ValueError('Unrecognized boolean value description: {}'.format(arg_value))
elif type(arg_value2) is int:
setattr(args, arg_name, int(arg_value))
elif type(arg_value2) is float:
setattr(args, arg_name, float(arg_value))
else:
raise ValueError('Unrecognized attribute type: {}: {}'.format(arg_name, type(arg_value2)))
else:
raise ValueError('Unrecognized argument: {}'.format(arg_name))
return args
def run_experiment(args):
if args.test:
if 'NELL' in args.data_dir:
dataset = os.path.basename(args.data_dir)
args.distmult_state_dict_path = data_utils.change_to_test_model_path(dataset, args.distmult_state_dict_path)
args.complex_state_dict_path = data_utils.change_to_test_model_path(dataset, args.complex_state_dict_path)
args.conve_state_dict_path = data_utils.change_to_test_model_path(dataset, args.conve_state_dict_path)
args.data_dir += '.test'
if args.process_data:
# Process knowledge graph data
process_data()
else:
with torch.set_grad_enabled(args.train or args.search_random_seed or args.grid_search):
if args.search_random_seed:
# Search for best random seed
# search log file
task = os.path.basename(os.path.normpath(args.data_dir))
out_log = '{}.{}.rss'.format(task, args.model)
o_f = open(out_log, 'w')
print('** Search Random Seed **')
o_f.write('** Search Random Seed **\n')
o_f.close()
num_runs = 5
hits_at_1s = {}
hits_at_10s = {}
mrrs = {}
mrrs_search = {}
for i in range(num_runs):
o_f = open(out_log, 'a')
random_seed = random.randint(0, 1e16)
print("\nRandom seed = {}\n".format(random_seed))
o_f.write("\nRandom seed = {}\n\n".format(random_seed))
torch.manual_seed(random_seed)
torch.cuda.manual_seed_all(args, random_seed)
initialize_model_directory(args, random_seed)
lf = construct_model(args)
lf.cuda()
train(lf)
metrics = inference(lf)
hits_at_1s[random_seed] = metrics['test']['hits_at_1']
hits_at_10s[random_seed] = metrics['test']['hits_at_10']
mrrs[random_seed] = metrics['test']['mrr']
mrrs_search[random_seed] = metrics['dev']['mrr']
# print the results of the hyperparameter combinations searched so far
print('------------------------------------------')
print('Random Seed\t@1\t@10\tMRR')
for key in hits_at_1s:
print('{}\t{:.3f}\t{:.3f}\t{:.3f}'.format(
key, hits_at_1s[key], hits_at_10s[key], mrrs[key]))
print('------------------------------------------')
o_f.write('------------------------------------------\n')
o_f.write('Random Seed\t@1\t@10\tMRR\n')
for key in hits_at_1s:
o_f.write('{}\t{:.3f}\t{:.3f}\t{:.3f}\n'.format(
key, hits_at_1s[key], hits_at_10s[key], mrrs[key]))
o_f.write('------------------------------------------\n')
# compute result variance
import numpy as np
hits_at_1s_ = list(hits_at_1s.values())
hits_at_10s_ = list(hits_at_10s.values())
mrrs_ = list(mrrs.values())
print('Hits@1 mean: {:.3f}\tstd: {:.6f}'.format(np.mean(hits_at_1s_), np.std(hits_at_1s_)))
print('Hits@10 mean: {:.3f}\tstd: {:.6f}'.format(np.mean(hits_at_10s_), np.std(hits_at_10s_)))
print('MRR mean: {:.3f}\tstd: {:.6f}'.format(np.mean(mrrs_), np.std(mrrs_)))
o_f.write('Hits@1 mean: {:.3f}\tstd: {:.6f}\n'.format(np.mean(hits_at_1s_), np.std(hits_at_1s_)))
o_f.write('Hits@10 mean: {:.3f}\tstd: {:.6f}\n'.format(np.mean(hits_at_10s_), np.std(hits_at_10s_)))
o_f.write('MRR mean: {:.3f}\tstd: {:.6f}\n'.format(np.mean(mrrs_), np.std(mrrs_)))
o_f.close()
# find best random seed
best_random_seed, best_mrr = sorted(mrrs_search.items(), key=lambda x: x[1], reverse=True)[0]
print('* Best Random Seed = {}'.format(best_random_seed))
print('* @1: {:.3f}\t@10: {:.3f}\tMRR: {:.3f}'.format(
hits_at_1s[best_random_seed],
hits_at_10s[best_random_seed],
mrrs[best_random_seed]))
with open(out_log, 'a'):
o_f.write('* Best Random Seed = {}\n'.format(best_random_seed))
o_f.write('* @1: {:.3f}\t@10: {:.3f}\tMRR: {:.3f}\n'.format(
hits_at_1s[best_random_seed],
hits_at_10s[best_random_seed],
mrrs[best_random_seed])
)
o_f.close()
elif args.grid_search:
# Grid search
# search log file
task = os.path.basename(os.path.normpath(args.data_dir))
out_log = '{}.{}.gs'.format(task, args.model)
o_f = open(out_log, 'w')
print("** Grid Search **")
o_f.write("** Grid Search **\n")
hyperparameters = args.tune.split(',')
if args.tune == '' or len(hyperparameters) < 1:
print("No hyperparameter specified.")
sys.exit(0)
grid = hp_range[hyperparameters[0]]
for hp in hyperparameters[1:]:
grid = itertools.product(grid, hp_range[hp])
hits_at_1s = {}
hits_at_10s = {}
mrrs = {}
grid = list(grid)
print('* {} hyperparameter combinations to try'.format(len(grid)))
o_f.write('* {} hyperparameter combinations to try\n'.format(len(grid)))
o_f.close()
for i, grid_entry in enumerate(list(grid)):
o_f = open(out_log, 'a')
if not (type(grid_entry) is list or type(grid_entry) is list):
grid_entry = [grid_entry]
grid_entry = flatten(grid_entry)
print('* Hyperparameter Set {}:'.format(i))
o_f.write('* Hyperparameter Set {}:\n'.format(i))
signature = ''
for j in range(len(grid_entry)):
hp = hyperparameters[j]
value = grid_entry[j]
if hp == 'bandwidth':
setattr(args, hp, int(value))
else:
setattr(args, hp, float(value))
signature += ':{}'.format(value)
print('* {}: {}'.format(hp, value))
initialize_model_directory(args)
lf = construct_model(args)
lf.cuda()
train(lf)
metrics = inference(lf)
hits_at_1s[signature] = metrics['dev']['hits_at_1']
hits_at_10s[signature] = metrics['dev']['hits_at_10']
mrrs[signature] = metrics['dev']['mrr']
# print the results of the hyperparameter combinations searched so far
print('------------------------------------------')
print('Signature\t@1\t@10\tMRR')
for key in hits_at_1s:
print('{}\t{:.3f}\t{:.3f}\t{:.3f}'.format(
key, hits_at_1s[key], hits_at_10s[key], mrrs[key]))
print('------------------------------------------\n')
o_f.write('------------------------------------------\n')
o_f.write('Signature\t@1\t@10\tMRR\n')
for key in hits_at_1s:
o_f.write('{}\t{:.3f}\t{:.3f}\t{:.3f}\n'.format(
key, hits_at_1s[key], hits_at_10s[key], mrrs[key]))
o_f.write('------------------------------------------\n')
# find best hyperparameter set
best_signature, best_mrr = sorted(mrrs.items(), key=lambda x:x[1], reverse=True)[0]
print('* best hyperparameter set')
o_f.write('* best hyperparameter set\n')
best_hp_values = best_signature.split(':')[1:]
for i, value in enumerate(best_hp_values):
hp_name = hyperparameters[i]
hp_value = best_hp_values[i]
print('* {}: {}'.format(hp_name, hp_value))
print('* @1: {:.3f}\t@10: {:.3f}\tMRR: {:.3f}'.format(
hits_at_1s[best_signature],
hits_at_10s[best_signature],
mrrs[best_signature]
))
o_f.write('* @1: {:.3f}\t@10: {:.3f}\tMRR: {:.3f}\ns'.format(
hits_at_1s[best_signature],
hits_at_10s[best_signature],
mrrs[best_signature]
))
o_f.close()
elif args.run_ablation_studies:
run_ablation_studies(args)
else:
initialize_model_directory(args)
lf = construct_model(args)
lf.cuda()
if args.train:
train(lf)
elif args.inference:
inference(lf)
elif args.eval_by_relation_type:
inference(lf)
elif args.eval_by_seen_queries:
inference(lf)
elif args.export_to_embedding_projector:
export_to_embedding_projector(lf)
elif args.export_reward_shaping_parameters:
export_reward_shaping_parameters(lf)
elif args.compute_fact_scores:
compute_fact_scores(lf)
elif args.export_fuzzy_facts:
export_fuzzy_facts(lf)
elif args.export_error_cases:
export_error_cases(lf)
if __name__ == '__main__':
run_experiment(args)