sampler.py 23.9 KB
Newer Older
1
2
3
4
5
6
import math
import numpy as np
import scipy as sp
import dgl.backend as F
import dgl
import os
7
import sys
8
9
10
import pickle
import time

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
def SoftRelationPartition(edges, n, threshold=0.05):
    """This partitions a list of edges to n partitions according to their
    relation types. For any relation with number of edges larger than the
    threshold, its edges will be evenly distributed into all partitions.
    For any relation with number of edges smaller than the threshold, its
    edges will be put into one single partition.

    Algo:
    For r in relations:
        if r.size() > threadold
            Evenly divide edges of r into n parts and put into each relation.
        else
            Find partition with fewest edges, and put edges of r into 
            this partition.

    Parameters
    ----------
    edges : (heads, rels, tails) triple
        Edge list to partition
    n : int
        Number of partitions
    threshold : float
        The threshold of whether a relation is LARGE or SMALL
        Default: 5%

    Returns
    -------
    List of np.array
        Edges of each partition
    List of np.array
        Edge types of each partition
    bool
        Whether there exists some relations belongs to multiple partitions
    """
    heads, rels, tails = edges
    print('relation partition {} edges into {} parts'.format(len(heads), n))
    uniq, cnts = np.unique(rels, return_counts=True)
    idx = np.flip(np.argsort(cnts))
    cnts = cnts[idx]
    uniq = uniq[idx]
    assert cnts[0] > cnts[-1]
    edge_cnts = np.zeros(shape=(n,), dtype=np.int64)
    rel_cnts = np.zeros(shape=(n,), dtype=np.int64)
    rel_dict = {}
    rel_parts = []
    cross_rel_part = []
    for _ in range(n):
        rel_parts.append([])

    large_threshold = int(len(rels) * threshold)
    capacity_per_partition = int(len(rels) / n)
    # ensure any relation larger than the partition capacity will be split
    large_threshold = capacity_per_partition if capacity_per_partition < large_threshold \
                      else large_threshold
    num_cross_part = 0
    for i in range(len(cnts)):
        cnt = cnts[i]
        r = uniq[i]
        r_parts = []
        if cnt > large_threshold:
            avg_part_cnt = (cnt // n) + 1
            num_cross_part += 1
            for j in range(n):
                part_cnt = avg_part_cnt if cnt > avg_part_cnt else cnt
                r_parts.append([j, part_cnt])
                rel_parts[j].append(r)
                edge_cnts[j] += part_cnt
                rel_cnts[j] += 1
                cnt -= part_cnt
            cross_rel_part.append(r)
        else:
            idx = np.argmin(edge_cnts)
            r_parts.append([idx, cnt])
            rel_parts[idx].append(r)
            edge_cnts[idx] += cnt
            rel_cnts[idx] += 1
        rel_dict[r] = r_parts

    for i, edge_cnt in enumerate(edge_cnts):
        print('part {} has {} edges and {} relations'.format(i, edge_cnt, rel_cnts[i]))
    print('{}/{} duplicated relation across partitions'.format(num_cross_part, len(cnts)))

    parts = []
    for i in range(n):
        parts.append([])
        rel_parts[i] = np.array(rel_parts[i])

    for i, r in enumerate(rels):
        r_part = rel_dict[r][0]
        part_idx = r_part[0]
        cnt = r_part[1]
        parts[part_idx].append(i)
        cnt -= 1
        if cnt == 0:
            rel_dict[r].pop(0)
        else:
            rel_dict[r][0][1] = cnt

    for i, part in enumerate(parts):
        parts[i] = np.array(part, dtype=np.int64)
    shuffle_idx = np.concatenate(parts)
    heads[:] = heads[shuffle_idx]
    rels[:] = rels[shuffle_idx]
    tails[:] = tails[shuffle_idx]

    off = 0
    for i, part in enumerate(parts):
        parts[i] = np.arange(off, off + len(part))
        off += len(part)
    cross_rel_part = np.array(cross_rel_part)

    return parts, rel_parts, num_cross_part > 0, cross_rel_part

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
def BalancedRelationPartition(edges, n):
    """This partitions a list of edges based on relations to make sure
    each partition has roughly the same number of edges and relations.
    Algo:
    For r in relations:
      Find partition with fewest edges
      if r.size() > num_of empty_slot
         put edges of r into this partition to fill the partition,
         find next partition with fewest edges to put r in.
      else
         put edges of r into this partition.

    Parameters
    ----------
    edges : (heads, rels, tails) triple
        Edge list to partition
    n : int
        number of partitions

    Returns
    -------
    List of np.array
        Edges of each partition
    List of np.array
        Edge types of each partition
    bool
        Whether there exists some relations belongs to multiple partitions
    """
152
153
154
    heads, rels, tails = edges
    print('relation partition {} edges into {} parts'.format(len(heads), n))
    uniq, cnts = np.unique(rels, return_counts=True)
155
156
157
158
159
160
161
    idx = np.flip(np.argsort(cnts))
    cnts = cnts[idx]
    uniq = uniq[idx]
    assert cnts[0] > cnts[-1]
    edge_cnts = np.zeros(shape=(n,), dtype=np.int64)
    rel_cnts = np.zeros(shape=(n,), dtype=np.int64)
    rel_dict = {}
162
163
164
    rel_parts = []
    for _ in range(n):
        rel_parts.append([])
165
166
167

    max_edges = (len(rels) // n) + 1
    num_cross_part = 0
168
169
170
    for i in range(len(cnts)):
        cnt = cnts[i]
        r = uniq[i]
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
        r_parts = []

        while cnt > 0:
            idx = np.argmin(edge_cnts)
            if edge_cnts[idx] + cnt <= max_edges:
                r_parts.append([idx, cnt])
                rel_parts[idx].append(r)
                edge_cnts[idx] += cnt
                rel_cnts[idx] += 1
                cnt = 0
            else:
                cur_cnt = max_edges - edge_cnts[idx]
                r_parts.append([idx, cur_cnt])
                rel_parts[idx].append(r)
                edge_cnts[idx] += cur_cnt
                rel_cnts[idx] += 1
                num_cross_part += 1
                cnt -= cur_cnt
        rel_dict[r] = r_parts

191
192
    for i, edge_cnt in enumerate(edge_cnts):
        print('part {} has {} edges and {} relations'.format(i, edge_cnt, rel_cnts[i]))
193
    print('{}/{} duplicated relation across partitions'.format(num_cross_part, len(cnts)))
194

195
    parts = []
196
    for i in range(n):
197
        parts.append([])
198
        rel_parts[i] = np.array(rel_parts[i])
199

200
    for i, r in enumerate(rels):
201
202
203
        r_part = rel_dict[r][0]
        part_idx = r_part[0]
        cnt = r_part[1]
204
        parts[part_idx].append(i)
205
206
207
208
209
210
        cnt -= 1
        if cnt == 0:
            rel_dict[r].pop(0)
        else:
            rel_dict[r][0][1] = cnt

211
212
    for i, part in enumerate(parts):
        parts[i] = np.array(part, dtype=np.int64)
213
214
215
216
217
218
219
220
221
222
223
    shuffle_idx = np.concatenate(parts)
    heads[:] = heads[shuffle_idx]
    rels[:] = rels[shuffle_idx]
    tails[:] = tails[shuffle_idx]

    off = 0
    for i, part in enumerate(parts):
        parts[i] = np.arange(off, off + len(part))
        off += len(part)

    return parts, rel_parts, num_cross_part > 0
224
225

def RandomPartition(edges, n):
226
227
228
229
230
231
232
233
234
235
236
237
238
239
    """This partitions a list of edges randomly across n partitions

    Parameters
    ----------
    edges : (heads, rels, tails) triple
        Edge list to partition
    n : int
        number of partitions

    Returns
    -------
    List of np.array
        Edges of each partition
    """
240
241
242
    heads, rels, tails = edges
    print('random partition {} edges into {} parts'.format(len(heads), n))
    idx = np.random.permutation(len(heads))
243
244
245
246
    heads[:] = heads[idx]
    rels[:] = rels[idx]
    tails[:] = tails[idx]

247
248
249
250
251
    part_size = int(math.ceil(len(idx) / n))
    parts = []
    for i in range(n):
        start = part_size * i
        end = min(part_size * (i + 1), len(idx))
252
253
        parts.append(idx[start:end])
        print('part {} has {} edges'.format(i, len(parts[-1])))
254
255
    return parts

256
def ConstructGraph(edges, n_entities, args):
257
258
259
260
261
262
263
264
265
266
267
    """Construct Graph for training

    Parameters
    ----------
    edges : (heads, rels, tails) triple
        Edge list
    n_entities : int
        number of entities
    args :
        Global configs.
    """
268
    pickle_name = 'graph_train.pickle'
269
270
271
272
273
    if args.pickle_graph and os.path.exists(os.path.join(args.data_path, args.dataset, pickle_name)):
        with open(os.path.join(args.data_path, args.dataset, pickle_name), 'rb') as graph_file:
            g = pickle.load(graph_file)
            print('Load pickled graph.')
    else:
274
        src, etype_id, dst = edges
275
276
        coo = sp.sparse.coo_matrix((np.ones(len(src)), (src, dst)), shape=[n_entities, n_entities])
        g = dgl.DGLGraph(coo, readonly=True, sort_csr=True)
277
        g.edata['tid'] = F.tensor(etype_id, F.int64)
278
279
280
281
282
283
        if args.pickle_graph:
            with open(os.path.join(args.data_path, args.dataset, pickle_name), 'wb') as graph_file:
                pickle.dump(g, graph_file)
    return g

class TrainDataset(object):
284
285
286
287
288
289
290
291
292
293
294
295
    """Dataset for training

    Parameters
    ----------
    dataset : KGDataset
        Original dataset.
    args :
        Global configs.
    ranks:
        Number of partitions.
    """
    def __init__(self, dataset, args, ranks=64):
296
        triples = dataset.train
297
298
        num_train = len(triples[0])
        print('|Train|:', num_train)
299

300
301
302
303
        if ranks > 1 and args.soft_rel_part:
            self.edge_parts, self.rel_parts, self.cross_part, self.cross_rels = \
            SoftRelationPartition(triples, ranks)
        elif ranks > 1 and args.rel_part:
304
305
            self.edge_parts, self.rel_parts, self.cross_part = \
                BalancedRelationPartition(triples, ranks)
306
        elif ranks > 1:
307
            self.edge_parts = RandomPartition(triples, ranks)
308
            self.cross_part = True
309
        else:
310
            self.edge_parts = [np.arange(num_train)]
311
312
            self.rel_parts = [np.arange(dataset.n_relations)]
            self.cross_part = False
313

314
        self.g = ConstructGraph(triples, dataset.n_entities, args)
315

316
    def create_sampler(self, batch_size, neg_sample_size=2, neg_chunk_size=None, mode='head', num_workers=32,
317
                       shuffle=True, exclude_positive=False, rank=0):
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
        """Create sampler for training

        Parameters
        ----------
        batch_size : int
            Batch size of each mini batch.
        neg_sample_size : int
            How many negative edges sampled for each node.
        neg_chunk_size : int
            How many edges in one chunk. We split one batch into chunks.
        mode : str
            Sampling mode.
        number_workers: int
            Number of workers used in parallel for this sampler
        shuffle : bool
            If True, shuffle the seed edges.
            If False, do not shuffle the seed edges.
            Default: False
        exclude_positive : bool
            If True, exlucde true positive edges in sampled negative edges
            If False, return all sampled negative edges even there are positive edges
            Default: False
        rank : int
            Which partition to sample.

        Returns
        -------
        dgl.contrib.sampling.EdgeSampler
            Edge sampler
        """
348
        EdgeSampler = getattr(dgl.contrib.sampling, 'EdgeSampler')
349
        assert batch_size % neg_sample_size == 0, 'batch_size should be divisible by B'
350
351
        return EdgeSampler(self.g,
                           seed_edges=F.tensor(self.edge_parts[rank]),
352
                           batch_size=batch_size,
353
                           neg_sample_size=int(neg_sample_size/neg_chunk_size),
354
                           chunk_size=neg_chunk_size,
355
356
357
358
359
360
                           negative_mode=mode,
                           num_workers=num_workers,
                           shuffle=shuffle,
                           exclude_positive=exclude_positive,
                           return_false_neg=False)

361

362
class ChunkNegEdgeSubgraph(dgl.subgraph.DGLSubGraph):
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
    """Wrapper for negative graph

        Parameters
        ----------
        neg_g : DGLGraph
            Graph holding negative edges.
        num_chunks : int
            Number of chunks in sampled graph.
        chunk_size : int
            Info of chunk_size.
        neg_sample_size : int
            Info of neg_sample_size.
        neg_head : bool
            If True, negative_mode is 'head'
            If False, negative_mode is 'tail'
    """
379
380
    def __init__(self, subg, num_chunks, chunk_size,
                 neg_sample_size, neg_head):
381
        super(ChunkNegEdgeSubgraph, self).__init__(subg._parent, subg.sgi)
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
        self.subg = subg
        self.num_chunks = num_chunks
        self.chunk_size = chunk_size
        self.neg_sample_size = neg_sample_size
        self.neg_head = neg_head

    @property
    def head_nid(self):
        return self.subg.head_nid

    @property
    def tail_nid(self):
        return self.subg.tail_nid


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
def create_neg_subgraph(pos_g, neg_g, chunk_size, neg_sample_size, is_chunked,
                        neg_head, num_nodes):
    """KG models need to know the number of chunks, the chunk size and negative sample size
    of a negative subgraph to perform the computation more efficiently.
    This function tries to infer all of these information of the negative subgraph
    and create a wrapper class that contains all of the information.

    Parameters
    ----------
    pos_g : DGLGraph
        Graph holding positive edges.
    neg_g : DGLGraph
        Graph holding negative edges.
    chunk_size : int
        Chunk size of negative subgrap.
    neg_sample_size : int
        Negative sample size of negative subgrap.
    is_chunked : bool
        If True, the sampled batch is chunked.
    neg_head : bool
        If True, negative_mode is 'head'
        If False, negative_mode is 'tail'
    num_nodes: int
        Total number of nodes in the whole graph.

    Returns
    -------
    ChunkNegEdgeSubgraph
        Negative graph wrapper
    """
427
428
429
430
431
432
433
    assert neg_g.number_of_edges() % pos_g.number_of_edges() == 0
    # We use all nodes to create negative edges. Regardless of the sampling algorithm,
    # we can always view the subgraph with one chunk.
    if (neg_head and len(neg_g.head_nid) == num_nodes) \
       or (not neg_head and len(neg_g.tail_nid) == num_nodes):
        num_chunks = 1
        chunk_size = pos_g.number_of_edges()
434
435
    elif is_chunked:
        if pos_g.number_of_edges() < chunk_size:
436
            return None
437
438
        else:
            # This is probably the last batch. Let's ignore it.
439
            if pos_g.number_of_edges() % chunk_size > 0:
440
                return None
441
            num_chunks = int(pos_g.number_of_edges() / chunk_size)
442
        assert num_chunks * chunk_size == pos_g.number_of_edges()
443
444
445
    else:
        num_chunks = pos_g.number_of_edges()
        chunk_size = 1
446
447
    return ChunkNegEdgeSubgraph(neg_g, num_chunks, chunk_size,
                                neg_sample_size, neg_head)
448
449

class EvalSampler(object):
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
    """Sampler for validation and testing

    Parameters
    ----------
    g : DGLGraph
        Graph containing KG graph
    edges : tensor
        Seed edges
    batch_size : int
        Batch size of each mini batch.
    neg_sample_size : int
        How many negative edges sampled for each node.
    neg_chunk_size : int
        How many edges in one chunk. We split one batch into chunks.
    mode : str
        Sampling mode.
    number_workers: int
        Number of workers used in parallel for this sampler
    filter_false_neg : bool
        If True, exlucde true positive edges in sampled negative edges
        If False, return all sampled negative edges even there are positive edges
        Default: True
    """
    def __init__(self, g, edges, batch_size, neg_sample_size, neg_chunk_size, mode, num_workers=32,
                 filter_false_neg=True):
475
476
477
478
479
        EdgeSampler = getattr(dgl.contrib.sampling, 'EdgeSampler')
        self.sampler = EdgeSampler(g,
                                   batch_size=batch_size,
                                   seed_edges=edges,
                                   neg_sample_size=neg_sample_size,
480
                                   chunk_size=neg_chunk_size,
481
482
483
484
                                   negative_mode=mode,
                                   num_workers=num_workers,
                                   shuffle=False,
                                   exclude_positive=False,
485
                                   relations=g.edata['tid'],
486
                                   return_false_neg=filter_false_neg)
487
488
489
490
        self.sampler_iter = iter(self.sampler)
        self.mode = mode
        self.neg_head = 'head' in mode
        self.g = g
491
        self.filter_false_neg = filter_false_neg
492
        self.neg_chunk_size = neg_chunk_size
493
        self.neg_sample_size = neg_sample_size
494
495
496
497
498

    def __iter__(self):
        return self

    def __next__(self):
499
500
501
502
503
504
505
506
507
        """Get next batch

        Returns
        -------
        DGLGraph
            Sampled positive graph
        ChunkNegEdgeSubgraph
            Negative graph wrapper
        """
508
509
        while True:
            pos_g, neg_g = next(self.sampler_iter)
510
511
            if self.filter_false_neg:
                neg_positive = neg_g.edata['false_neg']
512
513
514
515
516
517
            neg_g = create_neg_subgraph(pos_g, neg_g, 
                                        self.neg_chunk_size, 
                                        self.neg_sample_size, 
                                        'chunk' in self.mode, 
                                        self.neg_head, 
                                        self.g.number_of_nodes())
518
519
520
            if neg_g is not None:
                break

521
522
523
        pos_g.ndata['id'] = pos_g.parent_nid
        neg_g.ndata['id'] = neg_g.parent_nid
        pos_g.edata['id'] = pos_g._parent.edata['tid'][pos_g.parent_eid]
524
525
        if self.filter_false_neg:
            neg_g.edata['bias'] = F.astype(-neg_positive, F.float32)
526
527
528
        return pos_g, neg_g

    def reset(self):
529
530
        """Reset the sampler
        """
531
532
533
534
        self.sampler_iter = iter(self.sampler)
        return self

class EvalDataset(object):
535
536
537
538
539
540
541
542
543
    """Dataset for validation or testing

    Parameters
    ----------
    dataset : KGDataset
        Original dataset.
    args :
        Global configs.
    """
544
545
546
547
548
549
550
    def __init__(self, dataset, args):
        pickle_name = 'graph_all.pickle'
        if args.pickle_graph and os.path.exists(os.path.join(args.data_path, args.dataset, pickle_name)):
            with open(os.path.join(args.data_path, args.dataset, pickle_name), 'rb') as graph_file:
                g = pickle.load(graph_file)
                print('Load pickled graph.')
        else:
551
552
553
554
555
            src = np.concatenate((dataset.train[0], dataset.valid[0], dataset.test[0]))
            etype_id = np.concatenate((dataset.train[1], dataset.valid[1], dataset.test[1]))
            dst = np.concatenate((dataset.train[2], dataset.valid[2], dataset.test[2]))
            coo = sp.sparse.coo_matrix((np.ones(len(src)), (src, dst)),
                                       shape=[dataset.n_entities, dataset.n_entities])
556
            g = dgl.DGLGraph(coo, readonly=True, sort_csr=True)
557
            g.edata['tid'] = F.tensor(etype_id, F.int64)
558
559
560
561
            if args.pickle_graph:
                with open(os.path.join(args.data_path, args.dataset, pickle_name), 'wb') as graph_file:
                    pickle.dump(g, graph_file)
        self.g = g
562
563
564
        self.num_train = len(dataset.train[0])
        self.num_valid = len(dataset.valid[0])
        self.num_test = len(dataset.test[0])
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581

        if args.eval_percent < 1:
            self.valid = np.random.randint(0, self.num_valid,
                    size=(int(self.num_valid * args.eval_percent),)) + self.num_train
        else:
            self.valid = np.arange(self.num_train, self.num_train + self.num_valid)
        print('|valid|:', len(self.valid))

        if args.eval_percent < 1:
            self.test = np.random.randint(0, self.num_test,
                    size=(int(self.num_test * args.eval_percent,)))
            self.test += self.num_train + self.num_valid
        else:
            self.test = np.arange(self.num_train + self.num_valid, self.g.number_of_edges())
        print('|test|:', len(self.test))

    def get_edges(self, eval_type):
582
583
584
585
586
587
588
589
590
591
592
593
        """ Get all edges in this dataset

        Parameters
        ----------
        eval_type : str
            Sampling type, 'valid' for validation and 'test' for testing

        Returns
        -------
        np.array
            Edges
        """
594
595
596
597
598
599
600
        if eval_type == 'valid':
            return self.valid
        elif eval_type == 'test':
            return self.test
        else:
            raise Exception('get invalid type: ' + eval_type)

601
    def create_sampler(self, eval_type, batch_size, neg_sample_size, neg_chunk_size,
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
                       filter_false_neg, mode='head', num_workers=32, rank=0, ranks=1):
        """Create sampler for validation or testing

        Parameters
        ----------
        eval_type : str
            Sampling type, 'valid' for validation and 'test' for testing
        batch_size : int
            Batch size of each mini batch.
        neg_sample_size : int
            How many negative edges sampled for each node.
        neg_chunk_size : int
            How many edges in one chunk. We split one batch into chunks.
        filter_false_neg : bool
            If True, exlucde true positive edges in sampled negative edges
            If False, return all sampled negative edges even there are positive edges
        mode : str
            Sampling mode.
        number_workers: int
            Number of workers used in parallel for this sampler
        rank : int
            Which partition to sample.
        ranks : int
            Total number of partitions.

        Returns
        -------
        dgl.contrib.sampling.EdgeSampler
            Edge sampler
        """
632
633
634
635
        edges = self.get_edges(eval_type)
        beg = edges.shape[0] * rank // ranks
        end = min(edges.shape[0] * (rank + 1) // ranks, edges.shape[0])
        edges = edges[beg: end]
636
        return EvalSampler(self.g, edges, batch_size, neg_sample_size, neg_chunk_size,
637
                           mode, num_workers, filter_false_neg)
638
639

class NewBidirectionalOneShotIterator:
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
    """Grouped samper iterator

    Parameters
    ----------
    dataloader_head : dgl.contrib.sampling.EdgeSampler
        EdgeSampler in head mode
    dataloader_tail : dgl.contrib.sampling.EdgeSampler
        EdgeSampler in tail mode
    neg_chunk_size : int
        How many edges in one chunk. We split one batch into chunks.
    neg_sample_size : int
        How many negative edges sampled for each node.
    is_chunked : bool
        If True, the sampled batch is chunked.
    num_nodes : int
        Total number of nodes in the whole graph.
    """
    def __init__(self, dataloader_head, dataloader_tail, neg_chunk_size, neg_sample_size,
                 is_chunked, num_nodes):
659
660
        self.sampler_head = dataloader_head
        self.sampler_tail = dataloader_tail
661
662
        self.iterator_head = self.one_shot_iterator(dataloader_head, neg_chunk_size,
                                                    neg_sample_size, is_chunked,
663
                                                    True, num_nodes)
664
665
        self.iterator_tail = self.one_shot_iterator(dataloader_tail, neg_chunk_size,
                                                    neg_sample_size, is_chunked,
666
667
668
669
670
671
672
673
674
675
676
677
                                                    False, num_nodes)
        self.step = 0

    def __next__(self):
        self.step += 1
        if self.step % 2 == 0:
            pos_g, neg_g = next(self.iterator_head)
        else:
            pos_g, neg_g = next(self.iterator_tail)
        return pos_g, neg_g

    @staticmethod
678
679
    def one_shot_iterator(dataloader, neg_chunk_size, neg_sample_size, is_chunked,
                          neg_head, num_nodes):
680
681
        while True:
            for pos_g, neg_g in dataloader:
682
683
                neg_g = create_neg_subgraph(pos_g, neg_g, neg_chunk_size, neg_sample_size,
                                            is_chunked, neg_head, num_nodes)
684
685
686
                if neg_g is None:
                    continue

687
688
689
                pos_g.ndata['id'] = pos_g.parent_nid
                neg_g.ndata['id'] = neg_g.parent_nid
                pos_g.edata['id'] = pos_g._parent.edata['tid'][pos_g.parent_eid]
690
                yield pos_g, neg_g