File size: 39,679 Bytes
cdfecf8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from mmcv.cnn import ConvModule, xavier_init
from mmcv.runner import force_fp32

from mmdet.core import build_sampler, fast_nms, images_to_levels, multi_apply
from ..builder import HEADS, build_loss
from .anchor_head import AnchorHead


@HEADS.register_module()
class YOLACTHead(AnchorHead):
    """YOLACT box head used in https://arxiv.org/abs/1904.02689.

    Note that YOLACT head is a light version of RetinaNet head.
    Four differences are described as follows:

    1. YOLACT box head has three-times fewer anchors.
    2. YOLACT box head shares the convs for box and cls branches.
    3. YOLACT box head uses OHEM instead of Focal loss.
    4. YOLACT box head predicts a set of mask coefficients for each box.

    Args:
        num_classes (int): Number of categories excluding the background
            category.
        in_channels (int): Number of channels in the input feature map.
        anchor_generator (dict): Config dict for anchor generator
        loss_cls (dict): Config of classification loss.
        loss_bbox (dict): Config of localization loss.
        num_head_convs (int): Number of the conv layers shared by
            box and cls branches.
        num_protos (int): Number of the mask coefficients.
        use_ohem (bool): If true, ``loss_single_OHEM`` will be used for
            cls loss calculation. If false, ``loss_single`` will be used.
        conv_cfg (dict): Dictionary to construct and config conv layer.
        norm_cfg (dict): Dictionary to construct and config norm layer.
    """

    def __init__(self,
                 num_classes,
                 in_channels,
                 anchor_generator=dict(
                     type='AnchorGenerator',
                     octave_base_scale=3,
                     scales_per_octave=1,
                     ratios=[0.5, 1.0, 2.0],
                     strides=[8, 16, 32, 64, 128]),
                 loss_cls=dict(
                     type='CrossEntropyLoss',
                     use_sigmoid=False,
                     reduction='none',
                     loss_weight=1.0),
                 loss_bbox=dict(
                     type='SmoothL1Loss', beta=1.0, loss_weight=1.5),
                 num_head_convs=1,
                 num_protos=32,
                 use_ohem=True,
                 conv_cfg=None,
                 norm_cfg=None,
                 **kwargs):
        self.num_head_convs = num_head_convs
        self.num_protos = num_protos
        self.use_ohem = use_ohem
        self.conv_cfg = conv_cfg
        self.norm_cfg = norm_cfg
        super(YOLACTHead, self).__init__(
            num_classes,
            in_channels,
            loss_cls=loss_cls,
            loss_bbox=loss_bbox,
            anchor_generator=anchor_generator,
            **kwargs)
        if self.use_ohem:
            sampler_cfg = dict(type='PseudoSampler')
            self.sampler = build_sampler(sampler_cfg, context=self)
            self.sampling = False

    def _init_layers(self):
        """Initialize layers of the head."""
        self.relu = nn.ReLU(inplace=True)
        self.head_convs = nn.ModuleList()
        for i in range(self.num_head_convs):
            chn = self.in_channels if i == 0 else self.feat_channels
            self.head_convs.append(
                ConvModule(
                    chn,
                    self.feat_channels,
                    3,
                    stride=1,
                    padding=1,
                    conv_cfg=self.conv_cfg,
                    norm_cfg=self.norm_cfg))
        self.conv_cls = nn.Conv2d(
            self.feat_channels,
            self.num_anchors * self.cls_out_channels,
            3,
            padding=1)
        self.conv_reg = nn.Conv2d(
            self.feat_channels, self.num_anchors * 4, 3, padding=1)
        self.conv_coeff = nn.Conv2d(
            self.feat_channels,
            self.num_anchors * self.num_protos,
            3,
            padding=1)

    def init_weights(self):
        """Initialize weights of the head."""
        for m in self.head_convs:
            xavier_init(m.conv, distribution='uniform', bias=0)
        xavier_init(self.conv_cls, distribution='uniform', bias=0)
        xavier_init(self.conv_reg, distribution='uniform', bias=0)
        xavier_init(self.conv_coeff, distribution='uniform', bias=0)

    def forward_single(self, x):
        """Forward feature of a single scale level.

        Args:
            x (Tensor): Features of a single scale level.

        Returns:
            tuple:
                cls_score (Tensor): Cls scores for a single scale level \
                    the channels number is num_anchors * num_classes.
                bbox_pred (Tensor): Box energies / deltas for a single scale \
                    level, the channels number is num_anchors * 4.
                coeff_pred (Tensor): Mask coefficients for a single scale \
                    level, the channels number is num_anchors * num_protos.
        """
        for head_conv in self.head_convs:
            x = head_conv(x)
        cls_score = self.conv_cls(x)
        bbox_pred = self.conv_reg(x)
        coeff_pred = self.conv_coeff(x).tanh()
        return cls_score, bbox_pred, coeff_pred

    @force_fp32(apply_to=('cls_scores', 'bbox_preds'))
    def loss(self,
             cls_scores,
             bbox_preds,
             gt_bboxes,
             gt_labels,
             img_metas,
             gt_bboxes_ignore=None):
        """A combination of the func:``AnchorHead.loss`` and
        func:``SSDHead.loss``.

        When ``self.use_ohem == True``, it functions like ``SSDHead.loss``,
        otherwise, it follows ``AnchorHead.loss``. Besides, it additionally
        returns ``sampling_results``.

        Args:
            cls_scores (list[Tensor]): Box scores for each scale level
                Has shape (N, num_anchors * num_classes, H, W)
            bbox_preds (list[Tensor]): Box energies / deltas for each scale
                level with shape (N, num_anchors * 4, H, W)
            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
            gt_labels (list[Tensor]): Class indices corresponding to each box
            img_metas (list[dict]): Meta information of each image, e.g.,
                image size, scaling factor, etc.
            gt_bboxes_ignore (None | list[Tensor]): Specify which bounding
                boxes can be ignored when computing the loss. Default: None

        Returns:
            tuple:
                dict[str, Tensor]: A dictionary of loss components.
                List[:obj:``SamplingResult``]: Sampler results for each image.
        """
        featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores]
        assert len(featmap_sizes) == self.anchor_generator.num_levels

        device = cls_scores[0].device

        anchor_list, valid_flag_list = self.get_anchors(
            featmap_sizes, img_metas, device=device)
        label_channels = self.cls_out_channels if self.use_sigmoid_cls else 1
        cls_reg_targets = self.get_targets(
            anchor_list,
            valid_flag_list,
            gt_bboxes,
            img_metas,
            gt_bboxes_ignore_list=gt_bboxes_ignore,
            gt_labels_list=gt_labels,
            label_channels=label_channels,
            unmap_outputs=not self.use_ohem,
            return_sampling_results=True)
        if cls_reg_targets is None:
            return None
        (labels_list, label_weights_list, bbox_targets_list, bbox_weights_list,
         num_total_pos, num_total_neg, sampling_results) = cls_reg_targets

        if self.use_ohem:
            num_images = len(img_metas)
            all_cls_scores = torch.cat([
                s.permute(0, 2, 3, 1).reshape(
                    num_images, -1, self.cls_out_channels) for s in cls_scores
            ], 1)
            all_labels = torch.cat(labels_list, -1).view(num_images, -1)
            all_label_weights = torch.cat(label_weights_list,
                                          -1).view(num_images, -1)
            all_bbox_preds = torch.cat([
                b.permute(0, 2, 3, 1).reshape(num_images, -1, 4)
                for b in bbox_preds
            ], -2)
            all_bbox_targets = torch.cat(bbox_targets_list,
                                         -2).view(num_images, -1, 4)
            all_bbox_weights = torch.cat(bbox_weights_list,
                                         -2).view(num_images, -1, 4)

            # concat all level anchors to a single tensor
            all_anchors = []
            for i in range(num_images):
                all_anchors.append(torch.cat(anchor_list[i]))

            # check NaN and Inf
            assert torch.isfinite(all_cls_scores).all().item(), \
                'classification scores become infinite or NaN!'
            assert torch.isfinite(all_bbox_preds).all().item(), \
                'bbox predications become infinite or NaN!'

            losses_cls, losses_bbox = multi_apply(
                self.loss_single_OHEM,
                all_cls_scores,
                all_bbox_preds,
                all_anchors,
                all_labels,
                all_label_weights,
                all_bbox_targets,
                all_bbox_weights,
                num_total_samples=num_total_pos)
        else:
            num_total_samples = (
                num_total_pos +
                num_total_neg if self.sampling else num_total_pos)

            # anchor number of multi levels
            num_level_anchors = [anchors.size(0) for anchors in anchor_list[0]]
            # concat all level anchors and flags to a single tensor
            concat_anchor_list = []
            for i in range(len(anchor_list)):
                concat_anchor_list.append(torch.cat(anchor_list[i]))
            all_anchor_list = images_to_levels(concat_anchor_list,
                                               num_level_anchors)
            losses_cls, losses_bbox = multi_apply(
                self.loss_single,
                cls_scores,
                bbox_preds,
                all_anchor_list,
                labels_list,
                label_weights_list,
                bbox_targets_list,
                bbox_weights_list,
                num_total_samples=num_total_samples)

        return dict(
            loss_cls=losses_cls, loss_bbox=losses_bbox), sampling_results

    def loss_single_OHEM(self, cls_score, bbox_pred, anchors, labels,
                         label_weights, bbox_targets, bbox_weights,
                         num_total_samples):
        """"See func:``SSDHead.loss``."""
        loss_cls_all = self.loss_cls(cls_score, labels, label_weights)

        # FG cat_id: [0, num_classes -1], BG cat_id: num_classes
        pos_inds = ((labels >= 0) & (labels < self.num_classes)).nonzero(
            as_tuple=False).reshape(-1)
        neg_inds = (labels == self.num_classes).nonzero(
            as_tuple=False).view(-1)

        num_pos_samples = pos_inds.size(0)
        if num_pos_samples == 0:
            num_neg_samples = neg_inds.size(0)
        else:
            num_neg_samples = self.train_cfg.neg_pos_ratio * num_pos_samples
            if num_neg_samples > neg_inds.size(0):
                num_neg_samples = neg_inds.size(0)
        topk_loss_cls_neg, _ = loss_cls_all[neg_inds].topk(num_neg_samples)
        loss_cls_pos = loss_cls_all[pos_inds].sum()
        loss_cls_neg = topk_loss_cls_neg.sum()
        loss_cls = (loss_cls_pos + loss_cls_neg) / num_total_samples
        if self.reg_decoded_bbox:
            # When the regression loss (e.g. `IouLoss`, `GIouLoss`)
            # is applied directly on the decoded bounding boxes, it
            # decodes the already encoded coordinates to absolute format.
            bbox_pred = self.bbox_coder.decode(anchors, bbox_pred)
        loss_bbox = self.loss_bbox(
            bbox_pred,
            bbox_targets,
            bbox_weights,
            avg_factor=num_total_samples)
        return loss_cls[None], loss_bbox

    @force_fp32(apply_to=('cls_scores', 'bbox_preds', 'coeff_preds'))
    def get_bboxes(self,
                   cls_scores,
                   bbox_preds,
                   coeff_preds,
                   img_metas,
                   cfg=None,
                   rescale=False):
        """"Similiar to func:``AnchorHead.get_bboxes``, but additionally
        processes coeff_preds.

        Args:
            cls_scores (list[Tensor]): Box scores for each scale level
                with shape (N, num_anchors * num_classes, H, W)
            bbox_preds (list[Tensor]): Box energies / deltas for each scale
                level with shape (N, num_anchors * 4, H, W)
            coeff_preds (list[Tensor]): Mask coefficients for each scale
                level with shape (N, num_anchors * num_protos, H, W)
            img_metas (list[dict]): Meta information of each image, e.g.,
                image size, scaling factor, etc.
            cfg (mmcv.Config | None): Test / postprocessing configuration,
                if None, test_cfg would be used
            rescale (bool): If True, return boxes in original image space.
                Default: False.

        Returns:
            list[tuple[Tensor, Tensor, Tensor]]: Each item in result_list is
                a 3-tuple. The first item is an (n, 5) tensor, where the
                first 4 columns are bounding box positions
                (tl_x, tl_y, br_x, br_y) and the 5-th column is a score
                between 0 and 1. The second item is an (n,) tensor where each
                item is the predicted class label of the corresponding box.
                The third item is an (n, num_protos) tensor where each item
                is the predicted mask coefficients of instance inside the
                corresponding box.
        """
        assert len(cls_scores) == len(bbox_preds)
        num_levels = len(cls_scores)

        device = cls_scores[0].device
        featmap_sizes = [cls_scores[i].shape[-2:] for i in range(num_levels)]
        mlvl_anchors = self.anchor_generator.grid_anchors(
            featmap_sizes, device=device)

        det_bboxes = []
        det_labels = []
        det_coeffs = []
        for img_id in range(len(img_metas)):
            cls_score_list = [
                cls_scores[i][img_id].detach() for i in range(num_levels)
            ]
            bbox_pred_list = [
                bbox_preds[i][img_id].detach() for i in range(num_levels)
            ]
            coeff_pred_list = [
                coeff_preds[i][img_id].detach() for i in range(num_levels)
            ]
            img_shape = img_metas[img_id]['img_shape']
            scale_factor = img_metas[img_id]['scale_factor']
            bbox_res = self._get_bboxes_single(cls_score_list, bbox_pred_list,
                                               coeff_pred_list, mlvl_anchors,
                                               img_shape, scale_factor, cfg,
                                               rescale)
            det_bboxes.append(bbox_res[0])
            det_labels.append(bbox_res[1])
            det_coeffs.append(bbox_res[2])
        return det_bboxes, det_labels, det_coeffs

    def _get_bboxes_single(self,
                           cls_score_list,
                           bbox_pred_list,
                           coeff_preds_list,
                           mlvl_anchors,
                           img_shape,
                           scale_factor,
                           cfg,
                           rescale=False):
        """"Similiar to func:``AnchorHead._get_bboxes_single``, but
        additionally processes coeff_preds_list and uses fast NMS instead of
        traditional NMS.

        Args:
            cls_score_list (list[Tensor]): Box scores for a single scale level
                Has shape (num_anchors * num_classes, H, W).
            bbox_pred_list (list[Tensor]): Box energies / deltas for a single
                scale level with shape (num_anchors * 4, H, W).
            coeff_preds_list (list[Tensor]): Mask coefficients for a single
                scale level with shape (num_anchors * num_protos, H, W).
            mlvl_anchors (list[Tensor]): Box reference for a single scale level
                with shape (num_total_anchors, 4).
            img_shape (tuple[int]): Shape of the input image,
                (height, width, 3).
            scale_factor (ndarray): Scale factor of the image arange as
                (w_scale, h_scale, w_scale, h_scale).
            cfg (mmcv.Config): Test / postprocessing configuration,
                if None, test_cfg would be used.
            rescale (bool): If True, return boxes in original image space.

        Returns:
            tuple[Tensor, Tensor, Tensor]: The first item is an (n, 5) tensor,
                where the first 4 columns are bounding box positions
                (tl_x, tl_y, br_x, br_y) and the 5-th column is a score between
                0 and 1. The second item is an (n,) tensor where each item is
                the predicted class label of the corresponding box. The third
                item is an (n, num_protos) tensor where each item is the
                predicted mask coefficients of instance inside the
                corresponding box.
        """
        cfg = self.test_cfg if cfg is None else cfg
        assert len(cls_score_list) == len(bbox_pred_list) == len(mlvl_anchors)
        mlvl_bboxes = []
        mlvl_scores = []
        mlvl_coeffs = []
        for cls_score, bbox_pred, coeff_pred, anchors in \
                zip(cls_score_list, bbox_pred_list,
                    coeff_preds_list, mlvl_anchors):
            assert cls_score.size()[-2:] == bbox_pred.size()[-2:]
            cls_score = cls_score.permute(1, 2,
                                          0).reshape(-1, self.cls_out_channels)
            if self.use_sigmoid_cls:
                scores = cls_score.sigmoid()
            else:
                scores = cls_score.softmax(-1)
            bbox_pred = bbox_pred.permute(1, 2, 0).reshape(-1, 4)
            coeff_pred = coeff_pred.permute(1, 2,
                                            0).reshape(-1, self.num_protos)
            nms_pre = cfg.get('nms_pre', -1)
            if nms_pre > 0 and scores.shape[0] > nms_pre:
                # Get maximum scores for foreground classes.
                if self.use_sigmoid_cls:
                    max_scores, _ = scores.max(dim=1)
                else:
                    # remind that we set FG labels to [0, num_class-1]
                    # since mmdet v2.0
                    # BG cat_id: num_class
                    max_scores, _ = scores[:, :-1].max(dim=1)
                _, topk_inds = max_scores.topk(nms_pre)
                anchors = anchors[topk_inds, :]
                bbox_pred = bbox_pred[topk_inds, :]
                scores = scores[topk_inds, :]
                coeff_pred = coeff_pred[topk_inds, :]
            bboxes = self.bbox_coder.decode(
                anchors, bbox_pred, max_shape=img_shape)
            mlvl_bboxes.append(bboxes)
            mlvl_scores.append(scores)
            mlvl_coeffs.append(coeff_pred)
        mlvl_bboxes = torch.cat(mlvl_bboxes)
        if rescale:
            mlvl_bboxes /= mlvl_bboxes.new_tensor(scale_factor)
        mlvl_scores = torch.cat(mlvl_scores)
        mlvl_coeffs = torch.cat(mlvl_coeffs)
        if self.use_sigmoid_cls:
            # Add a dummy background class to the backend when using sigmoid
            # remind that we set FG labels to [0, num_class-1] since mmdet v2.0
            # BG cat_id: num_class
            padding = mlvl_scores.new_zeros(mlvl_scores.shape[0], 1)
            mlvl_scores = torch.cat([mlvl_scores, padding], dim=1)
        det_bboxes, det_labels, det_coeffs = fast_nms(mlvl_bboxes, mlvl_scores,
                                                      mlvl_coeffs,
                                                      cfg.score_thr,
                                                      cfg.iou_thr, cfg.top_k,
                                                      cfg.max_per_img)
        return det_bboxes, det_labels, det_coeffs


@HEADS.register_module()
class YOLACTSegmHead(nn.Module):
    """YOLACT segmentation head used in https://arxiv.org/abs/1904.02689.

    Apply a semantic segmentation loss on feature space using layers that are
    only evaluated during training to increase performance with no speed
    penalty.

    Args:
        in_channels (int): Number of channels in the input feature map.
        num_classes (int): Number of categories excluding the background
            category.
        loss_segm (dict): Config of semantic segmentation loss.
    """

    def __init__(self,
                 num_classes,
                 in_channels=256,
                 loss_segm=dict(
                     type='CrossEntropyLoss',
                     use_sigmoid=True,
                     loss_weight=1.0)):
        super(YOLACTSegmHead, self).__init__()
        self.in_channels = in_channels
        self.num_classes = num_classes
        self.loss_segm = build_loss(loss_segm)
        self._init_layers()
        self.fp16_enabled = False

    def _init_layers(self):
        """Initialize layers of the head."""
        self.segm_conv = nn.Conv2d(
            self.in_channels, self.num_classes, kernel_size=1)

    def init_weights(self):
        """Initialize weights of the head."""
        xavier_init(self.segm_conv, distribution='uniform')

    def forward(self, x):
        """Forward feature from the upstream network.

        Args:
            x (Tensor): Feature from the upstream network, which is
                a 4D-tensor.

        Returns:
            Tensor: Predicted semantic segmentation map with shape
                (N, num_classes, H, W).
        """
        return self.segm_conv(x)

    @force_fp32(apply_to=('segm_pred', ))
    def loss(self, segm_pred, gt_masks, gt_labels):
        """Compute loss of the head.

        Args:
            segm_pred (list[Tensor]): Predicted semantic segmentation map
                with shape (N, num_classes, H, W).
            gt_masks (list[Tensor]): Ground truth masks for each image with
                the same shape of the input image.
            gt_labels (list[Tensor]): Class indices corresponding to each box.

        Returns:
            dict[str, Tensor]: A dictionary of loss components.
        """
        loss_segm = []
        num_imgs, num_classes, mask_h, mask_w = segm_pred.size()
        for idx in range(num_imgs):
            cur_segm_pred = segm_pred[idx]
            cur_gt_masks = gt_masks[idx].float()
            cur_gt_labels = gt_labels[idx]
            segm_targets = self.get_targets(cur_segm_pred, cur_gt_masks,
                                            cur_gt_labels)
            if segm_targets is None:
                loss = self.loss_segm(cur_segm_pred,
                                      torch.zeros_like(cur_segm_pred),
                                      torch.zeros_like(cur_segm_pred))
            else:
                loss = self.loss_segm(
                    cur_segm_pred,
                    segm_targets,
                    avg_factor=num_imgs * mask_h * mask_w)
            loss_segm.append(loss)
        return dict(loss_segm=loss_segm)

    def get_targets(self, segm_pred, gt_masks, gt_labels):
        """Compute semantic segmentation targets for each image.

        Args:
            segm_pred (Tensor): Predicted semantic segmentation map
                with shape (num_classes, H, W).
            gt_masks (Tensor): Ground truth masks for each image with
                the same shape of the input image.
            gt_labels (Tensor): Class indices corresponding to each box.

        Returns:
            Tensor: Semantic segmentation targets with shape
                (num_classes, H, W).
        """
        if gt_masks.size(0) == 0:
            return None
        num_classes, mask_h, mask_w = segm_pred.size()
        with torch.no_grad():
            downsampled_masks = F.interpolate(
                gt_masks.unsqueeze(0), (mask_h, mask_w),
                mode='bilinear',
                align_corners=False).squeeze(0)
            downsampled_masks = downsampled_masks.gt(0.5).float()
            segm_targets = torch.zeros_like(segm_pred, requires_grad=False)
            for obj_idx in range(downsampled_masks.size(0)):
                segm_targets[gt_labels[obj_idx] - 1] = torch.max(
                    segm_targets[gt_labels[obj_idx] - 1],
                    downsampled_masks[obj_idx])
            return segm_targets


@HEADS.register_module()
class YOLACTProtonet(nn.Module):
    """YOLACT mask head used in https://arxiv.org/abs/1904.02689.

    This head outputs the mask prototypes for YOLACT.

    Args:
        in_channels (int): Number of channels in the input feature map.
        proto_channels (tuple[int]): Output channels of protonet convs.
        proto_kernel_sizes (tuple[int]): Kernel sizes of protonet convs.
        include_last_relu (Bool): If keep the last relu of protonet.
        num_protos (int): Number of prototypes.
        num_classes (int): Number of categories excluding the background
            category.
        loss_mask_weight (float): Reweight the mask loss by this factor.
        max_masks_to_train (int): Maximum number of masks to train for
            each image.
    """

    def __init__(self,
                 num_classes,
                 in_channels=256,
                 proto_channels=(256, 256, 256, None, 256, 32),
                 proto_kernel_sizes=(3, 3, 3, -2, 3, 1),
                 include_last_relu=True,
                 num_protos=32,
                 loss_mask_weight=1.0,
                 max_masks_to_train=100):
        super(YOLACTProtonet, self).__init__()
        self.in_channels = in_channels
        self.proto_channels = proto_channels
        self.proto_kernel_sizes = proto_kernel_sizes
        self.include_last_relu = include_last_relu
        self.protonet = self._init_layers()

        self.loss_mask_weight = loss_mask_weight
        self.num_protos = num_protos
        self.num_classes = num_classes
        self.max_masks_to_train = max_masks_to_train
        self.fp16_enabled = False

    def _init_layers(self):
        """A helper function to take a config setting and turn it into a
        network."""
        # Possible patterns:
        # ( 256, 3) -> conv
        # ( 256,-2) -> deconv
        # (None,-2) -> bilinear interpolate
        in_channels = self.in_channels
        protonets = nn.ModuleList()
        for num_channels, kernel_size in zip(self.proto_channels,
                                             self.proto_kernel_sizes):
            if kernel_size > 0:
                layer = nn.Conv2d(
                    in_channels,
                    num_channels,
                    kernel_size,
                    padding=kernel_size // 2)
            else:
                if num_channels is None:
                    layer = InterpolateModule(
                        scale_factor=-kernel_size,
                        mode='bilinear',
                        align_corners=False)
                else:
                    layer = nn.ConvTranspose2d(
                        in_channels,
                        num_channels,
                        -kernel_size,
                        padding=kernel_size // 2)
            protonets.append(layer)
            protonets.append(nn.ReLU(inplace=True))
            in_channels = num_channels if num_channels is not None \
                else in_channels
        if not self.include_last_relu:
            protonets = protonets[:-1]
        return nn.Sequential(*protonets)

    def init_weights(self):
        """Initialize weights of the head."""
        for m in self.protonet:
            if isinstance(m, nn.Conv2d):
                xavier_init(m, distribution='uniform')

    def forward(self, x, coeff_pred, bboxes, img_meta, sampling_results=None):
        """Forward feature from the upstream network to get prototypes and
        linearly combine the prototypes, using masks coefficients, into
        instance masks. Finally, crop the instance masks with given bboxes.

        Args:
            x (Tensor): Feature from the upstream network, which is
                a 4D-tensor.
            coeff_pred (list[Tensor]): Mask coefficients for each scale
                level with shape (N, num_anchors * num_protos, H, W).
            bboxes (list[Tensor]): Box used for cropping with shape
                (N, num_anchors * 4, H, W). During training, they are
                ground truth boxes. During testing, they are predicted
                boxes.
            img_meta (list[dict]): Meta information of each image, e.g.,
                image size, scaling factor, etc.
            sampling_results (List[:obj:``SamplingResult``]): Sampler results
                for each image.

        Returns:
            list[Tensor]: Predicted instance segmentation masks.
        """
        prototypes = self.protonet(x)
        prototypes = prototypes.permute(0, 2, 3, 1).contiguous()

        num_imgs = x.size(0)
        # Training state
        if self.training:
            coeff_pred_list = []
            for coeff_pred_per_level in coeff_pred:
                coeff_pred_per_level = \
                    coeff_pred_per_level.permute(0, 2, 3, 1)\
                    .reshape(num_imgs, -1, self.num_protos)
                coeff_pred_list.append(coeff_pred_per_level)
            coeff_pred = torch.cat(coeff_pred_list, dim=1)

        mask_pred_list = []
        for idx in range(num_imgs):
            cur_prototypes = prototypes[idx]
            cur_coeff_pred = coeff_pred[idx]
            cur_bboxes = bboxes[idx]
            cur_img_meta = img_meta[idx]

            # Testing state
            if not self.training:
                bboxes_for_cropping = cur_bboxes
            else:
                cur_sampling_results = sampling_results[idx]
                pos_assigned_gt_inds = \
                    cur_sampling_results.pos_assigned_gt_inds
                bboxes_for_cropping = cur_bboxes[pos_assigned_gt_inds].clone()
                pos_inds = cur_sampling_results.pos_inds
                cur_coeff_pred = cur_coeff_pred[pos_inds]

            # Linearly combine the prototypes with the mask coefficients
            mask_pred = cur_prototypes @ cur_coeff_pred.t()
            mask_pred = torch.sigmoid(mask_pred)

            h, w = cur_img_meta['img_shape'][:2]
            bboxes_for_cropping[:, 0] /= w
            bboxes_for_cropping[:, 1] /= h
            bboxes_for_cropping[:, 2] /= w
            bboxes_for_cropping[:, 3] /= h

            mask_pred = self.crop(mask_pred, bboxes_for_cropping)
            mask_pred = mask_pred.permute(2, 0, 1).contiguous()
            mask_pred_list.append(mask_pred)
        return mask_pred_list

    @force_fp32(apply_to=('mask_pred', ))
    def loss(self, mask_pred, gt_masks, gt_bboxes, img_meta, sampling_results):
        """Compute loss of the head.

        Args:
            mask_pred (list[Tensor]): Predicted prototypes with shape
                (num_classes, H, W).
            gt_masks (list[Tensor]): Ground truth masks for each image with
                the same shape of the input image.
            gt_bboxes (list[Tensor]): Ground truth bboxes for each image with
                shape (num_gts, 4) in [tl_x, tl_y, br_x, br_y] format.
            img_meta (list[dict]): Meta information of each image, e.g.,
                image size, scaling factor, etc.
            sampling_results (List[:obj:``SamplingResult``]): Sampler results
                for each image.

        Returns:
            dict[str, Tensor]: A dictionary of loss components.
        """
        loss_mask = []
        num_imgs = len(mask_pred)
        total_pos = 0
        for idx in range(num_imgs):
            cur_mask_pred = mask_pred[idx]
            cur_gt_masks = gt_masks[idx].float()
            cur_gt_bboxes = gt_bboxes[idx]
            cur_img_meta = img_meta[idx]
            cur_sampling_results = sampling_results[idx]

            pos_assigned_gt_inds = cur_sampling_results.pos_assigned_gt_inds
            num_pos = pos_assigned_gt_inds.size(0)
            # Since we're producing (near) full image masks,
            # it'd take too much vram to backprop on every single mask.
            # Thus we select only a subset.
            if num_pos > self.max_masks_to_train:
                perm = torch.randperm(num_pos)
                select = perm[:self.max_masks_to_train]
                cur_mask_pred = cur_mask_pred[select]
                pos_assigned_gt_inds = pos_assigned_gt_inds[select]
                num_pos = self.max_masks_to_train
            total_pos += num_pos

            gt_bboxes_for_reweight = cur_gt_bboxes[pos_assigned_gt_inds]

            mask_targets = self.get_targets(cur_mask_pred, cur_gt_masks,
                                            pos_assigned_gt_inds)
            if num_pos == 0:
                loss = cur_mask_pred.sum() * 0.
            elif mask_targets is None:
                loss = F.binary_cross_entropy(cur_mask_pred,
                                              torch.zeros_like(cur_mask_pred),
                                              torch.zeros_like(cur_mask_pred))
            else:
                cur_mask_pred = torch.clamp(cur_mask_pred, 0, 1)
                loss = F.binary_cross_entropy(
                    cur_mask_pred, mask_targets,
                    reduction='none') * self.loss_mask_weight

                h, w = cur_img_meta['img_shape'][:2]
                gt_bboxes_width = (gt_bboxes_for_reweight[:, 2] -
                                   gt_bboxes_for_reweight[:, 0]) / w
                gt_bboxes_height = (gt_bboxes_for_reweight[:, 3] -
                                    gt_bboxes_for_reweight[:, 1]) / h
                loss = loss.mean(dim=(1,
                                      2)) / gt_bboxes_width / gt_bboxes_height
                loss = torch.sum(loss)
            loss_mask.append(loss)

        if total_pos == 0:
            total_pos += 1  # avoid nan
        loss_mask = [x / total_pos for x in loss_mask]

        return dict(loss_mask=loss_mask)

    def get_targets(self, mask_pred, gt_masks, pos_assigned_gt_inds):
        """Compute instance segmentation targets for each image.

        Args:
            mask_pred (Tensor): Predicted prototypes with shape
                (num_classes, H, W).
            gt_masks (Tensor): Ground truth masks for each image with
                the same shape of the input image.
            pos_assigned_gt_inds (Tensor): GT indices of the corresponding
                positive samples.
        Returns:
            Tensor: Instance segmentation targets with shape
                (num_instances, H, W).
        """
        if gt_masks.size(0) == 0:
            return None
        mask_h, mask_w = mask_pred.shape[-2:]
        gt_masks = F.interpolate(
            gt_masks.unsqueeze(0), (mask_h, mask_w),
            mode='bilinear',
            align_corners=False).squeeze(0)
        gt_masks = gt_masks.gt(0.5).float()
        mask_targets = gt_masks[pos_assigned_gt_inds]
        return mask_targets

    def get_seg_masks(self, mask_pred, label_pred, img_meta, rescale):
        """Resize, binarize, and format the instance mask predictions.

        Args:
            mask_pred (Tensor): shape (N, H, W).
            label_pred (Tensor): shape (N, ).
            img_meta (dict): Meta information of each image, e.g.,
                image size, scaling factor, etc.
            rescale (bool): If rescale is False, then returned masks will
                fit the scale of imgs[0].
        Returns:
            list[ndarray]: Mask predictions grouped by their predicted classes.
        """
        ori_shape = img_meta['ori_shape']
        scale_factor = img_meta['scale_factor']
        if rescale:
            img_h, img_w = ori_shape[:2]
        else:
            img_h = np.round(ori_shape[0] * scale_factor[1]).astype(np.int32)
            img_w = np.round(ori_shape[1] * scale_factor[0]).astype(np.int32)

        cls_segms = [[] for _ in range(self.num_classes)]
        if mask_pred.size(0) == 0:
            return cls_segms

        mask_pred = F.interpolate(
            mask_pred.unsqueeze(0), (img_h, img_w),
            mode='bilinear',
            align_corners=False).squeeze(0) > 0.5
        mask_pred = mask_pred.cpu().numpy().astype(np.uint8)

        for m, l in zip(mask_pred, label_pred):
            cls_segms[l].append(m)
        return cls_segms

    def crop(self, masks, boxes, padding=1):
        """Crop predicted masks by zeroing out everything not in the predicted
        bbox.

        Args:
            masks (Tensor): shape [H, W, N].
            boxes (Tensor): bbox coords in relative point form with
                shape [N, 4].

        Return:
            Tensor: The cropped masks.
        """
        h, w, n = masks.size()
        x1, x2 = self.sanitize_coordinates(
            boxes[:, 0], boxes[:, 2], w, padding, cast=False)
        y1, y2 = self.sanitize_coordinates(
            boxes[:, 1], boxes[:, 3], h, padding, cast=False)

        rows = torch.arange(
            w, device=masks.device, dtype=x1.dtype).view(1, -1,
                                                         1).expand(h, w, n)
        cols = torch.arange(
            h, device=masks.device, dtype=x1.dtype).view(-1, 1,
                                                         1).expand(h, w, n)

        masks_left = rows >= x1.view(1, 1, -1)
        masks_right = rows < x2.view(1, 1, -1)
        masks_up = cols >= y1.view(1, 1, -1)
        masks_down = cols < y2.view(1, 1, -1)

        crop_mask = masks_left * masks_right * masks_up * masks_down

        return masks * crop_mask.float()

    def sanitize_coordinates(self, x1, x2, img_size, padding=0, cast=True):
        """Sanitizes the input coordinates so that x1 < x2, x1 != x2, x1 >= 0,
        and x2 <= image_size. Also converts from relative to absolute
        coordinates and casts the results to long tensors.

        Warning: this does things in-place behind the scenes so
        copy if necessary.

        Args:
            _x1 (Tensor): shape (N, ).
            _x2 (Tensor): shape (N, ).
            img_size (int): Size of the input image.
            padding (int): x1 >= padding, x2 <= image_size-padding.
            cast (bool): If cast is false, the result won't be cast to longs.

        Returns:
            tuple:
                x1 (Tensor): Sanitized _x1.
                x2 (Tensor): Sanitized _x2.
        """
        x1 = x1 * img_size
        x2 = x2 * img_size
        if cast:
            x1 = x1.long()
            x2 = x2.long()
        x1 = torch.min(x1, x2)
        x2 = torch.max(x1, x2)
        x1 = torch.clamp(x1 - padding, min=0)
        x2 = torch.clamp(x2 + padding, max=img_size)
        return x1, x2


class InterpolateModule(nn.Module):
    """This is a module version of F.interpolate.

    Any arguments you give it just get passed along for the ride.
    """

    def __init__(self, *args, **kwargs):
        super().__init__()

        self.args = args
        self.kwargs = kwargs

    def forward(self, x):
        """Forward features from the upstream network."""
        return F.interpolate(x, *self.args, **self.kwargs)