forked from hachreak/unbiased-teacher
-
Notifications
You must be signed in to change notification settings - Fork 0
/
contributes.diff
1153 lines (1114 loc) · 48.7 KB
/
contributes.diff
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
diff --git a/README.md b/README.md
index 08239f1..f413719 100644
--- a/README.md
+++ b/README.md
@@ -1,173 +1,69 @@
-# Unbiased Teacher for Semi-Supervised Object Detection
+## Improving Localization for Semi-Supervised Object Detection
-<img src="teaser/pytorch-logo-dark.png" width="10%"> [](https://opensource.org/licenses/MIT)
+### Installation
-This is the PyTorch implementation of our paper: <br>
-**Unbiased Teacher for Semi-Supervised Object Detection**<br>
-[Yen-Cheng Liu](https://ycliu93.github.io/), [Chih-Yao Ma](https://chihyaoma.github.io/), [Zijian He](https://research.fb.com/people/he-zijian/), [Chia-Wen Kuo](https://sites.google.com/view/chiawen-kuo/home), [Kan Chen](https://kanchen.info/), [Peizhao Zhang](https://scholar.google.com/citations?user=eqQQkM4AAAAJ&hl=en), [Bichen Wu](https://scholar.google.com/citations?user=K3QJPdMAAAAJ&hl=en), [Zsolt Kira](https://www.cc.gatech.edu/~zk15/), [Peter Vajda](https://sites.google.com/site/vajdap)<br>
-International Conference on Learning Representations (ICLR), 2021 <br>
+See [unbiased-teacher](https://github.com/facebookresearch/unbiased-teacher/tree/ba543ed)
+official pages to know the installation procedure.
-[[arXiv](https://arxiv.org/abs/2102.09480)] [[OpenReview](https://openreview.net/forum?id=MJIve1zgR_)] [[Project](https://ycliu93.github.io/projects/unbiasedteacher.html)]
+### Models
-<p align="center">
-<img src="teaser/figure_teaser.gif" width="85%">
-</p>
+Most important files.
-# Installation
+- [trainer.py](ubteacher/engine/trainer.py)
+- [fast_rcnn](ubteacher/modeling/roi_heads/fast_rcnn.py) which contains the
+ BBox IoU classification branch.
-## Prerequisites
+## Experiments
-- Linux or macOS with Python ≥ 3.6
-- PyTorch ≥ 1.5 and torchvision that matches the PyTorch installation.
+The following configuration files have been used to run each experiment:
-## Install PyTorch in Conda env
+Table 1 / Figure 3a:
-```shell
-# create conda env
-conda create -n detectron2 python=3.6
-# activate the enviorment
-conda activate detectron2
-# install PyTorch >=1.5 with GPU
-conda install pytorch torchvision -c pytorch
-```
+| Row | Beta | AP |
+| :--: | :--: | :--: |
+| | [base](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-base.yaml) | 31.027 |
+| 0 | [0.5](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7_4.yaml) | 31.775 |
+| 1 | [1](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7.yaml) | 31.947 |
+| 2 | [2](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7_2.yaml) | 31.754 |
+| 3 | [4](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7_3.yaml) | 30.445 |
-## Build Detectron2 from Source
-Follow the [INSTALL.md](https://github.com/facebookresearch/detectron2/blob/master/INSTALL.md) to install Detectron2.
+Table 2 / Figure 3c:
-## Dataset download
+| Row | Model | AP |
+| :--: | :--: | :--: |
+| 1 | [UT](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-base.yaml) | 31.027 |
+| 2 | [Ours (with filter)](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7_4.yaml) | 31.605 |
+| 3 | [Ours (w/out filter)](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7.yaml) | 31.509 |
-1. Download COCO dataset
+Table 3 / Figure 3d:
-```shell
-# download images
-wget http://images.cocodataset.org/zips/train2017.zip
-wget http://images.cocodataset.org/zips/val2017.zip
+| Row | Mu | AP |
+| :--: | :--: | :--: |
+| 1 | [0.5](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_2.yaml) | 31.199 |
+| 2 | [0.6](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28.yaml) | 31.128 |
+| 3 | [0.7](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_3.yaml) | 31.461 |
+| 4 | [0.75](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v13_5.yaml) | 31.604 |
+| 5 | [0.8](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_4.yaml) | 31.336 |
+| 6 | [0.9](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_5.yaml) | 27.125 |
-# download annotations
-wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip
-```
+Table 4:
-2. Organize the dataset as following:
+| Row | Theta | AP |
+| :--: | :--: | :--: |
+| 0 | [0.3](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31_4.yaml) | 31.404 |
+| 1 | [0.4](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31_3.yaml) | 31.630 |
+| 2 | [0.5](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v13_5.yaml) | 31.604 |
+| 3 | [0.6](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31.yaml) | 31.158 |
+| 4 | [0.7](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31_2.yaml) | 30.649 |
-```shell
-unbiased_teacher/
-└── datasets/
- └── coco/
- ├── train2017/
- ├── val2017/
- └── annotations/
- ├── instances_train2017.json
- └── instances_val2017.json
-```
+Table 5:
-## Training
-
-- Train the Unbiased Teacher under 1% COCO-supervision
-
-```shell
-python train_net.py \
- --num-gpus 8 \
- --config configs/coco_supervision/faster_rcnn_R_50_FPN_sup1_run1.yaml \
- SOLVER.IMG_PER_BATCH_LABEL 16 SOLVER.IMG_PER_BATCH_UNLABEL 16
-```
-
-- Train the Unbiased Teacher under 2% COCO-supervision
-
-```shell
-python train_net.py \
- --num-gpus 8 \
- --config configs/coco_supervision/faster_rcnn_R_50_FPN_sup2_run1.yaml \
- SOLVER.IMG_PER_BATCH_LABEL 16 SOLVER.IMG_PER_BATCH_UNLABEL 16
-```
-
-- Train the Unbiased Teacher under 5% COCO-supervision
-
-```shell
-python train_net.py \
- --num-gpus 8 \
- --config configs/coco_supervision/faster_rcnn_R_50_FPN_sup5_run1.yaml \
- SOLVER.IMG_PER_BATCH_LABEL 16 SOLVER.IMG_PER_BATCH_UNLABEL 16
-```
-
-- Train the Unbiased Teacher under 10% COCO-supervision
-
-```shell
-python train_net.py \
- --num-gpus 8 \
- --config configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1.yaml \
- SOLVER.IMG_PER_BATCH_LABEL 16 SOLVER.IMG_PER_BATCH_UNLABEL 16
-```
-
-## Resume the training
-
-```shell
-python train_net.py \
- --resume \
- --num-gpus 8 \
- --config configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1.yaml \
- SOLVER.IMG_PER_BATCH_LABEL 16 SOLVER.IMG_PER_BATCH_UNLABEL 16 MODEL.WEIGHTS <your weight>.pth
-```
-
-## Evaluation
-
-```shell
-python train_net.py \
- --eval-only \
- --num-gpus 8 \
- --config configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1.yaml \
- SOLVER.IMG_PER_BATCH_LABEL 16 SOLVER.IMG_PER_BATCH_UNLABEL 16 MODEL.WEIGHTS <your weight>.pth
-```
-
-## Model Weights
-
-For the following results, we use 16 labeled images + 16 unlabeled images on 8 GPUs (single node).
-
-Faster-RCNN:
-
-| Model | Supervision | Batch size | AP | Model Weights |
-| :----------: | :------------: | :---------------------------------: | :---: |:-----: |
-| R50-FPN | 1% | 16 labeled img + 16 unlabeled imgs | 20.16 | [link](https://drive.google.com/file/d/1NQs5SrQ2-ODEVn_ZdPU_2xv9mxdY6MPq/view?usp=sharing) |
-| R50-FPN | 2% | 16 labeled img + 16 unlabeled imgs | 24.16 | [link](https://drive.google.com/file/d/12q-LB4iDvgXGW50Q-bYOahpalUvO3SIa/view?usp=sharing) |
-| R50-FPN | 5% | 16 labeled img + 16 unlabeled imgs | 27.84 | [link](https://drive.google.com/file/d/1IJQeRP9wHPU0J27YTea-y3lIW96bMAUu/view?usp=sharing) |
-| R50-FPN | 10% | 16 labeled img + 16 unlabeled imgs | 31.39 | [link](https://drive.google.com/file/d/1U9tnJGvzRFSOnOfIHOnelFmlvEfyayha/view?usp=sharing) |
-
-## FAQ
-
-1. Q: Using the lower batch size and fewer GPUs cannot achieve the results presented in the paper?
-
-- A: We train the model with 32 labeled images + 32 unlabeled images per batch for the results presented in the paper, and using the lower batch size leads to lower accuracy. For example, in the 1% COCO-supervision setting, the model trained with 16 labeled images + 16 unlabeled images achieves 19.9 AP as shown in the following table.
-
-| Experiment GPUs | Batch size per node | Batch size | AP |
-| :------------------: | :---------------------------------: | :---------------------------------: | :---: |
-| 8 GPUs/node; 4 nodes | 8 labeled imgs + 8 unlabeled imgs | 32 labeled img + 32 unlabeled imgs | 20.75 |
-| 8 GPUs/node; 1 node | 16 labeled imgs + 16 unlabeled imgs | 16 labeled imgs + 16 unlabeled imgs | 20.16 |
-
-## Citing Unbiased Teacher
-
-If you use Unbiased Teacher in your research or wish to refer to the results published in the paper, please use the following BibTeX entry.
-
-```BibTeX
-@inproceedings{liu2021unbiased,
- title={Unbiased Teacher for Semi-Supervised Object Detection},
- author={Liu, Yen-Cheng and Ma, Chih-Yao and He, Zijian and Kuo, Chia-Wen and Chen, Kan and Zhang, Peizhao and Wu, Bichen and Kira, Zsolt and Vajda, Peter},
- booktitle={Proceedings of the International Conference on Learning Representations (ICLR)},
- year={2021},
-}
-```
-
-Also, if you use Detectron2 in your research, please use the following BibTeX entry.
-
-```BibTeX
-@misc{wu2019detectron2,
- author = {Yuxin Wu and Alexander Kirillov and Francisco Massa and
- Wan-Yen Lo and Ross Girshick},
- title = {Detectron2},
- howpublished = {\url{https://github.com/facebookresearch/detectron2}},
- year = {2019}
-}
-```
-
-## License
-
-This project is licensed under [MIT License](LICENSE), as found in the LICENSE file.
+| Row | L<sub>reg</sub><sup>unsup</sup> | x<sup>sh</sup> | scores | deltas| Model | AP |
+| :--: | :--: | :--: | :--: | :--: | :--: | :--: |
+| 1 | | | | | [model](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-base.yaml) | 31.027 |
+| 2 | x | | | | [model](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7.yaml) | 31.947 |
+| 3 | x | x | | | [model](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v36_3.yaml) | 31.754 |
+| 4 | x | x | x | | [model](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v36_2.yaml) | 32.166 |
+| 5 | x | x | x | x | [model](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v36.yaml) | 31.923 |
+| 6 | | x | x | x | [model](configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31_3.yaml) | 31.630 |
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-base.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-base.yaml
new file mode 100644
index 0000000..836742f
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-base.yaml
@@ -0,0 +1,7 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1.yaml"
+SOLVER:
+ IMG_PER_BATCH_LABEL: 12
+ IMG_PER_BATCH_UNLABEL: 12
+ BASE_LR: 0.0075
+TEST:
+ EVAL_PERIOD: 10000
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v13_5.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v13_5.yaml
new file mode 100644
index 0000000..03e4c5b
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v13_5.yaml
@@ -0,0 +1,5 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-base.yaml"
+SEMISUPNET:
+ IOU_PRED_THRESHOLD: 0.75
+ IOU_PRED_INFERENCE_THRESHOLD: 0.5
+ WITH_IOU_PRED: True
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v13_8.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v13_8.yaml
new file mode 100644
index 0000000..3069857
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v13_8.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-v13_5.yaml"
+SEMISUPNET:
+ FILTER_WITH_BBOX_IOU: False
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v13_9.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v13_9.yaml
new file mode 100644
index 0000000..d59f8be
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v13_9.yaml
@@ -0,0 +1,6 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-base.yaml"
+SEMISUPNET:
+ IOU_PRED_THRESHOLD: 0.75
+ IOU_PRED_INFERENCE_THRESHOLD: 0.5
+ WITH_IOU_PRED: True
+ WITH_IOU_PRED_ELU: True
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v22.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v22.yaml
new file mode 100644
index 0000000..3ff8a72
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v22.yaml
@@ -0,0 +1,6 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-base.yaml"
+SEMISUPNET:
+ IOU_PRED_THRESHOLD: 0.75
+ IOU_PRED_INFERENCE_THRESHOLD: 0.5
+ WITH_IOU_PRED: True
+ WITH_BBOX_IOU_REGRES: True
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v22_2.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v22_2.yaml
new file mode 100644
index 0000000..ac0b0e3
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v22_2.yaml
@@ -0,0 +1,8 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-base.yaml"
+SEMISUPNET:
+ IOU_PRED_THRESHOLD: 0.75
+ IOU_PRED_INFERENCE_THRESHOLD: 0.5
+ #DEBUG_BBOX_LABELS: True
+ WITH_IOU_PRED: True
+ WITH_BBOX_IOU_REGRES: True
+ USE_SMOOTHL1: True
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28.yaml
new file mode 100644
index 0000000..16d0c70
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-v13_9.yaml"
+SEMISUPNET:
+ IOU_PRED_THRESHOLD: 0.6
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_2.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_2.yaml
new file mode 100644
index 0000000..71dc5b8
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_2.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-v13_9.yaml"
+SEMISUPNET:
+ IOU_PRED_THRESHOLD: 0.5
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_3.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_3.yaml
new file mode 100644
index 0000000..fe95af7
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_3.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-v13_9.yaml"
+SEMISUPNET:
+ IOU_PRED_THRESHOLD: 0.7
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_4.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_4.yaml
new file mode 100644
index 0000000..beea2d7
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_4.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-v13_9.yaml"
+SEMISUPNET:
+ IOU_PRED_THRESHOLD: 0.8
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_5.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_5.yaml
new file mode 100644
index 0000000..45b0eb9
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v28_5.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-v13_9.yaml"
+SEMISUPNET:
+ IOU_PRED_THRESHOLD: 0.9
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31.yaml
new file mode 100644
index 0000000..c31b23b
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-v13_9.yaml"
+SEMISUPNET:
+ IOU_PRED_INFERENCE_THRESHOLD: 0.6
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31_2.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31_2.yaml
new file mode 100644
index 0000000..96779ec
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31_2.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-v13_9.yaml"
+SEMISUPNET:
+ IOU_PRED_INFERENCE_THRESHOLD: 0.7
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31_3.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31_3.yaml
new file mode 100644
index 0000000..cede470
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31_3.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-v13_9.yaml"
+SEMISUPNET:
+ IOU_PRED_INFERENCE_THRESHOLD: 0.4
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31_4.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31_4.yaml
new file mode 100644
index 0000000..d80fe64
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v31_4.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-v13_9.yaml"
+SEMISUPNET:
+ IOU_PRED_INFERENCE_THRESHOLD: 0.3
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v36.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v36.yaml
new file mode 100644
index 0000000..24abf1d
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v36.yaml
@@ -0,0 +1,7 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-base.yaml"
+SEMISUPNET:
+ UNSUP_LOSS_WEIGHT_REG: 1.0
+ IOU_PRED_INFERENCE_THRESHOLD: 0.4
+ IOU_PRED_THRESHOLD: 0.75
+ WITH_IOU_PRED: True
+ WITH_IOU_PRED_ELU: True
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v36_2.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v36_2.yaml
new file mode 100644
index 0000000..8f546f4
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v36_2.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-v36.yaml"
+SEMISUPNET:
+ WITH_REGR_VALUES: False
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v36_3.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v36_3.yaml
new file mode 100644
index 0000000..bc20872
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v36_3.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-v36_2.yaml"
+SEMISUPNET:
+ WITH_SCORE_VALUES: False
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7.yaml
new file mode 100644
index 0000000..992c418
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-base.yaml"
+SEMISUPNET:
+ UNSUP_LOSS_WEIGHT_REG: 1.0
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7_2.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7_2.yaml
new file mode 100644
index 0000000..7647063
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7_2.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-base.yaml"
+SEMISUPNET:
+ UNSUP_LOSS_WEIGHT_REG: 2.0
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7_3.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7_3.yaml
new file mode 100644
index 0000000..a7dd564
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7_3.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-base.yaml"
+SEMISUPNET:
+ UNSUP_LOSS_WEIGHT_REG: 4.0
diff --git a/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7_4.yaml b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7_4.yaml
new file mode 100644
index 0000000..3916eb4
--- /dev/null
+++ b/configs/coco_supervision/faster_rcnn_R_50_FPN_sup10_run1-v7_4.yaml
@@ -0,0 +1,3 @@
+_BASE_: "faster_rcnn_R_50_FPN_sup10_run1-base.yaml"
+SEMISUPNET:
+ UNSUP_LOSS_WEIGHT_REG: 0.5
diff --git a/ubteacher/config.py b/ubteacher/config.py
index ef4fe0f..ee91d7d 100644
--- a/ubteacher/config.py
+++ b/ubteacher/config.py
@@ -38,6 +38,21 @@ def add_ubteacher_config(cfg):
_C.SEMISUPNET.SUP_LOSS_WEIGHT = 0.5
_C.SEMISUPNET.LOSS_WEIGHT_TYPE = "standard"
+ # loss regression
+ _C.SEMISUPNET.UNSUP_LOSS_WEIGHT_REG = 0.
+
+ # bbox iou branch
+ _C.SEMISUPNET.IOU_PRED_THRESHOLD = 0.5
+ _C.SEMISUPNET.IOU_PRED_INFERENCE_THRESHOLD = 0.25
+ _C.SEMISUPNET.WITH_IOU_PRED = False
+ _C.SEMISUPNET.WITH_IOU_PRED_ELU = False
+ _C.SEMISUPNET.FILTER_WITH_BBOX_IOU = True
+ _C.SEMISUPNET.WITH_SCORE_VALUES = True
+ _C.SEMISUPNET.WITH_REGR_VALUES = True
+ _C.SEMISUPNET.WITH_BBOX_IOU_REGRES = False
+ _C.SEMISUPNET.USE_SMOOTHL1 = False
+ _C.SEMISUPNET.DEBUG_BBOX_LABELS = False
+
# dataloader
# supervision level
_C.DATALOADER.SUP_PERCENT = 100.0 # 5 = 5% dataset as labeled set
diff --git a/ubteacher/engine/trainer.py b/ubteacher/engine/trainer.py
index b463428..038283c 100644
--- a/ubteacher/engine/trainer.py
+++ b/ubteacher/engine/trainer.py
@@ -16,8 +16,10 @@ from detectron2.utils.events import EventStorage
from detectron2.evaluation import COCOEvaluator, verify_results
from detectron2.data.dataset_mapper import DatasetMapper
from detectron2.engine import hooks
+from detectron2.structures import pairwise_iou
from detectron2.structures.boxes import Boxes
from detectron2.structures.instances import Instances
+from detectron2.utils.env import TORCH_VERSION
from ubteacher.data.build import (
build_detection_semisup_train_loader,
@@ -68,6 +70,32 @@ class BaselineTrainer(DefaultTrainer):
self.register_hooks(self.build_hooks())
+ def resume_or_load(self, resume=True):
+ """
+ If `resume==True` and `cfg.OUTPUT_DIR` contains the last checkpoint (defined by
+ a `last_checkpoint` file), resume from the file. Resuming means loading all
+ available states (eg. optimizer and scheduler) and update iteration counter
+ from the checkpoint. ``cfg.MODEL.WEIGHTS`` will not be used.
+ Otherwise, this is considered as an independent training. The method will load model
+ weights from the file `cfg.MODEL.WEIGHTS` (but will not load other states) and start
+ from iteration 0.
+ Args:
+ resume (bool): whether to do resume or not
+ """
+ checkpoint = self.checkpointer.resume_or_load(
+ self.cfg.MODEL.WEIGHTS, resume=resume
+ )
+ if resume and self.checkpointer.has_checkpoint():
+ self.start_iter = checkpoint.get("iteration", -1) + 1
+ # The checkpoint stores the training iteration that just finished, thus we start
+ # at the next iteration (or iter zero if there's no checkpoint).
+ if isinstance(self.model, DistributedDataParallel):
+ # broadcast loaded data/model from the first rank, because other
+ # machines may not have access to the checkpoint file
+ if TORCH_VERSION >= (1, 7):
+ self.model._sync_params_and_buffers()
+ self.start_iter = comm.all_gather(self.start_iter)[0]
+
def train_loop(self, start_iter: int, max_iter: int):
"""
Args:
@@ -267,6 +295,32 @@ class UBTeacherTrainer(DefaultTrainer):
self.register_hooks(self.build_hooks())
+ def resume_or_load(self, resume=True):
+ """
+ If `resume==True` and `cfg.OUTPUT_DIR` contains the last checkpoint (defined by
+ a `last_checkpoint` file), resume from the file. Resuming means loading all
+ available states (eg. optimizer and scheduler) and update iteration counter
+ from the checkpoint. ``cfg.MODEL.WEIGHTS`` will not be used.
+ Otherwise, this is considered as an independent training. The method will load model
+ weights from the file `cfg.MODEL.WEIGHTS` (but will not load other states) and start
+ from iteration 0.
+ Args:
+ resume (bool): whether to do resume or not
+ """
+ checkpoint = self.checkpointer.resume_or_load(
+ self.cfg.MODEL.WEIGHTS, resume=resume
+ )
+ if resume and self.checkpointer.has_checkpoint():
+ self.start_iter = checkpoint.get("iteration", -1) + 1
+ # The checkpoint stores the training iteration that just finished, thus we start
+ # at the next iteration (or iter zero if there's no checkpoint).
+ if isinstance(self.model, DistributedDataParallel):
+ # broadcast loaded data/model from the first rank, because other
+ # machines may not have access to the checkpoint file
+ if TORCH_VERSION >= (1, 7):
+ self.model._sync_params_and_buffers()
+ self.start_iter = comm.all_gather(self.start_iter)[0]
+
@classmethod
def build_evaluator(cls, cfg, dataset_name, output_folder=None):
if output_folder is None:
@@ -380,6 +434,47 @@ class UBTeacherTrainer(DefaultTrainer):
unlabel_datum["instances"] = lab_inst
return unlabled_data
+ def eval_label(self, gtlab, newlab):
+
+ def get_err(iou, gt_cls, new_cls, threshold):
+ filt = (iou>threshold) & (iou<threshold+0.1)
+ b_filt = len(iou[filt])
+ c_wrong = len(gt_cls[filt & (gt_cls != new_cls)])
+ return b_filt, c_wrong
+
+ logger = logging.getLogger('fvcore.common.checkpoint')
+ for gt, new in zip(gtlab, newlab):
+ if len(new) > 0:
+ b1 = gt["instances"].gt_boxes
+ b1.tensor = b1.tensor.cuda()
+ c1 = gt["instances"].gt_classes.cuda()
+ iou, idx = pairwise_iou(b1, new.gt_boxes).max(1)
+ b_wrong = len(iou[iou<0.5])
+ b_wronger = len(iou[iou<0.3])
+ b_wrongest = len(iou[iou<0.1])
+ c_wrong = b_wrong + len(c1[iou>=0.5][c1[iou>=0.5] != new.gt_classes[idx][iou>=0.5]])
+ count = float(len(iou))
+ # get error per iou range
+ err = sum([get_err(iou, c1, new.gt_classes[idx], th/10.) for th in range(0, 10)], ())
+ debug = ' '.join([str(s) for s in ["eval_labels", str(self.iter), str(len(iou)),
+ b_wrongest, '{:.2f}'.format(b_wrongest * 100 / count),
+ b_wronger, '{:.2f}'.format(b_wronger * 100 / count),
+ b_wrong, '{:.2f}'.format(b_wrong * 100 / count),
+ c_wrong, '{:.2f}'.format(c_wrong * 100 / count),
+ ' - ', ' '.join([str(i) for i in err]), 'more']])
+ logger.info(debug)
+ b1.tensor = b1.tensor.cpu()
+ c1 = c1.cpu()
+ else:
+ count = str(len(gt["instances"].gt_boxes))
+ debug = ' '.join([str(s) for s in ["eval_labels", str(self.iter), count,
+ count, 100.,
+ count, 100.,
+ count, 100.,
+ count, 100.,
+ ' - ', ' '.join(['0' for i in range(20)]), 'zero']])
+ logger.info(debug)
+
# =====================================================
# =================== Training Flow ===================
# =====================================================
@@ -448,6 +543,12 @@ class UBTeacherTrainer(DefaultTrainer):
)
joint_proposal_dict["proposals_pseudo_roih"] = pesudo_proposals_roih_unsup_k
+ # debug pseudo labels
+ if self.cfg.SEMISUPNET.DEBUG_BBOX_LABELS:
+ self.eval_label(
+ unlabel_data_q,
+ joint_proposal_dict["proposals_pseudo_roih"])
+
# add pseudo-label to unlabeled data
unlabel_data_q = self.remove_label(unlabel_data_q)
unlabel_data_k = self.remove_label(unlabel_data_k)
@@ -476,13 +577,17 @@ class UBTeacherTrainer(DefaultTrainer):
]
record_dict.update(new_record_all_unlabel_data)
+ # define unsupervised weight regression loss
+ weight_unsup_loss_reg = self.cfg.SEMISUPNET.UNSUP_LOSS_WEIGHT_REG
+
# weight losses
loss_dict = {}
for key in record_dict.keys():
if key[:4] == "loss":
if key == "loss_rpn_loc_pseudo" or key == "loss_box_reg_pseudo":
# pseudo bbox regression <- 0
- loss_dict[key] = record_dict[key] * 0
+ loss_dict[key] = record_dict[key] * \
+ weight_unsup_loss_reg
elif key[-6:] == "pseudo": # unsupervised loss
loss_dict[key] = (
record_dict[key] * self.cfg.SEMISUPNET.UNSUP_LOSS_WEIGHT
diff --git a/ubteacher/modeling/roi_heads/backports.py b/ubteacher/modeling/roi_heads/backports.py
new file mode 100644
index 0000000..6b1cb33
--- /dev/null
+++ b/ubteacher/modeling/roi_heads/backports.py
@@ -0,0 +1,162 @@
+import torch
+
+from detectron2.layers import ShapeSpec, batched_nms, cat, cross_entropy, nonzero_tuple
+from fvcore.nn import giou_loss, smooth_l1_loss
+from detectron2.modeling.box_regression import Box2BoxTransform
+from detectron2.structures import Boxes
+
+
+class FastRCNNOutputs:
+ """
+ An internal implementation that stores information about outputs of a Fast R-CNN head,
+ and provides methods that are used to decode the outputs of a Fast R-CNN head.
+ """
+
+ def __init__(
+ self,
+ box2box_transform,
+ pred_class_logits,
+ pred_proposal_deltas,
+ proposals,
+ smooth_l1_beta=0.0,
+ box_reg_loss_type="smooth_l1",
+ ):
+ """
+ Args:
+ box2box_transform (Box2BoxTransform/Box2BoxTransformRotated):
+ box2box transform instance for proposal-to-detection transformations.
+ pred_class_logits (Tensor): A tensor of shape (R, K + 1) storing the predicted class
+ logits for all R predicted object instances.
+ Each row corresponds to a predicted object instance.
+ pred_proposal_deltas (Tensor): A tensor of shape (R, K * B) or (R, B) for
+ class-specific or class-agnostic regression. It stores the predicted deltas that
+ transform proposals into final box detections.
+ B is the box dimension (4 or 5).
+ When B is 4, each row is [dx, dy, dw, dh (, ....)].
+ When B is 5, each row is [dx, dy, dw, dh, da (, ....)].
+ proposals (list[Instances]): A list of N Instances, where Instances i stores the
+ proposals for image i, in the field "proposal_boxes".
+ When training, each Instances must have ground-truth labels
+ stored in the field "gt_classes" and "gt_boxes".
+ The total number of all instances must be equal to R.
+ smooth_l1_beta (float): The transition point between L1 and L2 loss in
+ the smooth L1 loss function. When set to 0, the loss becomes L1. When
+ set to +inf, the loss becomes constant 0.
+ box_reg_loss_type (str): Box regression loss type. One of: "smooth_l1", "giou"
+ """
+ self.box2box_transform = box2box_transform
+ self.num_preds_per_image = [len(p) for p in proposals]
+ self.pred_class_logits = pred_class_logits
+ self.pred_proposal_deltas = pred_proposal_deltas
+ self.smooth_l1_beta = smooth_l1_beta
+ self.box_reg_loss_type = box_reg_loss_type
+
+ self.image_shapes = [x.image_size for x in proposals]
+
+ if len(proposals):
+ box_type = type(proposals[0].proposal_boxes)
+ # cat(..., dim=0) concatenates over all images in the batch
+ self.proposals = box_type.cat([p.proposal_boxes for p in proposals])
+ assert (
+ not self.proposals.tensor.requires_grad
+ ), "Proposals should not require gradients!"
+
+ # "gt_classes" exists if and only if training. But other gt fields may
+ # not necessarily exist in training for images that have no groundtruth.
+ if proposals[0].has("gt_classes"):
+ self.gt_classes = cat([p.gt_classes for p in proposals], dim=0)
+
+ # If "gt_boxes" does not exist, the proposals must be all negative and
+ # should not be included in regression loss computation.
+ # Here we just use proposal_boxes as an arbitrary placeholder because its
+ # value won't be used in self.box_reg_loss().
+ gt_boxes = [
+ p.gt_boxes if p.has("gt_boxes") else p.proposal_boxes for p in proposals
+ ]
+ self.gt_boxes = box_type.cat(gt_boxes)
+ else:
+ self.proposals = Boxes(torch.zeros(0, 4, device=self.pred_proposal_deltas.device))
+ self._no_instances = len(self.proposals) == 0 # no instances found
+
+ def softmax_cross_entropy_loss(self):
+ """
+ Deprecated
+ """
+ _log_classification_stats(self.pred_class_logits, self.gt_classes)
+ return cross_entropy(self.pred_class_logits, self.gt_classes, reduction="mean")
+
+ def box_reg_loss(self):
+ """
+ Deprecated
+ """
+ if self._no_instances:
+ return 0.0 * self.pred_proposal_deltas.sum()
+
+ box_dim = self.proposals.tensor.size(1) # 4 or 5
+ cls_agnostic_bbox_reg = self.pred_proposal_deltas.size(1) == box_dim
+ device = self.pred_proposal_deltas.device
+
+ bg_class_ind = self.pred_class_logits.shape[1] - 1
+ # Box delta loss is only computed between the prediction for the gt class k
+ # (if 0 <= k < bg_class_ind) and the target; there is no loss defined on predictions
+ # for non-gt classes and background.
+ # Empty fg_inds should produce a valid loss of zero because reduction=sum.
+ fg_inds = nonzero_tuple((self.gt_classes >= 0) & (self.gt_classes < bg_class_ind))[0]
+
+ if cls_agnostic_bbox_reg:
+ # pred_proposal_deltas only corresponds to foreground class for agnostic
+ gt_class_cols = torch.arange(box_dim, device=device)
+ else:
+ # pred_proposal_deltas for class k are located in columns [b * k : b * k + b],
+ # where b is the dimension of box representation (4 or 5)
+ # Note that compared to Detectron1,
+ # we do not perform bounding box regression for background classes.
+ gt_class_cols = box_dim * self.gt_classes[fg_inds, None] + torch.arange(
+ box_dim, device=device
+ )
+
+ if self.box_reg_loss_type == "smooth_l1":
+ gt_proposal_deltas = self.box2box_transform.get_deltas(
+ self.proposals.tensor, self.gt_boxes.tensor
+ )
+ loss_box_reg = smooth_l1_loss(
+ self.pred_proposal_deltas[fg_inds[:, None], gt_class_cols],
+ gt_proposal_deltas[fg_inds],
+ self.smooth_l1_beta,
+ reduction="sum",
+ )
+ elif self.box_reg_loss_type == "giou":
+ fg_pred_boxes = self.box2box_transform.apply_deltas(
+ self.pred_proposal_deltas[fg_inds[:, None], gt_class_cols],
+ self.proposals.tensor[fg_inds],
+ )
+ loss_box_reg = giou_loss(
+ fg_pred_boxes,
+ self.gt_boxes.tensor[fg_inds],
+ reduction="sum",
+ )
+ else:
+ raise ValueError(f"Invalid bbox reg loss type '{self.box_reg_loss_type}'")
+
+ loss_box_reg = loss_box_reg / self.gt_classes.numel()
+ return loss_box_reg
+
+ def losses(self):
+ """
+ Deprecated
+ """
+ return {"loss_cls": self.softmax_cross_entropy_loss(), "loss_box_reg": self.box_reg_loss()}
+
+ def predict_boxes(self):
+ """
+ Deprecated
+ """
+ pred = self.box2box_transform.apply_deltas(self.pred_proposal_deltas, self.proposals.tensor)
+ return pred.split(self.num_preds_per_image, dim=0)
+
+ def predict_probs(self):
+ """
+ Deprecated
+ """
+ probs = F.softmax(self.pred_class_logits, dim=-1)
+ return probs.split(self.num_preds_per_image, dim=0)
diff --git a/ubteacher/modeling/roi_heads/fast_rcnn.py b/ubteacher/modeling/roi_heads/fast_rcnn.py
index 31441ad..1f2af3a 100644
--- a/ubteacher/modeling/roi_heads/fast_rcnn.py
+++ b/ubteacher/modeling/roi_heads/fast_rcnn.py
@@ -1,19 +1,86 @@
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
+import logging
import torch
from torch import nn
from torch.nn import functional as F
+from typing import Tuple, List
+from detectron2.structures import Boxes, Instances
+from detectron2.layers import nonzero_tuple
from detectron2.modeling.roi_heads.fast_rcnn import (
FastRCNNOutputLayers,
- FastRCNNOutputs,
+ fast_rcnn_inference
)
+try:
+ from detectron2.modeling.roi_heads.fast_rcnn import FastRCNNOutputs
+except ImportError:
+ from .backports import FastRCNNOutputs
+
+
# focal loss
class FastRCNNFocaltLossOutputLayers(FastRCNNOutputLayers):
def __init__(self, cfg, input_shape):
super(FastRCNNFocaltLossOutputLayers, self).__init__(cfg, input_shape)
self.num_classes = cfg.MODEL.ROI_HEADS.NUM_CLASSES
+ # config bbox iou branch
+ self.with_iou_pred = cfg.SEMISUPNET.WITH_IOU_PRED
+ self.iou_threshold = cfg.SEMISUPNET.IOU_PRED_THRESHOLD
+ self.iou_inference_threshold = cfg.SEMISUPNET.IOU_PRED_INFERENCE_THRESHOLD
+ self.filter_with_bbox_iou = cfg.SEMISUPNET.FILTER_WITH_BBOX_IOU
+ self.with_score_values = cfg.SEMISUPNET.WITH_SCORE_VALUES
+ self.with_iou_pred_elu = cfg.SEMISUPNET.WITH_IOU_PRED_ELU
+ self.with_regr_values = cfg.SEMISUPNET.WITH_REGR_VALUES
+ self.with_bbox_iou_regres = cfg.SEMISUPNET.WITH_BBOX_IOU_REGRES
+ self.use_smoothl1 = cfg.SEMISUPNET.USE_SMOOTHL1
+
+ if self.with_iou_pred:
+ inter_channels = input_shape.channels
+ in_channels = inter_channels
+
+ if self.with_score_values:
+ in_channels += (self.num_classes + 1)
+
+ if self.with_regr_values:
+ in_channels += (self.num_classes * 4)
+
+ if not self.with_iou_pred_elu:
+ self.iou_pred = torch.nn.Sequential(
+ torch.nn.Linear(in_channels, inter_channels),
+ torch.nn.Linear(inter_channels, self.num_classes + 1)
+ )
+ else:
+ self.iou_pred = torch.nn.Sequential(
+ torch.nn.Linear(in_channels, inter_channels),
+ nn.ELU(),
+ torch.nn.Linear(inter_channels, self.num_classes + 1)
+ )
+
+ def init_bbox_iou_weights(m):
+ if type(m) == nn.Linear:
+ nn.init.normal_(m.weight, std=0.01)
+ nn.init.constant_(m.bias, 0)
+
+ self.iou_pred.apply(init_bbox_iou_weights)
+
+ def forward(self, x):
+ scores, proposal_deltas = super(FastRCNNFocaltLossOutputLayers, self).forward(x)
+
+ iou_scores = None
+ if self.with_iou_pred:
+ deltas = proposal_deltas
+
+ iou_input = x
+ if self.with_score_values:
+ iou_input = torch.cat([iou_input, scores], dim=1)
+ if self.with_regr_values:
+ iou_input = torch.cat([iou_input, deltas], dim=1)
+
+ iou_scores = self.iou_pred(iou_input)
+
+ return scores, proposal_deltas, iou_scores
+
def losses(self, predictions, proposals):
"""
Args:
@@ -21,7 +88,8 @@ class FastRCNNFocaltLossOutputLayers(FastRCNNOutputLayers):
proposals (list[Instances]): proposals that match the features
that were used to compute predictions.
"""
- scores, proposal_deltas = predictions
+ scores, proposal_deltas, iou_scores = predictions
+
losses = FastRCNNFocalLoss(
self.box2box_transform,
scores,
@@ -30,10 +98,95 @@ class FastRCNNFocaltLossOutputLayers(FastRCNNOutputLayers):
self.smooth_l1_beta,
self.box_reg_loss_type,
num_classes=self.num_classes,
+ iou_threshold=self.iou_threshold,
+ iou_scores=iou_scores,
+ with_bbox_iou_regres=self.with_bbox_iou_regres,
+ use_smoothl1=self.use_smoothl1,
).losses()
return losses
+ def inference(self, predictions: Tuple[torch.Tensor, torch.Tensor], proposals: List[Instances], branch: str = ""):
+ """
+ Args:
+ predictions: return values of :meth:`forward()`.
+ proposals (list[Instances]): proposals that match the features that were
+ used to compute predictions. The ``proposal_boxes`` field is expected.
+ Returns:
+ list[Instances]: same as `fast_rcnn_inference`.
+ list[Tensor]: same as `fast_rcnn_inference`.
+ """
+ logger = logging.getLogger('fvcore.common.checkpoint')
+ boxes = self.predict_boxes(predictions[:2], proposals)
+ scores = self.predict_probs(predictions[:2], proposals)
+ # filter by iou score
+ if len(predictions) > 2 and predictions[2] is not None:
+ iou_scores = self.predict_iou(predictions[-1], proposals)
+ if self.filter_with_bbox_iou:
+ clss = []
+ for s in scores:
+ clss.append(s.max(1)[1] if len(s) > 0 else torch.tensor([], dtype=torch.long, device=s.device))
+ counts = [b.shape[0] for b in boxes]
+ logger.info("bbox_iou bg {}".format(sum([len(s[s==self.num_classes]) for s in clss])))
+ boxes = self.filter(boxes, iou_scores, clss)
+ logger.info("bbox_iou filtering {}".format(sum([c-b.shape[0] for (c, b) in zip(counts, boxes)])))
+ scores = self.filter(scores, iou_scores, clss)
+ iou_scores = self.filter(iou_scores, iou_scores, clss)
+
+ image_shapes = [x.image_size for x in proposals]
+
+ # only on teacher pseudo-labeling!
+ if branch == 'unsup_data_weak' and self.with_teacher_filter_preds:
+ th = self.teacher_filter_preds_min_delta
+ boxes, scores = self._filter_preds(boxes, scores, th)
+
+ instances, idxs = fast_rcnn_inference(
+ boxes,
+ scores,
+ image_shapes,
+ self.test_score_thresh,
+ self.test_nms_thresh,
+ self.test_topk_per_image,
+ )
+
+ if len(predictions) > 2 and predictions[2] is not None:
+ instances = self.add_iou_scores(iou_scores, instances, idxs)
+
+ return instances, idxs
+
+ def _filter_preds(self, boxes, scores, min_delta):
+ boxes = list(boxes)
+ scores = list(scores)
+ for idx in range(len(boxes)):
+ top2 = scores[idx].topk(2)[0]
+ filter_ = (top2[:, 0] - top2[:, 1]) > min_delta
+ boxes[idx] = boxes[idx][filter_]
+ scores[idx] = scores[idx][filter_]
+ return boxes, scores
+
+ def predict_iou(self, preds, proposals):
+ num_inst_per_image = [len(p) for p in proposals]
+ preds = preds.sigmoid()
+ return preds.split(num_inst_per_image, dim=0)
+
+ def add_iou_scores(self, iou_scores, instances, idxs):
+ for i, instance in enumerate(instances):
+ ious = iou_scores[i][idxs[i]]
+ counter = torch.arange(0, len(instance))
+ ious = ious[counter, instance.pred_classes]
+ instance.pred_iou_scores = ious
+ return instances
+
+ def filter(self, x, iou_scores, scores):
+ res = []
+ for idx, val in enumerate(x):
+ ious = iou_scores[idx][torch.arange(0, scores[idx].shape[0]), scores[idx]]
+ fit = (
+ (scores[idx] < self.num_classes) & (ious > self.iou_inference_threshold)
+ ) | (scores[idx] == self.num_classes)
+ res.append(val[fit])
+ return tuple(res)
+
class FastRCNNFocalLoss(FastRCNNOutputs):