-
Notifications
You must be signed in to change notification settings - Fork 2
/
bevfusion_changes.patch
4031 lines (4005 loc) · 152 KB
/
bevfusion_changes.patch
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/.gitignore b/.gitignore
index b6e4761..fe21cf8 100644
--- a/.gitignore
+++ b/.gitignore
@@ -5,6 +5,11 @@ __pycache__/
# C extensions
*.so
+data/
+
+*.pkl
+*.pth
+.vscode/
# Distribution / packaging
.Python
diff --git a/extend/__init__.py b/extend/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/extend/custom_func.py b/extend/custom_func.py
new file mode 100644
index 0000000..94e5057
--- /dev/null
+++ b/extend/custom_func.py
@@ -0,0 +1,102 @@
+# only for bevfusion
+import torch
+import torchvision.transforms as transforms
+import torch.nn.functional as F
+import torchvision
+from torchvision.utils import save_image
+import copy
+
+
+def custom_data_preprocess(data):
+ return data
+
+
+def custom_data_postprocess_eval(data):
+ return data
+
+
+def custom_data_work(data):
+ metas = data["metas"]._data[0][0]
+ img_path_list = metas["filename"]
+ img_org_np = metas["img_org"]
+ img_processed = data["img"]._data[0].clone()
+ gt_labels_3d = data["gt_labels_3d"]._data[0][0]
+ return metas, img_path_list, img_org_np, img_processed, gt_labels_3d
+
+
+def custom_data_work_point(data):
+ metas = data["metas"]._data[0][0]
+ img_path_list = metas["filename"]
+ img_org_np = metas["img_org"]
+ img_processed = data["img"]._data[0].clone()
+ gt_labels_3d = data["gt_labels_3d"]._data[0][0]
+ points_tensor = data["points"]._data[0][0].clone()
+ return metas, img_path_list, img_org_np, img_processed, gt_labels_3d, points_tensor
+
+
+def custom_result_postprocess(result):
+ return result
+
+
+def custom_img_read_from_img_org(img_org_np, device):
+ img_org_np_255_rgb_hwcn_uint8 = img_org_np # PIL读取 RGB 直接转 numpy
+ img_org_tensor_rgb_255_hwcn = torch.from_numpy(
+ img_org_np_255_rgb_hwcn_uint8
+ ).float()
+ img_org_tensor_rgb_255 = img_org_tensor_rgb_255_hwcn.permute(3, 2, 0, 1)
+ img_tensor_rgb_6chw_0to1 = (img_org_tensor_rgb_255 / 255.0).to(device)
+ return img_tensor_rgb_6chw_0to1
+
+
+def custom_differentiable_transform(img_tensor_rgb_6chw_0to1, img_metas):
+ """Alternative Data Preparation for Original Model
+
+ Args:
+ img_tensor (torch.tensor): (6xCxHxW), tensors of original imgs
+ """
+
+ assert len(img_tensor_rgb_6chw_0to1.shape) == 4
+ assert img_tensor_rgb_6chw_0to1.shape[0] == 6
+ assert img_tensor_rgb_6chw_0to1.shape[1] == 3
+ assert img_tensor_rgb_6chw_0to1.max() <= 1.0
+ assert img_tensor_rgb_6chw_0to1.min() >= 0.0
+ assert img_tensor_rgb_6chw_0to1.dtype == torch.float32
+ assert img_tensor_rgb_6chw_0to1.is_cuda
+ img_tensor = img_tensor_rgb_6chw_0to1
+
+ device = img_tensor_rgb_6chw_0to1.device
+ mean = [0.485, 0.456, 0.406]
+ std = [0.229, 0.224, 0.225]
+ mean = torch.tensor(mean).to(device)[None, None, :, None, None]
+ std = torch.tensor(std).to(device)[None, None, :, None, None]
+
+ ############ resize norm pad
+ ######## resize
+ resize_size = (432, 768)
+ img_tensor_resize = F.interpolate(
+ img_tensor, resize_size, mode="bilinear", align_corners=False
+ )
+
+ ######## crop
+ crop_size = (32, 176, 736, 432)
+ img_tensor_crop = img_tensor_resize[
+ ..., crop_size[1] : crop_size[3], crop_size[0] : crop_size[2]
+ ]
+
+ ######## norm
+ img_tensor_norm = (img_tensor_crop - mean) / std
+
+ return img_tensor_norm
+
+
+def custom_image_data_give(data, image_ready):
+ data_copy = copy.deepcopy(data)
+ data_copy["img"]._data[0] = image_ready
+ return data_copy
+
+
+def custom_image_data_give_point(data, image_ready, points_ready):
+ data_copy = copy.deepcopy(data)
+ data_copy["img"]._data[0] = image_ready
+ data_copy["points"]._data[0][0] = points_ready
+ return data_copy
diff --git a/extend_common b/extend_common
new file mode 120000
index 0000000..0633dd9
--- /dev/null
+++ b/extend_common
@@ -0,0 +1 @@
+../extend_common/
\ No newline at end of file
diff --git a/mmdet3d/apis_common b/mmdet3d/apis_common
new file mode 120000
index 0000000..b44e193
--- /dev/null
+++ b/mmdet3d/apis_common
@@ -0,0 +1 @@
+../../apis_common/
\ No newline at end of file
diff --git a/mmdet3d/datasets/pipelines/formating.py b/mmdet3d/datasets/pipelines/formating.py
index 3e781ce..34ac202 100644
--- a/mmdet3d/datasets/pipelines/formating.py
+++ b/mmdet3d/datasets/pipelines/formating.py
@@ -157,6 +157,8 @@ class Collect3D:
"pcd_rotation",
"lidar_path",
"transformation_3d_flow",
+ # zzj api
+ 'img_org'
),
):
self.keys = keys
diff --git a/mmdet3d/datasets/pipelines/loading.py b/mmdet3d/datasets/pipelines/loading.py
index cac5b6d..8b57313 100644
--- a/mmdet3d/datasets/pipelines/loading.py
+++ b/mmdet3d/datasets/pipelines/loading.py
@@ -70,6 +70,9 @@ class LoadMultiViewImageFromFiles:
results["pad_shape"] = images[0].size
results["scale_factor"] = 1.0
+ # zzj add
+ results["img_org"] = np.stack([np.asarray(image) for image in images],axis=-1)
+
return results
def __repr__(self):
diff --git a/mmdet3d/models/fusion_models/bevfusion.py b/mmdet3d/models/fusion_models/bevfusion.py
index 2931b48..301f5dd 100644
--- a/mmdet3d/models/fusion_models/bevfusion.py
+++ b/mmdet3d/models/fusion_models/bevfusion.py
@@ -126,13 +126,20 @@ class BEVFusion(Base3DFusionModel):
)
return x
+ # def extract_lidar_features(self, x) -> torch.Tensor:
+ # feats, coords, sizes = self.voxelize(x)
+ # batch_size = coords[-1, 0] + 1
+ # x = self.encoders["lidar"]["backbone"](feats, coords, batch_size, sizes=sizes)
+ # return x
def extract_lidar_features(self, x) -> torch.Tensor:
feats, coords, sizes = self.voxelize(x)
+ # zzj api voxelization
+ coords = coords[:,[0,3,2,1]] # z,y,x -->> x,y,z plz check mmcvfull1.6.1: mmcv\ops\csrc\common\cuda\voxelization_cuda_kernel.cuh L-45
batch_size = coords[-1, 0] + 1
- x = self.encoders["lidar"]["backbone"](feats, coords, batch_size, sizes=sizes)
+ x = self.encoders["lidar"]["backbone"](feats, coords, batch_size, sizes=sizes) # SparseEncoder
return x
- @torch.no_grad()
+ # @torch.no_grad()
@force_fp32()
def voxelize(self, points):
feats, coords, sizes = [], [], []
@@ -261,7 +268,8 @@ class BEVFusion(Base3DFusionModel):
x = self.decoder["backbone"](x)
x = self.decoder["neck"](x)
- if self.training:
+ # if self.training:
+ if kwargs['return_loss']:
outputs = {}
for type, head in self.heads.items():
if type == "object":
diff --git a/mmdet3d/models/vtransforms/base.py b/mmdet3d/models/vtransforms/base.py
index 400054c..2094e18 100644
--- a/mmdet3d/models/vtransforms/base.py
+++ b/mmdet3d/models/vtransforms/base.py
@@ -240,7 +240,7 @@ class BaseDepthTransform(BaseTransform):
)
for b in range(batch_size):
- cur_coords = points[b][:, :3]
+ cur_coords = points[b][:, :3].detach()
cur_img_aug_matrix = img_aug_matrix[b]
cur_lidar_aug_matrix = lidar_aug_matrix[b]
cur_lidar2image = lidar2image[b]
@@ -272,7 +272,7 @@ class BaseDepthTransform(BaseTransform):
& (cur_coords[..., 1] < self.image_size[1])
& (cur_coords[..., 1] >= 0)
)
- for c in range(on_img.shape[0]):
+ for c in range(on_img.shape[0]): # 把 有lidar的点 在depth 上 打出来,数值为真实深度(m)
masked_coords = cur_coords[c, on_img[c]].long()
masked_dist = dist[c, on_img[c]]
depth[b, c, 0, masked_coords[:, 0], masked_coords[:, 1]] = masked_dist
diff --git a/mmdet3d/models/vtransforms/depth_lss.py b/mmdet3d/models/vtransforms/depth_lss.py
index b7cd45d..263aa4a 100644
--- a/mmdet3d/models/vtransforms/depth_lss.py
+++ b/mmdet3d/models/vtransforms/depth_lss.py
@@ -82,11 +82,11 @@ class DepthLSSTransform(BaseDepthTransform):
def get_cam_feats(self, x, d):
B, N, C, fH, fW = x.shape
- d = d.view(B * N, *d.shape[2:])
+ d = d.view(B * N, *d.shape[2:]) # 这里的d 可是从LiDAR 投影过来的,是GT啊!
x = x.view(B * N, C, fH, fW)
- d = self.dtransform(d)
- x = torch.cat([d, x], dim=1)
+ d = self.dtransform(d) # 但gt的 d 还是被投入 模型运算 了
+ x = torch.cat([d, x], dim=1) # 处理后的d 作为特征的一部分,供深度预测了
x = self.depthnet(x)
depth = x[:, : self.D].softmax(dim=1)
diff --git a/mmdet3d/ops/voxel/src/voxelization.h b/mmdet3d/ops/voxel/src/voxelization.h
index 765b30a..1e96c4e 100644
--- a/mmdet3d/ops/voxel/src/voxelization.h
+++ b/mmdet3d/ops/voxel/src/voxelization.h
@@ -27,6 +27,11 @@ int hard_voxelize_gpu(const at::Tensor &points, at::Tensor &voxels,
const std::vector<float> voxel_size,
const std::vector<float> coors_range,
const int max_points, const int max_voxels,
+ // zzj api 20220224
+ // : add input valuables for return
+ at::Tensor &point_to_pointidx,
+ at::Tensor &point_to_voxelidx,
+ at::Tensor &coor_to_voxelidx,
const int NDim = 3);
int nondisterministic_hard_voxelize_gpu(const at::Tensor &points, at::Tensor &voxels,
@@ -60,12 +65,22 @@ inline int hard_voxelize(const at::Tensor &points, at::Tensor &voxels,
const std::vector<float> voxel_size,
const std::vector<float> coors_range,
const int max_points, const int max_voxels,
+ // zzj api 20220224
+ // : add input valuables for return
+ at::Tensor &point_to_pointidx,
+ at::Tensor &point_to_voxelidx,
+ at::Tensor &coor_to_voxelidx,
const int NDim = 3, const bool deterministic = true) {
if (points.device().is_cuda()) {
#ifdef WITH_CUDA
if (deterministic) {
return hard_voxelize_gpu(points, voxels, coors, num_points_per_voxel,
voxel_size, coors_range, max_points, max_voxels,
+ // zzj api 20220224
+ // : add input valuables for return
+ point_to_pointidx,
+ point_to_voxelidx,
+ coor_to_voxelidx,
NDim);
}
return nondisterministic_hard_voxelize_gpu(points, voxels, coors, num_points_per_voxel,
diff --git a/mmdet3d/ops/voxel/src/voxelization_cpu.cpp b/mmdet3d/ops/voxel/src/voxelization_cpu.cpp
index 1f87e26..6bcec40 100644
--- a/mmdet3d/ops/voxel/src/voxelization_cpu.cpp
+++ b/mmdet3d/ops/voxel/src/voxelization_cpu.cpp
@@ -27,7 +27,7 @@ void dynamic_voxelize_kernel(const torch::TensorAccessor<T, 2> points,
failed = true;
break;
}
- coor[j] = c;
+ coor[ndim_minus_1 - j] = c;
}
for (int k = 0; k < NDim; ++k) {
diff --git a/mmdet3d/ops/voxel/src/voxelization_cuda.cu b/mmdet3d/ops/voxel/src/voxelization_cuda.cu
index 8191cba..b3b4af5 100644
--- a/mmdet3d/ops/voxel/src/voxelization_cuda.cu
+++ b/mmdet3d/ops/voxel/src/voxelization_cuda.cu
@@ -53,9 +53,9 @@ __global__ void dynamic_voxelize_kernel(
coors_offset[1] = -1;
coors_offset[2] = -1;
} else {
- coors_offset[0] = c_x;
+ coors_offset[0] = c_z;
coors_offset[1] = c_y;
- coors_offset[2] = c_z;
+ coors_offset[2] = c_x;
}
}
}
@@ -166,10 +166,10 @@ __global__ void determin_voxel_num(
int voxelidx = voxel_num[0];
if (voxel_num[0] >= max_voxels) continue;
voxel_num[0] += 1;
- coor_to_voxelidx[i] = voxelidx;
+ coor_to_voxelidx[i] = voxelidx; // coor_to_voxelidx 应该改名为 point_to_voxelidx
num_points_per_voxel[voxelidx] = 1;
} else {
- int point_idx = point_to_pointidx[i];
+ int point_idx = point_to_pointidx[i]; // 从当前点,跳到体素的第一个点,再用第一个点查询体素的index
int voxelidx = coor_to_voxelidx[point_idx];
if (voxelidx != -1) {
coor_to_voxelidx[i] = voxelidx;
@@ -233,6 +233,11 @@ int hard_voxelize_gpu(const at::Tensor& points, at::Tensor& voxels,
const std::vector<float> voxel_size,
const std::vector<float> coors_range,
const int max_points, const int max_voxels,
+ // zzj api 20220224
+ // : add for return
+ at::Tensor &point_to_pointidx,
+ at::Tensor &point_to_voxelidx,
+ at::Tensor &coor_to_voxelidx,
const int NDim = 3) {
// current version tooks about 0.04s for one frame on cpu
// check device
@@ -280,16 +285,20 @@ int hard_voxelize_gpu(const at::Tensor& points, at::Tensor& voxels,
// 2. map point to the idx of the corresponding voxel, find duplicate coor
// create some temporary variables
- auto point_to_pointidx = -at::ones(
- {
- num_points,
- },
- points.options().dtype(at::kInt));
- auto point_to_voxelidx = -at::ones(
- {
- num_points,
- },
- points.options().dtype(at::kInt));
+
+ // zzj api 20220224
+ // : use outside coming values defined in python
+ // not defined in C
+ // auto point_to_pointidx = -at::ones(
+ // {
+ // num_points,
+ // },
+ // points.options().dtype(at::kInt));
+ // auto point_to_voxelidx = -at::ones(
+ // {
+ // num_points,
+ // },
+ // points.options().dtype(at::kInt));
dim3 map_grid(std::min(at::cuda::ATenCeilDiv(num_points, 512), 4096));
dim3 map_block(512);
@@ -307,11 +316,16 @@ int hard_voxelize_gpu(const at::Tensor& points, at::Tensor& voxels,
// 3. determin voxel num and voxel's coor index
// make the logic in the CUDA device could accelerate about 10 times
- auto coor_to_voxelidx = -at::ones(
- {
- num_points,
- },
- points.options().dtype(at::kInt));
+
+ // zzj api 20220224
+ // : use outside coming values defined in python
+ // not defined in C
+ // auto coor_to_voxelidx = -at::ones(
+ // {
+ // num_points,
+ // },
+ // points.options().dtype(at::kInt));
+
auto voxel_num = at::zeros(
{
1,
diff --git a/mmdet3d/ops/voxel/src_backup/scatter_points_cpu.cpp b/mmdet3d/ops/voxel/src_backup/scatter_points_cpu.cpp
new file mode 100644
index 0000000..c22b8ae
--- /dev/null
+++ b/mmdet3d/ops/voxel/src_backup/scatter_points_cpu.cpp
@@ -0,0 +1,122 @@
+#include <ATen/TensorUtils.h>
+#include <torch/extension.h>
+// #include "voxelization.h"
+
+namespace {
+
+template <typename T_int>
+void determin_max_points_kernel(
+ torch::TensorAccessor<T_int, 2> coor,
+ torch::TensorAccessor<T_int, 1> point_to_voxelidx,
+ torch::TensorAccessor<T_int, 1> num_points_per_voxel,
+ torch::TensorAccessor<T_int, 3> coor_to_voxelidx, int& voxel_num,
+ int& max_points, const int num_points) {
+ int voxelidx, num;
+ for (int i = 0; i < num_points; ++i) {
+ if (coor[i][0] == -1) continue;
+ voxelidx = coor_to_voxelidx[coor[i][0]][coor[i][1]][coor[i][2]];
+
+ // record voxel
+ if (voxelidx == -1) {
+ voxelidx = voxel_num;
+ voxel_num += 1;
+ coor_to_voxelidx[coor[i][0]][coor[i][1]][coor[i][2]] = voxelidx;
+ }
+
+ // put points into voxel
+ num = num_points_per_voxel[voxelidx];
+ point_to_voxelidx[i] = num;
+ num_points_per_voxel[voxelidx] += 1;
+
+ // update max points per voxel
+ max_points = std::max(max_points, num + 1);
+ }
+
+ return;
+}
+
+template <typename T, typename T_int>
+void scatter_point_to_voxel_kernel(
+ const torch::TensorAccessor<T, 2> points,
+ torch::TensorAccessor<T_int, 2> coor,
+ torch::TensorAccessor<T_int, 1> point_to_voxelidx,
+ torch::TensorAccessor<T_int, 3> coor_to_voxelidx,
+ torch::TensorAccessor<T, 3> voxels,
+ torch::TensorAccessor<T_int, 2> voxel_coors, const int num_features,
+ const int num_points, const int NDim) {
+ for (int i = 0; i < num_points; ++i) {
+ int num = point_to_voxelidx[i];
+ int voxelidx = coor_to_voxelidx[coor[i][0]][coor[i][1]][coor[i][2]];
+ for (int k = 0; k < num_features; ++k) {
+ voxels[voxelidx][num][k] = points[i][k];
+ }
+ for (int k = 0; k < NDim; ++k) {
+ voxel_coors[voxelidx][k] = coor[i][k];
+ }
+ }
+}
+
+} // namespace
+
+namespace voxelization {
+
+std::vector<at::Tensor> dynamic_point_to_voxel_cpu(
+ const at::Tensor& points, const at::Tensor& voxel_mapping,
+ const std::vector<float> voxel_size, const std::vector<float> coors_range) {
+ // current version tooks about 0.02s_0.03s for one frame on cpu
+ // check device
+ AT_ASSERTM(points.device().is_cpu(), "points must be a CPU tensor");
+
+ const int NDim = voxel_mapping.size(1);
+ const int num_points = points.size(0);
+ const int num_features = points.size(1);
+
+ std::vector<int> grid_size(NDim);
+ for (int i = 0; i < NDim; ++i) {
+ grid_size[i] =
+ round((coors_range[NDim + i] - coors_range[i]) / voxel_size[i]);
+ }
+
+ at::Tensor num_points_per_voxel = at::zeros(
+ {
+ num_points,
+ },
+ voxel_mapping.options());
+ at::Tensor coor_to_voxelidx = -at::ones(
+ {grid_size[2], grid_size[1], grid_size[0]}, voxel_mapping.options());
+ at::Tensor point_to_voxelidx = -at::ones(
+ {
+ num_points,
+ },
+ voxel_mapping.options());
+
+ int voxel_num = 0;
+ int max_points = 0;
+ AT_DISPATCH_ALL_TYPES(voxel_mapping.scalar_type(), "determin_max_point", [&] {
+ determin_max_points_kernel<scalar_t>(
+ voxel_mapping.accessor<scalar_t, 2>(),
+ point_to_voxelidx.accessor<scalar_t, 1>(),
+ num_points_per_voxel.accessor<scalar_t, 1>(),
+ coor_to_voxelidx.accessor<scalar_t, 3>(), voxel_num, max_points,
+ num_points);
+ });
+
+ at::Tensor voxels =
+ at::zeros({voxel_num, max_points, num_features}, points.options());
+ at::Tensor voxel_coors =
+ at::zeros({voxel_num, NDim}, points.options().dtype(at::kInt));
+
+ AT_DISPATCH_ALL_TYPES(points.scalar_type(), "scatter_point_to_voxel", [&] {
+ scatter_point_to_voxel_kernel<scalar_t, int>(
+ points.accessor<scalar_t, 2>(), voxel_mapping.accessor<int, 2>(),
+ point_to_voxelidx.accessor<int, 1>(),
+ coor_to_voxelidx.accessor<int, 3>(), voxels.accessor<scalar_t, 3>(),
+ voxel_coors.accessor<int, 2>(), num_features, num_points, NDim);
+ });
+
+ at::Tensor num_points_per_voxel_out =
+ num_points_per_voxel.slice(/*dim=*/0, /*start=*/0, /*end=*/voxel_num);
+ return {voxels, voxel_coors, num_points_per_voxel_out};
+}
+
+} // namespace voxelization
diff --git a/mmdet3d/ops/voxel/src_backup/scatter_points_cuda.cu b/mmdet3d/ops/voxel/src_backup/scatter_points_cuda.cu
new file mode 100644
index 0000000..2ed1869
--- /dev/null
+++ b/mmdet3d/ops/voxel/src_backup/scatter_points_cuda.cu
@@ -0,0 +1,310 @@
+#include <ATen/ATen.h>
+#include <ATen/cuda/CUDAContext.h>
+#include <torch/types.h>
+
+#include <ATen/cuda/CUDAApplyUtils.cuh>
+
+typedef enum { SUM = 0, MEAN = 1, MAX = 2 } reduce_t;
+
+#define CHECK_CUDA(x) \
+ TORCH_CHECK(x.device().is_cuda(), #x " must be a CUDA tensor")
+#define CHECK_CONTIGUOUS(x) \
+ TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
+#define CHECK_INPUT(x) \
+ CHECK_CUDA(x); \
+ CHECK_CONTIGUOUS(x)
+
+namespace {
+int const threadsPerBlock = 512;
+int const maxGridDim = 50000;
+} // namespace
+
+__device__ __forceinline__ static void reduceMax(float *address, float val) {
+ int *address_as_i = reinterpret_cast<int *>(address);
+ int old = *address_as_i, assumed;
+ do {
+ assumed = old;
+ old = atomicCAS(address_as_i, assumed,
+ __float_as_int(fmaxf(val, __int_as_float(assumed))));
+ } while (assumed != old || __int_as_float(old) < val);
+}
+
+__device__ __forceinline__ static void reduceMax(double *address, double val) {
+ unsigned long long *address_as_ull =
+ reinterpret_cast<unsigned long long *>(address);
+ unsigned long long old = *address_as_ull, assumed;
+ do {
+ assumed = old;
+ old = atomicCAS(
+ address_as_ull, assumed,
+ __double_as_longlong(fmax(val, __longlong_as_double(assumed))));
+ } while (assumed != old || __longlong_as_double(old) < val);
+}
+
+// get rid of meaningless warnings when compiling host code
+#ifdef __CUDA_ARCH__
+__device__ __forceinline__ static void reduceAdd(float *address, float val) {
+#if (__CUDA_ARCH__ < 200)
+#warning \
+ "compute capability lower than 2.x. fall back to use CAS version of atomicAdd for float32"
+ int *address_as_i = reinterpret_cast<int *>(address);
+ int old = *address_as_i, assumed;
+ do {
+ assumed = old;
+ old = atomicCAS(address_as_i, assumed,
+ __float_as_int(val + __int_as_float(assumed)));
+ } while (assumed != old);
+#else
+ atomicAdd(address, val);
+#endif
+}
+
+__device__ __forceinline__ static void reduceAdd(double *address, double val) {
+#if (__CUDA_ARCH__ < 600)
+#warning \
+ "compute capability lower than 6.x. fall back to use CAS version of atomicAdd for float64"
+ unsigned long long *address_as_ull =
+ reinterpret_cast<unsigned long long *>(address);
+ unsigned long long old = *address_as_ull, assumed;
+ do {
+ assumed = old;
+ old = atomicCAS(address_as_ull, assumed,
+ __double_as_longlong(val + __longlong_as_double(assumed)));
+ } while (assumed != old);
+#else
+ atomicAdd(address, val);
+#endif
+}
+#endif
+
+template <typename T>
+__global__ void
+feats_reduce_kernel(const T *feats, const int32_t *coors_map,
+ T *reduced_feats, // shall be 0 at initialization
+ const int num_input, const int num_feats,
+ const reduce_t reduce_type) {
+ for (int x = blockIdx.x * blockDim.x + threadIdx.x; x < num_input;
+ x += gridDim.x * blockDim.x) {
+ int32_t reduce_to = coors_map[x];
+ if (reduce_to == -1) continue;
+
+ const T *feats_offset = feats + x * num_feats;
+ T *reduced_feats_offset = reduced_feats + reduce_to * num_feats;
+ if (reduce_type == reduce_t::MAX) {
+ for (int i = 0; i < num_feats; i++) {
+ reduceMax(&reduced_feats_offset[i], feats_offset[i]);
+ }
+ } else {
+ for (int i = 0; i < num_feats; i++) {
+ reduceAdd(&reduced_feats_offset[i], feats_offset[i]);
+ }
+ }
+ }
+}
+
+template <typename T>
+__global__ void add_reduce_traceback_grad_kernel(
+ T *grad_feats, const T *grad_reduced_feats, const int32_t *coors_map,
+ const int32_t *reduce_count, const int num_input, const int num_feats,
+ const reduce_t reduce_type) {
+ for (int x = blockIdx.x * blockDim.x + threadIdx.x; x < num_input;
+ x += gridDim.x * blockDim.x) {
+ int32_t reduce_to = coors_map[x];
+ if (reduce_to == -1) {
+ continue;
+ }
+
+ const int input_offset = x * num_feats;
+ T *grad_feats_offset = grad_feats + input_offset;
+ const int reduced_offset = reduce_to * num_feats;
+ const T *grad_reduced_feats_offset = grad_reduced_feats + reduced_offset;
+
+ if (reduce_type == reduce_t::SUM) {
+ for (int i = 0; i < num_feats; i++) {
+ grad_feats_offset[i] = grad_reduced_feats_offset[i];
+ }
+ } else if (reduce_type == reduce_t::MEAN) {
+ for (int i = 0; i < num_feats; i++) {
+ grad_feats_offset[i] = grad_reduced_feats_offset[i] /
+ static_cast<T>(reduce_count[reduce_to]);
+ }
+ }
+ }
+}
+
+template <typename T>
+__global__ void max_reduce_traceback_scatter_idx_kernel(
+ const T *feats, const T *reduced_feats, int32_t *reduce_from,
+ const int32_t *coors_map, const int num_input, const int num_feats) {
+ for (int x = blockIdx.x * blockDim.x + threadIdx.x; x < num_input;
+ x += gridDim.x * blockDim.x) {
+ int32_t reduce_to = coors_map[x];
+
+ const int input_offset = x * num_feats;
+ const T *feats_offset = feats + input_offset;
+
+ if (reduce_to == -1) {
+ continue;
+ }
+
+ const int reduced_offset = reduce_to * num_feats;
+ const T *reduced_feats_offset = reduced_feats + reduced_offset;
+ int32_t *reduce_from_offset = reduce_from + reduced_offset;
+
+ for (int i = 0; i < num_feats; i++) {
+ if (feats_offset[i] == reduced_feats_offset[i]) {
+ atomicMin(&reduce_from_offset[i], static_cast<int32_t>(x));
+ }
+ }
+ }
+}
+
+template <typename T>
+__global__ void max_reduce_scatter_grad_kernel(T *grad_feats,
+ const T *grad_reduced_feats,
+ const int32_t *reduce_from,
+ const int num_reduced,
+ const int num_feats) {
+ for (int x = blockIdx.x * blockDim.x + threadIdx.x; x < num_reduced;
+ x += gridDim.x * blockDim.x) {
+ const int reduced_offset = x * num_feats;
+ const int32_t *scatter_to_offset = reduce_from + reduced_offset;
+ const T *grad_reduced_feats_offset = grad_reduced_feats + reduced_offset;
+
+ for (int i = 0; i < num_feats; i++) {
+ grad_feats[scatter_to_offset[i] * num_feats + i] =
+ grad_reduced_feats_offset[i];
+ }
+ }
+}
+
+namespace voxelization {
+
+std::vector<at::Tensor> dynamic_point_to_voxel_forward_gpu(
+ const at::Tensor &feats, const at::Tensor &coors,
+ const reduce_t reduce_type) {
+ CHECK_INPUT(feats);
+ CHECK_INPUT(coors);
+
+ const int num_input = feats.size(0);
+ const int num_feats = feats.size(1);
+
+ if (num_input == 0)
+ return {feats.clone().detach(),
+ coors.clone().detach(),
+ coors.new_empty({0}, torch::kInt32),
+ coors.new_empty({0}, torch::kInt32)};
+
+ at::Tensor out_coors;
+ at::Tensor coors_map;
+ at::Tensor reduce_count;
+
+ auto coors_clean = coors.masked_fill(coors.lt(0).any(-1, true), -1);
+
+ std::tie(out_coors, coors_map, reduce_count) =
+ at::unique_dim(coors_clean, 0, true, true, true);
+
+ if (out_coors.index({0, 0}).lt(0).item<bool>()) {
+ // the first element of out_coors (-1,-1,-1) and should be removed
+ out_coors = out_coors.slice(0, 1);
+ reduce_count = reduce_count.slice(0, 1);
+ coors_map = coors_map - 1;
+ }
+
+ coors_map = coors_map.to(torch::kInt32);
+ reduce_count = reduce_count.to(torch::kInt32);
+
+ auto reduced_feats =
+ at::empty({out_coors.size(0), num_feats}, feats.options());
+
+ AT_DISPATCH_FLOATING_TYPES(
+ feats.scalar_type(), "feats_reduce_kernel", ([&] {
+ if (reduce_type == reduce_t::MAX)
+ reduced_feats.fill_(-std::numeric_limits<scalar_t>::infinity());
+ else
+ reduced_feats.fill_(static_cast<scalar_t>(0));
+
+ dim3 blocks(std::min(at::cuda::ATenCeilDiv(num_input, threadsPerBlock),
+ maxGridDim));
+ dim3 threads(threadsPerBlock);
+ feats_reduce_kernel<<<blocks, threads>>>(
+ feats.data_ptr<scalar_t>(), coors_map.data_ptr<int32_t>(),
+ reduced_feats.data_ptr<scalar_t>(), num_input, num_feats, reduce_type);
+ if (reduce_type == reduce_t::MEAN)
+ reduced_feats /= reduce_count.unsqueeze(-1).to(reduced_feats.dtype());
+ }));
+ AT_CUDA_CHECK(cudaGetLastError());
+
+ return {reduced_feats, out_coors, coors_map, reduce_count};
+}
+
+void dynamic_point_to_voxel_backward_gpu(at::Tensor &grad_feats,
+ const at::Tensor &grad_reduced_feats,
+ const at::Tensor &feats,
+ const at::Tensor &reduced_feats,
+ const at::Tensor &coors_map,
+ const at::Tensor &reduce_count,
+ const reduce_t reduce_type) {
+ CHECK_INPUT(grad_feats);
+ CHECK_INPUT(grad_reduced_feats);
+ CHECK_INPUT(feats);
+ CHECK_INPUT(reduced_feats);
+ CHECK_INPUT(coors_map);
+ CHECK_INPUT(reduce_count);
+
+ const int num_input = feats.size(0);
+ const int num_reduced = reduced_feats.size(0);
+ const int num_feats = feats.size(1);
+
+ grad_feats.fill_(0);
+ // copy voxel grad to points
+
+ if (num_input == 0 || num_reduced == 0) return;
+
+ if (reduce_type == reduce_t::MEAN || reduce_type == reduce_t::SUM) {
+ AT_DISPATCH_FLOATING_TYPES(
+ grad_reduced_feats.scalar_type(), "add_reduce_traceback_grad_kernel",
+ ([&] {
+ dim3 blocks(std::min(
+ at::cuda::ATenCeilDiv(num_input, threadsPerBlock), maxGridDim));
+ dim3 threads(threadsPerBlock);
+ add_reduce_traceback_grad_kernel<<<blocks, threads>>>(
+ grad_feats.data_ptr<scalar_t>(),
+ grad_reduced_feats.data_ptr<scalar_t>(),
+ coors_map.data_ptr<int32_t>(), reduce_count.data_ptr<int32_t>(),
+ num_input, num_feats, reduce_type);
+ }));
+ AT_CUDA_CHECK(cudaGetLastError());
+ } else {
+ auto reduce_from = at::full({num_reduced, num_feats}, num_input,
+ coors_map.options().dtype(torch::kInt32));
+ AT_DISPATCH_FLOATING_TYPES(
+ grad_reduced_feats.scalar_type(),
+ "max_reduce_traceback_scatter_idx_kernel", ([&] {
+ dim3 blocks(std::min(
+ at::cuda::ATenCeilDiv(num_input, threadsPerBlock), maxGridDim));
+ dim3 threads(threadsPerBlock);
+ max_reduce_traceback_scatter_idx_kernel<<<blocks, threads>>>(
+ feats.data_ptr<scalar_t>(), reduced_feats.data_ptr<scalar_t>(),
+ reduce_from.data_ptr<int32_t>(), coors_map.data_ptr<int32_t>(),
+ num_input, num_feats);
+ }));
+ AT_CUDA_CHECK(cudaGetLastError());
+
+ AT_DISPATCH_FLOATING_TYPES(
+ grad_reduced_feats.scalar_type(),
+ "max_reduce_traceback_scatter_idx_kernel", ([&] {
+ dim3 blocks(std::min(
+ at::cuda::ATenCeilDiv(num_reduced, threadsPerBlock), maxGridDim));
+ dim3 threads(threadsPerBlock);
+ max_reduce_scatter_grad_kernel<<<blocks, threads>>>(
+ grad_feats.data_ptr<scalar_t>(),
+ grad_reduced_feats.data_ptr<scalar_t>(),
+ reduce_from.data_ptr<int32_t>(), num_reduced, num_feats);
+ }));
+ AT_CUDA_CHECK(cudaGetLastError());
+ }
+ return;
+}
+
+} // namespace voxelization
diff --git a/mmdet3d/ops/voxel/src_backup/voxelization.cpp b/mmdet3d/ops/voxel/src_backup/voxelization.cpp
new file mode 100644
index 0000000..f83348e
--- /dev/null
+++ b/mmdet3d/ops/voxel/src_backup/voxelization.cpp
@@ -0,0 +1,13 @@
+#include <torch/extension.h>
+#include "voxelization.h"
+
+namespace voxelization {
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+ m.def("hard_voxelize", &hard_voxelize, "hard voxelize");
+ m.def("dynamic_voxelize", &dynamic_voxelize, "dynamic voxelization");
+ m.def("dynamic_point_to_voxel_forward", &dynamic_point_to_voxel_forward, "dynamic point to voxel forward");
+ m.def("dynamic_point_to_voxel_backward", &dynamic_point_to_voxel_backward, "dynamic point to voxel backward");
+}
+
+} // namespace voxelization
diff --git a/mmdet3d/ops/voxel/src_backup/voxelization.h b/mmdet3d/ops/voxel/src_backup/voxelization.h
new file mode 100644
index 0000000..765b30a
--- /dev/null
+++ b/mmdet3d/ops/voxel/src_backup/voxelization.h
@@ -0,0 +1,142 @@
+#pragma once
+#include <torch/extension.h>
+
+typedef enum { SUM = 0, MEAN = 1, MAX = 2 } reduce_t;
+
+namespace voxelization {
+
+int hard_voxelize_cpu(const at::Tensor &points, at::Tensor &voxels,
+ at::Tensor &coors, at::Tensor &num_points_per_voxel,
+ const std::vector<float> voxel_size,
+ const std::vector<float> coors_range,
+ const int max_points, const int max_voxels,
+ const int NDim = 3);
+
+void dynamic_voxelize_cpu(const at::Tensor &points, at::Tensor &coors,
+ const std::vector<float> voxel_size,
+ const std::vector<float> coors_range,
+ const int NDim = 3);
+
+std::vector<at::Tensor> dynamic_point_to_voxel_cpu(
+ const at::Tensor &points, const at::Tensor &voxel_mapping,
+ const std::vector<float> voxel_size, const std::vector<float> coors_range);
+
+#ifdef WITH_CUDA
+int hard_voxelize_gpu(const at::Tensor &points, at::Tensor &voxels,
+ at::Tensor &coors, at::Tensor &num_points_per_voxel,
+ const std::vector<float> voxel_size,
+ const std::vector<float> coors_range,
+ const int max_points, const int max_voxels,
+ const int NDim = 3);
+
+int nondisterministic_hard_voxelize_gpu(const at::Tensor &points, at::Tensor &voxels,
+ at::Tensor &coors, at::Tensor &num_points_per_voxel,
+ const std::vector<float> voxel_size,
+ const std::vector<float> coors_range,
+ const int max_points, const int max_voxels,
+ const int NDim = 3);
+
+void dynamic_voxelize_gpu(const at::Tensor &points, at::Tensor &coors,
+ const std::vector<float> voxel_size,
+ const std::vector<float> coors_range,
+ const int NDim = 3);
+
+std::vector<torch::Tensor> dynamic_point_to_voxel_forward_gpu(const torch::Tensor &feats,
+ const torch::Tensor &coors,
+ const reduce_t reduce_type);
+
+void dynamic_point_to_voxel_backward_gpu(torch::Tensor &grad_feats,
+ const torch::Tensor &grad_reduced_feats,
+ const torch::Tensor &feats,
+ const torch::Tensor &reduced_feats,
+ const torch::Tensor &coors_idx,
+ const torch::Tensor &reduce_count,
+ const reduce_t reduce_type);
+#endif
+
+// Interface for Python
+inline int hard_voxelize(const at::Tensor &points, at::Tensor &voxels,
+ at::Tensor &coors, at::Tensor &num_points_per_voxel,
+ const std::vector<float> voxel_size,
+ const std::vector<float> coors_range,
+ const int max_points, const int max_voxels,
+ const int NDim = 3, const bool deterministic = true) {
+ if (points.device().is_cuda()) {
+#ifdef WITH_CUDA
+ if (deterministic) {
+ return hard_voxelize_gpu(points, voxels, coors, num_points_per_voxel,
+ voxel_size, coors_range, max_points, max_voxels,
+ NDim);
+ }
+ return nondisterministic_hard_voxelize_gpu(points, voxels, coors, num_points_per_voxel,
+ voxel_size, coors_range, max_points, max_voxels,
+ NDim);
+#else
+ AT_ERROR("Not compiled with GPU support");
+#endif
+ }
+ return hard_voxelize_cpu(points, voxels, coors, num_points_per_voxel,
+ voxel_size, coors_range, max_points, max_voxels,
+ NDim);
+}
+
+inline void dynamic_voxelize(const at::Tensor &points, at::Tensor &coors,
+ const std::vector<float> voxel_size,
+ const std::vector<float> coors_range,
+ const int NDim = 3) {
+ if (points.device().is_cuda()) {
+#ifdef WITH_CUDA
+ return dynamic_voxelize_gpu(points, coors, voxel_size, coors_range, NDim);
+#else
+ AT_ERROR("Not compiled with GPU support");
+#endif
+ }
+ return dynamic_voxelize_cpu(points, coors, voxel_size, coors_range, NDim);
+}
+
+inline reduce_t convert_reduce_type(const std::string &reduce_type) {
+ if (reduce_type == "max")
+ return reduce_t::MAX;
+ else if (reduce_type == "sum")
+ return reduce_t::SUM;
+ else if (reduce_type == "mean")
+ return reduce_t::MEAN;
+ else TORCH_CHECK(false, "do not support reduce type " + reduce_type)
+ return reduce_t::SUM;
+}
+
+inline std::vector<torch::Tensor> dynamic_point_to_voxel_forward(const torch::Tensor &feats,
+ const torch::Tensor &coors,
+ const std::string &reduce_type) {
+ if (feats.device().is_cuda()) {
+#ifdef WITH_CUDA
+ return dynamic_point_to_voxel_forward_gpu(feats, coors, convert_reduce_type(reduce_type));
+#else
+ TORCH_CHECK(false, "Not compiled with GPU support");
+#endif
+ }
+ TORCH_CHECK(false, "do not support cpu yet");
+ return std::vector<torch::Tensor>();
+}
+
+inline void dynamic_point_to_voxel_backward(torch::Tensor &grad_feats,
+ const torch::Tensor &grad_reduced_feats,
+ const torch::Tensor &feats,
+ const torch::Tensor &reduced_feats,
+ const torch::Tensor &coors_idx,
+ const torch::Tensor &reduce_count,
+ const std::string &reduce_type) {
+ if (grad_feats.device().is_cuda()) {
+#ifdef WITH_CUDA
+ dynamic_point_to_voxel_backward_gpu(
+ grad_feats, grad_reduced_feats, feats, reduced_feats, coors_idx, reduce_count,
+ convert_reduce_type(reduce_type));
+ return;
+#else
+ TORCH_CHECK(false, "Not compiled with GPU support");
+#endif