-
Notifications
You must be signed in to change notification settings - Fork 4
/
losses.py
739 lines (636 loc) · 28.8 KB
/
losses.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
# ------------------------------------------------------------------------
# Modification: EDA
# Created: 05/21/2022
# Author: Yanmin Wu
# E-mail: wuyanminmax@gmail.com
# https://github.com/yanmin-wu/EDA
# ------------------------------------------------------------------------
# BEAUTY DETR
# Copyright (c) 2022 Ayush Jain & Nikolaos Gkanatsios
# Licensed under CC-BY-NC [see LICENSE for details]
# All Rights Reserved
# ------------------------------------------------------------------------
# Parts adapted from Group-Free
# Copyright (c) 2021 Ze Liu. All Rights Reserved.
# Licensed under the MIT License.
# ------------------------------------------------------------------------
from scipy.optimize import linear_sum_assignment
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.distributed as dist
def is_dist_avail_and_initialized():
if not dist.is_available():
return False
if not dist.is_initialized():
return False
return True
def box_cxcyczwhd_to_xyzxyz(x):
x_c, y_c, z_c, w, h, d = x.unbind(-1)
w = torch.clamp(w, min=1e-6)
h = torch.clamp(h, min=1e-6)
d = torch.clamp(d, min=1e-6)
assert (w < 0).sum() == 0
assert (h < 0).sum() == 0
assert (d < 0).sum() == 0
b = [(x_c - 0.5 * w), (y_c - 0.5 * h), (z_c - 0.5 * d),
(x_c + 0.5 * w), (y_c + 0.5 * h), (z_c + 0.5 * d)]
return torch.stack(b, dim=-1)
def _volume_par(box):
return (
(box[:, 3] - box[:, 0])
* (box[:, 4] - box[:, 1])
* (box[:, 5] - box[:, 2])
)
def _intersect_par(box_a, box_b):
xA = torch.max(box_a[:, 0][:, None], box_b[:, 0][None, :])
yA = torch.max(box_a[:, 1][:, None], box_b[:, 1][None, :])
zA = torch.max(box_a[:, 2][:, None], box_b[:, 2][None, :])
xB = torch.min(box_a[:, 3][:, None], box_b[:, 3][None, :])
yB = torch.min(box_a[:, 4][:, None], box_b[:, 4][None, :])
zB = torch.min(box_a[:, 5][:, None], box_b[:, 5][None, :])
return (
torch.clamp(xB - xA, 0)
* torch.clamp(yB - yA, 0)
* torch.clamp(zB - zA, 0)
)
def _iou3d_par(box_a, box_b):
intersection = _intersect_par(box_a, box_b)
vol_a = _volume_par(box_a)
vol_b = _volume_par(box_b)
union = vol_a[:, None] + vol_b[None, :] - intersection
return intersection / union, union
# BRIEF 3DIoU loss
def generalized_box_iou3d(boxes1, boxes2):
"""
Generalized IoU from https://giou.stanford.edu/
The boxes should be in [x0, y0, x1, y1] format
Returns a [N, M] pairwise matrix, where N = len(boxes1)
and M = len(boxes2)
"""
# degenerate boxes gives inf / nan results
# so do an early check
assert (boxes1[:, 3:] >= boxes1[:, :3]).all()
assert (boxes2[:, 3:] >= boxes2[:, :3]).all()
iou, union = _iou3d_par(boxes1, boxes2)
lt = torch.min(boxes1[:, None, :3], boxes2[:, :3])
rb = torch.max(boxes1[:, None, 3:], boxes2[:, 3:])
wh = (rb - lt).clamp(min=0) # [N,M,3]
volume = wh[:, :, 0] * wh[:, :, 1] * wh[:, :, 2]
return iou - (volume - union) / volume
class SigmoidFocalClassificationLoss(nn.Module):
"""
Sigmoid focal cross entropy loss.
This class is taken from Group-Free code.
"""
def __init__(self, gamma=2.0, alpha=0.25):
"""
Args:
gamma: Weighting parameter for hard and easy examples.
alpha: Weighting parameter for positive and negative examples.
"""
super().__init__()
self.alpha = alpha
self.gamma = gamma
@staticmethod
def sigmoid_cross_entropy_with_logits(input, target):
"""
PyTorch Implementation for tf.nn.sigmoid_cross_entropy_with_logits:
max(x, 0) - x * z + log(1 + exp(-abs(x))) in
Args:
input: (B, #proposals, #classes) float tensor.
Predicted logits for each class
target: (B, #proposals, #classes) float tensor.
One-hot encoded classification targets
Returns:
loss: (B, #proposals, #classes) float tensor.
Sigmoid cross entropy loss without reduction
"""
loss = (
torch.clamp(input, min=0) - input * target
+ torch.log1p(torch.exp(-torch.abs(input)))
)
return loss
def forward(self, input, target, weights):
"""
Args:
input: (B, #proposals, #classes) float tensor.
Predicted logits for each class
target: (B, #proposals, #classes) float tensor.
One-hot encoded classification targets
weights: (B, #proposals) float tensor.
Anchor-wise weights.
Returns:
weighted_loss: (B, #proposals, #classes) float tensor
"""
pred_sigmoid = torch.sigmoid(input)
alpha_weight = target * self.alpha + (1 - target) * (1 - self.alpha)
pt = target * (1.0 - pred_sigmoid) + (1.0 - target) * pred_sigmoid
focal_weight = alpha_weight * torch.pow(pt, self.gamma)
bce_loss = self.sigmoid_cross_entropy_with_logits(input, target)
loss = focal_weight * bce_loss
loss = loss.squeeze(-1)
assert weights.shape.__len__() == loss.shape.__len__()
return loss * weights
def compute_points_obj_cls_loss_hard_topk(end_points, topk):
box_label_mask = end_points['box_label_mask']
seed_inds = end_points['seed_inds'].long() # B, K
seed_xyz = end_points['seed_xyz'] # B, K, 3
seeds_obj_cls_logits = end_points['seeds_obj_cls_logits'] # B, 1, K
gt_center = end_points['center_label'][:, :, :3] # B, G=132, 3
gt_size = end_points['size_gts'][:, :, :3] # B, G, 3
B = gt_center.shape[0] # batch size
K = seed_xyz.shape[1] # number if points from p++ output 1024
G = gt_center.shape[1] # number of gt boxes (with padding) 132
# Assign each point to a GT object
point_instance_label = end_points['point_instance_label'] # B, num_points=5000
obj_assignment = torch.gather(point_instance_label, 1, seed_inds) # B, K=1024
obj_assignment[obj_assignment < 0] = G - 1 # bg points to last gt
obj_assignment_one_hot = torch.zeros((B, K, G)).to(seed_xyz.device)
obj_assignment_one_hot.scatter_(2, obj_assignment.unsqueeze(-1), 1)
# Normalized distances of points and gt centroids
delta_xyz = seed_xyz.unsqueeze(2) - gt_center.unsqueeze(1) # (B, K, G, 3)
delta_xyz = delta_xyz / (gt_size.unsqueeze(1) + 1e-6) # (B, K, G, 3)
new_dist = torch.sum(delta_xyz ** 2, dim=-1)
euclidean_dist1 = torch.sqrt(new_dist + 1e-6) # BxKxG
euclidean_dist1 = (
euclidean_dist1 * obj_assignment_one_hot
+ 100 * (1 - obj_assignment_one_hot)
) # BxKxG
euclidean_dist1 = euclidean_dist1.transpose(1, 2).contiguous()
# Find the points that lie closest to each gt centroid
topk_inds = (
torch.topk(euclidean_dist1, topk, largest=False)[1]
* box_label_mask[:, :, None]
+ (box_label_mask[:, :, None] - 1)
) # BxGxtopk
topk_inds = topk_inds.long() # BxGxtopk
topk_inds = topk_inds.view(B, -1).contiguous() # B, Gxtopk
batch_inds = torch.arange(B)[:, None].repeat(1, G*topk).to(seed_xyz.device)
batch_topk_inds = torch.stack([
batch_inds,
topk_inds
], -1).view(-1, 2).contiguous()
# Topk points closest to each centroid are marked as true objects
objectness_label = torch.zeros((B, K + 1)).long().to(seed_xyz.device)
objectness_label[batch_topk_inds[:, 0], batch_topk_inds[:, 1]] = 1
objectness_label = objectness_label[:, :K]
objectness_label_mask = torch.gather(point_instance_label, 1, seed_inds)
objectness_label[objectness_label_mask < 0] = 0
# Compute objectness loss
criterion = SigmoidFocalClassificationLoss()
cls_weights = (objectness_label >= 0).float()
cls_normalizer = cls_weights.sum(dim=1, keepdim=True).float()
cls_weights /= torch.clamp(cls_normalizer, min=1.0)
cls_loss_src = criterion(
seeds_obj_cls_logits.view(B, K, 1),
objectness_label.unsqueeze(-1),
weights=cls_weights
)
objectness_loss = cls_loss_src.sum() / B
return objectness_loss
class HungarianMatcher(nn.Module):
"""
Assign targets to predictions.
This class is taken from MDETR and is modified for our purposes.
For efficiency reasons, the [targets don't include the no_object].
Because of this, in general, there are [more predictions than targets].
In this case, we do a 1-to-1 matching of the best predictions,
while the others are un-matched (and thus treated as non-objects).
"""
def __init__(self, cost_class=1, cost_bbox=5, cost_giou=2,
soft_token=False):
"""
Initialize matcher.
Args:
cost_class: relative weight of the classification error
cost_bbox: relative weight of the L1 bounding box regression error
cost_giou: relative weight of the giou loss of the bounding box
soft_token: whether to use soft-token prediction
"""
super().__init__()
self.cost_class = cost_class
self.cost_bbox = cost_bbox
self.cost_giou = cost_giou
assert cost_class != 0 or cost_bbox != 0 or cost_giou != 0
self.soft_token = soft_token
@torch.no_grad()
def forward(self, outputs, targets):
"""
Perform the matching.
Args:
outputs: This is a dict that contains at least these entries:
"pred_logits" (tensor): [batch_size, num_queries, num_classes]
"pred_boxes" (tensor): [batch_size, num_queries, 6], cxcyczwhd
targets: list (len(targets) = batch_size) of dict:
"labels" (tensor): [num_target_boxes]
(where num_target_boxes is the no. of ground-truth objects)
"boxes" (tensor): [num_target_boxes, 6], cxcyczwhd
"positive_map" (tensor): [num_target_boxes, 256]
Returns:
A list of size batch_size, containing tuples of (index_i, index_j):
- index_i is the indices of the selected predictions
- index_j is the indices of the corresponding selected targets
For each batch element, it holds:
len(index_i) = len(index_j) = min(num_queries, num_target_boxes)
"""
# Notation: {B: batch_size, Q: num_queries, C: num_classes}
bs, num_queries = outputs["pred_logits"].shape[:2] # Q: num_queries = 256
# We flatten to compute the cost matrices in a batch
out_prob = outputs["pred_logits"].flatten(0, 1).softmax(-1) # [B*Q, C=256]
out_bbox = outputs["pred_boxes"].flatten(0, 1) # [B*Q, 6]
# Also concat the target labels and boxes
positive_map = torch.cat([t["positive_map"] for t in targets]) # (B, 256)
tgt_ids = torch.cat([v["labels"] for v in targets]) # (B)
tgt_bbox = torch.cat([v["boxes"] for v in targets]) # (B, 6)
if self.soft_token:
# pad if necessary
if out_prob.shape[-1] != positive_map.shape[-1]:
positive_map = positive_map[..., :out_prob.shape[-1]]
cost_class = -torch.matmul(out_prob, positive_map.transpose(0, 1)) # (256, 1)
else:
# Compute the classification cost.
# Contrary to the loss, we don't use the NLL,
# but approximate it in 1 - proba[target class].
# The 1 is a constant that doesn't change the matching,
# it can be ommitted. DETR
# out_prob = out_prob * out_objectness.view(-1, 1)
cost_class = -out_prob[:, tgt_ids]
# Compute the L1 cost between boxes
cost_bbox = torch.cdist(out_bbox, tgt_bbox, p=1) # ([B*Q, 2])
# Compute the giou cost betwen boxes
cost_giou = -generalized_box_iou3d( # ([B*Q, 2])
box_cxcyczwhd_to_xyzxyz(out_bbox),
box_cxcyczwhd_to_xyzxyz(tgt_bbox)
)
# Final cost matrix
C = (
self.cost_bbox * cost_bbox # 0 *
+ self.cost_class * cost_class # 1 * ([B*Q, 2])
+ self.cost_giou * cost_giou # 2 * ([B*Q, 2])
).view(bs, num_queries, -1).cpu()
sizes = [len(v["boxes"]) for v in targets]
indices = [
linear_sum_assignment(c[i])
for i, c in enumerate(C.split(sizes, -1))
]
return [
(
torch.as_tensor(i, dtype=torch.int64), # matched pred boxes
torch.as_tensor(j, dtype=torch.int64) # corresponding gt boxes
)
for i, j in indices
]
# BRIEF Compute loss
class SetCriterion(nn.Module):
def __init__(self, matcher, losses={}, eos_coef=0.1, temperature=0.07):
"""
Parameters:
matcher: module that matches targets and proposals
losses: list of all the losses to be applied
eos_coef: weight of the no-object category
temperature: used to sharpen the contrastive logits
"""
super().__init__()
self.matcher = matcher
self.eos_coef = eos_coef # 0.1
self.losses = losses
self.temperature = temperature
#####################################
# BRIEF dense position-aligned loss #
#####################################
def loss_pos_align(self, outputs, targets, indices, num_boxes, auxi_indices):
logits = outputs["pred_logits"].log_softmax(-1)
# text position label
positive_map = torch.cat([t["positive_map"] for t in targets]) # main object
modify_positive_map = torch.cat([t["modify_positive_map"] for t in targets]) # attribute(modify)
pron_positive_map = torch.cat([t["pron_positive_map"] for t in targets]) # pron
other_entity_map = torch.cat([t["other_entity_map"] for t in targets]) # other(auxi)
rel_positive_map = torch.cat([t["rel_positive_map"] for t in targets]) # relation
# Trick to get target indices across batches
src_idx = self._get_src_permutation_idx(indices)
tgt_idx = []
offset = 0
for i, (_, tgt) in enumerate(indices):
tgt_idx.append(tgt + offset)
offset += len(targets[i]["boxes"])
tgt_idx = torch.cat(tgt_idx)
# NOTE constract the position label of the target object
tgt_pos = positive_map[tgt_idx]
mod_pos = modify_positive_map[tgt_idx]
pron_pos = pron_positive_map[tgt_idx]
other_pos = other_entity_map[tgt_idx]
rel_pos = rel_positive_map[tgt_idx]
# TODO ScanRefer & NR3D
tgt_weight_pos = tgt_pos * 0.6 + mod_pos * 0.2 + pron_pos * 0.2 + rel_pos*0.1
# TODO SR3D (5:1:1:1)/8 = 0.625: 0.125: 0.125: 0.125
if outputs["language_dataset"][0] == "sr3d":
tgt_weight_pos = tgt_pos * 0.625 + mod_pos * 0.125 + pron_pos * 0.125 + rel_pos * 0.125
# mask, keep the positive term
pos_mask = tgt_pos + mod_pos + pron_pos + rel_pos + other_pos
target_mask = torch.zeros_like(logits)
target_mask[:, :, -1] = 1
target_mask[src_idx] = pos_mask
target_sim = torch.zeros_like(logits)
target_sim[:, :, -1] = 1
target_sim[src_idx] = tgt_weight_pos
# STEP Compute entropy
entropy = torch.log(target_sim + 1e-6) * target_sim
loss_ce = (entropy - logits * target_sim).sum(-1)
# Weight less 'no_object'
eos_coef = torch.full(
loss_ce.shape, self.eos_coef,
device=target_sim.device
)
eos_coef[src_idx] = 1
loss_ce = loss_ce * eos_coef
loss_ce = loss_ce.sum() / num_boxes
losses = {"loss_ce": loss_ce}
return losses
# BRIEF object detection loss.
def loss_boxes(self, outputs, targets, indices, num_boxes, auxi_indices):
"""Compute bbox losses."""
assert 'pred_boxes' in outputs
idx = self._get_src_permutation_idx(indices)
src_boxes = outputs['pred_boxes'][idx]
target_boxes = torch.cat([
t['boxes'][i] for t, (_, i) in zip(targets, indices)
], dim=0)
loss_bbox = (
F.l1_loss(
src_boxes[..., :3], target_boxes[..., :3],
reduction='none'
)
+ 0.2 * F.l1_loss(
src_boxes[..., 3:], target_boxes[..., 3:],
reduction='none'
)
)
losses = {}
loss_giou = 1 - torch.diag(generalized_box_iou3d(
box_cxcyczwhd_to_xyzxyz(src_boxes),
box_cxcyczwhd_to_xyzxyz(target_boxes)))
losses['loss_bbox'] = loss_bbox.sum() / num_boxes
losses['loss_giou'] = loss_giou.sum() / num_boxes
return losses
############################
# BRIEF semantic alignment #
############################
def loss_sem_align(self, outputs, targets, indices, num_boxes, auxi_indices):
tokenized = outputs["tokenized"]
# step 1. Contrastive logits
norm_text_emb = outputs["proj_tokens"] # B, num_tokens=L, dim=64
norm_img_emb = outputs["proj_queries"] # B, num_queries=256, dim=64
logits = (
torch.matmul(norm_img_emb, norm_text_emb.transpose(-1, -2))
/ self.temperature
) # [[B, num_queries, num_tokens]
# step 2. positive map
# construct a map such that positive_map[k, i, j] = True
# iff query i is associated to token j in batch item k
positive_map = torch.zeros(logits.shape, device=logits.device) # ([B, 256, L])
# handle 'not mentioned'
inds = tokenized['attention_mask'].sum(1) - 1
positive_map[torch.arange(len(inds)), :, inds] = 0.5
positive_map[torch.arange(len(inds)), :, inds - 1] = 0.5
# handle true mentions
pmap = torch.cat([
t['positive_map'][i] for t, (_, i) in zip(targets, indices)
], dim=0)[..., :logits.shape[-1]]
idx = self._get_src_permutation_idx(indices)
positive_map[idx] = pmap
positive_map = positive_map > 0
modi_positive_map = torch.zeros(logits.shape, device=logits.device)
pron_positive_map = torch.zeros(logits.shape, device=logits.device)
other_positive_map = torch.zeros(logits.shape, device=logits.device)
rel_positive_map = torch.zeros(logits.shape, device=logits.device)
# [positive, 256] --> [positive, L]
pmap_modi = torch.cat([
t['modify_positive_map'][i] for t, (_, i) in zip(targets, indices)
], dim=0)[..., :logits.shape[-1]]
pmap_pron = torch.cat([
t['pron_positive_map'][i] for t, (_, i) in zip(targets, indices)
], dim=0)[..., :logits.shape[-1]]
pmap_other = torch.cat([
t['other_entity_map'][i] for t, (_, i) in zip(targets, indices)
], dim=0)[..., :logits.shape[-1]]
pmap_rel = torch.cat([
t['rel_positive_map'][i] for t, (_, i) in zip(targets, indices)
], dim=0)[..., :logits.shape[-1]]
modi_positive_map[idx] = pmap_modi
pron_positive_map[idx] = pmap_pron
other_positive_map[idx] = pmap_other
rel_positive_map[idx] = pmap_rel
# step object mask
# Mask for matches <> 'not mentioned'
mask = torch.full(
logits.shape[:2],
self.eos_coef,
dtype=torch.float32, device=logits.device
)
mask[idx] = 1.0
# step text mask
# Token mask for matches <> 'not mentioned'
tmask = torch.full(
(len(logits), logits.shape[-1]),
self.eos_coef,
dtype=torch.float32, device=logits.device
) # [B, L]
tmask[torch.arange(len(inds)), inds] = 1.0
# Positive logits are those who correspond to a match
positive_logits = -logits.masked_fill(~positive_map, 0)
negative_logits = logits
other_entity_neg_term = negative_logits.masked_fill(~(other_positive_map>0), 0)
modi_positive_logits = -logits.masked_fill(~(modi_positive_map>0), 0)
pron_positive_logits = -logits.masked_fill(~(pron_positive_map>0), 0)
rel_positive_logits = -logits.masked_fill(~(rel_positive_map>0), 0)
pos_modi_term = modi_positive_logits.sum(2)
pos_pron_term = pron_positive_logits.sum(2)
pos_rel_term = rel_positive_logits.sum(2)
# number of the token
nb_modi_pos_token = (modi_positive_map>0).sum(2) + 1e-6
nb_pron_pos_token = (pron_positive_map>0).sum(2) + 1e-6
nb_rel_pos_token = (rel_positive_map>0).sum(2) + 1e-6
###############################
# NOTE loss1: object --> text #
###############################
boxes_with_pos = positive_map.any(2)
pos_term = positive_logits.sum(2)
# note negative term
neg_term = (negative_logits+other_entity_neg_term).logsumexp(2)
nb_pos_token = positive_map.sum(2) + 1e-6
entropy = -torch.log(nb_pos_token+1e-6) / nb_pos_token
box_to_token_loss_ = (
pos_term/nb_pos_token \
+ 0.2*pos_modi_term/nb_modi_pos_token \
+ 0.2*pos_pron_term/nb_pron_pos_token \
+ 0.1*pos_rel_term/nb_rel_pos_token \
+ neg_term
).masked_fill(~boxes_with_pos, 0)
box_to_token_loss = (box_to_token_loss_ * mask).sum()
###############################
# NOTE loss2: text --> object #
###############################
tokens_with_pos = (positive_map + (modi_positive_map>0) + (pron_positive_map>0) + (rel_positive_map>0)).any(1)
tmask[positive_map.any(1)] = 1.0
tmask[(modi_positive_map>0).any(1)] = 0.2
tmask[(pron_positive_map>0).any(1)] = 0.2
tmask[(rel_positive_map>0).any(1)] = 0.1
tmask[torch.arange(len(inds)), inds-1] = 0.1
pos_term = positive_logits.sum(1)
pos_modi_term = modi_positive_logits.sum(1)
pos_pron_term = pron_positive_logits.sum(1)
pos_rel_term = rel_positive_logits.sum(1)
# note
pos_term = pos_term + pos_modi_term + pos_pron_term + pos_rel_term
neg_term = negative_logits.logsumexp(1)
nb_pos_obj = positive_map.sum(1) + modi_positive_map.sum(1) + pron_positive_map.sum(1) \
+ rel_positive_map.sum(1) + 1e-6
entropy = -torch.log(nb_pos_obj+1e-6) / nb_pos_obj
token_to_box_loss = (
(entropy + pos_term / nb_pos_obj + neg_term)
).masked_fill(~tokens_with_pos, 0)
token_to_box_loss = (token_to_box_loss * tmask).sum()
# total loss
tot_loss = (box_to_token_loss + token_to_box_loss) / 2
return {"loss_sem_align": tot_loss / num_boxes}
def _get_src_permutation_idx(self, indices):
# permute predictions following indices
batch_idx = torch.cat([
torch.full_like(src, i) for i, (src, _) in enumerate(indices)
])
src_idx = torch.cat([src for (src, _) in indices])
return batch_idx, src_idx
def _get_tgt_permutation_idx(self, indices):
# permute targets following indices
batch_idx = torch.cat([
torch.full_like(tgt, i) for i, (_, tgt) in enumerate(indices)
])
tgt_idx = torch.cat([tgt for (_, tgt) in indices])
return batch_idx, tgt_idx
# BRIEF get loss.
def get_loss(self, loss, outputs, targets, indices, num_boxes, auxi_indices, **kwargs):
loss_map = {
'boxes': self.loss_boxes, # box loss
'labels': self.loss_pos_align, # position alignment
'contrastive_align': self.loss_sem_align # semantic alignment
}
assert loss in loss_map, f'do you really want to compute {loss} loss?'
return loss_map[loss](outputs, targets, indices, num_boxes, auxi_indices, **kwargs)
def forward(self, outputs, targets):
"""
Perform the loss computation.
Parameters:
outputs: dict of tensors
targets: list of dicts, such that len(targets) == batch_size.
"""
# STEP Retrieve the matching between outputs and targets
indices = self.matcher(outputs, targets)
# auxi object
auxi_target = [
{
"labels": targets[b]["labels"],
"boxes": targets[b]["auxi_box"],
"positive_map": targets[b]["auxi_entity_positive_map"]
}
for b in range(outputs["pred_boxes"].shape[0])
]
auxi_indices = self.matcher(outputs, auxi_target)
num_boxes = sum(len(inds[1]) for inds in indices)
num_boxes = torch.as_tensor(
[num_boxes], dtype=torch.float,
device=next(iter(outputs.values())).device
)
if is_dist_avail_and_initialized():
torch.distributed.all_reduce(num_boxes)
# Compute all the requested losses
losses = {}
for loss in self.losses:
losses.update(self.get_loss(
loss, outputs, targets, indices, num_boxes, auxi_indices
))
return losses, indices
# BRIEF loss
def compute_hungarian_loss(end_points, num_decoder_layers, set_criterion,
query_points_obj_topk=5):
"""Compute Hungarian matching loss containing CE, bbox and giou."""
prefixes = ['last_'] + [f'{i}head_' for i in range(num_decoder_layers - 1)]
prefixes = ['proposal_'] + prefixes # 6+1: 'proposal_' 'last_' '0head_' '1head_' '2head_' '3head_' '4head_'
# STEP target GT box
gt_center = end_points['center_label'][:, :, 0:3]
gt_size = end_points['size_gts']
gt_labels = end_points['sem_cls_label']
gt_bbox = torch.cat([gt_center, gt_size], dim=-1)
# text
positive_map = end_points['positive_map'] # main obj.
modify_positive_map = end_points['modify_positive_map'] # attribute(modify)
pron_positive_map = end_points['pron_positive_map'] # pron
other_entity_map = end_points['other_entity_map'] # other(auxi)
rel_positive_map = end_points['rel_positive_map'] # relation
box_label_mask = end_points['box_label_mask'] # (132,) target object mask
auxi_entity_positive_map = end_points['auxi_entity_positive_map']
auxi_box = end_points['auxi_box']
target = [
{
"labels": gt_labels[b, box_label_mask[b].bool()],
"boxes": gt_bbox[b, box_label_mask[b].bool()],
"positive_map": positive_map[b, box_label_mask[b].bool()],
"modify_positive_map": modify_positive_map[b, box_label_mask[b].bool()],
"pron_positive_map": pron_positive_map[b, box_label_mask[b].bool()],
"other_entity_map": other_entity_map[b, box_label_mask[b].bool()],
"rel_positive_map": rel_positive_map[b, box_label_mask[b].bool()],
"auxi_entity_positive_map": auxi_entity_positive_map[b, 0].unsqueeze(0),
"auxi_box": auxi_box[b]
}
for b in range(gt_labels.shape[0])
]
loss_ce, loss_bbox, loss_giou, loss_sem_align = 0, 0, 0, 0
for prefix in prefixes:
output = {}
if 'proj_tokens' in end_points:
output['proj_tokens'] = end_points['proj_tokens']
output['proj_queries'] = end_points[f'{prefix}proj_queries']
output['tokenized'] = end_points['tokenized']
# STEP Get predicted boxes and labels
pred_center = end_points[f'{prefix}center'] # B, K, 3
pred_size = end_points[f'{prefix}pred_size'] # (B,K,3) (l,w,h)
pred_bbox = torch.cat([pred_center, pred_size], dim=-1)
pred_logits = end_points[f'{prefix}sem_cls_scores'] # (B, Q, n_class)
output['pred_logits'] = pred_logits
output["pred_boxes"] = pred_bbox
output["language_dataset"] = end_points["language_dataset"] # dataset
# NOTE Compute all the requested losses, forward
losses, _ = set_criterion(output, target)
for loss_key in losses.keys():
end_points[f'{prefix}_{loss_key}'] = losses[loss_key]
loss_ce += losses.get('loss_ce', 0)
loss_bbox += losses['loss_bbox']
loss_giou += losses.get('loss_giou', 0)
if 'proj_tokens' in end_points:
loss_sem_align += losses['loss_sem_align']
if 'seeds_obj_cls_logits' in end_points.keys():
query_points_generation_loss = compute_points_obj_cls_loss_hard_topk(
end_points, query_points_obj_topk
)
else:
query_points_generation_loss = 0.0
# total loss
weight = 1
if end_points["language_dataset"][0] == "scanrefer":
weight = 0.5
loss = (
8 * query_points_generation_loss
+ 1.0 / (num_decoder_layers + 1) * (
weight * loss_ce
+ 5 * loss_bbox
+ loss_giou
+ weight * loss_sem_align
)
)
end_points['loss_ce'] = loss_ce
end_points['loss_bbox'] = loss_bbox
end_points['loss_giou'] = loss_giou
end_points['query_points_generation_loss'] = query_points_generation_loss
end_points['loss_sem_align'] = loss_sem_align
end_points['loss'] = loss
return loss, end_points