-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_depth.py
928 lines (748 loc) · 45 KB
/
train_depth.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
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
"""
This script is used to train the online refinement model outside the full SLAM pipeline.
We used this to develop the online refinement module.
"""
import os
import cv2
import yaml
import json
import torch # Sorry to the google supervisors
import shutil
import numpy as np
from tqdm import tqdm
import torch.nn.functional as F
import matplotlib.pyplot as plt
from collections import OrderedDict
from tensorboardX import SummaryWriter
from torch.utils.data import DataLoader
from torchviz import make_dot, make_dot_from_trace
# Imports from our files
from loss.losses import *
from utils.training_utils import *
from utils.arguments import arguments
from depth_estimation.networks import *
from utils.modify_images import corrupt_rgbd
from slam.custom_slam import image_recover_slam
from utils.yaml_configs import load_yaml, save_yaml
from utils.advanced_vis import plotly_map_update_visualization
from depth_estimation.view_synthesis import BackprojectDepth, Project3D
# GradSLAM Imports
import gradslam as gs
from gradslam.datasets import ICL, TUM
from gradslam.slam import ICPSLAM
from gradslam.slam import PointFusion
from chamferdist import ChamferDistance
from gradslam import Pointclouds, RGBDImages
class Depth_Estimation:
def __init__(self, arguments):
# Initialization of required functions
self.args = arguments
self.device = torch.device("cuda" if self.args.SETTINGS.device == "cuda" else "cpu")
self.sequence_length = len(self.args.DATA.frames)
self.color_map = plt.cm.get_cmap("magma").reversed()
self.writer = SummaryWriter("tensorboard_outputs")
self.dataset_init()
self.model_init()
self.view_reconstruction_init()
self.losses_init()
if self.args.ABLATION.scale_intrinsics:
print("Scaling Intrinsics")
if self.args.ABLATION.scaled_depth:
print("Scaling Depth Maps")
def dataset_init(self):
"""
Initialize datasets in this function
"""
print("Loading Images of Size {} x {}".format(self.args.DATA.width, self.args.DATA.height))
self.data_path = os.path.join(self.args.DATA.data_path, self.args.DATA.name)
if self.args.DATA.name == "ICL":
self.dataset = ICL(basedir=self.data_path,
seqlen=self.sequence_length,
height=self.args.DATA.height,
width=self.args.DATA.width,
dilation=self.args.DATA.dilation,
stride=self.args.DATA.stride,
start=self.args.DATA.start)
elif self.args.DATA.name == "TUM":
self.dataset = TUM(basedir=self.data_path,
seqlen=self.sequence_length,
height=self.args.DATA.height,
width=self.args.DATA.width,
dilation=self.args.DATA.dilation,
stride=self.args.DATA.stride,
start=self.args.DATA.start)
self.train_loader = DataLoader(dataset=self.dataset,
batch_size=self.args.OPTIMIZATION.batch_size,
shuffle=False,
num_workers=self.args.SETTINGS.num_workers,
pin_memory=True,
drop_last=True)
print("{} Dataset Loaded".format(self.args.DATA.name))
def model_init(self):
"""
Define all training models in here
"""
self.models = {}
self.train_params = []
print("Initializing Models")
if self.args.MODEL.slam == "ICPSLAM":
self.models["SLAM"] = ICPSLAM(odom=self.args.MODEL.odom,
numiters=self.args.MODEL.numiters,
device=self.device)
self.models["GT_SLAM"] = ICPSLAM(odom="gt", device=self.device)
elif self.args.MODEL.slam == "PointFusion":
self.models["SLAM"] = PointFusion(odom=self.args.MODEL.odom,
dist_th=self.args.MODEL.dist_th,
angle_th= self.args.MODEL.angle_th,
sigma= self.args.MODEL.sigma,
numiters=self.args.MODEL.numiters,
device=self.device)
self.models["GT_SLAM"] = PointFusion(odom="gt", device=self.device)
print("Using the {} for SLAM".format(self.args.MODEL.slam))
"Resnet Encoder"
if self.args.MODEL.depth_network == "monodepth2":
self.models["depth_encoder"] = ResnetEncoder(self.args.MODEL.num_layers,
self.args.MODEL.weights_init_encoder == "imagenet")
self.models["depth_encoder"].to(self.device)
self.train_params += list(self.models["depth_encoder"].parameters())
"Depth Decoder"
self.models["depth_decoder"] = DepthDecoder(self.models["depth_encoder"].num_ch_enc,
self.args.DATA.scales)
self.models["depth_decoder"].to(self.device)
print("Loaded ResNet{} based depth network".format(self.args.MODEL.num_layers))
self.train_params += list(self.models["depth_decoder"].parameters())
if self.args.VIZ.tensorboard:
print("Requested tensorboard output.")
for module in self.models["depth_decoder"].decoder:
if type(module.conv) == Conv3x3:
print("Submodule: ", module.conv.conv)
module.conv.conv.register_backward_hook(self.hook_fn)
print("Hook registered.")
elif type(module.conv) == torch.nn.modules.conv.Conv2d:
print("Submodule: ", module.conv)
module.conv.register_backward_hook(self.hook_fn)
print("Hook registered.")
elif self.args.MODEL.depth_network == "indoor":
self.models["depth"] = DispResNet_Indoor(num_layers=self.args.MODEL.num_layers,
pretrained=self.args.MODEL.weights_init_encoder == "imagenet")
self.train_params += list(self.models["depth"].parameters())
self.models["depth"].to(self.device)
if self.args.VIZ.tensorboard:
print("Requested tensorboard output.")
for module in self.models["depth"].decoder.decoder:
if type(module.conv) == Conv3x3:
print("Submodule: ", module.conv.conv)
module.conv.conv.register_backward_hook(self.hook_fn)
print("Hook registered.")
elif type(module.conv) == torch.nn.modules.conv.Conv2d:
print("Submodule: ", module.conv)
module.conv.register_backward_hook(self.hook_fn)
print("Hook registered.")
print("Using indoor pretrained model.")
else:
raise ValueError("Given {} is not a valid depth network option".format(self.args.MODEL.depth_network))
if self.args.MODEL.use_pretrained_models and self.args.MODEL.depth_network == "monodepth2":
self.load_model()
elif self.args.MODEL.use_pretrained_models and self.args.MODEL.depth_network == "indoor":
self.load_model_indoor()
self.optimizer = define_optim(self.args, self.train_params)
self.schedular = define_schedular(self.args, self.optimizer)
if self.args.OPTIMIZATION.load_optimizer and self.args.MODEL.load_depth_path:
self.load_optimizer()
elif self.args.OPTIMIZATION.load_optimizer and not self.args.MODEL.load_depth_path:
raise ValueError("Load optimizer only if pretrained depth is used !! Set Flag off!")
def view_reconstruction_init(self):
"""
Initialize the novel view synthesis functions
"""
self.backproject_depth = BackprojectDepth(self.args.OPTIMIZATION.batch_size,
self.args.DATA.height,
self.args.DATA.width)
self.backproject_depth.to(self.device)
self.project_3d = Project3D(self.args.OPTIMIZATION.batch_size,
self.args.DATA.height,
self.args.DATA.width)
self.project_3d.to(self.device)
def losses_init(self):
"""
Initialize loss functions
"""
self.ssim = SSIM()
self.ssim = self.ssim.to(self.device)
self.chamfer = ChamferDistance()
self.chamfer = self.chamfer.to(self.device)
def set_refinement_mode(self):
"""Convert depth model to refinement mode:
batch norm is in eval mode + frozen params
"""
# try using frozen batch norm?
for m in self.models.values():
m.eval()
for name, param in m.named_parameters():
if name.find("bn") != -1:
param.requires_grad = False
def process_disparity(self, inputs, index):
"""
This function does an additional processing step on the disparity maps by aggregating the original and a flipped version
"""
left = inputs[("disp", index, 0)][:1]
right = torch.flip(inputs[("disp", index, 0)][1:], [3])
middle = 0.5 * (left + right)
h, w = left.shape[2], left.shape[3]
l_mesh, _ = torch.meshgrid(torch.linspace(0, 1, h), torch.linspace(0, 1, w))
l_mesh = l_mesh.to(self.device)
l_mask = (1.0 - torch.clip(20 * (l_mesh - 0.05), 0, 1)).unsqueeze(0).unsqueeze(0)
r_mask = torch.flip(l_mask, [3])
inputs[("disp", index , 0)] = r_mask * left + l_mask * right + (1.0 - l_mask - r_mask) * middle
return inputs
def train(self):
"""
Main Training Loop
"""
self.epoch = 0
self.step = 0
if self.args.MODEL.refinement_mode:
self.set_refinement_mode()
print("SLAM Reconstruction Started")
"Initialize a small sequence, "
for iter, batch in enumerate(self.train_loader):
colors, gt_depths, intrinsics, poses, transform = batch[0], batch[1], batch[2], batch[3], batch[4]
colors /= 255.0
colors, gt_depths, intrinsics, poses, transform = colors.to(self.device), \
gt_depths.to(self.device), \
intrinsics.to(self.device), \
poses.to(self.device), \
transform.to(self.device)
# Convert to an RGB-D object for SLAM
rgbd = RGBDImages(colors, gt_depths, intrinsics, poses)
# Obtain GT Reconstruction from SLAM
self.gt_reconstruction, _ = self.models["GT_SLAM"](rgbd)
self.gt_reconstruction = self.gt_reconstruction.detach()
scale = 0 # Only using 1 scale for now.
self.initial_depths = {}
# This is the main refinement loop
for refine_step in range(self.args.OPTIMIZATION.refinement_steps):
inputs = OrderedDict() # Used for the depth part
encoder_features = []
depth_tensor = []
# counters for naming of tensorboard outputs (reset for every refinement step)
if self.args.VIZ.tensorboard:
if self.args.MODEL.depth_network == "indoor":
self.hook_counter = 33
elif self.args.MODEL.depth_network == "monodepth2":
self.hook_counter = 10
self.refinement_step = refine_step
# Predict depth for a sequence of frames.
for index in range(self.sequence_length):
if self.args.MODEL.depth_network == "monodepth2":
# We no longer use Monodepth2 but it's still here incase we decide to do outdoor
encoder_features.append(self.models["depth_encoder"](colors[:, index, ...]))
inputs.update(self.models["depth_decoder"](encoder_features[index], index))
# Convert Disparity into Depth
inputs[("depth", index, scale)] = convert_disp_to_depth(inputs[("disp", index, scale)],
self.args.DATA.min_depth,
self.args.DATA.max_depth)
if self.args.ABLATION.scale_intrinsics:
# Intrinsics scaling was proposed by CNN-SLAM [Keisuke et al] but we no longer use it since
# we do median scaling.
focal_data = intrinsics[0, 0, 0, 0]
focal_pretrain = self.args.ABLATION.focal_pretrain
inputs[("depth", index, scale)] = scale_by_f(focal_data=focal_data,
focal_pretrain=focal_pretrain,
depth= inputs[("depth", index, scale)])
if self.args.ABLATION.scaled_depth:
inputs[("depth", index, scale)] *= self.args.ABLATION.scaling_depth
if self.args.OPTIMIZATION.refinement == "PFT" and self.args.LOSS.depth_regularizer and refine_step == 0:
self.initial_depths[("initial_depth", index, scale)] = inputs[("depth", index, scale)].copy().detach()
inputs[("gt_depth", index, scale)] = gt_depths[:, index, ...]
depth_tensor.append(inputs[("depth", index, scale)].unsqueeze(1)) # Unsqueeze to create Sequence Dimension
elif self.args.MODEL.depth_network == "indoor":
if self.args.ABLATION.dual_disparity:
# Additional processing step that aggregrates two disparity maps into a final one.
inputs.update(self.models["depth"](torch.cat([colors[:, index, ...],
torch.flip(colors[:, index, ...], [2])], 0),
index))
inputs.update(self.process_disparity(inputs, index))
else:
inputs.update(self.models["depth"](colors[:, index, ...], index))
# depth is the inverse of disparity
inputs[("depth", index, scale)] = 1 / inputs[("disp", index, scale)]
if self.args.ABLATION.scale_intrinsics:
# Intrinsics scaling was proposed by CNN-SLAM [Keisuke et al] but we no longer use it since
# we do median scaling.
focal_data = intrinsics[0, 0, 0, 0]
focal_pretrain = self.args.ABLATION.focal_pretrain
inputs[("depth", index, scale)] = scale_by_f(focal_data=focal_data,
focal_pretrain=focal_pretrain,
depth= inputs[("depth", index, scale)])
if self.args.ABLATION.scaled_depth:
# Median Scaling
inputs[("depth", index, scale)] *= self.args.ABLATION.scaling_depth
if self.args.OPTIMIZATION.refinement == "PFT" and self.args.LOSS.depth_regularizer and refine_step == 0:
# Store initial depth predictions for depth regularization
self.initial_depths[("initial_depth", index, scale)] = inputs[("depth", index, scale)].clone().detach()
depth_tensor.append(inputs[("depth", index, scale)].unsqueeze(1)) # Unsqueeze to create Sequence Dimension
inputs[("gt_depth", index, scale)] = gt_depths[:, index, ...]
del encoder_features # Free Space
depth_tensor = torch.cat(depth_tensor, dim=1) # use this for SLAM!
depth_tensor = depth_tensor.permute(0, 1, 3, 4, 2) # Change to channel last representation
if self.args.DATA.use_gt_pose:
new_poses = poses
else:
if self.step == 0: # Initially the pose is taken as identity
new_poses = torch.eye(4, device=self.device).view(1, 1, 4, 4).repeat(self.args.OPTIMIZATION.batch_size,
self.sequence_length, 1, 1)
# Create a noisy RGB-D item (contains depth predictions)
noisy_rgbd = RGBDImages(rgb_image=colors,
depth_image=depth_tensor,
intrinsics=intrinsics,
poses=new_poses)
# Pass noisy RGB-D through SLAM to obtain a 3D reconstruction and corresponding poses.
if self.args.DATA.use_gt_pose:
"""
We get rid of poses from SLAM if we using GT Pose, otherwise we will take the new pose from the SLAM
and feed it back into the network.
"""
noisy_reconstruction, _ = self.models["SLAM"](noisy_rgbd)
new_transform = transform
else:
noisy_reconstruction, new_poses = self.models["SLAM"](noisy_rgbd)
new_transform = torch_poses_to_transforms(new_poses)
noisy_pointcloud = noisy_reconstruction.points_list[0].unsqueeze(0).contiguous() # Only optimizing points not color!
inputs[("noisy_pointcloud")] = noisy_pointcloud
# Viz Funcs
if self.step == 0 and self.args.VIZ.plot_first_step:
plt.imshow(inputs[("depth", 1, 0)][0].detach().cpu().squeeze(), cmap=self.color_map)
plt.xlabel("Initial Predicted Depth")
plt.show()
noisy_reconstruction.plotly(0).show()
"--- Refinement of Depth Maps ---"
total_loss = self.depth_refinement(colors, inputs, intrinsics, new_transform)
new_poses = new_poses.detach()
self.step += 1
# Refinement Metrics and print progress
if self.args.DEBUG.print_metrics:
abs_rel, sq_rel, rmse, rmse_log, a1, a2, a3 = depth_metrics(dataset=self.args.DATA.name,
gt=gt_depths[0][1],
pred=inputs[("depth", 1, 0)][0])
print("Iter:", iter,
"Refine_Step:", refine_step,
"Total_Loss:", round(total_loss, 5),
"abs_rel: ", round(abs_rel.item(), 5),
"rmse: ", round(rmse.item(), 5),
"a1: ", round(a1.item(), 5))
else:
print("Iter:", iter,
"Refine_Step:", refine_step,
"Total_Loss:", round(total_loss, 5))
self.schedular.step()
# When we are only training on one sequence, we break off early.
if self.args.DEBUG.early_stop and iter == self.args.DEBUG.iter_stop:
break
# Final plot after training. [Maybe we should perform validation step on other scenes]
if self.args.VIZ.plot_final_step:
plt.imshow(inputs[("depth", 1, 0)][0].detach().cpu().squeeze(), cmap=self.color_map)
plt.xlabel("Final Refined Depth")
plt.show()
noisy_reconstruction.plotly(0).show()
if self.args.VIZ.plot_gt:
self.gt_reconstruction.plotly(0).show()
def depth_refinement(self, colors, inputs, intrinsics, poses):
"""
Essentially self-supervised depth estimation pipeline in a few refinement steps.
outputs only contain the outputs from the view synthesis module, NOT initial data.
"""
outputs = {}
# Process to make inputs compatiable with our self-supervised depth estimation pipeline
inputs.update(self.process_inputs(colors, inputs, intrinsics, poses))
outputs.update(self.novel_view_synthesis(inputs))
loss = self.compute_losses(inputs, outputs)
return loss
def process_inputs(self, colors, inputs, intrinsics, poses):
"""
Process the inputs suitable for view sythesis. Easier to understand for others then.
NOTE: This function assumes sequence length of 2 or 3. For larger sequence lengths change appropriately.
Making everything channel first representation too for PyTorch.
"""
if self.sequence_length == 3:
inputs["source_frame", -1] = colors[:, 0, ...].permute(0, 3, 1, 2)
inputs["target_frame"] = colors[:, 1, ...].permute(0, 3, 1, 2)
inputs["source_frame", 1] = colors[:, 2, ...].permute(0, 3, 1, 2)
if self.args.MODEL.depth_network == "monodepth2" and self.args.DATA.normalize_intrinsics:
# TODO: SLAM is giving errors if i do this normalization.
inputs["K"] = normalize_intrinsics(self.args, intrinsics[:, 0,...]) # In a particular sequence, the intrinsic should be constant.
inputs["Inverse_K"] = torch.pinverse(inputs["K"])
else:
inputs["K"] = intrinsics[:, 0, ...] # In a particular sequence, the intrinsic should be constant.
inputs["Inverse_K"] = torch.pinverse(inputs["K"])
inputs["source_depth", -1] = inputs["depth", 0, 0]
inputs["target_depth"] = inputs["depth", 1, 0]
inputs["source_depth", 1] = inputs["depth", 2, 0]
inputs["source_disp", -1] = inputs["disp", 0, 0]
inputs["target_disp"] = inputs["disp", 1, 0]
inputs["source_disp", 1] = inputs["disp", 2, 0]
# IF SEQUENCE LENGTH IS GREATER THAN 3, handle pose appropriately.
inputs["T", -1] = poses[:, 1, ...] # This represents the pose 0 -> -1
inputs["T", 1] = inverse_T_matrix(poses[:, 2, ...]) # This represents the pose 0 -> 1
# GT Depths
scale = 0
for index in range(len(self.args.DATA.frames)):
inputs[("sparse_gt_depth", index, scale)], inputs[("sparse_mask", index, scale)] = sparse_sampling(self.args.LOSS.sampling_type,
self.args.LOSS.sampling_prob,
inputs[("gt_depth", index, scale)])
elif self.sequence_length == 2:
if self.args.DATA.frames[1] < 0:
inputs["source_frame", -1] = colors[:, 0, ...].permute(0, 3, 1, 2)
inputs["target_frame"] = colors[:, 1, ...].permute(0, 3, 1, 2)
if self.args.MODEL.depth_network == "monodepth2" and self.args.DATA.normalize_intrinsics:
# TODO: SLAM is giving errors if i do this normalization.
inputs["K"] = normalize_intrinsics(self.args, intrinsics[:, 0,
...]) # In a particular sequence, the intrinsic should be constant.
inputs["Inverse_K"] = torch.pinverse(inputs["K"])
else:
inputs["K"] = intrinsics[:, 0, ...] # In a particular sequence, the intrinsic should be constant.
inputs["Inverse_K"] = torch.pinverse(inputs["K"])
inputs["source_depth", -1] = inputs["depth", 0, 0]
inputs["target_depth"] = inputs["depth", 1, 0]
inputs["source_disp", -1] = inputs["disp", 0, 0]
inputs["target_disp"] = inputs["disp", 1, 0]
inputs["T", -1] = poses[:, 1, ...] # This represents the pose 0 -> -1
scale = 0
for index in range(len(self.args.DATA.frames)):
inputs[("sparse_gt_depth", index, scale)], inputs[("sparse_mask", index, scale)] = sparse_sampling(
self.args.LOSS.sampling_type,
self.args.LOSS.sampling_prob,
inputs[("gt_depth", index, scale)])
elif self.args.DATA.frames[1] > 0:
inputs["target_frame"] = colors[:, 0, ...].permute(0, 3, 1, 2)
inputs["source_frame", 1] = colors[:, 1, ...].permute(0, 3, 1, 2)
if self.args.MODEL.depth_network == "monodepth2" and self.args.DATA.normalize_intrinsics:
# TODO: SLAM is giving errors if i do this normalization.
inputs["K"] = normalize_intrinsics(self.args, intrinsics[:, 0,
...]) # In a particular sequence, the intrinsic should be constant.
inputs["Inverse_K"] = torch.pinverse(inputs["K"])
else:
inputs["K"] = intrinsics[:, 0, ...] # In a particular sequence, the intrinsic should be constant.
inputs["Inverse_K"] = torch.pinverse(inputs["K"])
inputs["target_depth"] = inputs["depth", 0, 0]
inputs["source_depth", 1] = inputs["depth", 1, 0]
inputs["target_disp"] = inputs["disp", 0, 0]
inputs["source_disp", 1] = inputs["disp", 1, 0]
inputs["T", 1] = inverse_T_matrix(poses[:, 1, ...])
scale = 0
for index in range(len(self.args.DATA.frames)):
inputs[("sparse_gt_depth", index, scale)], inputs[("sparse_mask", index, scale)] = sparse_sampling(
self.args.LOSS.sampling_type,
self.args.LOSS.sampling_prob,
inputs[("gt_depth", index, scale)])
else:
raise ValueError("Sequence Length of 2 and 3 is only supported")
return inputs
def novel_view_synthesis(self, inputs):
outputs = {}
for frame in self.args.DATA.frames[1:]:
# project the points into 3D
camera_points = self.backproject_depth(inputs["target_depth"], inputs["Inverse_K"])
if self.args.DEBUG.plot and self.step % 10 == 0:
"Plots photometric error maps and saves them in the given location"
plt.imshow(inputs["target_depth"].detach().cpu().squeeze(), cmap=self.color_map)
plot_path = os.path.join(self.args.DEBUG.plot_path + "/depth_{}".format(self.step))
plt.axis('off')
plt.savefig(plot_path, bbox_inches="tight", pad_inches=0, dpi=1200)
if self.args.LOSS.geometric:
# we decided to not used geometric loss in the end so proceed to else:
pixel_coordinates, warped_depth, valid_mask = self.project_3d(points=camera_points,
K=inputs["K"],
T=inputs["T", frame],
geometric=True)
outputs[("warped_depth", frame)] = warped_depth
outputs[("valid_mask", frame)] = valid_mask
outputs[("synthesized_frame", frame)] = F.grid_sample(inputs["source_frame", frame],
pixel_coordinates,
padding_mode=self.args.MODEL.padding_mode,
align_corners=True)
outputs[("interpolated_depth", frame)] = F.grid_sample(inputs["source_depth", frame],
pixel_coordinates,
padding_mode=self.args.MODEL.padding_mode,
align_corners=False)
else:
# Create projection flow that we will use to sample on top off from the source frame
pixel_coordinates, valid_mask = self.project_3d(points=camera_points,
K=inputs["K"],
T=inputs["T", frame],
geometric=False)
# Photometric mask that contains only valid points
outputs[("valid_mask", frame)] = valid_mask
# Sample from source frame onto the projection flow
outputs[("synthesized_frame", frame)] = F.grid_sample(inputs["source_frame", frame],
pixel_coordinates,
padding_mode=self.args.MODEL.padding_mode,
align_corners=False)
if self.args.DEBUG.plot and self.step == 0:
"Plots photometric error maps and saves them in the given location"
plt.imshow(inputs["target_frame"].detach().cpu().squeeze().permute(1,2,0))
plot_path = os.path.join(self.args.DEBUG.plot_path + "/tF_{}".format(self.step))
plt.axis('off')
plt.savefig(plot_path, bbox_inches="tight", pad_inches=0, dpi=1200)
if self.args.DEBUG.plot and self.step == 0:
"Plots photometric error maps and saves them in the given location"
plt.imshow(inputs["source_frame", frame].detach().cpu().squeeze().permute(1,2,0))
plot_path = os.path.join(self.args.DEBUG.plot_path + "/sF_{}".format(self.step))
plt.axis('off')
plt.savefig(plot_path, bbox_inches="tight", pad_inches=0, dpi=1200)
if self.args.DEBUG.plot and self.step % 10 == 0:
"Plots photometric error maps and saves them in the given location"
plt.imshow(outputs["synthesized_frame", frame].detach().cpu().squeeze().permute(1,2,0))
plot_path = os.path.join(self.args.DEBUG.plot_path + "/synthF_{}".format(self.step))
plt.axis('off')
plt.savefig(plot_path, bbox_inches="tight", pad_inches=0, dpi=1200)
plt.close("all")
return outputs
def compute_losses(self, inputs, outputs):
losses = {}
loss = 0
self.optimizer.zero_grad()
photmetric = self.compute_photometric_loss(inputs=inputs,
outputs=outputs)
if self.args.LOSS.min_reprojection:
# we decided to not use min reprojection because it was designed to tackle occlusion issues and in
# indoor we don't face alot of those.
photmetric = photmetric
else:
photmetric = photmetric.mean(1, keepdim=True)
if self.args.DEBUG.plot and self.args.LOSS.min_reprojection and self.step % 10 == 0:
"Plots photometric error maps and saves them in the given location"
plt.imshow(torch.min(photmetric, dim=1)[0].detach().cpu().squeeze(), cmap="binary")
plot_path = os.path.join(self.args.DEBUG.plot_path + "/pE_{}".format(self.step))
plt.axis('off')
plt.savefig(plot_path, bbox_inches="tight", pad_inches=0, dpi=1200)
elif self.args.DEBUG.plot and self.step % 10 == 0:
plt.imshow(photmetric[0].detach().cpu().squeeze(), cmap="binary")
plot_path = os.path.join(self.args.DEBUG.plot_path + "/pE_{}".format(self.step))
plt.axis('off')
plt.savefig(plot_path, bbox_inches="tight", pad_inches=0, dpi=1200)
if self.args.LOSS.auto_masking:
# We decided to not use auto_masking technique because it was defined for outdoor scenes.
auto_masking = self.compute_automasking_loss(inputs=inputs,
outputs=outputs)
if self.args.LOSS.min_reprojection:
auto_masking += torch.randn(auto_masking.shape).cuda() * 0.00001 # Break tie's
else:
auto_masking = auto_masking.mean(1, keepdim=True)
photmetric = torch.cat((auto_masking, photmetric), dim=1)
if photmetric.shape[1] == 1:
optimize = photmetric
optimize = optimize.mean()
else:
optimize, indexs = torch.min(photmetric, dim=1)
optimize = optimize.mean()
loss += optimize
losses["photometric_loss"] = optimize.item()
if self.args.LOSS.geometric:
geometric = self.compute_geometric_loss(outputs=outputs)
geometric = geometric.mean()
loss += geometric * self.args.LOSS.geometric_weight
losses["geometric_loss"] = geometric.item()
if self.args.LOSS.smoothness:
smooth_loss = self.compute_smoothness_loss(inputs=inputs)
loss += smooth_loss * self.args.LOSS.smoothness_weight
losses["smoothn_loss"] = smooth_loss.item()
if self.args.LOSS.depth_regularizer:
depth_reg = self.compute_depth_regularizer(inputs=inputs)
loss += depth_reg * self.args.LOSS.depth_regularizer_weight
losses["depth regularizer"] = depth_reg.item()
if self.args.LOSS.knn_points:
knn_loss, indexs = knn_points_loss(gt_pointcloud=self.gt_reconstruction.points_list[0].unsqueeze(0).contiguous(),
noisy_pointcloud=inputs["noisy_pointcloud"])
loss += knn_loss * self.args.LOSS.knn_points_weight
print("knn_loss", knn_loss.item()) # Turn off unless debug
if self.args.LOSS.chamfer_distance:
chamfer_dist = 0.5 * self.chamfer(inputs["noisy_pointcloud"],
self.gt_reconstruction.points_list[0].unsqueeze(0).contiguous(),
bidirectional=True)
loss += chamfer_dist * self.args.LOSS.chamfer_weight
print("chamfer_loss", chamfer_dist.item()) # Turn off unless debug
if self.args.LOSS.supervise_depth:
gt_loss = self.compute_gt_depth_loss(inputs=inputs)
loss += gt_loss * self.args.LOSS.gt_depth_weight
losses["gt_depth_loss"] = gt_loss.item()
loss.backward()
self.optimizer.step()
return loss.item()
def compute_photometric_loss(self, inputs, outputs):
"""
Compute the photometric loss here
"""
photometric_losses = []
for frame in self.args.DATA.frames[1:]:
if self.args.LOSS.photometric_mask:
masked_prediction = outputs[("synthesized_frame", frame)] * outputs["valid_mask", frame]
masked_target = inputs[("target_frame")] * outputs["valid_mask", frame]
photometric_losses.append(photometric_loss(ssim=self.ssim,
prediction=masked_prediction,
target=masked_target))
else:
prediction = outputs[("synthesized_frame", frame)]
target = inputs[("target_frame")]
photometric_losses.append(photometric_loss(ssim=self.ssim,
prediction=prediction,
target=target))
photometric_losses = torch.cat(photometric_losses, 1)
return photometric_losses
def compute_automasking_loss(self, inputs, outputs):
"""
Compute identity losses for auto-masking technique by monodepth2
"""
auto_masking_losses = []
for frame in self.args.DATA.frames[1:]:
if self.args.LOSS.photometric_mask:
masked_prediction = inputs[("source_frame", frame)] * outputs["valid_mask", frame]
masked_target = inputs[("target_frame")] * outputs["valid_mask", frame]
auto_masking_losses.append(photometric_loss(ssim=self.ssim,
prediction=masked_prediction,
target=masked_target))
else:
prediction = inputs[("source_frame", frame)]
target = inputs[("target_frame")]
auto_masking_losses.append(photometric_loss(ssim=self.ssim,
prediction=prediction,
target=target))
auto_masking_losses = torch.cat(auto_masking_losses, 1)
return auto_masking_losses
def compute_geometric_loss(self, outputs):
"""
Compute the geometric loss here
"""
geometric_losses = []
for frame in self.args.DATA.frames[1:]:
geometric_losses.append(geometric_consistency_loss(outputs, frame, self.device))
geometric_losses = torch.stack(geometric_losses, dim=0)
return geometric_losses
def compute_smoothness_loss(self, inputs):
"""
compute the smoothness loss here on the disparities! Not depths.
"""
disparity = inputs[("disp", 0, 0)]
mean_disparity = disparity.mean(2, True).mean(3, True)
norm_disparity = disparity / (mean_disparity + 1e-7)
smooth_loss = disparity_smoothness_loss(disp=norm_disparity,
img=inputs[("target_frame")])
return smooth_loss
def compute_depth_regularizer(self, inputs):
"""
This function computes a regularizer such that the refined depth does not sway too far from the initial estimate of depth.
"""
scale = 0
reg = 0
for frame in range(len(self.args.DATA.frames)): # TODO: Improve convention of how initial_depth and predicted depth are stored maybe.
reg += depth_reguralizer(initial_depth=self.initial_depths[("initial_depth", frame, scale)],
refined_depth=inputs[("depth", frame, scale)],
loss_func=self.args.LOSS.depth_regularizer_type)
return reg
def compute_gt_depth_loss(self, inputs):
gt_loss = 0
scale = 0
for frame in range(len(self.args.DATA.frames)):
gt_loss += depth_gt_loss(prediction=inputs[("depth", frame, scale)],
sparse_groundtruth=inputs[("sparse_gt_depth", frame, scale)],
sparse_mask=inputs[("sparse_mask", frame, scale)])
return gt_loss
def load_model(self):
"""
Load pretrained models from disk.
Variables To Consider:
self.models: Contains List of All Trainable Models.
Flags To Consider:
MODEL.load_depth_path: Path to pretrained models and their optimizers for resume training
MODEL.pretrained_models_list: List of models to load
"""
self.args.MODEL.load_depth_path = os.path.expanduser(self.args.MODEL.load_depth_path)
assert os.path.isdir(self.args.MODEL.load_depth_path), "Cannot find folder {}".format(self.args.MODEL.load_depth_path)
print("loading model from folder {}".format(self.args.MODEL.load_depth_path))
for n in self.args.MODEL.pretrained_models_list:
print("Loading {} weights...".format(n))
path = os.path.join(self.args.MODEL.load_depth_path, "{}.pth".format(n))
model_dict = self.models[n].state_dict()
pretrained_dict = torch.load(path)
pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
model_dict.update(pretrained_dict)
self.models[n].load_state_dict(model_dict)
def load_model_indoor(self):
"""
Load pretrained models from disk.
Variables To Consider:
self.models: Contains List of All Trainable Models.
Flags To Consider:
MODEL.load_depth_path: Path to pretrained models and their optimizers for resume training
MODEL.pretrained_models_list: List of models to load
"""
self.args.MODEL.load_depth_path = os.path.expanduser(self.args.MODEL.load_depth_path)
assert os.path.isdir(self.args.MODEL.load_depth_path), "Cannot find folder {}".format(self.args.MODEL.load_depth_path)
print("loading model from folder {}".format(self.args.MODEL.load_depth_path))
n = "depth"
print("Loading {} weights...".format(n))
path = os.path.join(self.args.MODEL.load_depth_path, "{}.pth.tar".format(n))
pretrained_dict = torch.load(path)
self.models[n].load_state_dict(pretrained_dict["state_dict"])
print("Loaded Indoor Depth Model")
#TODO: Add Save_Model
def load_optimizer(self):
"""
Load Optimizer state dict if resuming training.
Flags To Consider:
MODEL.load_depth_path: Path to pretrained models and their optimizers for resume training
MODEL.pretrained_models_list: List of models to load
"""
optimizer_load_path = os.path.join(self.args.MODEL.load_depth_path, "{}.pth".format(self.args.OPTIMIZATION.optimizer))
if os.path.isfile(optimizer_load_path):
print("Loading Optimizer Weights")
optimizer_dict = torch.load(optimizer_load_path)
self.optimizer.load_state_dict(optimizer_dict)
else:
print("Optimizer Not Found. Randomly Initialized")
def hook_fn(self, layer, inputs, outputs):
"""INFO: Outputs of gradients w.r.t. to inputs and outputs are basically the same, therefore only one metric is
used."""
# tensorboard needs numpy array as input
o = outputs[0].clone().detach().cpu().numpy()
if self.args.VIZ.tensorboard_scaled:
# scale the convolution layers
scale_o = np.min(o) + np.max(o)
if scale_o == 0.0:
scale_o = 1.0
o /= scale_o
if self.args.MODEL.depth_network == "indoor":
decoder_idx = 20
image_idx = 33
if self.hook_counter > decoder_idx:
self.writer.add_histogram("Decoder_Layer_{}".format(self.hook_counter), o, self.refinement_step)
print("Added histogram for convolution layer {}.".format(self.hook_counter))
if self.hook_counter == image_idx:
self.writer.add_images("Image_Layer_{}_Step_{}_Outputs".format(self.hook_counter, self.refinement_step),
o, dataformats='NCHW')
print("Added images for convolution layer {}.".format(self.hook_counter))
self.hook_counter -= 1
elif self.hook_counter > 0:
self.writer.add_histogram("Encoder_Layer_{}".format(self.hook_counter), o, self.refinement_step)
print("Added histogram for convolution layer {}.".format(self.hook_counter))
self.hook_counter -= 1
else:
self.hook_counter = image_idx
elif self.args.MODEL.depth_network == "monodepth2":
image_idx = 10
self.writer.add_histogram("Decoder_Layer_{}".format(self.hook_counter), o, self.refinement_step)
print("Added histogram for convolution layer {}.".format(self.hook_counter))
if self.hook_counter == image_idx:
self.writer.add_images("Image_Layer_{}_Step_{}_Outputs".format(self.hook_counter, self.refinement_step), o, dataformats='NCHW')
print("Added images for convolution layer {}.".format(self.hook_counter))
if self.hook_counter > 0:
self.hook_counter -= 1
else:
self.hook_counter = image_idx
if __name__ == "__main__":
args = arguments()
config_path = args['config_path']
config_dict = load_yaml(config_path)
config_dict.SETTINGS.name = args['name']
SLAM = Depth_Estimation(config_dict)
SLAM.train()