-
Notifications
You must be signed in to change notification settings - Fork 0
/
datamodule.py
648 lines (539 loc) · 21.3 KB
/
datamodule.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
import glob
import gzip
import os
import pickle
import cv2
# import matplotlib as mpl
# mpl.use('agg') # Must be before pyplot import to avoid memory leak
# import matplotlib.pyplot as plt
import numpy as np
import pytorch_lightning as pl
import torch
from torch.utils.data import DataLoader, Dataset
from world_model import WorldModel
def biternion_to_angle(x, y):
'''Converts biternion tensor representation to positive angle tensor.
Args:
x: Biternion 'x' component of shape (batch_n, n, n)
y: Biternion 'y' component of shape (batch_n, n, n)
'''
ang = torch.atan2(y, x)
# Add 360 deg to negative angle elements
mask = (ang < 0).float()
ang = ang + 2.0*np.pi*mask
return ang
def add_ang_to_dict(ang_dict, ang, i, j):
if (i, j) not in ang_dict.keys():
ang_dict[(i, j)] = []
ang_dict[(i, j)].append(ang)
return ang_dict
def get_angs_from_dict(ang_dict, i, j):
return ang_dict[(i, j)]
class BEVDataset(Dataset):
'''
Returns (input_tensors, label_tensors) torch.Tensor.
input_tensors[0] --> Road semantic 0|1
input_tensors[1] --> Road intensity (0,1)
label_tensors[0] --> Path label 0|1
label_tensors[1] --> Directional x label (-1,1)
label_tensors[2] --> Directional y label (-1,1)
'''
def __init__(self,
abs_root_path,
world_model,
do_rotation=False,
num_samples=8,
batch_size=8,
prob_use_pred=0.,
input_type='present',
get_gt_lanes=False,
rm_static_dyn_obj=False,
wm_conditioning='traj_full',
wm_temp=1.,
skip_wm=False):
'''
'''
self.abs_root_path = abs_root_path
self.sample_paths = glob.glob(
os.path.join(self.abs_root_path, '*', '*.pkl.gz'))
self.sample_paths = [
os.path.relpath(path, self.abs_root_path)
for path in self.sample_paths
]
self.sample_paths.sort()
self.num_samples = num_samples
self.batch_size = batch_size
self.world_model = world_model
self.wm_conditioning = wm_conditioning
self.wm_temp = wm_temp
self.skip_wm = skip_wm
self.do_rotation = do_rotation
self.temp = 1.
self.prob_use_pred = prob_use_pred
self.input_type = input_type
self.get_gt_lanes = get_gt_lanes
self.rm_static_dyn_obj = rm_static_dyn_obj
# Road marking intensity transformation
self.int_scaler = 20
self.int_sep_scaler = 20
self.int_mid_threshold = 0.5
# Representation size
self.res = 256
self.min_elements = 0.01 * 256 * 256
self.static_dyn_obj_threshold = 0.5 # Range (0,1)
def __len__(self):
return len(self.sample_paths)
def create_traj_label_set(self, trajs):
# 1. (x,y) --> angle
# 2. Add angle by img coord
#####################################
# Create set of trajectory labels
#####################################
traj_labels = []
dir_x_labels = []
dir_y_labels = []
for traj in trajs:
# traj: (N, 2) matrix with (i, j) coordinates
traj = traj[:, 0:2]
traj[:, 1] = 255 - traj[:, 1]
# Convert to point list
n = traj.shape[0]
traj = [(int(traj[idx, 0]), int(traj[idx, 1])) for idx in range(n)]
traj = self.remove_duplicate_pnts(traj)
# intensity = self.road_marking_transform(intensity)
traj_label = self.draw_trajectory(traj, 256, 256, traj_width=1)
dir_x_label, dir_y_label = self.draw_directional_trajectory(
traj, 256, 256, traj_width=2)
traj_labels.append(traj_label)
dir_x_labels.append(dir_x_label)
dir_y_labels.append(dir_y_label)
##################################################
# Merge labels to multimodal trajectory labels
##################################################
# Add all labels into one
traj_label_all = np.zeros((self.res, self.res))
for traj_label in traj_labels:
traj_label_all = np.logical_or(traj_label_all, traj_label)
# Create a dict with list of angles indexed by coord.
# ang_dict = {}
max_elems = 256 * 256
angs = -1 * np.ones((max_elems, 3))
ang_idx = 0
for idx in range(len(dir_x_labels)):
dir_x_label = dir_x_labels[idx]
dir_y_label = dir_y_labels[idx]
mask = np.abs(dir_x_label) + np.abs(dir_y_label) > 0
dir_ijs = np.argwhere(mask) # (N,2)
dir_ijs = dir_ijs.tolist() # [(i,j), ... ]
ang = biternion_to_angle(torch.tensor(dir_x_label),
torch.tensor(dir_y_label))
ang = ang.numpy()
for dir_ij in dir_ijs:
i = dir_ij[0]
j = dir_ij[1]
# ang_dict = add_ang_to_dict(ang_dict, ang[i, j].item(), i, j)
# angs.append((i, j, ang[i, j]))
angs[ang_idx, 0] = i
angs[ang_idx, 1] = j
angs[ang_idx, 2] = ang[i, j]
ang_idx += 1
if ang_idx > max_elems:
raise Exception('Directional elements over limit')
return traj_label_all, angs
def process_sample(self, sample: dict):
'''
Args:
sample:
Returns:
x: VDVAE posterior matching input tensor(6,256,256) in [0, 1]
dynamic: (256,256) in [0, 1]
traj_label_all: (256,256) in {0, 1}
ang_dict: key (i,j) --> value 'ang' in rad
'''
if self.input_type == 'full':
road = sample['road_full'].astype(np.float32)
intensity = sample['intensity_full'].astype(np.float32)
rgb = sample['rgb_full'].astype(np.float32)
dynamic = sample['dynamic_full'].astype(np.float32)
# trajs = sample['trajs_full']
elif self.input_type == 'present':
road = sample['road_present'].astype(np.float32)
intensity = sample['intensity_present'].astype(np.float32)
rgb = sample['rgb_present'].astype(np.float32)
dynamic = sample['dynamic_present'].astype(np.float32)
# trajs = sample['trajs_present']
elif self.input_type == 'future':
road = sample['road_future'].astype(np.float32)
intensity = sample['intensity_future'].astype(np.float32)
rgb = sample['rgb_future'].astype(np.float32)
dynamic = sample['dynamic_future'].astype(np.float32)
# trajs = sample['trajs_future']
else:
raise IOError(f'Undefined type ({self.input_type})')
trajs_present = sample['trajs_present']
trajs_full = sample['trajs_full']
gt_lanes = sample['gt_lanes']
traj_label_present, angs_present = self.create_traj_label_set(
trajs_present)
traj_label_full, angs_full = self.create_traj_label_set(trajs_full)
# Make non-road intensity 0.5 (i.e. unknown)
road_mask = road > 0.5
intensity[~road_mask] = 0.5
# Remove static dynamic objects from observations
if self.rm_static_dyn_obj:
mask = dynamic > self.static_dyn_obj_threshold
road[mask] = 1.
intensity[mask] = 0.5
mask_3ch = np.tile(np.expand_dims(mask, 0), (3, 1, 1))
rgb[mask_3ch] = np.zeros_like(rgb)[mask_3ch]
obs_mask = ~(road == 0.5)
# VDVAE posterior matching input tensor 'x' (6, H, W)
x = np.concatenate([
np.expand_dims(road, 0),
np.expand_dims(intensity, 0),
rgb,
np.expand_dims(obs_mask, 0),
])
x = torch.tensor(x)
dynamic = torch.tensor(dynamic)
traj_label_present = torch.tensor(traj_label_present)
ang_present = torch.tensor(angs_present)
traj_label_full = torch.tensor(traj_label_full)
ang_full = torch.tensor(angs_full)
sample = [
x, dynamic, traj_label_present, ang_present, traj_label_full,
ang_full
]
if self.get_gt_lanes:
gt_lanes_label, angs_gt_lanes = self.create_traj_label_set(
gt_lanes)
gt_lanes_label = torch.tensor(gt_lanes_label)
angs_gt_lanes = torch.tensor(angs_gt_lanes)
sample += [gt_lanes_label, angs_gt_lanes]
return sample
def __getitem__(self, idx):
while True:
sample_path = self.sample_paths[idx]
sample_path = os.path.join(self.abs_root_path, sample_path)
sample = self.read_compressed_pickle(sample_path)
num_obs_elem_present = np.sum(sample['road_present'] != 0.5)
num_obs_elem_future = np.sum(sample['road_future'] != 0.5)
num_obs_elem_full = np.sum(sample['road_full'] != 0.5)
if (num_obs_elem_present < self.min_elements
or num_obs_elem_future < self.min_elements
or num_obs_elem_full < self.min_elements):
idx = self.get_random_sample_idx()
else:
break
out = self.process_sample(sample)
x = out[0]
dynamic = out[1]
traj_label_present = out[2]
angs_present = out[3]
traj_label_full = out[4]
angs_full = out[5]
if self.get_gt_lanes:
gt_lanes_label = out[6]
angs_gt_lanes = out[7]
# Future observed trajectory conditioning
x_cond = x.clone()
if self.wm_conditioning == 'traj_full':
mask = traj_label_full == 1
elif self.wm_conditioning == 'gt_lanes':
mask = gt_lanes_label == 1
elif self.wm_conditioning is None:
mask = torch.zeros_like(traj_label_full)
else:
raise IOError('Undefined world model conditioning')
# TODO Make mask bigger by dillation (model may ignore conditioning)
kernel = np.ones((3, 3), np.uint8)
mask = cv2.dilate(mask.numpy().astype(float), kernel)
mask = torch.tensor(mask, dtype=torch.bool)
x_cond[0][mask] = 1
################
# Completion
################
if not self.skip_wm:
# Transform value range [0,1] --> [-1,1] (except obs mask)
x_cond[:5] = 2 * x_cond[:5] - 1
# TODO Add trajectory all as guidance for prediction
x_hat = self.world_model.sat_sample(x_cond,
traj_label_full,
self.batch_size,
temp=self.wm_temp)
else:
x_hat = x_cond
label = {
# NOTE Sample should be generated as dynamic yes|no
'dynamic': dynamic,
'traj_present': traj_label_present,
'angs_present': angs_present,
'traj_full': traj_label_full,
'angs_full': angs_full,
'scene_idx': sample['scene_idx'],
'map': sample['map'],
'ego_global_x': sample['ego_global_x'],
'ego_global_y': sample['ego_global_y'],
}
if self.get_gt_lanes:
label.update({
'gt_lanes': gt_lanes_label,
'gt_angs': angs_gt_lanes
})
return x, x_hat, label
def get_random_sample_idx(self):
return np.random.randint(0, self.num_samples)
def road_marking_transform(self, intensity_map):
'''
Args:
intensity_map: Value interval (0, 1)
'''
intensity_map = self.int_scaler * self.sigmoid(
self.int_sep_scaler * (intensity_map - self.int_mid_threshold))
# Normalization
intensity_map[intensity_map > 1.] = 1.
return intensity_map
def draw_trajectory(self,
traj: np.ndarray,
I: int,
J: int,
traj_width: int = 5):
'''
Args:
traj: (N,2) matrix with (i, j) coordinates
'''
label = np.zeros((I, J))
for idx in range(len(traj) - 1):
pnt_0 = traj[idx]
pnt_1 = traj[idx + 1]
# pnt_0 = poses[idx].astype(int)
# pnt_1 = poses[idx + 1].astype(int)
# pnt_0 = tuple(pnt_0)
# pnt_1 = tuple(pnt_1)
cv2.line(label, pnt_0, pnt_1, 1, traj_width)
return label
def draw_directional_trajectory(self, traj, I, J, traj_width=5):
'''
'''
circle_radius = int(np.ceil(0.5 * (traj_width + 1)))
label_x = np.zeros((I, J))
label_y = np.zeros((I, J))
for idx in range(len(traj) - 1):
pnt_0 = traj[idx]
pnt_1 = traj[idx + 1]
if pnt_0 == pnt_1:
continue
vec_x, vec_y = self.cal_norm_vector(pnt_0, pnt_1)
cv2.line(label_x, pnt_0, pnt_1, vec_x, traj_width)
cv2.line(label_y, pnt_0, pnt_1, vec_y, traj_width)
# Average mid-points
for idx in range(1, len(traj) - 1):
pnt_0 = traj[idx - 1]
pnt_1 = traj[idx]
pnt_2 = traj[idx + 1]
# Calculate the normalized average vector between two
vec_x_before, vec_y_before = self.cal_norm_vector(pnt_0, pnt_1)
vec_x_after, vec_y_after = self.cal_norm_vector(pnt_1, pnt_2)
vec_x_avg = (vec_x_before + vec_x_after)
vec_y_avg = (vec_y_before + vec_y_after)
vec_avg_len = np.sqrt(vec_x_avg**2 + vec_y_avg**2)
if vec_avg_len < 1e-9:
continue
vec_x_avg = vec_x_avg / vec_avg_len
vec_y_avg = vec_y_avg / vec_avg_len
cv2.circle(label_x, pnt_1, circle_radius, vec_x_avg, -1)
cv2.circle(label_y, pnt_1, circle_radius, vec_y_avg, -1)
return label_x, label_y
@staticmethod
def remove_duplicate_pnts(sequence):
seen = set()
return [x for x in sequence if not (x in seen or seen.add(x))]
@staticmethod
def cal_norm_vector(pnt_0, pnt_1):
dx = pnt_1[0] - pnt_0[0]
dy = pnt_1[1] - pnt_0[1]
length = np.sqrt(dx**2 + dy**2)
vec_x = dx / length
vec_y = -dy / length # NOTE Inverted y-axis!
if (length == 0):
print(pnt_0)
print(pnt_1)
print(dx)
print(dy)
print(length)
print(vec_x)
print(vec_y)
return vec_x, vec_y
@staticmethod
def sigmoid(z):
return 1 / (1 + np.exp(-z))
@staticmethod
def read_compressed_pickle(path):
try:
with gzip.open(path, "rb") as f:
pkl_obj = f.read()
obj = pickle.loads(pkl_obj)
return obj
except IOError as error:
print(error)
# class BEVDummyDataset(BEVDataset):
#
# def __init__(self):
# pass
class BEVDataModule(pl.LightningDataModule):
def __init__(self,
world_model,
train_data_dir: str = "./",
val_data_dir: str = "./",
batch_size: int = 128,
num_samples: int = 8,
num_workers: int = 0,
persistent_workers=False,
do_rotation: bool = False,
prob_use_pred: float = 0.,
input_type: str = 'present',
get_gt_lanes: bool = False,
rm_static_dyn_obj: bool = False,
wm_conditioning: str = 'traj_full',
wm_temp: float = 1.,
skip_wm: bool = False):
super().__init__()
self.train_data_dir = train_data_dir
self.val_data_dir = val_data_dir
self.batch_size = batch_size
self.num_samples = num_samples
self.num_workers = num_workers
self.persistent_workers = persistent_workers
self.bev_dataset_train = BEVDataset(
self.train_data_dir,
world_model,
do_rotation=do_rotation,
num_samples=num_samples,
batch_size=batch_size,
prob_use_pred=prob_use_pred,
input_type=input_type,
get_gt_lanes=get_gt_lanes,
rm_static_dyn_obj=rm_static_dyn_obj,
wm_conditioning=wm_conditioning,
wm_temp=wm_temp,
skip_wm=skip_wm)
self.bev_dataset_val = BEVDataset(self.val_data_dir,
world_model,
num_samples=num_samples,
batch_size=batch_size,
prob_use_pred=prob_use_pred,
input_type=input_type,
get_gt_lanes=get_gt_lanes,
rm_static_dyn_obj=rm_static_dyn_obj,
wm_conditioning=wm_conditioning,
wm_temp=wm_temp,
skip_wm=skip_wm)
def train_dataloader(self, shuffle=True):
return DataLoader(
self.bev_dataset_train,
batch_size=self.batch_size,
num_workers=self.num_workers,
persistent_workers=self.persistent_workers,
shuffle=shuffle,
)
def val_dataloader(self, shuffle=False):
return DataLoader(
self.bev_dataset_val,
batch_size=self.batch_size,
num_workers=self.num_workers,
persistent_workers=self.persistent_workers,
shuffle=shuffle,
)
def test_dataloader(self, shuffle=False):
return DataLoader(
self.bev_dataset_val,
batch_size=self.batch_size,
num_workers=self.num_workers,
persistent_workers=self.persistent_workers,
shuffle=shuffle,
)
def write_compressed_pickle(obj, filename, write_dir):
'''Converts an object into byte representation and writes a compressed file.
Args:
obj: Generic Python object.
filename: Name of file without file ending.
write_dir (str): Output path.
'''
path = os.path.join(write_dir, f"{filename}.gz")
pkl_obj = pickle.dumps(obj)
try:
with gzip.open(path, "wb") as f:
f.write(pkl_obj)
except IOError as error:
print(error)
if __name__ == '__main__':
'''
For visualizing dataset tensors.
NOTE Needs to be run with VDVAE input arguments.
'''
# import matplotlib.pyplot as plt
from viz.viz_dataset import viz_dataset_sample
# Load model as global variable to avoid duplicate loadings by 'train' and
# 'val' dataloader
world_model = WorldModel()
batch_size = 1
num_samples = 8
prob_use_pred = 0.
input_type = 'present'
get_gt_lanes = True
rm_static_dyn_obj = True
wm_conditioning = None # 'gt_lanes'
wm_temp = 0.6 # Better to keep temp < 1. ?
skip_wm = True
###################################
# Generate preprocessed dataset
###################################
bev = BEVDataModule(
world_model,
'/media/robin/Drive2/bev_nuscenes_256px_v01_boston_seaport_gt_huracan',
'/media/robin/Drive2/bev_nuscenes_256px_v01_boston_seaport_gt_huracan',
# '/home/robin/projects/pc-accumulation-lib/bev_kitti360_256px_aug_gt_3_rev',
# '/home/robin/projects/pc-accumulation-lib/bev_kitti360_256px_aug_gt_3_rev',
# '/home/robin/projects/vdvae/bev_kitti360_256px_aug_gt_3_rev_preproc_val_completed',
# '/home/robin/projects/vdvae/bev_kitti360_256px_aug_gt_3_rev_preproc_val_completed',
batch_size,
num_samples,
prob_use_pred=prob_use_pred,
input_type=input_type,
get_gt_lanes=get_gt_lanes,
rm_static_dyn_obj=rm_static_dyn_obj,
wm_conditioning=wm_conditioning,
wm_temp=wm_temp,
skip_wm=skip_wm)
dataloader = bev.train_dataloader(shuffle=False)
num_samples = len(bev.bev_dataset_train)
bev_idx = 0
subdir_idx = 0
savedir = '/media/robin/Drive2/bev_nuscenes_256px_v01_boston_seaport_gt_huracan_raw_preproc'
# Three process example
# num_processes = 3
# process_idx = 0|1|2
num_processes = 1
process_idx = 0
for idx, batch in enumerate(dataloader):
x, x_hat, label = batch
# Remove batch index (B,C,H,W) --> (C,H,W)
x = x[0]
x_hat = x_hat[0]
for key in label.keys():
label[key] = label[key][0]
if bev_idx >= 1000:
bev_idx = 0
subdir_idx += 1
output_path = f'{savedir}/subdir{subdir_idx:03d}'
if not os.path.isdir(output_path):
os.makedirs(output_path)
bev_idx_str = str(bev_idx).zfill(3)
filename = f'bev_{bev_idx_str}.pkl'
sample = (x_hat, label)
write_compressed_pickle(sample, filename, output_path)
filepath = os.path.join(output_path, f'viz_{bev_idx_str}.png')
viz_dataset_sample(x, x_hat, label, filepath, get_gt_lanes)
bev_idx += 1
# if idx % 100 == 0:
print(f'idx {idx} / {num_samples} ({idx/num_samples*100:.2f}%)')