-
Notifications
You must be signed in to change notification settings - Fork 1
/
utils.py
526 lines (441 loc) · 17.5 KB
/
utils.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
import os
from os.path import join as ospj
import sys
import time
import math
import shutil
from datetime import datetime
import random
import numpy as np
import torch
import torch.nn as nn
import imageio
import matplotlib.pyplot as plt
def get_data_dir():
dataroot=os.environ["MY_DATA_DIR"]
if len(dataroot)<1:
exit("CANNOT FIND ENV VARIABLE for 'MY_DATA_DIR':%s"%(dataroot))
else:
dataroot=os.path.join(dataroot, "nuscenes")
return dataroot
def get_exp_dir():
return "exps/"
def get_model_path(pretrained_path):
return ospj(get_exp_dir(), smart_path(pretrained_path))
def find_path(path):
return os.path.join(get_exp_dir(), path)
def find_npz_path(path):
if ".npz" not in path:
path = os.path.join(path, "cache.npz")
if path.startswith("/"):
path = path
else:
path = os.path.join(get_exp_dir(), path)
return path
def smart_path(s):
if ".ckpt" not in s:
s = s+"/models/model_last.ckpt"
return s
def to_np(x):
return x.detach().cpu().numpy()
def to_torch(x):
return torch.from_numpy(x).float().cuda()
def uniform_tensor(amin, amax, size):
return torch.rand(size) * (amax - amin) + amin
def rand_choice_tensor(choices, size):
return torch.from_numpy(np.random.choice(choices, size)).float()
def to_np_dict(di):
di_np = {}
for key in di:
di_np[key] = to_np(di[key])
return di_np
def dict_to_cuda(batch):
cuda_batch = {}
for key in batch:
cuda_batch[key] = batch[key]
if hasattr(batch[key], "device"):
cuda_batch[key] = cuda_batch[key].cuda()
return cuda_batch
def dict_to_torch(batch, keep_keys=[]):
torch_batch = {}
for key in batch:
if key in keep_keys:
torch_batch[key] = batch[key]
else:
torch_batch[key] = torch.from_numpy(batch[key])
return torch_batch
def save_model_freq_last(state_dict, model_dir, epi, save_freq, epochs):
if epi % save_freq == 0 or epi == epochs-1:
torch.save(state_dict, "%s/model_%05d.ckpt"%(model_dir, epi))
if epi % 10 == 0 or epi == epochs-1:
torch.save(state_dict, "%s/model_last.ckpt"%(model_dir))
def plt_save_close(img_path, bbox_inches='tight', pad_inches=0.1):
plt.savefig(img_path, bbox_inches=bbox_inches, pad_inches=pad_inches)
plt.close()
def build_relu_nn(input_dim, output_dim, hiddens, activation_fn=torch.nn.ReLU, last_fn=None):
n_neurons = [input_dim] + hiddens + [output_dim]
layers = []
for i in range(len(n_neurons)-1):
layers.append(nn.Linear(n_neurons[i], n_neurons[i+1]))
layers.append(activation_fn())
if last_fn is not None:
layers[-1] = last_fn()
else:
del layers[-1]
return nn.Sequential(*layers)
def build_relu_nn1(input_output_dim, hiddens, activation_fn, last_fn=None):
return build_relu_nn(input_output_dim[0], input_output_dim[1], hiddens, activation_fn, last_fn=last_fn)
def generate_gif(gif_path, duration, fs_list):
with imageio.get_writer(gif_path, mode='I', duration=duration) as writer:
for filename in fs_list:
image = imageio.imread(filename)
writer.append_data(image)
class MyTimer():
def __init__(self):
self.timestamp = {}
self.count = {}
self.profile = {}
self.left = {}
self.right = {}
self.last = None
def add(self, key, new_name=None):
self.timestamp[key] = time.time()
if key not in self.count:
self.count[key] = 0
self.count[key] += 1
if self.last is not None and self.count[key]==self.count[self.last]:
if new_name is None:
new_name = "%s-%s"%(key, self.last)
self.left[new_name] = key
self.right[new_name] = self.last
dt = self.timestamp[key] - self.timestamp[self.last]
if new_name not in self.profile:
self.profile[new_name] = 0
self.profile[new_name] += dt
self.last = key
def print_profile(self):
s=""
for key in self.profile:
left = self.left[key]
right = self.right[key]
tsum = self.profile[key]
cnt = self.count[left]
s += "%s:%.3f "%(key, tsum/ cnt)
print(s)
class EtaEstimator():
def __init__(self, start_iter, end_iter, check_freq=1, epochs=None, total_train_bs=None, total_val_bs=None, batch_size=None, viz_freq=None, num_workers=1):
self.start_iter = start_iter
num_workers = 1# if num_workers is None else num_workers
self.end_iter = end_iter//num_workers
self.check_freq = check_freq
self.curr_iter = start_iter
self.start_timer = None
self.interval = 0
self.eta_t = 0
self.num_workers = num_workers
self.viz_freq = viz_freq
self.epochs = epochs
self.batch_size = batch_size
self.prev_is_viz = False
self.prev_timer = time.time()
self.nn_stat_train = []
self.nn_train_bs = []
self.nn_stat_val = []
self.nn_val_bs = []
self.viz_stat = []
self.total_train_bs = total_train_bs
self.total_val_bs = total_val_bs
def update(self):
if self.start_timer is None:
self.start_timer = time.time()
self.curr_iter += 1
# if self.curr_iter % (max(1,self.check_freq//self.num_workers)) == 0:
self.interval = self.elapsed() / (self.curr_iter - self.start_iter)
self.eta_t = self.interval * (self.end_iter - self.curr_iter)
def update_viz_time(self, duration):
self.viz_stat.append(duration)
def smart_update(self, epi, duration=None, bs=None, mode=None, bi=None, is_viz=False):
self.curr_epi = epi
self.curr_stage = mode
self.curr_bi = bi
nn_time = duration
if mode=="train":
self.nn_stat_train.append(nn_time)
self.nn_train_bs.append(bs)
elif mode=="val":
self.nn_stat_val.append(nn_time)
self.nn_val_bs.append(bs)
else:
raise NotImplementedError
if len(self.nn_stat_train)>0:
if len(self.nn_stat_train)>1:
nn_train_sum = np.sum(self.nn_stat_train[1:])
nn_train_bs = np.sum(self.nn_train_bs[1:])
else:
nn_train_sum = np.sum(self.nn_stat_train)
nn_train_bs = np.sum(self.nn_train_bs)
nn_train_avg_per_sample = nn_train_sum / nn_train_bs
if len(self.nn_stat_val)>0:
nn_val_sum = np.sum(self.nn_stat_val)
nn_val_bs = np.sum(self.nn_val_bs)
nn_val_avg_per_sample = nn_val_sum / nn_val_bs
else:
nn_val_avg_per_sample = nn_train_avg_per_sample
remain_epis = self.epochs - self.curr_epi - 1
if self.curr_stage == "train":
remain_train_bs = self.total_train_bs - (self.curr_bi+1) * self.batch_size
remain_val_bs = self.total_val_bs
elif self.curr_stage == "val":
remain_train_bs = 0
remain_val_bs = self.total_val_bs - (self.curr_bi+1) * self.batch_size
remain_single_time = remain_train_bs * nn_train_avg_per_sample+ remain_val_bs * nn_val_avg_per_sample
if len(self.viz_stat)>0:
avg_viz_time = np.mean(self.viz_stat)
else:
avg_viz_time = 1 * nn_train_avg_per_sample
viz_cnt=0
for ii in range(self.curr_epi, self.epochs):
if ii % self.viz_freq == 0 or ii == self.epochs-1:
viz_cnt += 1
remain_viz_time = viz_cnt * avg_viz_time
self.eta_t_smart = remain_epis * (nn_train_avg_per_sample*self.total_train_bs + nn_val_avg_per_sample*self.total_val_bs) +\
remain_single_time + remain_viz_time
if is_viz==False:
self.update()
self.prev_is_viz = is_viz
self.prev_timer = time.time()
def elapsed(self):
return time.time() - self.start_timer
def eta(self):
return self.eta_t
def elapsed_str(self):
return time_format(self.elapsed())
def interval_str(self):
return time_format(self.interval)
def eta_str(self):
return time_format(self.eta_t)
def eta_str_smart(self):
return time_format(self.eta_t_smart)
def time_format(secs):
_s = secs % 60
_m = secs % 3600 // 60
_h = secs % 86400 // 3600
_d = secs // 86400
if _d != 0:
return "%02dD%02dh%02dm%02ds"%(_d, _h, _m, _s)
else:
if _h != 0:
return "%02dH%02dm%02ds"%(_h, _m, _s)
else:
if _m != 0:
return "%02dm%02ds"%(_m, _s)
else:
return "%05.2fs"%(_s)
def uniform(a, b, size):
return torch.rand(*size) * (b - a) + a
def linspace(a, b, size):
return torch.from_numpy(np.linspace(a, b, size)).float()
# TODO logger
class Logger(object):
def __init__(self):
self._terminal = sys.stdout
self._timestr = datetime.fromtimestamp(time.time()).strftime("%m%d-%H%M%S")
def create_log(self, log_path):
self.log = open(log_path + "/log-%s.txt" % self._timestr, "a", 1)
def write(self, message):
self._terminal.write(message)
self.log.write(message)
def flush(self):
pass
def write_cmd_to_file(log_dir, argv):
with open(ospj(log_dir, "cmd.txt"), "w") as f:
f.write("python " + " ".join(argv))
# TODO create the exp directory
def setup_exp_and_logger(args, set_gpus=True, test=False):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
sys.stdout = logger = Logger()
EXP_ROOT_DIR = get_exp_dir()
if test:
if hasattr(args, "rl") and args.rl:
tuples = args.rl_path.split("/")
else:
tuples = args.net_pretrained_path.split("/")
if ".ckpt" in tuples[-1] or ".zip" in tuples[-1] :
EXP_ROOT_DIR = ospj(EXP_ROOT_DIR, tuples[-3])
else:
EXP_ROOT_DIR = ospj(EXP_ROOT_DIR, tuples[-1])
if hasattr(args, "suffix") and args.suffix is not None:
suffix="_"+args.suffix
else:
suffix=""
args.exp_dir_full = os.path.join(EXP_ROOT_DIR, "test_%s%s" % (logger._timestr, suffix))
else:
if args.exp_name.startswith("e"):
args.exp_dir_full = os.path.join(EXP_ROOT_DIR, args.exp_name)
else:
args.exp_dir_full = os.path.join(EXP_ROOT_DIR, "g%s_%s" % (logger._timestr, args.exp_name))
args.viz_dir = os.path.join(args.exp_dir_full, "viz")
args.src_dir = os.path.join(args.exp_dir_full, "src")
args.model_dir = os.path.join(args.exp_dir_full, "models")
os.makedirs(args.viz_dir, exist_ok=True)
os.makedirs(args.src_dir, exist_ok=True)
os.makedirs(args.model_dir, exist_ok=True)
for fname in os.listdir('./'):
if fname.endswith('.py'):
shutil.copy(fname, os.path.join(args.src_dir, fname))
logger.create_log(args.exp_dir_full)
write_cmd_to_file(args.exp_dir_full, sys.argv)
np.savez(os.path.join(args.exp_dir_full, 'args'), args=args)
if set_gpus and hasattr(args, "gpus") and args.gpus is not None:
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpus
return args
# TODO metrics
class MeterDict:
def __init__(self):
self.d = {}
def reset(self):
del self.d
self.d = {}
def update(self, key, val):
if key not in self.d:
# curr, count, avg
self.d[key] = [val, 1, val]
else:
_, count, avg = self.d[key]
self.d[key][0] = val
self.d[key][1] = count+1
ratio = 1 / (count+1)
self.d[key][2] = avg * (1-ratio) + val * ratio
def get_val(self, key):
return self.d[key][0]
def __getitem__(self, key):
return self.get_val(key)
def get_avg(self, key):
return self.d[key][2]
def __contains__(self, key):
return key in self.d
def __call__(self, key):
return self.get_avg(key)
def compute_entropy(x, mask, n_bins=10, x_min=None, x_max=None):
# x (N, m)
# mask (N, m)
# return (N, )
assert len(x.shape)==len(mask.shape)==2
BIG_NUM = float("Inf")
SMALL_NUM = float("-Inf")
CLIP_VAL=1e-5
x_aug_min = x * 1.0
x_aug_min[mask==0] = SMALL_NUM
x_aug_max = x * 1.0
x_aug_max[mask==0] = BIG_NUM
if x_min is None:
xmin = torch.min(x_aug_max, dim=1)[0] - CLIP_VAL
xmax = torch.max(x_aug_min, dim=1)[0] + CLIP_VAL
else:
xmin = x_min * torch.ones_like(x[:,0])
xmax = x_max * torch.ones_like(x[:,0])
# gap = (xmax - xmin) / n_bins
alphas = torch.linspace(0.0, 1.0, n_bins+1)[None, :].to(x.device)
bins = xmin[:, None] * (1 - alphas) + xmax[:, None] * alphas # (N, 11)
# probs = torch.floor((x - xmin) / gap)
spotted = torch.logical_and(x_aug_max[:, :, None]>=bins[:, None, :-1], x_aug_max[:, :, None]<bins[:, None, 1:]) # (N, m, 10)
counts = torch.sum(spotted.float(), dim=1) # (N, m)
probs = counts / torch.clip(torch.sum(counts, dim=-1, keepdim=True), CLIP_VAL)
entropy = torch.sum(-probs * torch.log2(torch.clip(probs, CLIP_VAL)), dim=-1)
return entropy
def euler_from_quaternion(quat):
"""
Convert a quaternion into euler angles (roll, pitch, yaw)
roll is rotation around x in radians (counterclockwise)
pitch is rotation around y in radians (counterclockwise)
yaw is rotation around z in radians (counterclockwise)
"""
x, y, z, w = quat
t0 = +2.0 * (w * x + y * z)
t1 = +1.0 - 2.0 * (x * x + y * y)
roll_x = math.atan2(t0, t1)
t2 = +2.0 * (w * y - z * x)
t2 = +1.0 if t2 > +1.0 else t2
t2 = -1.0 if t2 < -1.0 else t2
pitch_y = math.asin(t2)
t3 = +2.0 * (w * z + x * y)
t4 = +1.0 - 2.0 * (y * y + z * z)
yaw_z = math.atan2(t3, t4)
return roll_x, pitch_y, yaw_z # in radians
def generate_bbox(x, y, theta, L, W):
# (2, 5)
bbox=np.array([
[L/2, W/2],
[L/2, -W/2],
[-L/2, -W/2],
[-L/2, W/2],
]).T
rot = np.array([
[np.cos(theta), -np.sin(theta)],
[np.sin(theta), np.cos(theta)],
])
# (2, 1)
trans = np.array([[
x, y
]]).T
new_bbox = (rot @ bbox) + trans
return new_bbox.T
def get_anchor_point(x, y, th, L, W, num_L, num_W):
x1 = L/2
y1 = W/2
x2 = -L/2
y2 = W/2
x3 = -x1
y3 = -y1
x4 = -x2
y4 = -y2
r_l = L / num_L / 2
r_w = W / num_W / 2
r = torch.minimum(torch.maximum(r_l, r_w), W / 2)
poly = torch.stack([x1, y1, x2, y2, x3, y3, x4, y4], dim=-1).reshape(list(x1.shape) + [4, 2])
poly_x = poly[..., 0] * torch.cos(th[..., None]) - poly[..., 1] * torch.sin(th[..., None]) + x[..., None]
poly_y = poly[..., 0] * torch.sin(th[..., None]) + poly[..., 1] * torch.cos(th[..., None]) + y[..., None]
poly = torch.stack([poly_x, poly_y], dim=-1)
alpha = torch.linspace(0, 1, num_L).to(x1.device)
beta = torch.linspace(0, 1, num_W).to(x1.device)
xs_ = (x2 + r)[..., None] * (1 - alpha) + (x1 - r)[..., None] * alpha # (N, T, k1)
ys_ = (y3 + r)[..., None] * (1 - beta) + (y2 - r)[..., None] * beta # (N, T, k2)
# xys_ = torch.stack(torch.meshgrid(xs_, ys_), dim=-1).reshape(list(xs.shape[:-1]) + [num_L*num_W, 2])
batch_size = list(x1.shape)
xs_ = xs_[..., None].expand(batch_size+ [num_L, num_W]).reshape(batch_size +[num_L*num_W])
ys_ = ys_[..., None, :].expand(batch_size+ [num_L, num_W]).reshape(batch_size +[num_L*num_W])
# print(xs_.shape, ys_.shape, th.shape, th[..., None].shape)
xs = xs_ * torch.cos(th[..., None]) - ys_ * torch.sin(th[..., None]) + x[..., None]
ys = xs_ * torch.sin(th[..., None]) + ys_ * torch.cos(th[..., None]) + y[..., None]
# xys = torch.stack(torch.meshgrid(xs, ys), dim=-1).reshape(list(xs.shape[:-1]) + [num_L*num_W, 2]) # (N, T, k1*k2, 2)
xys = torch.stack([xs, ys], dim=-1)
return poly, xys, r
def dist_between_two_cars(x1, y1, th1, L1, W1, x2, y2, th2, L2, W2, num_L, num_W, debug=False, full=False):
poly1, xys1, rs1 = get_anchor_point(x1, y1, th1, L1, W1, num_L, num_W) # (K, 1, n, k, 2), (K, 1, n, )
poly2, xys2, rs2 = get_anchor_point(x2, y2, th2, L2, W2, num_L, num_W) # (K, m, 1, k, 2), (K, m, 1, )
dist = torch.norm(xys1[..., None, :] - xys2[..., None, :, :], dim=-1)
dist = dist.reshape(list(dist.shape[:-2]) + [num_L * num_W * num_L * num_W])
# (K, 1, n, k, 1, 2) - (K, m, 1, 1, k, 2) -> (K, m, n, k, k) -> (K, m, n)
min_dist = torch.min(dist, dim=-1)[0]
car_dist = min_dist - rs1 - rs2
if full:
return car_dist, min_dist, rs1 + rs2
else:
return car_dist
def dist_between_two_cars_stack(state1, state2, num_L, num_W, debug=False, ego_L=None, ego_W=None, full=False):
if ego_L is not None:
assert 6>= state2.shape[-1] >= 5
return dist_between_two_cars(
state1[..., 0], state1[..., 1], state1[..., 2], ego_L * torch.ones_like(state1[..., 0]), ego_W * torch.ones_like(state1[..., 0]),
state2[..., 0], state2[..., 1], state2[..., 2], state2[..., -2], state2[..., -1],
num_L, num_W, debug, full)
else:
# print(state1.shape, state2.shape)
assert 6>= state1.shape[-1] >= 5
assert 6>= state2.shape[-1] >= 5
return dist_between_two_cars(
state1[..., 0], state1[..., 1], state1[..., 2], state1[..., -2], state1[..., -1],
state2[..., 0], state2[..., 1], state2[..., 2], state2[..., -2], state2[..., -1],
num_L, num_W, debug, full)