-
Notifications
You must be signed in to change notification settings - Fork 47
/
utils.py
386 lines (306 loc) · 14.8 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
"""
DIP: training, evaluating and running of deep inertial poser.
Copyright (C) 2018 ETH Zurich, Emre Aksan, Manuel Kaufmann
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <http://www.gnu.org/licenses/>.
"""
import time
import os
from tensorflow.python.client import timeline
import numpy as np
import quaternion
import cv2
SMPL_MAJOR_JOINTS = [1, 2, 3, 4, 5, 6, 9, 12, 13, 14, 15, 16, 17, 18, 19]
SMPL_NR_JOINTS = 24
SMPL_PARENTS = [-1, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 9, 9, 12, 13, 14, 16, 17, 18, 19, 20, 21]
def get_model_dir_timestamp(prefix="", suffix="", connector="_"):
"""
Creates a directory name based on timestamp.
Args:
prefix:
suffix:
connector: one connector character between prefix, timestamp and suffix.
Returns:
"""
return prefix+connector+str(int(time.time()))+connector+suffix
def create_tf_timeline(model_dir, run_metadata):
"""
This is helpful for profiling slow Tensorflow code.
Args:
model_dir:
run_metadata:
Returns:
"""
tl = timeline.Timeline(run_metadata.step_stats)
ctf = tl.generate_chrome_trace_format()
timeline_file_path = os.path.join(model_dir,'timeline.json')
with open(timeline_file_path, 'w') as f:
f.write(ctf)
def split_data_dictionary(dictionary, split_indices, keys_frozen=[], verbose=1):
"""
Splits the data dictionary of lists into smaller chunks. All (key,value) pairs must have the same number of
elements. If there is an index error, then the corresponding (key, value) pair is copied directly to the new
dictionaries.
Args:
dictionary (dict): data dictionary.
split_indices (list): Each element contains a list of indices for one split. Multiple splits are supported.
keys_frozen (list): list of keys that are to be copied directly. The remaining (key,value) pairs will be
used in splitting. If not provided, all keys will be used.
verbose (int): status messages.
Returns:
(tuple): a tuple containing chunks of new dictionaries.
"""
# Find <key, value> pairs having an entry per sample.
sample_level_keys = []
dataset_level_keys = []
num_samples = sum([len(l) for l in split_indices])
for key, value in dictionary.items():
if not(key in keys_frozen) and ((isinstance(value, np.ndarray) or isinstance(value, list)) and (len(value) == num_samples)):
sample_level_keys.append(key)
else:
dataset_level_keys.append(key)
print(str(key) + " is copied.")
chunks = []
for chunk_indices in split_indices:
dict = {}
for key in dataset_level_keys:
dict[key] = dictionary[key]
for key in sample_level_keys:
dict[key] = [dictionary[key][i] for i in chunk_indices]
chunks.append(dict)
return tuple(chunks)
def aa_to_rot_matrix(data):
"""
Converts the orientation data to represent angle axis as rotation matrices. `data` is expected in format
(seq_length, n*3). Returns an array of shape (seq_length, n*9).
"""
# reshape to have sensor values explicit
data_c = np.array(data, copy=True)
seq_length, n = data_c.shape[0], data_c.shape[1] // 3
data_r = np.reshape(data_c, [seq_length, n, 3])
qs = quaternion.from_rotation_vector(data_r)
rot = np.reshape(quaternion.as_rotation_matrix(qs), [seq_length, n, 9])
return np.reshape(rot, [seq_length, 9*n])
def rot_matrix_to_aa(data):
"""
Converts the orientation data given in rotation matrices to angle axis representation. `data` is expected in format
(seq_length, n*9). Returns an array of shape (seq_length, n*3).
"""
seq_length, n_joints = data.shape[0], data.shape[1]//9
data_r = np.reshape(data, [seq_length, n_joints, 3, 3])
data_c = np.zeros([seq_length, n_joints, 3])
for i in range(seq_length):
for j in range(n_joints):
data_c[i, j] = np.ravel(cv2.Rodrigues(data_r[i, j])[0])
return np.reshape(data_c, [seq_length, n_joints*3])
def get_seq_len_histogram(sequence_length_array, num_bins=10, collapse_first_and_last_bins=None):
"""
Creates a histogram of sequence-length.
Args:
sequence_length_array: numpy array of sequence length for all samples.
num_bins:
collapse_first_and_last_bins: selects bin edges between the provided indices by discarding from the first and
last bins.
Returns:
(list): bin edges.
"""
collapse_first_and_last_bins = collapse_first_and_last_bins or [1, -1]
h, bins = np.histogram(sequence_length_array, bins=num_bins)
if collapse_first_and_last_bins is not None:
return [int(b) for b in bins[collapse_first_and_last_bins[0]:collapse_first_and_last_bins[1]]]
else:
return [int(b) for b in bins]
def rot_to_aa_representation(sample_list):
"""
Args:
sample_list (list): of motion samples with shape of (seq_len, SMPL_NR_JOINTS*9).
Returns:
"""
out = [np.reshape(s, [-1, SMPL_NR_JOINTS, 3, 3]) for s in sample_list]
aa = [quaternion.as_rotation_vector(quaternion.from_rotation_matrix(s)) for s in out]
aa = [np.reshape(s, (-1, 72)) for s in aa]
return aa
def rad2deg(v):
"""
Convert from radians to degrees.
"""
return v * 180.0 / np.pi
def padded_array_to_list(data, mask):
"""
Converts a padded numpy array to a list of un-padded numpy arrays. `data` is expected in shape
(n, max_seq_length, ...) and `mask` in shape (n, max_seq_length). The returned value is a list of size n, each
element being an np array of shape (dynamic_seq_length, ...).
"""
converted = []
seq_lengths = np.array(np.sum(mask, axis=1), dtype=np.int)
for i in range(data.shape[0]):
converted.append(data[i, 0:seq_lengths[i], ...])
return converted
def smpl_reduced_to_full(smpl_reduced):
"""
Converts an np array that uses the reduced smpl representation into the full representation by filling in
the identity rotation for the missing joints. Can handle either rotation input (dof = 9) or quaternion input
(dof = 4).
:param smpl_full: An np array of shape (seq_length, n_joints_reduced*dof)
:return: An np array of shape (seq_length, 24*dof)
"""
dof = smpl_reduced.shape[1] // len(SMPL_MAJOR_JOINTS)
assert dof == 9 or dof == 4
seq_length = smpl_reduced.shape[0]
smpl_full = np.zeros([seq_length, SMPL_NR_JOINTS * dof])
for idx in range(SMPL_NR_JOINTS):
if idx in SMPL_MAJOR_JOINTS:
red_idx = SMPL_MAJOR_JOINTS.index(idx)
smpl_full[:, idx * dof:(idx + 1) * dof] = smpl_reduced[:, red_idx * dof:(red_idx + 1) * dof]
else:
if dof == 9:
identity = np.repeat(np.eye(3, 3)[np.newaxis, ...], seq_length, axis=0)
else:
identity = np.concatenate([np.array([[1.0, 0.0, 0.0, 0.0]])] * seq_length, axis=0)
smpl_full[:, idx * dof:(idx + 1) * dof] = np.reshape(identity, [-1, dof])
return smpl_full
def smpl_rot_to_global(smpl_rotations_local):
"""
Converts local smpl rotations into global rotations by "unrolling" the kinematic chain.
:param smpl_rotations_local: np array of rotation matrices of shape (..., N, 3, 3), or (..., 216) where N
corresponds to the amount of joints in SMPL (currently 24)
:return: The global rotations as an np array of the same shape as the input.
"""
in_shape = smpl_rotations_local.shape
do_reshape = in_shape[-1] != 3
if do_reshape:
assert in_shape[-1] == 216
rots = np.reshape(smpl_rotations_local, in_shape[:-1] + (SMPL_NR_JOINTS, 3, 3))
else:
rots = smpl_rotations_local
out = np.zeros_like(rots)
dof = rots.shape[-3]
for j in range(dof):
if SMPL_PARENTS[j] < 0:
out[..., j, :, :] = rots[..., j, :, :]
else:
parent_rot = out[..., SMPL_PARENTS[j], :, :]
local_rot = rots[..., j, :, :]
out[..., j, :, :] = np.matmul(parent_rot, local_rot)
if do_reshape:
out = np.reshape(out, in_shape)
return out
def joint_angle_error(predicted_pose_params, target_pose_params):
"""
Computes the distance in joint angles between predicted and target joints for every given frame. Currently,
this function can only handle input pose parameters represented as rotation matrices.
:param predicted_pose_params: An np array of shape `(seq_length, dof)` where `dof` is 216, i.e. a 3-by-3 rotation
matrix for each of the 24 joints.
:param target_pose_params: An np array of the same shape as `predicted_pose_params` representing the target poses.
:return: An np array of shape `(seq_length, 24)` containing the joint angle error in Radians for each joint.
"""
seq_length, dof = predicted_pose_params.shape[0], predicted_pose_params.shape[1]
assert dof == 216, 'unexpected number of degrees of freedom'
assert target_pose_params.shape[0] == seq_length and target_pose_params.shape[1] == dof, 'target_pose_params must match predicted_pose_params'
# reshape to have rotation matrices explicit
n_joints = dof // 9
p1 = np.reshape(predicted_pose_params, [-1, n_joints, 3, 3])
p2 = np.reshape(target_pose_params, [-1, n_joints, 3, 3])
# compute R1 * R2.T, if prediction and target match, this will be the identity matrix
r1 = np.reshape(p1, [-1, 3, 3])
r2 = np.reshape(p2, [-1, 3, 3])
r2t = np.transpose(r2, [0, 2, 1])
r = np.matmul(r1, r2t)
# convert `r` to angle-axis representation and extract the angle, which is our measure of difference between
# the predicted and target orientations
angles = []
for i in range(r1.shape[0]):
aa, _ = cv2.Rodrigues(r[i])
angles.append(np.linalg.norm(aa))
return np.reshape(np.array(angles), [seq_length, n_joints])
def compute_metrics(prediction, target, compute_positional_error=False):
"""
Compute the metrics on the predictions. The function can handle variable sequence lengths for each pair of
prediction-target array. The pose parameters can either be represented as rotation matrices (dof = 9) or
quaternions (dof = 4)
:param prediction: a list of np arrays of size (seq_length, 24*dof)
:param target: a list of np arrays of size (seq_length, 24*dof)
:param compute_positional_error: if set, the euclidean pose error is calculated which can take some time.
"""
assert len(prediction) == len(target)
dof = prediction[0].shape[1] // SMPL_NR_JOINTS
assert dof == 9 or dof == 4
# because we are interested in difference per frame, flatten inputs
pred = np.concatenate(prediction, axis=0)
targ = np.concatenate(target, axis=0)
if dof == 4:
def to_rot(x):
seq_length = x.shape[0]
x_ = np.reshape(x, [seq_length, -1, dof])
x_ = quaternion.as_rotation_matrix(quaternion.as_quat_array(x_))
return np.reshape(x_, [seq_length, -1])
# convert quaternions to rotation matrices
pred = to_rot(pred)
targ = to_rot(targ)
pred_g = smpl_rot_to_global(pred)
targ_g = smpl_rot_to_global(targ)
angles = joint_angle_error(pred_g, targ_g)
mm = np.zeros([1, SMPL_NR_JOINTS]) # Ignore positional loss in Python 3.
return angles, mm
class Stats(object):
def __init__(self, tracking_joints=None, sip_evaluation_joints=None, evaluation_joints=None, logger=None):
self.tracking_joints = tracking_joints or list(range(SMPL_NR_JOINTS))
self.sip_evaluation_joints = sip_evaluation_joints or list(range(SMPL_NR_JOINTS))
self.evaluation_joints = evaluation_joints or list(range(SMPL_NR_JOINTS))
self.logger = logger
def __enter__(self):
self.joint_angle_diffs = None
self.euclidean_diffs = None
return self
def __exit__(self, exc_type, exc_value, traceback):
if exc_type is not None:
print(exc_type, exc_value, traceback)
return False
sip_stats = self.get_stats(self.sip_evaluation_joints)
tracking_stats = self.get_stats(self.tracking_joints)
eval_stats = self.get_stats(self.evaluation_joints)
self.logger.print('\n*** SIP Evaluation Error ***\n')
self.logger.print('average joint angle error (deg): {:.4f} (+/- {:.3f})\n'.format(sip_stats[0], sip_stats[1]))
self.logger.print('average positional error (m) : {:.4f} (+/- {:.3f})\n'.format(sip_stats[2], sip_stats[3]))
self.logger.print('\n*** Tracking Error ***\n')
self.logger.print('average joint angle error (deg): {:.4f} (+/- {:.3f})\n'.format(tracking_stats[0], tracking_stats[1]))
self.logger.print('average positional error (m) : {:.4f} (+/- {:.3f})\n'.format(tracking_stats[2], tracking_stats[3]))
self.logger.print('\n*** Remaining Evaluation Error ***\n')
self.logger.print('average joint angle error (deg): {:.4f} (+/- {:.3f})\n'.format(eval_stats[0], eval_stats[1]))
self.logger.print('average positional error (m) : {:.4f} (+/- {:.3f})\n'.format(eval_stats[2], eval_stats[3]))
return True
def get_stats(self, joint_idxs):
total_joint_angle_diff = np.mean(rad2deg(self.joint_angle_diffs[:, joint_idxs]))
std_per_joint = np.std(rad2deg(self.joint_angle_diffs[:, joint_idxs]), axis=0)
joint_angle_std = np.mean(std_per_joint)
total_mm_diff = np.mean(self.euclidean_diffs[:, joint_idxs])
std_per_joint = np.std(self.euclidean_diffs[:, joint_idxs], axis=0)
mm_std = np.mean(std_per_joint)
return total_joint_angle_diff, joint_angle_std, total_mm_diff, mm_std
def add(self, ja_diffs, euc_diffs):
if self.joint_angle_diffs is None:
self.joint_angle_diffs = np.zeros([0, ja_diffs.shape[-1]])
if self.euclidean_diffs is None:
self.euclidean_diffs = np.zeros([0, euc_diffs.shape[-1]])
self.joint_angle_diffs = np.concatenate([self.joint_angle_diffs, ja_diffs])
self.euclidean_diffs = np.concatenate([self.euclidean_diffs, euc_diffs])
def get_sip_stats(self):
return self.get_stats(self.sip_evaluation_joints)
class Logger:
def __init__(self, filename, stdout=None):
self.stdout = stdout
self.logfile = open(filename, 'w')
def print(self, text):
if self.stdout is not None:
self.stdout.write(text)
self.logfile.write(text)
def close(self):
self.logfile.close()