File size: 15,634 Bytes
1e45055 |
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 |
# Copyright (c) 2018-present, Facebook, Inc.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
#
from itertools import zip_longest
import numpy as np
class ChunkedGenerator_Seq:
"""
Batched data generator, used for training.
The sequences are split into equal-length chunks and padded as necessary.
Arguments:
batch_size -- the batch size to use for training
cameras -- list of cameras, one element for each video (optional, used for semi-supervised training)
poses_3d -- list of ground-truth 3D poses, one element for each video (optional, used for supervised training)
poses_2d -- list of input 2D keypoints, one element for each video
chunk_length -- number of output frames to predict for each training example (usually 1)
pad -- 2D input padding to compensate for valid convolutions, per side (depends on the receptive field)
causal_shift -- asymmetric padding offset when causal convolutions are used (usually 0 or "pad")
shuffle -- randomly shuffle the dataset before each epoch
random_seed -- initial seed to use for the random generator
augment -- augment the dataset by flipping poses horizontally
kps_left and kps_right -- list of left/right 2D keypoints if flipping is enabled
joints_left and joints_right -- list of left/right 3D joints if flipping is enabled
"""
def __init__(self, batch_size, cameras, poses_3d, poses_2d,
chunk_length, pad=0, causal_shift=0,
shuffle=True, random_seed=1234,
augment=False, kps_left=None, kps_right=None, joints_left=None, joints_right=None,
endless=False):
assert poses_3d is None or len(poses_3d) == len(poses_2d), (len(poses_3d), len(poses_2d))
assert cameras is None or len(cameras) == len(poses_2d)
# Build lineage info
pairs = [] # (seq_idx, start_frame, end_frame, flip) tuples
for key in poses_2d.keys():
assert poses_3d is None or poses_2d[key].shape[0] == poses_3d[key].shape[0]
n_chunks = (poses_2d[key].shape[0] + chunk_length - 1) // chunk_length
offset = (n_chunks * chunk_length - poses_2d[key].shape[0]) // 2
bounds = np.arange(n_chunks+1)*chunk_length - offset
augment_vector = np.full(len(bounds - 1), False, dtype=bool)
keys = np.tile(np.array(key).reshape([1, 3]), (len(bounds - 1), 1))
pairs += zip(keys, bounds[:-1], bounds[1:], augment_vector)
if augment:
pairs += zip(keys, bounds[:-1], bounds[1:], ~augment_vector)
# Initialize buffers
if cameras is not None:
self.batch_cam = np.empty((batch_size, cameras[0].shape[-1]))
if poses_3d is not None:
self.batch_3d = np.empty((batch_size, chunk_length, poses_3d[key].shape[-2], poses_3d[key].shape[-1]))
self.batch_2d = np.empty((batch_size, chunk_length, poses_2d[key].shape[-2], poses_2d[key].shape[-1]))
self.num_batches = (len(pairs) + batch_size - 1) // batch_size
self.batch_size = batch_size
self.random = np.random.RandomState(random_seed)
self.pairs = pairs
self.shuffle = shuffle
self.pad = pad
self.causal_shift = causal_shift
self.endless = endless
self.state = None
self.cameras = cameras
self.poses_3d = poses_3d
self.poses_2d = poses_2d
self.augment = augment
self.kps_left = kps_left
self.kps_right = kps_right
self.joints_left = joints_left
self.joints_right = joints_right
def num_frames(self):
return self.num_batches * self.batch_size
def batch_num(self):
return self.num_batches
def random_state(self):
return self.random
def set_random_state(self, random):
self.random = random
def augment_enabled(self):
return self.augment
def next_pairs(self):
if self.state is None:
if self.shuffle:
np.warnings.filterwarnings('ignore', category=np.VisibleDeprecationWarning)
pairs = self.random.permutation(self.pairs)
else:
pairs = self.pairs
return 0, pairs
else:
return self.state
def next_epoch(self):
enabled = True
while enabled:
start_idx, pairs = self.next_pairs()
for b_i in range(start_idx, self.num_batches):
chunks = pairs[b_i*self.batch_size : (b_i+1)*self.batch_size]
for i, (seq_i, start_3d, end_3d, flip) in enumerate(chunks):
subject, seq, cam_index = seq_i
seq_name = (subject, seq, cam_index)
# start_2d = start_3d - self.pad - self.causal_shift
start_2d = start_3d
# end_2d = end_3d + self.pad - self.causal_shift
end_2d = end_3d
# 2D poses
seq_2d = self.poses_2d[seq_name]
low_2d = max(start_2d, 0)
high_2d = min(end_2d, seq_2d.shape[0])
pad_left_2d = low_2d - start_2d
pad_right_2d = end_2d - high_2d
if pad_left_2d != 0 or pad_right_2d != 0:
self.batch_2d[i] = np.pad(seq_2d[low_2d:high_2d], ((pad_left_2d, pad_right_2d), (0, 0), (0, 0)), 'edge')
else:
self.batch_2d[i] = seq_2d[low_2d:high_2d]
if flip:
# Flip 2D keypoints
self.batch_2d[i, :, :, 0] *= -1
self.batch_2d[i, :, self.kps_left + self.kps_right] = self.batch_2d[i, :, self.kps_right + self.kps_left]
# 3D poses
if self.poses_3d is not None:
seq_3d = self.poses_3d[seq_name]
low_3d = max(start_3d, 0)
high_3d = min(end_3d, seq_3d.shape[0])
pad_left_3d = low_3d - start_3d
pad_right_3d = end_3d - high_3d
if pad_left_3d != 0 or pad_right_3d != 0:
self.batch_3d[i] = np.pad(seq_3d[low_3d:high_3d], ((pad_left_3d, pad_right_3d), (0, 0), (0, 0)), 'edge')
else:
self.batch_3d[i] = seq_3d[low_3d:high_3d]
if flip:
# Flip 3D joints
self.batch_3d[i, :, :, 0] *= -1
self.batch_3d[i, :, self.joints_left + self.joints_right] = \
self.batch_3d[i, :, self.joints_right + self.joints_left]
# Cameras
if self.cameras is not None:
self.batch_cam[i] = self.cameras[seq_name]
if flip:
# Flip horizontal distortion coefficients
self.batch_cam[i, 2] *= -1
self.batch_cam[i, 7] *= -1
if self.endless:
self.state = (b_i + 1, pairs)
if self.poses_3d is None and self.cameras is None:
yield None, None, self.batch_2d[:len(chunks)]
elif self.poses_3d is not None and self.cameras is None:
yield None, self.batch_3d[:len(chunks)], self.batch_2d[:len(chunks)]
elif self.poses_3d is None:
yield self.batch_cam[:len(chunks)], None, self.batch_2d[:len(chunks)]
else:
yield self.batch_cam[:len(chunks)], self.batch_3d[:len(chunks)], self.batch_2d[:len(chunks)]
if self.endless:
self.state = None
else:
enabled = False
class UnchunkedGenerator_Seq:
"""
Non-batched data generator, used for testing.
Sequences are returned one at a time (i.e. batch size = 1), without chunking.
If data augmentation is enabled, the batches contain two sequences (i.e. batch size = 2),
the second of which is a mirrored version of the first.
Arguments:
cameras -- list of cameras, one element for each video (optional, used for semi-supervised training)
poses_3d -- list of ground-truth 3D poses, one element for each video (optional, used for supervised training)
poses_2d -- list of input 2D keypoints, one element for each video
pad -- 2D input padding to compensate for valid convolutions, per side (depends on the receptive field)
causal_shift -- asymmetric padding offset when causal convolutions are used (usually 0 or "pad")
augment -- augment the dataset by flipping poses horizontally
kps_left and kps_right -- list of left/right 2D keypoints if flipping is enabled
joints_left and joints_right -- list of left/right 3D joints if flipping is enabled
"""
def __init__(self, cameras, poses_3d, poses_2d, pad=0, causal_shift=0,
augment=False, kps_left=None, kps_right=None, joints_left=None, joints_right=None, valid_frame=None):
assert poses_3d is None or len(poses_3d) == len(poses_2d)
assert cameras is None or len(cameras) == len(poses_2d)
self.augment = False
self.kps_left = kps_left
self.kps_right = kps_right
self.joints_left = joints_left
self.joints_right = joints_right
self.pad = pad
self.causal_shift = causal_shift
self.cameras = [] if cameras is None else cameras
self.poses_3d = [] if poses_3d is None else poses_3d
self.valid_frame = [] if valid_frame is None else valid_frame
self.poses_2d = poses_2d
def num_frames(self):
count = 0
for p in self.poses_2d:
count += self.poses_2d[p].shape[0]
return count
def batch_num(self):
return self.num_batches
def augment_enabled(self):
return self.augment
def set_augment(self, augment):
self.augment = augment
def next_epoch(self):
for (k_3d,seq_3d), (k_2d,seq_2d), (k_v,valid_f) in zip_longest(self.poses_3d.items(), self.poses_2d.items(), self.valid_frame.items()):
batch_cam = None
batch_3d = None if seq_3d is None else np.expand_dims(seq_3d, axis=0)
batch_2d = None if seq_2d is None else np.expand_dims(seq_2d, axis=0)
batch_valid = None if valid_f is None else valid_f
# batch_2d = np.expand_dims(np.pad(seq_2d,
# ((self.pad + self.causal_shift, self.pad - self.causal_shift), (0, 0), (0, 0)),
# 'edge'), axis=0)
if self.augment:
# Append flipped version
if batch_cam is not None:
batch_cam = np.concatenate((batch_cam, batch_cam), axis=0)
batch_cam[1, 2] *= -1
batch_cam[1, 7] *= -1
if batch_3d is not None:
batch_3d = np.concatenate((batch_3d, batch_3d), axis=0)
batch_3d[1, :, :, 0] *= -1
batch_3d[1, :, self.joints_left + self.joints_right] = batch_3d[1, :, self.joints_right + self.joints_left]
batch_2d = np.concatenate((batch_2d, batch_2d), axis=0)
batch_2d[1, :, :, 0] *= -1
batch_2d[1, :, self.kps_left + self.kps_right] = batch_2d[1, :, self.kps_right + self.kps_left]
# print(batch_2d.shape)
if batch_valid is None:
yield batch_cam, batch_3d, batch_2d
else:
yield batch_cam, batch_3d, batch_2d, batch_valid, k_3d
class UnchunkedGenerator_Seq2Seq:
"""
Non-batched data generator, used for testing.
Sequences are returned one at a time (i.e. batch size = 1), without chunking.
If data augmentation is enabled, the batches contain two sequences (i.e. batch size = 2),
the second of which is a mirrored version of the first.
Arguments:
cameras -- list of cameras, one element for each video (optional, used for semi-supervised training)
poses_3d -- list of ground-truth 3D poses, one element for each video (optional, used for supervised training)
poses_2d -- list of input 2D keypoints, one element for each video
pad -- 2D input padding to compensate for valid convolutions, per side (depends on the receptive field)
causal_shift -- asymmetric padding offset when causal convolutions are used (usually 0 or "pad")
augment -- augment the dataset by flipping poses horizontally
kps_left and kps_right -- list of left/right 2D keypoints if flipping is enabled
joints_left and joints_right -- list of left/right 3D joints if flipping is enabled
"""
def __init__(self, cameras, poses_3d, poses_2d, pad=0, causal_shift=0,
augment=False, kps_left=None, kps_right=None, joints_left=None, joints_right=None):
assert poses_3d is None or len(poses_3d) == len(poses_2d)
assert cameras is None or len(cameras) == len(poses_2d)
self.augment = False
self.kps_left = kps_left
self.kps_right = kps_right
self.joints_left = joints_left
self.joints_right = joints_right
self.pad = pad
self.causal_shift = causal_shift
self.cameras = [] if cameras is None else cameras
self.poses_3d = [] if poses_3d is None else poses_3d
self.poses_2d = poses_2d
def num_frames(self):
count = 0
for p in self.poses_2d:
count += p.shape[0]
return count
def augment_enabled(self):
return self.augment
def batch_num(self):
return self.num_batches
def set_augment(self, augment):
self.augment = augment
def next_epoch(self):
for seq_cam, seq_3d, seq_2d in zip_longest(self.cameras, self.poses_3d, self.poses_2d):
batch_cam = None if seq_cam is None else np.expand_dims(seq_cam, axis=0)
batch_3d = None if seq_3d is None else np.expand_dims(np.pad(seq_3d,
((self.pad + self.causal_shift, self.pad - self.causal_shift), (0, 0), (0, 0)),
'edge'), axis=0)
batch_2d = np.expand_dims(np.pad(seq_2d,
((self.pad + self.causal_shift, self.pad - self.causal_shift), (0, 0), (0, 0)),
'edge'), axis=0)
if self.augment:
# Append flipped version
if batch_cam is not None:
batch_cam = np.concatenate((batch_cam, batch_cam), axis=0)
batch_cam[1, 2] *= -1
batch_cam[1, 7] *= -1
if batch_3d is not None:
batch_3d = np.concatenate((batch_3d, batch_3d), axis=0)
batch_3d[1, :, :, 0] *= -1
batch_3d[1, :, self.joints_left + self.joints_right] = batch_3d[1, :, self.joints_right + self.joints_left]
batch_2d = np.concatenate((batch_2d, batch_2d), axis=0)
batch_2d[1, :, :, 0] *= -1
batch_2d[1, :, self.kps_left + self.kps_right] = batch_2d[1, :, self.kps_right + self.kps_left]
yield batch_cam, batch_3d, batch_2d
|