File size: 16,392 Bytes
7ca9b42 |
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 |
import sys, os
thismodule = sys.modules[__name__]
from lib.util.motion import preprocess_mixamo, rotate_motion_3d, limb_scale_motion_2d, normalize_motion, get_change_of_basis, localize_motion, scale_limbs
import torch
import glob
import numpy as np
import random
from torch.utils.data import Dataset, DataLoader
from easydict import EasyDict as edict
from tqdm import tqdm
view_angles = np.array([ i * np.pi / 6 for i in range(-3, 4)])
def get_dataloader(phase, config):
config.data.batch_size = config.batch_size
config.data.seq_len = config.seq_len
dataset_cls_name = config.data.train_cls if phase == 'train' else config.data.eval_cls
dataset_cls = getattr(thismodule, dataset_cls_name)
dataset = dataset_cls(phase, config.data)
dataloader = DataLoader(dataset, shuffle=(phase=='train'),
batch_size=config.batch_size,
num_workers=(config.data.num_workers if phase == 'train' else 1),
worker_init_fn=lambda _: np.random.seed(),
drop_last=True)
return dataloader
class _MixamoDatasetBase(Dataset):
def __init__(self, phase, config):
super(_MixamoDatasetBase, self).__init__()
assert phase in ['train', 'test']
self.phase = phase
self.data_root = config.train_dir if phase=='train' else config.test_dir
self.meanpose_path = config.train_meanpose_path if phase=='train' else config.test_meanpose_path
self.stdpose_path = config.train_stdpose_path if phase=='train' else config.test_stdpose_path
self.unit = config.unit
self.aug = (phase == 'train')
self.character_names = sorted(os.listdir(self.data_root))
items = glob.glob(os.path.join(self.data_root, self.character_names[0], '*/motions/*.npy'))
self.motion_names = ['/'.join(x.split('/')[-3:]) for x in items]
self.meanpose, self.stdpose = get_meanpose(phase, config)
self.meanpose = self.meanpose.astype(np.float32)
self.stdpose = self.stdpose.astype(np.float32)
if 'preload' in config and config.preload:
self.preload()
self.cached = True
else:
self.cached = False
def build_item(self, mot_name, char_name):
"""
:param mot_name: animation_name/motions/xxx.npy
:param char_name: character_name
:return:
"""
return os.path.join(self.data_root, char_name, mot_name)
def load_item(self, item):
if self.cached:
data = self.cache[item]
else:
data = np.load(item)
return data
def preload(self):
print("pre-loading into memory")
pbar = tqdm(total=len(self))
self.cache = {}
for motion_name in self.motion_names:
for character_name in self.character_names:
item = self.build_item(motion_name, character_name)
motion3d = np.load(item)
self.cache[item] = motion3d
pbar.update(1)
@staticmethod
def gen_aug_params(rotate=False):
if rotate:
params = {'ratio': np.random.uniform(0.8, 1.2),
'roll': np.random.uniform((-np.pi / 9, -np.pi / 9, -np.pi / 6), (np.pi / 9, np.pi / 9, np.pi / 6))}
else:
params = {'ratio': np.random.uniform(0.5, 1.5)}
return edict(params)
@staticmethod
def augmentation(data, params=None):
"""
:param data: numpy array of size (joints, 3, len_frames)
:return:
"""
if params is None:
return data, params
# rotate
if 'roll' in params.keys():
cx, cy, cz = np.cos(params.roll)
sx, sy, sz = np.sin(params.roll)
mat33_x = np.array([
[1, 0, 0],
[0, cx, -sx],
[0, sx, cx]
], dtype='float')
mat33_y = np.array([
[cy, 0, sy],
[0, 1, 0],
[-sy, 0, cy]
], dtype='float')
mat33_z = np.array([
[cz, -sz, 0],
[sz, cz, 0],
[0, 0, 1]
], dtype='float')
data = mat33_x @ mat33_y @ mat33_z @ data
# scale
if 'ratio' in params.keys():
data = data * params.ratio
return data, params
def __getitem__(self, index):
raise NotImplementedError
def __len__(self):
return len(self.motion_names) * len(self.character_names)
def get_meanpose(phase, config):
meanpose_path = config.train_meanpose_path if phase == "train" else config.test_meanpose_path
stdpose_path = config.train_stdpose_path if phase == "train" else config.test_stdpose_path
if os.path.exists(meanpose_path) and os.path.exists(stdpose_path):
meanpose = np.load(meanpose_path)
stdpose = np.load(stdpose_path)
else:
meanpose, stdpose = gen_meanpose(phase, config)
np.save(meanpose_path, meanpose)
np.save(stdpose_path, stdpose)
print("meanpose saved at {}".format(meanpose_path))
print("stdpose saved at {}".format(stdpose_path))
if meanpose.shape[-1] == 2:
mean_x, mean_y = meanpose[:, 0], meanpose[:, 1]
meanpose = np.stack([mean_x, mean_x, mean_y], axis=1)
if stdpose.shape[-1] == 2:
std_x, std_y = stdpose[:, 0], stdpose[:, 1]
stdpose = np.stack([std_x, std_x, std_y], axis=1)
return meanpose, stdpose
def gen_meanpose(phase, config, n_samp=20000):
data_dir = config.train_dir if phase == "train" else config.test_dir
all_paths = glob.glob(os.path.join(data_dir, '*/*/motions/*.npy'))
random.shuffle(all_paths)
all_paths = all_paths[:n_samp]
all_joints = []
print("computing meanpose and stdpose")
for path in tqdm(all_paths):
motion = np.load(path)
if motion.shape[1] == 3:
basis = None
if sum(config.rotation_axes) > 0:
x_angles = view_angles if config.rotation_axes[0] else np.array([0])
z_angles = view_angles if config.rotation_axes[1] else np.array([0])
y_angles = view_angles if config.rotation_axes[2] else np.array([0])
x_angles, z_angles, y_angles = np.meshgrid(x_angles, z_angles, y_angles)
angles = np.stack([x_angles.flatten(), z_angles.flatten(), y_angles.flatten()], axis=1)
i = np.random.choice(len(angles))
basis = get_change_of_basis(motion, angles[i])
motion = preprocess_mixamo(motion)
motion = rotate_motion_3d(motion, basis)
motion = localize_motion(motion)
all_joints.append(motion)
else:
motion = preprocess_mixamo(motion)
motion = rotate_motion_3d(motion, basis)
motion = localize_motion(motion)
all_joints.append(motion)
else:
motion = motion * 128
motion_proj = localize_motion(motion)
all_joints.append(motion_proj)
all_joints = np.concatenate(all_joints, axis=2)
meanpose = np.mean(all_joints, axis=2)
stdpose = np.std(all_joints, axis=2)
stdpose[np.where(stdpose == 0)] = 1e-9
return meanpose, stdpose
class MixamoDataset(_MixamoDatasetBase):
def __init__(self, phase, config):
super(MixamoDataset, self).__init__(phase, config)
x_angles = view_angles if config.rotation_axes[0] else np.array([0])
z_angles = view_angles if config.rotation_axes[1] else np.array([0])
y_angles = view_angles if config.rotation_axes[2] else np.array([0])
x_angles, z_angles, y_angles = np.meshgrid(x_angles, z_angles, y_angles)
angles = np.stack([x_angles.flatten(), z_angles.flatten(), y_angles.flatten()], axis=1)
self.view_angles = angles
def preprocessing(self, motion3d, view_angle=None, params=None):
"""
:param item: filename built from self.build_tiem
:return:
"""
if self.aug: motion3d, params = self.augmentation(motion3d, params)
basis = None
if view_angle is not None: basis = get_change_of_basis(motion3d, view_angle)
motion3d = preprocess_mixamo(motion3d)
motion3d = rotate_motion_3d(motion3d, basis)
motion3d = localize_motion(motion3d)
motion3d = normalize_motion(motion3d, self.meanpose, self.stdpose)
motion2d = motion3d[:, [0, 2], :]
motion3d = motion3d.reshape([-1, motion3d.shape[-1]])
motion2d = motion2d.reshape([-1, motion2d.shape[-1]])
motion3d = torch.from_numpy(motion3d).float()
motion2d = torch.from_numpy(motion2d).float()
return motion3d, motion2d
def __getitem__(self, index):
# select two motions
idx_a, idx_b = np.random.choice(len(self.motion_names), size=2, replace=False)
mot_a, mot_b = self.motion_names[idx_a], self.motion_names[idx_b]
# select two characters
idx_a, idx_b = np.random.choice(len(self.character_names), size=2, replace=False)
char_a, char_b = self.character_names[idx_a], self.character_names[idx_b]
idx_a, idx_b = np.random.choice(len(self.view_angles), size=2, replace=False)
view_a, view_b = self.view_angles[idx_a], self.view_angles[idx_b]
if self.aug:
param_a = self.gen_aug_params(rotate=False)
param_b = self.gen_aug_params(rotate=False)
else:
param_a = param_b = None
item_a = self.load_item(self.build_item(mot_a, char_a))
item_b = self.load_item(self.build_item(mot_b, char_b))
item_ab = self.load_item(self.build_item(mot_a, char_b))
item_ba = self.load_item(self.build_item(mot_b, char_a))
X_a, x_a = self.preprocessing(item_a, view_a, param_a)
X_b, x_b = self.preprocessing(item_b, view_b, param_b)
X_aab, x_aab = self.preprocessing(item_a, view_b, param_a)
X_bba, x_bba = self.preprocessing(item_b, view_a, param_b)
X_aba, x_aba = self.preprocessing(item_ab, view_a, param_b)
X_bab, x_bab = self.preprocessing(item_ba, view_b, param_a)
X_abb, x_abb = self.preprocessing(item_ab, view_b, param_b)
X_baa, x_baa = self.preprocessing(item_ba, view_a, param_a)
return {"X_a": X_a, "X_b": X_b,
"X_aab": X_aab, "X_bba": X_bba,
"X_aba": X_aba, "X_bab": X_bab,
"X_abb": X_abb, "X_baa": X_baa,
"x_a": x_a, "x_b": x_b,
"x_aab": x_aab, "x_bba": x_bba,
"x_aba": x_aba, "x_bab": x_bab,
"x_abb": x_abb, "x_baa": x_baa,
"mot_a": mot_a, "mot_b": mot_b,
"char_a": char_a, "char_b": char_b,
"view_a": view_a, "view_b": view_b,
"meanpose": self.meanpose, "stdpose": self.stdpose}
class MixamoLimbScaleDataset(_MixamoDatasetBase):
def __init__(self, phase, config):
super(MixamoLimbScaleDataset, self).__init__(phase, config)
self.global_range = config.global_range
self.local_range = config.local_range
x_angles = view_angles if config.rotation_axes[0] else np.array([0])
z_angles = view_angles if config.rotation_axes[1] else np.array([0])
y_angles = view_angles if config.rotation_axes[2] else np.array([0])
x_angles, z_angles, y_angles = np.meshgrid(x_angles, z_angles, y_angles)
angles = np.stack([x_angles.flatten(), z_angles.flatten(), y_angles.flatten()], axis=1)
self.view_angles = angles
def preprocessing(self, motion3d, view_angle=None, params=None):
if self.aug: motion3d, params = self.augmentation(motion3d, params)
basis = None
if view_angle is not None: basis = get_change_of_basis(motion3d, view_angle)
motion3d = preprocess_mixamo(motion3d)
motion3d = rotate_motion_3d(motion3d, basis)
motion2d = motion3d[:, [0, 2], :]
motion2d_scale = limb_scale_motion_2d(motion2d, self.global_range, self.local_range)
motion2d = localize_motion(motion2d)
motion2d_scale = localize_motion(motion2d_scale)
motion2d = normalize_motion(motion2d, self.meanpose, self.stdpose)
motion2d_scale = normalize_motion(motion2d_scale, self.meanpose, self.stdpose)
motion2d = motion2d.reshape([-1, motion2d.shape[-1]])
motion2d_scale = motion2d_scale.reshape((-1, motion2d_scale.shape[-1]))
motion2d = torch.from_numpy(motion2d).float()
motion2d_scale = torch.from_numpy(motion2d_scale).float()
return motion2d, motion2d_scale
def __getitem__(self, index):
# select two motions
motion_idx = np.random.choice(len(self.motion_names))
motion = self.motion_names[motion_idx]
# select two characters
char_idx = np.random.choice(len(self.character_names))
character = self.character_names[char_idx]
view_idx = np.random.choice(len(self.view_angles))
view = self.view_angles[view_idx]
if self.aug:
param = self.gen_aug_params(rotate=True)
else:
param = None
item = self.build_item(motion, character)
x, x_s = self.preprocessing(self.load_item(item), view, param)
return {"x": x, "x_s": x_s, "mot": motion, "char": character, "view": view,
"meanpose": self.meanpose, "stdpose": self.stdpose}
class SoloDanceDataset(Dataset):
def __init__(self, phase, config):
super(SoloDanceDataset, self).__init__()
self.global_range = config.global_range
self.local_range = config.local_range
assert phase in ['train', 'test']
self.data_root = config.train_dir if phase=='train' else config.test_dir
self.phase = phase
self.unit = config.unit
self.meanpose_path = config.train_meanpose_path if phase == 'train' else config.test_meanpose_path
self.stdpose_path = config.train_stdpose_path if phase == 'train' else config.test_stdpose_path
self.character_names = sorted(os.listdir(self.data_root))
self.items = glob.glob(os.path.join(self.data_root, '*/*/motions/*.npy'))
self.meanpose, self.stdpose = get_meanpose(phase, config)
self.meanpose = self.meanpose.astype(np.float32)
self.stdpose = self.stdpose.astype(np.float32)
if 'preload' in config and config.preload:
self.preload()
self.cached = True
else:
self.cached = False
def load_item(self, item):
if self.cached:
data = self.cache[item]
else:
data = np.load(item)
return data
def preload(self):
print("pre-loading into memory")
pbar = tqdm(total=len(self))
self.cache = {}
for item in self.items:
motion = np.load(item)
self.cache[item] = motion
pbar.update(1)
def preprocessing(self, motion):
motion = motion * self.unit
motion[1, :, :] = (motion[2, :, :] + motion[5, :, :]) / 2
motion[8, :, :] = (motion[9, :, :] + motion[12, :, :]) / 2
global_scale = self.global_range[0] + np.random.random() * (self.global_range[1] - self.global_range[0])
local_scales = self.local_range[0] + np.random.random([8]) * (self.local_range[1] - self.local_range[0])
motion_scale = scale_limbs(motion, global_scale, local_scales)
motion = localize_motion(motion)
motion_scale = localize_motion(motion_scale)
motion = normalize_motion(motion, self.meanpose, self.stdpose)
motion_scale = normalize_motion(motion_scale, self.meanpose, self.stdpose)
motion = motion.reshape((-1, motion.shape[-1]))
motion_scale = motion_scale.reshape((-1, motion_scale.shape[-1]))
motion = torch.from_numpy(motion).float()
motion_scale = torch.from_numpy(motion_scale).float()
return motion, motion_scale
def __len__(self):
return len(self.items)
def __getitem__(self, index):
item = self.items[index]
motion = self.load_item(item)
x, x_s = self.preprocessing(motion)
return {"x": x, "x_s": x_s, "meanpose": self.meanpose, "stdpose": self.stdpose}
|