def __init__(self, config, subset_name, data_path=None, extra_data_dir=None):
super(MotionNorm, self).__init__()
np.random.seed(2020)
self.skel = Skel() # TD: add config
if data_path is None:
data_path = config.data_path
dataset = np.load(data_path, allow_pickle=True)[subset_name].item()
motions, labels, metas = dataset["motion"], dataset["style"], dataset["meta"]
self.label_i = labels
self.len = len(self.label_i)
self.metas = [{key: metas[key][i] for key in metas.keys()} for i in range(self.len)]
self.motion_i, self.foot_i = [], []
content, style3d, style2d = [], [], []
self.labels = []
self.data_dict = {}
self.diff_labels_dict = {}
for i, motion in enumerate(motions):
label = labels[i]
anim = AnimationData(motion, skel=self.skel)
if label not in self.labels:
self.labels.append(label)
self.data_dict[label] = []
self.data_dict[label].append(i)
self.motion_i.append(anim)
self.foot_i.append(anim.get_foot_contact(transpose=True)) # [4, T]
content.append(anim.get_content_input())
style3d.append(anim.get_style3d_input())
view_angles, scales = [], []
for v in range(10):
view_angles.append(self.random_view_angle())
scales.append(self.random_scale())
style2d.append(anim.get_projections(view_angles, scales))
# calc diff labels
for x in self.labels:
self.diff_labels_dict[x] = [y for y in self.labels if y != x]
if extra_data_dir is None:
extra_data_dir = config.extra_data_dir
norm_cfg = config.dataset_norm_config
norm_data = []
for key, raw in zip(["content", "style3d", "style2d"], [content, style3d, style2d]):
prefix = norm_cfg[subset_name][key]
pre_computed = prefix is not None
if prefix is None:
prefix = subset_name
norm_data.append(NormData(prefix + "_" + key, pre_computed, raw,
config, extra_data_dir, keep_raw=(key != "style2d")))
self.content, self.style3d, self.style2d = norm_data
self.device = config.device
self.rand = random.SystemRandom()
def __init__(self,
config,
subset_name,
data_path=None,
extra_data_dir=None,
panda=False):
super(MotionNorm, self).__init__()
np.random.seed(2020)
self.skel = Skel() # TD: add config
if panda:
self.skel = PandaSkel()
if data_path is None:
data_path = config.data_path
dataset = np.load(data_path, allow_pickle=True)[subset_name].item()
'''
motions: arrays of T x 132
labels: array of integer values to denote the 'style' of the motion
metas:
- style: array of string labels e.g. 'angry', 'childlike'
- content: array of string labels e.g. 'walk'
- phase: array of floats
'''
motions, labels, metas = dataset["motion"], dataset["style"], dataset[
"meta"]
self.label_i = labels
self.len = len(self.label_i)
self.metas = [{key: metas[key][i]
for key in metas.keys()} for i in range(self.len)]
self.motion_i = []
# self.foot_i = []
content, style3d, style2d = [], [], []
self.labels = []
'''data_dict contains mapping of style label to indices belonging to this label'''
self.data_dict = {}
self.diff_labels_dict = {}
for i, motion in enumerate(motions):
label = labels[i]
anim = AnimationData(motion, skel=self.skel, panda=panda)
if label not in self.labels:
self.labels.append(label)
self.data_dict[label] = []
self.data_dict[label].append(i)
self.motion_i.append(anim)
# self.foot_i.append(anim.get_foot_contact(transpose=True)) # [4, T]
content.append(anim.get_content_input())
style3d.append(anim.get_style3d_input())
# TODO: FIGURE OUT HOW TO CALCULATE ROOT ROTATION THEN CAN DO STYLE_2D
# view_angles, scales = [], []
# for v in range(10):
# view_angles.append(self.random_view_angle())
# scales.append(self.random_scale())
# style2d.append(anim.get_projections(view_angles, scales))
# calc diff labels
for x in self.labels:
self.diff_labels_dict[x] = [y for y in self.labels if y != x]
if extra_data_dir is None:
extra_data_dir = config.extra_data_dir
norm_cfg = config.dataset_norm_config
norm_data = []
# for key, raw in zip(["content", "style3d", "style2d"], [content, style3d, style2d]):
# prefix = norm_cfg[subset_name][key]
# pre_computed = prefix is not None
# if prefix is None:
# prefix = subset_name
# norm_data.append(NormData(prefix + "_" + key, pre_computed, raw,
# config, extra_data_dir, keep_raw=(key != "style2d")))
# self.content, self.style3d, self.style2d = norm_data
for key, raw in zip(["content", "style3d"], [content, style3d]):
prefix = norm_cfg[subset_name][key]
pre_computed = prefix is not None
if prefix is None:
prefix = subset_name
norm_data.append(
NormData(prefix + "_" + key,
pre_computed,
raw,
config,
extra_data_dir,
keep_raw=(key != "style2d")))
self.content, self.style3d = norm_data
self.device = config.device
self.rand = random.SystemRandom()