def __init__(self,
seq_type,
root_dir,
data_list,
cache_path=None,
step_size=100,
window_size=400,
random_shift=0,
transform=None,
**kwargs):
super(SequenceToSequenceDataset, self).__init__()
self.seq_type = seq_type
self.feature_dim = seq_type.feature_dim
self.target_dim = seq_type.target_dim
self.aux_dim = seq_type.aux_dim
self.window_size = window_size
self.step_size = step_size
self.random_shift = random_shift
self.transform = transform
self.data_path = [osp.join(root_dir, data) for data in data_list]
self.index_map = []
self.features, self.targets, aux = load_cached_sequences(
seq_type, root_dir, data_list, cache_path, **kwargs)
# Optionally smooth the sequence
feat_sigma = kwargs.get('feature_sigma,', -1)
targ_sigma = kwargs.get('target_sigma,', -1)
if feat_sigma > 0:
self.features = [
gaussian_filter1d(feat, sigma=feat_sigma, axis=0)
for feat in self.features
]
if targ_sigma > 0:
self.targets = [
gaussian_filter1d(targ, sigma=targ_sigma, axis=0)
for targ in self.targets
]
max_norm = kwargs.get('max_velocity_norm', 3.0)
self.ts, self.orientations, self.gt_pos, self.local_v = [], [], [], []
for i in range(len(data_list)):
self.features[i] = self.features[i][:-1]
self.targets[i] = self.targets[i]
self.ts.append(aux[i][:-1, :1])
self.orientations.append(aux[i][:-1, 1:5])
self.gt_pos.append(aux[i][:-1, 5:8])
velocity = np.linalg.norm(
self.targets[i], axis=1) # Remove outlier ground truth data
bad_data = velocity > max_norm
for j in range(window_size + random_shift,
self.targets[i].shape[0], step_size):
if not bad_data[j - window_size - random_shift:j +
random_shift].any():
self.index_map.append([i, j])
if kwargs.get('shuffle', True):
random.shuffle(self.index_map)
def __init__(self, seq_type, root_dir, data_list, cache_path=None, step_size=10, window_size=200,
random_shift=0, transform=None, **kwargs):
super().__init__()
self.feature_dim = seq_type.feature_dim
self.target_dim = seq_type.target_dim
self.aux_dim = seq_type.aux_dim
self.window_size = window_size
self.step_size = step_size
self.random_shift = random_shift
self.transform = transform
self.data_path = [osp.join(root_dir, data) for data in data_list]
self.index_map = []
self.ts, self.orientations, self.gt_pos = [], [], []
self.features, self.targets, aux = load_cached_sequences(
seq_type, root_dir, data_list, cache_path, interval=1, **kwargs)
# Optionally smooth the sequence
feat_sigma = kwargs.get('feature_sigma,', -1)
targ_sigma = kwargs.get('target_sigma,', -1)
if feat_sigma > 0:
self.features = [gaussian_filter1d(feat, sigma=feat_sigma, axis=0) for feat in self.features]
if targ_sigma > 0:
self.targets = [gaussian_filter1d(targ, sigma=targ_sigma, axis=0) for targ in self.targets]
for i in range(len(data_list)):
self.ts.append(aux[i][:, 0])
self.orientations.append(aux[i][:, 1:5])
self.gt_pos.append(aux[i][:, -3:])
self.index_map += [[i, j] for j in range(window_size, self.targets[i].shape[0], step_size)]
if kwargs.get('shuffle', True):
random.shuffle(self.index_map)
def __init__(self, seq_type, root_dir, data_list, cache_path=None, step_size=10, window_size=200,
random_shift=0, transform=None, **kwargs):
super(StridedSequenceDataset, self).__init__()
self.feature_dim = seq_type.feature_dim
self.target_dim = seq_type.target_dim
self.aux_dim = seq_type.aux_dim
self.window_size = window_size
self.step_size = step_size
self.random_shift = random_shift
self.transform = transform
self.interval = kwargs.get('interval', window_size)
self.data_path = [osp.join(root_dir, data) for data in data_list]
self.index_map = []
self.ts, self.orientations, self.gt_pos = [], [], []
self.features, self.targets, aux = load_cached_sequences(
seq_type, root_dir, data_list, cache_path, interval=self.interval, **kwargs)
for i in range(len(data_list)):
self.ts.append(aux[i][:, 0])
self.orientations.append(aux[i][:, 1:5])
self.gt_pos.append(aux[i][:, -3:])
self.index_map += [[i, j]
for j in range(0, self.targets[i].shape[0], step_size)]
if kwargs.get('shuffle', True):
random.shuffle(self.index_map)