def get_predicted_features(self, pos_past, traj, height, orient_pred,
quat_pred):
num_samples = quat_pred.shape[0]
num_frames = quat_pred.shape[1]
orient_pred = orient_pred.view(num_samples, num_frames, 1, -1)
quat_pred = quat_pred.view(num_samples, num_frames, self.V - 1, -1)
quats_world = torch.cat((orient_pred, quat_pred), dim=-2)
root_pred = torch.zeros(
(num_samples, num_frames, self.C)).cuda().float()
root_pred[:, :, [0, 2]] = traj
root_pred[:, :, 1] = height
pos_pred = MocapDataset.forward_kinematics(
quats_world, root_pred, self.joint_parents,
torch.from_numpy(self.joint_offsets).float().cuda())
affs_pred = torch.tensor(
MocapDataset.get_affective_features(
pos_pred.detach().cpu().numpy())).cuda().float()
spline = []
for s in range(num_samples):
data_pred_curr = dict()
data_pred_curr['positions'] = torch.cat(
(pos_past[s], pos_pred[s]), dim=0).detach().cpu().numpy()
data_pred_curr[
'trans_and_controls'] = MocapDataset.compute_translations_and_controls(
data_pred_curr)
spline.append(
Spline.extract_spline_features(
MocapDataset.compute_splines(data_pred_curr))[0][-1:])
spline_pred = torch.from_numpy(np.stack(spline, axis=0)).cuda().float()
return pos_pred, affs_pred, spline_pred
def return_batch(self, batch_size, dataset):
if len(batch_size) > 1:
rand_keys = np.copy(batch_size)
batch_size = len(batch_size)
else:
batch_size = batch_size[0]
probs = []
for k in dataset.keys():
if 'spline' not in dataset[k]:
raise KeyError(
'No splines found. Perhaps you forgot to compute them?'
)
probs.append(dataset[k]['spline'].size())
probs = np.array(probs) / np.sum(probs)
rand_keys = np.random.choice(len(dataset),
size=batch_size,
replace=False,
p=probs)
batch_pos = np.zeros((batch_size, self.T, self.V, self.C),
dtype='float32')
batch_traj = np.zeros((batch_size, self.T, self.C), dtype='float32')
batch_quat = np.zeros((batch_size, self.T, (self.V - 1) * self.D),
dtype='float32')
batch_orient = np.zeros((batch_size, self.T, self.O), dtype='float32')
batch_affs = np.zeros((batch_size, self.T, self.A), dtype='float32')
batch_spline = np.zeros((batch_size, self.T, self.S), dtype='float32')
batch_phase_and_root_speed = np.zeros((batch_size, self.T, self.PRS),
dtype='float32')
batch_labels = np.zeros((batch_size, 1, self.num_labels[0]),
dtype='float32')
for i, k in enumerate(rand_keys):
pos = dataset[str(k)]['positions_world']
traj = dataset[str(k)]['trajectory']
quat = dataset[str(k)]['rotations']
orient = dataset[str(k)]['orientations']
affs = dataset[str(k)]['affective_features']
spline, phase = Spline.extract_spline_features(
dataset[str(k)]['spline'])
root_speed = dataset[str(k)]['trans_and_controls'][:, -1].reshape(
-1, 1)
labels = dataset[str(k)]['labels'][:self.num_labels[0]]
batch_pos[i] = pos
batch_traj[i] = traj
batch_quat[i] = quat.reshape(self.T, -1)
batch_orient[i] = orient.reshape(self.T, -1)
batch_affs[i] = affs
batch_spline[i] = spline
batch_phase_and_root_speed[i] = np.concatenate((phase, root_speed),
axis=-1)
batch_labels[i] = np.expand_dims(labels, axis=0)
return batch_pos, batch_traj, batch_quat, batch_orient, batch_affs, batch_spline,\
batch_phase_and_root_speed, batch_labels
def return_batch(self, batch_size, dataset, randomized=True):
if len(batch_size) > 1:
rand_keys = np.copy(batch_size)
batch_size = len(batch_size)
else:
batch_size = batch_size[0]
probs = []
for k in dataset.keys():
if 'spline' not in dataset[k]:
raise KeyError('No splines found. Perhaps you forgot to compute them?')
probs.append(dataset[k]['spline'].size())
probs = np.array(probs) / np.sum(probs)
if randomized:
rand_keys = np.random.choice(len(dataset), size=batch_size, replace=False, p=probs)
else:
rand_keys = np.arange(batch_size)
batch_pos = np.zeros((batch_size, self.T, self.V, self.C), dtype='float32')
batch_quat = np.zeros((batch_size, self.T, (self.V - 1) * self.D), dtype='float32')
batch_orient = np.zeros((batch_size, self.T, self.O), dtype='float32')
batch_z_mean = np.zeros((batch_size, self.Z), dtype='float32')
batch_z_dev = np.zeros((batch_size, self.T, self.Z), dtype='float32')
batch_root_speed = np.zeros((batch_size, self.T, self.RS), dtype='float32')
batch_affs = np.zeros((batch_size, self.T, self.A), dtype='float32')
batch_spline = np.zeros((batch_size, self.T, self.S), dtype='float32')
batch_labels = np.zeros((batch_size, 1, self.num_labels[0]), dtype='float32')
pseudo_passes = (len(dataset) + batch_size - 1) // batch_size
for i, k in enumerate(rand_keys):
pos = dataset[str(k)]['positions'][:self.T]
quat = dataset[str(k)]['rotations'][:self.T, 1:]
orient = dataset[str(k)]['rotations'][:self.T, 0]
affs = dataset[str(k)]['affective_features'][:self.T]
spline, phase = Spline.extract_spline_features(dataset[str(k)]['spline'])
spline = spline[:self.T]
phase = phase[:self.T]
z = dataset[str(k)]['trans_and_controls'][:, 1][:self.T]
z_mean = np.mean(z[:self.prefix_length])
z_dev = z - z_mean
root_speed = dataset[str(k)]['trans_and_controls'][:, -1][:self.T]
labels = dataset[str(k)]['labels'][:self.num_labels[0]]
batch_pos[i] = pos
batch_quat[i] = quat.reshape(self.T, -1)
batch_orient[i] = orient.reshape(self.T, -1)
batch_z_mean[i] = z_mean.reshape(-1, 1)
batch_z_dev[i] = z_dev.reshape(self.T, -1)
batch_root_speed[i] = root_speed.reshape(self.T, 1)
batch_affs[i] = affs
batch_spline[i] = spline
batch_labels[i] = np.expand_dims(labels, axis=0)
return batch_pos, batch_quat, batch_orient, batch_z_mean, batch_z_dev,\
batch_root_speed, batch_affs, batch_spline, batch_labels
def get_predicted_features(self, pos_past, orient_past, traj, height,
quat_pred, orient_pred):
num_samples = quat_pred.shape[0]
num_frames = quat_pred.shape[1]
offsets = torch.from_numpy(self.joint_offsets).cuda().float(). \
unsqueeze(0).unsqueeze(0).repeat(num_samples, num_frames, 1, 1)
quat_pred = quat_pred.view(num_samples, num_frames, self.V - 1, -1)
zeros = torch.zeros_like(orient_pred)
quats_world = quat_pred.clone()
quats_world = torch.cat((expmap_to_quaternion(
torch.cat((zeros, orient_pred, zeros),
dim=-1)).unsqueeze(-2), quats_world),
dim=-2)
pos_pred = torch.zeros(
(num_samples, num_frames, self.V, self.C)).cuda().float()
pos_pred[:, :, 0, [0, 2]] = traj
pos_pred[:, :, 0, 1] = height
# for joint in range(self.V):
# if self.joint_parents[joint] == -1:
# quats_world[:, :, joint] = expmap_to_quaternion(torch.cat((zeros, orient_pred, zeros), dim=-1))
# else:
# pos_pred[:, :, joint] = qrot(quats_world[:, :, self.joint_parents[joint]], offsets[:, :, joint]) \
# + pos_pred[:, :, self.joint_parents[joint]]
# quats_world[:, :, joint] = qmul(quats_world[:, :, self.joint_parents[joint]],
# quat_pred[:, :, joint - 1])
for joint in range(1, self.V):
pos_pred[:, :, joint] = qrot(quats_world[:, :, joint], offsets[:, :, joint]) \
+ pos_pred[:, :, self.joint_parents[joint]]
affs_pred = torch.tensor(
MocapDataset.get_affective_features(
pos_pred.detach().cpu().numpy())).cuda().float()
spline = []
for s in range(num_samples):
data_pred_curr = dict()
data_pred_curr['positions_world'] = torch.cat(
(pos_past[s], pos_pred[s]), dim=0).detach().cpu().numpy()
data_pred_curr['trajectory'] = data_pred_curr['positions_world'][:,
0]
data_pred_curr['orientations'] = torch.cat(
(orient_past[s], orient_pred[s]),
dim=0).squeeze().detach().cpu().numpy()
data_pred_curr[
'trans_and_controls'] = MocapDataset.compute_translations_and_controls(
data_pred_curr)
spline.append(
Spline.extract_spline_features(
MocapDataset.compute_splines(data_pred_curr))[0][-1:])
spline_pred = torch.from_numpy(np.stack(spline, axis=0)).cuda().float()
return pos_pred, affs_pred, spline_pred