def to_tensor(self, to_rot=True, *args, **kwargs):
if not torch.is_tensor(self.left_hand_pose):
if self.left_hand_pose is not None:
self.left_hand_pose = torch.from_numpy(self.left_hand_pose)
if not torch.is_tensor(self.right_hand_pose):
if self.right_hand_pose is not None:
self.right_hand_pose = torch.from_numpy(self.right_hand_pose)
if to_rot:
if self.left_hand_pose is not None:
self.left_hand_pose = batch_rodrigues(
self.left_hand_pose.view(-1, 3)).view(-1, 3, 3)
if self.right_hand_pose is not None:
self.right_hand_pose = batch_rodrigues(
self.right_hand_pose.view(-1, 3)).view(-1, 3, 3)
for k, v in self.extra_fields.items():
if isinstance(v, GenericTarget):
v.to_tensor(*args, **kwargs)
def to_tensor(self, to_rot=True, *args, **kwargs):
self.body_pose = torch.from_numpy(self.body_pose)
if to_rot:
self.body_pose = batch_rodrigues(self.body_pose.view(-1, 3)).view(
-1, 3, 3)
for k, v in self.extra_fields.items():
if isinstance(v, GenericTarget):
v.to_tensor(*args, **kwargs)
def forward(self, module_input, prev_val):
if self.append_params:
if self.to_aa:
prev_val = batch_rodrigues(prev_val)
prev_val = prev_val[:, :, :2].contiguous().view(
-1, self.num_angles * 6)
module_input = torch.cat([module_input, prev_val], dim=-1)
cont_repr = super(ContinuousRotReprRegressor,
self).forward(module_input)
output = self.repr_decoder(cont_repr).view(-1, self.num_angles, 3, 3)
return output
def forward(self, pca_coeffs):
batch_size = pca_coeffs.shape[0]
decoded = torch.einsum('bi,ij->bj',
[pca_coeffs, self.pca_basis]) + self.mean
return batch_rodrigues(decoded.view(-1, 3)).view(batch_size, -1, 3, 3)
def forward(self, module_input):
output = batch_rodrigues(module_input.view(-1, 3)).view(
-1, self.num_angles, 3, 3)
return output
def forward(self, features):
aa_vectors = super(AARegressor, self).forward(features).view(
-1, self.num_angles, 3)
return batch_rodrigues(aa_vectors.view(-1, 3)).view(
-1, self.num_angles, 3, 3)