Spaces:
Sleeping
Sleeping
File size: 7,805 Bytes
5ac1897 |
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 |
from lib.kits.basic import *
from lib.utils.camera import perspective_projection
from lib.modeling.losses import compute_poses_angle_prior_loss
from .utils import (
gmof,
guess_cam_z,
estimate_kp2d_scale,
get_kp_active_jids,
get_params_active_q_masks,
)
def build_closure(
self,
cfg,
optimizer,
inputs,
focal_length : float,
gt_kp2d,
log_data,
):
B = len(gt_kp2d)
act_parts = instantiate(cfg.parts)
act_q_masks = None
if not (act_parts == 'all' or 'all' in act_parts):
act_q_masks = get_params_active_q_masks(act_parts)
# Shortcuts for the inference of the skeleton model.
def inference_skel(inputs):
poses_active = torch.cat([inputs['poses_orient'], inputs['poses_body']], dim=-1) # (B, 46)
if act_q_masks is not None:
poses_hidden = poses_active.clone().detach() # (B, 46)
poses = poses_active * act_q_masks + poses_hidden * (1 - act_q_masks) # (B, 46)
else:
poses = poses_active
skel_params = {
'poses' : poses, # (B, 46)
'betas' : inputs['betas'], # (B, 10)
}
skel_output = self.skel_model(**skel_params, skelmesh=False)
return skel_params, skel_output
# Estimate the camera depth as an initialization if depth loss is enabled.
gs_cam_z = None
if 'w_depth' in cfg.losses:
with torch.no_grad():
_, skel_output = inference_skel(inputs)
gs_cam_z = guess_cam_z(
pd_kp3d = skel_output.joints,
gt_kp2d = gt_kp2d,
focal_length = focal_length,
)
# Prepare the focal length for the perspective projection.
focal_length_xy = np.ones((B, 2)) * focal_length # (B, 2)
def closure():
optimizer.zero_grad()
# 📦 Data preparation.
with PM.time_monitor('SKEL-forward'):
skel_params, skel_output = inference_skel(inputs)
with PM.time_monitor('reproj'):
pd_kp2d = perspective_projection(
points = to_tensor(skel_output.joints, device=self.device),
translation = to_tensor(inputs['cam_t'], device=self.device),
focal_length = to_tensor(focal_length_xy, device=self.device),
)
with PM.time_monitor('compute_losses'):
loss, losses = compute_losses(
# Loss configuration.
loss_cfg = instantiate(cfg.losses),
parts = act_parts,
# Data inputs.
gt_kp2d = gt_kp2d,
pd_kp2d = pd_kp2d,
pd_params = skel_params,
pd_cam_z = inputs['cam_t'][:, 2],
gs_cam_z = gs_cam_z,
)
with PM.time_monitor('visualization'):
VISUALIZE = True
if VISUALIZE:
# For visualize the optimization process.
kp2d_err = torch.sum((pd_kp2d - gt_kp2d[..., :2]) ** 2, dim=-1) * gt_kp2d[..., 2] # (B, J)
kp2d_err = kp2d_err.sum(dim=-1) / (torch.sum(gt_kp2d[..., 2], dim=-1) + 1e-6) # (B,)
# Store logging data.
if self.tb_logger is not None:
log_data.update({
'losses' : losses,
'pd_kp2d' : pd_kp2d[:self.n_samples].detach().clone(),
'pd_verts' : skel_output.skin_verts[:self.n_samples].detach().clone(),
'cam_t' : inputs['cam_t'][:self.n_samples].detach().clone(),
'optim_betas' : inputs['betas'][:self.n_samples].detach().clone(),
'kp2d_err' : kp2d_err[:self.n_samples].detach().clone(),
})
with PM.time_monitor('backwards'):
loss.backward()
return loss.item()
return closure
def compute_losses(
loss_cfg : Dict[str, Union[bool, float]],
parts : List[str],
gt_kp2d : torch.Tensor,
pd_kp2d : torch.Tensor,
pd_params : Dict[str, torch.Tensor],
pd_cam_z : torch.Tensor,
gs_cam_z : Optional[torch.Tensor] = None,
):
'''
### Args
- loss_cfg: Dict[str, Union[bool, float]]
- Special option flags (`f_xxx`) or loss weights (`w_xxx`).
- parts: List[str]
- The list of the active joint parts groups.
- Among {'all', 'torso', 'torso-lite', 'limbs', 'head', 'limbs_proximal', 'limbs_distal'}.
- gt_kp2d: torch.Tensor (B, 44, 3)
- The ground-truth 2D keypoints with confidence.
- pd_kp2d: torch.Tensor (B, 44, 2)
- The predicted 2D keypoints.
- pd_params: Dict[str, torch.Tensor]
- poses: torch.Tensor (B, 46)
- betas: torch.Tensor (B, 10)
- pd_cam_z: torch.Tensor (B,)
- The predicted camera depth translation.
- gs_cam_z: Optional[torch.Tensor] (B,)
- The guessed camera depth translation.
### Returns
- loss: torch.Tensor (,)
- The weighted loss value for optimization.
- losses: Dict[str, float]
- The dictionary of the loss values for logging.
'''
losses = {}
loss = torch.tensor(0.0, device=gt_kp2d.device)
kp2d_conf = gt_kp2d[:, :, 2] # (B, J)
gt_kp2d = gt_kp2d[:, :, :2] # (B, J, 2)
# Special option flags.
f_normalize_kp2d = loss_cfg.get('f_normalize_kp2d', False)
if f_normalize_kp2d:
scale2mean = loss_cfg.get('f_normalize_kp2d_to_mean', False)
scale2d = estimate_kp2d_scale(gt_kp2d) # (B,)
pd_kp2d = pd_kp2d / (scale2d[:, None, None] + 1e-6) # (B, J, 2)
gt_kp2d = gt_kp2d / (scale2d[:, None, None] + 1e-6) # (B, J, 2)
if scale2mean:
scale2d_mean = scale2d.mean()
pd_kp2d = pd_kp2d * scale2d_mean # (B, J, 2)
gt_kp2d = gt_kp2d * scale2d_mean # (B, J, 2)
# Mask the keypoints.
act_jids = get_kp_active_jids(parts)
kp2d_conf = kp2d_conf[:, act_jids] # (B, J)
gt_kp2d = gt_kp2d[:, act_jids, :] # (B, J, 2)
pd_kp2d = pd_kp2d[:, act_jids, :] # (B, J, 2)
# Calculate weighted losses.
w_depth = loss_cfg.get('w_depth', None)
w_reprojection = loss_cfg.get('w_reprojection', None)
w_shape_prior = loss_cfg.get('w_shape_prior', None)
w_angle_prior = loss_cfg.get('w_angle_prior', None)
if w_depth:
assert gs_cam_z is not None, 'The guessed camera depth is required for the depth loss.'
depth_loss = (gs_cam_z - pd_cam_z).pow(2) # (B,)
loss += (w_depth ** 2) * depth_loss.mean() # (,)
losses['depth'] = (w_depth ** 2) * depth_loss.mean().item() # float
if w_reprojection:
reproj_err_j = gmof(pd_kp2d - gt_kp2d).sum(dim=-1) # (B, J)
reproj_err_j = kp2d_conf.pow(2) * reproj_err_j # (B, J)
reproj_loss = reproj_err_j.sum(-1) # (B,)
loss += (w_reprojection ** 2) * reproj_loss.mean() # (,)
losses['reprojection'] = (w_reprojection ** 2) * reproj_loss.mean().item() # float
if w_shape_prior:
shape_prior_loss = pd_params['betas'].pow(2).sum(dim=-1) # (B,)
loss += (w_shape_prior ** 2) * shape_prior_loss.mean() # (,)
losses['shape_prior'] = (w_shape_prior ** 2) * shape_prior_loss.mean().item() # float
if w_angle_prior:
w_angle_prior *= loss_cfg.get('w_angle_prior_scale', 1.0)
angle_prior_loss = compute_poses_angle_prior_loss(pd_params['poses']) # (B,)
loss += (w_angle_prior ** 2) * angle_prior_loss.mean() # (,)
losses['angle_prior'] = (w_angle_prior ** 2) * angle_prior_loss.mean().item() # float
losses['weighted_sum'] = loss.item() # float
return loss, losses |