AiOS / detrsmpl /models /losses /prior_loss.py
ttxskk
update
d7e58f0
raw
history blame
29.1 kB
import itertools
import os
import pickle
import sys
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from detrsmpl.core.conventions.joints_mapping.standard_joint_angles import (
STANDARD_JOINT_ANGLE_LIMITS,
TRANSFORMATION_AA_TO_SJA,
TRANSFORMATION_SJA_TO_AA,
)
from detrsmpl.utils.keypoint_utils import search_limbs
from detrsmpl.utils.transforms import aa_to_rot6d, aa_to_sja
class ShapePriorLoss(nn.Module):
"""Prior loss for body shape parameters.
Args:
reduction (str, optional): The method that reduces the loss to a
scalar. Options are "none", "mean" and "sum".
loss_weight (float, optional): The weight of the loss. Defaults to 1.0
"""
def __init__(self, reduction='mean', loss_weight=1.0):
super().__init__()
assert reduction in (None, 'none', 'mean', 'sum')
self.reduction = reduction
self.loss_weight = loss_weight
def forward(self,
betas,
loss_weight_override=None,
reduction_override=None):
"""Forward function of loss.
Args:
betas (torch.Tensor): The body shape parameters
loss_weight_override (float, optional): The weight of loss used to
override the original weight of loss
reduction_override (str, optional): The reduction method used to
override the original reduction method of the loss.
Defaults to None
Returns:
torch.Tensor: The calculated loss
"""
assert reduction_override in (None, 'none', 'mean', 'sum')
reduction = (reduction_override
if reduction_override else self.reduction)
loss_weight = (loss_weight_override if loss_weight_override is not None
else self.loss_weight)
shape_prior_loss = loss_weight * betas**2
if reduction == 'mean':
shape_prior_loss = shape_prior_loss.mean()
elif reduction == 'sum':
shape_prior_loss = shape_prior_loss.sum()
return shape_prior_loss
class ShapeThresholdPriorLoss(nn.Module):
"""Threshold loss for betas. Soft constraint to prevent parameters for
leaving feasible set. Implements a penalty constraint that encourages the
parameters to stay in the feasible set of solutions.
Args:
margin (int, optional): The threshold value
norm (str, optional): The loss method. Options are 'l1', l2'
loss_weight (float, optional): The weight of the loss. Defaults to 1.0
"""
def __init__(self, margin=1, norm='l2', epsilon=1e-7, loss_weight=1.0):
super().__init__()
self.margin = margin
assert norm in ['l1', 'l2'], 'Norm variable must me l1 or l2'
self.norm = norm
self.epsilon = epsilon
self.loss_weight = loss_weight
def forward(self, betas):
"""Forward function of loss.
Args:
betas (torch.Tensor): The body shape parameters
Returns:
torch.Tensor: The calculated loss
"""
abs_values = betas.abs()
mask = abs_values.gt(self.margin)
invalid_values = torch.masked_select(betas, mask)
if self.norm == 'l1':
return self.loss_weight * invalid_values.abs().sum() / (
mask.to(dtype=betas.dtype).sum() + self.epsilon)
elif self.norm == 'l2':
return self.loss_weight * invalid_values.pow(2).sum() / (
mask.to(dtype=betas.dtype).sum() + self.epsilon)
class PoseRegLoss(nn.Module):
"""Regulizer loss for body pose parameters.
Args:
reduction (str, optional): The method that reduces the loss to a
scalar. Options are "none", "mean" and "sum".
loss_weight (float, optional): The weight of the loss. Defaults to 1.0
"""
def __init__(self, reduction='mean', loss_weight=1.0):
super().__init__()
assert reduction in (None, 'none', 'mean', 'sum')
self.reduction = reduction
self.loss_weight = loss_weight
def forward(self,
body_pose,
weight=None,
avg_factor=None,
loss_weight_override=None,
reduction_override=None):
reduction = (reduction_override
if reduction_override else self.reduction)
loss_weight = (loss_weight_override if loss_weight_override is not None
else self.loss_weight)
pose_prior_loss = loss_weight * (body_pose**2)
if reduction == 'mean':
pose_prior_loss = pose_prior_loss.mean()
elif reduction == 'sum':
pose_prior_loss = pose_prior_loss.sum()
return pose_prior_loss
class LimbLengthLoss(nn.Module):
"""Limb length loss for body shape parameters. As betas are associated with
the height of a person, fitting on limb length help determine body shape
parameters. It penalizes the L2 distance between target limb length and
pred limb length. Note that it should take keypoints3d as input, as limb
length computed from keypoints2d varies with camera.
Args:
convention (str): Limb convention to search for keypoint connections.
reduction (str, optional): The method that reduces the loss to a
scalar. Options are "none", "mean" and "sum".
loss_weight (float, optional): The weight of the loss. Defaults to 1.0
eps (float, optional): epsilon for computing normalized limb vector.
Defaults to 1e-4.
"""
def __init__(self,
convention,
reduction='mean',
loss_weight=1.0,
eps=1e-4):
super().__init__()
assert reduction in (None, 'none', 'mean', 'sum')
self.reduction = reduction
self.loss_weight = loss_weight
self.eps = eps
limb_idxs, _ = search_limbs(data_source=convention)
limb_idxs = sorted(limb_idxs['body'])
self.limb_idxs = np.array(
list(x for x, _ in itertools.groupby(limb_idxs)))
def _compute_limb_length(self, keypoints3d):
kp_src = keypoints3d[:, self.limb_idxs[:, 0], :3]
kp_dst = keypoints3d[:, self.limb_idxs[:, 1], :3]
limb_vec = kp_dst - kp_src
limb_length = torch.norm(limb_vec, dim=2)
return limb_length
def _keypoint_conf_to_limb_conf(self, keypoint_conf):
limb_conf = torch.min(keypoint_conf[:, self.limb_idxs[:, 1]],
keypoint_conf[:, self.limb_idxs[:, 0]])
return limb_conf
def forward(self,
pred,
target,
pred_conf=None,
target_conf=None,
loss_weight_override=None,
reduction_override=None):
"""Forward function of LimbLengthLoss.
Args:
pred (torch.Tensor): The predicted smpl keypoints3d.
Shape should be (N, K, 3).
B: batch size. K: number of keypoints.
target (torch.Tensor): The ground-truth keypoints3d.
Shape should be (N, K, 3).
pred_conf (torch.Tensor, optional): Confidence of
predicted keypoints. Shape should be (N, K).
target_conf (torch.Tensor, optional): Confidence of
target keypoints. Shape should be (N, K).
loss_weight_override (float, optional): The weight of loss used to
override the original weight of loss. Defaults to None.
reduction_override (str, optional): The reduction method used to
override the original reduction method of the loss.
Defaults to None
Returns:
torch.Tensor: The calculated loss
"""
assert pred.dim() == 3 and pred.shape[-1] == 3
assert pred.shape == target.shape
if pred_conf is not None:
assert pred_conf.dim() == 2
assert pred_conf.shape == pred.shape[:2]
if target_conf is not None:
assert target_conf.dim() == 2
assert target_conf.shape == target.shape[:2]
assert reduction_override in (None, 'none', 'mean', 'sum')
reduction = (reduction_override
if reduction_override else self.reduction)
loss_weight = (loss_weight_override if loss_weight_override is not None
else self.loss_weight)
limb_len_target = self._compute_limb_length(target)
limb_len_pred = self._compute_limb_length(pred)
if target_conf is None:
target_conf = torch.ones_like(target[..., 0])
if pred_conf is None:
pred_conf = torch.ones_like(pred[..., 0])
limb_conf_target = self._keypoint_conf_to_limb_conf(target_conf)
limb_conf_pred = self._keypoint_conf_to_limb_conf(pred_conf)
limb_conf = limb_conf_target * limb_conf_pred
diff_len = limb_len_target - limb_len_pred
loss = diff_len**2 * limb_conf
if reduction == 'mean':
loss = loss.mean()
elif reduction == 'sum':
loss = loss.sum()
loss *= loss_weight
return loss
class JointPriorLoss(nn.Module):
"""Prior loss for joint angles.
Args:
reduction (str, optional): The method that reduces the loss to a
scalar. Options are "none", "mean" and "sum".
loss_weight (float, optional): The weight of the loss. Defaults to 1.0
use_full_body (bool, optional): Use full set of joint constraints
(in standard joint angles).
smooth_spine (bool, optional): Ensuring smooth spine rotations
smooth_spine_loss_weight (float, optional): An additional weight
factor multiplied on smooth spine loss
"""
def __init__(self,
reduction='mean',
loss_weight=1.0,
use_full_body=False,
smooth_spine=False,
smooth_spine_loss_weight=1.0):
super().__init__()
assert reduction in (None, 'none', 'mean', 'sum')
self.reduction = reduction
self.loss_weight = loss_weight
self.use_full_body = use_full_body
self.smooth_spine = smooth_spine
self.smooth_spine_loss_weight = smooth_spine_loss_weight
if self.use_full_body:
self.register_buffer('R_t', TRANSFORMATION_AA_TO_SJA)
self.register_buffer('R_t_inv', TRANSFORMATION_SJA_TO_AA)
self.register_buffer('sja_limits', STANDARD_JOINT_ANGLE_LIMITS)
def forward(self,
body_pose,
loss_weight_override=None,
reduction_override=None):
"""Forward function of loss.
Args:
body_pose (torch.Tensor): The body pose parameters
loss_weight_override (float, optional): The weight of loss used to
override the original weight of loss
reduction_override (str, optional): The reduction method used to
override the original reduction method of the loss.
Defaults to None
Returns:
torch.Tensor: The calculated loss
"""
assert reduction_override in (None, 'none', 'mean', 'sum')
reduction = (reduction_override
if reduction_override else self.reduction)
loss_weight = (loss_weight_override if loss_weight_override is not None
else self.loss_weight)
if self.use_full_body:
batch_size = body_pose.shape[0]
body_pose_reshape = body_pose.reshape(batch_size, -1, 3)
assert body_pose_reshape.shape[1] in (21, 23) # smpl-x, smpl
body_pose_reshape = body_pose_reshape[:, :21, :]
body_pose_sja = aa_to_sja(body_pose_reshape, self.R_t,
self.R_t_inv)
lower_limits = self.sja_limits[:, :, 0] # shape: (21, 3)
upper_limits = self.sja_limits[:, :, 1] # shape: (21, 3)
lower_loss = (torch.exp(F.relu(lower_limits - body_pose_sja)) -
1).pow(2)
upper_loss = (torch.exp(F.relu(body_pose_sja - upper_limits)) -
1).pow(2)
standard_joint_angle_prior_loss = (lower_loss + upper_loss).view(
body_pose.shape[0], -1) # shape: (n, 3)
joint_prior_loss = standard_joint_angle_prior_loss
else:
# default joint prior loss applied on elbows and knees
joint_prior_loss = (torch.exp(
body_pose[:, [55, 58, 12, 15]] *
torch.tensor([1., -1., -1, -1.], device=body_pose.device)) -
1)**2
if self.smooth_spine:
spine1 = body_pose[:, [9, 10, 11]]
spine2 = body_pose[:, [18, 19, 20]]
spine3 = body_pose[:, [27, 28, 29]]
smooth_spine_loss_12 = (torch.exp(F.relu(-spine1 * spine2)) -
1).pow(2) * self.smooth_spine_loss_weight
smooth_spine_loss_23 = (torch.exp(F.relu(-spine2 * spine3)) -
1).pow(2) * self.smooth_spine_loss_weight
joint_prior_loss = torch.cat(
[joint_prior_loss, smooth_spine_loss_12, smooth_spine_loss_23],
axis=1)
joint_prior_loss = loss_weight * joint_prior_loss
if reduction == 'mean':
joint_prior_loss = joint_prior_loss.mean()
elif reduction == 'sum':
joint_prior_loss = joint_prior_loss.sum()
return joint_prior_loss
class SmoothJointLoss(nn.Module):
"""Smooth loss for joint angles.
Args:
reduction (str, optional): The method that reduces the loss to a
scalar. Options are "none", "mean" and "sum".
loss_weight (float, optional): The weight of the loss. Defaults to 1.0
degree (bool, optional): The flag which represents whether the input
tensor is in degree or radian.
"""
def __init__(self,
reduction='mean',
loss_weight=1.0,
degree=False,
loss_func='L1'):
super().__init__()
assert reduction in (None, 'none', 'mean', 'sum')
assert loss_func in ('L1', 'L2')
self.reduction = reduction
self.loss_weight = loss_weight
self.degree = degree
self.loss_func = loss_func
def forward(self,
body_pose,
loss_weight_override=None,
reduction_override=None):
"""Forward function of SmoothJointLoss.
Args:
body_pose (torch.Tensor): The body pose parameters
loss_weight_override (float, optional): The weight of loss used to
override the original weight of loss
reduction_override (str, optional): The reduction method used to
override the original reduction method of the loss.
Defaults to None
Returns:
torch.Tensor: The calculated loss
"""
assert reduction_override in (None, 'none', 'mean', 'sum')
reduction = (reduction_override
if reduction_override else self.reduction)
loss_weight = (loss_weight_override if loss_weight_override is not None
else self.loss_weight)
theta = body_pose.reshape(body_pose.shape[0], -1, 3)
if self.degree:
theta = torch.deg2rad(theta)
rot_6d = aa_to_rot6d(theta)
rot_6d_diff = rot_6d[1:] - rot_6d[:-1]
if self.loss_func == 'L2':
smooth_joint_loss = (rot_6d_diff**2).sum(dim=[1, 2])
elif self.loss_func == 'L1':
smooth_joint_loss = rot_6d_diff.abs().sum(dim=[1, 2])
else:
raise TypeError(f'{self.func} is not defined')
# add zero padding to retain original batch_size
smooth_joint_loss = torch.cat(
[torch.zeros_like(smooth_joint_loss)[:1], smooth_joint_loss])
if reduction == 'mean':
smooth_joint_loss = smooth_joint_loss.mean()
elif reduction == 'sum':
smooth_joint_loss = smooth_joint_loss.sum()
smooth_joint_loss *= loss_weight
return smooth_joint_loss
class SmoothPelvisLoss(nn.Module):
"""Smooth loss for pelvis angles.
Args:
reduction (str, optional): The method that reduces the loss to a
scalar. Options are "none", "mean" and "sum".
loss_weight (float, optional): The weight of the loss. Defaults to 1.0
degree (bool, optional): The flag which represents whether the input
tensor is in degree or radian.
"""
def __init__(self, reduction='mean', loss_weight=1.0, degree=False):
super().__init__()
assert reduction in (None, 'none', 'mean', 'sum')
self.reduction = reduction
self.loss_weight = loss_weight
self.degree = degree
def forward(self,
global_orient,
loss_weight_override=None,
reduction_override=None):
"""Forward function of SmoothPelvisLoss.
Args:
global_orient (torch.Tensor): The global orientation parameters
loss_weight_override (float, optional): The weight of loss used to
override the original weight of loss
reduction_override (str, optional): The reduction method used to
override the original reduction method of the loss.
Defaults to None
Returns:
torch.Tensor: The calculated loss
"""
assert reduction_override in (None, 'none', 'mean', 'sum')
reduction = (reduction_override
if reduction_override else self.reduction)
loss_weight = (loss_weight_override if loss_weight_override is not None
else self.loss_weight)
if self.degree:
global_orient = torch.deg2rad(global_orient)
pelvis = global_orient.unsqueeze(1)
rot_6d = aa_to_rot6d(pelvis)
rot_6d_diff = rot_6d[1:] - rot_6d[:-1]
smooth_pelvis_loss = rot_6d_diff.abs().sum(dim=-1)
# add zero padding to retain original batch_size
smooth_pelvis_loss = torch.cat(
[torch.zeros_like(smooth_pelvis_loss)[:1],
smooth_pelvis_loss]).sum(dim=-1)
smooth_pelvis_loss = loss_weight * smooth_pelvis_loss
if reduction == 'mean':
smooth_pelvis_loss = smooth_pelvis_loss.mean()
elif reduction == 'sum':
smooth_pelvis_loss = smooth_pelvis_loss.sum()
return smooth_pelvis_loss
class SmoothTranslationLoss(nn.Module):
"""Smooth loss for translations.
Args:
reduction (str, optional): The method that reduces the loss to a
scalar. Options are "none", "mean" and "sum".
loss_weight (float, optional): The weight of the loss. Defaults to 1.0
"""
def __init__(self, reduction='mean', loss_weight=1.0):
super().__init__()
assert reduction in (None, 'none', 'mean', 'sum')
self.reduction = reduction
self.loss_weight = loss_weight
def forward(self,
translation,
loss_weight_override=None,
reduction_override=None):
"""Forward function of loss.
Args:
translation (torch.Tensor): The body translation parameters
loss_weight_override (float, optional): The weight of loss used to
override the original weight of loss
reduction_override (str, optional): The reduction method used to
override the original reduction method of the loss.
Defaults to None
Returns:
torch.Tensor: The calculated loss
"""
assert reduction_override in (None, 'none', 'mean', 'sum')
reduction = (reduction_override
if reduction_override else self.reduction)
loss_weight = (loss_weight_override if loss_weight_override is not None
else self.loss_weight)
translation_diff = translation[1:] - translation[:-1]
smooth_translation_loss = translation_diff.abs().sum(dim=-1,
keepdim=True)
# add zero padding to retain original batch_size
smooth_translation_loss = torch.cat([
torch.zeros_like(smooth_translation_loss)[:1],
smooth_translation_loss
]).sum(dim=-1)
smooth_translation_loss *= 1e3
smooth_translation_loss = loss_weight * \
smooth_translation_loss
if reduction == 'mean':
smooth_translation_loss = smooth_translation_loss.mean()
elif reduction == 'sum':
smooth_translation_loss = smooth_translation_loss.sum()
return smooth_translation_loss
class CameraPriorLoss(nn.Module):
"""Prior loss for predicted camera.
Args:
reduction (str, optional): The method that reduces the loss to a
scalar. Options are "none", "mean" and "sum".
scale (float, optional): The scale coefficient for regularizing camera
parameters. Defaults to 10
loss_weight (float, optional): The weight of the loss. Defaults to 1.0
"""
def __init__(self, scale=10, reduction='mean', loss_weight=1.0):
super().__init__()
self.scale = scale
self.reduction = reduction
self.loss_weight = loss_weight
def forward(self,
cameras,
loss_weight_override=None,
reduction_override=None):
"""Forward function of loss.
Args:
cameras (torch.Tensor): The predicted camera parameters
loss_weight_override (float, optional): The weight of loss used to
override the original weight of loss
reduction_override (str, optional): The reduction method used to
override the original reduction method of the loss.
Defaults to None
Returns:
torch.Tensor: The calculated loss
"""
assert reduction_override in (None, 'none', 'mean', 'sum')
reduction = (reduction_override
if reduction_override else self.reduction)
loss_weight = (loss_weight_override if loss_weight_override is not None
else self.loss_weight)
camera_prior_loss = torch.exp(-cameras[:, 0] * self.scale)
camera_prior_loss = torch.pow(camera_prior_loss, 2) * loss_weight
if reduction == 'mean':
camera_prior_loss = camera_prior_loss.mean()
elif reduction == 'sum':
camera_prior_loss = camera_prior_loss.sum()
return camera_prior_loss
class MaxMixturePrior(nn.Module):
"""Ref: SMPLify-X
https://github.com/vchoutas/smplify-x/blob/master/smplifyx/prior.py
"""
def __init__(self,
prior_folder='data',
num_gaussians=8,
dtype=torch.float32,
epsilon=1e-16,
use_merged=True,
reduction=None,
loss_weight=1.0):
super(MaxMixturePrior, self).__init__()
assert reduction in (None, 'none', 'mean', 'sum')
self.reduction = reduction
self.loss_weight = loss_weight
if dtype == torch.float32:
np_dtype = np.float32
elif dtype == torch.float64:
np_dtype = np.float64
else:
print('Unknown float type {}, exiting!'.format(dtype))
sys.exit(-1)
self.num_gaussians = num_gaussians
self.epsilon = epsilon
self.use_merged = use_merged
gmm_fn = 'gmm_{:02d}.pkl'.format(num_gaussians)
full_gmm_fn = os.path.join(prior_folder, gmm_fn)
if not os.path.exists(full_gmm_fn):
print('The path to the mixture prior "{}"'.format(full_gmm_fn) +
' does not exist, exiting!')
sys.exit(-1)
with open(full_gmm_fn, 'rb') as f:
gmm = pickle.load(f, encoding='latin1')
if type(gmm) == dict:
means = gmm['means'].astype(np_dtype)
covs = gmm['covars'].astype(np_dtype)
weights = gmm['weights'].astype(np_dtype)
elif 'sklearn.mixture.gmm.GMM' in str(type(gmm)):
means = gmm.means_.astype(np_dtype)
covs = gmm.covars_.astype(np_dtype)
weights = gmm.weights_.astype(np_dtype)
else:
print('Unknown type for the prior: {}, exiting!'.format(type(gmm)))
sys.exit(-1)
self.register_buffer('means', torch.tensor(means, dtype=dtype))
self.register_buffer('covs', torch.tensor(covs, dtype=dtype))
precisions = [np.linalg.inv(cov) for cov in covs]
precisions = np.stack(precisions).astype(np_dtype)
self.register_buffer('precisions', torch.tensor(precisions,
dtype=dtype))
# The constant term:
sqrdets = np.array([(np.sqrt(np.linalg.det(c)))
for c in gmm['covars']])
const = (2 * np.pi)**(69 / 2.)
nll_weights = np.asarray(gmm['weights'] / (const *
(sqrdets / sqrdets.min())))
nll_weights = torch.tensor(nll_weights, dtype=dtype).unsqueeze(dim=0)
self.register_buffer('nll_weights', nll_weights)
weights = torch.tensor(gmm['weights'], dtype=dtype).unsqueeze(dim=0)
self.register_buffer('weights', weights)
self.register_buffer('pi_term',
torch.log(torch.tensor(2 * np.pi, dtype=dtype)))
cov_dets = [
np.log(np.linalg.det(cov.astype(np_dtype)) + epsilon)
for cov in covs
]
self.register_buffer('cov_dets', torch.tensor(cov_dets, dtype=dtype))
# The dimensionality of the random variable
self.random_var_dim = self.means.shape[1]
def get_mean(self):
"""Returns the mean of the mixture."""
mean_pose = torch.matmul(self.weights, self.means)
return mean_pose
def merged_log_likelihood(self, pose):
diff_from_mean = pose.unsqueeze(dim=1) - self.means
prec_diff_prod = torch.einsum('mij,bmj->bmi',
[self.precisions, diff_from_mean])
diff_prec_quadratic = (prec_diff_prod * diff_from_mean).sum(dim=-1)
curr_loglikelihood = 0.5 * diff_prec_quadratic - \
torch.log(self.nll_weights)
# curr_loglikelihood = 0.5 * (self.cov_dets.unsqueeze(dim=0) +
# self.random_var_dim * self.pi_term +
# diff_prec_quadratic
# ) - torch.log(self.weights)
min_likelihood, _ = torch.min(curr_loglikelihood, dim=1)
return min_likelihood
def log_likelihood(self, pose):
"""Create graph operation for negative log-likelihood calculation."""
likelihoods = []
for idx in range(self.num_gaussians):
mean = self.means[idx]
prec = self.precisions[idx]
cov = self.covs[idx]
diff_from_mean = pose - mean
curr_loglikelihood = torch.einsum('bj,ji->bi',
[diff_from_mean, prec])
curr_loglikelihood = torch.einsum(
'bi,bi->b', [curr_loglikelihood, diff_from_mean])
cov_term = torch.log(torch.det(cov) + self.epsilon)
curr_loglikelihood += 0.5 * (cov_term +
self.random_var_dim * self.pi_term)
likelihoods.append(curr_loglikelihood)
log_likelihoods = torch.stack(likelihoods, dim=1)
min_idx = torch.argmin(log_likelihoods, dim=1)
weight_component = self.nll_weights[:, min_idx]
weight_component = -torch.log(weight_component)
return weight_component + log_likelihoods[:, min_idx]
def forward(self,
body_pose,
loss_weight_override=None,
reduction_override=None):
assert reduction_override in (None, 'none', 'mean', 'sum')
reduction = (reduction_override
if reduction_override else self.reduction)
loss_weight = (loss_weight_override if loss_weight_override is not None
else self.loss_weight)
if self.use_merged:
pose_prior_loss = self.merged_log_likelihood(body_pose)
else:
pose_prior_loss = self.log_likelihood(body_pose)
pose_prior_loss = loss_weight * pose_prior_loss
if reduction == 'mean':
pose_prior_loss = pose_prior_loss.mean()
elif reduction == 'sum':
pose_prior_loss = pose_prior_loss.sum()
return pose_prior_loss