[英]Einstein summation in numpy for non-linear function
我正在嘗試建立以下非線性成本函數(1)和(2)
通過使用numpy einsum函數。
我嘗試將它們轉換為python,其中R
是形狀(14,3,3)的數組, v
, vt
(從等式(1)表示v')的形狀為(14,3)。 分別表示p
和p'
old_points
和new_points
的形狀為(6890,3)。
def costfunc(x, old_points, new_points, weights, n_joints):
"""
Set up non-linear cost functions by using equations from LBS:
(1) p'_i = sum{j}(w_ji (R_j p_i + v_j))
(2) p_i - sum{j}(w_ji (R_j p'_i + v'_j))
where Rt denotes the transpose of R.
:param old_points: original vertex positions
:param new_points: transformed vertex positions
:param weights: weight matrix obtained from spectral clustering
:param n_joints: number of joints
:return: non-linear cost functions as in (1), (2) to find the root of
"""
# Extract rotations R, Rt and offsets v, v' from rv
R = np.array([(np.array(x[j * 15:j * 15 + 9]).reshape(3, 3)) for j in range(n_joints)])
Rt = np.array([R[j].T for j in range(n_joints)])
v = np.array([(np.array(x[j * 15 + 9:j * 15 + 12])) for j in range(n_joints)])
vt = np.array([(np.array(x[j * 15 + 12:j * 15 + 15])) for j in range(n_joints)])
## Use equations (1) and (2) for the non-linear pass.
# R_j p_i
Rp = np.einsum('jkl,il', R, old_points)
Rtv = np.einsum('jkl,il', Rt, v)
# Rt_j p'_i
Rtp = np.einsum('jkl,il', Rt, new_points)
Rvt = np.einsum('jkl,il', R, vt)
# w_ji (Rp_ij - Rtv_j)
wRpv = np.einsum('ji,ijk->ik', weights, Rp - Rvt)
# w_ji (Rtp'_ij - Rv'_j)
wRtpv = np.einsum('ji,ijk->ik', weights, Rtp - Rtv)
# Set up a non-linear cost function, then compute the squared norm.
d = new_points - wRpv
dt = old_points - wRtpv
norm = np.linalg.norm(d, axis=1)
normt = np.linalg.norm(dt, axis=1)
result = np.concatenate([norm, normt])
return np.power(result, 2)
目前,在計算wRpv
和wRtpv
的行中存在錯誤ValueError: operands could not be broadcast together with shapes (6890,14,3) (14,14,3)
。 我該如何解決? 很感謝任何形式的幫助!
我知道了 這是解決方案:
def costfunc(x, old_points, new_points, weights, n_joints):
"""
Set up non-linear cost functions by using equations from LBS:
(1) p'_i = sum{j}(w_ji (R_j p_i + v_j))
(2) p_i - sum{j}(w_ji (R_j p'_i + v'_j))
where Rt denotes the transpose of R.
:param old_points: original vertex positions
:param new_points: transformed vertex positions
:param weights: weight matrix obtained from spectral clustering
:param n_joints: number of joints
:return: non-linear cost functions as in (1), (2) to find the root of
"""
# Extract rotations R, Rt and offsets v, v' from rv
R = np.array([(np.array(x[j * 15:j * 15 + 9]).reshape(3, 3)) for j in range(n_joints)])
Rt = np.array([R[j].T for j in range(n_joints)])
v = np.array([(np.array(x[j * 15 + 9:j * 15 + 12])) for j in range(n_joints)])
vt = np.array([(np.array(x[j * 15 + 12:j * 15 + 15])) for j in range(n_joints)])
## Use equations (1) and (2) for the non-linear pass.
# R_j p_i
Rp = np.einsum('jkl,il', R, old_points)
# Rt_j p'_i
Rtp = np.einsum('jkl,il', Rt, new_points)
# R_j v'_j
Rvt = np.array([R[i] @ vt[i] for i in range(n_joints)])
# Rt_j v_j
Rtv = np.array([Rt[i] @ v[i] for i in range(n_joints)])
# w_ji (Rp_ij - Rtv_j)
wRpv = np.einsum('ji,ijk->ik', weights, Rp - Rvt)
# w_ji (Rtp'_ij - Rv'_j)
wRtpv = np.einsum('ji,ijk->ik', weights, Rtp - Rtv)
# Set up a non-linear cost function, then compute the squared norm.
d = new_points - wRpv
dt = old_points - wRtpv
norm = np.linalg.norm(d, axis=1)
normt = np.linalg.norm(dt, axis=1)
result = np.concatenate([norm, normt])
return np.power(result, 2)
聲明:本站的技術帖子網頁,遵循CC BY-SA 4.0協議,如果您需要轉載,請注明本站網址或者原文地址。任何問題請咨詢:yoyou2525@163.com.