def _e_step(Xt, Wt, T, L, N, U, B, z0, psi0, sgm0, sgmO, sgmR, sgmV):
# matricize U and B
matU = kronecker(U, reverse=True)
matB = kronecker(B, reverse=True)
# workspace
Lp = np.prod(L)
Np = np.prod(N)
P = np.zeros((T, Lp, Lp))
J = np.zeros((T, Lp, Lp))
mu_ = np.zeros((T, Lp))
psi = np.zeros((T, Lp, Lp))
mu_h = np.zeros((T, Lp))
psih = np.zeros((T, Lp, Lp))
# forward algorithm
for t in trange(T, desc='forward'):
ot = tensor_to_vec(
Wt[t]) # indices of the observed entries of a tensor X
xt = tensor_to_vec(Xt[t])[ot]
lt = sum(ot) # of observed samples
Ht = matU[ot, :]
if t == 0:
K = sgm0 * Ht.T @ pinv(sgm0 * Ht @ Ht.T + sgmR * np.eye(lt))
# K = psi0 @ Ht.T @ pinv(Ht @ psi0 @ Ht.T + sgmR * np.eye(lt))
# psi[0] = sgm0 * np.eye(Lp) - K[0] @ Ht
psi[0] = sgm0 * (np.eye(Lp) - K[0] @ Ht)
# psi[0] = (np.eye(Lp) - K @ Ht) @ psi0
mu_[0] = z0 + K @ (xt - Ht @ z0)
else:
P[t - 1] = matB @ psi[t - 1] @ matB.T + sgmO * np.eye(Lp)
K = P[t - 1] @ Ht.T @ pinv(Ht @ P[t - 1] @ Ht.T +
sgmR * np.eye(lt))
mu_[t] = matB @ mu_[t - 1] + K @ (xt - Ht @ matB @ mu_[t - 1])
psi[t] = (np.eye(Lp) - K @ Ht) @ P[t - 1]
# backward
mu_h[-1] = mu_[-1]
psih[-1] = psi[-1]
for t in tqdm(list(reversed(range(T - 1))), desc='backward'):
J[t] = psi[t] @ matB.T @ pinv(P[t])
mu_h[t] = mu_[t] + J[t] @ (mu_h[t + 1] - matB @ mu_[t])
psih[t] = psi[t] + J[t] @ (psih[t + 1] - P[t]) @ J[t].T
# compute expectations
ztt = np.zeros((T, Lp, Lp))
zt_ = np.zeros((T, Lp, Lp))
cov_zt_ = np.zeros((T, Lp, Lp))
for t in trange(T, desc='compute expectations'):
if t > 0:
cov_zt_[t] = psih[t] @ J[t - 1].T
zt_[t] = cov_zt_[t] + np.outer(mu_h[t], mu_h[t - 1])
ztt[t] = psih[t] + np.outer(mu_h[t], mu_h[t])
zt = mu_h
cov_ztt = psih
return zt, cov_ztt, cov_zt_, ztt, zt_
def predict(self, X, Y):
"""Compute the HOPLS for X and Y wrt the parameters R, Ln and Km.
Parameters:
X: tensorly Tensor, The tensor we wish to do a prediction from.
Of shape [i1, ... iN], N >= 3.
Y: tensorly Tensor, used only for the shape of the prediction.
Returns:
Y_pred: tensorly Tensor, The predicted Y from the model.
"""
_, Q, D, _, W = self.model
best_q2 = +np.inf
if len(Y.shape) > 2:
Q_star = []
for r in range(self.R):
Qkron = kronecker([Q[r][self.M - m - 1] for m in range(self.M)])
Q_star.append(torch.mm(matricize(D[r][np.newaxis, ...]), Qkron.t()))
Q_star = torch.cat(Q_star)
q2s = []
for r in range(1, self.R + 1):
if len(Y.shape) == 2:
Q_star = torch.mm(D[:r, :r], Q[:, :r].t())
inter = torch.mm(W[:, :r], Q_star[:r])
Y_pred = np.reshape(torch.mm(matricize(X), inter), Y.shape, order="F")
Q2 = metric_nmse(Y.numpy(), Y_pred.numpy())
q2s.append(Q2)
if Q2 < best_q2:
best_q2 = Q2
best_r = r
best_Y_pred = Y_pred
return best_Y_pred, best_r, q2s
def _m_step(Xt, Wt, T, S, L, N, U, B, z0, sgm0, sgmO, sgmR, sgmV, xi, _lambda,
Ev, Evv, Ez, Ezzt, Ezz_, mode):
"""
Eq. (12)
"""
Umat = kronecker(U, reverse=True)
Bmat = kronecker(B, reverse=True)
Lm = np.prod(L)
z0_new = deepcopy(Ez[0])
psi0_new = Ezzt[0] - np.outer(Ez[0], Ez[0])
sgm0_new = np.trace(Ezzt[0] - np.outer(Ez[0], Ez[0])) / Lm
if sgm0_new < 1.e-7:
sgm0_new = 1.e-7
res = np.trace(
sum(Ezzt[1:]) - sum(Ezz_[1:]) @ Bmat.T - (sum(Ezz_[1:]) @ Bmat.T).T +
Bmat @ sum(Ezzt[:-1]) @ Bmat.T)
sgmO_new = res / ((T - 1) * Lm)
res = 0
for t in range(T):
Wvec = Wt[t].reshape(-1)
Xvec = Xt[t].reshape(-1)[Wvec]
# Wvec = tensor_to_vec(Wt[t])
# Xvec = tensor_to_vec(Xt[t])[Wvec]
Uobs = Umat[Wvec, :]
res += np.trace(Uobs @ Ezzt[t] @ Uobs.T)
res += Xvec @ Xvec - 2 * Xvec @ (Uobs @ Ez[t])
print(res, Wt.sum())
sgmR_new = res / Wt.sum()
if _lambda > 0:
sgmV_new = sum([np.trace(Evv[mode][j]) for j in range(N[mode])])
sgmV_new /= (N[mode] * L[mode])
xi_new = sum([
S[:, j] @ S[:, j] - 2 * S[:, j] @ U[mode] @ Ev[mode][j]
for j in range(N[mode])
])
xi_new += np.trace(U[mode] @ sum(Evv[mode]) @ U[mode].T)
xi_new /= N[mode]**2
else:
sgmV_new = sgmV[mode]
xi_new = xi[mode]
return z0_new, psi0_new, sgm0_new, sgmO_new, sgmR_new, sgmV_new, xi_new
def update_observation_tensor(mode, Xt, Wt, T, S, L, M, N, U, _lambda, EZ, Ev,
Evv, cov_ZZt, z0, sgm0, sgmO, sgmR, sgmV, xi):
"""
Eq. (19), (20)
"""
G = kronecker(U, skip_matrix=mode, reverse=True).T
for i in trange(N[mode], desc=f"update U[{mode}]"):
A_11, A_12, A_21, A_22 = _compute_A(_lambda, mode, i, G, Xt, Wt, T, S,
L, M, N, EZ, Ev, Evv, cov_ZZt)
numer = _lambda * A_11 / xi + (1 - _lambda) * A_12 / sgmR
denom = _lambda * A_21 / xi + (1 - _lambda) * A_22 / sgmR
U[mode][i, :] = numer @ pinv(denom)
return U[mode]
def Tensor_MM_kron(Tx, modej, U_list):
"""
Calculate X_{[j]}U^{(-modej)}T
Tx: input tensor
modej: index for mode j; here consider mode from 0, 1, ..., d - 1
U_list: a list of projection matrix, length = d
"""
U_temp = U_list[:]
U_temp.pop(modej) #Delete first, then reverse and do kroncker product.
U_temp = U_temp[::-1]
res = kronecker(U_temp)
les = tl.unfold(Tx, mode=modej)
ans = np.dot(les, res.T)
return ans
def update_transition_tensor(mode, L, B, covZZt, covZZ_, EZ):
T = len(covZZt)
F = kronecker(B, skip_matrix=mode, reverse=True).T
Lm = L[mode]
Ln = int(np.prod(L) / Lm)
C1 = np.zeros((Lm, Lm))
C2 = np.zeros((Lm, Lm))
for t in trange(1, T, desc=f'update B[{mode}]'):
for j in range(Ln):
C1 += _compute_b(F, covZZt[t - 1], j) # t = 1..T-1
C1 += np.outer(EZ[t - 1] @ F[:, j], EZ[t - 1] @ F[:, j])
C2 += _compute_a(F, covZZ_[t], j) # t = 2..T
C2 += np.outer(EZ[t, :, j], EZ[t - 1] @ F[:, j])
return C2 @ pinv(C1)
def predict(self, X, Y):
"""Compute the HOPLS for X and Y wrt the parameters R, Ln and Km.
Parameters:
X: tensorly Tensor, The tensor we wish to do a prediction from.
Of shape [i1, ... iN], N >= 3.
Y: tensorly Tensor, used only for the shape of the prediction.
Returns:
Y_pred: tensorly Tensor, The predicted Y from the model.
"""
_, Q, D, _, W = self.model
best_q2 = -np.inf
if len(Y.shape) > 2:
Q_star = []
for r in range(self.R):
Qkron = kronecker(
[Q[r][self.M - m - 1] for m in range(self.M)])
Q_star.append(
np.matmul(matricize(D[r][np.newaxis, ...]), Qkron.T))
Q_star = np.concatenate(Q_star)
for r in range(1, self.R + 1):
if len(Y.shape) == 2:
Q_star = np.matmul(D[:r, :r], Q[:, :r].T)
inter = np.matmul(W[:, :r], Q_star[:r])
Y_pred = np.reshape(np.matmul(matricize(X), inter),
Y.shape,
order="F")
Q2 = qsquared(Y, Y_pred)
if Q2 > best_q2:
best_q2 = Q2
best_r = r
best_Y_pred = Y_pred
return Y_pred, best_r, best_q2
T = X.shape[-1]
M = X.ndim - 1
z = np.loadtxt(outdir + "vec_z.txt")
U = [np.loadtxt(outdir + f"U_{i}.txt") for i in range(M)]
ranks = [U[i].shape[1] for i in range(M)]
Z = np.zeros((T, *ranks))
for t in range(T):
Z[t] = z[t].reshape(ranks)
plt.plot(z)
plt.show()
for m in range(M):
pred = np.zeros((T, X.shape[m]))
for t in range(T):
print(unfold(Z[t], m).shape, U[m].shape)
print(kronecker(U, skip_matrix=m, reverse=True).T.shape)
X_n = U[m] @ unfold(Z[t], m) @ kronecker(
U, skip_matrix=m, reverse=True).T
pred[t] = X_n[:, m]
# plt.plot((X, -1)[:, i])
plt.plot(pred)
plt.show()
matU = kronecker(U, reverse=True)
predict = z @ matU.T
print(predict.shape)
print(unfold(X, -1).shape)
for i in range(unfold(X, -1).shape[1]):
plt.plot(unfold(X, -1)[:, i])
plt.plot(predict[:, i])
plt.show()
def _fit_2d(self, X, Y):
"""
Compute the HOPLS for X and Y wrt the parameters R, Ln and Km for the special case mode_Y = 2.
Parameters:
X: tensorly Tensor, The target tensor of shape [i1, ... iN], N = 2.
Y: tensorly Tensor, The target tensor of shape [j1, ... jM], M >= 3.
Returns:
G: Tensor, The core Tensor of the HOPLS for X, of shape (R, L2, ..., LN).
P: List, The N-1 loadings of X.
D: Tensor, The core Tensor of the HOPLS for Y, of shape (R, K2, ..., KN).
Q: List, The N-1 loadings of Y.
ts: Tensor, The latent vectors of the HOPLS, of shape (i1, R).
"""
# Initialization
Er, Fr = X, Y
P, T, W, Q = [], [], [], []
D = tl.zeros((self.R, self.R))
G = []
# Beginning of the algorithm
# Gr, _ = tucker(Er, ranks=[1] + self.Ln)
for r in range(self.R):
if torch.norm(Er) > self.epsilon and torch.norm(Fr) > self.epsilon:
# computing the covariance
Cr = mode_dot(Er, Fr.t(), 0)
# HOOI tucker decomposition of C
Gr_C, latents = tucker(Cr, rank=[1] + self.Ln)
# Getting P and Q loadings
qr = latents[0]
qr /= torch.norm(qr)
# Pr = latents[1:]
Pr = [a / torch.norm(a) for a in latents[1:]]
P.append(Pr)
tr = multi_mode_dot(Er, Pr, list(range(1, len(Pr) + 1)), transpose=True)
# Gr_pi = torch.pinverse(matricize(Gr))
# tr = torch.mm(matricize(tr), Gr_pi)
GrC_pi = torch.pinverse(matricize(Gr_C))
tr = torch.mm(matricize(tr), GrC_pi)
tr /= torch.norm(tr)
# recomposition of the core tensor of Y
ur = torch.mm(Fr, qr)
dr = torch.mm(ur.t(), tr)
D[r, r] = dr
Pkron = kronecker([Pr[self.N - n - 1] for n in range(self.N)])
# P.append(torch.mm(matricize(Gr), Pkron.t()).t())
# W.append(torch.mm(Pkron, Gr_pi))
Q.append(qr)
T.append(tr)
Gd = tl.tucker_to_tensor([Er, [tr] + Pr], transpose_factors=True)
Gd_pi = torch.pinverse(matricize(Gd))
W.append(torch.mm(Pkron, Gd_pi))
# Deflation
# X_hat = torch.mm(torch.cat(T, dim=1), torch.cat(P, dim=1).t())
# Er = X - np.reshape(X_hat, (Er.shape), order="F")
Er = Er - tl.tucker_to_tensor([Gd, [tr] + Pr])
Fr = Fr - dr * torch.mm(tr, qr.t())
else:
break
Q = torch.cat(Q, dim=1)
T = torch.cat(T, dim=1)
# P = torch.cat(P, dim=1)
W = torch.cat(W, dim=1)
self.model = (P, Q, D, T, W)
return self
def fit(self, X, Y):
"""
Compute the HOPLS for X and Y wrt the parameters R, Ln and Km.
Parameters:
X: tensorly Tensor, The target tensor of shape [i1, ... iN], N >= 3.
Y: tensorly Tensor, The target tensor of shape [j1, ... jM], M >= 3.
Returns:
G: Tensor, The core Tensor of the HOPLS for X, of shape (R, L2, ..., LN).
P: List, The N-1 loadings of X.
D: Tensor, The core Tensor of the HOPLS for Y, of shape (R, K2, ..., KN).
Q: List, The N-1 loadings of Y.
ts: Tensor, The latent vectors of the HOPLS, of shape (i1, R).
"""
# check parameters
X_mode = len(X.shape)
Y_mode = len(Y.shape)
assert Y_mode >= 2, "Y need to be mode 2 minimum."
assert X_mode >= 3, "X need to be mode 3 minimum."
assert (
len(self.Ln) == X_mode - 1
), f"The list of ranks for the decomposition of X (Ln) need to be equal to the mode of X -1: {X_mode-1}."
if Y_mode == 2:
return self._fit_2d(X, Y)
assert (
len(self.Km) == Y_mode - 1
), f"The list of ranks for the decomposition of Y (Km) need to be equal to the mode of Y -1: {Y_mode-1}."
# Initialization
Er, Fr = X, Y
In = X.shape
T, G, P, Q, D, W = [], [], [], [], [], []
# Beginning of the algorithm
for r in range(self.R):
if torch.norm(Er) > self.epsilon and torch.norm(Fr) > self.epsilon:
Cr = torch.Tensor(np.tensordot(Er, Fr, (0, 0)))
# HOOI tucker decomposition of C
_, latents = tucker(Cr, ranks=self.Ln + self.Km)
# Getting P and Q loadings
Pr = latents[: len(Er.shape) - 1]
Qr = latents[len(Er.shape) - 1 :]
# computing product of Er by latents of X
tr = multi_mode_dot(Er, Pr, list(range(1, len(Pr))), transpose=True)
# Getting t as the first leading left singular vector of the product
tr = torch.svd(matricize(tr))[0][:, 0]
tr = tr[..., np.newaxis]
# recomposition of the core tensors
Gr = tl.tucker_to_tensor(Er, [tr] + Pr, transpose_factors=True)
Dr = tl.tucker_to_tensor(Fr, [tr] + Qr, transpose_factors=True)
Pkron = kronecker([Pr[self.N - n - 1] for n in range(self.N)])
Gr_pi = torch.pinverse(matricize(Gr))
W.append(torch.mm(Pkron, Gr_pi))
# Gathering of
P.append(Pr)
Q.append(Qr)
G.append(Gr)
D.append(Dr)
T.append(tr)
# Deflation
Er = Er - tl.tucker_to_tensor(Gr, [tr] + Pr)
Fr = Fr - tl.tucker_to_tensor(Dr, [tr] + Qr)
else:
break
T = torch.cat(T, dim=1)
W = torch.cat(W, dim=1)
self.model = (P, Q, D, T, W)
return self
def compose_tensor(n, vec): vecs = n_unit_norm_vecs(n, vec.shape[0]) print(type(vecs)) print(vecs) return kronecker(np.array(vecs, vecs))
def reconstruct_matrix(U, Z, mode):
# Lemma 3.2
index = np.ones(len(U), dtype=bool)
index[mode] = False
return U[mode] @ Z @ kronecker(U[ind][::-1]).T
