def als(tensor,rank,factors=None,it_max=100,tol=1e-7,list_factors=False,error_fast=True,time_rec=False):
"""
ALS methode of CP decomposition
Parameters
----------
tensor : tensor
rank : int
factors : list of matrices, optional
an initial factor matrices. The default is None.
it_max : int, optional
maximal number of iteration. The default is 100.
tol : float, optional
error tolerance. The default is 1e-7.
list_factors : boolean, optional
If true, then return factor matrices of each iteration. The default is False.
error_fast : boolean, optional
If true, use err_fast to compute data fitting error, otherwise, use err. The default is True.
time_rec : boolean, optional
If true, return computation time of each iteration. The default is False.
Returns
-------
the CP decomposition, number of iteration and termination criterion.
list_fac and list_time are optional.
"""
N=tl.ndim(tensor) # order of tensor
norm_tensor=tl.norm(tensor) # norm of tensor
if time_rec == True : list_time=[]
if list_factors==True : list_fac=[] # list of factor matrices
if (factors==None): factors=svd_init_fac(tensor,rank)
weights=None
it=0
if list_factors==True : list_fac.append(copy.deepcopy(factors))
error=[err(tensor,weights,factors)/norm_tensor]
while (error[len(error)-1]>tol and it<it_max):
if time_rec == True : tic=time.time()
for n in range(N):
V=np.ones((rank,rank))
for i in range(len(factors)):
if i != n : V=V*tl.dot(tl.transpose(factors[i]),factors[i])
W=tl.cp_tensor.unfolding_dot_khatri_rao(tensor, (None,factors), n)
factors[n]= tl.transpose(tl.solve(tl.transpose(V),tl.transpose(W)))
if list_factors==True : list_fac.append(copy.deepcopy(factors))
it=it+1
if (error_fast==False) : error.append(err(tensor,weights,factors)/norm_tensor)
else : error.append(err_fast(norm_tensor,factors[N-1],V,W)/norm_tensor)
if time_rec == True :
toc=time.time()
list_time.append(toc-tic)
# weights,factors=tl.cp_tensor.cp_normalize((None,factors))
if list_factors==True and time_rec==True: return(weights,factors,it,error,list_fac,list_time)
if time_rec==True : return(weights,factors,it,error,list_time)
if list_factors==True : return(weights,factors,it,error,list_fac)
return(weights,factors,it,error)
def test_err():
"""
Test of err
"""
# create a kruskal tensor
# factor matrices
A = np.arange(9).reshape(3, 3)
B = np.arange(6).reshape(2, 3) + 9
C = np.arange(6).reshape(2, 3) + 15
factors = []
factors += [A]
factors += [B]
factors += [C]
t_krus = tl.cp_to_tensor((None, factors))
weights_cp, factors_cp = tl.cp_normalize((None, factors))
print(base.err(t_krus, weights_cp, factors_cp))
def nn_her_Als(tensor,
rank,
factors=None,
it_max=100,
tol=1e-7,
beta=0.5,
eta=1.5,
gamma=1.05,
gamma_bar=1.01,
list_factors=False,
error_fast=True,
time_rec=False):
"""
her ALS methode of CP decomposition for non negative case
Parameters
----------
tensor : tensor
rank : int
factors : list of matrices, optional
an initial non negative factor matrices. The default is None.
it_max : int, optional
maximal number of iteration. The default is 100.
tol : float, optional
error tolerance. The default is 1e-7.
beta : float, optional
extrapolation parameter. The default is 0.5.
eta : float, optional
decrease coefficient of beta. The default is 1.5.
gamma : float, optional
increase coefficient of beta. The default is 1.05.
gamma_bar : float, optional
increase coeefficient of beta_bar. The default is 1.01.
list_factors : boolean, optional
If true, then return factor matrices of each iteration. The default is False.
error_fast : boolean, optional
If true, use err_fast to compute data fitting error, otherwise, use err. The default is True.
time_rec : boolean, optional
If true, return computation time of each iteration. The default is False.
Returns
-------
the CP decomposition, number of iteration, error and restart pourcentage.
list_fac and list_time are optional.
"""
beta_bar = 1
N = tl.ndim(tensor) # order of tensor
norm_tensor = tl.norm(tensor) # norm of tensor
weights = None
if time_rec == True: list_time = []
if list_factors == True: list_fac = []
if (factors == None): factors = svd_init_fac(tensor, rank)
# Initialization of factor hat matrices by factor matrices
factors_hat = factors
if list_factors == True: list_fac.append(copy.deepcopy(factors))
it = 0
cpt = 0
F_hat_bf = err(tensor, None, factors) # cost
error = [F_hat_bf / norm_tensor]
while (error[len(error) - 1] > tol and it < it_max):
if time_rec == True: tic = time.time()
for n in range(N):
V = np.ones((rank, rank))
for i in range(len(factors)):
if i != n:
V = V * tl.dot(tl.transpose(factors_hat[i]),
factors_hat[i])
W = tl.cp_tensor.unfolding_dot_khatri_rao(tensor,
(None, factors_hat), n)
factor_bf = factors[n]
# update
fac, _, _, _ = hals_nnls(tl.transpose(W), V,
tl.transpose(factors[n]))
factors[n] = tl.transpose(fac)
# extrapolate
factors_hat[n] = tl.clip(factors[n] + beta *
(factors[n] - factor_bf),
a_min=0.0)
if (error_fast == False):
F_hat_new = err(tensor, None, factors_hat) # cost update
else:
F_hat_new = err_fast(norm_tensor, factors[N - 1], V, W)
if (F_hat_new > F_hat_bf):
factors_hat = factors
beta_bar = beta
beta = beta / eta
cpt = cpt + 1
else:
factors = factors_hat
beta_bar = min(1, beta_bar * gamma_bar)
beta = min(beta_bar, gamma * beta)
F_hat_bf = F_hat_new
it = it + 1
if list_factors == True: list_fac.append(copy.deepcopy(factors))
error.append(F_hat_new / norm_tensor)
if time_rec == True:
toc = time.time()
list_time.append(toc - tic)
if time_rec == True and list_factors == True:
return (weights, factors, it, error, cpt / it, list_fac, list_time)
if list_factors == True:
return (weights, factors, it, error, cpt / it, list_fac)
if time_rec == True:
return (weights, factors, it, error, cpt / it, list_time)
return (weights, factors, it, error, cpt / it)
def her_CPRAND(tensor,
rank,
n_samples,
factors=None,
exact_err=True,
it_max=100,
err_it_max=20,
tol=1e-7,
beta=0.1,
eta=3,
gamma=1.01,
gamma_bar=1.005,
list_factors=False,
time_rec=False):
"""
herCPRAND for CP-decomposition
return also exact error
Parameters
----------
tensor : tensor
rank : int
n_samples : int
sample size
factors : list of matrices, optional
an initial factor matrices. The default is None.
exact_err : boolean, optional
whether use err or err_rand_fast for terminaison criterion. The default is False.
(not useful for this version)
it_max : int, optional
maximal number of iteration. The default is 100.
err_it_max : int, optional
maximal of iteration if terminaison critirion is not improved. The default is 20.
tol : float, optional
error tolerance. The default is 1e-7.
beta : float, optional
extrapolation parameter. The default is 0.1.
eta : float, optional
decrease coefficient of beta. The default is 3.
gamma : float, optional
increase coefficient of beta. The default is 1.01.
gamma_bar : float, optional
increase coeefficient of beta_bar. The default is 1.005.
list_factors : boolean, optional
If true, then return factor matrices of each iteration. The default is False.
time_rec : boolean, optional
If true, return computation time of each iteration. The default is False.
Returns
-------
the CP decomposition, number of iteration and exact / estimated termination criterion.
list_fac and list_time are optional.
"""
beta_bar = 1
N = tl.ndim(tensor) # order of tensor
norm_tensor = tl.norm(tensor) # norm of tensor
if list_factors == True: list_fac = []
if (time_rec == True): list_time = []
if (factors == None): factors = svd_init_fac(tensor, rank)
# Initialization of factor hat matrice by factor matrices
factors_hat = factors
if list_factors == True: list_fac.append(copy.deepcopy(factors))
weights = None
it = 0
err_it = 0
cpt = 0
########################################
######### error initialization #########
########################################
F_hat_bf, ind_bf = err_rand(tensor, None, factors, n_samples)
F_hat_bf_ex = err(tensor, None, factors) # exact cost
rng = tl.random.check_random_state(None)
error = [F_hat_bf / norm_tensor]
error_ex = [F_hat_bf_ex / norm_tensor]
min_err = error[len(error) - 1]
while (min_err > tol and it < it_max and err_it < err_it_max):
if time_rec == True: tic = time.time()
factors_hat_bf = factors_hat
for n in range(N):
Zs, indices = sample_khatri_rao(factors_hat,
n_samples,
skip_matrix=n,
random_state=rng)
indices_list = [i.tolist() for i in indices]
indices_list.insert(n, slice(None, None, None))
indices_list = tuple(indices_list)
V = tl.dot(tl.transpose(Zs), Zs)
# J'ai du mal avec la syntaxe tensor[indices_list],
# Ca renvoie une matrices et non un tenseur?
if (n == 0): sampled_unfolding = tensor[indices_list]
else: sampled_unfolding = tl.transpose(tensor[indices_list])
W = tl.dot(sampled_unfolding, Zs)
factor_bf = factors[n]
# update
factors[n] = tl.transpose(
tl.solve(V, tl.transpose(W))
) # solve needs a squared full rank matrix, if rank>nb_sampls ok
# if (n==N-1) : F_hat_new=tl.norm(tl.dot(Zs,tl.transpose(factors[n]))-sampled_unfolding,2) # cost update
# extrapolate
factors_hat[n] = factors[n] + beta * (factors[n] - factor_bf)
########################################
######### error update #########
########################################
matrices = factors_hat_bf[:-1]
Zs_bf = tl.ones((n_samples, rank), **tl.context(matrices[0]))
for indices, matrix in zip(indices_list, matrices):
Zs_bf = Zs_bf * matrix[indices, :]
V_bf = tl.dot(tl.transpose(Zs_bf), Zs_bf)
W_bf = tl.dot(tl.transpose(tensor[indices_list]), Zs_bf)
F_hat_bf, a = err_rand_fast(tensor, factor_bf, V_bf, W_bf,
indices_list, n_samples)
F_hat_new, _ = err_rand_fast(tensor, factors[N - 1], V, W,
indices_list, n_samples)
if (F_hat_new > F_hat_bf):
factors_hat = factors
beta_bar = beta
beta = beta / eta
cpt = cpt + 1
else:
factors = factors_hat
beta_bar = min(1, beta_bar * gamma_bar)
beta = min(beta_bar, gamma * beta)
########################################
######### update for next it #########
########################################
it = it + 1
if list_factors == True: list_fac.append(copy.deepcopy(factors))
error.append(F_hat_new / norm_tensor)
if (error[len(error) - 1] < min_err):
min_err = error[len(error) - 1] # err update
else:
err_it = err_it + 1
if time_rec == True:
toc = time.time()
list_time.append(toc - tic)
error_ex.append(err(tensor, None, factors) /
norm_tensor) # exact cost update
# weights,factors=tl.cp_normalize((None,factors))
if list_factors == True and time_rec == True:
return (weights, factors, it, error_ex, error, cpt / it, list_fac,
list_time)
if list_factors == True:
return (weights, factors, it, error_ex, error, cpt / it, list_fac)
if time_rec == True:
return (weights, factors, it, error_ex, error, cpt / it, list_time)
return (weights, factors, it, error_ex, error, cpt / it)
def test_err_fast():
"""
Test err_fast for 3 tensors.
plot the terminaison criterion (obtained by err_fast) and exact error.
"""
# create a kruskal tensor
# factor matrices
A = np.arange(9).reshape(3, 3)
B = np.arange(6).reshape(2, 3) + 9
C = np.arange(6).reshape(2, 3) + 15
factors = []
factors += [A]
factors += [B]
factors += [C]
t_krus = tl.cp_to_tensor((None, factors))
weights, factors, it, error, l = als(t_krus, 3, list_factors=True)
err_ex = []
for i in l:
err_ex += [err(t_krus, weights, i)]
plt.figure(0)
plt.plot(range(len(err_ex)), err_ex / tl.norm(t_krus), 'b-', label="exact")
plt.plot(range(len(error)), error, 'r--', label="err fast")
plt.xlabel('it')
plt.yscale('log')
plt.title('als for t_krus')
plt.ylabel('terminaison criterion')
plt.legend(loc='best')
# create an complicated random tensor
I = 50
J = 50
K = 50
r = 10 # rank
fac_true, noise = init_factors(I, J, K, r, True)
t = tl.cp_to_tensor((None, fac_true)) + noise
weights, factors, it, error, l = als(t, r, list_factors=True)
err_ex = []
for i in l:
err_ex += [err(t, weights, i)]
plt.figure(1)
plt.plot(range(len(err_ex)), err_ex / tl.norm(t), 'b-', label="exact")
plt.plot(range(len(error)), error, 'r--', label="err fast")
plt.xlabel('it')
plt.yscale('log')
plt.title('als for complicated case')
plt.ylabel('terminaison criterion')
plt.legend(loc='best')
# create a simple random tensor
fac_true, noise = init_factors(I, J, K, r, False)
t = tl.cp_to_tensor((None, fac_true)) + noise
weights, factors, it, error, l = als(t, r, list_factors=True)
err_ex = []
for i in l:
err_ex += [err(t, weights, i)]
plt.figure(2)
plt.plot(range(len(err_ex)), err_ex / tl.norm(t), 'b-', label="exact")
plt.plot(range(len(error)), error, 'r--', label="err fast")
plt.xlabel('it')
plt.yscale('log')
plt.title('als for simple case')
plt.ylabel('terminaison criterion')
plt.legend(loc='best')
def CPRAND(tensor,
rank,
n_samples,
factors=None,
exact_err=False,
it_max=100,
err_it_max=20,
tol=1e-7,
list_factors=False,
time_rec=False):
"""
CPRAND for CP-decomposition
return also exact error
Parameters
----------
tensor : tensor
rank : int
n_samples : int
sample size
factors : list of matrices, optional
initial factor matrices. The default is None.
exact_err : boolean, optional
whether use err or err_rand_fast for terminaison criterion. The default is False.
(not useful for this version)
it_max : int, optional
maximal number of iteration. The default is 100.
err_it_max : int, optional
maximal of iteration if terminaison critirion is not improved. The default is 20.
tol : float, optional
error tolerance. The default is 1e-7.
list_factors : boolean, optional
If true, then return factor matrices of each iteration. The default is False.
time_rec : boolean, optional
If true, return computation time of each iteration. The default is False.
Returns
-------
the CP decomposition, number of iteration and exact / estimated termination criterion.
list_fac and list_time are optional.
"""
N = tl.ndim(tensor) # order of tensor
norm_tensor = tl.norm(tensor) # norm of tensor
if list_factors == True: list_fac = []
if time_rec == True: list_time = []
if (factors == None): factors = svd_init_fac(tensor, rank)
# weights,factors=tl.cp_tensor.cp_normalize((None,factors))
if list_factors == True: list_fac.append(copy.deepcopy(factors))
weights = None
it = 0
err_it = 0
########################################
######### error initialization #########
########################################
temp, ind_err = err_rand(tensor, weights, factors, 400)
error = [temp / norm_tensor]
error_ex = [err(tensor, weights, factors) / norm_tensor]
min_err = error[len(error) - 1]
rng = tl.random.check_random_state(None)
while (min_err > tol and it < it_max and err_it < err_it_max):
if time_rec == True: tic = time.time()
for n in range(N):
Zs, indices = sample_khatri_rao(factors,
n_samples,
skip_matrix=n,
random_state=rng)
indices_list = [i.tolist() for i in indices]
indices_list.insert(n, slice(None, None, None))
indices_list = tuple(indices_list)
if (n == 0): sampled_unfolding = tensor[indices_list]
else: sampled_unfolding = tl.transpose(tensor[indices_list])
V = tl.dot(tl.transpose(Zs), Zs)
W = tl.dot(sampled_unfolding, Zs)
# update
factors[n] = tl.transpose(tl.solve(
V, tl.transpose(W))) # solve needs a squared matrix
if list_factors == True: list_fac.append(copy.deepcopy(factors))
it = it + 1
################################
######### error update #########
################################
error.append(
err_rand_fast(tensor, factors[N - 1], V, W, indices_list,
n_samples)[0] / norm_tensor
) # same indices used as for Random Lesat Square Calculation
if (error[len(error) - 1] < min_err):
min_err = error[len(error) - 1] # err update
else:
err_it = err_it + 1
if time_rec == True:
toc = time.time()
list_time.append(toc - tic)
error_ex.append(err(tensor, weights, factors) / norm_tensor)
# weights,factors=tl.cp_tensor.cp_normalize((None,factors))
if time_rec == True and list_factors == True:
return (weights, factors, it, error_ex, error, list_fac, list_time)
if list_factors == True:
return (weights, factors, it, error_ex, error, list_fac)
if time_rec == True:
return (weights, factors, it, error_ex, error, list_time)
return (weights, factors, it, error_ex, error)
def her_CPRAND5(tensor,
rank,
n_samples,
factors=None,
exact_err=True,
it_max=100,
err_it_max=20,
tol=1e-7,
beta=0.1,
eta=3,
gamma=1.01,
gamma_bar=1.005,
list_factors=False,
time_rec=False):
"""
different err sample taking mean value
"""
beta_bar = 1
N = tl.ndim(tensor) # order of tensor
norm_tensor = tl.norm(tensor) # norm of tensor
if list_factors == True: list_fac = []
if (time_rec == True): list_time = []
if (factors == None): factors = svd_init_fac(tensor, rank)
# Initialization of factor hat matrice by factor matrices
factors_hat = factors
if list_factors == True: list_fac.append(copy.deepcopy(factors))
list_F_hat_bf = []
weights = None
it = 0
err_it = 0
cpt = 0
n_samples_err = 400 # assuming that miu max = 1
########################################
######### error initialization #########
########################################
F_hat_bf, ind_bf = err_rand(tensor, None, factors, n_samples_err)
list_F_hat_bf.append(F_hat_bf)
F_hat_bf_ex = err(tensor, None, factors) # exact cost
rng = tl.check_random_state(None)
error = [F_hat_bf / norm_tensor]
error_ex = [F_hat_bf_ex / norm_tensor]
min_err = error[len(error) - 1]
while (min_err > tol and it < it_max and err_it < err_it_max):
if time_rec == True: tic = time.time()
for n in range(N):
Zs, indices = sample_khatri_rao(factors_hat,
n_samples,
skip_matrix=n,
random_state=rng)
indices_list = [i.tolist() for i in indices]
indices_list.insert(n, slice(None, None, None))
indices_list = tuple(indices_list)
V = tl.dot(tl.transpose(Zs), Zs)
# J'ai du mal avec la syntaxe tensor[indices_list],
# Ca renvoie une matrices et non un tenseur?
if (n == 0): sampled_unfolding = tensor[indices_list]
else: sampled_unfolding = tl.transpose(tensor[indices_list])
W = tl.dot(sampled_unfolding, Zs)
factor_bf = factors[n]
# update
factors[n] = tl.transpose(
tl.solve(V, tl.transpose(W))
) # solve needs a squared full rank matrix, if rank>nb_sampls ok
# if (n==N-1) : F_hat_new=tl.norm(tl.dot(Zs,tl.transpose(factors[n]))-sampled_unfolding,2) # cost update
# extrapolate
factors_hat[n] = factors[n] + beta * (factors[n] - factor_bf)
########################################
######### error update #########
########################################
F_hat_new, _ = err_rand(tensor,
weights,
factors,
n_samples_err,
indices_list=ind_bf)
#
# added
# a new sample
ind_bf = [
np.random.choice(tl.shape(m)[0], n_samples_err) for m in factors
]
ind_bf = [i.tolist() for i in ind_bf]
ind_bf = tuple(ind_bf)
list_F_hat_bf.append(F_hat_new)
if (F_hat_new > F_hat_bf):
factors_hat = factors
beta_bar = beta
beta = beta / eta
cpt = cpt + 1
else:
factors = factors_hat
beta_bar = min(1, beta_bar * gamma_bar)
beta = min(beta_bar, gamma * beta)
########################################
######### update for next it #########
########################################
it = it + 1
if it < 10:
F_hat_bf = np.mean(list_F_hat_bf)
else:
F_hat_bf = np.mean(list_F_hat_bf[(len(list_F_hat_bf) -
10):(len(list_F_hat_bf) - 1)])
if list_factors == True: list_fac.append(copy.deepcopy(factors))
error.append(F_hat_new / norm_tensor)
if (error[len(error) - 1] < min_err):
min_err = error[len(error) - 1] # err update
else:
err_it = err_it + 1
if time_rec == True:
toc = time.time()
list_time.append(toc - tic)
error_ex.append(err(tensor, None, factors) /
norm_tensor) # exact cost update
# weights,factors=tl.cp_normalize((None,factors))
if list_factors == True and time_rec == True:
return (weights, factors, it, error_ex, error, cpt / it, list_fac,
list_time)
if list_factors == True:
return (weights, factors, it, error_ex, error, cpt / it, list_fac)
if time_rec == True:
return (weights, factors, it, error_ex, error, cpt / it, list_time)
return (weights, factors, it, error_ex, error, cpt / it)
def nn_her_CPRAND(tensor,
rank,
n_samples,
n_samples_err=400,
factors=None,
exact_err=False,
it_max=100,
err_it_max=20,
tol=1e-7,
beta=0.1,
eta=3,
gamma=1.01,
gamma_bar=1.005,
list_factors=False,
time_rec=False,
filter=10):
"""
herCPRAND for nonnegative CP-decomposition
same err sample taking mean value of last filter values
Parameters
----------
tensor : tensor
rank : int
n_samples : int
sample size
n_samples_err : int, optional
sample size used for error estimation. The default is 400.
factors : list of matrices, optional
an initial factor matrices. The default is None.
exact_err : boolean, optional
whether use err or err_rand_fast for terminaison criterion. The default is False.
it_max : int, optional
maximal number of iteration. The default is 100.
err_it_max : int, optional
maximal of iteration if terminaison critirion is not improved. The default is 20.
tol : float, optional
error tolerance. The default is 1e-7.
beta : float, optional
extrapolation parameter. The default is 0.5.
eta : float, optional
decrease coefficient of beta. The default is 1.5.
gamma : float, optional
increase coefficient of beta. The default is 1.05.
gamma_bar : float, optional
increase coeefficient of beta_bar. The default is 1.01.
list_factors : boolean, optional
If true, then return factor matrices of each iteration. The default is False.
time_rec : boolean, optional
If true, return computation time of each iteration. The default is False.
filter : int, optional
The filter size used for the mean value
Returns
-------
the CP decomposition, number of iteration, error and restart pourcentage.
list_fac and list_time are optional.
"""
beta_bar = 1
N = tl.ndim(tensor) # order of tensor
norm_tensor = tl.norm(tensor) # norm of tensor
if list_factors == True: list_fac = []
if (time_rec == True): list_time = []
if (factors == None): factors = svd_init_fac(tensor, rank)
# Initialization of factor hat matrice by factor matrices
factors_hat = factors
if list_factors == True: list_fac.append(copy.deepcopy(factors))
list_F_hat_bf = []
weights = None
it = 0
err_it = 0
cpt = 0
########################################
######### error initialization #########
########################################
if (exact_err == True):
F_hat_bf = err(tensor, weights, factors)
else:
F_hat_bf, ind_bf = err_rand(tensor, None, factors, n_samples_err)
list_F_hat_bf.append(F_hat_bf)
rng = tl.check_random_state(None)
error = [F_hat_bf / norm_tensor]
min_err = error[len(error) - 1]
while (min_err > tol and it < it_max and err_it < err_it_max):
if time_rec == True: tic = time.time()
for n in range(N):
Zs, indices = sample_khatri_rao(factors_hat,
n_samples,
skip_matrix=n,
random_state=rng)
indices_list = [i.tolist() for i in indices]
indices_list.insert(n, slice(None, None, None))
indices_list = tuple(indices_list)
V = tl.dot(tl.transpose(Zs), Zs)
if (n == 0): sampled_unfolding = tensor[indices_list]
else: sampled_unfolding = tl.transpose(tensor[indices_list])
W = tl.dot(sampled_unfolding, Zs)
factor_bf = factors[n]
# update
fac, _, _, _ = hals_nnls(tl.transpose(W), V,
tl.transpose(factors[n]))
factors[n] = tl.transpose(
fac
) # solve needs a squared full rank matrix, if rank>nb_sampls ok
# extrapolate
factors_hat[n] = tl.clip(factors[n] + beta *
(factors[n] - factor_bf),
a_min=0.0)
########################################
######### error update #########
########################################
if (exact_err == False):
F_hat_new, _ = err_rand(tensor,
weights,
factors,
n_samples_err,
indices_list=ind_bf)
else:
F_hat_new = err(tensor, weights, factors)
list_F_hat_bf.append(F_hat_new)
if (F_hat_new > F_hat_bf):
factors_hat = factors
beta_bar = beta
beta = beta / eta
cpt = cpt + 1
else:
factors = factors_hat
beta_bar = min(1, beta_bar * gamma_bar)
beta = min(beta_bar, gamma * beta)
########################################
######### update for next it #########
########################################
it = it + 1
if (exact_err == False):
if it < filter:
F_hat_bf = np.mean(list_F_hat_bf)
else:
F_hat_bf = np.mean(list_F_hat_bf[(len(list_F_hat_bf) -
filter):(len(list_F_hat_bf) -
1)])
else:
F_hat_bf = F_hat_new
if list_factors == True: list_fac.append(copy.deepcopy(factors))
error.append(F_hat_new / norm_tensor)
if (error[len(error) - 1] < min_err):
min_err = error[len(error) - 1] # err update
else:
err_it = err_it + 1
if time_rec == True:
toc = time.time()
list_time.append(toc - tic)
if list_factors == True and time_rec == True:
return (weights, factors, it, error, cpt / it, list_fac, list_time)
if list_factors == True:
return (weights, factors, it, error, cpt / it, list_fac)
if time_rec == True:
return (weights, factors, it, error, cpt / it, list_time)
return (weights, factors, it, error, cpt / it)
