def get_RE(com_rank, duplicate=10):
ind = 0
XErList = []
ThetaErList = np.zeros((duplicate, len(missList)))
for miss in missList:
RETheta = []
REX = []
for dup in range(duplicate):
np.random.seed(dup * 4)
#每次都用同一份数据去做
data = dat.copy()
#观测值：丢失部分数据的
Omega = (np.random.random(siz) > miss) * 1
data[Omega == 0] -= data[Omega == 0]
#补全时候用的rank
print('missing ratio: {0}'.format(miss))
data = pyten.tenclass.tensor.Tensor(data)
self2 = PoissonAirCP(data,
omega=Omega,
rank=com_rank,
max_iter=3000,
tol=1e-5,
OnlyObs=True,
TrueValue=true_data,
alpha=[10, 10, 10],
lmbda=10,
sim_mats=simMat2)
self2.run()
#这里看对原始数据的补全准不准
[EErr, EReErr1, EReErr2] = tenerror(self2.X, true_data, Omega)
REX.append(EReErr1)
#这里看对原始theta矩阵的估计准不准
pred2 = pyten.tenclass.Ktensor(np.ones(com_rank),
us=self2.U).totensor()
ThetaRE = np.linalg.norm(pred2.data - theta.data) / np.linalg.norm(
theta.data)
RETheta.append(ThetaRE)
# print(np.mean(RECP),np.mean(RE1),np.mean(RE2))
XErList.append(np.mean(REX))
ThetaErList[:, ind] = RETheta
ind = ind + 1
print(XErList)
print(ThetaErList)
return ((XErList, ThetaErList))
tol=1e-5,
OnlyObs=True,
TrueValue=true_data,
sim_mats=aux,
alpha=para_alpha,
lmbda=para_lmbda)
self2.run()
temp_aux = cons_similarity(self2.X.data)
simerror = np.max((np.linalg.norm(aux[0] - temp_aux[0]),
np.linalg.norm(aux[1] - temp_aux[1]),
np.linalg.norm(aux[2] - temp_aux[2])))
aux = temp_aux
Iter = Iter + 1
print('ExpAirCP loop with similarity error: {0}'.format(simerror))
[EEr, EReEr1, EReEr2] = tenerror(self2.X, true_data, Omega)
if Iter == 2:
RE22.append(EReEr1)
print(EReEr1)
# 到这里为止
[EErr, EReErr1, EReErr2] = tenerror(self2.X, true_data, Omega)
print('ExpAirCP Completion Error: {0}, {1}, {2}'.format(
EErr, EReErr1, EReErr2))
RE2.append(EReErr1)
finalList22.append(np.mean(RE22))
finalList2.append(np.mean(RE2))
for miss in missList:
aux = [
eta=1e-4)
self3 = TNCP(data, Omega, rank=com_rank, alpha=[3, 3, 3], lmbda=3)
rX1 = falrtc(data, Omega, max_iter=100)
try:
#print(data.data)
self1.run()
#print(data.data)
self2.run()
self3.run()
except:
continue
#这里看对原始数据的补全准不准
[Err, ReErr1, ReErr2] = tenerror(self1.X, true_data, Omega)
[EErr, EReErr1, EReErr2] = tenerror(self2.X, true_data, Omega)
[ErrTNCP, ReErrTNCP, ReErr2TNCP] = tenerror(self3.X, true_data, Omega)
#[Errfal, ReErrfal, ReErr2fal] = tenerror(rX1, true_data, Omega)
print('AirCP Completion Error: {0}, {1}, {2}'.format(
Err, ReErr1, ReErr2))
print('ExpAirCP Completion Error: {0}, {1}, {2}'.format(
EErr, EReErr1, EReErr2))
print('TNCP Completion Error: {0}, {1}, {2}'.format(
ErrTNCP, ReErrTNCP, ReErr2TNCP))
#print ('falrtc Completion Error: {0}, {1}, {2}'.format(Errfal, ReErrfal, ReErr2fal))
#这里看对原始theta矩阵的估计准不准
pred2 = pyten.tenclass.Ktensor(np.ones(com_rank),
us=self2.U).totensor()
OnlyObs=True,
TrueValue=true_data,
sim_mats=simMat,
alpha=[3, 3, 3],
lmbda=3)
self3 = TNCP(data, Omega, rank=com_rank, alpha=[3, 3, 3], lmbda=3)
#rX1 = falrtc(data, Omega, max_iter=100)
self1.run()
self2.run()
self3.run()
#这里看对原始数据的补全准不准
[Err, ReErr1, ReErr2] = tenerror(self1.X, true_data, Omega)
[EErr, EReErr1, EReErr2] = tenerror(self2.X, true_data, Omega)
[Errcp, ReErrcp, ReErr2cp] = tenerror(self3.X, true_data, Omega)
#[Errfal, ReErrfal, ReErr2fal] = tenerror(rX1, true_data, Omega)
print('AirCP Completion Error: {0}, {1}, {2}'.format(
Err, ReErr1, ReErr2))
print('ExpAirCP Completion Error: {0}, {1}, {2}'.format(
EErr, EReErr1, EReErr2))
print('TNCP Completion Error: {0}, {1}, {2}'.format(
Errcp, ReErrcp, ReErr2cp))
#print ('falrtc Completion Error: {0}, {1}, {2}'.format(Errfal, ReErrfal, ReErr2fal))
pred2 = pyten.tenclass.Ktensor(np.ones(com_rank),
us=self2.U).totensor()
print('Theta Error for ExpAirCP: {0}'.format(
failuresR=failuresR2,
omega=Omega,
rank=com_rank,
max_iter=3000,
tol=1e-5,
OnlyObs=True,
TrueValue=true_data,
sim_mats=simMat,
alpha=[3, 3, 3],
lmbda=3)
self2.run()
self3.run()
#这里看对原始数据的补全准不准
[EErr, EReErr1, EReErr2] = tenerror(self2.X, true_data, Omega)
[EErr2, EReErr3, EReErr4] = tenerror(self3.X, true_data, Omega)
RE2.append(EReErr1)
RE3.append(EReErr3)
# #
# x = self1.X.data-true_data.data
# y = self2.X.data-true_data.data
# all_data=[x.reshape(np.prod(siz)),y.reshape(np.prod(siz))]
#
# figure,axes=plt.subplots() #得到画板、轴
# axes.boxplot(all_data) #描点上色
# plt.show()
finalList2[:, ind] = RE2
# Tucker-ALS
[T2, rX2] = tucker_als(X1, R, Omega1) # if no missing data just omit Omega1
# FalRTC, SiLRTC, HaLRTC
rX3 = falrtc(X1, Omega1)
rX4 = silrtc(X1, Omega1)
rX5 = halrtc(X1, Omega1)
# TNCP
self1 = TNCP(X1, Omega1, rank=r)
self1.run()
# Error Testing
from pyten.tools import tenerror
realX = sol1.totensor()
[Err1, ReErr11, ReErr21] = tenerror(rX1, realX, Omega1)
[Err2, ReErr12, ReErr22] = tenerror(rX2, realX, Omega1)
[Err3, ReErr13, ReErr23] = tenerror(rX3, realX, Omega1)
[Err4, ReErr14, ReErr24] = tenerror(rX4, realX, Omega1)
[Err5, ReErr15, ReErr25] = tenerror(rX5, realX, Omega1)
[Err6, ReErr16, ReErr26] = tenerror(self1.X, realX, Omega1)
print '\n', 'The Relative Error of the Six Methods are:', ReErr21, ReErr22, ReErr23, ReErr24, ReErr25, ReErr26
# 2. Real Problem - Image Recovery
import matplotlib.image as mpimg # Use it to load image
import numpy as np
lena = mpimg.imread("./test/testImg.png")
im = np.double(np.uint8(lena * 255))
im = im[0:50, 0:50, 0:3]
simerror = 1
Iter = 1
while (simerror > 1e-3 and Iter < 10):
self1 = PoissonAirCP(data,
omega=Omega,
rank=com_rank,
max_iter=3000,
tol=1e-5,
OnlyObs=True,
TrueValue=true_data,
sim_mats=aux,
alpha=para_alpha,
lmbda=para_lmbda)
self1.run()
temp_aux = cons_similarity(self1.X.data)
simerror = np.max((np.linalg.norm(aux[0] - temp_aux[0]),
np.linalg.norm(aux[1] - temp_aux[1]),
np.linalg.norm(aux[2] - temp_aux[2])))
aux = temp_aux
Iter = Iter + 1
print(simerror)
# 到这里为止
[EEr, EReEr1, EReEr2] = tenerror(self1.X, true_data, Omega)
print('ExpAirCP with aux: {0}, {1}, {2}'.format(
EEr, EReEr1, EReEr2))
RE1.append(EReEr1)
result[i, j, k] = np.mean(RE1)
print(result)
np.save('GDELTstep5.npy', result)
