def mvm_form_prod_reorder(self,X,xfactors):
xdata=[[],[],[]] ## row index, column index , value
(m,n)=X.shape
X2=np.dot(X,X.T)
if m==400:
print(m,n)
ctensor=tensor_decomp.cls_tensor_decomp()
ctensor.load(X2,xfactors)
X2=ctensor.reorder_in_x(X2)
(m2,n2)=X2.shape
xmean=np.mean(X2)
xstd=np.std(X2)
lbound=xmean-5*xstd
ubound=xmean+5*xstd
k=0
icount=0
for i in range(m2):
for j in range(n2):
v=X2[i,j]
if v>=lbound and v<=ubound:
xdata[0].append(i)
xdata[1].append(j)
xdata[2].append(v)
k+=1
icount+=1
xdata=mvm_prepare.sort_table(xdata,ifloat=1)
return(xdata,m2,n2)
def invert(self,xmatre):
ctensor=tensor_decomp.cls_tensor_decomp()
xmatout=ctensor.invert_reorder_in_x(xmatre,self.xfactors)
return(xmatout)
def load_mmr(self, cData, lviews):
ldata_labels = self.read_raw_txt(self.sbasedir, self.labelfile)
if self.loadFurther == 1:
ldata_labels.append([self.evaldir, 1])
mdata = len(ldata_labels)
## !!!!!!!!!!!!!!!!!!!!!!!!!!
## fixed number of categories od sides
ncategory = 4
## ncategory=3
category_recode = [0, 1, 2, 3]
## ####################################
X_out = np.zeros((mdata, ncategory))
for i in range(mdata):
## X_out[i,int(ldata_labels[i][1])-1]=1
icategory = int(ldata_labels[i][1]) - 1
if icategory == 0:
X_out[i, category_recode[icategory]] = 1
else:
if ncategory == 3:
X_out[i, category_recode[icategory]] = 1
else:
X_out[i, category_recode[icategory]] = 1
## number of views: joint force, cart force, cart torque
nview = len(self.linputmask)
## find common minimum length of trajectories
xlength = np.zeros(mdata)
for i in range(mdata):
X = np.loadtxt(ldata_labels[i][0], skiprows=1)
xlength[i] = X.shape[0]
ixin_type = 0 ## =0 simple vectorized =1 angles
ixin_decompose = 0 ## =0 no, =1 tensor decomposition
icumulant = 0 ## =0 no cumulants, =1 cumulants are the feature components
nmin = xlength.min()
## !!!!!!!!!!!!!!!!!!!!!!
if nmin > 240:
if ixin_decompose == 1:
nmin = 240 ## 8*6*5
tfactors = (30, 8, 1, 1)
nfactor = len(tfactors)
nlayer = 1
ncumulant = 5
if ncumulant > tfactors[0]:
ncumulant = tfactors[0]
if icumulant == 0:
nitem = tfactors[0]
else:
nitem = ncumulant
if ixin_type == 0:
nboot = 1
if ixin_decompose == 0:
X_in = []
for i in range(nview):
nslice = self.linputmask[i][1] - self.linputmask[i][0]
X_in.append(np.zeros((mdata, nboot * nmin * nslice)))
elif ixin_decompose == 1:
X_in = []
for i in range(nview):
nslice = self.linputmask[i][1] - self.linputmask[i][0]
X_in.append(np.zeros((mdata, nlayer * nitem * nslice)))
for i in range(mdata):
## if i==18:
## print(i)
X = np.loadtxt(ldata_labels[i][0], skiprows=1)
X = X[:nmin, :]
if ixin_decompose == 0:
for j in range(nview):
i0 = self.linputmask[j][0]
ik = self.linputmask[j][1]
## fX=X[:,i0:ik].ravel()
## fX=np.sort(fX)
## X_in[j][i]=fX
X_in[j][i] = X[:, i0:ik].ravel()
elif ixin_decompose == 1:
for j in range(nview):
i0 = self.linputmask[j][0]
ik = self.linputmask[j][1]
xmatrix = X[:, i0:ik]
## xamtrix=spsignal.savgol_filter(xmatrix,7,5,deriv=0,axis=0, \
## mode='interp')
## xfactors assume 240 rows
xfactors=np.array([[tfactors[0],ik-i0],[tfactors[1],1], \
[tfactors[2],1],[tfactors[3],1]])
ctensor = tensor_decomp.cls_tensor_decomp()
ctensor.load(xmatrix, xfactors)
niter = 10
ctensor.decompose2(niter)
nslice = self.linputmask[j][1] - self.linputmask[j][0]
fX = np.zeros(nlayer * nitem * nslice)
ipos = 0
for ilayer in range(nlayer):
for ifactor in range(1):
zfact = ctensor.lcomponents[ilayer][ifactor]
(mzfact, nzfact) = zfact.shape
for icol in range(nzfact):
if np.mean(zfact[:, icol]) < 0:
zfact[:, icol] = -zfact[:, icol]
if icumulant == 1:
for icol in range(nzfact):
xcumulants = cumulants(zfact[:, icol],
ncumulant,
icentral=1)
fX[ipos:ipos + nitem] = xcumulants
ipos += nitem
else:
## (u,s,v)=np.linalg.svd(zfact)
## zfact=np.dot(u[:,:nzfact],np.diag(s[:nzfact]))
for icol in range(nzfact):
fX[ipos:ipos + mzfact] = zfact[:, icol]
ipos += mzfact
## fX=np.sort(fX)
X_in[j][i] = fX
elif ixin_type == 1:
X_in = [np.zeros((mdata, nmin - 1)) for i in range(nview)]
for i in range(mdata):
X = np.loadtxt(self.sbasedir + ldata_labels[i][0], skiprows=1)
X = X[:nmin, :]
for j in range(nview):
i0 = self.linputmask[j][0]
ik = self.linputmask[j][1]
X_in[j][i] = angle_moment(X[:, i0:ik])
## subspace output kernel
cData.YKernel = mmr_kernel_explicit.cls_feature(ifeature=0)
cData.YKernel.load_data(X_out, ifeature=0)
cData.YKernel.ifeature = 0
cData.YKernel.title = 'output'
## setting output parameters
cparams = cls_initial_params()
cData.YKernel.kernel_params.set(cparams.get_yparams('kernel', 0))
## cData.YKernel.prekernel_params.set(cparams.get_yinparams('kernel',0))
cData.YKernel.crossval.set(cparams.get_yparams('cross', 0))
cData.YKernel.norm.set(cparams.get_yparams('norm', 0))
idata = 0
for iview in range(nview):
if iview in lviews:
cData.XKernel[idata] = mmr_kernel_explicit.cls_feature(
ifeature=0)
cData.XKernel[idata].load_data(X_in[iview], ifeature=0)
cData.XKernel[idata].title = 'input_' + str(iview)
## setting input parameters
cData.XKernel[idata].kernel_params.set(cparams. \
get_xparams('kernel',idata))
## cData.XKernel[idata].prekernel_params.set(cparams. \
## get_xinparams('kernel',idata))
cData.XKernel[idata].crossval.set(
cparams.get_xparams('cross', idata))
cData.XKernel[idata].norm.set(
cparams.get_xparams('norm', idata))
idata += 1
cData.ninputview = idata ## set active views
cData.mdata = cData.YKernel.dataraw.shape[0]
cData.nfold = 2
cData.nrepeat = 2
cData.kmode = 1 ## =0 additive (feature concatenation)
## =1 multiplicative (fetaure tensor product)
return
def load(self,ifile):
if ifile==0:
sdir='/home/szedmak/data/image_text/labeled_images/'
sfile='/cambridge/Image007.png'
xmatpil=pilimg.open(sdir+sfile)
xmatrix=np.array(xmatpil.resize((768,512))) ## 504,756 -> 512,768
self.xfactors=np.array([[32,48,1],[16,16,1],[1,1,3]])
elif ifile==1:
sdir='/home/szedmak/src/python/algebra/tensor_decomp/'
sfile='handwritten.png'
xmatpil=pilimg.open(sdir+sfile)
xmatrix=np.array(xmatpil.resize((1024,1024))) ## 504,756 -> 512,768
self.xfactors=np.array([[32,32],[32,32]])
elif ifile==2:
sdir='/home/szedmak/src/python/algebra/tensor_decomp/'
sfile='tilling.png'
xmatpil=pilimg.open(sdir+sfile)
xmatrix=np.array(xmatpil.resize((576,512))) ## 504,756 -> 512,768
## xmatrix=np.mean(xmatrix,axis=2)
self.xfactors=np.array([[32,36,1],[16,16,1],[1,1,3]]) ## tilling
elif ifile==3:
sdir='/home/szedmak/src/python/algebra/tensor_decomp/'
sfile='poligon.png'
xmatpil=pilimg.open(sdir+sfile)
xmatrix=np.array(xmatpil.resize((512,512))) ## 504,756 -> 512,768
self.xfactors=np.array([[32,32,1],[16,16,1],[1,1,3]]) ## poligon
elif ifile==4:
sdir='/home/szedmak/src/python/algebra/tensor_decomp/'
sfile='2008toyroom.png'
xmatpil=pilimg.open(sdir+sfile)
xmatrix=np.array(xmatpil.resize((768,512))) ## 504,756 -> 512,768
## xmatrix=np.mean(xmatrix,axis=2)
self.xfactors=np.array([[32,48,1],[16,16,1],[1,1,3]])
elif ifile==5:
sdir='/home/szedmak/src/python/algebra/tensor_decomp/'
sfile='PLayroom_How_to_Organize_Kids_Room_2013.png'
xmatpil=pilimg.open(sdir+sfile)
xmatrix=np.array(xmatpil.resize((768,512))) ## 504,756 -> 512,768
self.xfactors=np.array([[32,48,1],[16,16,1],[1,1,3]])
print(sdir,sfile)
print(self.xfactors)
self.baseline(xmatrix)
ctensor=tensor_decomp.cls_tensor_decomp()
ctensor.load(xmatrix,self.xfactors)
X=ctensor.reorder_in_x(xmatrix)
## X=X.T
tshape=X.shape
mi=tshape[0]
mj=tshape[1]
## if len(tshape)>2:
## mk=tshape[2]
## X=X/self.image_scale
xdata=[[],[],[]] ## row index, column index , value
for i in range(mi):
for j in range(mj):
xdata[0].append(i)
xdata[1].append(int(np.mean(X[i,j])))
xdata[2].append(X[i,j]/self.image_scale)
xdata=mvm_prepare.sort_table(xdata,ifloat=1)
m1=mi
## m2=mj
m2=256
return(xdata,m1,m2)
def load_mmr(self, cData, lviews):
ldata_labels = self.read_raw_txt(self.sbasedir,self.labelfile)
if self.loadFurther==1:
ldata_labels.append([self.evaldir, 1])
mdata=len(ldata_labels)
## !!!!!!!!!!!!!!!!!!!!!!!!!!
## fixed number of categories od sides
ncategory=4
## ncategory=3
category_recode=[0,1,2,3]
## ####################################
X_out=np.zeros((mdata,ncategory))
for i in range(mdata):
## X_out[i,int(ldata_labels[i][1])-1]=1
icategory=int(ldata_labels[i][1])-1
if icategory==0:
X_out[i,category_recode[icategory]]=1
else:
if ncategory==3:
X_out[i,category_recode[icategory]]=1
else:
X_out[i,category_recode[icategory]]=1
## number of views: joint force, cart force, cart torque
nview=len(self.linputmask)
## find common minimum length of trajectories
xlength=np.zeros(mdata)
for i in range(mdata):
X=np.loadtxt(ldata_labels[i][0],skiprows=1)
xlength[i]=X.shape[0]
ixin_type=0 ## =0 simple vectorized =1 angles
ixin_decompose=0 ## =0 no, =1 tensor decomposition
icumulant=0 ## =0 no cumulants, =1 cumulants are the feature components
nmin=xlength.min()
## !!!!!!!!!!!!!!!!!!!!!!
if nmin>240:
if ixin_decompose==1:
nmin=240 ## 8*6*5
tfactors=(30,8,1,1)
nfactor=len(tfactors)
nlayer=1
ncumulant=5
if ncumulant>tfactors[0]:
ncumulant=tfactors[0]
if icumulant==0:
nitem=tfactors[0]
else:
nitem=ncumulant
if ixin_type==0:
nboot=1
if ixin_decompose==0:
X_in=[]
for i in range(nview):
nslice=self.linputmask[i][1]-self.linputmask[i][0]
X_in.append(np.zeros((mdata,nboot*nmin*nslice)))
elif ixin_decompose==1:
X_in=[]
for i in range(nview):
nslice=self.linputmask[i][1]-self.linputmask[i][0]
X_in.append(np.zeros((mdata,nlayer*nitem*nslice)))
for i in range(mdata):
## if i==18:
## print(i)
X=np.loadtxt(ldata_labels[i][0],skiprows=1)
X=X[:nmin,:]
if ixin_decompose==0:
for j in range(nview):
i0=self.linputmask[j][0]
ik=self.linputmask[j][1]
## fX=X[:,i0:ik].ravel()
## fX=np.sort(fX)
## X_in[j][i]=fX
X_in[j][i]=X[:,i0:ik].ravel()
elif ixin_decompose==1:
for j in range(nview):
i0=self.linputmask[j][0]
ik=self.linputmask[j][1]
xmatrix=X[:,i0:ik]
## xamtrix=spsignal.savgol_filter(xmatrix,7,5,deriv=0,axis=0, \
## mode='interp')
## xfactors assume 240 rows
xfactors=np.array([[tfactors[0],ik-i0],[tfactors[1],1], \
[tfactors[2],1],[tfactors[3],1]])
ctensor=tensor_decomp.cls_tensor_decomp()
ctensor.load(xmatrix,xfactors)
niter=10
ctensor.decompose2(niter)
nslice=self.linputmask[j][1]-self.linputmask[j][0]
fX=np.zeros(nlayer*nitem*nslice)
ipos=0
for ilayer in range(nlayer):
for ifactor in range(1):
zfact=ctensor.lcomponents[ilayer][ifactor]
(mzfact,nzfact)=zfact.shape
for icol in range(nzfact):
if np.mean(zfact[:,icol])<0:
zfact[:,icol]=-zfact[:,icol]
if icumulant==1:
for icol in range(nzfact):
xcumulants=cumulants(zfact[:,icol],ncumulant,icentral=1)
fX[ipos:ipos+nitem]=xcumulants
ipos+=nitem
else:
## (u,s,v)=np.linalg.svd(zfact)
## zfact=np.dot(u[:,:nzfact],np.diag(s[:nzfact]))
for icol in range(nzfact):
fX[ipos:ipos+mzfact]=zfact[:,icol]
ipos+=mzfact
## fX=np.sort(fX)
X_in[j][i]=fX
elif ixin_type==1:
X_in=[ np.zeros((mdata,nmin-1)) for i in range(nview)]
for i in range(mdata):
X=np.loadtxt(self.sbasedir+ldata_labels[i][0],skiprows=1)
X=X[:nmin,:]
for j in range(nview):
i0=self.linputmask[j][0]
ik=self.linputmask[j][1]
X_in[j][i]=angle_moment(X[:,i0:ik])
## subspace output kernel
cData.YKernel=mmr_kernel_explicit.cls_feature(ifeature=0)
cData.YKernel.load_data(X_out,ifeature=0)
cData.YKernel.ifeature=0
cData.YKernel.title='output'
## setting output parameters
cparams=cls_initial_params()
cData.YKernel.kernel_params.set(cparams.get_yparams('kernel',0))
## cData.YKernel.prekernel_params.set(cparams.get_yinparams('kernel',0))
cData.YKernel.crossval.set(cparams.get_yparams('cross',0))
cData.YKernel.norm.set(cparams.get_yparams('norm',0))
idata=0
for iview in range(nview):
if iview in lviews:
cData.XKernel[idata]=mmr_kernel_explicit.cls_feature(ifeature=0)
cData.XKernel[idata].load_data(X_in[iview],ifeature=0)
cData.XKernel[idata].title='input_'+str(iview)
## setting input parameters
cData.XKernel[idata].kernel_params.set(cparams. \
get_xparams('kernel',idata))
## cData.XKernel[idata].prekernel_params.set(cparams. \
## get_xinparams('kernel',idata))
cData.XKernel[idata].crossval.set(cparams.get_xparams('cross',idata))
cData.XKernel[idata].norm.set(cparams.get_xparams('norm',idata))
idata+=1
cData.ninputview=idata ## set active views
cData.mdata=cData.YKernel.dataraw.shape[0]
cData.nfold=2
cData.nrepeat=2
cData.kmode=1 ## =0 additive (feature concatenation)
## =1 multiplicative (fetaure tensor product)
return
