def mmr_train(self):
if self.iperceptron == 0:
mtra = self.mtrain
KX=mmr_kernel(self,self.itrain,self.itrain,ioutput=0, \
itraintest=0,itraindata=0,itensor=self.kmode)[0]
KY=mmr_kernel(self,self.itrain,self.itrain,ioutput=1, \
itraintest=0,itraindata=0)[0]
cOptDual = base.cls_dual(None, None)
self.dual = cOptDual
self.csolver = mmr_solver_cls.cls_mmr_solver()
self.dual.alpha=self.csolver.mmr_solver(KX,KY,self.penalty.c, \
self.penalty.d)
## estimate the bias for linear output kernel
if self.YKernel.ifeature == 0:
YTrain = self.YKernel.get_train_norm(self.itrain)
ZW=np.dot(YTrain.T, \
(np.outer(self.dual.alpha,np.ones(mtra))*KX))
xbias = np.median(YTrain - ZW.T, axis=0)
else:
KY = self.YKernel.Kcross
ZW = np.dot(KY.T,
(np.outer(self.dual.alpha, np.ones(mtra)) * KX))
xbias = np.zeros(mtra)
self.dual.bias = xbias
elif self.iperceptron == 1:
cOptDual = base.cls_dual(None, None)
XTrainNorm0 = self.XKernel[0].get_train_norm()
YTrainNorm = self.YKernel.get_train_norm(self.itrain)
cOptDual.W=mmr_perceptron_primal(XTrainNorm0,YTrainNorm, \
self.perceptron.margin, \
self.perceptron.stepsize, \
self.perceptron.nrepeat)
elif self.iperceptron == 2:
cOptDual = base.cls_dual(None, None)
XTrainNorm0 = self.XKernel[0].get_train_norm()
YTrainNorm = self.YKernel.get_train_norm(self.itrain)
cOptDual.alpha=mmr_perceptron_dual(XTrainNorm0,YTrainNorm, \
self.XKernel[0].kernel_params.ipar1, \
self.XKernel[0].kernel_params.ipar2, \
self.XKernel[0].kernel_type, \
self.perceptron.margin, \
self.perceptron.stepsize, \
self.perceptron.nrepeat)
return (cOptDual)
def mmr_train(self):
if self.iperceptron==0:
mtra=self.mtrain
KX=mmr_kernel(self,self.itrain,self.itrain,ioutput=0, \
itraintest=0,itraindata=0,itensor=self.kmode)[0]
KY=mmr_kernel(self,self.itrain,self.itrain,ioutput=1, \
itraintest=0,itraindata=0)[0]
cOptDual=base.cls_dual(None,None)
self.dual=cOptDual
self.csolver=mmr_solver_cls.cls_mmr_solver()
self.dual.alpha=self.csolver.mmr_solver(KX,KY,self.penalty.c, \
self.penalty.d)
## estimate the bias for linear output kernel
if self.YKernel.ifeature==0:
YTrain=self.YKernel.get_train_norm(self.itrain)
ZW=np.dot(YTrain.T, \
(np.outer(self.dual.alpha,np.ones(mtra))*KX))
xbias=np.median(YTrain-ZW.T,axis=0)
else:
KY=self.YKernel.Kcross
ZW=np.dot(KY.T,(np.outer(self.dual.alpha,np.ones(mtra))*KX))
xbias=np.zeros(mtra)
self.dual.bias=xbias
elif self.iperceptron==1:
cOptDual=base.cls_dual(None,None)
XTrainNorm0=self.XKernel[0].get_train_norm()
YTrainNorm=self.YKernel.get_train_norm(self.itrain)
cOptDual.W=mmr_perceptron_primal(XTrainNorm0,YTrainNorm, \
self.perceptron.margin, \
self.perceptron.stepsize, \
self.perceptron.nrepeat)
elif self.iperceptron==2:
cOptDual=base.cls_dual(None,None)
XTrainNorm0=self.XKernel[0].get_train_norm()
YTrainNorm=self.YKernel.get_train_norm(self.itrain)
cOptDual.alpha=mmr_perceptron_dual(XTrainNorm0,YTrainNorm, \
self.XKernel[0].kernel_params.ipar1, \
self.XKernel[0].kernel_params.ipar2, \
self.XKernel[0].kernel_type, \
self.perceptron.margin, \
self.perceptron.stepsize, \
self.perceptron.nrepeat)
return(cOptDual)
def mvm_train(self):
"""
execute the trianing procedure
Inputs:
"""
## print('Generate kernels')
time0 = time.time()
self.compute_kernels()
if self.verbose == 1:
print('Kernel computation:', time.time() - time0)
## print('Solve optimization problem')
time0 = time.time()
self.KX=mmr_kernel(self,self.itrain,self.itrain,ioutput=0, \
itraintest=0, itensor=self.kmode)[0]
self.KY = mmr_kernel(self, self.itrain, self.itrain, ioutput=1)[0]
if self.verbose == 1:
print('Kernel merge computation:', time.time() - time0)
## self.solvertime=time.time()-time0
## t0=time.clock()
cOptDual = base.cls_dual(None, None)
self.dual = cOptDual
time0 = time.time()
cmvm_solver = mvm_solver_cls.cls_mvm_solver()
self.dual.alpha = cmvm_solver.mvm_solver(self)
self.solvertime = time.time() - time0
return (cOptDual)
def mvm_train(self):
"""
execute the trianing procedure
Inputs:
"""
## print('Generate kernels')
time0=time.time()
self.compute_kernels()
if self.verbose==1:
print('Kernel computation:',time.time()-time0)
## print('Solve optimization problem')
time0=time.time()
self.KX=mmr_kernel(self,self.itrain,self.itrain,ioutput=0, \
itraintest=0, itensor=self.kmode)[0]
self.KY=mmr_kernel(self,self.itrain,self.itrain,ioutput=1)[0]
if self.verbose==1:
print('Kernel merge computation:',time.time()-time0)
## self.solvertime=time.time()-time0
## t0=time.clock()
cOptDual=base.cls_dual(None,None)
self.dual=cOptDual
time0=time.time()
cmvm_solver=mvm_solver_cls.cls_mvm_solver()
self.dual.alpha=cmvm_solver.mvm_solver(self)
self.solvertime=time.time()-time0
return(cOptDual)
def load_mmr(self,cData):
X_in=np.loadtxt(self.training_input)
X_out=np.loadtxt(self.training_output)
alpha=np.loadtxt(self.dualParams)
kernelParams=np.loadtxt(self.kernelParams)
(m,nx)=X_in.shape
ny=X_out.shape[1]
cData.training_center=np.mean(X_in,axis=0)
## loading the offline training parameters
cData.dual=base.cls_dual(alpha,np.zeros(ny)+0.1)
## add an empty test row to the trainig,
## it will be replaced with the real test inputs
X_in=[np.vstack((X_in,np.zeros((1,nx))))] ## list of views
X_out=np.vstack((X_out,np.zeros((1,ny))))
## 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.crossval.set(cparams.get_yparams('cross',0))
cData.YKernel.norm.set(cparams.get_yparams('norm',0))
idata=0
nview=1
for iview in range(nview):
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))
## kernel parameters
cData.XKernel[idata].kernel_params.ipar1=kernelParams[0]
cData.XKernel[idata].kernel_params.ipar2=kernelParams[1]
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=1
cData.nrepeat=1
cData.kmode=1 ## =0 additive (feature concatenation)
## =1 multiplicative (fetaure tensor product)
return
