def mvm_test(self):
"""
select the potential test method
wrapper around the possible test evaluations of different losses
inputs:
cOptDual object containing the optimal values of the
dual variables
outputs:
cPredict object containing the prediction results
"""
cPredict = base.cls_predict()
itest_method = 0 # =0 orig
if itest_method == 0:
# matrix completition test
cPredict.Zrow = mvm_test_orig.mvm_test_orig(self, self.dual.alpha)
return (cPredict)
def mvm_test(self):
"""
select the potential test method
wrapper around the possible test evaluations of different losses
inputs:
cOptDual object containing the optimal values of the
dual variables
outputs:
cPredict object containing the prediction results
"""
cPredict=base.cls_predict()
itest_method=0 # =0 orig
if itest_method==0:
# matrix completition test
cPredict.Zrow=mvm_test_orig.mvm_test_orig(self,self.dual.alpha)
return(cPredict)
def test_dynamic_prog(cDataTest,cOptDual):
(mtrain,mtest)=cDataTest.KXcross.shape
ndim=cDataTest.YTrainNorm.shape[1]
cPredict=mmr_base_classes.cls_predict()
cPredict.zPred=np.zeros((mtest,ndim))
beta=cDataTest.KXcross*np.outer(cOptDual.alpha,np.ones(mtest))
beta=beta.T
## inverse covariance
## ymean=np.mean(cDataTest.YTrainNorm,axis=0)
Y=cDataTest.YTrainNorm
## Yn=Y-np.outer(np.ones(mtrain),ymean)
## Ycov=np.dot(Yn.T,Yn)/mtrain
## Ycovinv=np_lin.pinv(Ycov,rcond=1e-2)/(2*(params.output.ipar1**2))
Ycovinv=np.eye(ndim)/(2*(cDataTest.kernel_params.ipar1**2))
CY=np.dot(Ycovinv,Y.T)
cy0=np.sum(Y*CY.T,axis=1)
y0=np.zeros(ndim)
## y0c=np.zeros(ndim)
## e=np.ones(mtrain)
rtime=time.time()
ntime=5
xdynstack0=np.zeros((ntime,ndim))
xdynstack1=np.zeros((ntime,ndim,ndim))
xdynstack2=np.zeros((ntime,ndim,mtrain))
for it in range(mtest):
## setup of the dynamic programming
## load time 0
f=np.dot(np.exp(-cy0),beta[it])
for i in range(ntime):
for j in range(ndim):
xdynstack0[i,j]=f
xdynstack1[i,j,:]=y0
xdynstack2[i,j,:]=cy0
## print('>>>',f)
for t in range(0,ntime-1):
if t==0:
nitem=1
else:
nitem=ndim
for i1 in range(nitem):
## print(t,i1)
f=xdynstack0[t,i1]
y0=xdynstack1[t,i1,:]
cy0=xdynstack2[t,i1,:]
fmax=0
y0c=np.dot(Ycovinv,y0)
for i2 in range(ndim):
cy_new=cy0+Ycovinv[i2,i2]+y0c[i2]-CY[i2]
fnew=np.dot(np.exp(-cy_new),beta[it])
if fnew>xdynstack0[t+1,i2]:
## print(t,i1,i2,fnew)
y=np.copy(y0)
y[i2]=1
xdynstack0[t+1,i2]=fnew
xdynstack1[t+1,i2,:]=y
xdynstack2[t+1,i2,:]=cy_new
## else:
## y=np.copy(y0)
## dynstack[t+1][i2]=(f,y,y0c)
if fnew>fmax:
fmax=fnew
## print(f,fmax)
## print(t,i1)
## find the best one
fbest=0
ibest=-1
for i in range(ndim):
if fbest<xdynstack0[ntime-1,i]:
fbest=xdynstack0[ntime-1,i]
ibest=i
cPredict.zPred[it,:]=xdynstack1[ntime-1,ibest,:]
if it%10==0:
print(it,(time.time()-rtime)/(it+1))
return(cPredict)
def mmr_test(self,cOptDual,itraindata=1):
if self.YKernel.kernel_params.kernel_type>0:
self.itestmode=0
if itraindata==0:
itest=self.itrain
else:
itest=self.itest
if self.iperceptron in (0,2):
KXcross=mmr_kernel(self,self.itrain,itest,ioutput=0, \
itraintest=1, itraindata=itraindata, \
itensor=self.kmode)[0]
KYcross=mmr_kernel(self,self.itrain,self.itrain,ioutput=1, \
itraintest=1,itraindata=itraindata)[0]
mtest=KXcross.shape[1]
cPredict=base.cls_predict()
if self.YKernel.ifeature==0:
if self.YKernel.kernel_params.kernel_type==0:
YTrain=self.YKernel.get_train_norm(self.itrain)
Zw0=np.dot(YTrain.T,(np.outer(cOptDual.alpha, \
np.ones(mtest))*KXcross))
if self.csolver.ibias_estim==1:
Zw0=Zw0+np.outer(cOptDual.bias,np.ones(mtest))
xnorm=np.sqrt(np.sum(Zw0**2,axis=0))
xnorm=xnorm+(xnorm==0)
Zw=Zw0/np.outer(np.ones(Zw0.shape[0]),xnorm)
Y0=self.YKernel.get_Y0_norm(self.itrain)
cPredict.ZTest=np.dot(Y0,Zw)
else:
Zw0=np.dot(KYcross.T, \
(np.outer(cOptDual.alpha,np.ones(mtest))*KXcross))
cPredict.ZTest=Zw0
if self.csolver.ibias_estim==1:
cPredict.bias=cOptDual.bias
else:
cPredict.bias=0
Zw=Zw0
cPredict.Zw0=Zw0.T
cPredict.Zw=Zw.T
if self.itestmode in (0,2,3,4):
Y0=self.YKernel.get_Y0(self.itrain)
cPredict.zPremax=cPredict.ZTest.max(0)
cPredict.iPredCat=cPredict.ZTest.argmax(0)
cPredict.zPred=Y0[cPredict.iPredCat]
if self.itestmode==2 and self.YKernel.kernel_params.kernel_type==0:
cPredict.knnPredCat=np.argsort \
(-cPredict.ZTest,axis=0)[:self.testknn,:]
elif self.itestmode==3:
n=YTrain.shape[1]
nsort=5
for i in range(mtest):
cPredict.zPred[i,:]=np.zeros(n)
iones=np.argsort(-Zw[:,i])[:nsort]
for j in range(nsort-1):
if Zw[iones[j],i]<0.04 or \
Zw[iones[j],i]>10.0*Zw[iones[j+1],i]:
iones=iones[:(j+1)]
break
cPredict.zPred[i,iones]=1
elif self.itestmode==1:
YTrain=self.YKernel.get_train(self.itrain)
ndim=YTrain.shape[1]
i2boro=np.argsort(-Zw0,axis=0)[0:self.testknn,:].T
cPredict.zPred=np.zeros((mtest,ndim))
for i in range(mtest):
zsum=np.zeros(ndim)
for j in range(self.testknn):
zsum+=YTrain[i2boro[i,j]]
cPredict.zPred[i,:]=1*(zsum>=self.testknn/2)
elif self.YKernel.ifeature==1:
Y0=self.YKernel.get_Y0(self.itrain)
cPredict.ZTest=np.dot(KYcross.T,(np.outer(cOptDual.alpha, \
np.ones(mtest))*KXcross))
cPredict.zPremax=cPredict.ZTest.max(0)
cPredict.iPredCat=cPredict.ZTest.argmax(0)
cPredict.zPred=Y0[cPredict.iPredCat]
if self.itestmode==2:
cPredict.knnPredCat=np.argsort \
(-cPredict.ZTest,axis=0)[:self.testknn,:]
elif self.iperceptron==1:
if itraindata==0:
X=self.XKernel[0].get_train_norm(self.itrain)
for i in range(1,len(self.XKernel)):
X=np.hstack((X,self.XKernel[i].get_train_norm(self.itrain)))
else:
X=self.XKernel[0].get_test_norm(itest)
for i in range(1,len(self.XKernel)):
X=np.hstack((X,self.XKernel[i].get_test_norm(itest)))
cPredict=base.cls_predict()
xmean=np.mean(X)
mtest=X.shape[0]
Zw=np.dot(cOptDual.W, \
np.hstack((X,np.ones((mtest,1))*xmean)).T)
cPredict.ZTest=np.dot(YTrain,Zw)
cPredict.zPremax=cPredict.ZTest.max(0)
cPredict.iPredCat=cPredict.ZTest.argmax(0)
cPredict.zPred=YTrain[cPredict.iPredCat]
if self.itestmode==2:
cPredict.knnPredCat=np.argsort(-cPredict.ZTest, \
axis=0)[:self.testknn,:]
return(cPredict)
def mmr_test(self, cOptDual, itraindata=1):
if self.YKernel.kernel_params.kernel_type > 0:
self.itestmode = 0
if itraindata == 0:
itest = self.itrain
else:
itest = self.itest
if self.iperceptron in (0, 2):
KXcross=mmr_kernel(self,self.itrain,itest,ioutput=0, \
itraintest=1, itraindata=itraindata, \
itensor=self.kmode)[0]
KYcross=mmr_kernel(self,self.itrain,self.itrain,ioutput=1, \
itraintest=1,itraindata=itraindata)[0]
mtest = KXcross.shape[1]
cPredict = base.cls_predict()
if self.YKernel.ifeature == 0:
if self.YKernel.kernel_params.kernel_type == 0:
YTrain = self.YKernel.get_train_norm(self.itrain)
Zw0=np.dot(YTrain.T,(np.outer(cOptDual.alpha, \
np.ones(mtest))*KXcross))
if self.csolver.ibias_estim == 1:
Zw0 = Zw0 + np.outer(cOptDual.bias, np.ones(mtest))
xnorm = np.sqrt(np.sum(Zw0**2, axis=0))
xnorm = xnorm + (xnorm == 0)
Zw = Zw0 / np.outer(np.ones(Zw0.shape[0]), xnorm)
Y0 = self.YKernel.get_Y0_norm(self.itrain)
cPredict.ZTest = np.dot(Y0, Zw)
else:
Zw0=np.dot(KYcross.T, \
(np.outer(cOptDual.alpha,np.ones(mtest))*KXcross))
cPredict.ZTest = Zw0
if self.csolver.ibias_estim == 1:
cPredict.bias = cOptDual.bias
else:
cPredict.bias = 0
Zw = Zw0
cPredict.Zw0 = Zw0.T
cPredict.Zw = Zw.T
if self.itestmode in (0, 2, 3, 4):
Y0 = self.YKernel.get_Y0(self.itrain)
cPredict.zPremax = cPredict.ZTest.max(0)
cPredict.iPredCat = cPredict.ZTest.argmax(0)
cPredict.zPred = Y0[cPredict.iPredCat]
if self.itestmode == 2 and self.YKernel.kernel_params.kernel_type == 0:
cPredict.knnPredCat=np.argsort \
(-cPredict.ZTest,axis=0)[:self.testknn,:]
elif self.itestmode == 3:
n = YTrain.shape[1]
nsort = 5
for i in range(mtest):
cPredict.zPred[i, :] = np.zeros(n)
iones = np.argsort(-Zw[:, i])[:nsort]
for j in range(nsort - 1):
if Zw[iones[j],i]<0.04 or \
Zw[iones[j],i]>10.0*Zw[iones[j+1],i]:
iones = iones[:(j + 1)]
break
cPredict.zPred[i, iones] = 1
elif self.itestmode == 1:
YTrain = self.YKernel.get_train(self.itrain)
ndim = YTrain.shape[1]
i2boro = np.argsort(-Zw0, axis=0)[0:self.testknn, :].T
cPredict.zPred = np.zeros((mtest, ndim))
for i in range(mtest):
zsum = np.zeros(ndim)
for j in range(self.testknn):
zsum += YTrain[i2boro[i, j]]
cPredict.zPred[i, :] = 1 * (zsum >= self.testknn / 2)
elif self.YKernel.ifeature == 1:
Y0 = self.YKernel.get_Y0(self.itrain)
cPredict.ZTest=np.dot(KYcross.T,(np.outer(cOptDual.alpha, \
np.ones(mtest))*KXcross))
cPredict.zPremax = cPredict.ZTest.max(0)
cPredict.iPredCat = cPredict.ZTest.argmax(0)
cPredict.zPred = Y0[cPredict.iPredCat]
if self.itestmode == 2:
cPredict.knnPredCat=np.argsort \
(-cPredict.ZTest,axis=0)[:self.testknn,:]
elif self.iperceptron == 1:
if itraindata == 0:
X = self.XKernel[0].get_train_norm(self.itrain)
for i in range(1, len(self.XKernel)):
X = np.hstack(
(X, self.XKernel[i].get_train_norm(self.itrain)))
else:
X = self.XKernel[0].get_test_norm(itest)
for i in range(1, len(self.XKernel)):
X = np.hstack((X, self.XKernel[i].get_test_norm(itest)))
cPredict = base.cls_predict()
xmean = np.mean(X)
mtest = X.shape[0]
Zw=np.dot(cOptDual.W, \
np.hstack((X,np.ones((mtest,1))*xmean)).T)
cPredict.ZTest = np.dot(YTrain, Zw)
cPredict.zPremax = cPredict.ZTest.max(0)
cPredict.iPredCat = cPredict.ZTest.argmax(0)
cPredict.zPred = YTrain[cPredict.iPredCat]
if self.itestmode == 2:
cPredict.knnPredCat=np.argsort(-cPredict.ZTest, \
axis=0)[:self.testknn,:]
return (cPredict)
def test_dynamic_prog(cDataTest, cOptDual):
(mtrain, mtest) = cDataTest.KXcross.shape
ndim = cDataTest.YTrainNorm.shape[1]
cPredict = mmr_base_classes.cls_predict()
cPredict.zPred = np.zeros((mtest, ndim))
beta = cDataTest.KXcross * np.outer(cOptDual.alpha, np.ones(mtest))
beta = beta.T
## inverse covariance
## ymean=np.mean(cDataTest.YTrainNorm,axis=0)
Y = cDataTest.YTrainNorm
## Yn=Y-np.outer(np.ones(mtrain),ymean)
## Ycov=np.dot(Yn.T,Yn)/mtrain
## Ycovinv=np_lin.pinv(Ycov,rcond=1e-2)/(2*(params.output.ipar1**2))
Ycovinv = np.eye(ndim) / (2 * (cDataTest.kernel_params.ipar1**2))
CY = np.dot(Ycovinv, Y.T)
cy0 = np.sum(Y * CY.T, axis=1)
y0 = np.zeros(ndim)
## y0c=np.zeros(ndim)
## e=np.ones(mtrain)
rtime = time.time()
ntime = 5
xdynstack0 = np.zeros((ntime, ndim))
xdynstack1 = np.zeros((ntime, ndim, ndim))
xdynstack2 = np.zeros((ntime, ndim, mtrain))
for it in range(mtest):
## setup of the dynamic programming
## load time 0
f = np.dot(np.exp(-cy0), beta[it])
for i in range(ntime):
for j in range(ndim):
xdynstack0[i, j] = f
xdynstack1[i, j, :] = y0
xdynstack2[i, j, :] = cy0
## print('>>>',f)
for t in range(0, ntime - 1):
if t == 0:
nitem = 1
else:
nitem = ndim
for i1 in range(nitem):
## print(t,i1)
f = xdynstack0[t, i1]
y0 = xdynstack1[t, i1, :]
cy0 = xdynstack2[t, i1, :]
fmax = 0
y0c = np.dot(Ycovinv, y0)
for i2 in range(ndim):
cy_new = cy0 + Ycovinv[i2, i2] + y0c[i2] - CY[i2]
fnew = np.dot(np.exp(-cy_new), beta[it])
if fnew > xdynstack0[t + 1, i2]:
## print(t,i1,i2,fnew)
y = np.copy(y0)
y[i2] = 1
xdynstack0[t + 1, i2] = fnew
xdynstack1[t + 1, i2, :] = y
xdynstack2[t + 1, i2, :] = cy_new
## else:
## y=np.copy(y0)
## dynstack[t+1][i2]=(f,y,y0c)
if fnew > fmax:
fmax = fnew
## print(f,fmax)
## print(t,i1)
## find the best one
fbest = 0
ibest = -1
for i in range(ndim):
if fbest < xdynstack0[ntime - 1, i]:
fbest = xdynstack0[ntime - 1, i]
ibest = i
cPredict.zPred[it, :] = xdynstack1[ntime - 1, ibest, :]
if it % 10 == 0:
print(it, (time.time() - rtime) / (it + 1))
return (cPredict)