def pls_cal(dbfile,maskfile,outpath,which_elem,testfold,nc,normtype=1,mincomp=0,maxcomp=100,plstype='mlpy',keepfile=None,removefile=None,cal_dir=None,masterlist_file=None,compfile=None,name_sub_file=None,foldfile=None,nfolds=7,seed=None,n_bag=None,skscale=False,n_boost=None,max_samples=0.1,n_elems=9):
plstype_string=plstype
if n_bag!=None:
plstype_string=plstype+'_bag'
if n_boost!=None:
plstype_string=plstype+'_boost'
if skscale==True:
plstype_string=plstype+'_scale'
print('Reading database')
sys.stdout.flush()
spectra,comps,spect_index,names,labels,wvl=ccam.read_db(dbfile,compcheck=True,n_elems=n_elems)
oxides=labels[2:]
compindex=numpy.where(oxides==which_elem)[0]
print('Choosing spectra')
which_removed=outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_removed.csv'
spectra,names,spect_index,comps=ccam.choose_spectra(spectra,spect_index,names,comps,compindex,mincomp=mincomp,maxcomp=maxcomp,keepfile=keepfile,removefile=removefile,which_removed=which_removed)
print('Masking spectra')
spectra,wvl=ccam.mask(spectra,wvl,maskfile)
print('Normalizing spectra')
spectra=ccam.normalize(spectra,wvl,normtype=normtype)
print('Assigning Folds')
if foldfile!=None:
#if a fold file is specified, use it
folds=ccam.folds(foldfile,names)
else:
#otherwise, define random folds
folds=ccam.random_folds(names,nfolds,seed=seed)
names_nofold=names[(folds==0)]
spect_index_nofold=spect_index[(folds==0)]
#write a file containing the samples not assigned to folds
with open(which_removed,'ab') as writefile:
writer=csv.writer(writefile,delimiter=',',)
for i in range(len(names_nofold)):
writer.writerow([names_nofold[i],spect_index_nofold[i],'No Fold'])
#remove spectra that are not assigned to any fold
spectra=spectra[(folds!=0),:]
spect_index=spect_index[(folds!=0)]
names=names[(folds!=0)]
comps=comps[(folds!=0),:]
folds=folds[(folds!=0)]
print('Defining Training and Test Sets')
spectra_train=spectra[(folds!=testfold)]
spect_index_train=spect_index[(folds!=testfold)]
names_train=names[(folds!=testfold)]
comps_train=comps[(folds!=testfold),compindex]
folds_train=folds[(folds!=testfold)]
folds_train_unique=numpy.unique(folds_train)
spectra_test=spectra[(folds==testfold)]
spect_index_test=spect_index[(folds==testfold)]
names_test=names[(folds==testfold)]
comps_test=comps[(folds==testfold),compindex]
folds_test=folds[(folds==testfold)]
print('Do Leave One Label Out (LOLO) cross validation with all folds but the test set')
#define array to hold cross validation predictions and RMSEs
train_predict_cv=numpy.zeros((len(names_train),nc))
RMSECV=numpy.zeros(nc)
for i in folds_train_unique:
print('Holding out fold #'+str(i))
if skscale==False:
#mean center those spectra left in
#X_cv_in1,X_cv_in_mean1=meancenter.ccam_meancenter(spectra_train[(folds_train!=i),:])
X_cv_in,X_cv_in_mean=ccam.meancenter(spectra_train[(folds_train!=i),:])
#and those left out
X_cv_out=ccam.meancenter(spectra_train[(folds_train==i),:],X_mean=X_cv_in_mean)[0]
#mean center compositions left in
Y_cv_in,Y_cv_in_mean=ccam.meancenter(comps_train[(folds_train!=i)])
if skscale==True:
X_cv_in=spectra_train[(folds_train!=i),:]
X_cv_out=spectra_train[(folds_train==i),:]
Y_cv_in=comps_train[(folds_train!=i)]
Y_cv_in_mean=0
#step through each number of components
for j in range(1,nc+1):
print('Training Model for '+str(j)+' components')
#train the model
if plstype=='mlpy':
PLS1model=ccam.mlpy_pls.PLS(j)
PLS1model.learn(X_cv_in,Y_cv_in)
#predict the samples held out
train_predict_cv[(folds_train==i),j-1]=PLS1model.pred(X_cv_out)+Y_cv_in_mean
if plstype=='sklearn':
PLS1model=PLSRegression(n_components=j,scale=skscale)
if n_bag==None and n_boost==None:
PLS1model.fit(X_cv_in,Y_cv_in)
train_predict_cv[(folds_train==i),j-1]=numpy.squeeze(PLS1model.predict(X_cv_out)+Y_cv_in_mean)
if n_bag!=None:
PLS1bagged=ensemble.BaggingRegressor(PLS1model,n_estimators=n_bag,max_samples=max_samples,verbose=1)
PLS1bagged.fit(X_cv_in,Y_cv_in)
train_predict_cv[(folds_train==i),j-1]=numpy.squeeze(PLS1bagged.predict(X_cv_out)+Y_cv_in_mean)
if n_boost!=None:
PLS1boosted=ensemble.AdaBoostRegressor(PLS1model,n_estimators=n_boost)
PLS1boosted.fit(X_cv_in,Y_cv_in)
train_predict_cv[(folds_train==i),j-1]=numpy.squeeze(PLS1boosted.predict(X_cv_out)+Y_cv_in_mean)
#calculate RMSECV
for i in range(0,nc):
sqerr=(train_predict_cv[:,i]-comps_train)**2.0
RMSECV[i]=numpy.sqrt(numpy.mean(sqerr))
#mean center full model
if skscale==False:
X,X_mean=ccam.meancenter(spectra_train)
X_test=ccam.meancenter(spectra_test,X_mean=X_mean)[0]
X_all=ccam.meancenter(spectra,X_mean=X_mean)[0]
Y,Y_mean=ccam.meancenter(comps_train)
if skscale==True:
X=spectra_train
X_test=spectra_test
X_all=spectra
Y=comps_train
Y_mean=0
#create arrays for results and RMSEs
trainset_results=numpy.zeros((len(names_train),nc))
testset_results=numpy.zeros((len(names_test),nc))
results=numpy.zeros((len(names),nc))
RMSEP=numpy.zeros(nc)
RMSEC=numpy.zeros(nc)
beta=numpy.zeros((len(X[0,:]),nc))
Q_res=numpy.zeros((len(X[:,0]),nc))
T2=numpy.zeros((len(X[:,0]),nc))
[a,evals,b]=numpy.linalg.svd(numpy.cov(numpy.dot(X,X.transpose())))
evals=numpy.diag(evals**2)
if cal_dir!=None:
print('Reading cal target data')
cal_data,cal_wvl,cal_filelist=ccam.read_ccs(cal_dir)
cal_data,cal_wvl=ccam.mask(cal_data,cal_wvl,maskfile)
cal_data=ccam.normalize(cal_data,cal_wvl,normtype=normtype)
if skscale==True:
cal_data_centered=cal_data
if skscale==False:
cal_data_centered=ccam.meancenter(cal_data,X_mean=X_mean)[0]
RMSEP_cal=numpy.zeros(nc)
RMSEP_cal_good=numpy.zeros(nc)
RMSEP_KGAMEDS=numpy.zeros(nc)
RMSEP_MACUSANITE=numpy.zeros(nc)
RMSEP_NAU2HIS=numpy.zeros(nc)
RMSEP_NAU2LOS=numpy.zeros(nc)
RMSEP_NAU2MEDS=numpy.zeros(nc)
RMSEP_NORITE=numpy.zeros(nc)
RMSEP_PICRITE=numpy.zeros(nc)
RMSEP_SHERGOTTITE=numpy.zeros(nc)
targets,dists,amps,nshots=ccam.target_lookup(cal_filelist,masterlist_file,name_sub_file)
target_comps=ccam.target_comp_lookup(targets,compfile,which_elem)
cal_results=numpy.zeros((len(targets),nc))
model_list=[]
#Now step through each # of components with the full model
for j in range(1,nc+1):
print('Training full model for '+str(j)+' components')
if plstype=='mlpy':
PLS1model=ccam.mlpy_pls.PLS(j)
PLS1model.learn(X,Y)
beta[:,j-1]=PLS1model.beta()
model_list.append([PLS1model])
trainset_results[:,j-1]=PLS1model.pred(X)+Y_mean
testset_results[:,j-1]=PLS1model.pred(X_test)+Y_mean
results[:,j-1]=PLS1model.pred(X_all)+Y_mean
if cal_dir != None:
comps_copy=copy.copy(target_comps)
# if skscale==True:
# cal_results[:,j-1]=PLS1model.pred(cal_data)
# if skscale==False:
cal_results[:,j-1]=PLS1model.pred(cal_data_centered)+Y_mean
RMSEP_KGAMEDS[j-1],RMSEP_MACUSANITE[j-1],RMSEP_NAU2HIS[j-1],RMSEP_NAU2LOS[j-1],RMSEP_NAU2MEDS[j-1],RMSEP_NORITE[j-1],RMSEP_PICRITE[j-1],RMSEP_SHERGOTTITE[j-1],RMSEP_cal_good[j-1]=cal_rmses(targets,nc,target_comps,j,cal_data_centered,Y_mean,mincomp,maxcomp,cal_results)
if plstype=='sklearn':
PLS1model=PLSRegression(n_components=j,scale=skscale)
if n_bag==None and n_boost==None:
PLS1model.fit(X,Y)
T=PLS1model.x_scores_
#There's probably a more efficient way to calculate T2...
for k in range(len(X[:,0])):
T2[k,j-1]=numpy.dot(T[k,:],numpy.dot(numpy.linalg.inv(numpy.dot(T.transpose(),T)),T[k,:]))
E=X-numpy.dot(PLS1model.x_scores_,PLS1model.x_loadings_.transpose())
Q_res[:,j-1]=numpy.dot(E,E.transpose()).diagonal()
trainset_results[:,j-1]=numpy.squeeze(PLS1model.predict(X)+Y_mean)
testset_results[:,j-1]=numpy.squeeze(PLS1model.predict(X_test)+Y_mean)
results[:,j-1]=numpy.squeeze(PLS1model.predict(X_all)+Y_mean)
beta[:,j-1]=numpy.squeeze(PLS1model.coefs)
model_list.append([PLS1model])
if cal_dir != None:
comps_copy=copy.copy(target_comps)
cal_results[:,j-1]=numpy.squeeze(PLS1model.predict(cal_data_centered)+Y_mean)
RMSEP_KGAMEDS[j-1],RMSEP_MACUSANITE[j-1],RMSEP_NAU2HIS[j-1],RMSEP_NAU2LOS[j-1],RMSEP_NAU2MEDS[j-1],RMSEP_NORITE[j-1],RMSEP_PICRITE[j-1],RMSEP_SHERGOTTITE[j-1],RMSEP_cal_good[j-1]=cal_rmses(targets,nc,target_comps,j,cal_data_centered,Y_mean,mincomp,maxcomp,cal_results)
if n_bag!=None:
PLS1bagged=ensemble.BaggingRegressor(PLS1model,n_estimators=n_bag,max_samples=max_samples,verbose=1)
PLS1bagged.fit(X,Y)
trainset_results[:,j-1]=numpy.squeeze(PLS1bagged.predict(X)+Y_mean)
testset_results[:,j-1]=numpy.squeeze(PLS1bagged.predict(X_test)+Y_mean)
results[:,j-1]=numpy.squeeze(PLS1bagged.predict(X_all)+Y_mean)
beta[:,j-1]=None
model_list.append([PLS1bagged])
if cal_dir != None:
comps_copy=copy.copy(target_comps)
cal_results[:,j-1]=numpy.squeeze(PLS1bagged.predict(cal_data_centered)+Y_mean)
RMSEP_KGAMEDS[j-1],RMSEP_MACUSANITE[j-1],RMSEP_NAU2HIS[j-1],RMSEP_NAU2LOS[j-1],RMSEP_NAU2MEDS[j-1],RMSEP_NORITE[j-1],RMSEP_PICRITE[j-1],RMSEP_SHERGOTTITE[j-1],RMSEP_cal_good[j-1]=cal_rmses(targets,nc,target_comps,j,cal_data_centered,Y_mean,mincomp,maxcomp,cal_results)
if n_boost!=None:
PLS1boosted=ensemble.AdaBoostRegressor(PLS1model,n_estimators=n_boost)
PLS1boosted.fit(X,Y)
trainset_results[:,j-1]=numpy.squeeze(PLS1boosted.predict(X)+Y_mean)
testset_results[:,j-1]=numpy.squeeze(PLS1boosted.predict(X_test)+Y_mean)
results[:,j-1]=numpy.squeeze(PLS1boosted.predict(X_all)+Y_mean)
beta[:,j-1]=None
model_list.append([PLS1boosted])
if cal_dir != None:
comps_copy=copy.copy(target_comps)
cal_results[:,j-1]=numpy.squeeze(PLS1boosted.predict(cal_data_centered)+Y_mean)
RMSEP_KGAMEDS[j-1],RMSEP_MACUSANITE[j-1],RMSEP_NAU2HIS[j-1],RMSEP_NAU2LOS[j-1],RMSEP_NAU2MEDS[j-1],RMSEP_NORITE[j-1],RMSEP_PICRITE[j-1],RMSEP_SHERGOTTITE[j-1],RMSEP_cal_good[j-1]=cal_rmses(targets,nc,target_comps,j,cal_data_centered,Y_mean,mincomp,maxcomp,cal_results)
RMSEC[j-1]=numpy.sqrt(numpy.mean((trainset_results[:,j-1]-comps_train)**2.0))
RMSEP[j-1]=numpy.sqrt(numpy.mean((testset_results[:,j-1]-comps_test)**2.0))
with open(outpath+which_elem+'_'+plstype_string+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'.pkl','wb') as picklefile:
pickle.dump(model_list,picklefile)
#if cal_dir is specified, read cal target data and calculate RMSEs
if cal_dir!=None:
n_good_cal=numpy.sum(numpy.array([RMSEP_KGAMEDS,RMSEP_MACUSANITE,RMSEP_NAU2HIS,RMSEP_NAU2LOS,RMSEP_NAU2MEDS,RMSEP_NORITE,RMSEP_PICRITE,RMSEP_SHERGOTTITE])[:,0]!=0)
print(n_good_cal)
RMSEP_cal=(RMSEP_KGAMEDS+RMSEP_MACUSANITE+RMSEP_NAU2HIS+RMSEP_NAU2LOS+RMSEP_NAU2MEDS+RMSEP_NORITE+RMSEP_PICRITE+RMSEP_SHERGOTTITE)/n_good_cal
RMSEP_single_cals=[RMSEP_KGAMEDS,RMSEP_MACUSANITE,RMSEP_NAU2HIS,RMSEP_NAU2LOS,RMSEP_NAU2MEDS,RMSEP_NORITE,RMSEP_PICRITE,RMSEP_SHERGOTTITE,RMSEP_cal]
with open(outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_caltargets_predict.csv','w',newline='') as writefile:
writer=csv.writer(writefile,delimiter=',')
row=['File','Target','Laser Energy','True_Comp']
row.extend(list(range(1,nc+1)))
writer.writerow(row)
for i in range(0,len(targets)):
row=[cal_filelist[i],targets[i],amps[i],target_comps[i]]
row.extend(cal_results[i,:])
writer.writerow(row)
with open(outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_RMSEP_caltargets.csv','w',newline='') as writefile:
writer=csv.writer(writefile,delimiter=',')
writer.writerow(['NC','RMSEP Cal Targets (wt.%)'])
for i in range(0,nc):
writer.writerow([i+1,RMSEP_cal[i]])
ccam.RMSE(RMSECV,RMSEP,RMSEC,which_elem+' RMSEs',outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_RMSE_plot_cal.png',RMSEP_cals=RMSEP_single_cals)
ccam.RMSE(RMSECV,RMSEP,RMSEC,which_elem+' RMSEs',outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_RMSE_plot_cal_good.png',RMSEP_good=RMSEP_cal_good)
# plot RMSEs
ccam.RMSE(RMSECV,RMSEP,RMSEC,which_elem+' RMSEs',outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_RMSE_plot.png')
#Write output info to files
with open(outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_Q_res.csv','w',newline='') as writefile:
writer=csv.writer(writefile,delimiter=',')
row=["Sample","Spectrum","Fold","True Comp"]
row.extend(range(1,nc+1))
writer.writerow(row)
for i in range(0,len(names_train)):
row=[names_train[i],spect_index_train[i],folds_train[i],comps_train[i]]
row.extend(Q_res[i,:])
writer.writerow(row)
with open(outpath+which_elem+'_'+str(mincomp)+'-'+str(maxcomp)+'_quartiles.csv','w',newline='') as writefile:
writer=csv.writer(writefile,delimiter=',')
row=[which_elem]
writer.writerow(row)
row=['Min',numpy.percentile(comps[:,compindex],0)]
writer.writerow(row)
row=['1st Quartile',numpy.percentile(comps[:,compindex],25)]
writer.writerow(row)
row=['Median',numpy.percentile(comps[:,compindex],50)]
writer.writerow(row)
row=['3rd Quartile',numpy.percentile(comps[:,compindex],75)]
writer.writerow(row)
row=['Max',numpy.percentile(comps[:,compindex],100)]
writer.writerow(row)
with open(outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_HotellingT2.csv','w',newline='') as writefile:
writer=csv.writer(writefile,delimiter=',')
row=["Sample","Spectrum","Fold","True Comp"]
row.extend(range(1,nc+1))
writer.writerow(row)
for i in range(0,len(names_train)):
row=[names_train[i],spect_index_train[i],folds_train[i],comps_train[i]]
row.extend(T2[i,:])
writer.writerow(row)
with open(outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_RMSECV.csv','w',newline='') as writefile:
writer=csv.writer(writefile,delimiter=',')
writer.writerow(['NC','RMSECV (wt.%)'])
for i in range(0,nc):
writer.writerow([i+1,RMSECV[i]])
with open(outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_RMSEC.csv','w',newline='') as writefile:
writer=csv.writer(writefile,delimiter=',')
writer.writerow(['NC','RMSEC (wt.%)'])
for i in range(0,nc):
writer.writerow([i+1,RMSEC[i]])
with open(outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_RMSEP.csv','w',newline='') as writefile:
writer=csv.writer(writefile,delimiter=',')
writer.writerow(['NC','RMSEP (wt.%)'])
for i in range(0,nc):
writer.writerow([i+1,RMSEP[i]])
with open(outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_cv_predict.csv','w',newline='') as writefile:
writer=csv.writer(writefile,delimiter=',')
row=['Sample','Spectrum','Fold','True_Comp']
row.extend(range(1,nc+1))
writer.writerow(row)
for i in range(0,len(names_train)):
row=[names_train[i],spect_index_train[i],folds_train[i],comps_train[i]]
row.extend(train_predict_cv[i,:])
writer.writerow(row)
with open(outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_train_predict.csv','w',newline='') as writefile:
writer=csv.writer(writefile,delimiter=',')
row=['Sample','Spectrum','Fold','True_Comp']
row.extend(range(1,nc+1))
writer.writerow(row)
for i in range(0,len(names_train)):
row=[names_train[i],spect_index_train[i],folds_train[i],comps_train[i]]
row.extend(trainset_results[i,:])
writer.writerow(row)
with open(outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_test_predict.csv','w',newline='') as writefile:
writer=csv.writer(writefile,delimiter=',')
row=['Sample','Spectrum','Fold','True_Comp']
row.extend(range(1,nc+1))
writer.writerow(row)
for i in range(0,len(names_test)):
row=[names_test[i],spect_index_test[i],folds_test[i],comps_test[i]]
row.extend(testset_results[i,:])
writer.writerow(row)
with open(outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_all_predict.csv','w',newline='') as writefile:
writer=csv.writer(writefile,delimiter=',')
row=['Sample','Spectrum','Fold','True_Comp']
row.extend(range(1,nc+1))
writer.writerow(row)
for i in range(0,len(names)):
row=[names[i],spect_index[i],folds[i],comps[i,compindex]]
row.extend(results[i,:])
writer.writerow(row)
with open(outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_beta_coeffs.csv','w',newline='') as writefile:
writer=csv.writer(writefile,delimiter=',')
row=['wvl']
row.extend(range(1,nc+1))
writer.writerow(row)
for i in range(0,len(wvl)):
row=[wvl[i]]
row.extend(beta[i,:])
writer.writerow(row)
if skscale==False:
with open(outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_meancenters.csv','w',newline='') as writefile:
writer=csv.writer(writefile,delimiter=',')
writer.writerow([which_elem+' mean',Y_mean])
for i in range(0,len(wvl)):
row=[wvl[i],X_mean[i]]
writer.writerow(row)
with open(outpath+which_elem+'_'+plstype_string+'_nc'+str(nc)+'_norm'+str(normtype)+'_'+str(mincomp)+'-'+str(maxcomp)+'_inputinfo.csv','w',newline='') as writefile:
writer=csv.writer(writefile,delimiter=',')
writer.writerow(['Spectral database =',dbfile])
writer.writerow(['Spectra Kept =',keepfile])
writer.writerow(['Spectra Removed =',which_removed])
writer.writerow(['Fold Definition =',foldfile])
writer.writerow(['Test Fold =',maskfile])
writer.writerow(['Mask File =',maskfile])
writer.writerow(['Algorithm =',plstype_string])
writer.writerow(['# of components =',nc])
writer.writerow(['Normalization Type =',normtype])
writer.writerow(['Composition Min. =',mincomp])
writer.writerow(['Composition Max. =',maxcomp])
def pls_cal(dbfile,
foldfile,
maskfile,
outpath,
which_elem,
testfold,
nc,
normtype=3,
mincomp=0,
maxcomp=100,
plstype='mlpy',
keepfile=None,
removefile=None,
cal_dir=None,
masterlist_file=None,
compfile=None,
name_sub_file=None):
print 'Reading database'
sys.stdout.flush()
spectra, comps, spect_index, names, labels, wvl = ccam.read_db(
dbfile, compcheck=True)
oxides = labels[2:]
compindex = numpy.where(oxides == which_elem)[0]
print 'Choosing spectra'
which_removed = outpath + which_elem + '_' + plstype + '_nc' + str(
nc) + '_norm' + str(normtype) + '_' + str(mincomp) + '-' + str(
maxcomp) + '_removed.csv'
spectra, names, spect_index, comps = ccam.choose_spectra(
spectra,
spect_index,
names,
comps,
compindex,
mincomp=mincomp,
maxcomp=maxcomp,
keepfile=keepfile,
removefile=removefile,
which_removed=which_removed)
print 'Masking spectra'
spectra, wvl = ccam.mask(spectra, wvl, maskfile)
print 'Normalizing spectra'
spectra = ccam.normalize(spectra, wvl, normtype=normtype)
print 'Assigning Folds'
folds = ccam.folds(foldfile, names)
names_nofold = names[(folds == 0)]
spect_index_nofold = spect_index[(folds == 0)]
#write a file containing the samples not assigned to folds
with open(which_removed, 'ab') as writefile:
writer = csv.writer(
writefile,
delimiter=',',
)
for i in range(len(names_nofold)):
writer.writerow(
[names_nofold[i], spect_index_nofold[i], 'No Fold'])
#remove spectra that are not assigned to any fold
spectra = spectra[(folds != 0), :]
spect_index = spect_index[(folds != 0)]
names = names[(folds != 0)]
comps = comps[(folds != 0), :]
folds = folds[(folds != 0)]
print 'Defining Training and Test Sets'
spectra_train = spectra[(folds != testfold)]
spect_index_train = spect_index[(folds != testfold)]
names_train = names[(folds != testfold)]
comps_train = comps[(folds != testfold), compindex]
folds_train = folds[(folds != testfold)]
folds_train_unique = numpy.unique(folds_train)
spectra_test = spectra[(folds == testfold)]
spect_index_test = spect_index[(folds == testfold)]
names_test = names[(folds == testfold)]
comps_test = comps[(folds == testfold), compindex]
folds_test = folds[(folds == testfold)]
print 'Do Leave One Label Out (LOLO) cross validation with all folds but the test set'
#define array to hold cross validation predictions and RMSEs
train_predict_cv = numpy.zeros((len(names_train), nc))
RMSECV = numpy.zeros(nc)
for i in folds_train_unique:
print 'Holding out fold #' + str(i)
#mean center those spectra left in
#X_cv_in1,X_cv_in_mean1=meancenter.ccam_meancenter(spectra_train[(folds_train!=i),:])
X_cv_in, X_cv_in_mean = ccam.meancenter(
spectra_train[(folds_train != i), :])
#and those left out
X_cv_out = ccam.meancenter(spectra_train[(folds_train == i), :],
X_mean=X_cv_in_mean)[0]
#mean center compositions left in
Y_cv_in, Y_cv_in_mean = ccam.meancenter(
comps_train[(folds_train != i)])
#step through each number of components
for j in range(1, nc + 1):
print 'Training PLS Model for ' + str(j) + ' components'
#train the model
if plstype == 'mlpy':
PLS1model = mlpy.pls.PLS(j)
PLS1model.learn(X_cv_in, Y_cv_in)
#predict the samples held out
train_predict_cv[(folds_train == i), j -
1] = PLS1model.pred(X_cv_out) + Y_cv_in_mean
if plstype == 'sklearn':
PLS1model = PLSRegression(n_components=nc)
PLS1model.fit(X_cv_in, Y_cv_in)
train_predict_cv[
(folds_train == i),
j - 1] = PLS1model.predict(X_cv_out) + Y_cv_in_mean
#calculate RMSECV
for i in range(0, nc):
sqerr = (train_predict_cv[:, i] - comps_train)**2.0
RMSECV[i] = numpy.sqrt(numpy.mean(sqerr))
#mean center full model
X, X_mean = ccam.meancenter(spectra_train)
X_test = ccam.meancenter(spectra_test, X_mean=X_mean)[0]
Y, Y_mean = ccam.meancenter(comps_train)
#create arrays for results and RMSEs
trainset_results = numpy.zeros((len(names_train), nc))
testset_results = numpy.zeros((len(names_test), nc))
RMSEP = numpy.zeros(nc)
RMSEC = numpy.zeros(nc)
beta = numpy.zeros((len(X_mean), nc))
#Now step through each # of components with the full model
for j in range(1, nc + 1):
print 'Training full model for ' + str(j) + ' components'
if plstype == 'mlpy':
PLS1model = mlpy.pls.PLS(j)
PLS1model.learn(X, Y)
beta[:, j - 1] = PLS1model.beta()
trainset_results[:, j - 1] = PLS1model.pred(X) + Y_mean
testset_results[:, j - 1] = PLS1model.pred(X_test) + Y_mean
if plstype == 'sklearn':
PLS1model = PLSRegression(n_components=nc)
PLS1model.fit(X, Y)
print 'stop'
RMSEC[j - 1] = numpy.sqrt(
numpy.mean((trainset_results[:, j - 1] - comps_train)**2.0))
RMSEP[j - 1] = numpy.sqrt(
numpy.mean((testset_results[:, j - 1] - comps_test)**2.0))
#if cal_dir is specified, read cal target data and calculate RMSEs
if cal_dir != None:
cal_data, cal_wvl, cal_filelist = ccam.read_ccs(cal_dir)
cal_data, cal_wvl = ccam.mask(cal_data, cal_wvl, maskfile)
cal_data = ccam.normalize(cal_data, cal_wvl, normtype=normtype)
RMSEP_cal = numpy.zeros(nc)
RMSEP_KGAMEDS = numpy.zeros(nc)
RMSEP_MACUSANITE = numpy.zeros(nc)
RMSEP_NAU2HIS = numpy.zeros(nc)
RMSEP_NAU2LOS = numpy.zeros(nc)
RMSEP_NAU2MEDS = numpy.zeros(nc)
RMSEP_NORITE = numpy.zeros(nc)
RMSEP_PICRITE = numpy.zeros(nc)
RMSEP_SHERGOTTITE = numpy.zeros(nc)
targets, dists, amps = ccam.target_lookup(cal_filelist,
masterlist_file,
name_sub_file)
target_comps = ccam.target_comp_lookup(targets, compfile, which_elem)
cal_results = numpy.zeros((len(targets), nc))
for i in range(nc):
comps_copy = copy.copy(target_comps)
cal_results[:, i] = ccam.pls_unk(cal_data,
i + 1,
beta=beta[:, i],
X_mean=X_mean,
Y_mean=Y_mean)
#RMSEP_cal[i]=numpy.sqrt(numpy.mean((cal_results[:,i]-target_comps)**2))
cal_results[(comps_copy < mincomp), i] = 0
cal_results[(comps_copy > maxcomp), i] = 0
comps_copy[(comps_copy < mincomp)] = 0
comps_copy[(comps_copy > maxcomp)] = 0
RMSEP_KGAMEDS[i] = numpy.sqrt(
numpy.mean((cal_results[(targets == 'KGAMEDS'), i] -
comps_copy[(targets == 'KGAMEDS')])**2))
RMSEP_MACUSANITE[i] = numpy.sqrt(
numpy.mean((cal_results[(targets == 'MACUSANITE'), i] -
comps_copy[(targets == 'MACUSANITE')])**2))
RMSEP_NAU2HIS[i] = numpy.sqrt(
numpy.mean((cal_results[(targets == 'NAU2HIS'), i] -
comps_copy[(targets == 'NAU2HIS')])**2))
RMSEP_NAU2LOS[i] = numpy.sqrt(
numpy.mean((cal_results[(targets == 'NAU2LOS'), i] -
comps_copy[(targets == 'NAU2LOS')])**2))
RMSEP_NAU2MEDS[i] = numpy.sqrt(
numpy.mean((cal_results[(targets == 'NAU2MEDS'), i] -
comps_copy[(targets == 'NAU2MEDS')])**2))
RMSEP_NORITE[i] = numpy.sqrt(
numpy.mean((cal_results[(targets == 'NORITE'), i] -
comps_copy[(targets == 'NORITE')])**2))
RMSEP_PICRITE[i] = numpy.sqrt(
numpy.mean((cal_results[(targets == 'PICRITE'), i] -
comps_copy[(targets == 'PICRITE')])**2))
RMSEP_SHERGOTTITE[i] = numpy.sqrt(
numpy.mean((cal_results[(targets == 'SHERGOTTITE'), i] -
comps_copy[(targets == 'SHERGOTTITE')])**2))
n_good_cal = len(numpy.unique(comps_copy)) - 1
RMSEP_cal = (RMSEP_KGAMEDS + RMSEP_MACUSANITE + RMSEP_NAU2HIS +
RMSEP_NAU2LOS + RMSEP_NAU2MEDS + RMSEP_NORITE +
RMSEP_PICRITE + RMSEP_SHERGOTTITE) / n_good_cal
RMSEP_single_cals = [
RMSEP_KGAMEDS, RMSEP_MACUSANITE, RMSEP_NAU2HIS, RMSEP_NAU2LOS,
RMSEP_NAU2MEDS, RMSEP_NORITE, RMSEP_PICRITE, RMSEP_SHERGOTTITE,
RMSEP_cal
]
with open(
outpath + which_elem + '_' + str(mincomp) + '-' +
str(maxcomp) + '_' + plstype + '_nc' + str(nc) + '_norm' +
str(normtype) + '_caltargets_predict.csv', 'wb') as writefile:
writer = csv.writer(writefile, delimiter=',')
row = ['File', 'Target', 'Laser Energy', 'True_Comp']
row.extend(range(1, nc + 1))
writer.writerow(row)
for i in range(0, len(targets)):
row = [cal_filelist[i], targets[i], amps[i], target_comps[i]]
row.extend(cal_results[i, :])
writer.writerow(row)
with open(
outpath + which_elem + '_' + str(mincomp) + '-' +
str(maxcomp) + '_' + plstype + '_nc' + str(nc) + '_norm' +
str(normtype) + '_RMSECP_caltargets.csv', 'wb') as writefile:
writer = csv.writer(writefile, delimiter=',')
writer.writerow(['NC', 'RMSECP Cal Targets (wt.%)'])
for i in range(0, nc):
writer.writerow([i + 1, RMSEP_cal[i]])
ccam.plots.RMSE(RMSECV,
RMSEP,
RMSEC,
which_elem + ' RMSEs',
outpath + which_elem + '_' + str(mincomp) + '-' +
str(maxcomp) + '_' + plstype + '_nc' + str(nc) +
'_norm' + str(normtype) + '_RMSE_plot_cal.png',
RMSEP_cals=RMSEP_single_cals)
# plot RMSEs
ccam_plots.ccam_plot_RMSE(
RMSECV, RMSEP, RMSEC, which_elem + 'RMSEs', outpath + which_elem +
'_' + plstype + '_nc' + str(nc) + '_norm' + str(normtype) + '_' +
str(mincomp) + '-' + str(maxcomp) + '_RMSE_plot.png')
#Write output info to files
print outpath + which_elem + '_' + plstype + '_nc' + str(
nc) + '_norm' + str(normtype) + '_' + str(mincomp) + '-' + str(
maxcomp) + '_RMSECV.csv'
with open(
outpath + which_elem + '_' + plstype + '_nc' + str(nc) + '_norm' +
str(normtype) + '_' + str(mincomp) + '-' + str(maxcomp) +
'_RMSECV.csv', 'wb') as writefile:
writer = csv.writer(writefile, delimiter=',')
writer.writerow(['NC', 'RMSECV (wt.%)'])
for i in range(0, nc):
writer.writerow([i + 1, RMSECV[i]])
with open(
outpath + which_elem + '_' + plstype + '_nc' + str(nc) + '_norm' +
str(normtype) + '_' + str(mincomp) + '-' + str(maxcomp) +
'_RMSEC.csv', 'wb') as writefile:
writer = csv.writer(writefile, delimiter=',')
writer.writerow(['NC', 'RMSEC (wt.%)'])
for i in range(0, nc):
writer.writerow([i + 1, RMSEC[i]])
with open(
outpath + which_elem + '_' + plstype + '_nc' + str(nc) + '_norm' +
str(normtype) + '_' + str(mincomp) + '-' + str(maxcomp) +
'_RMSEP.csv', 'wb') as writefile:
writer = csv.writer(writefile, delimiter=',')
writer.writerow(['NC', 'RMSEP (wt.%)'])
for i in range(0, nc):
writer.writerow([i + 1, RMSEP[i]])
with open(
outpath + which_elem + '_' + plstype + '_nc' + str(nc) + '_norm' +
str(normtype) + '_' + str(mincomp) + '-' + str(maxcomp) +
'_cv_predict.csv', 'wb') as writefile:
writer = csv.writer(writefile, delimiter=',')
row = ['Sample', 'Spectrum', 'Fold', 'True_Comp']
row.extend(range(1, nc + 1))
writer.writerow(row)
for i in range(0, len(names_train)):
row = [
names_train[i], spect_index_train[i], folds_train[i],
comps_train[i]
]
row.extend(train_predict_cv[i, :])
writer.writerow(row)
with open(
outpath + which_elem + '_' + plstype + '_nc' + str(nc) + '_norm' +
str(normtype) + '_' + str(mincomp) + '-' + str(maxcomp) +
'_train_predict.csv', 'wb') as writefile:
writer = csv.writer(writefile, delimiter=',')
row = ['Sample', 'Spectrum', 'Fold', 'True_Comp']
row.extend(range(1, nc + 1))
writer.writerow(row)
for i in range(0, len(names_train)):
row = [
names_train[i], spect_index_train[i], folds_train[i],
comps_train[i]
]
row.extend(trainset_results[i, :])
writer.writerow(row)
with open(
outpath + which_elem + '_' + plstype + '_nc' + str(nc) + '_norm' +
str(normtype) + '_' + str(mincomp) + '-' + str(maxcomp) +
'_test_predict.csv', 'wb') as writefile:
writer = csv.writer(writefile, delimiter=',')
row = ['Sample', 'Spectrum', 'Fold', 'True_Comp']
row.extend(range(1, nc + 1))
writer.writerow(row)
for i in range(0, len(names_test)):
row = [
names_test[i], spect_index_test[i], folds_test[i],
comps_test[i]
]
row.extend(testset_results[i, :])
writer.writerow(row)
with open(
outpath + which_elem + '_' + plstype + '_nc' + str(nc) + '_norm' +
str(normtype) + '_' + str(mincomp) + '-' + str(maxcomp) +
'_beta_coeffs.csv', 'wb') as writefile:
writer = csv.writer(writefile, delimiter=',')
row = ['wvl']
row.extend(range(1, nc + 1))
writer.writerow(row)
for i in range(0, len(wvl)):
row = [wvl[i]]
row.extend(beta[i, :])
writer.writerow(row)
with open(
outpath + which_elem + '_' + plstype + '_nc' + str(nc) + '_norm' +
str(normtype) + '_' + str(mincomp) + '-' + str(maxcomp) +
'_meancenters.csv', 'wb') as writefile:
writer = csv.writer(writefile, delimiter=',')
writer.writerow([which_elem + ' mean', Y_mean])
for i in range(0, len(wvl)):
row = [wvl[i], X_mean[i]]
writer.writerow(row)
with open(
outpath + which_elem + '_' + plstype + '_nc' + str(nc) + '_norm' +
str(normtype) + '_' + str(mincomp) + '-' + str(maxcomp) +
'_inputinfo.csv', 'wb') as writefile:
writer = csv.writer(writefile, delimiter=',')
writer.writerow(['Spectral database =', dbfile])
writer.writerow(['Spectra Kept =', keepfile])
writer.writerow(['Spectra Removed =', which_removed])
writer.writerow(['Fold Definition =', foldfile])
writer.writerow(['Test Fold =', maskfile])
writer.writerow(['Mask File =', maskfile])
writer.writerow(['Algorithm =', plstype])
writer.writerow(['# of components =', nc])
writer.writerow(['Normalization Type =', normtype])
writer.writerow(['Composition Min. =', mincomp])
writer.writerow(['Composition Max. =', maxcomp])