def pls_cv(self,ncomp_range=range(1,21),plot=False,verbose=False,
osc_params=(10,1)):
# Separating X from Y for PLS
X=self.df[self.freqs].to_numpy()
Y=self.df[self.y_name].to_numpy().reshape(-1, 1)
sample_std=np.std(self.df[self.y_name])
# CV based on measurement day
if self.cval=="MD":
cv = LeaveOneGroupOut()
folds=list(cv.split(X=X,y=Y,groups=self.df[self.date_name]))
# kfold CV
elif self.cval=="kfold":
cv = KFold(n_splits=self.cval_param)
folds=list(cv.split(X))
else:
raise InputError("Invalid CV type!")
# Array for storing CV errors
cv_RMSE_all=np.zeros([len(folds),len(ncomp_range)])
i=0
for train, val in folds:
# If OSC model specified
if len(osc_params)==2:
osc=OSC(nicomp=osc_params[0],ncomp=osc_params[1])
osc.fit(X[train], Y[train])
X_train_osc=osc.X_osc
X_val_osc=osc.transform(X[val])
j=0
for ncomp in ncomp_range:
pls = PLSRegression(n_components=ncomp,scale=False)
if len(osc_params)==2:
pls.fit(X_train_osc, Y[train])
cv_RMSE_all[i,j]=metrics.mean_squared_error(
Y[val], pls.predict(X_val_osc))**0.5
else:
pls.fit(X[train], Y[train])
cv_RMSE_all[i,j]=metrics.mean_squared_error(
Y[val], pls.predict(X[val]))**0.5
j=j+1
i=i+1
# Printing and plotting CV results
cv_RMSE_ncomp=np.mean(cv_RMSE_all,axis=0)
cv_RPD_ncomp=sample_std/cv_RMSE_ncomp
if plot:
fig = plt.figure(figsize=(12,8))
plt.gca().xaxis.grid(True)
plt.xticks(ncomp_range)
plt.ylabel("RPD")
plt.xlabel("Number of components")
plt.plot(ncomp_range,cv_RPD_ncomp)
# Best model
rpd_best=max(cv_RPD_ncomp)
ncomp_best=ncomp_range[cv_RMSE_ncomp.argmin()]
if verbose:
print("Best RMSE: ",min(cv_RMSE_ncomp))
print("Best RPD: ",max(cv_RPD_ncomp))
print("Number of latent components: ",ncomp_range[cv_RMSE_ncomp.argmin()])
return (ncomp_best,rpd_best)
def fssregression_cv(self,inner_cv="kfold",inner_cv_param=5,maxvar=2,verbose=False,
osc_params=(10,1)):
#inner CV can be "kfold" or "none"
# Separating X from Y for PLS
X=self.df[self.freqs]
Y=self.df[self.y_name]
# Create list for selected variables
best_vars=[]
reg = FSSRegression(inner_cv,inner_cv_param,maxvar)
# CV based on measurement day
if self.cval=="MD":
cv = LeaveOneGroupOut()
folds=list(cv.split(X=X,y=Y,groups=self.df[self.date_name]))
# kfold CV
elif self.cval=="kfold":
cv = KFold(n_splits=self.cval_param)
folds=list(cv.split(X))
else:
raise InputError("Invalid CV type!")
i=0
#Array for cv values
cv_RMSE_all=np.zeros([len(folds)])
for train,val in folds:
# If OSC model specified
if len(osc_params)==2:
osc=OSC(nicomp=osc_params[0],ncomp=osc_params[1])
# FSSR needs column names, so it uses pandas, but osc uses numpy arrays
osc.fit(X.iloc[train].to_numpy(), Y.iloc[train].to_numpy().reshape(-1,1))
X_train_osc=pd.DataFrame(data=osc.X_osc,columns=self.freqs)
X_val_osc=pd.DataFrame(data=osc.transform(X.iloc[val].to_numpy()),columns=self.freqs)
# Fit and predict
reg.fit(X_train_osc, Y.iloc[train])
cv_RMSE_all[i]=metrics.mean_squared_error(
Y.iloc[val], reg.predict(X_val_osc))**0.5
best_vars.append(reg.bestvar)
else:
reg.fit(X.iloc[train], Y.iloc[train])
cv_RMSE_all[i]=metrics.mean_squared_error(
Y.iloc[val], reg.predict(X.iloc[val]))**0.5
best_vars.append(reg.bestvar)
i=i+1
cv_RMSE=np.mean(cv_RMSE_all)
rpd=np.std(self.df[self.y_name])/cv_RMSE
if verbose:
print("RMSE: ",cv_RMSE)
print("RPD: ",rpd)
print("Selected freqs: ",best_vars)
k=0
for day in self.df[self.date_name].unique():
print("Date: {0}, Measurements: {1:.0f}, RMSE: {2:.2f}, selected vars: {3}"
.format(
np.datetime_as_string(day,unit='D'),
sum(self.df[self.date_name]==day),
cv_RMSE_all[k],
len(best_vars[k])))
k=k+1
return(rpd)
def __init__(self):
super(LongWebW, self).__init__()
self.grid = QGridLayout(self)
self.osc = OSC()
self.osc.clear()
self.view_list = [LongWebView(), LongWebView()]
self.view_list[0].setSizePolicy(QSizePolicy.MinimumExpanding,QSizePolicy.Preferred)
self.view_list[1].setSizePolicy(QSizePolicy.MinimumExpanding,QSizePolicy.Preferred)
self.onData_list = [[], []] # [[(cur_set, onset),], [(cur_set, onset),]
self.offData_list = [[], []] # [[(cur_set, offset),], [(cur_set, offset),]
self.onOff_list = [[], []]
self.grid.addWidget(self.view_list[0], 0,0,1,1)
self.grid.addWidget(self.view_list[1], 1,0,1,1)
self.setMinimumWidth(1200)
self.setFixedHeight(700)
self.cur_notes_set = set()
self.lock = threading.RLock()
self.idx = 0
self.auto = False
self.renewParserT("molihua.abc")
def ipls_cv(self,version="basic",nint_list=[8,16,32],ncomp_range=range(1,10),
inner_cv="kfold",inner_cv_param=5,verbose=True,
osc_params=(10,1)):
X=self.df[self.freqs]
Y=self.df[self.y_name]
# CV based on measurement day
if self.cval=="MD":
cv = LeaveOneGroupOut()
folds=list(cv.split(X=X,y=Y,groups=self.df[self.date_name]))
# kfold CV
elif self.cval=="kfold":
cv = KFold(n_splits=self.cval_param)
folds=list(cv.split(X))
else:
raise InputError("Invalid CV type!")
#Array for cv values
cv_RMSE_all=np.zeros([len(folds),len(ncomp_range),len(nint_list)])
i=0
for train,val in folds:
# If OSC model specified
if len(osc_params)==2:
osc=OSC(nicomp=osc_params[0],ncomp=osc_params[1])
# IPLS needs column names, so it uses pandas, but osc uses numpy arrays
osc.fit(X.iloc[train].to_numpy(), Y.iloc[train].to_numpy().reshape(-1,1))
X_train_osc=pd.DataFrame(data=osc.X_osc,columns=self.freqs)
X_val_osc=pd.DataFrame(data=osc.transform(X.iloc[val].to_numpy()),columns=self.freqs)
j=0
for ncomp in ncomp_range:
k=0
for nint in nint_list:
ipls_obj=IntervalPLSRegression(ncomp=ncomp,nint=nint,
cv_type=inner_cv,cv_param=inner_cv_param)
if len(osc_params)==2:
ipls_obj.fit(X_train_osc, Y.iloc[train])
cv_RMSE_all[i,j,k]=metrics.mean_squared_error(
Y.iloc[val], ipls_obj.predict(X_val_osc))**0.5
else:
ipls_obj.fit(X.iloc[train], Y.iloc[train])
cv_RMSE_all[i,j,k]=metrics.mean_squared_error(
Y.iloc[val], ipls_obj.predict(X.iloc[val]))**0.5
k=k+1
j=j+1
i=i+1
cv_RMSE=np.mean(cv_RMSE_all,axis=0)
RMSE_best=np.amin(cv_RMSE)
rpd_best=np.std(self.df[self.y_name])/RMSE_best
# Best model
ncomp_best=ncomp_range[np.where(
cv_RMSE==RMSE_best)[0][0]]
nint_best=nint_list[np.where(
cv_RMSE==RMSE_best)[1][0]]
if verbose:
print("Best RMSE: ",RMSE_best)
print("Best RPD: ",rpd_best)
print("Number of components:",ncomp_best)
print("Number of intervals:",nint_best)
return (ncomp_best,nint_best,rpd_best)
def mcw_pls_cv(self,ncomp_range=range(1,21),sig_start=0.1,optimization="grid",
plot=False,verbose=True,
osc_params=(10,1)):
# Separating X from Y for PLS
# Needs to be converted to numpy array from pandas df
X=self.df[self.freqs].to_numpy()
# Y need to be converted to numpy array from pandas series and reshaped to (N,1) from (N,)
Y=self.df[self.y_name].to_numpy().reshape(-1, 1)
sample_std=np.std(self.df[self.y_name])
# CV based on measurement day
if self.cval=="MD":
cv = LeaveOneGroupOut()
folds=list(cv.split(X=X,y=Y,groups=self.df[self.date_name]))
# kfold CV
elif self.cval=="kfold":
cv = KFold(n_splits=self.cval_param)
folds=list(cv.split(X))
else:
raise InputError("Invalid CV type!")
if optimization=="grid":
# Create a search vector from starting values for gridsearch
sig_list=np.linspace(sig_start/10,sig_start*10,30)
rpd_best_all=0
non_improve=0
repeat=True
while repeat:
# Array for storing CV errors
cv_RMSE_all=np.zeros([len(folds),len(ncomp_range),len(sig_list)])
i=0
for train,val in folds:
# If OSC model specified
if len(osc_params)==2:
osc=OSC(nicomp=osc_params[0],ncomp=osc_params[1])
osc.fit(X[train], Y[train])
X_train_osc=osc.X_osc
X_val_osc=osc.transform(X[val])
j=0
for ncomp in ncomp_range:
k=0
for sig in sig_list:
if len(osc_params)==2:
pls = mcw_pls_sklearn(n_components=ncomp, max_iter=30, R_initial=None, scale_sigma2=sig)
pls.fit(X_train_osc, Y[train])
cv_RMSE_all[i,j,k]=metrics.mean_squared_error(
Y[val], pls.predict(X_val_osc))**0.5
else:
pls = mcw_pls_sklearn(n_components=ncomp, max_iter=30, R_initial=None, scale_sigma2=sig)
pls.fit(X[train], Y[train])
cv_RMSE_all[i,j,k]=metrics.mean_squared_error(
Y[val], pls.predict(X[val]))**0.5
k=k+1
j=j+1
i=i+1
cv_RMSE_ncomp_sigs=np.mean(cv_RMSE_all,axis=0)
# Best model
ncomp_best=ncomp_range[np.where(
cv_RMSE_ncomp_sigs==np.amin(cv_RMSE_ncomp_sigs))[0][0]]
sig_best=sig_list[np.where(
cv_RMSE_ncomp_sigs==np.amin(cv_RMSE_ncomp_sigs))[1][0]]
rpd_best=sample_std/np.amin(cv_RMSE_ncomp_sigs)
if verbose:
print("Best RMSE: ",np.amin(cv_RMSE_ncomp_sigs))
print("Best RPD: ",rpd_best)
print("Number of latent components: ",ncomp_best)
print("Best sigma: ",sig_best)
# Check against all time best
if rpd_best>rpd_best_all:
ncomp_best_all = ncomp_best
sig_best_all = sig_best
rpd_best_all= rpd_best
else:
# Increase counter if there is no improvement
non_improve=non_improve+1
repeat=False
# Check if best value is in IQ range
if sig_best<np.quantile(sig_list,0.2) or sig_best>np.quantile(sig_list,0.8):
# If not, move the search interval based on the magnitude of the best value
scale=math.floor(math.log10(sig_best))-1
lower=sig_best-(10**scale)*5
upper=sig_best+(10**scale)*5
# If best value is at the extreme of the interval expand it by a lot that way
if min(sig_list)==sig_best:
lower=sig_best/2
elif max(sig_list)==sig_best:
upper=sig_best*2
# Create new search vector
sig_list=np.linspace(lower,upper,10)
# Repeat evaluation
repeat=True
# Terminate early if no improvements in 10 iterations
if non_improve>10:
repeat=False
print("No improvement, terminate early.")
if repeat:
print("new iteration")
# Set final values to all time best
ncomp_best=ncomp_best_all
sig_best=sig_best_all
rpd_best=rpd_best_all
elif optimization=="simple":
# Array for storing CV errors
sig_list=sig_start
cv_RMSE_all=np.zeros([len(folds),len(ncomp_range),len(sig_list)])
i=0
for ncomp in ncomp_range:
j=0
for sig in sig_list:
pls = mcw_pls_sklearn(n_components=ncomp, max_iter=30, R_initial=None, scale_sigma2=sig)
k=0
for train,val in folds:
pls.fit(X[train], Y[train])
cv_RMSE_all[k,i,j]=metrics.mean_squared_error(
Y[val], pls.predict(X[val]))**0.5
k=k+1
j=j+1
i=i+1
# Printing and plotting CV results
cv_RMSE_ncomp_sigs=np.mean(cv_RMSE_all,axis=0)
if plot:
cv_RPD_ncomp_sigs=sample_std/cv_RMSE_ncomp_sigs
fig = plt.figure(figsize=(10,5))
ax = plt.axes(projection="3d")
# Cartesian indexing (x,y) transposes matrix indexing (i,j)
x, y = np.meshgrid(list(sig_list),list(ncomp_range))
z=cv_RPD_ncomp_sigs
ls = LightSource(270, 45)
rgb = ls.shade(z, cmap=cm.gist_earth, vert_exag=0.1, blend_mode='soft')
surf = ax.plot_surface(x, y, z, rstride=1, cstride=1, facecolors=rgb,
linewidth=0, antialiased=False, shade=False)
plt.show()
# Best model
ncomp_best=ncomp_range[np.where(
cv_RMSE_ncomp_sigs==np.amin(cv_RMSE_ncomp_sigs))[0][0]]
sig_best=sig_list[np.where(
cv_RMSE_ncomp_sigs==np.amin(cv_RMSE_ncomp_sigs))[1][0]]
rpd_best=sample_std/np.amin(cv_RMSE_ncomp_sigs)
print("Best RMSE: ",np.amin(cv_RMSE_ncomp_sigs))
print("Best RPD: ",rpd_best)
print("Number of latent components: ",ncomp_best)
print("Best sigma: ",sig_best)
return (ncomp_best,sig_best,rpd_best)
def svr_cv(self,gam_start=0.001,
c_start=100,
eps_start=0.1,
optimization="grid",gridscale=5,non_improve_lim=10,verbose=False,
osc_params=None):
# Separating X from Y for PLS
X=self.df[self.freqs].to_numpy()
Y=self.df[self.y_name].to_numpy().reshape(-1, 1)
sample_std=np.std(self.df[self.y_name])
# CV based on measurement day
if self.cval=="MD":
cv = LeaveOneGroupOut()
folds=list(cv.split(X=X,y=Y,groups=self.df[self.date_name]))
# kfold CV
elif self.cval=="kfold":
cv = KFold(n_splits=self.cval_param)
folds=list(cv.split(X))
else:
raise InputError("Invalid CV type!")
if optimization=="none":
cv_RMSE=np.zeros(len(folds))
# Only use RBF kernels, also standardize data
pipe = Pipeline([('scaler', StandardScaler()),
('support vector regression',
SVR(kernel="rbf",gamma=gam_start,C=c_start,epsilon=eps_start))])
l=0
for train, val in folds:
pipe.fit(X[train], Y[train])
cv_RMSE[l]=metrics.mean_squared_error(
Y[val], pipe.predict(X[val]))**0.5
l=l+1
gam_best=gam_start
c_best=c_start
eps_best=eps_start
rpd_best=sample_std/np.mean(cv_RMSE)
elif optimization=="grid":
# Create a search vector from starting values for gridsearch
gam_list=np.linspace(gam_start/gridscale,gam_start*gridscale,10)
c_list=np.linspace(c_start/gridscale,c_start*gridscale,10)
eps_list=np.linspace(eps_start/gridscale,eps_start*gridscale,10)
# Create list of ndarrays from parameter search vectors,
# it will help with making the cood more tidy
param_lists=[gam_list,c_list,eps_list]
param_best=np.zeros(3)
rpd_best_all=0
non_improve=0
repeat=True
while repeat:
# Array for storing CV errors
cv_RMSE_all=np.zeros([len(folds),len(gam_list),len(c_list),len(eps_list)])
# Put the CV iteration outside to save time when using OSC
i=0
for train, val in folds:
# If OSC model specified
if len(osc_params)==2:
osc=OSC(nicomp=osc_params[0],ncomp=osc_params[1])
osc.fit(X[train], Y[train])
X_train_osc=osc.X_osc
X_val_osc=osc.transform(X[val])
j=0
for gam in param_lists[0]:
k=0
for c in param_lists[1]:
l=0
for eps in param_lists[2]:
pipe = Pipeline([('scaler', StandardScaler()),
('support vector regression', SVR(kernel="rbf",gamma=gam,C=c,epsilon=eps))])
if len(osc_params)==2:
pipe.fit(X_train_osc, Y[train])
cv_RMSE_all[i,j,k,l]=metrics.mean_squared_error(
Y[val], pipe.predict(X_val_osc))**0.5
else:
pipe.fit(X[train], Y[train])
cv_RMSE_all[i,j,k,l]=metrics.mean_squared_error(
Y[val], pipe.predict(X[val]))**0.5
l=l+1
k=k+1
j=j+1
i=i+1
cv_RMSE=np.mean(cv_RMSE_all,axis=0)
# Best model
param_best[0]=param_lists[0][np.where(
cv_RMSE==np.amin(cv_RMSE))[0][0]]
param_best[1]=param_lists[1][np.where(
cv_RMSE==np.amin(cv_RMSE))[1][0]]
param_best[2]=param_lists[2][np.where(
cv_RMSE==np.amin(cv_RMSE))[2][0]]
rpd_best=sample_std/np.amin(cv_RMSE)
# Check against all time best
if rpd_best>rpd_best_all:
param_best_all = param_best.copy()
rpd_best_all=rpd_best
else:
# Increase counter if there is no improvement
non_improve=non_improve+1
if verbose==True:
print("Best RMSE: ",np.amin(cv_RMSE))
print("Best RPD: ",rpd_best)
print("Gamma: ",param_best[0])
print("C: ",param_best[1])
print("Epsilon: ",param_best[2])
repeat=False
for index,p in enumerate(param_best):
# Check if best value is in IQ range
if p<np.quantile(param_lists[index],0.2) or p>np.quantile(param_lists[index],0.8):
# If not, move the search interval based on the magnitude of the best value
scale=math.floor(math.log10(p))-1
lower=p-(10**scale)*5
upper=p+(10**scale)*5
# If best value is at the extreme of the interval expand it by a lot that way
if min(param_lists[index])==p:
lower=min(param_lists[index])/2
elif max(param_lists[index])==p:
upper=max(param_lists[index])*2
# Create new search vector
param_lists[index]=np.linspace(lower,upper,10)
# Repeat evaluation
repeat=True
# Terminate early if no improvements in 10 iterations
if non_improve>non_improve_lim:
repeat=False
print("No improvement, terminate early.")
if repeat:
print("new iteration")
# Set final values to all time best
gam_best=param_best_all[0]
c_best=param_best_all[1]
eps_best=param_best_all[2]
rpd_best=rpd_best_all
# Simulated annealing
elif optimization=="sa":
# Number of cycles
cycles = 100
# Trials per cycle
trials = 100
# Number of accepted solutions
n_accepted = 0.0
# Probability of accepting worse solution at the start
p_start = 0.3
# Probability of accepting worse solution at the end
p_end = 0.001
# Initial temperature
t_start = -1.0/math.log(p_start)
# Final temperature
t_end = -1.0/math.log(p_end)
# Use geometric temp reduction
frac = (t_end/t_start)**(1.0/(cycles-1.0))
# Starting values
t=t_start
dE_mean = 0.0
gam=gam_start
c=c_start
eps=eps_start
# Calculate starting cost
cv_RMSE=np.zeros(len(folds))
pipe = Pipeline([('scaler', StandardScaler()),
('support vector regression',
SVR(kernel="rbf",gamma=gam,C=c,epsilon=eps))])
L=0
for train, val in folds:
pipe.fit(X[train], Y[train])
cv_RMSE[L]=metrics.mean_squared_error(
Y[val], pipe.predict(X[val]))**0.5
L=L+1
cost=np.mean(cv_RMSE)
rpd=sample_std/cost
print("starting RPD:",rpd)
# Best results
gam_old = gam
c_old = c
eps_old = eps
cost_old=cost
rpd_old=rpd
# All time best result
gam_best = gam
c_best = c
eps_best = eps
cost_best=cost
rpd_best = rpd
for i in range(cycles):
if verbose and i%10==0 and i>0:
print('Cycle: ', i ,' with Temperature: ', t)
print('RPD=',rpd_old,'Gamma=' ,gam_old,', C=' ,c_old,', epsilon=',eps_old)
for j in range(trials):
# Generate new trial points
gam = gam_old + (random.random()-0.5)*2/1000
c = c_old + (random.random()-0.5)*2*10
eps = eps_old + (random.random()-0.5)*2/100
# Enforce lower bounds
gam = max(gam,0.0000001)
c = max(c,0.0000001)
eps = max(eps,0)
# Calculate cost
cv_RMSE=np.zeros(len(folds))
pipe = Pipeline([('scaler', StandardScaler()),
('support vector regression',
SVR(kernel="rbf",gamma=gam,C=c,epsilon=eps))])
L=0
for train, val in folds:
pipe.fit(X[train], Y[train])
cv_RMSE[L]=metrics.mean_squared_error(
Y[val], pipe.predict(X[val]))**0.5
L=L+1
cost=np.mean(cv_RMSE)
rpd=sample_std/cost
dE = cost-cost_old
# If new cost is higher
if dE > 0:
if (i==0 and j==0): dE_mean = dE
# Generate probability of acceptance
p = math.exp(-dE/(dE_mean * t))
# Determine whether to accept worse point
if (random.random()<p):
accept = True
else:
accept = False
else:
# New cost is lower, automatically accept
accept = True
# Check if cost is lower than all time best
if cost<cost_best:
# If new best, store the parameters, cost and RPD
gam_best=gam
c_best=c
eps_best=eps
cost_best=cost
rpd_best=rpd
if accept==True:
# Update parameters, cost and RPD
gam_old = gam
c_old = c
eps_old = eps
cost_old=cost
rpd_old=rpd
# Increment number of accepted solutions
n_accepted = n_accepted + 1
# Update energy change
dE_mean = (dE_mean * (n_accepted-1) + abs(dE)) / n_accepted
# Lower the temperature for next cycle
t = frac * t
# Return the best setting found
else:
raise InputError("Invalid optimization strategy!")
return (gam_best,c_best,eps_best,rpd_best)
def osc_cv(self,nicomp_range=range(10,130,10),ncomp_range=range(1,5),epsilon = 10e-6,
max_iters = 20,model="pls",model_parameter_range=range(1,11)):
# Separating X from Y for PLS
# Needs to be converted to numpy array from pandas df
X=self.df[self.freqs].to_numpy()
# Y need to be converted to numpy array from pandas series and reshaped to (N,1) from (N,)
Y=self.df[self.y_name].to_numpy().reshape(-1, 1)
# CV based on measurement day
if self.cval=="MD":
cv = LeaveOneGroupOut()
folds=list(cv.split(X=X,y=Y,groups=self.df[self.date_name]))
# kfold CV
elif self.cval=="kfold":
cv = KFold(n_splits=self.cval_param)
folds=list(cv.split(X))
else:
raise InputError("Invalid CV type!")
#Matrix for cv values for all the possible parameter combinations
cv_RMSE_all=np.zeros([len(folds),len(model_parameter_range),len(nicomp_range),len(ncomp_range)])
i=0
#possible internal component values for osc
for nicomp in nicomp_range:
j=0
#possible removed component values for osc
for ncomp in ncomp_range:
k=0
for train, val in folds:
# train osc
osc_obj=OSC("SWosc",nicomp,ncomp,epsilon, max_iters)
X_osc_train, W,P,mu_x=osc_obj.fit(X[train],Y[train])
# apply osc on validation set
# mean center data, alternatively the training set's mean can be used
# if you think it is a better estimate by mean="training"
X_osc_val=osc_obj.transform(X[val],mean="estimate")
l=0
#possible model patrameter values for pls
for param in model_parameter_range:
#setup pls model
pls = PLSRegression(param,scale=False)
#train pls
pls.fit(X_osc_train, Y[train])
#predict with pls and calculate error
cv_RMSE_all[k,l,i,j]=metrics.mean_squared_error(
Y[val], pls.predict(X_osc_val))**0.5
l=l+1
k=k+1
j=j+1
i=i+1
# Calculate mean performance across the folds
cv_RMSE_mean=np.mean(cv_RMSE_all,axis=0)
# Find maximum for every osc paremeter combination
cv_RMSE=np.amax(cv_RMSE_mean, axis=0)
cv_RPD=np.std(self.df[self.y_name])/cv_RMSE
fig = plt.figure(figsize=(10,5))
ax = plt.axes(projection="3d")
# Cartesian indexing (x,y) transposes matrix indexing (i,j)
x, y = np.meshgrid(list(ncomp_range),list(nicomp_range))
z=cv_RPD
ls = LightSource(200, 45)
rgb = ls.shade(z, cmap=cm.gist_earth, vert_exag=0.1, blend_mode='soft')
surf = ax.plot_surface(x, y, z, rstride=1, cstride=1, facecolors=rgb,
linewidth=0, antialiased=False, shade=False)
plt.show()
# Best model
print("Best RMSE: ",np.amin(cv_RMSE))
print("Best RPD: ",np.std(self.df[self.y_name])/np.amin(cv_RMSE))
print("Number of internal components: ",nicomp_range[np.where(
cv_RMSE==np.amin(cv_RMSE))[0][0]])
print("Number of removed components: ",ncomp_range[np.where(
cv_RMSE==np.amin(cv_RMSE))[1][0]])
return cv_RMSE