def build (self, X, Y, quantitative=False, autoscale=False, nestimators=0, features='', random=False, tune=False):
"""Build a new RF model with the X and Y numpy matrices
"""
nobj, nvarx= np.shape(X)
self.nobj = nobj
self.nvarx = nvarx
self.quantitative = quantitative
self.autoscale = autoscale
self.estimators = nestimators
self.features = features
self.random = random
self.X = X.copy()
self.Y = Y.copy()
if autoscale:
self.X, self.mux = center(self.X)
self.X, self.wgx = scale(self.X, autoscale)
if random :
RANDOM_STATE = None
else:
RANDOM_STATE = 1226 # no reason to pick this number
if tune :
self.estimators, self.features = self.optimize (X,Y)
if self.features=='none':
self.features = None
#print self.estimators
if self.quantitative:
print "Building Quantitative RF model"
self.clf = RandomForestRegressor(n_estimators = int(self.estimators),
warm_start=False,
max_features=self.features,
oob_score=True,
random_state=RANDOM_STATE)
else:
print "Building Qualitative RF_model"
self.clf = RandomForestClassifier(n_estimators = int(self.estimators),
warm_start=False,
max_features=self.features,
oob_score=True,
random_state=RANDOM_STATE)
self.clf.fit(self.X, self.Y)
# Regenerate the X and Y, since they might have been centered/scaled
self.X = X.copy()
self.Y = Y.copy()
def build (self, X, targetA, autoscale=False):
nobj, nvar= np.shape(X)
self.nobj = nobj
self.nvar = nvar
self.X = X
X, mu = center(X)
X, wg = scale (X, autoscale)
self.mu = mu
self.wg = wg
self.autoscale = autoscale
SSXac=0.0
for a in range(targetA):
# extracts LV
t, p = self.extractPC(X)
self.t.append(t)
self.p.append(p)
# deflates X
X, SSX, SSXex = self.deflatePC(X,t,p)
SSXac += SSXex
self.SSXex.append(SSXex)
self.SSXac.append(SSXac)
if a==0:
self.SSX = SSX
self.A = targetA
def getLOO (self, X, Y, Xout):
clf = None
if self.autoscale:
X, mux = center(X)
X, wgx = scale(X, self.autoscale)
RANDOM_STATE = 1226 # no reason to pick this number
if self.quantitative:
clf = RandomForestRegressor(n_estimators = int(self.estimators),
warm_start=False,
max_features=self.features,
oob_score=True,
random_state=RANDOM_STATE)
else:
clf = RandomForestClassifier(n_estimators = int(self.estimators),
warm_start=False,
max_features=self.features,
oob_score=True,
random_state=RANDOM_STATE)
clf.fit(X, Y)
return ( clf.predict(Xout), clf.oob_score_ )
def validateLOO(self, A, gui=False):
""" Validates A dimensions of an already built PLS model, using Leave-One-Out cross-validation
Returns nothing. The results of the cv (SSY, SDEP and Q2) are stored internally
"""
if self.X == None or self.Y == None:
return
X = self.X
Y = self.Y
nobj, nvarx = np.shape(X)
SSY0 = 0.0
for i in range(nobj):
SSY0 += np.square(Y[i] - np.mean(Y))
SSY = np.zeros(A, dtype=np.float64)
YP = np.zeros((nobj, A + 1), dtype=np.float64)
if gui: updateProgress(0.0)
for i in range(nobj):
# build reduced X and Y matrices removing i object
Xr = np.delete(X, i, axis=0)
Yr = np.delete(Y, i)
Xr, muxr = center(Xr)
Xr, wgxr = scale(Xr, self.autoscale)
Yr, muyr = center(Yr)
xp = np.copy(X[i, :])
xp -= muxr
xp *= wgxr
# predicts y for the i object, using A LV
yp = self.getLOO(Xr, Yr, xp, A)
yp += muyr
# updates SSY with the object i errors
YP[i, 0] = Y[i]
for a in range(A):
SSY[a] += np.square(yp[a] - Y[i])
YP[i, a + 1] = yp[a]
if gui: updateProgress(float(i) / float(nobj))
if gui: print
self.SSY = SSY
self.SDEP = [np.sqrt(i / nobj) for i in SSY]
self.Q2 = [1.00 - (i / SSY0) for i in SSY]
self.Av = A
return (YP)
def build(self, X, Y, targetA=0, targetSSX=0.0, autoscale=False):
"""Build a new PLS model with the X and Y numpy matrice provided using NIPALS algorithm
The dimensionality of the model can be defined either providing
1. directly the number of LV to extract (targetA)
2. the fraction of SSX that the model will explain (targetSSX)
The X and Y matrices are centered but no other scaling transform is applied
Does not return anything, but updates internals vectors and variables
"""
nobj, nvarx = np.shape(X)
## for i in range (nobj):
## for j in range (nvarx):
## print X[i,j],
## print
self.nobj = nobj
self.nvarx = nvarx
self.X = X.copy()
self.Y = Y.copy()
self.X, self.mux = center(self.X)
self.Y, self.muy = center(self.Y)
self.X, self.wgx = scale(self.X, autoscale)
## self.mux = mux
## self.muy = muy
## self.wgx = wgx
self.autoscale = autoscale
SSXac = 0.0
SSYac = 0.0
SSX0, SSY0, null = self.computeSS(self.X, self.Y)
SSXold = SSX0
SSYold = SSY0
a = 0
while True:
t, p, w, c = self.extractLV(self.X, self.Y)
self.t.append(t)
self.p.append(p)
self.w.append(w)
self.c.append(c)
self.X, self.Y = self.deflateLV(self.X, self.Y, t, p, c)
SSXnew, SSYnew, dmodx = self.computeSS(self.X, self.Y)
SSXex = (SSXold - SSXnew) / SSX0
SSXac += SSXex
SSYex = (SSYold - SSYnew) / SSY0
SSYac += SSYex
SDEC = np.sqrt(SSYnew / nobj)
dof = nvarx - a
if dof <= 0: dof = 1
dmodx = [np.sqrt(d / dof) for d in dmodx]
SSXold = SSXnew
SSYold = SSYnew
self.SSXex.append(SSXex)
self.SSXac.append(SSXac)
self.SSYex.append(SSYex)
self.SSYac.append(SSYac)
self.SDEC.append(SDEC)
self.dmodx.append(dmodx)
a += 1
if targetA > 0:
if a == targetA: break
if targetSSX > 0.0:
if SSXac > targetSSX: break
# prevents to extract a meaningless number of LV
if a > min(20, nobj / 5): break
self.Am = a
# NIPALS is destructive, so we must retrieve X and Y from original data for validation
self.X = X.copy()
self.Y = Y.copy()
self.cutoff = np.zeros(self.Am, dtype=np.float64)
self.TP = np.zeros(self.Am)
self.TN = np.zeros(self.Am)
self.FP = np.zeros(self.Am)
self.FN = np.zeros(self.Am)
self.TPpred = np.zeros(self.Am)
self.TNpred = np.zeros(self.Am)
self.FPpred = np.zeros(self.Am)
self.FNpred = np.zeros(self.Am)
def build(self,
X,
Y,
quantitative=False,
autoscale=False,
nestimators=0,
features='',
random=False,
tune=False,
class_weight="balanced",
cv='loo',
n=2,
p=1,
lc=True,
vpath=''):
"""Build a new RF model with the X and Y numpy matrices
"""
nobj, nvarx = np.shape(X)
self.nobj = nobj
self.nvarx = nvarx
self.quantitative = quantitative
self.autoscale = autoscale
self.estimators = nestimators
self.features = features
self.random = random
self.class_weight = class_weight
self.learning_curve = lc
self.n = n
self.p = p
self.cv = cv
self.X = X.copy()
self.Y = Y.copy()
self.vpath = vpath
#print self.vpath
if autoscale:
self.X, self.mux = center(self.X)
self.X, self.wgx = scale(self.X, autoscale)
if random:
RANDOM_STATE = None
else:
RANDOM_STATE = 1226 # no reason to pick this number
if self.cv:
self.cv = getCrossVal(self.cv, RANDOM_STATE, self.n, self.p)
if tune:
self.estimators, self.features = self.optimize(self.X, self.Y)
if self.features == 'none':
self.features = None
#print self.estimators
if self.quantitative:
print "Building Quantitative RF model"
self.clf = RandomForestRegressor(n_estimators=int(self.estimators),
warm_start=False,
max_features=self.features,
oob_score=True,
random_state=RANDOM_STATE)
else:
print "Building Qualitative RF_model"
self.clf = RandomForestClassifier(n_estimators=int(
self.estimators),
warm_start=False,
max_features=self.features,
oob_score=True,
random_state=RANDOM_STATE,
class_weight=self.class_weight)
self.clf.fit(self.X, self.Y)
print 'Building Learning Curves'
if self.learning_curve:
title = "Learning Curves (RF)"
# SVC is more expensive so we do a lower number of CV iterations:
cv = ShuffleSplit(n_splits=10, test_size=0.2, random_state=0)
estimator = self.clf
plot = plot_learning_curve(estimator,
title,
self.X,
self.Y, (0.0, 1.01),
cv=cv)
plot.savefig(self.vpath + "/RF-learning_curves.png", format='png')
plot.savefig("./RF-learning_curves.png", format='png')
# Regenerate the X and Y, since they might have been centered/scaled
self.X = X.copy()
self.Y = Y.copy()
def build (self, X, Y, quantitative=False, autoscale=False,
nestimators=0, features='', random=False, tune=False, class_weight="balanced",
cv='loo', n=2, p=1, lc=True, vpath = ''):
"""Build a new RF model with the X and Y numpy matrices
"""
nobj, nvarx= np.shape(X)
self.nobj = nobj
self.nvarx = nvarx
self.quantitative = quantitative
self.autoscale = autoscale
self.estimators = nestimators
self.features = features
self.random = random
self.class_weight = class_weight
self.learning_curve = lc
self.n = n
self.p = p
self.cv = cv
self.X = X.copy()
self.Y = Y.copy()
self.vpath = vpath
#print self.vpath
if autoscale:
self.X, self.mux = center(self.X)
self.X, self.wgx = scale(self.X, autoscale)
if random :
RANDOM_STATE = None
else:
RANDOM_STATE = 1226 # no reason to pick this number
if self.cv:
self.cv = getCrossVal(self.cv, RANDOM_STATE, self.n, self.p)
if tune :
self.estimators, self.features = self.optimize (self.X, self.Y)
if self.features=='none':
self.features = None
#print self.estimators
if self.quantitative:
print "Building Quantitative RF model"
self.clf = RandomForestRegressor(n_estimators = int(self.estimators),
warm_start=False,
max_features=self.features,
oob_score=True,
random_state=RANDOM_STATE)
else:
print "Building Qualitative RF_model"
self.clf = RandomForestClassifier(n_estimators = int(self.estimators),
warm_start=False,
max_features=self.features,
oob_score=True,
random_state=RANDOM_STATE,
class_weight=self.class_weight)
self.clf.fit(self.X, self.Y)
print 'Building Learning Curves'
if self.learning_curve:
title = "Learning Curves (RF)"
# SVC is more expensive so we do a lower number of CV iterations:
cv = ShuffleSplit(n_splits=10, test_size=0.2, random_state=0)
estimator = self.clf
plot = plot_learning_curve(estimator, title, self.X, self.Y, (0.0, 1.01), cv=cv)
plot.savefig(self.vpath+"/RF-learning_curves.png", format='png')
plot.savefig("./RF-learning_curves.png", format='png')
# Regenerate the X and Y, since they might have been centered/scaled
self.X = X.copy()
self.Y = Y.copy()
def validateLOO (self, A, gui=False):
""" Validates A dimensions of an already built PLS model, using Leave-One-Out cross-validation
Returns nothing. The results of the cv (SSY, SDEP and Q2) are stored internally
"""
if self.X == None or self.Y == None:
return
X = self.X
Y = self.Y
nobj,nvarx = np.shape (X)
SSY0 = 0.0
for i in range (nobj):
SSY0+=np.square(Y[i]-np.mean(Y))
SSY = np.zeros(A,dtype=np.float64)
YP = np.zeros ((nobj,A+1),dtype=np.float64)
if gui: updateProgress (0.0)
for i in range (nobj):
# build reduced X and Y matrices removing i object
Xr = np.delete(X,i,axis=0)
Yr = np.delete(Y,i)
Xr,muxr = center(Xr)
Xr,wgxr = scale (Xr, self.autoscale)
Yr,muyr = center(Yr)
xp = np.copy(X[i,:])
xp -= muxr
xp *= wgxr
# predicts y for the i object, using A LV
yp = self.getLOO(Xr,Yr,xp,A)
yp += muyr
# updates SSY with the object i errors
YP[i,0]=Y[i]
for a in range(A):
SSY[a]+= np.square(yp[a]-Y[i])
YP[i,a+1]=yp[a]
if gui : updateProgress (float(i)/float(nobj))
if gui : print
self.SSY = SSY
self.SDEP = [np.sqrt(i/nobj) for i in SSY]
self.Q2 = [1.00-(i/SSY0) for i in SSY]
self.Av = A
return (YP)
def build (self, X, Y, targetA=0, targetSSX=0.0, autoscale=False):
"""Build a new PLS model with the X and Y numpy matrice provided using NIPALS algorithm
The dimensionality of the model can be defined either providing
1. directly the number of LV to extract (targetA)
2. the fraction of SSX that the model will explain (targetSSX)
The X and Y matrices are centered but no other scaling transform is applied
Does not return anything, but updates internals vectors and variables
"""
nobj, nvarx= np.shape(X)
## for i in range (nobj):
## for j in range (nvarx):
## print X[i,j],
## print
self.nobj = nobj
self.nvarx = nvarx
self.X = X.copy()
self.Y = Y.copy()
self.X, self.mux = center(self.X)
self.Y, self.muy = center(self.Y)
self.X, self.wgx = scale(self.X, autoscale)
## self.mux = mux
## self.muy = muy
## self.wgx = wgx
self.autoscale = autoscale
SSXac=0.0
SSYac=0.0
SSX0,SSY0, null = self.computeSS(self.X,self.Y)
SSXold=SSX0
SSYold=SSY0
a=0
while True:
t, p, w, c = self.extractLV(self.X, self.Y)
self.t.append(t)
self.p.append(p)
self.w.append(w)
self.c.append(c)
self.X, self.Y = self.deflateLV(self.X, self.Y, t, p, c)
SSXnew, SSYnew, dmodx = self.computeSS(self.X, self.Y)
SSXex = (SSXold-SSXnew)/SSX0
SSXac+=SSXex
SSYex = (SSYold-SSYnew)/SSY0
SSYac+=SSYex
SDEC = np.sqrt(SSYnew/nobj)
dof = nvarx-a
if dof <= 0 : dof = 1
dmodx = [np.sqrt(d/dof) for d in dmodx]
SSXold=SSXnew
SSYold=SSYnew
self.SSXex.append(SSXex)
self.SSXac.append(SSXac)
self.SSYex.append(SSYex)
self.SSYac.append(SSYac)
self.SDEC.append(SDEC)
self.dmodx.append(dmodx)
a+=1
if targetA>0:
if a==targetA : break
if targetSSX>0.0:
if SSXac>targetSSX: break
# prevents to extract a meaningless number of LV
if a > min (20,nobj/5) : break
self.Am=a
# NIPALS is destructive, so we must retrieve X and Y from original data for validation
self.X = X.copy()
self.Y = Y.copy()
self.cutoff = np.zeros(self.Am, dtype=np.float64)
self.TP = np.zeros(self.Am)
self.TN = np.zeros(self.Am)
self.FP = np.zeros(self.Am)
self.FN = np.zeros(self.Am)
self.TPpred = np.zeros(self.Am)
self.TNpred = np.zeros(self.Am)
self.FPpred = np.zeros(self.Am)
self.FNpred = np.zeros(self.Am)