def testModel(self):
Y = np.random.random((10))
X = np.random.random((10, 100))
kwargs = {}
kwargs["Y"] = Y
kwargs["X"] = X
kwargs["regparam"] = 1
learner = RLS(**kwargs)
model = learner.predictor
print
#print 'Ten data points, single label '
model = mod.LinearPredictor(np.random.random((100)))
self.all_pred_cases(model)
model = mod.LinearPredictor(np.random.random((100, 2)))
self.all_pred_cases(model)
#model = mod.LinearPredictor(np.random.random((1, 2)))
#self.all_pred_cases(model)
kwargs["kernel"] = "GaussianKernel"
Y = np.random.random((10))
kwargs["Y"] = Y
learner = RLS(**kwargs)
model = learner.predictor
self.all_pred_cases(model)
kwargs["kernel"] = "GaussianKernel"
Y = np.random.random((10, 2))
kwargs["Y"] = Y
learner = RLS(**kwargs)
model = learner.predictor
self.all_pred_cases(model)
def callback(self, learner):
m = predictor.LinearPredictor(learner.A, learner.b)
P = m.predict(self.X_valid)
if self.qids_valid:
perfs = []
for query in self.qids_valid:
try:
perf = self.measure(self.Y_valid[query], P[query])
perfs.append(perf)
except UndefinedPerformance:
pass
perf = np.mean(perfs)
else:
perf = self.measure(self.Y_valid, P)
if self.bestperf is None or (self.measure.iserror
== (perf < self.bestperf)):
self.bestperf = perf
self.bestA = learner.A
self.last_update = 0
else:
self.iter += 1
self.last_update += 1
if self.last_update == self.maxiter:
learner.A = np.mat(self.bestA)
raise Finished("Done")
def createModel(self, svdlearner):
A = svdlearner.A
A = self.reducedSetTransformation(A)
fs = self.X
if self.basis_vectors is not None:
fs = self.basis_vectors
bias = self.bias
X = getPrimalDataMatrix(fs, bias)
#The hyperplane is a linear combination of the feature vectors of the basis examples
W = np.dot(X.T, A)
if bias != 0:
W_biaz = W[W.shape[0] - 1] * math.sqrt(bias)
W_features = W[range(W.shape[0] - 1)]
mod = predictor.LinearPredictor(W_features, W_biaz)
else:
mod = predictor.LinearPredictor(W, 0.)
return mod
def __init__(self,
X,
Y,
regparam=1.0,
bias=1.0,
callbackfun=None,
**kwargs):
self.Y = array_tools.as_2d_array(Y)
self.X = csc_matrix(X.T)
self.bias = bias
self.regparam = regparam
if self.bias != 0.:
bias_slice = sqrt(self.bias) * np.mat(
ones((1, self.X.shape[1]), dtype=np.float64))
self.X = sparse.vstack([self.X, bias_slice]).tocsc()
self.X_csr = self.X.tocsr()
self.callbackfun = callbackfun
self.results = {}
regparam = self.regparam
Y = self.Y
X = self.X
X_csr = self.X_csr
def mv(v):
return X.T * (X_csr * v) + regparam * v
G = LinearOperator((X.shape[1], X.shape[1]),
matvec=mv,
dtype=np.float64)
self.AA = []
if not self.callbackfun is None:
def cb(v):
self.A = np.mat(v).T
self.callbackfun.callback(self)
else:
cb = None
try:
self.A = np.mat(cg(G, Y, callback=cb)[0]).T
except Finished:
pass
if self.callbackfun is not None:
self.callbackfun.finished(self)
self.A = X_csr * self.A
if self.bias == 0.:
self.b = np.mat(np.zeros((1, 1)))
else:
self.b = sqrt(self.bias) * self.A[-1]
self.A = self.A[:-1]
#self.results['predictor'] = self.getModel()
self.predictor = predictor.LinearPredictor(self.A, self.b)
def __init__(self, X, train_preferences, regparam=1., **kwargs):
self.regparam = regparam
self.callbackfun = None
self.pairs = train_preferences
self.X = csc_matrix(X.T)
regparam = self.regparam
X = self.X.tocsc()
X_csr = X.tocsr()
vals = np.concatenate([
np.ones((self.pairs.shape[0]), dtype=np.float64), -np.ones(
(self.pairs.shape[0]), dtype=np.float64)
])
row = np.concatenate(
[np.arange(self.pairs.shape[0]),
np.arange(self.pairs.shape[0])])
col = np.concatenate([self.pairs[:, 0], self.pairs[:, 1]])
coo = coo_matrix((vals, (row, col)),
shape=(self.pairs.shape[0], X.shape[1]))
pairs_csr = coo.tocsr()
pairs_csc = coo.tocsc()
def mv(v):
vmat = np.mat(v).T
ret = np.array(X_csr * (pairs_csc.T *
(pairs_csr *
(X.T * vmat)))) + regparam * vmat
return ret
G = LinearOperator((X.shape[0], X.shape[0]),
matvec=mv,
dtype=np.float64)
self.As = []
M = np.mat(np.ones((self.pairs.shape[0], 1)))
if not self.callbackfun is None:
def cb(v):
self.A = np.mat(v).T
self.b = np.mat(np.zeros((1, 1)))
self.callbackfun.callback()
else:
cb = None
XLY = X_csr * (pairs_csc.T * M)
self.A = np.mat(cg(G, XLY, callback=cb)[0]).T
self.b = np.mat(np.zeros((1, self.A.shape[1])))
self.predictor = predictor.LinearPredictor(self.A, self.b)
def callback(self, learner):
A = learner.A
b = learner.bias
A = learner.X_csr * A
if b == 0:
b = np.mat(np.zeros((1, 1)))
else:
b = sqrt(b) * A[-1]
A = A[:-1]
m = predictor.LinearPredictor(A, b)
P = m.predict(self.X_valid)
perf = self.measure(self.Y_valid, P)
if self.bestperf is None or (self.measure.iserror
== (perf < self.bestperf)):
self.bestperf = perf
self.bestA = learner.A
self.last_update = 0
else:
self.iter += 1
self.last_update += 1
if self.last_update == self.maxiter:
learner.A = np.mat(self.bestA)
raise Finished("Done")
def _solve_bu(self, regparam):
self.regparam = regparam
X = self.X
Y = self.Y
bias_slice = np.sqrt(self.bias) * np.mat(
np.ones((1, X.shape[1]), dtype=np.float64))
su = np.sum(X, axis=1)
cc = 0
tsize = self.size
fsize = X.shape[0]
assert X.shape[1] == tsize
self.A = np.mat(np.zeros((fsize, Y.shape[1])))
rp = regparam
rpinv = 1. / rp
#Biaz
cv = np.sqrt(self.bias) * np.mat(np.ones((1, tsize)))
ca = rpinv * (1. / (1. + cv * rpinv * cv.T)) * (cv * rpinv)
self.dualvec = rpinv * Y - cv.T * rpinv * (
1. / (1. + cv * rpinv * cv.T)) * (cv * rpinv * Y)
XT = X.T
GXT = rpinv * XT - cv.T * rpinv * (1. / (1. + cv * rpinv * cv.T)) * (
(cv * rpinv) * XT)
diagG = []
for i in range(tsize):
diagGi = rpinv - cv.T[i, 0] * ca[0, i]
diagG.append(diagGi)
diagG = np.mat(diagG).T
listX = []
for ci in range(fsize):
listX.append(X[ci])
self.selected = []
currentfcount = 0
self.performances = []
while currentfcount < self.desiredfcount:
if not self.measure is None:
bestlooperf = None
else:
bestlooperf = 9999999999.
self.looperf = []
for ci in range(fsize):
ci_mapped = ci
if ci in self.selected: continue
cv = listX[ci_mapped]
GXT_ci = GXT[:, ci_mapped]
ca = GXT_ci * (1. / (1. + cv * GXT_ci))
updA = self.dualvec - ca * (cv * self.dualvec)
invupddiagG = 1. / (diagG - np.multiply(ca, GXT_ci))
if not self.measure is None:
loopred = Y - np.multiply(invupddiagG, updA)
looperf_i = self.measure(Y, loopred)
if bestlooperf is None:
bestlooperf = looperf_i
bestcind = ci
if looperf_i < bestlooperf:
bestcind = ci
bestlooperf = looperf_i
else:
#This default squared performance is a bit faster to compute than the one loaded separately.
loodiff = np.multiply(invupddiagG, updA)
looperf_i = np.mean(np.multiply(loodiff, loodiff))
if looperf_i < bestlooperf:
bestcind = ci
bestlooperf = looperf_i
self.looperf.append(looperf_i)
self.looperf = np.mat(self.looperf)
self.bestlooperf = bestlooperf
self.performances.append(bestlooperf)
ci_mapped = bestcind
cv = listX[ci_mapped]
GXT_bci = GXT[:, ci_mapped]
ca = GXT_bci * (1. / (1. + cv * GXT_bci))
self.dualvec = self.dualvec - ca * (cv * self.dualvec)
diagG = diagG - np.multiply(ca, GXT_bci)
GXT = GXT - ca * (cv * GXT)
self.selected.append(bestcind)
#print self.selected
#print bestlooperf
currentfcount += 1
#Linear predictor with bias
self.A[self.selected] = X[self.selected] * self.dualvec
self.b = bias_slice * self.dualvec * np.sqrt(self.bias)
self.predictor = predictor.LinearPredictor(self.A, self.b)
if not self.callbackfun is None:
self.callbackfun.callback(self)
if not self.callbackfun is None:
self.callbackfun.finished(self)
self.A[self.selected] = X[self.selected] * self.dualvec
self.b = bias_slice * self.dualvec * np.sqrt(self.bias)
self.results[SELECTED_FEATURES] = self.selected
self.results[GREEDYRLS_LOO_PERFORMANCES] = self.performances
self.predictor = predictor.LinearPredictor(self.A, self.b)
def _solve_new(self, regparam, floattype=np.float64):
#Legacy code. Works only with a single output but can work with given performance measures and is faster than _solve_bu
if not self.Y.shape[1] == 1:
raise Exception(
'This variation of GreedyRLS supports only one output at a time. The output matrix is now of shape '
+ str(self.Y.shape) + '.')
self.regparam = regparam
X = self.X
Y = np.mat(self.Y, dtype=floattype)
bias_slice = np.sqrt(self.bias) * np.mat(
np.ones((1, X.shape[1]), dtype=floattype))
tsize = self.size
fsize = X.shape[0]
assert X.shape[1] == tsize
self.A = np.mat(np.zeros((fsize, 1), dtype=floattype))
rp = regparam
rpinv = 1. / rp
#Biaz
cv = np.sqrt(self.bias) * np.mat(np.ones((1, tsize), dtype=floattype))
ca = np.mat(rpinv * (1. / (1. + cv * rpinv * cv.T)) * (cv * rpinv),
dtype=floattype)
self.dualvec = rpinv * Y - cv.T * rpinv * (
1. / (1. + cv * rpinv * cv.T)) * (cv * rpinv * Y)
GXT = cv.T * np.mat(
(rpinv * (1. / (1. + cv * rpinv * cv.T)) * (cv * rpinv)) * X.T,
dtype=floattype)
tempmatrix = np.mat(np.zeros(X.T.shape, dtype=floattype))
np.multiply(X.T, rpinv, tempmatrix)
#tempmatrix = rpinv * X.T
np.subtract(tempmatrix, GXT, GXT)
diagG = []
for i in range(tsize):
diagGi = rpinv - cv.T[i, 0] * ca[0, i]
diagG.append(diagGi)
diagG = np.mat(diagG, dtype=floattype).T
self.selected = []
self.performances = []
currentfcount = 0
temp2 = np.mat(np.zeros(tempmatrix.shape, dtype=floattype))
while currentfcount < self.desiredfcount:
np.multiply(X.T, GXT, tempmatrix)
XGXTdiag = np.sum(tempmatrix, axis=0)
XGXTdiag = 1. / (1. + XGXTdiag)
np.multiply(GXT, XGXTdiag, tempmatrix)
tempvec1 = np.multiply((X * self.dualvec).T, XGXTdiag)
np.multiply(GXT, tempvec1, temp2)
np.subtract(self.dualvec, temp2, temp2)
np.multiply(tempmatrix, GXT, tempmatrix)
np.subtract(diagG, tempmatrix, tempmatrix)
np.divide(1, tempmatrix, tempmatrix)
np.multiply(tempmatrix, temp2, tempmatrix)
if not self.measure is None:
np.subtract(Y, tempmatrix, tempmatrix)
np.multiply(temp2, 0, temp2)
np.add(temp2, Y, temp2)
looperf = self.measure.multiTaskPerformance(temp2, tempmatrix)
looperf = np.mat(looperf, dtype=floattype)
if self.measure.iserror:
looperf[0, self.selected] = float('inf')
bestcind = np.argmin(looperf)
self.bestlooperf = np.amin(looperf)
else:
looperf[0, self.selected] = -float('inf')
bestcind = np.argmax(looperf)
self.bestlooperf = np.amax(looperf)
else:
np.multiply(tempmatrix, tempmatrix, temp2)
looperf = np.sum(temp2, axis=0)
looperf[0, self.selected] = float('inf')
bestcind = np.argmin(looperf)
self.bestlooperf = np.amin(looperf)
self.loo_predictions = Y - tempmatrix[:, bestcind]
self.looperf = looperf #Needed in test_GreedyRLS module
self.performances.append(self.bestlooperf)
cv = X[bestcind]
GXT_bci = GXT[:, bestcind]
ca = GXT_bci * (1. / (1. + cv * GXT_bci))
self.dualvec = self.dualvec - ca * (cv * self.dualvec)
diagG = diagG - np.multiply(ca, GXT_bci)
np.multiply(tempmatrix, 0, tempmatrix)
np.add(tempmatrix, ca, tempmatrix)
tempvec1 = cv * GXT
np.multiply(tempmatrix, tempvec1, tempmatrix)
np.subtract(GXT, tempmatrix, GXT)
self.selected.append(bestcind)
currentfcount += 1
#Linear predictor with bias
self.A[self.selected] = X[self.selected] * self.dualvec
self.b = bias_slice * self.dualvec * np.sqrt(self.bias)
self.predictor = predictor.LinearPredictor(self.A, self.b)
if not self.callbackfun is None:
self.callbackfun.callback(self)
if not self.callbackfun is None:
self.callbackfun.finished(self)
self.A[self.selected] = X[self.selected] * self.dualvec
self.b = bias_slice * self.dualvec * np.sqrt(self.bias)
self.results[SELECTED_FEATURES] = self.selected
self.results[GREEDYRLS_LOO_PERFORMANCES] = self.performances
#self.results['predictor'] = self.getModel()
self.predictor = predictor.LinearPredictor(self.A, self.b)
def _solve_cython(self, regparam):
self.regparam = regparam
X = self.X
Y = self.Y
bias_slice = np.sqrt(self.bias) * np.mat(
np.ones((1, X.shape[1]), dtype=np.float64))
tsize = self.size
fsize = X.shape[0]
assert X.shape[1] == tsize
self.A = np.mat(np.zeros((fsize, Y.shape[1])))
rp = regparam
rpinv = 1. / rp
#Biaz
cv = np.sqrt(self.bias) * np.mat(np.ones((1, tsize)))
ca = rpinv * (1. / (1. + cv * rpinv * cv.T)) * (cv * rpinv)
self.dualvec = rpinv * Y - cv.T * rpinv * (
1. / (1. + cv * rpinv * cv.T)) * (cv * rpinv * Y)
XT = X.T
GXT = rpinv * XT - cv.T * rpinv * (1. / (1. + cv * rpinv * cv.T)) * (
(cv * rpinv) * XT)
diagG = []
for i in range(tsize):
diagGi = rpinv - cv.T[i, 0] * ca[0, i]
diagG.append(diagGi)
diagG = np.array(diagG)
listX = []
for ci in range(fsize):
listX.append(X[ci])
self.selected = []
currentfcount = 0
self.performances = []
selectedvec = np.zeros(fsize, dtype=np.int16)
tempvec1, tempvec2, tempvec3 = np.zeros(tsize), np.zeros(
Y.shape[1]), np.zeros((tsize, Y.shape[1]))
while currentfcount < self.desiredfcount:
if not self.measure is None:
self.bestlooperf = None
else:
self.bestlooperf = 9999999999.
self.looperf = np.ones(fsize) * float('Inf')
#'''
bestcind = _greedy_rls.find_optimal_feature(
np.array(Y), np.array(X), np.array(GXT), diagG,
np.array(self.dualvec), self.looperf, fsize, tsize, Y.shape[1],
selectedvec, tempvec1, tempvec2, tempvec3)
#Reference code
'''
diagG = np.mat(diagG).T
for ci in allinds:
if ci in self.selected: continue
cv = listX[ci_mapped]
GXT_ci = GXT[:, ci_mapped]
ca = GXT_ci * (1. / (1. + cv * GXT_ci))
updA = self.dualvec - ca * (cv * self.dualvec)
invupddiagG = 1. / (diagG - np.multiply(ca, GXT_ci))
if not self.measure is None:
loopred = Y - np.multiply(invupddiagG, updA)
looperf_i = self.measure.multiOutputPerformance(Y, loopred)
if self.bestlooperf is None:
self.bestlooperf = looperf_i
bestcind = ci
if self.measure.comparePerformances(looperf_i, self.bestlooperf) > 0:
bestcind = ci
self.bestlooperf = looperf_i
else:
#This default squared performance is a bit faster to compute than the one loaded separately.
loodiff = np.multiply(invupddiagG, updA)
print loodiff
foo
looperf_i = mean(np.multiply(loodiff, loodiff))
if looperf_i < self.bestlooperf:
bestcind = ci
self.bestlooperf = looperf_i
self.looperf[ci] = looperf_i
'''
#'''
self.bestlooperf = self.looperf[bestcind]
self.looperf = np.mat(self.looperf)
self.performances.append(self.bestlooperf)
ci_mapped = bestcind
cv = listX[ci_mapped]
GXT_bci = GXT[:, ci_mapped]
ca = GXT_bci * (1. / (1. + cv * GXT_bci))
self.dualvec = self.dualvec - ca * (cv * self.dualvec)
diagG = diagG - np.array(np.multiply(ca, GXT_bci)).reshape(
(self.size))
GXT = GXT - ca * (cv * GXT)
self.selected.append(bestcind)
currentfcount += 1
#Linear predictor with bias
self.A[self.selected] = X[self.selected] * self.dualvec
self.b = bias_slice * self.dualvec * np.sqrt(self.bias)
self.predictor = predictor.LinearPredictor(self.A, self.b)
if not self.callbackfun is None:
self.callbackfun.callback(self)
if not self.callbackfun is None:
self.callbackfun.finished(self)
self.A[self.selected] = X[self.selected] * self.dualvec
self.b = bias_slice * self.dualvec * np.sqrt(self.bias)
self.results[SELECTED_FEATURES] = self.selected
self.results[GREEDYRLS_LOO_PERFORMANCES] = self.performances
self.predictor = predictor.LinearPredictor(self.A, self.b)
def getModel(self):
return predictor.LinearPredictor(self.A, self.b)
def solve_weak(self, regparam):
X = self.X
Y = self.Y
tsize = self.size
fsize = X.shape[0]
assert X.shape[1] == tsize
self.A = np.mat(np.zeros((fsize, Y.shape[1])))
rp = regparam
rpinv = 1. / rp
desiredfcount = self.desiredfcount
if not fsize >= desiredfcount:
raise Exception('The overall number of features ' + str(fsize) +
' is smaller than the desired number ' +
str(desiredfcount) +
' of features to be selected.')
#Biaz
bias_slice = np.sqrt(self.bias) * np.mat(
np.ones((1, X.shape[1]), dtype=np.float64))
cv = np.sqrt(self.bias) * np.mat(np.ones((1, tsize)))
ca = rpinv * (1. / (1. + cv * rpinv * cv.T)) * (cv * rpinv)
self.dualvec = rpinv * Y - cv.T * rpinv * (
1. / (1. + cv * rpinv * cv.T)) * (cv * rpinv * Y)
diagG = []
for i in range(tsize):
diagGi = rpinv - cv.T[i, 0] * ca[0, i]
diagG.append(diagGi)
diagG = np.mat(diagG).T
U, S, VT = la.svd(cv, full_matrices=False)
U, S, VT = np.mat(U), np.mat(S), np.mat(VT)
Omega = 1. / (S * S + rp) - rpinv
self.selected = []
notselected = set(range(fsize))
currentfcount = 0
self.performances = []
while currentfcount < desiredfcount:
if not self.measure is None:
bestlooperf = None
else:
bestlooperf = float('inf')
X_s = X[self.selected]
self.looperf = []
#sample_60 = pyrandom.sample(notselected, len(notselected))
#sample_60 = sorted(sample_60)
#print sample_60
sample_60 = pyrandom.sample(notselected, 60)
for ci in sample_60:
cv = X[ci]
GXT_ci = VT.T * np.multiply(Omega.T,
(VT * cv.T)) + rpinv * cv.T
ca = GXT_ci * (1. / (1. + cv * GXT_ci))
updA = self.dualvec - ca * (cv * self.dualvec)
invupddiagG = 1. / (diagG - np.multiply(ca, GXT_ci))
if not self.measure is None:
loopred = Y - np.multiply(invupddiagG, updA)
looperf_i = self.measure.multiOutputPerformance(Y, loopred)
if bestlooperf is None:
bestlooperf = looperf_i
bestcind = ci
if self.measure.comparePerformances(
looperf_i, bestlooperf) > 0:
bestcind = ci
bestlooperf = looperf_i
else:
#This default squared performance is a bit faster to compute than the one loaded separately.
loodiff = np.multiply(invupddiagG, updA)
looperf_i = np.mean(
np.sum(np.multiply(loodiff, loodiff), axis=0))
if looperf_i < bestlooperf:
bestcind = ci
bestlooperf = looperf_i
bestlooperf = looperf_i
self.looperf.append(looperf_i)
self.looperf = np.mat(self.looperf)
self.bestlooperf = bestlooperf
print bestlooperf
self.performances.append(bestlooperf)
cv = X[bestcind]
GXT_bci = VT.T * np.multiply(Omega.T, (VT * cv.T)) + rpinv * cv.T
ca = GXT_bci * (1. / (1. + cv * GXT_bci))
self.dualvec = self.dualvec - ca * (cv * self.dualvec)
diagG = diagG - np.multiply(ca, GXT_bci)
#GXT = GXT - ca * (cv * GXT)
self.selected.append(bestcind)
notselected.remove(bestcind)
X_sel = X[self.selected]
if isinstance(X_sel, sp.base.spmatrix):
X_sel = X_sel.todense()
U, S, VT = la.svd(np.vstack([X_sel, bias_slice]),
full_matrices=False)
U, S, VT = np.mat(U), np.mat(S), np.mat(VT)
Omega = 1. / (np.multiply(S, S) + rp) - rpinv
currentfcount += 1
#Linear predictor with bias
self.A[self.selected] = X[self.selected] * self.dualvec
self.b = bias_slice * self.dualvec
if not self.callbackfun is None:
self.callbackfun.callback(self)
if not self.callbackfun is None:
self.callbackfun.finished(self)
self.A[self.selected] = X[self.selected] * self.dualvec
self.b = bias_slice * self.dualvec
self.results['selected_features'] = self.selected
self.results['GreedyRLS_LOO_performances'] = self.performances
self.predictor = predictor.LinearPredictor(self.A, self.b)
def solve_tradeoff(self, regparam):
"""Trains RLS with the given value of the regularization parameter
@param regparam: value of the regularization parameter
@type regparam: float
"""
self.regparam = regparam
X = self.X
Y = self.Y
if not hasattr(self, "bias"):
self.bias = 0.
tsize = self.size
fsize = X.shape[0]
assert X.shape[1] == tsize
self.A = np.mat(np.zeros((fsize, Y.shape[1])))
rp = regparam
rpinv = 1. / rp
if not self.resource_pool.has_key('subsetsize'):
raise Exception("Parameter 'subsetsize' must be given.")
desiredfcount = int(self.resource_pool['subsetsize'])
if not fsize >= desiredfcount:
raise Exception('The overall number of features ' + str(fsize) +
' is smaller than the desired number ' +
str(desiredfcount) +
' of features to be selected.')
#Biaz
bias_slice = np.sqrt(self.bias) * np.mat(
np.ones((1, X.shape[1]), dtype=np.float64))
cv = bias_slice
ca = rpinv * (1. / (1. + cv * rpinv * cv.T)) * (cv * rpinv)
self.dualvec = rpinv * Y - cv.T * rpinv * (
1. / (1. + cv * rpinv * cv.T)) * (cv * rpinv * Y)
diagG = []
for i in range(tsize):
diagGi = rpinv - cv.T[i, 0] * ca[0, i]
diagG.append(diagGi)
diagG = np.mat(diagG).T
#listX = []
#for ci in range(fsize):
# listX.append(X[ci])
U, S, VT = la.svd(cv, full_matrices=False)
U, S, VT = np.mat(U), np.mat(S), np.mat(VT)
Omega = 1. / (S * S + rp) - rpinv
self.selected = []
blocksize = 1000
blocks = []
blockcount = 0
while True:
startind = blockcount * blocksize
if (blockcount + 1) * blocksize < fsize:
print blockcount, fsize, (blockcount + 1) * blocksize
endind = (blockcount + 1) * blocksize
blocks.append(range(startind, endind))
blockcount += 1
else:
blocks.append(range(startind, fsize))
blockcount += 1
break
currentfcount = 0
self.performances = []
while currentfcount < desiredfcount:
if not self.measure is None:
self.bestlooperf = None
else:
self.bestlooperf = float('inf')
looperf = np.mat(np.zeros((1, fsize)))
for blockind in range(blockcount):
block = blocks[blockind]
tempmatrix = np.mat(np.zeros((tsize, len(block))))
temp2 = np.mat(np.zeros((tsize, len(block))))
X_block = X[block]
GXT_block = VT.T * np.multiply(
Omega.T, (VT * X_block.T)) + rpinv * X_block.T
np.multiply(X_block.T, GXT_block, tempmatrix)
XGXTdiag = sum(tempmatrix, axis=0)
XGXTdiag = 1. / (1. + XGXTdiag)
np.multiply(GXT_block, XGXTdiag, tempmatrix)
tempvec1 = np.multiply((X_block * self.dualvec).T, XGXTdiag)
np.multiply(GXT_block, tempvec1, temp2)
np.subtract(self.dualvec, temp2, temp2)
np.multiply(tempmatrix, GXT_block, tempmatrix)
np.subtract(diagG, tempmatrix, tempmatrix)
np.divide(1, tempmatrix, tempmatrix)
np.multiply(tempmatrix, temp2, tempmatrix)
if not self.measure is None:
np.subtract(Y, tempmatrix, tempmatrix)
np.multiply(temp2, 0, temp2)
np.add(temp2, Y, temp2)
looperf_block = self.measure.multiTaskPerformance(
temp2, tempmatrix)
looperf_block = np.mat(looperf_block)
else:
np.multiply(tempmatrix, tempmatrix, tempmatrix)
looperf_block = sum(tempmatrix, axis=0)
looperf[:, block] = looperf_block
if not self.measure is None:
if self.measure.isErrorMeasure():
looperf[0, self.selected] = float('inf')
bestcind = np.argmin(looperf)
self.bestlooperf = np.amin(looperf)
else:
looperf[0, self.selected] = -float('inf')
bestcind = np.argmax(looperf)
self.bestlooperf = np.amax(looperf)
else:
looperf[0, self.selected] = float('inf')
bestcind = np.argmin(looperf)
self.bestlooperf = np.amin(looperf)
self.looperf = looperf
self.performances.append(self.bestlooperf)
#cv = listX[bestcind]
cv = X[bestcind]
#GXT_bci = GXT[:, bestcind]
GXT_bci = VT.T * np.multiply(Omega.T, (VT * cv.T)) + rpinv * cv.T
ca = GXT_bci * (1. / (1. + cv * GXT_bci))
self.dualvec = self.dualvec - ca * (cv * self.dualvec)
diagG = diagG - np.multiply(ca, GXT_bci)
#GXT = GXT - ca * (cv * GXT)
self.selected.append(bestcind)
X_sel = X[self.selected]
if isinstance(X_sel, sp.base.spmatrix):
X_sel = X_sel.todense()
U, S, VT = la.svd(np.vstack([X_sel, bias_slice]),
full_matrices=False)
U, S, VT = np.mat(U), np.mat(S), np.mat(VT)
#print U.shape, S.shape, VT.shape
Omega = 1. / (np.multiply(S, S) + rp) - rpinv
#print self.selected
#print self.performances
currentfcount += 1
#Linear predictor with bias
self.A[self.selected] = X[self.selected] * self.dualvec
self.b = bias_slice * self.dualvec
self.callback()
#print who(locals())
self.finished()
self.A[self.selected] = X[self.selected] * self.dualvec
self.b = bias_slice * self.dualvec
self.results['selected_features'] = self.selected
self.results['GreedyRLS_LOO_performances'] = self.performances
#self.results['predictor'] = self.getModel()
self.predictor = predictor.LinearPredictor(self.A, self.b)
def __init__(self,
X,
Y,
regparam=1.0,
qids=None,
callbackfun=None,
**kwargs):
self.regparam = regparam
self.callbackfun = None
self.Y = array_tools.as_2d_array(Y)
#Number of training examples
self.size = Y.shape[0]
if self.Y.shape[1] > 1:
raise Exception(
'CGRankRLS does not currently work in multi-label mode')
self.learn_from_labels = True
self.callbackfun = callbackfun
self.X = csc_matrix(X.T)
if qids is not None:
self.qids = map_qids(qids)
self.splits = qids_to_splits(self.qids)
else:
self.qids = None
regparam = self.regparam
qids = self.qids
if qids is not None:
P = sp.lil_matrix((self.size, len(set(qids))))
for qidind in range(len(self.splits)):
inds = self.splits[qidind]
qsize = len(inds)
for i in inds:
P[i, qidind] = 1. / sqrt(qsize)
P = P.tocsr()
PT = P.tocsc().T
else:
P = 1. / sqrt(self.size) * (np.mat(
np.ones((self.size, 1), dtype=np.float64)))
PT = P.T
X = self.X.tocsc()
X_csr = X.tocsr()
def mv(v):
v = np.mat(v).T
return X_csr * (X.T * v) - X_csr * (P * (PT *
(X.T * v))) + regparam * v
G = LinearOperator((X.shape[0], X.shape[0]),
matvec=mv,
dtype=np.float64)
Y = self.Y
if not self.callbackfun is None:
def cb(v):
self.A = np.mat(v).T
self.b = np.mat(np.zeros((1, 1)))
self.callbackfun.callback(self)
else:
cb = None
XLY = X_csr * Y - X_csr * (P * (PT * Y))
try:
self.A = np.mat(cg(G, XLY, callback=cb)[0]).T
except Finished:
pass
self.b = np.mat(np.zeros((1, 1)))
self.predictor = predictor.LinearPredictor(self.A, self.b)
def createModel(self, svdlearner):
A = svdlearner.A
A = self.reducedSetTransformation(A)
mod = predictor.LinearPredictor(A, 0.)
return mod