def test_holdout(self):
for X in [self.Xtrain1, self.Xtrain2]:
for Y in [self.Ytrain1, self.Ytrain2]:
m = X.shape[0]
qids, L = generate_qids(m)
qids = np.array(qids)
hoindices = np.where(qids == 1)[0]
hocompl = list(set(range(m)) - set(hoindices))
#Holdout with linear kernel
rls1 = QueryRankRLS(X, Y, qids)
rls2 = QueryRankRLS(X[hocompl], Y[hocompl], qids[hocompl])
P1 = rls1.holdout(hoindices)
P2 = rls2.predict(X[hoindices])
assert_allclose(P1, P2)
#Holdout with bias
rls1 = QueryRankRLS(X, Y, qids, bias=3.0)
rls2 = QueryRankRLS(X[hocompl],
Y[hocompl],
qids[hocompl],
bias=3.0)
P1 = rls1.holdout(hoindices)
P2 = rls2.predict(X[hoindices])
assert_allclose(P1, P2)
#Fast regularization
for i in range(-5, 5):
rls1.solve(2**i)
rls2.solve(2**i)
P1 = rls1.holdout(hoindices)
P2 = rls2.predict(X[hoindices])
assert_allclose(P1, P2)
#Kernel holdout
rls1 = QueryRankRLS(X,
Y,
qids,
kernel="GaussianKernel",
gamma=0.01)
rls2 = QueryRankRLS(X[hocompl],
Y[hocompl],
qids[hocompl],
kernel="GaussianKernel",
gamma=0.01)
P1 = rls1.holdout(hoindices)
P2 = rls2.predict(X[hoindices])
assert_allclose(P1, P2)
for i in range(-15, 15):
rls1.solve(2**i)
rls2.solve(2**i)
P1 = rls1.holdout(hoindices)
P2 = rls2.predict(X[hoindices])
assert_allclose(P1, P2)
#Incorrect indices
I = [0, 3, 100]
self.assertRaises(IndexError, rls1.holdout, I)
I = [-1, 0, 2]
self.assertRaises(IndexError, rls1.holdout, I)
I = [1, 1, 2]
self.assertRaises(IndexError, rls1.holdout, I)
I = [0, 4, 8]
self.assertRaises(IndexError, rls1.holdout, I)
def test_holdout(self):
for X in [self.Xtrain1, self.Xtrain2]:
for Y in [self.Ytrain1, self.Ytrain2]:
m = X.shape[0]
qids, L = generate_qids(m)
qids = np.array(qids)
hoindices = np.where(qids == 1)[0]
hocompl = list(set(range(m)) - set(hoindices))
#Holdout with linear kernel
rls1 = QueryRankRLS(X, Y, qids)
rls2 = QueryRankRLS(X[hocompl], Y[hocompl], qids[hocompl])
P1 = rls1.holdout(hoindices)
P2 = rls2.predict(X[hoindices])
assert_allclose(P1, P2)
#Holdout with bias
rls1 = QueryRankRLS(X, Y, qids, bias = 3.0)
rls2 = QueryRankRLS(X[hocompl], Y[hocompl], qids[hocompl], bias = 3.0)
P1 = rls1.holdout(hoindices)
P2 = rls2.predict(X[hoindices])
assert_allclose(P1, P2)
#Fast regularization
for i in range(-5, 5):
rls1.solve(2**i)
rls2.solve(2**i)
P1 = rls1.holdout(hoindices)
P2 = rls2.predict(X[hoindices])
assert_allclose(P1, P2)
#Kernel holdout
rls1 = QueryRankRLS(X, Y, qids, kernel = "GaussianKernel", gamma = 0.01)
rls2 = QueryRankRLS(X[hocompl], Y[hocompl], qids[hocompl], kernel = "GaussianKernel", gamma = 0.01)
P1 = rls1.holdout(hoindices)
P2 = rls2.predict(X[hoindices])
assert_allclose(P1, P2)
for i in range(-15, 15):
rls1.solve(2**i)
rls2.solve(2**i)
P1 = rls1.holdout(hoindices)
P2 = rls2.predict(X[hoindices])
assert_allclose(P1, P2)
#Incorrect indices
I = [0, 3, 100]
self.assertRaises(IndexError, rls1.holdout, I)
I = [-1, 0, 2]
self.assertRaises(IndexError, rls1.holdout, I)
I = [1,1,2]
self.assertRaises(IndexError, rls1.holdout, I)
I = [0,4,8]
self.assertRaises(IndexError, rls1.holdout, I)
def train_rls():
#Select regparam with k-fold cross-validation,
#where instances related to a single sentence form
#together a fold
X_train = read_sparse("train_2000_x.txt")
Y_train = np.loadtxt("train_2000_y.txt")
X_test = read_sparse("test_2000_x.txt", X_train.shape[1])
Y_test = np.loadtxt("test_2000_y.txt")
#list of sentence ids
qids_train = np.loadtxt("train_2000_qids.txt")
qids_test = np.loadtxt("test_2000_qids.txt")
learner = QueryRankRLS(X_train, Y_train, qids_train)
P_test = learner.predict(X_test)
folds = map_ids(qids_train)
perfs = []
for fold in folds:
if np.var(Y_train[fold]) != 0:
P = learner.holdout(fold)
c = cindex(Y_train[fold], P)
perfs.append(c)
perf = np.mean(perfs)
print("leave-query-out cross-validation cindex %f" % perf)
partition = map_ids(qids_test)
test_perfs = []
#compute the ranking accuracy separately for each test query
for query in partition:
#skip such queries, where all instances have the same
#score, since in this case cindex is undefined
if np.var(Y_test[query]) != 0:
perf = cindex(Y_test[query], P_test[query])
test_perfs.append(perf)
test_perf = np.mean(test_perfs)
print("test cindex %f" % test_perf)
def train_rls():
#Select regparam with k-fold cross-validation,
#where instances related to a single sentence form
#together a fold
X_train = read_sparse("train_2000_x.txt")
Y_train = np.loadtxt("train_2000_y.txt")
X_test = read_sparse("test_2000_x.txt", X_train.shape[1])
Y_test = np.loadtxt("test_2000_y.txt")
#list of sentence ids
qids_train = np.loadtxt("train_2000_qids.txt")
qids_test = np.loadtxt("test_2000_qids.txt")
learner = QueryRankRLS(X_train, Y_train, qids_train)
P_test = learner.predict(X_test)
folds = map_ids(qids_train)
perfs = []
for fold in folds:
if np.var(Y_train[fold]) != 0:
P = learner.holdout(fold)
c = cindex(Y_train[fold], P)
perfs.append(c)
perf = np.mean(perfs)
print("leave-query-out cross-validation cindex %f" %perf)
partition = map_ids(qids_test)
test_perfs = []
#compute the ranking accuracy separately for each test query
for query in partition:
#skip such queries, where all instances have the same
#score, since in this case cindex is undefined
if np.var(Y_test[query]) != 0:
perf = cindex(Y_test[query], P_test[query])
test_perfs.append(perf)
test_perf = np.mean(test_perfs)
print("test cindex %f" %test_perf)
def testLabelRankRLS(self):
print("Testing the cross-validation routines of the QueryRankRLS module.\n")
np.random.seed(100)
floattype = np.float64
m, n = 100, 400 #data, features
Xtrain = np.mat(np.random.rand(m, n))
K = Xtrain * Xtrain.T
ylen = 1
Y = np.mat(np.zeros((m, ylen), dtype=floattype))
Y[:, 0] = np.sum(Xtrain, 1)
labelcount = 5
hoindices = range(labelcount)
hocompl = list(set(range(m)) - set(hoindices))
qidlist = [0 for i in range(100)]
for h in range(5, 12):
qidlist[h] = 1
for h in range(12, 32):
qidlist[h] = 2
for h in range(32, 34):
qidlist[h] = 3
for h in range(34, 85):
qidlist[h] = 4
for h in range(85, 100):
qidlist[h] = 5
qidlist_cv = qidlist[5: len(qidlist)]
objcount = max(qidlist) + 1
P = np.mat(np.zeros((m, objcount), dtype=np.float64))
for i in range(m):
qid = qidlist[i]
P[i, qid] = 1.
labelcounts = np.sum(P, axis=0)
P = np.divide(P, np.sqrt(labelcounts))
D = np.mat(np.ones((1, m), dtype=np.float64))
L = np.multiply(np.eye(m), D) - P * P.T
Kcv = K[np.ix_(hocompl, hocompl)]
Lcv = L[np.ix_(hocompl, hocompl)]
Xcv = Xtrain[hocompl]
Xtest = Xtrain[hoindices]
Yho = Y[hocompl]
rpool = {}
rpool["X"] = Xtrain
rpool["Y"] = Y
rpool["qids"] = qidlist
primalrls = QueryRankRLS(**rpool)
rpool = {}
rpool["X"] = K
rpool['kernel'] = 'PrecomputedKernel'
rpool["Y"] = Y
rpool["qids"] = qidlist
dualrls = QueryRankRLS(**rpool)
rpool = {}
rpool['X'] = Xcv
rpool['Y'] = Yho
rpool['qids'] = qidlist_cv
primalrls_naive = QueryRankRLS(**rpool)
rpool = {}
rpool['X'] = Kcv
rpool['kernel'] = 'PrecomputedKernel'
rpool['Y'] = Yho
#rpool['X'] = Xcv
rpool['qids'] = qidlist_cv
dualrls_naive = QueryRankRLS(**rpool)
testkm = K[np.ix_(hocompl, hoindices)]
loglambdas = range(-5, 5)
for j in range(0, len(loglambdas)):
regparam = 2. ** loglambdas[j]
print
print("Regparam 2^%1d" % loglambdas[j])
print(str(np.squeeze(np.array((testkm.T * la.inv(Lcv * Kcv + regparam * np.eye(Lcv.shape[0])) * Lcv * Yho).T))) + ' Dumb HO')
predhos = []
primalrls_naive.solve(regparam)
predho = primalrls_naive.predictor.predict(Xtest)
print(str(predho.T) + ' Naive HO (primal)')
predhos.append(predho)
dualrls_naive.solve(regparam)
predho = dualrls_naive.predictor.predict(testkm.T)
print(str(predho.T) + ' Naive HO (dual)')
predhos.append(predho)
primalrls.solve(regparam)
predho = np.squeeze(primalrls.holdout(hoindices))
print(str(predho.T) + ' Fast HO (primal)')
predhos.append(predho)
dualrls.solve(regparam)
predho = np.squeeze(dualrls.holdout(hoindices))
print(str(predho.T) + ' Fast HO (dual)')
predhos.append(predho)
predho0 = predhos.pop(0)
for predho in predhos:
self.assertEqual(predho0.shape, predho.shape)
for row in range(predho.shape[0]):
#for col in range(predho.shape[1]):
# self.assertAlmostEqual(predho0[row,col],predho[row,col], places=5)
self.assertAlmostEqual(predho0[row],predho[row], places=5)
def testLabelRankRLS(self):
print("Testing the cross-validation routines of the QueryRankRLS module.\n")
np.random.seed(100)
floattype = np.float64
m, n = 100, 400 #data, features
Xtrain = np.mat(np.random.rand(m, n))
K = Xtrain * Xtrain.T
ylen = 1
Y = np.mat(np.zeros((m, ylen), dtype=floattype))
Y[:, 0] = np.sum(Xtrain, 1)
labelcount = 5
hoindices = range(labelcount)
hocompl = list(set(range(m)) - set(hoindices))
qidlist = [0 for i in range(100)]
for h in range(5, 12):
qidlist[h] = 1
for h in range(12, 32):
qidlist[h] = 2
for h in range(32, 34):
qidlist[h] = 3
for h in range(34, 85):
qidlist[h] = 4
for h in range(85, 100):
qidlist[h] = 5
qidlist_cv = qidlist[5: len(qidlist)]
objcount = max(qidlist) + 1
P = np.mat(np.zeros((m, objcount), dtype=np.float64))
for i in range(m):
qid = qidlist[i]
P[i, qid] = 1.
labelcounts = np.sum(P, axis=0)
P = np.divide(P, np.sqrt(labelcounts))
D = np.mat(np.ones((1, m), dtype=np.float64))
L = np.multiply(np.eye(m), D) - P * P.T
Kcv = K[np.ix_(hocompl, hocompl)]
Lcv = L[np.ix_(hocompl, hocompl)]
Xcv = Xtrain[hocompl]
Xtest = Xtrain[hoindices]
Yho = Y[hocompl]
rpool = {}
rpool["X"] = Xtrain
rpool["Y"] = Y
rpool["qids"] = qidlist
primalrls = QueryRankRLS(**rpool)
rpool = {}
rpool["X"] = K
rpool['kernel'] = 'PrecomputedKernel'
rpool["Y"] = Y
rpool["qids"] = qidlist
dualrls = QueryRankRLS(**rpool)
rpool = {}
rpool['X'] = Xcv
rpool['Y'] = Yho
rpool['qids'] = qidlist_cv
primalrls_naive = QueryRankRLS(**rpool)
rpool = {}
rpool['X'] = Kcv
rpool['kernel'] = 'PrecomputedKernel'
rpool['Y'] = Yho
#rpool['X'] = Xcv
rpool['qids'] = qidlist_cv
dualrls_naive = QueryRankRLS(**rpool)
testkm = K[np.ix_(hocompl, hoindices)]
loglambdas = range(-5, 5)
for j in range(0, len(loglambdas)):
regparam = 2. ** loglambdas[j]
print
print("Regparam 2^%1d" % loglambdas[j])
print(str(np.squeeze(np.array((testkm.T * la.inv(Lcv * Kcv + regparam * np.eye(Lcv.shape[0])) * Lcv * Yho).T))) + ' Dumb HO')
predhos = []
primalrls_naive.solve(regparam)
predho = primalrls_naive.predictor.predict(Xtest)
print(str(predho.T) + ' Naive HO (primal)')
predhos.append(predho)
dualrls_naive.solve(regparam)
predho = dualrls_naive.predictor.predict(testkm.T)
print(str(predho.T) + ' Naive HO (dual)')
predhos.append(predho)
primalrls.solve(regparam)
predho = np.squeeze(primalrls.holdout(hoindices))
print(str(predho.T) + ' Fast HO (primal)')
predhos.append(predho)
dualrls.solve(regparam)
predho = np.squeeze(dualrls.holdout(hoindices))
print(str(predho.T) + ' Fast HO (dual)')
predhos.append(predho)
predho0 = predhos.pop(0)
for predho in predhos:
self.assertEqual(predho0.shape, predho.shape)
for row in range(predho.shape[0]):
#for col in range(predho.shape[1]):
# self.assertAlmostEqual(predho0[row,col],predho[row,col], places=5)
self.assertAlmostEqual(predho0[row],predho[row], places=5)
