def test_cross_val_generator_mask_indices_same():
# Test that the cross validation generators return the same results when
# indices=True and when indices=False
y = np.array([0, 0, 0, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2])
labels = np.array([1, 1, 2, 3, 3, 3, 4])
loo_mask = cval.LeaveOneOut(5, indices=False)
loo_ind = cval.LeaveOneOut(5, indices=True)
lpo_mask = cval.LeavePOut(10, 2, indices=False)
lpo_ind = cval.LeavePOut(10, 2, indices=True)
kf_mask = cval.KFold(10, 5, indices=False, shuffle=True, random_state=1)
kf_ind = cval.KFold(10, 5, indices=True, shuffle=True, random_state=1)
skf_mask = cval.StratifiedKFold(y, 3, indices=False)
skf_ind = cval.StratifiedKFold(y, 3, indices=True)
lolo_mask = cval.LeaveOneLabelOut(labels, indices=False)
lolo_ind = cval.LeaveOneLabelOut(labels, indices=True)
lopo_mask = cval.LeavePLabelOut(labels, 2, indices=False)
lopo_ind = cval.LeavePLabelOut(labels, 2, indices=True)
for cv_mask, cv_ind in [(loo_mask, loo_ind), (lpo_mask, lpo_ind),
(kf_mask, kf_ind), (skf_mask, skf_ind),
(lolo_mask, lolo_ind), (lopo_mask, lopo_ind)]:
for (train_mask, test_mask), (train_ind, test_ind) in \
zip(cv_mask, cv_ind):
assert_array_equal(np.where(train_mask)[0], train_ind)
assert_array_equal(np.where(test_mask)[0], test_ind)
def test_leave_label_out_changing_labels():
# Check that LeaveOneLabelOut and LeavePLabelOut work normally if
# the labels variable is changed before calling __iter__
labels = np.array([0, 1, 2, 1, 1, 2, 0, 0])
labels_changing = np.array(labels, copy=True)
lolo = cval.LeaveOneLabelOut(labels)
lolo_changing = cval.LeaveOneLabelOut(labels_changing)
lplo = cval.LeavePLabelOut(labels, p=2)
lplo_changing = cval.LeavePLabelOut(labels_changing, p=2)
labels_changing[:] = 0
for llo, llo_changing in [(lolo, lolo_changing), (lplo, lplo_changing)]:
for (train, test), (train_chan, test_chan) in zip(llo, llo_changing):
assert_array_equal(train, train_chan)
assert_array_equal(test, test_chan)
def train_svm(train_vector, train_labels, object_ids):
"""
train_svm - expects a vector of features and a nx1 set of
corresponding labels
Returns a trained SVM classifier
"""
# Create the obj_id_vector for cross validation
lpl = cross_validation.LeavePLabelOut(object_ids, p=1, indices=True)
# Grid search with nested cross-validation
parameters = {
'kernel': ['linear'],
'C': [1, 1e1, 1e2, 1e3, 1e4],
'gamma': [1, 1e-1, 1e-2, 1e-3, 1e-4],
'degree': [2, 3, 4, 5]
}
#parameters = {'kernel': ['linear'], 'C': [1, 1e1], 'gamma': [1, 1e-1, 1e-2]}
svm = GridSearchCV(SVC(probability=True),
parameters,
score_func=f1_score,
cv=lpl)
# Train the SVM using the best parameters
svm.fit(train_vector, train_labels)
svm_best = svm.best_estimator_
return svm_best
def train_gradient_boost(train_vector, train_labels, object_ids):
"""
train_svm - expects a vector of features and a nx1 set of
corresponding labels
Returns a trained Gradient Boosting classifier
"""
# Create the obj_id_vector for cross validation
lpl = cross_validation.LeavePLabelOut(object_ids, p=1, indices=True)
parameters = {'n_estimators': [1000], 'learn_rate': [1e-1, 1e-2, 1e-3]}
# Grid search with nested cross-validation
#parameters = {'kernel': ['rbf'], 'C': [1, 1e1, 1e2, 1e3, 1e4], 'gamma': [1, 1e-1, 1e-2, 1e-3, 1e-4]}
clf = GridSearchCV(GradientBoostingClassifier(max_depth=4),
parameters,
score_func=f1_score,
cv=lpl)
# Train the SVM using the best parameters
clf.fit(train_vector, train_labels)
clf_best = clf.best_estimator_
return clf_best
def main():
c = mdb.connect(host=HOST, user=USER, passwd=PASSWORD, db=DATABASE)
### CLASSIFYING USERS FROM KEYSTROKE DYNAMICS
#
## get list of insights and labels
#userlist = [3, 9]
#users = []
#for user in userlist:
# values = kt.get_insights_by_user(user, c).values[:,0].tolist()
# users.extend([user]*len(values))
#users = np.array(users)
#
#rfc = ensemble.RandomForestClassifier()
#train_and_test_model(userlist, users, rfc, c)
## PREDICTING HAPPINESS LEVEL FROM ALL DATA
#users = [45]
users = [33, 35, 37, 39, 41, 43, 45, 47, 49]
if len(sys.argv) > 1:
model_type = sys.argv[1]
else:
raise ValueError("must specify model type as argument")
split_by = 'hour'
# get all data
data = get_and_impute_data(users, True, split_by, c)
happiness = get_var_by_users("happiness_level", users, split_by, c)
energy = get_var_by_users("energy_level", users, split_by, c)
relax = get_var_by_users("relax_level", users, split_by, c)
insights = get_var_by_users("id", users, split_by, c)
response = np.vstack((happiness, energy, relax)).T
n_insights = len(np.unique(insights))
# split dataset into training and evaluation datasets
for train_idx, eval_idx in cross_validation.LeavePLabelOut(
insights, int(n_insights * 0.2)):
train_data = data[train_idx]
eval_data = data[eval_idx]
train_response = response[train_idx]
eval_response = response[eval_idx]
train_insights = insights[train_idx]
eval_insights = insights[eval_idx]
break
#model = train_and_test_model(train_data, train_response[:,1], train_insights, model_type, split_by, c, varname="energy level")
#evaluate_model(eval_data, eval_response[:,1], model, "%s model evaluation" % model_type, "energy level")
model = train_and_test_model(train_data,
train_response[:, 2],
train_insights,
model_type,
split_by,
c,
varname="relaxation level")
#evaluate_model(eval_data, eval_response[:,2], model, "%s model evaluation" % models[model_type][2], "relaxation level")
np.savetxt('response.txt', np.vstack((happiness, energy, relax)))
def leavePLabelOut(X, y_true, labels):
y_pred = y_true * 0.0
lpl = cross_validation.LeavePLabelOut(labels, p=1)
i = 0
for train_index, test_index in lpl:
i += 1
print "CrossVal " + str(i) + " of " + str(len(lpl))
X_train, X_test = X[train_index], X[test_index]
y_train, y_test = y_true[train_index], y_true[test_index]
clf = RFClassifier(n_estimators=200)
#clf = GBClassifier(n_estimators=5000,learning_rate=0.05, max_features=0.25)
clf.fit(X_train, y_train)
y_pred[test_index] = clf.predict_proba(X_test)
roc_auc(y_true, y_pred)
def train_knn(train_vector, train_labels, test_vector, test_labels, scaler):
"""
train_knn - expects a vector of features and a nx1 set of
corresponding labels. Finally the number of
neighbors used for comparison
Returns a trained knn classifier
"""
# Data scaling
train_vector_scaled = scaler.transform(train_vector[0])
test_vector_scaled = scaler.transform(test_vector[0])
# Create the obj_id_vector for cross validation
lpl = cross_validation.LeavePLabelOut(np.array(train_vector[1]),
p=10,
indices=True)
import pdb
pdb.set_trace()
pass
'''
# Check the split
for train_index, test_index in lpl:
print "TRAIN_INDEX: %s TEST_INDEX: %s" % (train_index, test_index)
# the train/test_vector_scaled need to by a array
train = train_vector_scaled[train_index]
test = train_vector_scaled[test_index]
import pdb; pdb.set_trace()
pass
'''
# Grid search with nested cross-validation
parameters = {
'n_neighbors': [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]
}
knn = GridSearchCV(KNeighborsClassifier(),
parameters,
score_func=f1_score,
cv=lpl)
knn.fit(train_vector_scaled, train_labels)
knn_best = knn.best_estimator_
score = f1_score(test_labels, knn.predict(test_vector_scaled))
proba = knn.predict_proba(test_vector_scaled)
report = classification_report(test_labels,
knn.predict(test_vector_scaled))
return (knn_best, proba, score, report)
def train_univariate_selection(train_X,
train_Y,
object_ids=None,
score_fun=f1_score,
verbose=0,
n_jobs=6,
scale=False):
'''
Cross validates on the best percentage of features to keep
when doing univariate feature selection
'''
# Setup cross validation
if (object_ids is None) or (sum(train_Y) <= 10):
print "Cannot perform leave one out cross validation"
cv = 3 # default 3 fold cross validation
else:
# Leave one object out cross validation
cv = cross_validation.LeavePLabelOut(object_ids, p=1, indices=True)
if scale is True:
scaler = preprocessing.StandardScaler().fit(train_X)
train_X = scaler.transform(train_X)
else:
scaler = False
univ_select = SelectPercentile(f_classif, percentile=10)
X_features = univ_select.fit(train_X, train_Y).transform(train_X)
# class weight normalizes the lack of positive examples
svm = LinearSVC(dual=False, class_weight='auto')
svm.fit(X_features, train_Y)
pipeline = Pipeline([("features", univ_select), ("svm", svm)])
param_grid = dict(features__percentile=[20, 40, 60, 80, 100],
svm__C=np.linspace(1, 1e6, 1000))
#svm__penalty = ['l1','l2'])
grid = GridSearchCV(pipeline,
param_grid,
cv=cv,
verbose=verbose,
n_jobs=n_jobs,
score_func=score_fun)
grid.fit(train_X, train_Y)
svm_best = grid.best_estimator_
return svm_best, scaler
def test_cross_val_generator_with_mask():
X = np.array([[1, 2], [3, 4], [5, 6], [7, 8]])
y = np.array([1, 1, 2, 2])
labels = np.array([1, 2, 3, 4])
loo = cval.LeaveOneOut(4, indices=False)
lpo = cval.LeavePOut(4, 2, indices=False)
kf = cval.KFold(4, 2, indices=False)
skf = cval.StratifiedKFold(y, 2, indices=False)
lolo = cval.LeaveOneLabelOut(labels, indices=False)
lopo = cval.LeavePLabelOut(labels, 2, indices=False)
ss = cval.ShuffleSplit(4, indices=False)
for cv in [loo, lpo, kf, skf, lolo, lopo, ss]:
for train, test in cv:
X_train, X_test = X[train], X[test]
y_train, y_test = y[train], y[test]
def sample(self, X, y, folds):
n = np.shape(X)[0]
self.lpl = cross_validation.LeavePLabelOut(folds, p=1)
if self.fold == 'all':
X_train = X
y_train = y
X_test = X
y_test = y
else:
for train_index, test_index in self.lpl:
if folds[test_index[0]] == self.fold:
X_train, X_test = X[train_index], X[test_index]
y_train, y_test = y[train_index], y[test_index]
self.train_folds, self.test_folds = folds[
train_index], folds[test_index]
return X_train, X_test, y_train, y_test
def train_svm(train_vector, train_labels, test_vector, test_labels, scaler,
cv_flag):
"""
train_svm - expects a vector of features and a nx1 set of
corresponding labels
Returns a trained SVM classifier
"""
# Data scaling
train_vector_scaled = scaler.transform(train_vector[0])
test_vector_scaled = scaler.transform(test_vector[0])
# Create the obj_id_vector for cross validation
lpl = cross_validation.LeavePLabelOut(np.array(train_vector[1]),
p=2,
indices=True)
#import pdb; pdb.set_trace()
#pass
if cv_flag == True:
cv_mode = lpl
else:
cv_mode = 4
# Grid search with nested cross-validation
parameters = {
'kernel': ['rbf'],
'C': [1, 1e1, 1e2, 1e3, 1e4],
'gamma': [1, 1e-1, 1e-2, 1e-3, 1e-4]
}
svm = GridSearchCV(SVC(probability=True),
parameters,
score_func=f1_score,
cv=cv_mode)
svm.fit(train_vector_scaled, train_labels)
svm_best = svm.best_estimator_
score = f1_score(test_labels, svm.predict(test_vector_scaled))
proba = svm.predict_proba(test_vector_scaled)
report = classification_report(test_labels,
svm.predict(test_vector_scaled))
#import pdb; pdb.set_trace()
#pass
return (svm_best, proba, score, report)
def test_cross_indices_exception():
X = coo_matrix(np.array([[1, 2], [3, 4], [5, 6], [7, 8]]))
y = np.array([1, 1, 2, 2])
labels = np.array([1, 2, 3, 4])
loo = cval.LeaveOneOut(4, indices=False)
lpo = cval.LeavePOut(4, 2, indices=False)
kf = cval.KFold(4, 2, indices=False)
skf = cval.StratifiedKFold(y, 2, indices=False)
lolo = cval.LeaveOneLabelOut(labels, indices=False)
lopo = cval.LeavePLabelOut(labels, 2, indices=False)
assert_raises(ValueError, cval.check_cv, loo, X, y)
assert_raises(ValueError, cval.check_cv, lpo, X, y)
assert_raises(ValueError, cval.check_cv, kf, X, y)
assert_raises(ValueError, cval.check_cv, skf, X, y)
assert_raises(ValueError, cval.check_cv, lolo, X, y)
assert_raises(ValueError, cval.check_cv, lopo, X, y)
def train_svm_gridsearch(train_X,
train_Y,
object_ids=None,
verbose=0,
n_jobs=6,
score_fun=f1_score,
scale=False):
'''
Performs cross validation using grid search
'''
# Setup cross validation
if (object_ids is None) or (sum(train_Y) <= 10):
print "Cannot perform leave one out cross validation"
cv = 3 # 5 fold cross validation
else:
# Leave one object out cross validation
cv = cross_validation.LeavePLabelOut(object_ids, p=1, indices=True)
parameters = {
'C': np.linspace(1, 1e6, 1000),
#'C': np.linspace(1,1e6,100),
#'C': (1e-3,1e-2,1e-1,1.0, 10, 100, 1000, 1e4, 1e5, 1e6),
'penalty': ('l1', 'l2'),
}
# class weight normalizes the lack of positive examples
clf = LinearSVC(dual=False, class_weight='auto')
if scale is True:
scaler = preprocessing.StandardScaler().fit(train_X)
train_X = scaler.transform(train_X)
else:
scaler = False
grid = GridSearchCV(clf,
parameters,
cv=cv,
verbose=verbose,
n_jobs=n_jobs,
score_func=score_fun)
grid.fit(train_X, train_Y)
svm_best = grid.best_estimator_
return svm_best, scaler
def test_cross_val_generator_with_default_indices():
X = np.array([[1, 2], [3, 4], [5, 6], [7, 8]])
y = np.array([1, 1, 2, 2])
labels = np.array([1, 2, 3, 4])
loo = cval.LeaveOneOut(4)
lpo = cval.LeavePOut(4, 2)
kf = cval.KFold(4, 2)
skf = cval.StratifiedKFold(y, 2)
lolo = cval.LeaveOneLabelOut(labels)
lopo = cval.LeavePLabelOut(labels, 2)
ss = cval.ShuffleSplit(2)
ps = cval.PredefinedSplit([1, 1, 2, 2])
for cv in [loo, lpo, kf, skf, lolo, lopo, ss, ps]:
for train, test in cv:
assert_not_equal(np.asarray(train).dtype.kind, 'b')
assert_not_equal(np.asarray(train).dtype.kind, 'b')
X[train], X[test]
y[train], y[test]
def test_cross_val_generator_with_default_indices():
X = np.array([[1, 2], [3, 4], [5, 6], [7, 8]])
y = np.array([1, 1, 2, 2])
labels = np.array([1, 2, 3, 4])
loo = cval.LeaveOneOut(4)
lpo = cval.LeavePOut(4, 2)
kf = cval.KFold(4, 2)
skf = cval.StratifiedKFold(y, 2)
lolo = cval.LeaveOneLabelOut(labels)
lopo = cval.LeavePLabelOut(labels, 2)
b = cval.Bootstrap(2) # only in index mode
ss = cval.ShuffleSplit(2)
for cv in [loo, lpo, kf, skf, lolo, lopo, b, ss]:
for train, test in cv:
assert_not_equal(np.asarray(train).dtype.kind, 'b')
assert_not_equal(np.asarray(train).dtype.kind, 'b')
X_train, X_test = X[train], X[test]
y_train, y_test = y[train], y[test]
def show_cross_val(method):
if method == "lolo":
labels = np.array(["summer", "winter", "summer", "winter", "spring"])
cv = cross_validation.LeaveOneLabelOut(labels)
elif method == "lplo":
labels = np.array(["summer", "winter", "summer", "winter", "spring"])
cv = cross_validation.LeavePLabelOut(labels, p=2)
elif method == "loo":
cv = cross_validation.LeaveOneOut(n=len(y))
elif method == "lpo":
cv = cross_validation.LeavePOut(n=len(y), p=3)
for train_index, test_index in cv:
print("TRAIN:", train_index, "TEST:", test_index)
X_train, X_test = X[train_index], X[test_index]
y_train, y_test = y[train_index], y[test_index]
print "X_train:", X_train
print "y_train:", y_train
print "X_test:", X_test
print "y_test:", y_test
def train_gradient_boost(train_X,
train_Y,
object_ids=None,
score_fun=f1_score,
verbose=0,
n_jobs=6,
scale=False):
'''
Performs cross validation using grid search and
gradient tree boosting
'''
# Setup cross validation
if (object_ids is None) or (sum(train_Y) <= 10):
print "Cannot perform leave one out cross validation"
cv = 5 # 10 fold cross validation
else:
# Leave one object out cross validation
cv = cross_validation.LeavePLabelOut(object_ids, p=1, indices=True)
if scale is True:
scaler = preprocessing.StandardScaler().fit(train_X)
train_X = scaler.transform(train_X)
else:
scaler = False
parameters = {
'n_estimators': [1000],
'learn_rate': [1e-1, 1e-2, 1, 1e-3]
#'max_depth':[4]
}
print "Beginning Grid Search"
grid = GridSearchCV(GradientBoostingClassifier(max_depth=4),
parameters,
score_func=score_fun,
cv=cv,
verbose=verbose,
n_jobs=n_jobs)
grid.fit(train_X, train_Y)
svm_best = grid.best_estimator_
return svm_best, scaler
dataset = mnist.load()
dim = dataset['train']['data'][0].size
N_train = len(dataset['train']['target'])
N_test = len(dataset['test']['target'])
test_data_dict = {
'data': dataset['test']['data'].reshape(N_test, dim).astype(np.float32),
'target': dataset['test']['target'].astype(np.int32)
}
#unlabeled_data_dict = {'data':dataset['train']['data'].reshape(N_train, dim).astype(np.float32),
# 'target':-np.ones(N_train)}
unlabeled_data_dict = {
'data': dataset['train']['data'].reshape(N_train, dim).astype(np.float32)
}
# making labeled data
lplo = cross_validation.LeavePLabelOut(labels=six.moves.range(N_train),
p=args.slabeled)
fold = 1
for i in six.moves.range(fold):
train_idx, test_idx = next(iter(lplo))
labeled_data_dict = {
'data': unlabeled_data_dict['data'][test_idx].astype(np.float32),
'target': dataset['train']['target'][test_idx].astype(np.int32)
}
#
labeled_data = datafeeders.SiameseFeeder(labeled_data_dict,
batchsize=args.lbatch)
unlabeled_data = DataFeeder(unlabeled_data_dict, batchsize=args.ubatch)
test_data = datafeeders.SiameseFeeder(test_data_dict, batchsize=args.valbatch)
labeled_data.hook_preprocess(mnist_preprocess)
unlabeled_data.hook_preprocess(mnist_preprocess_u)
