def gridsearch_for_linear(X, y):
"""
Parameter tuning for the linear classifier in two stages. First tuning is done on a coarse grid, second on a finer grid at the position of the optimal values of the first grid.
:param X: Data x
:param y: Labels y
:return: Best parameters
"""
n_cpu = multiprocessing.cpu_count()
IOHelper.write("Linear SVC: Starting coarse gridsearch.")
# LinSvm gridSearch
c_range = np.logspace(1, 10, 10, base=10.0)
param_grid = dict(C=c_range)
grid = GridSearchCV(LinearSVC(), param_grid=param_grid, n_jobs=n_cpu)
grid.fit(X, y)
_c = grid.best_params_['C']
IOHelper.write("Linear SVC: Finished coarse gridsearch with params: C: " + str(_c))
IOHelper.write("Linear SVC: Starting fine gridsearch:")
# c_range_2 = np.linspace(_c - 0.5 * _c, _c + 0.5 * _c, num=5)
c_range_2 = [_c - 0.5 * _c, _c, 2 * _c]
param_grid = dict(C=c_range_2)
grid = GridSearchCV(LinearSVC(), param_grid=param_grid, n_jobs=n_cpu)
grid.fit(X, y)
_c = grid.best_params_['C']
IOHelper.write("Linear SVC: Finished fine gridsearch with params: C: " + str(_c))
return _c
def gridsearch_for_gauss(X, y):
"""
Parameter tuning for the gauss classifier in two stages. First tuning is done on a coarse grid, second on a finer grid at the position of the optimal val
:param X: Data x
:param y: Labels y
:return: Best parameters
"""
n_cpu = multiprocessing.cpu_count()
print("Using multiprocessing. Avaiable cores: " + str(n_cpu))
IOHelper.write("Gauss SVC: Starting gridsearch for gaussian classifier.")
c_range = np.logspace(1, 10, 10, base=10.0)
gamma_range = np.logspace(-3, 2, 6, base=10.0)
# c_range = np.logspace(-4, 1, 6, base=10.0)
# gamma_range = np.logspace(-9, -3, 6, base=10.0)
param_grid = dict(gamma=gamma_range, C=c_range)
grid = GridSearchCV(SVC(kernel="rbf"), param_grid=param_grid, n_jobs=n_cpu)
grid.fit(X, y)
_c = grid.best_params_['C']
_gamma = grid.best_params_['gamma']
print("First search complete. Starting second search...")
IOHelper.write("Gauss SVC: Finished coarse gridsearch with params: C: " + str(_c) + " gamma: " + str(_gamma))
IOHelper.write("Gauss SVC: Starting fine for gaussian classifier.")
c_range_2 = [_c - 0.5 * _c, _c, 2 * _c]
gamma_range_2 = [_gamma - 0.5 * _gamma, _gamma, 2 * _gamma]
param_grid = dict(gamma=gamma_range_2, C=c_range_2)
grid = GridSearchCV(SVC(kernel="rbf"), param_grid=param_grid, n_jobs=n_cpu)
grid.fit(X, y)
_c = grid.best_params_['C']
_gamma = grid.best_params_['gamma']
IOHelper.write("Gauss SVC: Finished fine gridsearch with params: C: " + str(_c) + " gamma: " + str(_gamma))
return _c, _gamma
def predict(self, X):
"""
Predicts the labels for the given data vector X. Uses the range _gauss_distance defined in the fit()-method to determine which classifier should predict which element in the data vector x.
:param X: Data vector.
:return: Vector of predictions.
"""
time_start = time.time()
if self._verbose:
IOHelper.write("Starting predicting.")
"""
If-Construct to account for the border cases (all points for one classifier):
(1) margins = [0, 0]: All points used for the linear SVM.
(2) e.g. margins = [-0.3, 0.3] Points distributed between both. Standard case.
(3) margins = [1, -1]: All points used for the gauss SVM. (fit()-Method set margins to -1)
"""
if not self._use_distance and 0 < self._k < 1.0:
n = np.ceil(self._k * X.shape[0]) # Random ziehen.
gauss_indices = random.sample(np.arange(X.shape[0]), int(n))
lin_indices = np.setdiff1d(np.arange(X.shape[0]), gauss_indices)
lin_predictions = self._lin_svc.predict(X[lin_indices])
gauss_predictions = self._gauss_svc.predict(X[gauss_indices])
predictions = np.zeros(len(lin_predictions) + len(gauss_predictions))
predictions[lin_indices] = lin_predictions
predictions[gauss_indices] = gauss_predictions
self._time_predict = time.time() - time_start
if self._verbose:
IOHelper.write("Finished predicting.")
return predictions
if self._gauss_distance == 0.0: # (1)
predictions = self._lin_svc.predict(X)
self._time_predict = time.time() - time_start
if self._verbose:
IOHelper.write("Finished predicting.")
return predictions
if 0.0 < self._gauss_distance: # (2)
fx = abs(self._lin_svc.decision_function(X)) / np.linalg.norm(self._lin_svc.coef_[0])
gauss_indices = np.where(fx < self._gauss_distance)
lin_indices = np.where(fx >= self._gauss_distance)
lin_predictions = self._lin_svc.predict(X[lin_indices])
gauss_predictions = self._gauss_svc.predict(X[gauss_indices])
predictions = np.zeros(len(lin_predictions) + len(gauss_predictions))
predictions[lin_indices] = lin_predictions
predictions[gauss_indices] = gauss_predictions
self._time_predict = time.time() - time_start
if self._verbose:
IOHelper.write("Finished predicting.")
return predictions
if self._gauss_distance == -1: # (3)
predictions = self._gauss_svc.predict(X)
self._time_predict = time.time() - time_start
if self._verbose:
IOHelper.write("Finished predicting.")
return predictions
# If no condition matched
raise Exception("Fatal error: Count param")
def fit(self, X, y):
"""
Fit the model according to the given training data.
Fits a linear SVC on the given data.
Afterwards, certain datapoints are selected and given to a gaussian SVC. The selection is dependant on the attribute useFactor of this object.
:param X: Training vector
:param y: Target vector relative to X
:return: Returns self.
"""
if self._verbose:
IOHelper.write("Starting fitting process.\n")
IOHelper.write("Starting fitting process for linear SVC.")
time_start_lin = time.time()
self._lin_svc.fit(X, y)
self._time_fit_lin = time.time() - time_start_lin
if self._verbose:
IOHelper.write("Completed fitting process for linear SVC.")
IOHelper.write("Sorting points for classifiers.")
time_start_overhead = time.time()
x, y, gauss_distance = self.get_points_close_to_hyperplane_by_count(X, y, self._k)
try:
self._n_gauss = x.shape[0] # Measure the number of points for gauss classifier:
except AttributeError:
self._n_gauss = len(x)
self._gauss_distance = gauss_distance
self._time_overhead = time.time() - time_start_overhead
if (self._verbose):
IOHelper.write("Sorting finished.")
IOHelper.write("Starting fitting process for gaussian SVC.")
# Measure the number of points for linear classifier:
self._n_lin = X.shape[0] - self._n_gauss
time_start_gauss = time.time()
if self._n_gauss != 0:
self._gauss_svc.fit(x, y)
self._time_fit_gauss = time.time() - time_start_gauss
if self._verbose:
IOHelper.write("Completed fitting process for gaussian SVC.")
IOHelper.write("Finished fitting process.\n")
return self
def gridsearch_and_save(data):
'''
Method that searches the best parameters for the given data-set for the DualSvm for different k and saves it to a text file.
:param data: String, name of the data. Search is done automatically in the data directory.
:return: None. Output will be written to a textfile.
'''
IOHelper.write("Starting parameter tuning for " + data)
x, x_test, y, y_test = DataLoader.load_data(data)
file_string = "output/" + data + "-params.csv"
#file_string = "output/" + data + "-params.csv"
k = 0
n = 0
c_lin = 0
c_gauss = [0, 0, 0, 0, 0, 0, 0, 0, 0]
gamma = [0, 0, 0, 0, 0, 0, 0, 0, 0]
try:
output = open(file_string, 'w')
except Exception:
try:
output = open("../" + file_string, 'w')
except Exception:
output = open("error-params.txt", 'w')
for j in range(4): # Smaller steps from 0 to 20: 0, 5, 10, 15
n = j
IOHelper.write("Batch run " + str(j) + ", k = " + str(0.05 * j))
# Load the classifier
k = 0.05 * j
clf = DualSvm(use_distance=True)
clf.k = k
# Parameter Tuning
if j == 0: # In the first run, calculate best parameters for linear svm
c_lin = gridsearch_for_linear(x, y)
else:
clf.c_lin = c_lin
clf.fit_lin_svc(x,
y) # Fit linear classifier beforehand. This is necessary for the get_points method to work correctly.
x_gauss, y_gauss, margins = clf.get_points_close_to_hyperplane_by_count(x, y, k)
c_gauss[n], gamma[n] = gridsearch_for_gauss(x_gauss,
y_gauss) # In the following runs, do the same for the gaussian svm, as the subset of points for the classifier is changing
for i in range(5): # Bigger steps from 20 to 100: 20, 40, 60, 80, 100
n = 4 + i
IOHelper.write("Batch run " + str(i + 4) + ", k = " + str(0.2 * (i + 1)))
# Load the classifier
k = 0.2 * (i + 1)
clf = DualSvm(use_distance=True)
clf.k = k
if k <= 0.6:
clf.c_lin = c_lin
clf.fit_lin_svc(x, y)
x_gauss, y_gauss, margins = clf.get_points_close_to_hyperplane_by_count(x, y, k)
c_gauss[n], gamma[n] = gridsearch_for_gauss(x_gauss, y_gauss)
output.write(str(c_lin) + "\n")
for value in c_gauss:
output.write(str(value) + ",")
output.write("\n")
for value in gamma:
output.write(str(value) + ",")
output.write("\n")