def __init__(self,
kernel_type='rbf',
C=1,
gamma=1,
degree=3,
tolerance=0.1,
epsilon=0.1,
max_iter=100,
solver="smo"):
self.__kernel = Kernel(kernel_type, gamma, degree)
self.__C = C
self.__tol = tolerance
self.__error_cache = {}
self.__eps = epsilon
self.__max_iter = max_iter
self.__solver = solver
def example(num_samples=10, num_features=2, grid_size=20, filename="svm.pdf"):
samples = np.array(
np.random.normal(size=num_samples * num_features).reshape(
num_samples, num_features))
labels = 2 * (samples.sum(axis=1) > 0) - 1.0
data_dict = build_data_dict(samples, labels)
svm = SVM(data=data_dict, kernel=Kernel.linear(), c=0.1)
svm.fit()
plot(svm, samples, labels, grid_size, filename)
def custom_SVM_test_function(X_train, y_train, X_test):
data_dict = dataset_reader.build_data_dict(X_train, y_train)
svm = custom_svm.SVM(data=data_dict, kernel=Kernel.linear(), c=1)
svm.fit()
y_pred = []
print(X_test[0])
for x in X_test:
y_pred.append(svm.predict(x))
return y_pred
def train_eval(config, args, X_train, Y_train, X_test=None, Y_test=None):
seed = str(args['seed']) if not args['split_ready'] else ''
model_path = "%s%s_%s.pkl" % (args['model_path'], args['dataset'], seed)
ker = Kernel(config, args['kernel_type'])
logging.info('Training on dataset %s...' % args['dataset'])
logging.info('\tComputing %s kernel.' % args['kernel_type'])
K_train = ker.fit_transform(X_train)
lins = []
nans = []
for col in range(Y_train.shape[1]):
Y_train_all = Y_train[:, col]
K_train_notnan = K_train[~np.isnan(Y_train_all)][:, ~np.
isnan(Y_train_all)]
Y_train_notnan = Y_train_all[~np.isnan(Y_train_all)]
nans.append(np.isnan(Y_train_all))
if args['metric'] in ['ROC', 'PRC']:
logging.info('\tTraining classifier on task %d.' % (col + 1))
lin = svm.SVC(kernel='precomputed', C=10, probability=True)
lin.fit(K_train_notnan, Y_train_notnan)
else:
logging.info('\tTraining regressor on task %d.' % (col + 1))
lin = svm.SVR(kernel='precomputed', C=10)
lin.fit(K_train_notnan, Y_train_notnan)
lins.append(lin)
model = {'kernel': ker, 'linear': lins, 'nans': nans}
save_model(model, model_path)
logging.info('\tTrained model saved to \"%s\".' %
(model_path.split('/')[-1]))
if X_test is not None and Y_test is not None:
score = evaluate(args, X_test, Y_test)
logging.info('\tAll tasks averaged score (%s): %.6f.' %
(args['metric'], score))
return score
def plot_results(time_points, values):
axis_x = np.arange(0,5.1,0.1)
fig = plt.figure(0)
plt.axis([0,5,-2,2], facecolor = 'g')
plt.grid(color='w', linestyle='-', linewidth=0.5)
ax = fig.add_subplot(111)
ax.spines['top'].set_visible(False)
ax.spines['right'].set_visible(False)
ax.patch.set_facecolor('#E8E8F1')
# show mean
mu = np.zeros(axis_x.size)
var = np.zeros(axis_x.size)
ker = Kernel()
ker.SE(1,1)
gp = GP()
for i in range(axis_x.size):
mu[i],var[i],_ = gp.GPR(time_points = time_points,values = values, predict_point = axis_x[i], kernel = ker)
# show covariance
print mu
plt.fill_between(axis_x,mu + var,mu-var,color = '#D1D9F0')
# show mean
plt.plot(axis_x, mu, linewidth = 2, color = "#5B8CEB")
# show the points
plt.scatter(time, value,color = '#598BEB')
plt.show()
def example(grid_size=30):
samples, labels = generate_data()
# samples = np.matrix(np.random.normal(size=num_samples * num_features).reshape(num_samples, num_features))
# labels = 2 * (samples.sum(axis=1) > 0) - 1.0
shufled_data = np.concatenate((samples, labels), axis=1)
np.random.shuffle(shufled_data)
f1_scores = []
parts = 5
trainer = SVMTrainer(Kernel.gaussian_trans(0.5), 0.1)
for i in range(parts):
train_data, train_labels, test_data, test_labels = split_data(
shufled_data, parts, i)
predictor = trainer.train(train_data, train_labels)
f1 = f_measure(test_labels, predictor, test_data)
print(f1)
f1_scores.append(f1)
print(np.average(np.array(f1_scores)))
predictor = trainer.train(samples, labels)
plot(predictor, samples, labels, grid_size)
class TestKernel(unittest.TestCase):
def setUp(self):
self.python = Kernel('python', os.path.join(PATH, 'pykernel.py'))
self.ruby = Kernel('ruby', os.path.join(PATH, 'rubykernel.rb'))
self.kernels = [self.python, self.ruby]
def tearDown(self):
for kernel in self.kernels:
kernel.terminate()
def test_encode_message(self):
message = 'a\nb c'
for kernel in self.kernels:
encoded = kernel._encode(message).decode(kernel._ENCODING)
self.assertEqual(encoded, u'a\uffffb c')
def test_decode_message(self):
message = u'while 1:\n pass'
for kernel in self.kernels:
encoded = message.encode(kernel._ENCODING)
decoded = kernel._decode(encoded)
self.assertEqual(decoded, message)
def test_eval_python(self):
self.python.send('print range(5)')
result = self.python.recv()
self.assertEqual(result, '[0, 1, 2, 3, 4]\n')
def test_eval_ruby(self):
self.ruby.send('puts 2 + 3')
result = self.ruby.recv()
self.assertEqual(result, '5\n')
def test_eval_code(self):
for kernel in self.kernels:
result = kernel.eval_code('print 5+3')
self.assertEqual(result, 8)
def setUp(self):
self.python = Kernel('python', os.path.join(PATH, 'pykernel.py'))
self.ruby = Kernel('ruby', os.path.join(PATH, 'rubykernel.rb'))
self.kernels = [self.python, self.ruby]
class SVM:
"""
Binary Support Vector Classification
Parameters
----------
kernel_type : string, optional (default = "rbf")
- Kernel to be used to transform data
C : float, optional (default = 1)
- Coefficient of error term in Soft Margin Obj Function
gamma : float, optional (default = 1)
- Paramter in RBF and Sigmoid Kernel Functions
degree : float, optional (default = 3)
- Degree of Polynomial in Polynomial Kernel Function
tolerance : float, optional (default = 1e-4)
- tolerance for stopping criteria
epsilon : float, optional (defualt = 1e-4)
- UPDATE AFTER UNDERSTANDING
max_iter : int
- The maximum number of iterations of SMO.
solver : string, optional (default = "smo")
- Which optimization algorithm to use for the dual form of the Obj
"""
def __init__(self,
kernel_type='rbf',
C=1,
gamma=1,
degree=3,
tolerance=0.1,
epsilon=0.1,
max_iter=100,
solver="smo"):
self.__kernel = Kernel(kernel_type, gamma, degree)
self.__C = C
self.__tol = tolerance
self.__error_cache = {}
self.__eps = epsilon
self.__max_iter = max_iter
self.__solver = solver
def __del__(self):
pass
def fit(self, x, y):
"""
Parameters
----------
x : ndarray
- Data
y : ndarray
- Labels
Actions
-------
Creates the SVM model over x and y
- Calls SMO to solve for alphas and b
- Sets the support vectors
"""
self.__xs = x
self.__ys = y
self.__size = x.shape[0]
self.__kernel_mat = self.__get_kernel_matrix()
# Get the alphas and b from smo
self.__alphas, self.__b = self.__solve(self.__solver)
supp_idxs = np.nonzero(self.__alphas)[0]
self.__supp_x = np.take(self.__xs, supp_idxs, axis=0)
self.__supp_y = np.take(self.__ys, supp_idxs)
self.__supp_a = np.take(self.__alphas, supp_idxs)
def __get_kernel_matrix(self):
"""
Actions:
-------
Compute and store the kernel matrix KM over the input data
"""
self.__KM = np.zeros((self.__size, self.__size))
for i in range(self.__size):
for j in range(self.__size):
self.__KM[i, j] = self.__kernel.eval(self.__xs[i],
self.__xs[j])
def __initialize_error_cache(self):
"""
Actions
-------
Initialize the error over each training example
Since the model output is initially 0 for all xi,
the Error is 0 - yi = -yi
"""
for i in range(self.__size):
self.__error_cache[i] = -self.__ys[i]
def __solve(self, solver):
"""
Parameters
----------
solver : string
- The quadratic optimization algorithm to use
Actions:
--------
Calls the appropriate optimizer function
"""
if (solver == "smo"):
return self.__smo()
def __get_bounds(self, s, i1, i2):
"""
Parameters
----------
s : int
- 1 if x[i1] and x[i2] have same label, -1 otherwise
i1 : int
- Index of second chosen Lagrange multiplier
i2 : int
- Index of first chosen Lagrange multiplier
Returns:
--------
The minimum and maximum values that the new alpha2 can take
"""
return L, H
def __optimize_and_clip(self, a2_old, y2, E1, E2, eta, L, H):
"""
Parameters
----------
a2_old : float
- current values of alphas[i2]
y2 : int
- label of x[i2]
E1 : float
- Error on x[i1]
E2 : float
- Error on x[i2]
eta : floats
- Second derivative of dual objective
L : float
- Lower bound for new alpha2
H : float
- Upper bound for new alpha2
Returns:
--------
a2 : float
- The new optimal, clipped lagrangian multiplier
"""
a2 = a2_old + ((y2 * (E1 - E2)) / eta)
if (a2 < L):
a2 = L
elif (a2 > H):
a2 = H
return a2
def __take_step(self, i1, i2, E2):
"""
Parameters
----------
i1 : int
- Index of second chosen Lagrange multiplier
i2 : int
- Index of second first Lagrange multiplier
E2 : float
- Error on the first chosen Lagrange multiplier
Actions
-------
Update alphas by optimizing 2
Find optimal value for 1 alpha, clip so constraints are not violated, then solve for
Returns
-------
"""
# Same alphas or invalid alpha
if (i1 == i2 or i1 == -1):
return 0
# old alphas
alph1 = self.__alphas[i1]
alph2 = self.__alphas[i2]
# ys
y1 = self.__ys[i1]
y2 = self.__ys[i2]
# Error on alph1
E1 = self.__error_cache[i1]
s = y1 * y2
# Compute Lower and Upper bounds
if (s > 0):
L = max(0, self.__alphas[i2] - self.__alphas[i1] - self.__C)
H = min(self.__C, self.__alphas[i2] + self.__alphas[i1])
else:
L = max(0, self.__alphas[i2] - self.__alphas[i1])
H = min(self.__C, self.__C + self.__alphas[i2] - self.__alphas[i1])
# No feasible alphas here
if (L == H):
return 0
# eta = k11 + k22 - k12
k11 = self.__KM[i1, i1]
k22 = self.__KM[i2, i2]
k12 = self.__KM[i1, i2]
eta = k11 + k22 - (2 * k12)
# Optimum is valid. Set new a2 to optimal value and then clip
if (eta > 0):
a2 = alph2 + ((y2 * (E1 - E2)) / eta)
if (a2 < L):
a2 = L
elif (a2 > H):
a2 = H
# eta is equal to 0,so we can't get the optimum for a2
# settle for setting it to the max of the endpoints of the constraint line
else:
f1 = (y1 * (E1 + self.__b)) - (self.__alphas[i1] *
k11) - (s * self.__alphas[i2] * k12)
f2 = (y2 * (E2 + self.__b)) - (s * self.__alphas[i1] *
k12) - (self.__alphas[i2] * k22)
L1 = self.__alphas[i1] + (s * (self.__alphas[i2] - L))
H1 = self.__alphas[i1] + (s * (self.__alphas[i2] - H))
Lobj = L1 * f1 + L * f2 + 0.5 * pow(L1, 2) * k11 + 0.5 * pow(
L, 2) * k22 + s * L * L1 * k12
Hobj = H1 * f1 + H * f2 + 0.5 * pow(H1, 2) * k11 + 0.5 * pow(
H, 2) * k22 + s * H * H1 * k12
if (Lobj < Hobj - self.__eps):
a2 = L
elif (Lobj < Hobj + self.__eps):
a2 = H
else:
a2 = alph2
# Not a significant change in a2 so dont change it and return 0
if (abs(a2 - alph2) < self.__eps * (a2 + alph2 + self.__eps)):
return 0
# Compute a1 from box linear constraint
a1 = alph1 + s * (alph2 - a2)
# Make sure a1 is in feasible region
if (a1 < 0):
a2 += s * a1
a1 = 0
elif (a1 > self.__C):
a2 += s * (a1 - self.__C)
a1 = self.__C
# Update threshold to reflect change in Lagrange multipliers
b1 = self.__b + E1 + y1 * (a1 - alph1) * k11 + y2 * (a2 - alph2) * k12
b2 = self.__b + E2 + y1 * (a1 - alph1) * k12 + y2 * (a2 - alph2) * k22
b_new = (b1 + b2) / 2
delta_b = b_new = self.__b
self.__b = b_new
# Update error cache using new Lagrange multipliers
t1 = y1 * (a1 - alph1)
t2 = y2 * (a2 - alph2)
for i in range(self.__size):
if (self.__alphas[i] > 0 and self.__alphas[i] < self.__C):
self.__error_cache[i] += t1 * self.__KM[
i1, i] + t2 * self.__KM[i2, i] - delta_b
self.__error_cache[i1] = 0
self.__error_cache[i2] = 0
# store new alphas
self.__alphas[i1] = a1
self.__alphas[i2] = a2
return 1
def __second_choice_heuristic(self, E2):
"""
Parameters
----------
E2 : int
- Error on first selected alpha value
Returns
-------
i1 : int
- The index of the training example that maximizes the absolute value of errors E1 and E2
"""
tmax = 0
i1 = -1
for k in range(self.__size):
if (self.__alphas[k] > 0 and self.__alphas[k] < self.__C):
E1 = self.__error_cache[k]
temp = abs(E2 - E1)
if (temp > tmax):
tmax = temp
i1 = k
return i1
def __examine_example(self, i2):
"""
Paramters
---------
i2 : int
- Index of current training example
Actions:
-------
Checks if alphas[i2] violates the KKT conditions
If it does, use the heuristics to select the second alpha value
Then optimize them jointly by calling takeStep
Returns
------
int (0 or 1)
- Whether or not alphas[i2] was changed
"""
y2 = self.__ys[i2]
alph2 = self.__alphas[i2]
E2 = np.sign(self.__u(idx=i2)) - y2
r2 = E2 * y2
if ((r2 < -self.__tol and alph2 < self.__C)
or (r2 > self.__tol and alph2 > 0)):
i1 = self.__second_choice_heuristic(E2)
if (self.__take_step(i1, i2, E2)):
return 1
idx = random.randint(0, self.__size)
for i1 in list(range(idx, self.__size)) + list(range(0, idx)):
if (self.__alphas[i1] > 0 and self.__alphas[i1] < self.__C):
if (self.__take_step(i1, i2, E2)):
return 1
for i1 in list(range(idx, self.__size)) + list(range(0, idx)):
if (self.__take_step(i1, i2, E2)):
return 1
return 0
def __smo(self):
"""
Sequential Minimal Optimization of Dual Form of Obj Function
Returns
-------
self.__alphas : ndarray
- Optimal Lagrange Multipliers in Obj Function
self.__b : float
- straight line distance to optimal hyperplane
"""
self.__alphas = np.zeros(self.__size)
self.__b = 0
# num alphas changed in 1 pass of training set
num_changed = 0
# variable used to alternate between 1 pass over all
# training examples and 1 pass over examples whose
# Lagrange multipliers are not at their bounds
examine_all = 1
# Store errors to speed up algorithm
self.__initialize_error_cache()
steps = 0
while (num_changed > 0 or examine_all):
steps += 1
if (steps > self.__max_iter):
break
if (examine_all):
# loop over entire training set
for i in range(self.__size):
num_changed += self.__examine_example(i)
else:
for i in range(self.__size):
# loop over non boundary training examples
if (self.__alphas[i] != 0
and self.__alphas[i] != self.__C):
num_changed += self.__examine_example(i)
if (examine_all == 1):
examine_all = 0
elif (num_changed == 0):
examine_all = 1
return self.__alphas, self.__b
def __u(self, x=None, idx=None):
"""
Parameters
----------
x : ndarray, optional (default = None)
One data vector
idx : int, optional (default = None)
If not None, x = self.xs[idx]
Returns
-------
u : float
-The evaluation of the decision function at point x.
Note: When this function is called during training,
we use the precomputed kernel vector. When called
during testing, we compute the kernel vector.
"""
if (idx != None):
kernel_vector = self.__KM[idx, :]
return np.sum(kernel_vector * self.__ys * self.__alphas) - self.__b
else:
if (self.__supp_x.shape[0] == 0):
return -self.__b
kernel_vector = np.apply_along_axis(self.__kernel.eval,
1,
self.__supp_x,
x2=x)
mult = kernel_vector * self.__supp_y, *self.__supp_a
return np.sum(mult[0]) - self.__b
def predict(self, xs):
"""
Parameters
----------
xs : ndarray
- Input samples to predict the labels of
Returns
-------
ndarray
- Predicted labels of xs
"""
def sign(x):
return np.sign(self.__u(x=x))
return np.apply_along_axis(sign, 1, xs)
def __init__(self, max_num, domain):
self.max_num = max_num
rho0 = 1000. #rest density [ kg/m^3 ]
VF = .0262144 #simulation volume [ m^3 ]
VP = VF / max_num #particle volume [ m^3 ]
m = rho0 * VP #particle mass [ kg ]
re = (VP)**(1 / 3.) #particle radius [ m ]
#re = (VP)**(1/2.) #particle radius [ m ]
print "re, m, VP", re, m, VP
rest_distance = .87 * re #rest distance between particles [ m ]
smoothing_radius = 2.0 * rest_distance #smoothing radius for SPH Kernels
boundary_distance = .5 * rest_distance #for calculating collision with boundary
#the ratio between the particle radius in simulation space and world space
print "VF", VF
print "domain.V: ", domain.V
print "VF/domain.V", VF / domain.V
print "scale calc", (VF / domain.V)**(1 / 3.)
#print "scale calc", (VF/domain.V)**(1/2.)
sim_scale = (VF / domain.V)**(1 / 3.) #[m^3 / world m^3 ]
#sim_scale = (VF / domain.V)**(1/2.) #[m^2 / world m^2 ]
self.rho0 = rho0
self.VF = VF
self.mass = m
self.VP = VP
self.re = re
self.rest_distance = rest_distance
self.smoothing_radius = smoothing_radius
self.boundary_distance = boundary_distance
self.sim_scale = sim_scale
print "====================================================="
print "particle mass:", self.mass
print "Fluid Volume VF:", self.VF
print "simulation scale:", self.sim_scale
print "smoothing radius:", self.smoothing_radius
print "rest distance:", self.rest_distance
print "====================================================="
#Other parameters
self.K = 15. #Gas constant
self.boundary_stiffness = 20000.
self.boundary_dampening = 256.
#friction
self.friction_coef = 0.
self.restitution_coef = 0.
#not used yet
self.shear = 0.
self.attraction = 0.
self.spring = 0.
self.velocity_limit = 600.
self.xsph_factor = .05
self.viscosity = .01
self.gravity = -9.8
self.EPSILON = 10E-6
#Domain
self.domain = domain
self.domain.setup(self.smoothing_radius / self.sim_scale)
#Kernels
self.kernels = Kernel(self.smoothing_radius)
#Load Train and embedding path_data = 'dataset/Xtr0.csv' path_label = 'dataset/Ytr0.csv' train_dataset = datahandler(path_data, path_label,features_generated=False) train_dataset.Y[train_dataset.Y==0]=-1 train_dataset.compute_vocabulary(6) train_dataset.mismatch_embedding(6,1,train_dataset.vocab) X_train0, Y_train = train_dataset.X_embedded,train_dataset.Y kernel = Kernel(Kernel.dot_product()) K0 = kernel.kernel_matrix(X_train0) #Load Test and embedding path_data = 'dataset/Xte0.csv' path_label = 'dataset/Ytr0.csv' test_dataset = datahandler(path_data, path_label,features_generated=False) test_dataset.mismatch_embedding(6,1,train_dataset.vocab) X_test0 = test_dataset.X_embedded ########################################## #KERNEL 2 ########################################## train_dataset.compute_vocabulary(7)
# positives = [[4,4], [5,5]]
positives = [[5, 1], [6, -1], [7, 3]]
# positives = [[-0.23530383, 1.70585848],
# [ 0.8157926 , 0.04591391],
# [ 0.03237168, 1.36243792]]
# negatives = [[2,1], [1,2]]
negatives = [[1, 7], [2, 8], [3, 8]]
# negatives = [[-0.07810244, -0.65502153],
# [ 0.25648505, -0.79438534],
# [-0.83531028, -0.18554141],
# [ 0.41896733, -0.73003242],
# [-1.00007796, 0.00366544],
# [-1.58005843, 0.83875439],
# [ 0.77187267, -1.67242829]]
data_dict = {-1: np.array(negatives), 1: np.array(positives)}
num_samples = len(positives) + len(negatives)
samples = np.append(np.array(positives),
np.array(negatives)).reshape(num_samples, 2)
labels = np.append(np.ones(len(positives)), np.ones(len(negatives)) * -1)
svm = SVM(data=data_dict, kernel=Kernel.linear(), c=0.1)
svm.fit()
plot(svm, samples, labels, 20, "svm_run.pdf")
#print(svm.predict(test_data))
'metric' : 'rmse'}
bounds = [(1.0e-5, 1.0e-1), # learning rate
(0.5, 0.9999), # change of learning rate
(2, 1000)] # number of leaves
n_random_trials = 3 # initiate Bayesian optimization with 3 random draws
n_searches = 10
# Use my Bayesian Optimization
mdl = Model(data_mat, lags, n_oos, n_val, prediction_range,
target_vars_inds, params)
kernel = Kernel("rbf", 1)
bo = BayesianOptimization(mdl.obj_fun, bounds, kernel,
expected_improvement, n_random_trials)
ind, best_para_my, y = bo.search(n_searches, 2, 25)
# Use Ax Bayesian Optimization
n_random_trials = 5 # initiate Bayesian optimization with 3 random draws
n_searches = 20
mdl = Model(data_mat, lags, n_oos, n_val, prediction_range,
target_vars_inds, params)
import matplotlib.pyplot as plt
import svm
from kernels import Kernel
iris = datasets.load_iris()
y = iris.target
y = np.asarray([-1 if x == 0 else x for x in y])
sel = [y != 2]
y = y[sel]
X = iris.data[:, :2]
X = X[sel]
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33)
C = 0.1
trainer = svm.SVMTrainer(kernel=Kernel.linear(), c=C)
model = trainer.train(X_train, y_train)
x_min, x_max = X[:, 0].min() - 1, X[:, 0].max() + 1
y_min, y_max = X[:, 1].min() - 1, X[:, 1].max() + 1
h = 0.02
xx, yy = np.meshgrid(np.arange(x_min, x_max, h), np.arange(y_min, y_max, h))
xxyy = np.stack((xx.ravel(), yy.ravel()), axis=-1)
result = []
for i in range(len(xxyy)):
result.append(model.predict(xxyy[i]))
Z = np.array(result).reshape(xx.shape)
plt.plot(axis_x, mu, linewidth=2, color=mean_color[i % k_color])
# background color
# plt.axis([0,5,-2,2], facecolor = 'g')
# plt.grid(color='w', linestyle='-', linewidth=0.5)
# ax.patch.set_facecolor('#E8E8F1')
# show the points
for i in range(len(GPs)):
plt.scatter(GPs[i].time_points,
GPs[i].values,
color=mean_color[i % k_color],
marker='X')
plt.show()
if __name__ == '__main__':
X = np.matrix([1, 2, 3, 4]).reshape(-1, 1)
Y = np.sin(X) # np.random.randn(20,1)*0.05
X = X.reshape(4, )
Y = Y.reshape(4, )
k1 = Kernel("SE", np.sqrt(2), 1)
gp1 = GP(time_points=X.T, values=Y.T, kernel=k1)
gp1.optimize()
print gp1.se_function([0.697774447341, 1.61119536129, 7.64794567566e-09])
#print k1.cal_SE()
