def get_2d_intuition_data(n_samples=100, seed=10, L=10, C=1., **kwargs):
X, y = get_toy_data(n_samples=n_samples, seed=seed)
trainer = svm.SVMTrainer('linear', C)
predictor = trainer.train(X, y, remove_zero=False)
alpha, b = predictor._weights, predictor._bias
weight = (alpha * predictor._support_vector_labels).reshape(-1, 1) * predictor._support_vectors
weight = weight.sum(axis=0)
y_p = y.copy()
dist = np.abs(np.dot(X, weight.T) + b)
flip_inds = np.argsort(dist)[::-1][:L]
flip_pnts = X[flip_inds]
y_p[flip_inds] *= -1
# plt.figure()
# svm_plot(X, y_p)
# boundary_plot(X, predictor)
# plt.scatter(flip_pnts[:, 0], flip_pnts[:, 1], s=85 * 2, facecolors='none', edgecolors='green')
# plt.show()
return X, y_p, flip_pnts
def get_adv_data(n_samples=100,
seed=16,
C=1.,
R=25,
beta1=0.1,
beta2=0.1,
L=10,
**kwargs):
X, y = get_toy_data(
n_samples=n_samples,
seed=seed,
)
trainer = svm.SVMTrainer('linear', C)
predictor = trainer.train(X, y, remove_zero=False)
alpha, b = predictor._weights, predictor._bias
s = predictor.score(X)
s *= predictor._support_vector_labels
s = normalize(s)
# svm_plot(X, y)
# boundary_plot(X, predictor)
# plt.show()
error_hist = []
yp_hist = []
flip_pnts_hist = []
print('training error on untainted data is ', calc_error(s, y))
for i in range(R):
alpha_rnd = np.random.uniform(-C, C, size=alpha.shape)
b_rnd = np.random.uniform(-C, C)
predictor_rnd = svm.SVMPredictor(
weights=alpha_rnd,
support_vectors=predictor._support_vectors,
support_vector_labels=predictor._support_vector_labels,
bias=b_rnd,
sigma=predictor._sigma,
kernel=predictor._kernel)
# weight = (alpha_rnd * predictor_rnd._support_vector_labels).reshape(-1, 1) * predictor._support_vectors
# weight = weight.sum(axis=0)
print('training error of random svm is ',
calc_error(predictor_rnd.predict(X), y))
# svm_plot(X, y)
# boundary_plot(X, predictor_rnd)
# plt.show()
q = predictor_rnd.score(X)
q *= predictor._support_vector_labels
q = normalize(q)
v = alpha / C - beta1 * s - beta2 * q
# plt.figure()
# plt.plot(alpha,'o')
# plt.plot(s,'x')
# plt.plot(q,'.')
# plt.legend()
# plt.show()
k = np.argsort(v, axis=0)
y_p = y.copy()
y_p[k[0:L]] *= -1
predictor_new = trainer.train(X, y_p)
print('training error on tainted data is ',
calc_error(predictor_new.predict(X), y))
# plt.figure()
# svm_plot(X, y_p)
# boundary_plot(X, predictor_new)
flip_pnts = X[k[0:L]]
# plt.scatter(flip_pnts[:, 0], flip_pnts[:, 1], s=85 * 2, facecolors='none', edgecolors='green')
# plt.show()
error_hist.append(calc_error(predictor_new.predict(X), y))
yp_hist.append(y_p)
flip_pnts_hist.append(flip_pnts)
print(np.max(error_hist), error_hist)
y_p = yp_hist[np.argmax(error_hist)]
flip_pnts = flip_pnts_hist[np.argmax(error_hist)]
return X, y_p, flip_pnts
# from lz import *
import logging, numpy as np, matplotlib.pyplot as plt
logging.root.setLevel(logging.ERROR)
import svm, data
exp = 'proc'
if exp == 'proc':
C = 1.
kernel = 'rbf'
proc_train = np.loadtxt('proc-train', delimiter=' ')
X, y = proc_train[:, 1:], proc_train[:, 0]
X_test = np.loadtxt('proc-test', delimiter=' ')
trainer = svm.SVMTrainer(kernel, C)
predictor = trainer.train(X, y, remove_zero=True)
y_pred = predictor.predict(X_test)
np.savetxt('y-pred', y_pred, delimiter=' ')
elif exp == 'sonar':
n_samples = 208
n_features = 60
C = 1.
L = 208 // 10
kernel = 'rbf'
X, y = data.get_sonar_data()
X, y, X_val, y_val = data.split_train_test(X, y)
trainer = svm.SVMTrainer(kernel, C)
predictor = trainer.train(X, y, remove_zero=True)
print(predictor.error(X_val, y_val))
X, y, _ = data.apply_rand_flip(X, y, L)
import numpy as np
import pandas as pd
import svm
data = pd.read_table('wdbc.data', sep=',', header=None)
X = np.array(data.ix[:, 2:])
y = data.ix[:, 1]
y = np.array([1.0 if label == 'B' else -1.0 for label in y]).ravel()
n_samples, n_features = X.shape
c = 0.1
trainer = svm.SVMTrainer(svm.Kernel.linear(), c)
predictor = trainer.train(X, y)
test = X[0].reshape(1, n_features)
predictions = [
predictor.predict(sample.reshape(1, n_features)) for sample in X
]
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)
result.append(predictor.predict(point))
Z = np.array(result).reshape(xx.shape)
plt.contourf(xx, yy, Z,
cmap=cm.Paired,
levels=[-0.001, 0.001],
extend='both',
alpha=0.5)
plt.scatter(flatten(X[:, 0]), flatten(X[:, 1]),
c=flatten(y), cmap=cm.Paired)
plt.xlim(x_min, x_max)
plt.ylim(y_min, y_max)
plt.savefig(filename)
num_samples=100
num_features=2
grid_size=20
# Random 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
# Train and predict
trainer = svm.SVMTrainer(svm.Kernel.gaussian(0.5), 0.1)
predictor = trainer.train(samples, labels)
plot(predictor, samples, labels, grid_size, "svm-demo.pdf")
def test_on_dataset(dataset = 'sonar', kernel = 'rbf', n_samples = 208, n_features = 60):
l = 1;
fig = plt.figure(figsize=(8,3))
for C in [1, 10, 100, 1000]:
ax = fig.add_axes([0.25*l-0.20, 0.3,0.18,0.4])
acc_1 = []
acc_2 = []
acc_3 = []
for L in [0, n_samples//10, n_samples//5, int(n_samples//3.3333), int(n_samples//2.5)]:
X_, y_ = data.get_train_data(dataset, n_samples, n_features)
X_c, y_c, = data.split_train_test(X_, y_)
# trainer = svm.SVMTrainer(kernel, C)
# predictor = trainer.train(X, y, remove_zero=True)
# print(predictor.error(X_val, y_val))
# print(acc_1, acc_2, acc_3)
acc_1.append(0)
acc_2.append(0)
acc_3.append(0)
for i in range(5):
X, y = [], []
for j in range(5):
if j != i:
X.extend(X_c[j])
y.extend(y_c[j])
X_val, y_val = np.array(X_c[i]), np.array(y_c[i])
X, y = np.array(X), np.array(y)
X, y, flip_pnts = data.apply_rand_flip(X, y, L)
trainer = svm.SVMTrainer(kernel, C)
predictor = trainer.train(X, y, remove_zero=True)
acc_1[-1] += (1-predictor.error(X_val, y_val))
#print(acc_1[-1])
trainer = svm.SVMTrainer(kernel, C, ln_robust=True, mu=0.1)
predictor = trainer.train(X, y, remove_zero=True)
acc_2[-1] += (1-predictor.error(X_val, y_val))
#print(acc_2[-1])
trainer = svm.SVMTrainer(kernel, C, ln_robust=True, mu=0.5 - 1e-4)
predictor = trainer.train(X, y, remove_zero=True)
acc_3[-1] += (1-predictor.error(X_val, y_val))
#print(acc_3[-1])
acc_1[-1] /= 5
acc_2[-1] /= 5
acc_3[-1] /= 5
acc_1 = np.array(acc_1)
acc_2 = np.array(acc_2)
acc_3 = np.array(acc_3)
flip_ratio = np.linspace(0, 40,acc_1.shape[0])
plt.ylim((0.3,1))
print(acc_1)
ax.plot(flip_ratio,acc_1,color="blue", label='mu=0')
ax.plot(flip_ratio,acc_2,color="red", label='mu=0.1')
ax.plot(flip_ratio,acc_3,color="black", label='mu=0.5')
ax.set_xlabel('% flipped labels')
ax.set_ylabel('test acc')
ax.set_title("C = %d" % C)
ax.spines['right'].set_color('black')
ax.spines['top'].set_color('black')
ax.spines['left'].set_color('black')
ax.spines['bottom'].set_color('black')
ax.patch.set_facecolor("white")
ax.grid(color='r', linestyle='--',linewidth=1, alpha=0.3)
#if l == 1:
# ax.legend(facecolor='white')
l += 1
fig.suptitle(dataset, fontsize=12)
num_samples, num_features))
y = np.ravel(2 * (X.sum(axis=1) > 0) - 1.0)
linearly_separable = (X, y)
names = [
"SVM Linear c = 1", "SVM RBF c = 1", "SVM Linear c = 100",
"SVM RBF c = 100"
]
datasets = [
make_moons(noise=0.1, random_state=0),
make_circles(noise=0.1, factor=0.5, random_state=1),
linearly_separable,
]
classifiers = [
svm.SVMTrainer(kernel=linear(), c=1),
svm.SVMTrainer(kernel=radial_basis(gamma=1), c=1),
svm.SVMTrainer(kernel=linear(), c=100),
svm.SVMTrainer(kernel=radial_basis(gamma=1), c=100),
]
print(linear())
figure = plt.figure(figsize=(27, 9))
i = 1
for ds_cnt, ds in enumerate(datasets):
X, y = ds
## modify labels to correspond to SVM categories {-1, 1}
y[y == 0] = -1
X = StandardScaler().fit_transform(X)
import svm import numpy as np import cvxopt import plotFigure """ Some functions """ X = np.array([[1, 2], [2, 2], [0, 0], [-2, 3]]).copy() Y = np.array([-1., -1., 1., 1.]).copy() trainer = svm.SVMTrainer(svm.Kernel.gaussian(5), c=200) predictor = trainer.train(X, Y) plotFigure.plot(predictor, X, Y, grid_size=100, filename='TEST.pdf') print predictor.predict(np.array([0, 1]))