def UnBiasedHSIC(self, x, y, kernelx=vector.CLinearKernel(), \
kernely=vector.CLinearKernel()):
"""
Compute the UNbiased estimator of HSIC.
Args:
x: The data.
y: The labels.
kernelx: The kernel on the data, default to linear kernel.
kernely: The kernel on the labels, default to linear kernel.
Returns:
HSIC score
"""
nx = x.shape
ny = y.shape
assert nx[0] == ny[0], \
"Argument 1 and 2 have different number of data points"
kMat = kernelx.Dot(x, x)
setdiag0(kMat)
lMat = kernely.Dot(y, y)
setdiag0(lMat)
sK = kMat.sum(axis=1)
ssK = sK.sum()
sL = lMat.sum(axis=1)
ssL = sL.sum()
return (kMat.__imul__(lMat).sum() + \
(ssK * ssL) / ((nx[0] - 1) * (nx[0] - 2)) - \
2 * sK.__imul__(sL).sum() / (nx[0] - 2) \
) / (nx[0] * (nx[0] - 3))
def BiasedHSIC(self, x, y, kernelx=vector.CLinearKernel(), \
kernely=vector.CLinearKernel()):
nx = x.shape
ny = y.shape
assert nx[0] == ny[0], \
"Argument 1 and 2 have different number of data points"
if len(nx) > 1:
kMat = kernelx.Dot(x, x)
else:
kMat = numpy.outerproduct(x, x)
hlhMat = ComputeHLH(y, kernely)
return numpy.sum(numpy.sum(kMat * hlhMat)) / ((nx[0] - 1) *
(nx[0] - 1))
def ComputeHLH(self, y, kernely=vector.CLinearKernel()):
ny = y.shape
if len(ny) > 1:
lMat = kernely.Dot(y, y)
else:
lMat = numpy.outerproduct(y, y)
sL = numpy.sum(lMat, axis=1)
ssL = numpy.sum(sL)
# hlhMat
return lMat - numpy.add.outer(sL, sL) / ny[0] + ssL / (ny[0] * ny[0])
def BiasedHSIC(self, x, y, kernelx=vector.CLinearKernel(), kernely=vector.CLinearKernel()):
"""
Compute the biased estimator of HSIC.
@param x The data.
@param y The labels.
@param kernelx The kernel on the data, default to linear kernel.
@param kernely The kernel on the labels, default to linear kernel.
"""
nx = x.shape
ny = y.shape
assert nx[0] == ny[0], \
"Argument 1 and 2 have different number of data points"
if len(nx) > 1:
kMat = kernelx.Dot(x, x)
else:
kMat = numpy.outerproduct(x, x)
hlhMat = self.ComputeHLH(y, kernely)
return numpy.sum(numpy.sum(kMat * hlhMat)) / ((nx[0] - 1) * (nx[0] - 1))
def UnBiasedHSIC(self, x, y, kernelx=vector.CLinearKernel(), \
kernely=vector.CLinearKernel()):
nx = x.shape
ny = y.shape
assert nx[0] == ny[0], \
"Argument 1 and 2 have different number of data points"
kMat = kernelx.Dot(x, x)
setdiag0(kMat)
lMat = kernely.Dot(y, y)
setdiag0(lMat)
sK = kMat.sum(axis=1)
ssK = sK.sum()
sL = lMat.sum(axis=1)
ssL = sL.sum()
return ( kMat.__imul__(lMat).sum() + \
(ssK*ssL)/((nx[0]-1)*(nx[0]-2)) - \
2 * sK.__imul__(sL).sum() / (nx[0]-2) \
) / (nx[0]*(nx[0]-3))
def ComputeHLH(self, y, kernely=vector.CLinearKernel()):
"""
Compute HLH give the labels.
@param y The labels.
@param kernely The kernel on the labels, default to linear kernel.
"""
ny = y.shape
if len(ny) > 1:
lMat = kernely.Dot(y, y)
else:
lMat = numpy.outerproduct(y, y)
sL = numpy.sum(lMat, axis=1)
ssL = numpy.sum(sL)
# hlhMat
return lMat - numpy.add.outer(sL, sL) / ny[0] + ssL / (ny[0] * ny[0])
def main():
X = load('xdata.h5')
y1 = load('y1data.h5')
y2 = load('y2data.h5')
y3 = load('y3data.h5')
y4 = load('y4data.h5')
# print X.shape
# print y1.shape
data_no = X.shape[0]
y1 = normalize(y1, data_no)
y2 = normalize(y2, data_no)
y3 = normalize(y3, data_no)
y4 = normalize(y4, data_no)
# print y
# Normalize the data.
m = X.mean(0)
s = X.std(0)
# print m,s
X.__isub__(m).__idiv__(s)
# foo=np.copy(X)
xlabels = []
ylabels = []
foo = np.copy(X)
for features_tokeep in range(5000, 10001, 1000):
X = np.copy(foo)
xlabels.append(features_tokeep)
print xlabels
bahsic = CBAHSIC()
bhs = bahsic.BAHSICRaw(X, y1, vector.CLinearKernel(),
vector.CLinearKernel(), features_tokeep, 0.1)
hsicfeatures = np.zeros(shape=(data_no, features_tokeep))
for i in range(0, data_no):
for j in range(0, features_tokeep):
hsicfeatures[i][j] = X[i][bhs[features_tokeep + j]]
X = hsicfeatures
print X.shape
# y1=y1.reshape((y1.shape[0],))
# print y1.shape
#feature extraction
# test train split
# xlabels=[]
# ylabels=[]
# y1labels=[]
# y2labels=[]
# y3labels=[]
# y4labels=[]
# pca = PCA(n_components=pcanum)
# X=pca.fit_transform(foo)
# print X.shape
labels = list(zip(y1, y2, y3, y4))
xtrain, xtest, ytrain, ytest = train_test_split(X,
labels,
test_size=0.33,
random_state=42)
tmp = zip(*ytest)
ytest1 = tmp[0]
ytest2 = tmp[1]
ytest3 = tmp[2]
ytest4 = tmp[3]
ytest1 = np.array(ytest1)
ytest2 = np.array(ytest2)
ytest3 = np.array(ytest3)
ytest4 = np.array(ytest4)
tmp = zip(*ytrain)
ytrain1 = tmp[0]
ytrain2 = tmp[1]
ytrain3 = tmp[2]
ytrain4 = tmp[3]
ytrain1 = np.array(ytrain1)
ytrain2 = np.array(ytrain2)
ytrain3 = np.array(ytrain3)
ytrain4 = np.array(ytrain4)
# clf=SVR(kernel='rbf')
clf = LinearRegression()
clf.fit(xtrain, ytrain1)
vals1 = clf.predict(xtest)
# print vals1
# print accuracyregression(ytest1,vals1)
# clf=SVR(kernel='linear',C=1e3)
clf.fit(xtrain, ytrain2)
vals2 = clf.predict(xtest)
# print accuracyregression(ytest2,vals2)
# clf=SVR(kernel='linear',C=1e3)
clf.fit(xtrain, ytrain3)
vals3 = clf.predict(xtest)
# print accuracyregression(ytest3,vals3)
# clf=SVR(kernel='linear',C=1e3)
clf.fit(xtrain, ytrain4)
vals4 = clf.predict(xtest)
# print accuracyregression(ytest4,vals4)
truth = np.zeros((4, len(ytest1)))
predicted = np.zeros((4, len(vals1)))
for j in range(len(ytest1)):
truth[0][j] = ytest1[j]
predicted[0][j] = vals1[j]
truth[1][j] = ytest2[j]
predicted[1][j] = vals2[j]
truth[2][j] = ytest3[j]
predicted[2][j] = vals3[j]
truth[3][j] = ytest4[j]
predicted[3][j] = vals4[j]
ylabels.append(accuracyclassification(truth, predicted))
# print accuracyclassification(truth,predicted)
print xlabels
print ylabels
plt.plot(xlabels, ylabels)
plt.savefig("classificationbasic1.png")
# Normalize the labels.
y = 1.0 * y
tmp_no = numpy.sum(y)
pno = (data_no + tmp_no) / 2
nno = (data_no - tmp_no) / 2
y[y > 0] = y[y > 0] / pno
y[y < 0] = y[y < 0] / nno
# Normalize the data.
m = X.mean(0)
s = X.std(0)
X.__isub__(m).__idiv__(s)
t1 = time.clock()
tmp = bahsic.BAHSICRaw(X, y, vector.CLinearKernel(),
vector.CLinearKernel(), features_tokeep, 0.1)
t2 = time.clock()
print "time taken: " + str(t2 - t1)
print '--rank of the features'
print '--better features towards the end of the list:'
print tmp
hsicfeatures = numpy.zeros(shape=(data_no, features_tokeep))
for i in range(0, data_no):
for j in range(0, features_tokeep):
hsicfeatures[i][j] = X[i][tmp[features_tokeep + j]]
numpy.savetxt(file_out, hsicfeatures)
if (sys.argv == 5):
numpy.savetxt('original.csv', X)