def BAHSICRaw(self, x, y, kernelx, kernely, flg3, flg4):
assert len(x.shape) == 2, 'Argument 1 has wrong shape'
assert len(y.shape) == 2, 'Argument 2 has wrong shape'
assert x.shape[0] == y.shape[0], \
'Argument 1 and 2 have different number of data points'
print '--initializing...'
hsic = CHSIC()
L = kernely.Dot(y, y)
setdiag0(L)
sL = numpy.sum(L, axis=1)
ssL = numpy.sum(sL)
n = x.shape
eliminatedI = []
selectedI = set(numpy.arange(n[1]))
kernelx.CreateCacheKernel(x)
while True:
selectedI = selectedI - set(eliminatedI)
sI = numpy.array([j for j in selectedI])
m = len(sI)
print m
if (m == 1):
eliminatedI.append(selectedI.pop())
break
objj = []
for j in selectedI:
K = kernelx.DecDotCacheKernel(x, x[:,[j]])
setdiag0(K)
objj.append(hsic.UnBiasedHSICFast(K, L, sL, ssL))
if m > flg3:
maxj = numpy.argsort(objj)
num = int(flg4 * m)+1
if m-num <= flg3:
num = m - flg3
maxj = maxj[m:m-num-1:-1]
else:
maxj = numpy.array([numpy.argmax(objj)])
j = numpy.take(sI,maxj)
eliminatedI.extend(j)
kernelx.DecCacheKernel(x, x[:,j])
kernelx.ClearCacheKernel(x)
return eliminatedI
def BAHSICRaw(self, x, y, kernelx, kernely, flg3, flg4):
assert len(x.shape) == 2, 'Argument 1 has wrong shape'
assert len(y.shape) == 2, 'Argument 2 has wrong shape'
assert x.shape[0] == y.shape[0], \
'Argument 1 and 2 have different number of data points'
print '--initializing...'
hsic = CHSIC()
L = kernely.Dot(y, y)
setdiag0(L)
sL = numpy.sum(L, axis=1)
ssL = numpy.sum(sL)
n = x.shape
eliminatedI = []
selectedI = set(numpy.arange(n[1]))
kernelx.CreateCacheKernel(x)
while True:
selectedI = selectedI - set(eliminatedI)
sI = numpy.array([j for j in selectedI])
m = len(sI)
print m
if (m == 1):
eliminatedI.append(selectedI.pop())
break
objj = []
for j in selectedI:
K = kernelx.DecDotCacheKernel(x, x[:, [j]])
setdiag0(K)
objj.append(hsic.UnBiasedHSICFast(K, L, sL, ssL))
if m > flg3:
maxj = numpy.argsort(objj)
num = int(flg4 * m) + 1
if m - num <= flg3:
num = m - flg3
maxj = maxj[m:m - num - 1:-1]
else:
maxj = numpy.array([numpy.argmax(objj)])
j = numpy.take(sI, maxj)
eliminatedI.extend(j)
kernelx.DecCacheKernel(x, x[:, j])
kernelx.ClearCacheKernel(x)
return eliminatedI
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 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 BAHSICOpt(self, x, y, kernelx, kernely, flg3, flg4):
assert len(x.shape) == 2, 'Argument 1 has wrong shape'
assert len(y.shape) == 2, 'Argument 2 has wrong shape'
assert x.shape[0] == y.shape[0], \
'Argument 1 and 2 have different number of data points'
print '--initializing...'
hsic = CHSIC()
L = kernely.Dot(y, y)
setdiag0(L)
sL = numpy.sum(L, axis=1)
ssL = numpy.sum(sL)
n = x.shape
eliminatedI = []
selectedI = set(numpy.arange(n[1]))
kernelx.CreateCacheKernel(x)
sga = kernelx._typicalParam
sgaN = sga.shape
sgaN = sgaN[0]
while True:
selectedI = selectedI - set(eliminatedI)
sI = numpy.array([j for j in selectedI])
m = len(sI)
print m
if (m == 1):
eliminatedI.append(selectedI.pop())
break
sgaMat = []
hsicMat = []
for k in range(sgaN):
## bfgs in scipy is not working here
retval = optimize.fmin_cg(hsic.ObjUnBiasedHSIC, \
sga[[k],].ravel(), \
hsic.GradUnBiasedHSIC,\
args=[x, kernelx, L, sL, ssL], \
gtol=1e-6, maxiter=100, \
full_output=True, disp=False)
sgaMat.append(retval[0])
hsicMat.append(retval[1])
k = numpy.argmin(hsicMat)
sga0 = sgaMat[k]
objj = []
for j in selectedI:
K = kernelx.DecDotCacheKernel(x, x[:,[j]], sga0)
setdiag0(K)
objj.append(hsic.UnBiasedHSICFast(K, L, sL, ssL))
if m > flg3:
maxj = numpy.argsort(objj)
num = int(flg4 * m)+1
if m - num <= flg3:
num = m - flg3
maxj = maxj[m:m-num-1:-1]
else:
maxj = numpy.array([numpy.argmax(objj)])
j = numpy.take(sI,maxj)
eliminatedI.extend(j)
kernelx.DecCacheKernel(x, x[:,j])
kernelx.ClearCacheKernel(x)
return eliminatedI
def BAHSICOpt(self, x, y, kernelx, kernely, flg3, flg4):
assert len(x.shape) == 2, 'Argument 1 has wrong shape'
assert len(y.shape) == 2, 'Argument 2 has wrong shape'
assert x.shape[0] == y.shape[0], \
'Argument 1 and 2 have different number of data points'
print '--initializing...'
hsic = CHSIC()
L = kernely.Dot(y, y)
setdiag0(L)
sL = numpy.sum(L, axis=1)
ssL = numpy.sum(sL)
n = x.shape
eliminatedI = []
selectedI = set(numpy.arange(n[1]))
kernelx.CreateCacheKernel(x)
sga = kernelx._typicalParam
sgaN = sga.shape
sgaN = sgaN[0]
while True:
selectedI = selectedI - set(eliminatedI)
sI = numpy.array([j for j in selectedI])
m = len(sI)
print m
if (m == 1):
eliminatedI.append(selectedI.pop())
break
sgaMat = []
hsicMat = []
for k in range(sgaN):
## bfgs in scipy is not working here
retval = optimize.fmin_cg(hsic.ObjUnBiasedHSIC, \
sga[[k],].ravel(), \
hsic.GradUnBiasedHSIC,\
args=[x, kernelx, L, sL, ssL], \
gtol=1e-6, maxiter=100, \
full_output=True, disp=False)
sgaMat.append(retval[0])
hsicMat.append(retval[1])
k = numpy.argmin(hsicMat)
sga0 = sgaMat[k]
objj = []
for j in selectedI:
K = kernelx.DecDotCacheKernel(x, x[:, [j]], sga0)
setdiag0(K)
objj.append(hsic.UnBiasedHSICFast(K, L, sL, ssL))
if m > flg3:
maxj = numpy.argsort(objj)
num = int(flg4 * m) + 1
if m - num <= flg3:
num = m - flg3
maxj = maxj[m:m - num - 1:-1]
else:
maxj = numpy.array([numpy.argmax(objj)])
j = numpy.take(sI, maxj)
eliminatedI.extend(j)
kernelx.DecCacheKernel(x, x[:, j])
kernelx.ClearCacheKernel(x)
return eliminatedI
