class PreSmp(MainSmp):
def __init__(self, nadp, nsmp, gmdistr, rv_array,r_array, sl_array, gd_array, gcov_array):
self.nadp = nadp
MainSmp.__init__(self, nsmp, gmdistr, rv_array, r_array, sl_array, gd_array, gcov_array)
def setAdpNum(self, new_nadp):
self.nadp = new_nadp
def getAdpNum(self):
return self.nadp
def updateParameter(self, smps, responsibilities, gf2hv):
nV = self.gmdistr.means_.shape[1]
nCompo = self.gmdistr.get_params()['n_components']
# update weights
if nCompo == 1:
piV = np.array([1.0])
else:
if np.sum(gf2hv) == 0 or np.isnan(np.sum(gf2hv)) or np.isinf(np.sum(gf2hv)):
raise ValueError("[updating weights]: zero division")
piV = np.sum( np.dot(gf2hv[:, np.newaxis], np.ones((1,nCompo))) * responsibilities, axis=0 ) / np.sum(gf2hv)
piV = piV / np.sum(piV)
# update mean
#if nCompo == 1:
# rm_mean = self.r_array[0].stats(moments='m')
# rv_mean = self.r_array[1].stats(moments='m')
# muV = np.array([[rm_mean, rv_mean]])
#else:
muV = np.zeros(self.gmdistr.means_.shape)
for iCompo in range(nCompo):
tmps = gf2hv * responsibilities[:, iCompo]
if np.sum(tmps) == 0 or np.isnan(np.sum(tmps)) or np.isinf(np.sum(tmps)):
raise ValueError("[updating means]: zero division")
for iV in range(nV):
muV[iCompo, iV] = np.dot( smps[:, iV], tmps ) / np.sum(tmps)
# update covariance
covarV = np.zeros(self.gmdistr.covars_.shape)
muVold = self.gmdistr.means_
for iCompo in range(nCompo):
tmps = gf2hv * responsibilities[:, iCompo]
if np.sum(tmps) == 0 or np.isnan(np.sum(tmps)) or np.isinf(np.sum(tmps)):
raise ValueError("[updating covariance]: zero division")
for iV in range(nV):
for jV in range(nV):
smps_i = smps[:, iV]
smps_j = smps[:, jV]
covarV[iCompo, iV, jV] = np.dot( (smps_i-muVold[iCompo, iV]) * (smps_j-muVold[iCompo, jV]), tmps ) / np.sum(tmps)
self.gmdistr = GMM(n_components = nCompo, covariance_type='full')
self.gmdistr.weights_ = piV
self.gmdistr.means_ = muV
self.gmdistr.covars_ = covarV
def adaptation(self, ti):
kadp = np.ones(self.nadp)
kadp[0] = STD_RATIO1
kadp[1:3] = STD_RATIO2
kadp[3:5] = STD_RATIO3
sumX = 0; sumX2 = 0;
for iadp in range(self.nadp):
nCompo = self.gmdistr.get_params()['n_components']
tmp_gmdistr = GMM(n_components=nCompo,covariance_type='full')
tmp_gmdistr.weights_ = self.gmdistr.weights_
tmp_gmdistr.means_ = self.gmdistr.means_
tmp_gmdistr.covars_ = self.gmdistr.covars_ * kadp[iadp]**2
subSmp = MainSmp(self.nsmp, tmp_gmdistr, self.rv_array, self.r_array, self.sl_array, self.gd_array, self.gcov_array)
i = 0
while i<=MAX_NUM_UPDATE:
try:
smps = subSmp.sample()
f = subSmp.priorPDF(smps)
g = 1 - subSmp.condAvailability(smps, ti)
hv, responsibilities = subSmp.score_samples(smps)
gf2hv = g*f/hv
self.updateParameter(smps, responsibilities, gf2hv)
i += 1
break
except ValueError:
i += 1
continue
sumX += np.sum(gf2hv)
sumX2 += np.sum(gf2hv**2)
npre = self.nadp * self.nsmp
pfpre = 1./npre * sumX
Spf2pre = 1./(npre*(npre-1)) * (sumX2 - 2*sumX*pfpre + npre * pfpre**2)
return pfpre, Spf2pre
def getKopt(self, k_array, ti):
pf = np.zeros(k_array.shape)
for k in k_array:
nCompo = self.gmdistr.get_params()['n_components']
tmp_gmdistr = GMM(n_components=nCompo,covariance_type='full')
tmp_gmdistr.weights_ = self.gmdistr.weights_
tmp_gmdistr.means_ = self.gmdistr.means_
tmp_gmdistr.covars_ = self.gmdistr.covars_ * k**2
subSmp = MainSmp(self.nsmp, tmp_gmdistr, self.rv_array, self.r_array, self.sl_array, self.gd_array, self.gcov_array)
smps = subSmp.sample()
indx = np.where(k_array==k)[0][0]
pf[indx], dummy = subSmp.getPf(smps, ti)
if indx>0 and pf[indx] < KRATIO_CRITERIA * np.mean(pf[:indx]):
k_indx = indx-1
break
else:
k_indx = -1
return k_array[k_indx]
class PreSmp(MainSmp):
def __init__(self, nadp, nsmp, gmdistr, r_array, sl_array, gd_array):
self.nadp = nadp
MainSmp.__init__(self, nsmp, gmdistr, r_array, sl_array, gd_array)
def setAdpNum(self, new_nadp):
self.nadp = new_nadp
def getAdpNum(self):
return self.nadp
def updateParameter(self, smps, responsibilities, gf2hv):
nV = self.gmdistr.means_.shape[1]
nCompo = self.gmdistr.get_params()['n_components']
# update weights
if nCompo == 1:
piV = np.array([1.0])
else:
if np.sum(gf2hv) == 0 or np.isnan(np.sum(gf2hv)) or np.isinf(
np.sum(gf2hv)):
raise ValueError("[updating weights]: zero division")
piV = np.sum(np.dot(gf2hv[:, np.newaxis], np.ones(
(1, nCompo))) * responsibilities,
axis=0) / np.sum(gf2hv)
piV = piV / np.sum(piV)
# update mean
#if nCompo == 1:
# rm_mean = self.r_array[0].stats(moments='m')
# rv_mean = self.r_array[1].stats(moments='m')
# muV = np.array([[rm_mean, rv_mean]])
#else:
muV = np.zeros(self.gmdistr.means_.shape)
for iCompo in range(nCompo):
tmps = gf2hv * responsibilities[:, iCompo]
if np.sum(tmps) == 0 or np.isnan(np.sum(tmps)) or np.isinf(
np.sum(tmps)):
raise ValueError("[updating means]: zero division")
for iV in range(nV):
muV[iCompo, iV] = np.dot(smps[:, iV], tmps) / np.sum(tmps)
# update covariance
covarV = np.zeros(self.gmdistr.covars_.shape)
muVold = self.gmdistr.means_
for iCompo in range(nCompo):
tmps = gf2hv * responsibilities[:, iCompo]
if np.sum(tmps) == 0 or np.isnan(np.sum(tmps)) or np.isinf(
np.sum(tmps)):
raise ValueError("[updating covariance]: zero division")
for iV in range(nV):
for jV in range(nV):
smps_i = smps[:, iV]
smps_j = smps[:, jV]
covarV[iCompo, iV, jV] = np.dot(
(smps_i - muVold[iCompo, iV]) *
(smps_j - muVold[iCompo, jV]), tmps) / np.sum(tmps)
self.gmdistr = GMM(n_components=nCompo, covariance_type='full')
self.gmdistr.weights_ = piV
self.gmdistr.means_ = muV
self.gmdistr.covars_ = covarV
def adaptation(self, ti):
kadp = np.ones(self.nadp)
kadp[0] = STD_RATIO1
kadp[1:3] = STD_RATIO2
kadp[3:5] = STD_RATIO3
sumX = 0
sumX2 = 0
for iadp in range(self.nadp):
nCompo = self.gmdistr.get_params()['n_components']
tmp_gmdistr = GMM(n_components=nCompo, covariance_type='full')
tmp_gmdistr.weights_ = self.gmdistr.weights_
tmp_gmdistr.means_ = self.gmdistr.means_
tmp_gmdistr.covars_ = self.gmdistr.covars_ * kadp[iadp]**2
subSmp = MainSmp(self.nsmp, tmp_gmdistr, self.r_array,
self.sl_array, self.gd_array)
i = 0
while i <= MAX_NUM_UPDATE:
try:
smps = subSmp.sample()
f = subSmp.priorPDF(smps)
g = 1 - subSmp.condAvailability(smps, ti)
hv, responsibilities = subSmp.score_samples(smps)
gf2hv = g * f / hv
self.updateParameter(smps, responsibilities, gf2hv)
i += 1
break
except ValueError:
i += 1
continue
sumX += np.sum(gf2hv)
sumX2 += np.sum(gf2hv**2)
npre = self.nadp * self.nsmp
pfpre = 1. / npre * sumX
Spf2pre = 1. / (npre * (npre - 1)) * (sumX2 - 2 * sumX * pfpre +
npre * pfpre**2)
return pfpre, Spf2pre
def getKopt(self, k_array, ti):
pf = np.zeros(k_array.shape)
for k in k_array:
nCompo = self.gmdistr.get_params()['n_components']
tmp_gmdistr = GMM(n_components=nCompo, covariance_type='full')
tmp_gmdistr.weights_ = self.gmdistr.weights_
tmp_gmdistr.means_ = self.gmdistr.means_
tmp_gmdistr.covars_ = self.gmdistr.covars_ * k**2
subSmp = MainSmp(self.nsmp, tmp_gmdistr, self.r_array,
self.sl_array, self.gd_array)
smps = subSmp.sample()
indx = np.where(k_array == k)[0][0]
pf[indx], dummy = subSmp.getPf(smps, ti)
if indx > 0 and pf[indx] < KRATIO_CRITERIA * np.mean(pf[:indx]):
k_indx = indx - 1
break
else:
k_indx = -1
return k_array[k_indx]
