def init_kmean(self, init_data, niter=5):
""" Init the model using kmean."""
assert init_data.ndim == 1
k = self.gm.k
w = N.ones(k) / k
# Init the internal state of EM
self.cx = w * mean(init_data)
self.cxx = w * mean(init_data**2)
# w, mu and va init is the same that in the standard case
(code, label) = kmean(init_data[:, N.newaxis], \
init_data[0:k, N.newaxis], iter = niter)
mu = code.copy()
va = N.zeros((k, 1))
for i in range(k):
va[i] = N.cov(init_data[N.where(label == i)], rowvar=0)
self.gm.set_param(w, mu, va)
# c* are the parameters which are computed at every step (ie
# when a new frame is taken into account
self.cw = self.gm.w
self.cmu = self.gm.mu[:, 0]
self.cva = self.gm.va[:, 0]
# p* are the parameters used when computing gaussian densities
# they are the same than c* in the online case
# self.pw = self.cw.copy()
# self.pmu = self.cmu.copy()
# self.pva = self.cva.copy()
self.pw = self.cw
self.pmu = self.cmu
self.pva = self.cva
def init_kmean(self, data, niter=5):
""" Init the model with kmean."""
k = self.gm.k
d = self.gm.d
#init = data[0:k, :]
# XXX: This is bogus initialization should do better (in kmean with CV)
(code, label) = kmean(data, k, niter, minit='random')
w = N.ones(k) / k
mu = code.copy()
if self.gm.mode == 'diag':
va = N.zeros((k, d))
for i in range(k):
for j in range(d):
va[i, j] = N.cov(data[N.where(label == i), j], rowvar=0)
elif self.gm.mode == 'full':
va = N.zeros((k * d, d))
for i in range(k):
va[i*d:i*d+d, :] = \
N.cov(data[N.where(label==i)], rowvar = 0)
else:
raise GmmParamError("mode " + str(self.gm.mode) + \
" not recognized")
self.gm.set_param(w, mu, va)
self.isinit = True
def init_kmean(self, data, niter = 5):
""" Init the model with kmean."""
k = self.gm.k
d = self.gm.d
init = data[0:k, :]
# XXX: This is bogus initialization should do better (in kmean with CV)
(code, label) = kmean(data, init, niter, minit = 'matrix')
w = N.ones(k) / k
mu = code.copy()
if self.gm.mode == 'diag':
va = N.zeros((k, d))
for i in range(k):
for j in range(d):
va[i, j] = N.cov(data[N.where(label==i), j], rowvar = 0)
elif self.gm.mode == 'full':
va = N.zeros((k*d, d))
for i in range(k):
va[i*d:i*d+d, :] = \
N.cov(data[N.where(label==i)], rowvar = 0)
else:
raise GmmParamError("mode " + str(self.gm.mode) + \
" not recognized")
self.gm.set_param(w, mu, va)
self.isinit = True
def init_kmean(self, init_data, niter = 5):
""" Init the model using kmean."""
assert init_data.ndim == 1
k = self.gm.k
w = N.ones(k) / k
# Init the internal state of EM
self.cx = w * mean(init_data)
self.cxx = w * mean(init_data ** 2)
# w, mu and va init is the same that in the standard case
(code, label) = kmean(init_data[:, N.newaxis], \
init_data[0:k, N.newaxis], iter = niter)
mu = code.copy()
va = N.zeros((k, 1))
for i in range(k):
va[i] = N.cov(init_data[N.where(label==i)], rowvar = 0)
self.gm.set_param(w, mu, va)
# c* are the parameters which are computed at every step (ie
# when a new frame is taken into account
self.cw = self.gm.w
self.cmu = self.gm.mu[:, 0]
self.cva = self.gm.va[:, 0]
# p* are the parameters used when computing gaussian densities
# they are the same than c* in the online case
# self.pw = self.cw.copy()
# self.pmu = self.cmu.copy()
# self.pva = self.cva.copy()
self.pw = self.cw
self.pmu = self.cmu
self.pva = self.cva
def init_kmean(self, init_data, niter=5):
""" Init the model using kmean."""
k = self.gm.k
d = self.gm.d
if self.gm.mode == 'diag':
w = N.ones(k) / k
# Init the internal state of EM
self.cx = N.outer(w, mean(init_data, 0))
self.cxx = N.outer(w, mean(init_data**2, 0))
# w, mu and va init is the same that in the standard case
(code, label) = kmean(init_data,
init_data[0:k, :],
iter=niter,
minit='matrix')
mu = code.copy()
va = N.zeros((k, d))
for i in range(k):
for j in range(d):
va[i, j] = N.cov(init_data[N.where(label == i), j],
rowvar=0)
else:
raise OnGmmParamError(
"""init_online not implemented for
mode %s yet""", self.gm.mode)
self.gm.set_param(w, mu, va)
# c* are the parameters which are computed at every step (ie
# when a new frame is taken into account
self.cw = self.gm.w
self.cmu = self.gm.mu
self.cva = self.gm.va
# p* are the parameters used when computing gaussian densities
# they are the same than c* in the online case
# self.pw = self.cw.copy()
# self.pmu = self.cmu.copy()
# self.pva = self.cva.copy()
self.pw = self.cw
self.pmu = self.cmu
self.pva = self.cva
def init_kmean(self, init_data, niter = 5):
""" Init the model using kmean."""
k = self.gm.k
d = self.gm.d
if self.gm.mode == 'diag':
w = N.ones(k) / k
# Init the internal state of EM
self.cx = N.outer(w, mean(init_data, 0))
self.cxx = N.outer(w, mean(init_data ** 2, 0))
# w, mu and va init is the same that in the standard case
(code, label) = kmean(init_data, init_data[0:k, :],
iter = niter, minit = 'matrix')
mu = code.copy()
va = N.zeros((k, d))
for i in range(k):
for j in range(d):
va[i, j] = N.cov(init_data[N.where(label==i), j],
rowvar = 0)
else:
raise OnGmmParamError("""init_online not implemented for
mode %s yet""", self.gm.mode)
self.gm.set_param(w, mu, va)
# c* are the parameters which are computed at every step (ie
# when a new frame is taken into account
self.cw = self.gm.w
self.cmu = self.gm.mu
self.cva = self.gm.va
# p* are the parameters used when computing gaussian densities
# they are the same than c* in the online case
# self.pw = self.cw.copy()
# self.pmu = self.cmu.copy()
# self.pva = self.cva.copy()
self.pw = self.cw
self.pmu = self.cmu
self.pva = self.cva
