def __call__(self, data, gamma, mu_prior_mean, Phi0, nu0):
self.new_Sigma = np.zeros_like(self.Sigma)
self.new_mu = np.zeros((self.Sigma.shape[0], self.Sigma.shape[1]), dtype=np.float)
if isinstance(self.labels, list):
self.count = np.zeros((len(self.labels), len(self.comps)), dtype=np.int)
else:
self.count = np.zeros(len(self.comps), dtype=np.int)
jj = -1
for j in self.comps:
jj += 1
if isinstance(self.labels, list):
nobs = data.shape[0]
mask = np.zeros(nobs, dtype=np.bool)
cumobs = 0
ii = 0
for labs in self.labels:
submask = labs == j
mask[cumobs : (cumobs + len(labs))] = submask
self.count[ii, jj] = np.sum(submask)
cumobs += len(labs)
ii += 1
else:
mask = self.labels == j
self.count[jj] = np.sum(mask)
Xj = data[mask]
nj, ndim = Xj.shape
sumxj = Xj.sum(0)
gam = gamma[j]
mu_hyper = mu_prior_mean
post_mean = (mu_hyper / gam + sumxj) / (1 / gam + nj)
post_cov = 1 / (1 / gam + nj) * self.Sigma[jj]
new_mu = npr.multivariate_normal(post_mean, post_cov)
Xj_demeaned = Xj - new_mu
mu_SS = np.outer(new_mu - mu_hyper, new_mu - mu_hyper) / gam
data_SS = np.dot(Xj_demeaned.T, Xj_demeaned)
post_Phi = data_SS + mu_SS + Phi0[j]
# symmetrize
post_Phi = (post_Phi + post_Phi.T) / 2
# P(Sigma) ~ IW(nu + 2, nu * Phi)
# P(Sigma | theta, Y) ~
post_nu = nj + ndim + nu0 + 2
# pymc rinverse_wishart takes
new_Sigma = invwishartrand_prec(post_nu, post_Phi)
# store new results
self.new_Sigma[jj] = new_Sigma
self.new_mu[jj] = new_mu
del self.labels
del self.Sigma
def _set_initial_values(self, alpha0, nu0, Phi0, mu0, Sigma0, weights0, e0,
f0):
if nu0 is None:
nu0 = 1
if Phi0 is None:
Phi0 = np.empty((self.ncomp, self.ndim, self.ndim))
Phi0[:] = np.eye(self.ndim) * nu0
if Sigma0 is None:
# draw from prior .. bad idea for vague prior ???
Sigma0 = np.empty((self.ncomp, self.ndim, self.ndim))
for j in xrange(self.ncomp):
Sigma0[j] = invwishartrand_prec(nu0 + 1 + self.ndim, Phi0[j])
# starting values, are these sensible?
if mu0 is None:
mu0 = np.empty((self.ncomp, self.ndim))
for j in xrange(self.ncomp):
mu0[j] = npr.multivariate_normal(np.zeros(self.ndim),
self.gamma[j] * Sigma0[j])
if weights0 is None:
weights0 = (1 / self.ncomp) * np.ones((self.ncomp, 1))
self._alpha0 = alpha0
self.e = e0
self.f = f0
self._weights0 = weights0
self._mu0 = mu0
self._Sigma0 = Sigma0
self._nu0 = nu0 # prior degrees of freedom
self._Phi0 = Phi0 # prior location for Sigma_j's
def _set_initial_values(self, alpha0, nu0, Phi0, mu0, Sigma0, weights0,
e0, f0):
if nu0 is None:
nu0 = 1
if Phi0 is None:
Phi0 = np.empty((self.ncomp, self.ndim, self.ndim))
Phi0[:] = np.eye(self.ndim) * nu0
if Sigma0 is None:
# draw from prior .. bad idea for vague prior ???
Sigma0 = np.empty((self.ncomp, self.ndim, self.ndim))
for j in xrange(self.ncomp):
Sigma0[j] = invwishartrand_prec(nu0 + 1 + self.ndim, Phi0[j])
#Sigma0 = Phi0.copy()
# starting values, are these sensible?
if mu0 is None:
mu0 = np.empty((self.ncomp, self.ndim))
for j in xrange(self.ncomp):
mu0[j] = npr.multivariate_normal(np.zeros((self.ndim)),
self.gamma[j]*Sigma0[j])
if weights0 is None:
weights0 = (1/self.ncomp)*np.ones((self.ncomp, 1))
#_, weights0 = stick_break_proc(1, 1, size=self.ncomp - 1)
self._alpha0 = alpha0
self.e = e0
self.f = f0
self._weights0 = weights0
self._mu0 = mu0
self._Sigma0 = Sigma0
self._nu0 = nu0 # prior degrees of freedom
self._Phi0 = Phi0 # prior location for Sigma_j's
def __call__(self, data, gamma, mu_prior_mean, Phi0, nu0):
self.new_Sigma = np.zeros_like(self.Sigma)
self.new_mu = np.zeros((self.Sigma.shape[0], self.Sigma.shape[1]),
dtype=np.float)
if isinstance(self.labels, list):
self.count = np.zeros((len(self.labels), len(self.comps)),
dtype=np.int)
else:
self.count = np.zeros(len(self.comps), dtype=np.int)
jj = -1
for j in self.comps:
jj += 1
if isinstance(self.labels, list):
nobs = data.shape[0]
mask = np.zeros(nobs, dtype=np.bool)
cumobs = 0
ii = 0
for labs in self.labels:
submask = labs == j
mask[cumobs:(cumobs + len(labs))] = submask
self.count[ii, jj] = np.sum(submask)
cumobs += len(labs)
ii += 1
else:
mask = self.labels == j
self.count[jj] = np.sum(mask)
Xj = data[mask]
nj, ndim = Xj.shape
sumxj = Xj.sum(0)
gam = gamma[j]
mu_hyper = mu_prior_mean
post_mean = (mu_hyper / gam + sumxj) / (1 / gam + nj)
post_cov = 1 / (1 / gam + nj) * self.Sigma[jj]
new_mu = npr.multivariate_normal(post_mean, post_cov)
Xj_demeaned = Xj - new_mu
mu_SS = np.outer(new_mu - mu_hyper, new_mu - mu_hyper) / gam
data_SS = np.dot(Xj_demeaned.T, Xj_demeaned)
post_Phi = data_SS + mu_SS + Phi0[j]
# symmetrize
post_Phi = (post_Phi + post_Phi.T) / 2
# P(Sigma) ~ IW(nu + 2, nu * Phi)
# P(Sigma | theta, Y) ~
post_nu = nj + ndim + nu0 + 2
# pymc rinverse_wishart takes
new_Sigma = invwishartrand_prec(post_nu, post_Phi)
# store new results
self.new_Sigma[jj] = new_Sigma
self.new_mu[jj] = new_mu
del self.labels
del self.Sigma
N = 10000
nu = 6.0
phi = 2 * np.identity(4)
phi[1, 2] = 1
phi[2, 1] = 1
mu = np.zeros(4, dtype=np.float64)
Sigma = sampler.pyiwishartrand_prec(nu, phi)
res = np.zeros((4, 4))
t1 = time.time()
for i in range(N):
res += sampler.pyiwishartrand_prec(nu, phi)
print res / N
print time.time() - t1
t2 = time.time()
res = np.zeros((4, 4))
for i in range(N):
res += wishart.invwishartrand_prec(nu, phi)
print res / N
print time.time() - t2
t3 = time.time()
res = np.zeros(4)
for i in range(N):
s = sampler.pymvnorm(mu, Sigma)
res += s
print res / N
print time.time() - t3
def _update_mu_Sigma(self, Sigma, labels, other_dat=None):
is_hdp = isinstance(labels, list)
mu_output = np.zeros((self.ncomp, self.ndim))
Sigma_output = np.zeros((self.ncomp, self.ndim, self.ndim))
if is_hdp:
ct = np.zeros((len(labels), self.ncomp), dtype=np.int)
else:
ct = np.zeros(self.ncomp, dtype=np.int)
if other_dat is None:
data = self.data
else:
data = other_dat
if self.parallel:
num_jobs = len(self.compsdevmap)
for tid in self.compsdevmap:
rng = self.compsdevmap[tid]
self.work_queue[tid].put(
CompUpdate(np.arange(rng[0], rng[1]), labels,
Sigma[rng[0]:rng[1]]))
#while num_jobs:
for tid in self.compsdevmap:
result = self.result_queue[tid].get()
newcomps = result.comps
rng = (newcomps[0], newcomps[-1] + 1)
mu_output[rng[0]:rng[1]] = result.new_mu
Sigma_output[rng[0]:rng[1]] = result.new_Sigma
if is_hdp:
ct[:, rng[0]:rng[1]] = result.count
else:
ct[rng[0]:rng[1]] = result.count
num_jobs -= 1
else:
for j in xrange(self.ncomp):
if is_hdp:
nobs = data.shape[0]
mask = np.zeros(nobs, dtype=np.bool)
count = np.zeros(len(labels), dtype=np.int)
cumobs = 0
ii = 0
for labs in labels:
submask = labs == j
mask[cumobs:(cumobs + len(labs))] = submask
count[ii] = np.sum(submask)
cumobs += len(labs)
ii += 1
else:
mask = labels == j
count = np.sum(mask)
Xj = data[mask]
nj = len(Xj)
sumxj = Xj.sum(0)
gam = self.gamma[j]
mu_hyper = self.mu_prior_mean
post_mean = (mu_hyper / gam + sumxj) / (1 / gam + nj)
post_cov = 1 / (1 / gam + nj) * Sigma[j]
new_mu = npr.multivariate_normal(post_mean, post_cov)
Xj_demeaned = Xj - new_mu
mu_SS = np.outer(new_mu - mu_hyper, new_mu - mu_hyper) / gam
data_SS = np.dot(Xj_demeaned.T, Xj_demeaned)
post_Phi = data_SS + mu_SS + self._Phi0[j]
# symmetrize
post_Phi = (post_Phi + post_Phi.T) / 2
# P(Sigma) ~ IW(nu + 2, nu * Phi)
# P(Sigma | theta, Y) ~
post_nu = nj + self.ndim + self._nu0 + 2
new_Sigma = invwishartrand_prec(post_nu, post_Phi)
mu_output[j] = new_mu
Sigma_output[j] = new_Sigma
if is_hdp:
ct[:, j] = count
else:
ct[j] = count
return mu_output, Sigma_output, ct
if __name__ == '__main__':
N = 10000
nu = 6.0
phi = 2*np.identity(4)
phi[1,2]=1; phi[2,1]=1;
mu = np.zeros(4, dtype=np.float64)
Sigma = sampler.pyiwishartrand_prec(nu, phi)
res = np.zeros((4,4))
t1 = time.time()
for i in range(N):
res += sampler.pyiwishartrand_prec(nu, phi)
print res / N
print time.time() - t1
t2 = time.time()
res = np.zeros((4,4))
for i in range(N):
res += wishart.invwishartrand_prec(nu,phi)
print res / N
print time.time() - t2
t3 = time.time()
res = np.zeros(4)
for i in range(N):
s = sampler.pymvnorm(mu, Sigma)
res += s
print res / N
print time.time() - t3
def _update_mu_Sigma(self, Sigma, labels, other_dat=None):
is_hdp = isinstance(labels, list)
mu_output = np.zeros((self.ncomp, self.ndim))
Sigma_output = np.zeros((self.ncomp, self.ndim, self.ndim))
if is_hdp:
ct = np.zeros((len(labels), self.ncomp), dtype=np.int)
else:
ct = np.zeros(self.ncomp, dtype=np.int)
if other_dat is None:
data = self.data
else:
data = other_dat
if self.parallel:
num_jobs = len(self.compsdevmap)
for tid in self.compsdevmap:
rng = self.compsdevmap[tid]
self.work_queue[tid].put(CompUpdate(np.arange(rng[0], rng[1]), labels, Sigma[rng[0]:rng[1]]))
#while num_jobs:
for tid in self.compsdevmap:
result = self.result_queue[tid].get()
newcomps = result.comps
rng = (newcomps[0], newcomps[-1]+1)
mu_output[rng[0]:rng[1]] = result.new_mu
Sigma_output[rng[0]:rng[1]] = result.new_Sigma
if is_hdp:
ct[:,rng[0]:rng[1]] = result.count
else:
ct[rng[0]:rng[1]] = result.count
num_jobs -= 1
else:
for j in xrange(self.ncomp):
if is_hdp:
nobs = data.shape[0]
mask = np.zeros(nobs, dtype=np.bool)
count = np.zeros(len(labels), dtype=np.int)
cumobs = 0; ii = 0
for labs in labels:
submask = labs == j
mask[cumobs:(cumobs+len(labs))] = submask
count[ii] = np.sum(submask);
cumobs+=len(labs); ii+=1
else:
mask = labels == j
count = np.sum(mask)
Xj = data[mask]
nj = len(Xj)
sumxj = Xj.sum(0)
gam = self.gamma[j]
mu_hyper = self.mu_prior_mean
post_mean = (mu_hyper / gam + sumxj) / (1 / gam + nj)
post_cov = 1 / (1 / gam + nj) * Sigma[j]
new_mu = npr.multivariate_normal(post_mean, post_cov)
Xj_demeaned = Xj - new_mu
mu_SS = np.outer(new_mu - mu_hyper, new_mu - mu_hyper) / gam
data_SS = np.dot(Xj_demeaned.T, Xj_demeaned)
post_Phi = data_SS + mu_SS + self._Phi0[j]
# symmetrize
post_Phi = (post_Phi + post_Phi.T) / 2
# P(Sigma) ~ IW(nu + 2, nu * Phi)
# P(Sigma | theta, Y) ~
post_nu = nj + self.ndim + self._nu0 + 2
new_Sigma = invwishartrand_prec(post_nu, post_Phi)
mu_output[j] = new_mu
Sigma_output[j] = new_Sigma
if is_hdp:
ct[:, j] = count
else:
ct[j] = count
return mu_output, Sigma_output, ct
