def sample_delta_i(self, delta, r, zeta, sigma, alpha, beta, xi, tau, eta,
i):
_delta, _zeta, _sigma = self.clean_delta(delta, zeta, sigma, i)
_dmax = _delta.max()
njs = cu.counter(_delta, _dmax + 1 + self.m)
ljs = njs + (njs == 0) * eta / self.m
_zeta_new, _sigma_new = self.sample_zeta_sigma_new(
alpha, beta, xi, tau, self.m)
zeta_stack = np.vstack((_zeta, _zeta_new))
sigma_stack = np.vstack((_sigma, _sigma_new))
assert (zeta_stack.shape[0] == ljs.shape[0])
args = zip(
repeat(r[i] * self.data.V[i]),
zeta_stack,
sigma_stack,
)
res = map(logdgamma_wrapper, args)
lps = np.array(list(res))
lps[np.where(np.isnan(lps))[0]] = -np.inf
lps -= lps.max()
unnormalized = np.exp(lps) * ljs
normalized = unnormalized / unnormalized.sum()
dnew = choice(range(zeta_stack.shape[0]), 1, p=normalized)
if dnew > _dmax:
zeta = np.vstack((_zeta, zeta_stack[dnew]))
sigma = np.vstack((_sigma, sigma_stack[dnew]))
delta = np.insert(_delta, i, _dmax + 1)
else:
delta = np.insert(_delta, i, dnew)
zeta = _zeta.copy()
sigma = _sigma.copy()
return delta, zeta, sigma
def sample_delta_i(self, delta, r, alpha, beta, eta, mu, Sigma_chol, i):
_delta, _alpha, _beta = self.clean_delta(delta, alpha, beta, i)
_dmax = _delta.max()
njs = cu.counter(_delta, _dmax + 1 + self.m)
ljs = njs + (njs == 0) * (eta / self.m)
alpha_new, beta_new = self.sample_alpha_beta_new(mu, Sigma_chol, self.m)
alpha_stack = np.vstack((_alpha, alpha_new))
beta_stack = np.vstack((_beta, beta_new))
assert (alpha_stack.shape[0] == ljs.shape[0])
Y = self.data.Yl[i] * r[i]
args = zip(repeat(Y), alpha_stack, beta_stack)
lps = np.array(list(map(log_density_delta_i, args))) * self.inv_temper_temp
lps[np.where(np.isnan(lps))] = - np.inf
lps -= lps.max()
unnormalized = np.exp(lps) * ljs
normalized = unnormalized / unnormalized.sum()
dnew = choice(range(_dmax + self.m + 1), 1, p = normalized)
if dnew > _dmax:
_alpha_ = np.vstack((_alpha, alpha_stack[dnew]))
_beta_ = np.vstack((_beta, beta_stack[dnew]))
_delta_ = np.insert(_delta, i, _dmax + 1)
else:
_delta_ = np.insert(_delta, i, dnew)
_alpha_ = _alpha.copy()
_beta_ = _beta.copy()
return _delta_, _alpha_, _beta_
def generate_posterior_predictive_simplex_3d(self, cols, m = 20):
keep_idx = np.array(sample(range(self.nSamp), self.nDat), dtype = int)
delta = self.samples.delta[keep_idx]
alpha = [self.samples.alpha[i][:,cols] for i in keep_idx]
beta = [self.samples.beta[i][:,cols] for i in keep_idx]
eta = self.samples.eta[keep_idx]
mu = self.samples.mu[keep_idx][:,cols]
Sigma = self.samples.Sigma[keep_idx][:,cols][:,:,cols]
postpred = np.empty((self.nDat, len(cols)))
for n in range(self.nDat):
dmax = delta[n].max()
njs = cu.counter(delta[n], dmax + 1 + m)
ljs = njs + (njs == 0) * eta[n] / m
new_log_alpha = mu[n].reshape(1,-1) + \
(cholesky(Sigma[n]) @ normal.rvs(size = (len(cols), m))).T
new_alpha = np.exp(new_log_alpha)
if cols[0] == 0:
new_beta = np.hstack((
np.ones((m, 1)),
gamma.rvs(a = 2., scale = 1/2., size = (m, len(cols) - 1))
))
else:
new_beta = gamma.rvs(a = 2., scale = 1/2, size = (m, len(cols)))
prob = ljs / ljs.sum()
a = np.vstack((alpha[n], new_alpha))
b = np.vstack((beta[n], new_beta))
delta_new = np.random.choice(range(dmax + 1 + m), 1, p = prob)
postpred[n] = gamma.rvs(a = a[delta_new], scale = 1 / b[delta_new])
np.nan_to_num(postpred, copy = False)
return (postpred.T / postpred.sum(axis = 1)).T
def generate_posterior_predictive_gammas(self, n_per_sample=1, m=10):
new_gammas = []
for s in range(self.nSamp):
dmax = self.samples.delta[s].max()
njs = cu.counter(self.samples.delta[s], int(dmax + 1 + m))
ljs = njs + (njs == 0) * self.samples.eta[s] / m
new_zetas = gamma(
shape=self.samples.alpha[s],
scale=1. / self.samples.beta[s],
size=(m, self.nCol),
)
new_sigmas = np.hstack((
np.ones((m, 1)),
gamma(
shape=self.samples.xi[s],
scale=self.samples.tau[s],
size=(m, self.nCol - 1),
),
))
prob = ljs / ljs.sum()
deltas = cu.generate_indices(prob, n_per_sample)
zeta = np.vstack((self.samples.zeta[s], new_zetas))[deltas]
sigma = np.vstack((self.samples.sigma[s], new_sigmas))[deltas]
new_gammas.append(gamma(shape=zeta, scale=1 / sigma))
return np.vstack(new_gammas)
def sample_delta_i(self, delta, zeta, r, alpha, beta, eta, i):
_delta, _zeta = self.clean_delta(delta, zeta, i)
_dmax = _delta.max()
njs = cu.counter(_delta, _dmax + 1 + self.m)
ljs = njs + (njs == 0) * eta / self.m
_zeta_new = self.sample_zeta_new(self.m, alpha, beta)
zeta_stack = np.vstack((_zeta, _zeta_new))
assert (zeta_stack.shape[0] == ljs.shape[0])
args = zip(
repeat(r[i] * self.data.Yl[i]),
zeta_stack,
)
# res = self.pool.map(log_density_delta_i, args, chunksize = ceil(ljs.shape[0] / 8.))
res = map(log_density_delta_i, args)
lps = np.array(list(res))
lps[np.where(np.isnan(lps))[0]] = -np.inf
unnormalized = np.exp(lps) * ljs
normalized = unnormalized / unnormalized.sum()
dnew = choice(range(_dmax + 1 + self.m), 1, p=normalized)
if dnew > _dmax:
zeta = np.vstack((_zeta, zeta_stack[dnew]))
delta = np.insert(_delta, i, _dmax + 1)
else:
delta = np.insert(_delta, i, dnew)
zeta = _zeta.copy()
return delta, zeta
def sample_delta_i(self, deltas, mus, Sigmas, mu0, Sigma0, eta, i):
_deltas, _mus, _Sigmas = self.clean_delta(deltas, mus, Sigmas, i)
_dmax = _deltas.max()
njs = cu.counter(_deltas, _dmax + 1 + self.m)
ljs = njs + (njs == 0) * eta / self.m
mus_new, Sigmas_new = self.sample_mu_Sigma_new(mu0, Sigma0, self.m)
mu_stack = np.vstack((_mus, mus_new))
Sigma_stack = np.vstack((_Sigmas, Sigmas_new))
assert (mu_stack.shape[0] == ljs.shape[0])
args = zip(repeat(self.data.pVi[i]), mu_stack, Sigma_stack)
res = self.pool.map(log_density_mvnormal,
args,
chunksize=ceil(ljs.shape[0] / 8))
lps = np.array(list(res))
lps[np.where(np.isnan(lps))] = -np.inf
lps -= lps.max()
unnormalized = np.exp(lps) * ljs
normalized = unnormalized / unnormalized.sum()
dnew = choice(range(_dmax + self.m + 1), 1, p=normalized)
if dnew > _dmax:
mu = np.vstack((_mus, mu_stack[dnew]))
Sigma = np.vstack((_Sigmas, Sigma_stack[dnew]))
delta = np.insert(_deltas, i, _dmax + 1)
else:
delta = np.insert(_deltas, i, dnew)
mu = _mus.copy()
Sigma = _Sigmas.copy()
return delta, mu, Sigma
def clean_delta(self, delta, alpha, i):
assert (delta.max() + 1 == alpha.shape[0])
_delta = np.delete(delta, i)
nj = cu.counter(_delta, _delta.max() + 2)
fz = cu.first_zero(nj)
_alpha = alpha[np.where(nj > 0)[0]]
if (fz == delta[i]) and (fz <= _delta.max()):
_delta[_delta > fz] = _delta[_delta > fz] - 1
return _delta, _alpha
def clean_delta(self, deltas, mus, Sigmas, i):
assert (deltas.max() + 1 == mus.shape[0])
_deltas = np.delete(deltas, i)
nj = cu.counter(_deltas, _deltas.max() + 2)
fz = cu.first_zero(nj)
_mus = mus[np.where(nj > 0)[0]]
_Sigmas = Sigmas[np.where(nj > 0)[0]]
if (fz == deltas[i]) and (fz <= _deltas.max()):
_deltas[_deltas > fz] = _deltas[_deltas > fz] - 1
return _deltas, _mus, _Sigmas
def generate_posterior_predictive_gammas(self, n_per_sample = 10, m = 20):
new_gammas = []
for i in range(self.nSamp):
dmax = self.samples.delta[i].max()
njs = cu.counter(self.samples.delta[i], dmax + 1 + m)
ljs = njs + (njs == 0) * self.samples.eta[i] / m
# This part needs to be fixed.
new_alphas = self.generate_alpha_new(self.samples.mu[i], self.samples.Sigma[i], m)
prob = ljs / ljs.sum()
deltas = cu.generate_indices(prob, n_per_sample)
alpha = np.vstack((self.samples.alpha[i], new_alphas))[deltas]
new_gammas.append(gamma(shape = alpha))
new_gamma_arr = np.vstack(new_gammas)
return new_gamma_arr
def generate_posterior_predictive_gammas(self, n_per_sample=10, m=20):
new_gammas = []
for s in range(self.nSamp):
dmax = self.samples.delta[s].max()
njs = cu.counter(self.samples.delta[s], int(dmax + 1 + m))
ljs = njs + (njs == 0) * self.samples.eta[s] / m
gnew = binomial(size=(m, self.nCol), n=1, p=self.samples.rho[s])
anew = self.samples.alpha[s][0] * (
1 - gnew) + self.samples.alpha[s][1] * gnew
bnew = self.samples.beta[s][0] * (
1 - gnew) + self.samples.beta[s][1] * gnew
new_zetas = gamma(shape=anew, scale=1. / bnew, size=(m, self.nCol))
prob = ljs / ljs.sum()
deltas = cu.generate_indices(prob, n_per_sample)
zeta = np.vstack((self.samples.zeta[s], new_zetas))[deltas]
new_gammas.append(gamma(shape=zeta))
return np.vstack(new_gammas)
def generate_posterior_predictive_probit(self, n_per_sample, m=20):
new_pVis = []
for i in range(self.nSamp):
dmax = self.samples.delta[i].max()
# njs = cu.counter(self.samples.delta[i], dmax + 1 + m)
# ljs = njs + (njs == 0) * self.samples.eta[i] / m
njs = cu.counter(self.samples.delta[i], dmax + 1)
ljs = njs
# new_mus = self.samples.mu0[i] + \
# self.samples.Sigma0[i] @ normal.rvs(size = (m, self.nCol))
# new_Sigmas =
prob = ljs / ljs.sum()
deltas = cu.generate_indices(prob, n_per_sample)
mus = self.samples.mu[i][deltas]
Sigmas = self.samples.Sigma[i][deltas]
for j in range(n_per_sample):
new_pVis.append(
mus[j] + cholesky(Sigmas[j]) @ normal(size=self.nCol - 1))
return self.invprobit(np.vstack(new_pVis))
def generate_posterior_predictive_simplex_3d(self, cols, m = 20):
keep_idx = np.array(sample(range(self.nSamp), self.nDat), dtype = int)
delta = self.samples.delta[keep_idx]
alpha = [self.samples.alpha[i][:,cols] for i in keep_idx]
beta = [self.samples.beta[i][:,cols] for i in keep_idx]
eta = self.samples.eta[keep_idx]
alpha_shape = self.samples.alpha_shape[keep_idx][:,cols]
if cols[0] == 0:
# beta_shape = np.array(
# [1, self.samples.beta_shape[keep_idx][:,(np.array(cols[1:], dtype = int) - 1)]]
# )
beta_shape = np.hstack((
np.ones((alpha_shape.shape[0],1)),
self.samples.beta_shape[keep_idx][:,(np.array(cols[1:], dtype = int) - 1)]
))
else:
beta_shape = self.samples.beta_shape[keep_idx][:,(np.array(cols, dtype = int) - 1)]
postpred = np.empty((self.nDat, len(cols)))
for n in range(self.nDat):
dmax = delta[n].max()
njs = cu.counter(delta[n], dmax + 1 + m)
ljs = njs + (njs == 0) * eta[n] / m
new_alphas = gamma.rvs(a = alpha_shape[n], size = (m, len(cols)))
if cols[0] == 0:
new_betas = np.hstack((
np.ones((m, 1)),
gamma.rvs(a = beta_shape[n,1:], size = (m, len(cols) - 1))
))
else:
new_betas = gamma.rvs(a = beta_shape[n], size = (m, len(cols)))
prob = ljs / ljs.sum()
a = np.vstack((alpha[n], new_alphas))
b = np.vstack((beta[n], new_betas))
new_delta = np.random.choice(range(dmax + 1 + m), 1, p = prob)
postpred[n] = gamma.rvs(a = a[new_delta], scale = 1/b[new_delta])
np.nan_to_num(postpred, copy = False)
return (postpred.T / postpred.sum(axis = 1)).T
def sample_pi(self, delta):
shapes = cu.counter(delta, self.nMix) + self.priors.pi.a
unnormalized = gamma(shape = shapes)
return unnormalized / unnormalized.sum()