def make_2d_data(n_examples, noise="uniform"):
cov = [[2, 1], [1, 2]]
dist = 100
x1 = mvnormal([-dist, 0], cov, n_examples / 4)
x2 = mvnormal([0, dist], cov, n_examples / 4)
x3 = mvnormal([dist, 0], cov, n_examples / 4)
x4 = mvnormal([0, -dist], cov, n_examples / 4)
x = np.concatenate([x1, x2, x3, x4])
if noise == "uniform":
x += np.random.uniform(-4, 4, size=x.shape)
else:
x += np.random.normal(0, 3, size=x.shape)
y = np.zeros((n_examples, 4))
y[0 : n_examples / 4] = 1
y[n_examples / 4 : n_examples / 2, 1] = 1
y[n_examples / 2 : 3 * n_examples / 4, 2] = 1
y[3 * n_examples / 4 :, 3] = 1
idx = np.arange(len(x))
perm = np.random.permutation(idx)
x = x[perm]
y = y[perm]
return x, y
def make_2d_data(n_examples, noise='uniform'):
cov = [[2, 1], [1, 2]]
dist = 100
x1 = mvnormal([-dist, 0], cov, n_examples / 4)
x2 = mvnormal([0, dist], cov, n_examples / 4)
x3 = mvnormal([dist, 0], cov, n_examples / 4)
x4 = mvnormal([0, -dist], cov, n_examples / 4)
x = np.concatenate([x1, x2, x3, x4])
if noise == 'uniform':
x += np.random.uniform(-4, 4, size=x.shape)
else:
x += np.random.normal(0, 3, size=x.shape)
y = np.zeros((n_examples, 4))
y[0:n_examples / 4] = 1
y[n_examples / 4:n_examples / 2, 1] = 1
y[n_examples / 2:3 * n_examples / 4, 2] = 1
y[3 * n_examples / 4:, 3] = 1
idx = np.arange(len(x))
perm = np.random.permutation(idx)
x = x[perm]
y = y[perm]
return x, y
def reset(self):
"""
Reset priors and draw parameter estimates from prior.
"""
# priors
self.lbd_phi0 = self.h["lbd_phi0"]
self.alpha_s20 = self.h["alpha_s20"]
self.beta_s20 = self.h["beta_s20"]
self.alpha_w = self.h["alpha_w0"]
self.beta_w = self.h["beta_w0"]
self.sigma_phi0 = eye(self.pdata) * self.h["lbd_phi0"]
self.sigma_phi0_inv = eye(self.pdata) / self.h["lbd_phi0"]
self.mu_phi0 = ones(self.pdata) * self.h["mu_phi0"]
# initial parameter estimates drawn from prior
self.p = dict()
self.p["sigma2_1"] = 1.0 / gamma(self.alpha_s20 , (1.0 / self.beta_s20) ) # inverse gamma
self.p["phi_1"] = mvnormal(self.mu_phi0 , self.p["sigma2_1"] * self.sigma_phi0)
self.p["sigma2_2"] = 1.0 / gamma(self.alpha_s20 , (1.0 / self.beta_s20)) # inverse gamma
self.p["phi_2"] = mvnormal(self.mu_phi0, self.p["sigma2_2"] * self.sigma_phi0)
self.p["w"] = beta(self.alpha_w, self.beta_w) # beta distribution
def reset(self):
"""
Reset priors and draw parameter estimates from prior.
"""
# priors
self.lbd_phi0 = self.h["lbd_phi0"] #lambda0 = 1
self.alpha_s20 = self.h["alpha_s20"] #alpha(sigma^2) = 5
self.beta_s20 = self.h["beta_s20"] #beta(sigma^2) = 1
self.sigma_phi0 = eye(self.pdata) * self.h["lbd_phi0"] #covariate matrix of sigma square
self.sigma_phi0_inv = eye(self.pdata) / self.h["lbd_phi0"] #inverse
self.mu_phi0 = ones(self.pdata) * self.h["mu_phi0"] #0
self.alpha_w0 = self.h["alpha_w0"]
self.beta_w0 = self.h["beta_w0"]
# initial parameter estimates drawn from prior
self.p = dict()
self.p["sigma_square_1"] = 1.0 / gamma(self.alpha_s20, 1.0 / self.beta_s20)
self.p["sigma_square_2"] = 1.0 / gamma(self.alpha_s20, 1.0 / self.beta_s20)
self.p["phi_1"] = mvnormal(self.mu_phi0, self.p["sigma_square_1"] * self.sigma_phi0)
self.p["phi_2"] = mvnormal(self.mu_phi0, self.p["sigma_square_2"] * self.sigma_phi0)
self.p["w"] = beta(self.alpha_w0, self.beta_w0)
self.p["z"] = binomial(1, self.p["w"], self.ndata)
def reset(self):
"""
Reset priors and draw parameter estimates from prior.
"""
# priors
self.lbd_phi0 = self.h["lbd_phi0"]
self.alpha_s20 = self.h["alpha_s20"]
self.beta_s20 = self.h["beta_s20"]
self.sigma_phi0 = eye(self.pdata) * self.h["lbd_phi0"]
self.sigma_phi0_inv = eye(self.pdata) / self.h["lbd_phi0"]
self.mu_phi0 = ones(self.pdata) * self.h["mu_phi0"]
# initial parameter estimates drawn from prior
self.p = dict()
self.p["sigma2"] = 1.0 / gamma(self.alpha_s20, 1.0 / self.beta_s20) # inverse gamma
self.p["phi"] = mvnormal(self.mu_phi0, self.p["sigma2"] * self.sigma_phi0)
def reset(self):
"""
Reset priors and draw parameter estimates from prior.
"""
# priors
self.lbd_phi0 = self.h["lbd_phi0"] # lambda
self.alpha_s20 = self.h["alpha_s20"] # alpha sigma sq
self.beta_s20 = self.h["beta_s20"] # beta sigma sq
self.sigma_phi0 = eye(self.pdata) * self.h["lbd_phi0"]
self.sigma_phi0_inv = eye(self.pdata) / self.h["lbd_phi0"]
self.mu_phi0 = ones(self.pdata) * self.h["mu_phi0"]
# initial parameter estimates drawn from prior
self.p = dict()
self.p["sigma2"] = 1.0 / gamma(
self.alpha_s20, 1.0 /
self.beta_s20) # inverse gamma # possible numerical problems ?
self.p["phi"] = mvnormal(self.mu_phi0,
self.p["sigma2"] * self.sigma_phi0)
def generate(self):
return mvnormal(self.means, self.cov, self.size)
'''
def generate(self):
return mvnormal(self.means, self.cov, self.size)
'''
