def add_piecewise_mpe_support():
add_node_mpe(PiecewiseLinear, piecewise_bottom_up_ll, piecewise_top_down)
def add_cltree_mpe_support():
add_node_mpe(CLTree, cltree_bottom_up_log_ll, cltree_top_down)
def add_parametric_mpe_support():
def gaussian_mode(node):
return node.mean
add_node_mpe(
Gaussian,
get_parametric_bottom_up_log_ll(continuous_log_likelihood, gaussian_mode),
get_parametric_top_down_ll(gaussian_mode),
)
def gamma_mode(node):
return (node.alpha - 1) / node.beta
add_node_mpe(
Gamma, get_parametric_bottom_up_log_ll(gamma_log_likelihood, gamma_mode), get_parametric_top_down_ll(gamma_mode)
)
def lognormal_mode(node):
return np.exp(node.mean - node.variance)
add_node_mpe(
LogNormal,
get_parametric_bottom_up_log_ll(continuous_log_likelihood, lognormal_mode),
get_parametric_top_down_ll(lognormal_mode),
)
def poisson_mode(node):
return np.floor(node.mean)
add_node_mpe(
Poisson,
get_parametric_bottom_up_log_ll(discrete_log_likelihood, poisson_mode),
get_parametric_top_down_ll(poisson_mode),
)
def bernoulli_mode(node):
if node.p > 0.5:
return 1
else:
return 0
add_node_mpe(
Bernoulli,
get_parametric_bottom_up_log_ll(discrete_log_likelihood, bernoulli_mode),
get_parametric_top_down_ll(bernoulli_mode),
)
def categorical_mode(node):
return np.argmax(node.p)
add_node_mpe(
Categorical,
get_parametric_bottom_up_log_ll(categorical_log_likelihood, categorical_mode),
get_parametric_top_down_ll(categorical_mode),
)
def geometric_mode(node):
return 1
add_node_mpe(
Geometric,
get_parametric_bottom_up_log_ll(discrete_log_likelihood, geometric_mode),
get_parametric_top_down_ll(geometric_mode),
)
def negative_binomial_mode(node):
if node.n <= 1:
return 0
else:
return np.floor(node.p * (node.n - 1) / (1 - node.p))
add_node_mpe(
NegativeBinomial,
get_parametric_bottom_up_log_ll(discrete_log_likelihood, negative_binomial_mode),
get_parametric_top_down_ll(negative_binomial_mode),
)
def exponential_mode(node):
return 0
add_node_mpe(
Exponential,
get_parametric_bottom_up_log_ll(continuous_log_likelihood, exponential_mode),
get_parametric_top_down_ll(exponential_mode),
)
def hypergeometric_mode(node):
return np.floor((node.n + 1) * (node.K + 1 / (node.N + 2)))
add_node_mpe(
Hypergeometric,
get_parametric_bottom_up_log_ll(continuous_log_likelihood, hypergeometric_mode),
get_parametric_top_down_ll(hypergeometric_mode),
)
def categoricaldict_mode(node):
return node.params.keys()[np.argmax(node.params.values())]
add_node_mpe(
CategoricalDictionary,
get_parametric_bottom_up_log_ll(categorical_dictionary_log_likelihood, categoricaldict_mode),
get_parametric_top_down_ll(categoricaldict_mode),
)
##Compute the conditional distribution for a multivariate Gaussian when some entries are nan i.e. unseen##
def makeconditional(mean, cov):
def conditionalmodemvg(vec):
activeset = np.isnan(vec)
totalnans = np.sum(activeset)
if(totalnans == 0):
return mn.pdf(vec, mean, cov)
if(totalnans == (len(mean))):
return mn.pdf(mean, mean, cov)
cov1 = cov[activeset, :]
cov2 = cov[~activeset, :]
cov11, cov12 = cov1[:, activeset], cov1[:, ~activeset]
cov21, cov22 = cov2[:, activeset], cov2[:, ~activeset]
temp = np.matmul(cov12, np.linalg.inv(cov22))
schur = cov11 - np.matmul(temp, cov21)
return 1. / (np.sqrt(2 * 3.14 * np.linalg.det(schur)))
return conditionalmodemvg
##Infer the conditional mean when some entries are seen##
def conditionalmean(mean, cov):
def infercondnl(dvec):
for i in range(0, len(dvec)):
activeset = np.isnan(dvec[i])
totalnans = np.sum(activeset)
if(totalnans == 0):
continue
if(totalnans == (len(mean))):
dvec[i] = mean
else:
cov1 = cov[activeset, :]
cov2 = cov[~activeset, :]
cov11, cov12 = cov1[:, activeset], cov1[:, ~activeset]
cov21, cov22 = cov2[:, activeset], cov2[:, ~activeset]
mat = np.matmul(cov12, np.linalg.inv(cov22))
arr = dvec[i]
arr[activeset] = mean[activeset] + \
np.matmul(mat, (arr[~activeset] - mean[~activeset]))
return dvec
return infercondnl
def mvg_bu_ll(node, data, dtype=np.float64):
probs = np.ones((data.shape[0], 1))
effdat = data[:, node.scope]
for i in range(0, len(effdat)):
lambdacond = makeconditional(
np.asarray(
node.mean), np.asarray(
node.sigma))
probs[i] = lambdacond(effdat[i])
return probs
def mvg_td(
node,
input_vals,
data=None,
lls_per_node=None,
dtype=np.float64):
input_vals = input_vals[0]
if len(input_vals) == 0:
return None
temp = data[input_vals, :]
checksum = np.sum(temp[:, node.scope], axis=-1)
indices = np.isnan(checksum)
createcondmean = conditionalmean(
np.asarray(
node.mean), np.asarray(
node.sigma))
temp = data[input_vals[indices], :]
temp[:, node.scope] = createcondmean(temp[:, node.scope])
data[input_vals[indices], :] = temp
return
add_node_mpe(MultivariateGaussian, mvg_bu_ll, mvg_td)
def add_parametric_mpe_support():
def gaussian_mode(node):
return node.mean
add_node_mpe(
Gaussian,
get_parametric_bottom_up_ll(gaussian_likelihood, gaussian_mode),
get_parametric_top_down_ll(gaussian_mode))
def gamma_mode(node):
return (node.alpha - 1) / node.beta
add_node_mpe(Gamma,
get_parametric_bottom_up_ll(gamma_likelihood, gamma_mode),
get_parametric_top_down_ll(gamma_mode))
def lognormal_mode(node):
return np.exp(node.mean - node.variance)
add_node_mpe(
LogNormal,
get_parametric_bottom_up_ll(lognormal_likelihood, lognormal_mode),
get_parametric_top_down_ll(lognormal_mode))
def poisson_mode(node):
return np.floor(node.mean)
add_node_mpe(Poisson,
get_parametric_bottom_up_ll(poisson_likelihood, poisson_mode),
get_parametric_top_down_ll(poisson_mode))
def bernoulli_mode(node):
if node.p > 0.5:
return 1
else:
return 0
add_node_mpe(
Bernoulli,
get_parametric_bottom_up_ll(bernoulli_likelihood, bernoulli_mode),
get_parametric_top_down_ll(bernoulli_mode))
def categorical_mode(node):
return np.argmax(node.p)
add_node_mpe(
Categorical,
get_parametric_bottom_up_ll(categorical_likelihood, categorical_mode),
get_parametric_top_down_ll(categorical_mode))
def geometric_mode(node):
return 1
add_node_mpe(
Geometric,
get_parametric_bottom_up_ll(geometric_likelihood, geometric_mode),
get_parametric_top_down_ll(geometric_mode))
def negative_binomial_mode(node):
if node.n <= 1:
return 0
else:
return np.floor(node.p * (node.n - 1) / (1 - node.p))
add_node_mpe(
NegativeBinomial,
get_parametric_bottom_up_ll(geometric_likelihood,
negative_binomial_mode),
get_parametric_top_down_ll(negative_binomial_mode))
def exponential_mode(node):
return 0
add_node_mpe(
Exponential,
get_parametric_bottom_up_ll(exponential_likelihood, exponential_mode),
get_parametric_top_down_ll(exponential_mode))
def hypergeometric_mode(node):
return np.floor((node.n + 1) * (node.K + 1 / (node.N + 2)))
add_node_mpe(
Hypergeometric,
get_parametric_bottom_up_ll(exponential_likelihood,
hypergeometric_mode),
get_parametric_top_down_ll(hypergeometric_mode))
def categoricaldict_mode(node):
return node.params.keys()[np.argmax(node.params.values())]
add_node_mpe(
CategoricalDictionary,
get_parametric_bottom_up_ll(categorical_dictionary_likelihood,
categoricaldict_mode),
get_parametric_top_down_ll(categoricaldict_mode))
def add_histogram_mpe_support():
add_node_mpe(Histogram, histogram_bottom_up_ll, histogram_top_down)
def add_conditional_mpe_support():
add_node_mpe(SupervisedLeaf, supervised_leaf_bottom_up_mpe,
supervised_leaf_top_down_mpe)
add_node_mpe(SupervisedOr, conditional_supervised_likelihood,
supervised_or_top_down_mpe)
