def precompute_minimal(data, settings):
param = empty()
cache = {}
if settings.optype == 'class':
param.alpha = settings.alpha
param.alpha_per_class = float(param.alpha) / data['n_class']
cache['y_train_counts'] = hist_count(data['y_train'],
range(data['n_class']))
cache['range_n_class'] = range(data['n_class'])
param.base_measure = (np.ones(data['n_class']) + 0.) / data['n_class']
param.alpha_vec = param.base_measure * param.alpha
else:
cache['sum_y'] = float(np.sum(data['y_train']))
cache['sum_y2'] = float(np.sum(data['y_train']**2))
cache['n_points'] = len(data['y_train'])
warn('initializing prior mean and precision to their true values')
# FIXME: many of the following are relevant only for mondrian forests
param.prior_mean = np.mean(data['y_train'])
param.prior_variance = np.var(data['y_train'])
param.prior_precision = 1.0 / param.prior_variance
if not settings.smooth_hierarchically:
param.noise_variance = 0.01 # FIXME: hacky
else:
K = min(
1000, data['n_train']
) # FIXME: measurement noise set to fraction of unconditional variance
param.noise_variance = param.prior_variance / (
1. + K) # assume noise variance = prior_variance / (2K)
# NOTE: max_split_cost scales inversely with the number of dimensions
param.variance_coef = 2.0 * param.prior_variance
param.sigmoid_coef = data['n_dim'] / (2.0 * np.log2(data['n_train']))
param.noise_precision = 1.0 / param.noise_variance
return (param, cache)
def precompute_minimal(data, settings):
param = empty()
cache = {}
if settings.optype == 'class':
param.alpha = settings.alpha
param.alpha_per_class = float(param.alpha) / data['n_class']
cache['y_train_counts'] = hist_count(data['y_train'], range(data['n_class']))
cache['range_n_class'] = range(data['n_class'])
param.base_measure = (np.ones(data['n_class']) + 0.) / data['n_class']
param.alpha_vec = param.base_measure * param.alpha
else:
cache['sum_y'] = float(np.sum(data['y_train']))
cache['sum_y2'] = float(np.sum(data['y_train'] ** 2))
cache['n_points'] = len(data['y_train'])
warn('initializing prior mean and precision to their true values')
# FIXME: many of the following are relevant only for mondrian forests
param.prior_mean = np.mean(data['y_train'])
param.prior_variance = np.var(data['y_train'])
param.prior_precision = 1.0 / param.prior_variance
if not settings.smooth_hierarchically:
param.noise_variance = 0.01 # FIXME: hacky
else:
K = min(1000, data['n_train']) # FIXME: measurement noise set to fraction of unconditional variance
param.noise_variance = param.prior_variance / (1. + K) # assume noise variance = prior_variance / (2K)
# NOTE: max_split_cost scales inversely with the number of dimensions
param.variance_coef = 2.0 * param.prior_variance
param.sigmoid_coef = data['n_dim'] / (2.0 * np.log2(data['n_train']))
param.noise_precision = 1.0 / param.noise_variance
return (param, cache)
def precompute_minimal(data, settings):
param = empty()
cache = {}
assert settings.optype == 'class'
if settings.optype == 'class':
param.alpha = settings.alpha
param.alpha_per_class = float(param.alpha) / data['n_class']
cache['y_train_counts'] = hist_count(data['y_train'], range(data['n_class']))
cache['range_n_class'] = range(data['n_class'])
param.base_measure = (np.ones(data['n_class']) + 0.) / data['n_class']
param.alpha_vec = param.base_measure * param.alpha
return (param, cache)
def precompute_minimal(data, settings):
param = empty()
cache = {}
assert settings.optype == 'class'
if settings.optype == 'class':
param.alpha = settings.alpha
param.alpha_per_class = float(param.alpha) / data['n_class']
cache['y_train_counts'] = hist_count(data['y_train'],
range(data['n_class']))
cache['range_n_class'] = range(data['n_class'])
param.base_measure = (np.ones(data['n_class']) + 0.) / data['n_class']
param.alpha_vec = param.base_measure * param.alpha
return (param, cache)