def _allocate_households(households, persons, tract_controls):
# Only take nonzero weights
households = households[households[inputs.HOUSEHOLD_WEIGHT.name] > 0]
# Initial weights from PUMS
w = households[inputs.HOUSEHOLD_WEIGHT.name].as_matrix().T
allocation_inputs = [inputs.NUM_PEOPLE, inputs.NUM_VEHICLES] # Hard-coded for now
# Prepend column name to bin name to prevent bin collision
hh_columns = []
for a_input in allocation_inputs:
subset_values = households[a_input.name].unique().tolist()
hh_columns += HouseholdAllocator._str_broadcast(a_input.name, subset_values)
hh_columns = HouseholdAllocator._filter_sparse_columns(households, hh_columns)
hh_table = households[hh_columns].as_matrix()
A = tract_controls.data[hh_columns].as_matrix()
n_tracts, n_controls = A.shape
n_samples = len(households.index.values)
# Control importance weights
# < 1 means not important (thus relaxing the constraint in the solver)
mu = np.mat([1] * n_controls)
w_extend = np.tile(w, (n_tracts, 1))
mu_extend = np.mat(np.tile(mu, (n_tracts, 1)))
B = np.mat(np.dot(np.ones((1, n_tracts)), A)[0])
# Our trade-off coefficient gamma
# Low values (~1) mean we trust our initial weights, high values
# (~10000) mean want to fit the marginals.
gamma = 100.
# Meta-balancing coefficient
meta_gamma = 100.
hh_weights = balance_multi_cvx(
hh_table, A, B, w_extend, gamma * mu_extend.T, meta_gamma
)
# We're running discretization independently for each tract
tract_ids = tract_controls.data['TRACTCE'].values
total_weights = np.zeros(hh_weights.shape)
sample_weights_int = hh_weights.astype(int)
discretized_hh_weights = discretize_multi_weights(hh_table, hh_weights)
total_weights = sample_weights_int + discretized_hh_weights
# Extend households and add the weights and ids
households_extend = pandas.concat([households] * n_tracts)
households_extend[inputs.COUNT.name] = total_weights.flatten().T
tracts = np.repeat(tract_ids, n_samples)
households_extend[inputs.TRACT.name] = tracts
return households_extend, persons
def _allocate_households(households, persons, tract_controls):
# Only take nonzero weights
households = households[households[inputs.HOUSEHOLD_WEIGHT.name] > 0]
# Initial weights from PUMS
w = households[inputs.HOUSEHOLD_WEIGHT.name].as_matrix().T
hh_columns = ['1', '2', '3', '4+']
hh_table = households[hh_columns].as_matrix()
A = tract_controls.data[hh_columns].as_matrix()
n_tracts, n_controls = A.shape
n_samples = len(households.index.values)
# Control importance weights
# < 1 means not important (thus relaxing the contraint in the solver)
mu = np.mat([1] * n_controls)
w_extend = np.tile(w, (n_tracts, 1))
mu_extend = np.mat(np.tile(mu, (n_tracts, 1)))
B = np.mat(np.dot(np.ones((1, n_tracts)), A)[0])
# Our trade-off coefficient gamma
# Low values (~1) mean we trust our initial weights, high values
# (~10000) mean want to fit the marginals.
gamma = 100.
# Meta-balancing coefficient
meta_gamma = 100.
hh_weights, z, q = balance_multi_cvx(hh_table, A, B, w_extend,
gamma * mu_extend.T, meta_gamma)
# We're running discretization independently for each tract
tract_ids = tract_controls.data['TRACTCE'].values
total_weights = np.zeros(hh_weights.shape)
sample_weights_int = hh_weights.astype(int)
discretized_hh_weights = discretize_multi_weights(hh_table, hh_weights)
total_weights = sample_weights_int + discretized_hh_weights
# Extend households and add the weights and ids
households_extend = pandas.concat([households] * n_tracts)
households_extend['count'] = total_weights.flatten().T
tracts = np.repeat(tract_ids, n_samples)
households_extend['tract'] = tracts
return households_extend, persons
def test_discretize_multi_zero_weights(self):
hh_table, hh_weights, expected_hh_discretized = self._mock_hh_weights_zeroed(
)
hh_discretized = listbalancer.discretize_multi_weights(
hh_table, hh_weights)
np.testing.assert_array_equal(hh_discretized, expected_hh_discretized)