def _sample_by_agent_and_stratum(self, index1, index2, stratum, prob_array,
chosen_choice_index,
strata_sample_setting):
"""agent by agent and stratum by stratum stratified sampling, suitable for 2d prob_array and/or sample_size varies for agents
this method is slower than _sample_by_stratum, for simpler stratified sampling use _sample_by_stratum instead"""
rank_of_prob = rank(prob_array)
rank_of_strata = rank(strata_sample_setting)
J = self.__determine_sampled_index_size(strata_sample_setting,
rank_of_strata)
sampled_index = zeros((index1.size, J), dtype=DTYPE) - 1
self._sampling_probability = zeros((index1.size, J), dtype=float32)
self._stratum_id = ones((index1.size, J), dtype=DTYPE) * NO_STRATUM_ID
for i in range(index1.size):
if rank_of_strata == 3:
strata_sample_pairs = strata_sample_setting[i, :]
else:
strata_sample_pairs = strata_sample_setting
if rank_of_prob == 2:
prob = prob_array[i, :]
else:
prob = prob_array
j = 0
for (this_stratum, this_size) in strata_sample_pairs:
if this_size <= 0: continue
index_not_in_stratum = where(stratum != this_stratum)[0]
this_prob = copy.copy(prob)
this_prob[index_not_in_stratum] = 0.0
this_prob = normalize(this_prob)
if nonzerocounts(this_prob) < this_size:
logger.log_warning(
"weight array dosen't have enough non-zero counts, use sample with replacement"
)
# chosen_index_to_index2 = where(index2 == chosen_choice_index[i])[0]
#exclude_index passed to probsample_noreplace needs to be indexed to index2
this_sampled_index = probsample_noreplace(
index2,
sample_size=this_size,
prob_array=this_prob,
exclude_index=chosen_choice_index[i],
return_index=True)
sampled_index[i, j:j + this_size] = this_sampled_index
self._sampling_probability[
i, j:j + this_size] = this_prob[this_sampled_index]
self._stratum_id[i, j:j + this_size] = ones(
(this_sampled_index.size, ), dtype=DTYPE) * this_stratum
j += this_size
return index2[sampled_index]
def _sample_by_agent_and_stratum(
self, index1, index2, stratum, prob_array, chosen_choice_index, strata_sample_setting
):
"""agent by agent and stratum by stratum stratified sampling, suitable for 2d prob_array and/or sample_size varies for agents
this method is slower than _sample_by_stratum, for simpler stratified sampling use _sample_by_stratum instead"""
rank_of_prob = rank(prob_array)
rank_of_strata = rank(strata_sample_setting)
J = self.__determine_sampled_index_size(strata_sample_setting, rank_of_strata)
sampled_index = zeros((index1.size, J), dtype="int32") - 1
self._sampling_probability = zeros((index1.size, J), dtype=float32)
self._stratum_id = ones((index1.size, J), dtype="int32") * NO_STRATUM_ID
for i in range(index1.size):
if rank_of_strata == 3:
strata_sample_pairs = strata_sample_setting[i, :]
else:
strata_sample_pairs = strata_sample_setting
if rank_of_prob == 2:
prob = prob_array[i, :]
else:
prob = prob_array
j = 0
for (this_stratum, this_size) in strata_sample_pairs:
if this_size <= 0:
continue
index_not_in_stratum = where(stratum != this_stratum)[0]
this_prob = copy.copy(prob)
this_prob[index_not_in_stratum] = 0.0
this_prob = normalize(this_prob)
if nonzerocounts(this_prob) < this_size:
logger.log_warning("weight array dosen't have enough non-zero counts, use sample with replacement")
# chosen_index_to_index2 = where(index2 == chosen_choice_index[i])[0]
# exclude_index passed to probsample_noreplace needs to be indexed to index2
this_sampled_index = probsample_noreplace(
index2,
sample_size=this_size,
prob_array=this_prob,
exclude_index=chosen_choice_index[i],
return_index=True,
)
sampled_index[i, j : j + this_size] = this_sampled_index
self._sampling_probability[i, j : j + this_size] = this_prob[this_sampled_index]
self._stratum_id[i, j : j + this_size] = ones((this_sampled_index.size,), dtype="int32") * this_stratum
j += this_size
return index2[sampled_index]
def _sample_by_stratum(self, index1, index2, stratum, prob_array,
chosen_choice_index, strata_sample_setting):
"""stratum by stratum stratified sampling, suitable for 1d prob_array and sample_size is the same for all agents"""
if prob_array.ndim <> 1:
raise RuntimeError, "_sample_by_stratum only suitable for 1d prob_array"
sampled_index = zeros((index1.size, 1), dtype=DTYPE) - 1
self._sampling_probability = zeros((index1.size, 1), dtype=float32)
self._stratum_id = ones((index1.size, 1), dtype=DTYPE) * NO_STRATUM_ID
for this_stratum, this_size in strata_sample_setting:
index_not_in_stratum = where(stratum != this_stratum)[0]
this_prob = copy.copy(prob_array)
this_prob[index_not_in_stratum] = 0.0
this_prob = normalize(this_prob)
replace = False # non-repeat sampling
if nonzerocounts(this_prob) < this_size:
logger.log_warning(
"weight array dosen't have enough non-zero counts, sample with replacement"
)
replace = True
this_sampled_index = prob2dsample(
index2,
sample_size=(index1.size, this_size),
prob_array=this_prob,
exclude_index=chosen_choice_index,
replace=replace,
return_index=True)
sampled_index = concatenate((sampled_index, this_sampled_index),
axis=1)
self._sampling_probability = concatenate(
(self._sampling_probability, this_prob[this_sampled_index]),
axis=1)
self._stratum_id = concatenate(
(self._stratum_id,
ones((this_sampled_index.shape[0], 1), dtype=DTYPE) *
this_stratum),
axis=1)
self._sampling_probability = self._sampling_probability[:, 1:]
self._stratum_id = self._stratum_id[:, 1:]
return index2[sampled_index[:, 1:]]
def _sample_by_stratum(self, index1, index2, stratum, prob_array, chosen_choice_index, strata_sample_setting):
"""stratum by stratum stratified sampling, suitable for 1d prob_array and sample_size is the same for all agents"""
if prob_array.ndim <> 1:
raise RuntimeError, "_sample_by_stratum only suitable for 1d prob_array"
sampled_index = zeros((index1.size, 1), dtype="int32") - 1
self._sampling_probability = zeros((index1.size, 1), dtype=float32)
self._stratum_id = ones((index1.size, 1), dtype="int32") * NO_STRATUM_ID
for this_stratum, this_size in strata_sample_setting:
index_not_in_stratum = where(stratum != this_stratum)[0]
this_prob = copy.copy(prob_array)
this_prob[index_not_in_stratum] = 0.0
this_prob = normalize(this_prob)
replace = False # non-repeat sampling
if nonzerocounts(this_prob) < this_size:
logger.log_warning("weight array dosen't have enough non-zero counts, sample with replacement")
replace = True
this_sampled_index = prob2dsample(
index2,
sample_size=(index1.size, this_size),
prob_array=this_prob,
exclude_index=chosen_choice_index,
replace=replace,
return_index=True,
)
sampled_index = concatenate((sampled_index, this_sampled_index), axis=1)
self._sampling_probability = concatenate(
(self._sampling_probability, this_prob[this_sampled_index]), axis=1
)
self._stratum_id = concatenate(
(self._stratum_id, ones((this_sampled_index.shape[0], 1), dtype="int32") * this_stratum), axis=1
)
self._sampling_probability = self._sampling_probability[:, 1:]
self._stratum_id = self._stratum_id[:, 1:]
return index2[sampled_index[:, 1:]]
def run(self,
dataset1,
dataset2,
index1=None,
index2=None,
sample_size=10,
weight=None,
include_chosen_choice=False,
with_replacement=False,
resources=None,
dataset_pool=None):
"""this function samples number of sample_size (scalar value) alternatives from dataset2
for agent set specified by dataset1.
If index1 is not None, only samples alterantives for agents with indices in index1;
if index2 is not None, only samples alternatives from indices in index2.
sample_size specifies number of alternatives to be sampled for each agent.
weight, to be used as sampling weight, is either an attribute name of dataset2, or a 1d
array of the same length as index2 or 2d array of shape (index1.size, index2.size).
Also refer to document of interaction_dataset"""
if dataset_pool is None:
try:
sc = SessionConfiguration()
dataset_pool = sc.get_dataset_pool()
except:
dataset_pool = DatasetPool()
local_resources = Resources(resources)
local_resources.merge_if_not_None({
"dataset1":
dataset1,
"dataset2":
dataset2,
"index1":
index1,
"index2":
index2,
"sample_size":
sample_size,
"weight":
weight,
"with_replacement":
with_replacement,
"include_chosen_choice":
include_chosen_choice
})
local_resources.check_obligatory_keys(
['dataset1', 'dataset2', 'sample_size'])
agent = local_resources["dataset1"]
index1 = local_resources.get("index1", None)
if index1 is None:
index1 = arange(agent.size())
choice = local_resources["dataset2"]
index2 = local_resources.get("index2", None)
if index2 is None:
index2 = arange(choice.size())
if index1.size == 0 or index2.size == 0:
err_msg = "either choice size or agent size is zero, return None"
logger.log_warning(err_msg)
return None
include_chosen_choice = local_resources.get("include_chosen_choice",
False)
J = local_resources["sample_size"]
if include_chosen_choice:
J = J - 1
with_replacement = local_resources.get("with_replacement")
weight = local_resources.get("weight", None)
if isinstance(weight, str):
if weight in choice.get_known_attribute_names():
weight = choice.get_attribute(weight)
rank_of_weight = 1
else:
varname = VariableName(weight)
if varname.get_dataset_name() == choice.get_dataset_name():
weight = choice.compute_variables(
weight, dataset_pool=dataset_pool)
rank_of_weight = 1
elif varname.get_interaction_set_names() is not None:
## weights can be an interaction variable
interaction_dataset = InteractionDataset(local_resources)
weight = interaction_dataset.compute_variables(
weight, dataset_pool=dataset_pool)
rank_of_weight = 2
assert (len(weight.shape) >= rank_of_weight)
else:
err_msg = ("weight is neither a known attribute name "
"nor a simple variable from the choice dataset "
"nor an interaction variable: '%s'" % weight)
logger.log_error(err_msg)
raise ValueError, err_msg
elif isinstance(weight, ndarray):
rank_of_weight = weight.ndim
elif not weight: ## weight is None or empty string
weight = ones(index2.size)
rank_of_weight = 1
else:
err_msg = "unkown weight type"
logger.log_error(err_msg)
raise TypeError, err_msg
if (weight.size <> index2.size) and (weight.shape[rank_of_weight - 1]
<> index2.size):
if weight.shape[rank_of_weight - 1] == choice.size():
if rank_of_weight == 1:
weight = take(weight, index2)
if rank_of_weight == 2:
weight = take(weight, index2, axis=1)
else:
err_msg = "weight array size doesn't match to size of dataset2 or its index"
logger.log_error(err_msg)
raise ValueError, err_msg
prob = normalize(weight)
#chosen_choice = ones(index1.size) * UNPLACED_ID
chosen_choice_id = agent.get_attribute(choice.get_id_name()[0])[index1]
#index_of_placed_agent = where(greater(chosen_choice_id, UNPLACED_ID))[0]
chosen_choice_index = choice.try_get_id_index(
chosen_choice_id, return_value_if_not_found=UNPLACED_ID)
chosen_choice_index_to_index2 = lookup(chosen_choice_index,
index2,
index_if_not_found=UNPLACED_ID)
if rank_of_weight == 1: # if weight_array is 1d, then each agent shares the same weight for choices
replace = with_replacement # sampling with no replacement
non_zero_counts = nonzerocounts(weight)
if non_zero_counts < J:
logger.log_warning(
"weight array dosen't have enough non-zero counts, use sample with replacement"
)
replace = True
if non_zero_counts > 0:
sampled_index = prob2dsample(
index2,
sample_size=(index1.size, J),
prob_array=prob,
exclude_index=chosen_choice_index_to_index2,
replace=replace,
return_index=True)
else:
# all alternatives have a zero weight
sampled_index = zeros((index1.size, 0), dtype=DTYPE)
#return index2[sampled_index]
if rank_of_weight == 2:
sampled_index = zeros((index1.size, J), dtype=DTYPE) - 1
for i in range(index1.size):
replace = with_replacement # sampling with/without replacement
i_prob = prob[i, :]
if nonzerocounts(i_prob) < J:
logger.log_warning(
"weight array dosen't have enough non-zero counts, use sample with replacement"
)
replace = True
#exclude_index passed to probsample_noreplace needs to be indexed to index2
sampled_index[i, :] = probsample_noreplace(
index2,
sample_size=J,
prob_array=i_prob,
exclude_index=chosen_choice_index_to_index2[i],
return_index=True)
sampling_prob = take(prob, sampled_index)
sampled_index_within_prob = sampled_index.copy()
sampled_index = index2[sampled_index]
is_chosen_choice = zeros(sampled_index.shape, dtype="bool")
#chosen_choice = -1 * ones(chosen_choice_index.size, dtype="int32")
if include_chosen_choice:
sampled_index = column_stack(
(chosen_choice_index[:, newaxis], sampled_index))
is_chosen_choice = zeros(sampled_index.shape, dtype="bool")
is_chosen_choice[chosen_choice_index != UNPLACED_ID, 0] = 1
#chosen_choice[where(is_chosen_choice)[0]] = where(is_chosen_choice)[1]
## this is necessary because prob is indexed to index2, not to the choice set (as is chosen_choice_index)
sampling_prob_for_chosen_choices = take(
prob, chosen_choice_index_to_index2[:, newaxis])
## if chosen choice chosen equals unplaced_id then the sampling prob is 0
sampling_prob_for_chosen_choices[where(
chosen_choice_index == UNPLACED_ID)[0], ] = 0.0
sampling_prob = column_stack(
[sampling_prob_for_chosen_choices, sampling_prob])
interaction_dataset = self.create_interaction_dataset(
dataset1, dataset2, index1, sampled_index)
interaction_dataset.add_attribute(sampling_prob,
'__sampling_probability')
interaction_dataset.add_attribute(is_chosen_choice, 'chosen_choice')
if local_resources.get("include_mnl_bias_correction_term", False):
if include_chosen_choice:
sampled_index_within_prob = column_stack(
(chosen_choice_index_to_index2[:, newaxis],
sampled_index_within_prob))
interaction_dataset.add_mnl_bias_correction_term(
prob, sampled_index_within_prob)
## to get the older returns
#sampled_index = interaction_dataset.get_2d_index()
#chosen_choices = UNPLACED_ID * ones(index1.size, dtype="int32")
#where_chosen = where(interaction_dataset.get_attribute("chosen_choice"))
#chosen_choices[where_chosen[0]]=where_chosen[1]
#return (sampled_index, chosen_choice)
return interaction_dataset
def run(self, dataset1, dataset2, index1=None, index2=None, sample_size=10, weight=None,
include_chosen_choice=False, with_replacement=False, resources=None, dataset_pool=None):
"""this function samples number of sample_size (scalar value) alternatives from dataset2
for agent set specified by dataset1.
If index1 is not None, only samples alterantives for agents with indices in index1;
if index2 is not None, only samples alternatives from indices in index2.
sample_size specifies number of alternatives to be sampled for each agent.
weight, to be used as sampling weight, is either an attribute name of dataset2, or a 1d
array of the same length as index2 or 2d array of shape (index1.size, index2.size).
Also refer to document of interaction_dataset"""
if dataset_pool is None:
try:
sc = SessionConfiguration()
dataset_pool=sc.get_dataset_pool()
except:
dataset_pool = DatasetPool()
local_resources = Resources(resources)
local_resources.merge_if_not_None(
{"dataset1": dataset1, "dataset2": dataset2,
"index1":index1, "index2": index2,
"sample_size": sample_size, "weight": weight,
"with_replacement": with_replacement,
"include_chosen_choice": include_chosen_choice})
local_resources.check_obligatory_keys(['dataset1', 'dataset2', 'sample_size'])
agent = local_resources["dataset1"]
index1 = local_resources.get("index1", None)
if index1 is None:
index1 = arange(agent.size())
choice = local_resources["dataset2"]
index2 = local_resources.get("index2", None)
if index2 is None:
index2 = arange(choice.size())
if index1.size == 0 or index2.size == 0:
err_msg = "either choice size or agent size is zero, return None"
logger.log_warning(err_msg)
return None
include_chosen_choice = local_resources.get("include_chosen_choice", False)
J = local_resources["sample_size"]
if include_chosen_choice:
J = J - 1
with_replacement = local_resources.get("with_replacement")
weight = local_resources.get("weight", None)
if isinstance(weight, str):
if weight in choice.get_known_attribute_names():
weight=choice.get_attribute(weight)
rank_of_weight = 1
elif VariableName(weight).get_dataset_name() == choice.get_dataset_name():
weight=choice.compute_variables(weight, dataset_pool=dataset_pool)
rank_of_weight = 1
else:
## weights can be an interaction variable
interaction_dataset = InteractionDataset(local_resources)
weight=interaction_dataset.compute_variables(weight, dataset_pool=dataset_pool)
rank_of_weight = 2
elif isinstance(weight, ndarray):
rank_of_weight = weight.ndim
elif not weight: ## weight is None or empty string
weight = ones(index2.size)
rank_of_weight = 1
else:
err_msg = "unkown weight type"
logger.log_error(err_msg)
raise TypeError, err_msg
if (weight.size <> index2.size) and (weight.shape[rank_of_weight-1] <> index2.size):
if weight.shape[rank_of_weight-1] == choice.size():
if rank_of_weight == 1:
weight = take(weight, index2)
if rank_of_weight == 2:
weight = take(weight, index2, axis=1)
else:
err_msg = "weight array size doesn't match to size of dataset2 or its index"
logger.log_error(err_msg)
raise ValueError, err_msg
prob = normalize(weight)
#chosen_choice = ones(index1.size) * UNPLACED_ID
chosen_choice_id = agent.get_attribute(choice.get_id_name()[0])[index1]
#index_of_placed_agent = where(greater(chosen_choice_id, UNPLACED_ID))[0]
chosen_choice_index = choice.try_get_id_index(chosen_choice_id, return_value_if_not_found=UNPLACED_ID)
chosen_choice_index_to_index2 = lookup(chosen_choice_index, index2, index_if_not_found=UNPLACED_ID)
if rank_of_weight == 1: # if weight_array is 1d, then each agent shares the same weight for choices
replace = with_replacement # sampling with no replacement
if nonzerocounts(weight) < J:
logger.log_warning("weight array dosen't have enough non-zero counts, use sample with replacement")
replace = True
sampled_index = prob2dsample( index2, sample_size=(index1.size, J),
prob_array=prob, exclude_index=chosen_choice_index_to_index2,
replace=replace, return_index=True )
#return index2[sampled_index]
if rank_of_weight == 2:
sampled_index = zeros((index1.size,J), dtype="int32") - 1
for i in range(index1.size):
replace = with_replacement # sampling with/without replacement
i_prob = prob[i,:]
if nonzerocounts(i_prob) < J:
logger.log_warning("weight array dosen't have enough non-zero counts, use sample with replacement")
replace = True
#exclude_index passed to probsample_noreplace needs to be indexed to index2
sampled_index[i,:] = probsample_noreplace( index2, sample_size=J, prob_array=i_prob,
exclude_index=chosen_choice_index_to_index2[i],
return_index=True )
sampling_prob = take(prob, sampled_index)
sampled_index = index2[sampled_index]
is_chosen_choice = zeros(sampled_index.shape, dtype="bool")
#chosen_choice = -1 * ones(chosen_choice_index.size, dtype="int32")
if include_chosen_choice:
sampled_index = column_stack((chosen_choice_index[:,newaxis],sampled_index))
is_chosen_choice = zeros(sampled_index.shape, dtype="bool")
is_chosen_choice[chosen_choice_index!=UNPLACED_ID, 0] = 1
#chosen_choice[where(is_chosen_choice)[0]] = where(is_chosen_choice)[1]
## this is necessary because prob is indexed to index2, not to the choice set (as is chosen_choice_index)
sampling_prob_for_chosen_choices = take(prob, chosen_choice_index_to_index2[:, newaxis])
## if chosen choice chosen equals unplaced_id then the sampling prob is 0
sampling_prob_for_chosen_choices[where(chosen_choice_index==UNPLACED_ID)[0],] = 0.0
sampling_prob = column_stack([sampling_prob_for_chosen_choices, sampling_prob])
interaction_dataset = self.create_interaction_dataset(dataset1, dataset2, index1, sampled_index)
interaction_dataset.add_attribute(sampling_prob, '__sampling_probability')
interaction_dataset.add_attribute(is_chosen_choice, 'chosen_choice')
## to get the older returns
#sampled_index = interaction_dataset.get_2d_index()
#chosen_choices = UNPLACED_ID * ones(index1.size, dtype="int32")
#where_chosen = where(interaction_dataset.get_attribute("chosen_choice"))
#chosen_choices[where_chosen[0]]=where_chosen[1]
#return (sampled_index, chosen_choice)
return interaction_dataset
