def trip_schedule_calc_row_size(choosers, trace_label):
"""
rows_per_chunk calculator for trip_scheduler
"""
sizer = chunk.RowSizeEstimator(trace_label)
chooser_row_size = len(choosers.columns)
spec_columns = 3
sizer.add_elements(chooser_row_size + spec_columns, 'choosers')
row_size = sizer.get_hwm()
return row_size
def trip_purpose_calc_row_size(choosers, spec, trace_label):
"""
rows_per_chunk calculator for trip_purpose
"""
sizer = chunk.RowSizeEstimator(trace_label)
chooser_row_size = len(choosers.columns)
spec_columns = spec.shape[1] - len(PROBS_JOIN_COLUMNS)
sizer.add_elements(chooser_row_size + spec_columns, 'choosers')
row_size = sizer.get_hwm()
return row_size
def trip_scheduling_calc_row_size(trips, spec, trace_label):
sizer = chunk.RowSizeEstimator(trace_label)
# NOTE we chunk chunk_id
# scale row_size by average number of chooser rows per chunk_id
num_choosers = trips['chunk_id'].max() + 1
rows_per_chunk_id = len(trips) / num_choosers
# only non-initial trips require scheduling, segment handing first such trip in tour will use most space
outbound_chooser = (trips.trip_num == 2) & trips.outbound & (
trips.primary_purpose != 'atwork')
inbound_chooser = (trips.trip_num
== trips.trip_count - 1) & ~trips.outbound & (
trips.primary_purpose != 'atwork')
# furthermore, inbound and outbound are scheduled independently
if outbound_chooser.sum() > inbound_chooser.sum():
is_chooser = outbound_chooser
logger.debug(
f"{trace_label} {is_chooser.sum()} outbound_choosers of {len(trips)} require scheduling"
)
else:
is_chooser = inbound_chooser
logger.debug(
f"{trace_label} {is_chooser.sum()} inbound_choosers of {len(trips)} require scheduling"
)
chooser_fraction = is_chooser.sum() / len(trips)
logger.debug(f"{trace_label} chooser_fraction {chooser_fraction *100}%")
chooser_row_size = len(trips.columns) + len(
spec.columns) - len(PROBS_JOIN_COLUMNS)
sizer.add_elements(chooser_fraction * chooser_row_size, 'choosers')
# might be clipped to fewer but this is worst case
chooser_probs_row_size = len(spec.columns) - len(PROBS_JOIN_COLUMNS)
sizer.add_elements(chooser_fraction * chooser_probs_row_size,
'chooser_probs')
sizer.add_elements(chooser_fraction, 'choices')
sizer.add_elements(chooser_fraction, 'rands')
sizer.add_elements(chooser_fraction, 'failed')
row_size = sizer.get_hwm()
row_size = row_size * rows_per_chunk_id
return row_size
def accessibility_calc_row_size(accessibility_df, land_use_df, assignment_spec,
network_los, trace_label):
"""
rows_per_chunk calculator for accessibility
"""
sizer = chunk.RowSizeEstimator(trace_label)
# if there are skims, and zone_system is THREE_ZONE, and there are any
# then we want to estimate the per-row overhead tvpb skims
# (do this first to facilitate tracing of rowsize estimation below)
if network_los.zone_system == los.THREE_ZONE:
# DISABLE_TVPB_OVERHEAD
logger.debug("disable calc_row_size for THREE_ZONE with tap skims")
return 0
land_use_rows = len(land_use_df.index)
land_use_columns = len(land_use_df.columns)
od_columns = 2
# assignment spec has one row per value to assign
# count number of unique persistent assign_variables targets simultaneously resident during spec eval
# (since dict overwrites recurring targets, only count unique targets)
def is_persistent(target):
return not (assign.is_throwaway(target)
or assign.is_temp_scalar(target))
num_spec_values = len([
target for target in assignment_spec.target.unique()
if is_persistent(target)
])
sizer.add_elements(land_use_rows * od_columns, 'od_df')
# each od_df joins to all land_use zones
sizer.add_elements(land_use_rows * land_use_columns, 'land_use_choosers')
# and then we assign_variables to joined land_use from assignment_spec
sizer.add_elements(land_use_rows * num_spec_values, 'spec_values')
row_size = sizer.get_hwm()
return row_size
def initialize_tvpb_calc_row_size(choosers, network_los, trace_label):
"""
rows_per_chunk calculator for trip_purpose
"""
sizer = chunk.RowSizeEstimator(trace_label)
model_settings = \
network_los.setting(f'TVPB_SETTINGS.tour_mode_choice.tap_tap_settings')
attributes_as_columns = \
network_los.setting('TVPB_SETTINGS.tour_mode_choice.tap_tap_settings.attributes_as_columns', [])
# expression_values for each spec row
sizer.add_elements(len(choosers.columns), 'choosers')
# expression_values for each spec row
sizer.add_elements(len(attributes_as_columns), 'attributes_as_columns')
preprocessor_settings = model_settings.get('PREPROCESSOR')
if preprocessor_settings:
preprocessor_spec_name = preprocessor_settings.get('SPEC', None)
if not preprocessor_spec_name.endswith(".csv"):
preprocessor_spec_name = f'{preprocessor_spec_name}.csv'
expressions_spec = assign.read_assignment_spec(
config.config_file_path(preprocessor_spec_name))
sizer.add_elements(expressions_spec.shape[0], 'preprocessor')
# expression_values for each spec row
spec = simulate.read_model_spec(file_name=model_settings['SPEC'])
sizer.add_elements(spec.shape[0], 'expression_values')
# expression_values for each spec row
sizer.add_elements(spec.shape[1], 'utilities')
row_size = sizer.get_hwm()
return row_size
def tour_scheduling_calc_row_size(tours, persons_merged, alternatives, skims,
spec, model_settings, trace_label):
# this will not be consistent across mandatory tours (highest), non_mandatory tours, and atwork subtours (lowest)
TIMETABLE_AVAILABILITY_REDUCTION_FACTOR = 1
# this appears to be more stable
LOGSUM_DUPLICATE_REDUCTION_FACTOR = 0.5
sizer = chunk.RowSizeEstimator(trace_label)
# chooser is tours merged with persons_merged
chooser_row_size = len(tours.columns) + len(persons_merged.columns)
# e.g. start, end, duration, <chooser_column>
alt_row_size = alternatives.shape[1] + 1
# non-available alternatives will be sliced out so this is a over-estimate
# for atwork subtours this may be a gross over-estimate, but that is presumably ok since we are adaptive
sample_size = len(alternatives) * TIMETABLE_AVAILABILITY_REDUCTION_FACTOR
sizer.add_elements(chooser_row_size, 'tours') # tours_merged with persons
# alt_tdd tdd_interaction_dataset is cross join of choosers with alternatives
sizer.add_elements((chooser_row_size + alt_row_size) * sample_size,
'interaction_df')
# eval_interaction_utilities is parsimonious and doesn't create a separate column for each partial utility
sizer.add_elements(
sample_size,
'interaction_utilities') # <- this is probably always the HWM
sizer.drop_elements('interaction_df')
sizer.drop_elements('interaction_utilities')
sizer.add_elements(alt_row_size, 'utilities_df')
sizer.add_elements(alt_row_size, 'probs')
if 'LOGSUM_SETTINGS' in model_settings:
logsum_settings = config.read_model_settings(
model_settings['LOGSUM_SETTINGS'])
logsum_spec = simulate.read_model_spec(
file_name=logsum_settings['SPEC'])
logsum_nest_spec = config.get_logit_model_settings(logsum_settings)
if logsum_nest_spec is None:
# expression_values for each spec row
# utilities and probs for each alt
logsum_columns = logsum_spec.shape[0] + (2 * logsum_spec.shape[1])
else:
# expression_values for each spec row
# raw_utilities and base_probabilities) for each alt
# nested_exp_utilities, nested_probabilities for each nest
# less 1 as nested_probabilities lacks root
nest_count = logit.count_nests(logsum_nest_spec)
logsum_columns = logsum_spec.shape[0] + (
2 * logsum_spec.shape[1]) + (2 * nest_count) - 1
if USE_BRUTE_FORCE_TO_COMPUTE_LOGSUMS:
sizer.add_elements(logsum_columns * sample_size, 'logsum_columns')
else:
# if USE_BRUTE_FORCE_TO_COMPUTE_LOGSUMS is false compute_logsums prunes alt_tdd
# to only compute logsums for unique (tour_id, out_period, in_period, duration) in alt_tdd
# which cuts the number of alts by roughly 50% (44% for 100 hh mtctm1 test dataset)
# grep the log for USE_BRUTE_FORCE_TO_COMPUTE_LOGSUMS to check actual % savings
duplicate_sample_reduction = 0.5
sizer.add_elements(
logsum_columns * sample_size *
LOGSUM_DUPLICATE_REDUCTION_FACTOR, 'logsum_columns')
row_size = sizer.get_hwm()
if simulate.tvpb_skims(skims):
# DISABLE_TVPB_OVERHEAD
logger.debug("disable calc_row_size for THREE_ZONE with tap skims")
return 0
return row_size
