# -*- coding: utf-8 -*-
# vim: tabstop=4 shiftwidth=4 softtabstop=4
#
# Copyright (C) 2015-2025 GEM Foundation
#
# OpenQuake is free software: you can redistribute it and/or modify it
# under the terms of the GNU Affero General Public License as published
# by the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
#
# OpenQuake is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU Affero General Public License for more details.
#
# You should have received a copy of the GNU Affero General Public License
# along with OpenQuake. If not, see <http://www.gnu.org/licenses/>.
import os.path
import logging
import operator
from functools import partial
import numpy
import pandas
from scipy import sparse
from openquake.baselib import hdf5, performance, general, python3compat, config
from openquake.hazardlib import stats, InvalidFile
from openquake.commonlib.calc import starmap_from_gmfs, split
from openquake.risklib.scientific import (
total_losses, insurance_losses, MultiEventRNG, LOSSID)
from openquake.calculators import base, event_based
from openquake.calculators.post_risk import (
PostRiskCalculator, post_aggregate, fix_dtypes, fix_investigation_time)
U8 = numpy.uint8
U16 = numpy.uint16
U32 = numpy.uint32
U64 = numpy.uint64
F32 = numpy.float32
F64 = numpy.float64
TWO16 = 2 ** 16
TWO24 = 2 ** 24
TWO32 = U64(2 ** 32)
get_n_occ = operator.itemgetter(1)
[docs]def fast_agg(keys, values, correl, li, loss2):
"""
:param keys: an array of N uint64 numbers encoding (event_id, agg_id)
:param values: an array of (N, D) floats
:param correl: True if there is asset correlation
:param li: loss type index
:param loss2: dictionary unique key -> array(L, D)
"""
ukeys, avalues = general.fast_agg2(keys, values)
if correl: # restore the variances
avalues[:, 0] = avalues[:, 0] ** 2
for ukey, avalue in zip(ukeys, avalues):
loss2[ukey][li] += avalue
[docs]def update(loss3, lti, X, alt, rlz_id, collect_rlzs):
"""
Populate loss3 a dictionary {'aids': [], 'bids': [], 'loss': []}
:param alt: DataFrame agg_loss_table
:param rlz_id: effective realization index (usually 0)
:param collect_rlzs: usually True
"""
if collect_rlzs or len(numpy.unique(rlz_id)) == 1:
# fast lane
ldf = pandas.DataFrame(
dict(aid=alt.aid.to_numpy(), loss=alt.loss.to_numpy()))
tot = ldf.groupby('aid').loss.sum()
aids = tot.index.to_numpy()
rlzs = numpy.zeros_like(tot, U32)
else:
# rare case
ldf = pandas.DataFrame(
dict(aid=alt.aid.to_numpy(), loss=alt.loss.to_numpy(),
rlz=rlz_id[U32(alt.eid)])) # NB: without the U32 here
# the SURA calculation would fail with alt.eid being F64 (?)
tot = ldf.groupby(['aid', 'rlz']).loss.sum()
aids, rlzs = zip(*tot.index)
rlzs = U32(rlzs)
loss3['aids'].append(U32(aids))
loss3['bids'].append(rlzs * X + lti)
loss3['loss'].append(F32(tot))
[docs]def debugprint(ln, asset_loss_table, adf):
"""
Print risk_by_event in a reasonable format. To be used with --nd
"""
if '+' in ln or ln == 'claim':
df = asset_loss_table.set_index('aid').rename(columns={'loss': ln})
df['asset_id'] = python3compat.decode(adf.id[df.index].to_numpy())
del df['variance']
print(df)
[docs]def build_avg(loss3, A, X):
if loss3['aids']:
aids = numpy.concatenate(loss3['aids'], dtype=U32)
bids = numpy.concatenate(loss3['bids'], dtype=U32)
loss = numpy.concatenate(loss3['loss'], dtype=F32)
avg = sparse.coo_matrix((loss, (aids, bids)), (A, X))
else:
avg = sparse.coo_matrix((A, X), dtype=F32)
return avg
[docs]def build_alt(loss2, xtypes):
lis = range(len(xtypes))
dic = general.AccumDict(accum=[])
for ukey, arr in loss2.items():
eid, kid = divmod(ukey, TWO32)
for li in lis:
if arr[li].any():
dic['event_id'].append(eid)
dic['agg_id'].append(kid)
dic['loss_id'].append(LOSSID[xtypes[li]])
for c, col in enumerate(['variance', 'loss']):
dic[col].append(arr[li, c])
fix_dtypes(dic)
return pandas.DataFrame(dic)
[docs]def aggreg(outputs, loss2, loss3, crmodel, K, aggids, rlz_id, xtypes, monitor):
oq = crmodel.oqparam
correl = int(oq.asset_correlation)
X = len(xtypes)
value_cols = ['variance', 'loss']
for out in outputs:
for li, ln in enumerate(xtypes):
if ln not in out or len(out[ln]) == 0:
continue
alt = out[ln]
if oq.avg_losses: # fast
update(loss3, li, X, alt, rlz_id, oq.collect_rlzs)
if correl: # use sigma^2 = (sum sigma_i)^2
alt['variance'] = numpy.sqrt(alt.variance)
eids = alt.eid.to_numpy() * TWO32 # U64
values = numpy.array([alt[col] for col in value_cols]).T
# aggregate all assets
fast_agg(eids + U64(K), values, correl, li, loss2)
if len(aggids):
# aggregate assets for each tag combination
aids = alt.aid.to_numpy()
for kids in aggids[:, aids]:
fast_agg(eids + U64(kids), values, correl, li, loss2)
[docs]def ebr_from_gmfs(slice_by_event, oqparam, dstore, monitor):
"""
:param slice_by_event: composite array with fields 'start', 'stop'
:param oqparam: OqParam instance
:param dstore: DataStore instance from which to read the GMFs
:param monitor: a Monitor instance
:yields: dictionary of arrays, the output of event_based_risk
"""
if dstore.parent:
dstore.parent.open('r')
gmfcols = oqparam.gmf_data_dt().names
with monitor('reading crmodel', measuremem=True):
crmodel = monitor.read('crmodel')
risk_sids = monitor.read('sids') # this is fast
R = 1 if oqparam.collect_rlzs else len(monitor.read('weights'))
X = len(oqparam.ext_loss_types) + oqparam.ideduc
xtypes = oqparam.ext_loss_types
if oqparam.ideduc:
xtypes.append('claim')
assdic = read_assdic(slice(None), monitor)
loss3 = {'aids': [], 'bids': [], 'loss': []}
for sbe in split(slice_by_event, int(config.memory.max_gmvs_chunk)):
loss2 = general.AccumDict(accum=numpy.zeros((X, 2))) # u8idx->array
s0, s1 = sbe[0]['start'], sbe[-1]['stop']
with dstore:
haz_sids = dstore['gmf_data/sid'][s0:s1]
idx, = numpy.where(numpy.isin(haz_sids, risk_sids))
if len(idx) == 0:
yield {}
continue
with dstore, monitor('reading GMFs', measuremem=True):
start, stop = idx.min(), idx.max() + 1
gmfdic = {}
for col in gmfcols:
if col == 'sid':
gmfdic[col] = haz_sids[idx]
else:
dset = dstore['gmf_data/' + col]
gmfdic[col] = dset[s0+start:s0+stop][idx - start]
gmfdf = pandas.DataFrame(gmfdic) # few MB
dic = _event_based_risk(gmfdf, assdic, loss2, loss3, crmodel, monitor)
dic['alt'] = build_alt(loss2, xtypes)
yield dic
yield dict(avg=build_avg(loss3, oqparam.A, R*X))
[docs]def read_assdic(slc, monitor):
"""
:param slc: slice object
:returns: dictionary col->array
"""
with monitor('reading assets', measuremem=True):
# saving a lot of memory with respect to a simple read('assets')
cols = monitor.read_pdcolumns('assets').split()
assdic = {col: monitor.read('assets/' + col, slc) for col in cols}
return assdic
def _event_based_risk(df, assdic, loss2, loss3, crmodel, monitor):
if os.environ.get('OQ_DEBUG_SITE'):
print(df)
oq = crmodel.oqparam
aggids = monitor.read('aggids')
rlz_id = monitor.read('rlz_id')
xtypes = oq.ext_loss_types
if oq.ideduc:
xtypes.append('claim')
if oq.ignore_master_seed or oq.ignore_covs:
rng = None
else:
rng = MultiEventRNG(oq.master_seed, df.eid.unique(),
int(oq.asset_correlation))
outgen = output_gen(df, assdic, crmodel, rng, monitor)
with monitor('aggregating losses', measuremem=True) as agg_mon:
aggreg(outgen, loss2, loss3, crmodel, oq.K,
aggids, rlz_id, xtypes, monitor)
agg_mon.duration -= monitor.ctime # subtract the computing time
return dict(gmf_bytes=df.memory_usage().sum())
[docs]def output_gen(df, assdic, crmodel, rng, monitor):
"""
:param df: GMF dataframe (a slice of events)
:param assdic: a dictionary column name -> array
:param crmodel: CompositeRiskModel instance
:param rng: random number generator
:param monitor: Monitor instance
:yields: one output per taxonomy and slice of events
"""
risk_mon = monitor('computing risk', measuremem=False)
fil_mon = monitor('filtering GMFs', measuremem=False)
sids = df.sid.to_numpy()
monitor.ctime = 0
for id0taxo, s0, s1 in monitor.read('start-stop'):
adf = pandas.DataFrame({col: assdic[col][s0:s1] for col in assdic})
adf = adf.set_index('ordinal')
country = crmodel.countries[id0taxo // TWO24]
with fil_mon:
# *crucial* for the performance of the next step
gmf_df = df[numpy.isin(sids, adf.site_id.unique())]
if len(gmf_df) == 0: # common enough
continue
with risk_mon:
[out] = crmodel.get_outputs(
adf, gmf_df, crmodel.oqparam._sec_losses, rng, country)
monitor.ctime += fil_mon.dt + risk_mon.dt
yield out
def _tot_loss_unit_consistency(units, total_losses, loss_types):
total_losses_units = set()
for separate_lt in total_losses.split('+'):
assert separate_lt in loss_types, (separate_lt, loss_types)
for unit, lt in zip(units, loss_types):
if separate_lt == lt:
total_losses_units.add(unit)
if len(total_losses_units) != 1:
logging.warning(
'The units of the single components of the total losses'
' are not homogeneous: %s" ' % total_losses_units)
[docs]def set_oqparam(oq, assetcol, dstore):
"""
Set the attributes .M, .K, .A, .ideduc, ._sec_losses
"""
try:
K = len(dstore['agg_keys'])
except KeyError:
K = 0
sec_losses = [] # one insured loss for each loss type with a policy
try:
policy_df = dstore.read_df('policy')
except KeyError:
pass
else:
if 'reinsurance' not in oq.inputs:
sec_losses.append(
partial(insurance_losses, policy_df=policy_df))
ideduc = assetcol['ideductible'].any()
cc = dstore['exposure'].cost_calculator
if oq.total_losses and oq.total_loss_types and cc.cost_types:
# cc.cost_types is empty in scenario_damage/case_21 (consequences)
units = cc.get_units(oq.total_loss_types)
_tot_loss_unit_consistency(
units.split(), oq.total_losses, oq.total_loss_types)
sec_losses.append(
partial(total_losses, kind=oq.total_losses, ideduc=ideduc))
elif ideduc:
# subtract the insurance deductible for a single loss_type
[lt] = oq.loss_types
sec_losses.append(partial(total_losses, kind=lt, ideduc=ideduc))
oq._sec_losses = sec_losses
oq.ideduc = int(ideduc)
oq.M = len(oq.all_imts())
oq.K = K
oq.A = assetcol['ordinal'].max() + 1
def _expand3(arrayN3, maxsize):
# expand array with rows (id0taxo, start, stop) in chunks under
# maxsize
out = []
for idx, start, stop in arrayN3:
for slc in general.gen_slices(start, stop, maxsize):
out.append((idx, slc.start, slc.stop))
return U32(out)
[docs]def ebrisk(proxies, cmaker, sitecol, stations, dstore, monitor):
"""
:param proxies: list of RuptureProxies with the same trt_smr
:param cmaker: ContextMaker instance associated to the trt_smr
:param stations: empty pair or (station_data, station_sitecol)
:param monitor: a Monitor instance
:returns: a dictionary of arrays
"""
oq = cmaker.oq
oq.ground_motion_fields = True
weights = [1] if oq.collect_rlzs else monitor.read('weights')
xtypes = oq.ext_loss_types
if oq.ideduc:
xtypes.append('claim')
X = len(xtypes)
R = len(weights)
with monitor('reading crmodel', measuremem=True):
crmodel = monitor.read('crmodel')
assdic = read_assdic(slice(None), monitor)
for block in general.block_splitter(
proxies, 20_000, event_based.rup_weight):
for dic in event_based.event_based(
block, cmaker, sitecol, stations, dstore, monitor):
if len(dic['gmfdata']):
gmf_df = pandas.DataFrame(dic['gmfdata'])
loss3 = {'aids': [], 'bids': [], 'loss': []}
loss2 = general.AccumDict(accum=numpy.zeros((X, 2)))
dic = _event_based_risk(
gmf_df, assdic, loss2, loss3, crmodel, monitor)
dic['avg'] = build_avg(loss3, oq.A, R*X)
dic['alt'] = build_alt(loss2, xtypes)
yield dic
[docs]@performance.compile("(f4[:,:,:], i4[:], i4[:], f4[:], i8)")
def fast_add(avg_losses, row, col, data, X):
for r, c, d in zip(row, col, data):
rlz, lti = divmod(c, X)
avg_losses[r, lti, rlz] += d
[docs]@base.calculators.add('ebrisk', 'scenario_risk', 'event_based_risk')
class EventBasedRiskCalculator(event_based.EventBasedCalculator):
"""
Event based risk calculator generating event loss tables
"""
core_task = ebrisk
is_stochastic = True
precalc = 'event_based'
accept_precalc = ['scenario', 'event_based', 'event_based_risk', 'ebrisk']
[docs] def save_tmp(self, monitor):
"""
Save some useful data in the file calc_XXX_tmp.hdf5
"""
oq = self.oqparam
monitor.save('sids', self.sitecol.sids)
adf = self.assetcol.to_dframe()
del adf['id']
if 'ID_0' not in adf.columns:
adf['ID_0'] = 0
adf = adf.sort_values(['ID_0', 'taxonomy', 'ordinal'])
# NB: this is subtle! without the ordering by 'ordinal'
# the asset dataframe will be ordered differently on AMD machines
# with respect to Intel machines, depending on the machine, thus
# causing different losses
monitor.save('assets', adf)
if 'ID_0' in self.assetcol.tagnames:
self.crmodel.countries = self.assetcol.tagcol.ID_0
else:
self.crmodel.countries = ['?']
# storing start-stop indices in a smart way, so that the assets are
# read from the workers by taxonomy
id0taxo = TWO24 * adf.ID_0.to_numpy() + adf.taxonomy.to_numpy()
max_assets = int(config.memory.max_assets_chunk)
tss = _expand3(performance.idx_start_stop(id0taxo), max_assets)
monitor.save('start-stop', tss)
monitor.save('crmodel', self.crmodel)
monitor.save('rlz_id', self.rlzs)
monitor.save('weights', self.datastore['weights'][:])
if oq.K:
aggids, _ = self.assetcol.build_aggids(oq.aggregate_by)
else:
aggids = ()
monitor.save('aggids', aggids)
[docs] def pre_execute(self):
oq = self.oqparam
if oq.calculation_mode == 'ebrisk':
oq.ground_motion_fields = False
parent = self.datastore.parent
if parent:
self.datastore['full_lt'] = parent['full_lt']
self.parent_events = ne = len(parent['events'])
logging.info('There are %d ruptures and %d events',
len(parent['ruptures']), ne)
else:
self.parent_events = None
super().pre_execute()
parentdir = (os.path.dirname(self.datastore.ppath)
if self.datastore.ppath else None)
oq.hdf5path = self.datastore.filename
oq.parentdir = parentdir
logging.info(
'There are {:_d} ruptures and {:_d} events'.format(
len(self.datastore['ruptures']),
len(self.datastore['events'])))
self.events_per_sid = numpy.zeros(self.N, U32)
self.datastore.swmr_on()
set_oqparam(oq, self.assetcol, self.datastore)
self.A = A = len(self.assetcol)
self.L = L = len(oq.loss_types)
self.xtypes = oq.ext_loss_types
if self.assetcol['ideductible'].any():
self.xtypes.append('claim')
self.X = len(self.xtypes)
ELT = len(oq.ext_loss_types)
if oq.calculation_mode == 'event_based_risk' and oq.avg_losses:
R = 1 if oq.collect_rlzs else self.R
logging.info('Transfering %s * %d per task in avg_losses',
general.humansize(A * 8 * R), ELT)
if A * ELT * 8 > int(config.memory.avg_losses_max):
raise ValueError('For large exposures you must set '
'avg_losses=false')
elif A * ELT * self.R * 8 > int(config.memory.avg_losses_max):
raise ValueError('For large exposures you must set '
'collect_rlzs = true')
if (oq.aggregate_by and self.E * A > oq.max_potential_gmfs and
all(val == 0 for val in oq.minimum_asset_loss.values())):
logging.warning('The calculation is really big; consider setting '
'minimum_asset_loss')
base.create_risk_by_event(self)
self.rlzs = self.datastore['events']['rlz_id']
self.num_events = numpy.bincount(self.rlzs, minlength=self.R)
if oq.avg_losses:
self.create_avg_losses()
alt_nbytes = 4 * self.E * L
if alt_nbytes / (oq.concurrent_tasks or 1) > TWO32:
raise RuntimeError('The risk_by_event is too big to be transfer'
'ed with %d tasks' % oq.concurrent_tasks)
[docs] def create_avg_losses(self):
oq = self.oqparam
ws = self.datastore['weights']
R = 1 if oq.collect_rlzs else len(ws)
S = len(oq.hazard_stats())
fix_investigation_time(oq, self.datastore)
if oq.collect_rlzs:
if oq.investigation_time: # event_based
self.avg_ratio = numpy.array([oq.time_ratio / len(ws)])
else: # scenario
self.avg_ratio = numpy.array([1. / self.num_events.sum()])
else:
if oq.investigation_time: # event_based
self.avg_ratio = numpy.array([oq.time_ratio] * len(ws))
else: # scenario
self.avg_ratio = 1. / self.num_events
self.avg_losses = numpy.zeros((self.A, self.X, R), F32)
for lt in self.xtypes:
self.datastore.create_dset(
'avg_losses-rlzs/' + lt, F32, (self.A, R))
if S and R > 1:
self.datastore.create_dset(
'avg_losses-stats/' + lt, F32, (self.A, S))
[docs] def execute(self):
"""
Compute risk from GMFs or ruptures depending on what is stored
"""
oq = self.oqparam
self.gmf_bytes = 0
if oq.calculation_mode == 'ebrisk' or 'gmf_data' not in self.datastore:
# start from ruptures
if (oq.ground_motion_fields and
'gsim_logic_tree' not in oq.inputs and
oq.gsim == '[FromFile]'):
raise InvalidFile('%s: missing gsim or gsim_logic_tree_file'
% oq.inputs['job_ini'])
elif not hasattr(oq, 'maximum_distance'):
raise InvalidFile('Missing maximum_distance in %s'
% oq.inputs['job_ini'])
full_lt = self.datastore['full_lt']
smap = event_based.starmap_from_rups(
ebrisk, oq, full_lt, self.sitecol, self.datastore,
self.save_tmp)
smap.reduce(self.agg_dicts)
if self.gmf_bytes == 0:
raise RuntimeError(
'No GMFs were generated, perhaps they were '
'all below the minimum_intensity threshold')
logging.info(
'Produced %s of GMFs', general.humansize(self.gmf_bytes))
else: # start from GMFs
smap = starmap_from_gmfs(ebr_from_gmfs, oq, self.datastore,
self._monitor)
self.save_tmp(smap.monitor)
smap.reduce(self.agg_dicts)
if self.parent_events:
assert self.parent_events == len(self.datastore['events'])
return 1
[docs] def log_info(self, eids):
"""
Printing some information about the risk calculation
"""
logging.info('Processing {:_d} rows of gmf_data'.format(len(eids)))
E = len(numpy.unique(eids))
K = self.oqparam.K
logging.info('Risk parameters (rel_E={:_d}, K={:_d}, X={})'.
format(E, K, self.X))
[docs] def agg_dicts(self, dummy, dic):
"""
:param dummy: unused parameter
:param dic: dictionary with keys "avg", "alt"
"""
if not dic:
return
self.gmf_bytes += dic.pop('gmf_bytes', 0)
self.oqparam.ground_motion_fields = False # hack
if 'alt' in dic:
with self.monitor('saving risk_by_event'):
alt = dic.pop('alt')
for name in alt.columns:
dset = self.datastore['risk_by_event/' + name]
hdf5.extend(dset, alt[name].to_numpy())
if self.oqparam.avg_losses and 'avg' in dic:
# avg_losses are stored as coo matrices
with self.monitor('saving avg_losses'):
coo = dic.pop('avg')
# the non-numpy approach is 2-3x slower, causing
# a big data queue and then running out of memory
# rlzs, xlts = numpy.divmod(coo.col, self.X)
# self.avg_losses[coo.row, xlts, rlzs] += coo.data
fast_add(self.avg_losses, coo.row, coo.col, coo.data, self.X)
[docs] def post_execute(self, dummy):
"""
Compute and store average losses from the risk_by_event dataset,
and then loss curves and maps.
"""
oq = self.oqparam
K = self.datastore['risk_by_event'].attrs.get('K', 0)
upper_limit = self.E * (K + 1) * len(self.xtypes)
if upper_limit < 1E7:
# sanity check on risk_by_event if not too large
alt = self.datastore.read_df('risk_by_event')
size = len(alt)
assert size <= upper_limit, (size, upper_limit)
# sanity check on uniqueness by (agg_id, loss_id, event_id)
arr = alt[['agg_id', 'loss_id', 'event_id']].to_numpy()
uni, cnt = numpy.unique(arr, axis=0, return_counts=True)
if len(uni) < len(arr):
dupl = uni[cnt > 1] # (agg_id, loss_id, event_id)
raise RuntimeError(
'risk_by_event contains %d duplicates for event %s' %
(len(arr) - len(uni), dupl[0, 2]))
s = oq.hazard_stats()
if s:
_statnames, statfuncs = zip(*s.items())
weights = self.datastore['weights'][:]
if oq.avg_losses:
R = self.avg_losses.shape[2]
for li, lt in enumerate(self.xtypes):
al = self.avg_losses[:, li] # shape (A, R)
for r in range(R):
al[:, r] *= self.avg_ratio[r]
name = 'avg_losses-rlzs/' + lt
logging.info(f'Storing {name}')
self.datastore[name][:] = al
if s and self.R > 1:
self.datastore[name.replace('-rlzs', '-stats')][:] = \
stats.compute_stats2(al, statfuncs, weights)
self.build_aggcurves()
if oq.reaggregate_by:
post_aggregate(self.datastore.calc_id,
','.join(oq.reaggregate_by))
[docs] def build_aggcurves(self):
prc = PostRiskCalculator(self.oqparam, self.datastore.calc_id)
prc.assetcol = self.assetcol
if hasattr(self, 'exported'):
prc.exported = self.exported
prc.pre_execute()
res = prc.execute()
prc.post_execute(res)
