def search(self, queries, topk=None, **kwargs):
nq = queries.shape[0]
dsub = self.encoder.ecdat['dsub']
nsubq = self.encoder.ecdat['nsubq']
ksub = self.encoder.ecdat['ksub']
centroids = self.encoder.ecdat['centroids']
distab = np.zeros((nsubq, ksub), np.single)
dis = np.ones((nq, topk), np.single) * np.inf
ids = np.ones((nq, topk), np.int32) * -1
profiler = Profiler()
interval = 100 if nq >= 100 else 10
time_total = 0.0 # total time for all queries
logging.info('Start Querying ...')
for qry_id in xrange(nq):
profiler.start("distab") # time for computing distance table
# pre-compute the table of squared distance to centroids
for qnt_id in range(nsubq):
vsub = queries[qry_id:qry_id + 1,
qnt_id * dsub:(qnt_id + 1) * dsub]
distab[qnt_id:qnt_id + 1, :] = distFunc['euclidean'](
centroids[qnt_id], vsub)
profiler.end()
profiler.start("distance") # time for computing the distances
# add the tabulated distances to construct the distance estimators
idsquerybase, disquerybase = self.sumidxtab(distab)
profiler.end()
profiler.start("knn") # time for finding the kNN
realk = min(disquerybase.shape[0], topk)
cur_ids = pq_knn(disquerybase, realk)
profiler.end()
profiler.start("result") # time for getting final result
ids[qry_id, :] = idsquerybase[cur_ids]
dis[qry_id, :] = disquerybase[cur_ids]
profiler.end()
if (qry_id + 1) % interval == 0:
time_total += profiler.sum_overall()
logging.info('\t%d/%d: %.3fms per query' %
(qry_id + 1, nq, profiler.sum_average() * 1000))
logging.info("\t\t%s" % profiler.str_average())
profiler.reset()
logging.info('Querying Finished!')
time_total += profiler.sum_overall()
logging.info("Average querying time: %.3fms" %
(time_total * 1000 / nq))
return ids, dis
def search(self, queries, topk=None, **kwargs):
nq = queries.shape[0]
dsub = self.encoder.ecdat['dsub']
nsubq = self.encoder.ecdat['nsubq']
ksub = self.encoder.ecdat['ksub']
centroids = self.encoder.ecdat['centroids']
distab = np.zeros((nsubq, ksub), np.single)
dis = np.ones((nq, topk), np.single) * np.inf
ids = np.ones((nq, topk), np.int32) * -1
profiler = Profiler()
interval = 100 if nq >= 100 else 10
time_total = 0.0 # total time for all queries
logging.info('Start Querying ...')
for qry_id in xrange(nq):
profiler.start("distab") # time for computing distance table
# pre-compute the table of squared distance to centroids
for qnt_id in range(nsubq):
vsub = queries[qry_id:qry_id+1, qnt_id*dsub:(qnt_id+1)*dsub]
distab[qnt_id:qnt_id+1, :] = distFunc['euclidean'](
centroids[qnt_id], vsub)
profiler.end()
profiler.start("distance") # time for computing the distances
# add the tabulated distances to construct the distance estimators
idsquerybase, disquerybase = self.sumidxtab(distab)
profiler.end()
profiler.start("knn") # time for finding the kNN
realk = min(disquerybase.shape[0], topk)
cur_ids = pq_knn(disquerybase, realk)
profiler.end()
profiler.start("result") # time for getting final result
ids[qry_id, :] = idsquerybase[cur_ids]
dis[qry_id, :] = disquerybase[cur_ids]
profiler.end()
if (qry_id+1) % interval == 0:
time_total += profiler.sum_overall()
logging.info(
'\t%d/%d: %.3fms per query' %
(qry_id+1, nq, profiler.sum_average() * 1000))
logging.info("\t\t%s" % profiler.str_average())
profiler.reset()
logging.info('Querying Finished!')
time_total += profiler.sum_overall()
logging.info("Average querying time: %.3fms" % (time_total * 1000 / nq))
return ids, dis
def search(self, queries, topk=None, **kwargs):
nn_coa = kwargs.get("nn_coa", 8)
nq = queries.shape[0]
dsub = self.encoder.ecdat["dsub"]
nsubq = self.encoder.ecdat["nsubq"]
ksub = self.encoder.ecdat["ksub"]
centroids = self.encoder.ecdat["centroids"]
coa_centroids = self.encoder.ecdat["coa_centroids"]
centroids_l2norm = []
for i in xrange(nsubq):
centroids_l2norm.append((centroids[i] ** 2).sum(1))
coa_centroids_l2norm = (coa_centroids ** 2).sum(1)
distab = np.zeros((nsubq, ksub), np.single)
dis = np.ones((nq, topk), np.single) * np.inf
ids = np.ones((nq, topk), np.int32) * -1
profiler = Profiler()
interval = 100 if nq >= 100 else 10
time_total = 0.0 # total time for all queries
logging.info("Start Querying ...")
for qry_id in xrange(nq):
# Here `copy()` can ensure that you DONOT modify the queries
query = queries[qry_id : qry_id + 1, :].copy()
profiler.start("coa_knn")
coa_dist = fast_euclidean(coa_centroids, query, coa_centroids_l2norm).reshape(-1)
# profiler.end()
# profiler.start("coa_knn_knn")
coa_knn = pq_knn(coa_dist, nn_coa)
profiler.end()
profiler.start("distab+distance")
query = query - coa_centroids[coa_knn, :]
v_idsquerybase = []
v_disquerybase = []
for coa_idx in range(nn_coa):
# pre-compute the table of squared distance to centroids
for qnt_id in range(nsubq):
vsub = query[coa_idx : coa_idx + 1, qnt_id * dsub : (qnt_id + 1) * dsub]
distab[qnt_id : qnt_id + 1, :] = fast_euclidean(centroids[qnt_id], vsub, centroids_l2norm[qnt_id])
# construct the distance estimators from tabulated distances
idsquerybase, disquerybase = self.sumidxtab(distab, coa_knn[coa_idx])
v_idsquerybase.append(idsquerybase)
v_disquerybase.append(disquerybase)
idsquerybase = np.hstack(tuple(v_idsquerybase))
disquerybase = np.hstack(tuple(v_disquerybase))
profiler.end()
profiler.start("knn") # time for finding the kNN
realk = min(disquerybase.shape[0], topk)
cur_ids = pq_knn(disquerybase, realk)
profiler.end()
profiler.start("result") # time for getting final result
ids[qry_id, :realk] = idsquerybase[cur_ids]
dis[qry_id, :realk] = disquerybase[cur_ids]
profiler.end()
if (qry_id + 1) % interval == 0:
time_total += profiler.sum_overall()
logging.info("\t%d/%d: %.3fms per query" % (qry_id + 1, nq, profiler.sum_average() * 1000))
logging.info("\t\t%s" % profiler.str_average())
profiler.reset()
logging.info("Querying Finished!")
time_total += profiler.sum_overall()
logging.info("Average querying time: %.3fms" % (time_total * 1000 / nq))
return ids, dis
def search(self, queries, topk=None, **kwargs):
nn_coa = kwargs.get('nn_coa', 8)
nq = queries.shape[0]
dsub = self.encoder.ecdat['dsub']
nsubq = self.encoder.ecdat['nsubq']
ksub = self.encoder.ecdat['ksub']
centroids = self.encoder.ecdat['centroids']
coa_centroids = self.encoder.ecdat['coa_centroids']
centroids_l2norm = []
for i in xrange(nsubq):
centroids_l2norm.append((centroids[i]**2).sum(1))
coa_centroids_l2norm = (coa_centroids**2).sum(1)
distab = np.zeros((nsubq, ksub), np.single)
dis = np.ones((nq, topk), np.single) * np.inf
ids = np.ones((nq, topk), np.int32) * -1
profiler = Profiler()
interval = 100 if nq >= 100 else 10
time_total = 0.0 # total time for all queries
logging.info('Start Querying ...')
for qry_id in xrange(nq):
# Here `copy()` can ensure that you DONOT modify the queries
query = queries[qry_id:qry_id + 1, :].copy()
profiler.start("coa_knn")
coa_dist = fast_euclidean(coa_centroids, query,
coa_centroids_l2norm).reshape(-1)
# profiler.end()
# profiler.start("coa_knn_knn")
coa_knn = pq_knn(coa_dist, nn_coa)
profiler.end()
profiler.start("distab+distance")
query = query - coa_centroids[coa_knn, :]
v_idsquerybase = []
v_disquerybase = []
for coa_idx in range(nn_coa):
# pre-compute the table of squared distance to centroids
for qnt_id in range(nsubq):
vsub = query[coa_idx:coa_idx + 1,
qnt_id * dsub:(qnt_id + 1) * dsub]
distab[qnt_id:qnt_id + 1, :] = fast_euclidean(
centroids[qnt_id], vsub, centroids_l2norm[qnt_id])
# construct the distance estimators from tabulated distances
idsquerybase, disquerybase = self.sumidxtab(
distab, coa_knn[coa_idx])
v_idsquerybase.append(idsquerybase)
v_disquerybase.append(disquerybase)
idsquerybase = np.hstack(tuple(v_idsquerybase))
disquerybase = np.hstack(tuple(v_disquerybase))
profiler.end()
profiler.start("knn") # time for finding the kNN
realk = min(disquerybase.shape[0], topk)
cur_ids = pq_knn(disquerybase, realk)
profiler.end()
profiler.start("result") # time for getting final result
ids[qry_id, :realk] = idsquerybase[cur_ids]
dis[qry_id, :realk] = disquerybase[cur_ids]
profiler.end()
if (qry_id + 1) % interval == 0:
time_total += profiler.sum_overall()
logging.info('\t%d/%d: %.3fms per query' %
(qry_id + 1, nq, profiler.sum_average() * 1000))
logging.info("\t\t%s" % profiler.str_average())
profiler.reset()
logging.info('Querying Finished!')
time_total += profiler.sum_overall()
logging.info("Average querying time: %.3fms" %
(time_total * 1000 / nq))
return ids, dis