def time_measurement_for_ivpq_batch(con, cur, search_parameters, names, query,
k, param_query, parameter_variables,
num_queries, num_targets):
all_execution_times = defaultdict(list)
all_inner_times = defaultdict(list)
count = 0
data_size = ev.get_vector_dataset_size(cur)
for i, search_params in enumerate(search_parameters):
set_search_params(con, cur, search_params)
for elem in parameter_variables:
# set parameter variable
cur.execute(param_query.format(elem))
con.commit()
times = []
inner_times = []
for iteration in range(NUM_ITERATIONS):
samples = ev.get_samples(con, cur, num_queries, data_size)
target_samples = ev.get_samples(con, cur, num_targets,
data_size)
params = [(add_escapes(samples), k,
add_escapes(target_samples))]
trackings, execution_times = ev.create_track_statistics(
cur, con, query, params, log=False)
print(param_query.format(elem), search_params, execution_times)
times.append(execution_times)
inner_times.append(float(trackings[0]['total_time'][0][0]))
count += 1
print(str(
round((count * 100) /
(NUM_ITERATIONS * len(parameter_variables) *
len(search_parameters)), 2)) + '%',
end='\r')
all_execution_times[names[i]].append(mean(times))
all_inner_times[names[i]].append(mean(inner_times))
return all_execution_times, all_inner_times
# set parameters
cur.execute('SELECT set_pvf({:d});'.format(pvf))
cur.execute('SELECT set_alpha({:d});'.format(alpha))
cur.execute('SELECT set_method_flag({:d});'.format(method))
cur.execute("SELECT set_use_targetlist('{!s}');".format(use_target_list))
con.commit()
dynamic_sizes = list([i * step_size for i in range(1, 11)])
query = 'SELECT * FROM knn_in_iv_batch(' '{!s}' ', {:d}, ' '{!s}' ', \'ivpq_search_in\') LIMIT 1;'
convert = lambda x: "'{" + ",".join([
s.replace("'", "''").replace("\"", "\\\"").replace("{", "\}").replace(
"{", "\}").replace(",", "\,") for s in x
]) + "}'"
get_samples = lambda x: ev.get_samples(con, cur, x, data_size)
all_trackings = [defaultdict(list) for i in dynamic_sizes]
for iteration in range(num_iters):
if variable_parameter == 'query_size':
params = [(convert(get_samples(i)), k,
convert(get_samples(number_of_target_samples)))
for i in dynamic_sizes]
if variable_parameter == 'target_size':
params = [(convert(get_samples(number_of_query_samples)), k,
convert(get_samples(i))) for i in dynamic_sizes]
trackings, execution_times = ev.create_track_statistics(
cur, con, query, params)
for i in range(len(trackings)):
for key in trackings[i].keys():
all_trackings[i][key].append(trackings[i][key])
k = config['k']
alpha = config['alpha_start']
method_flag = config['method']
step_size = config['step_size']
query_template = 'SELECT query, target FROM knn_in_ivpq_batch(' '{!s}' ', {:d}, ' '{!s}' ');'
ev.set_vec_table_name(config['vecs_table_name'])
search_params = {'pvf': 1, 'alpha': alpha, 'method': method_flag}
ivpqEv.set_search_params(con, cur, search_params)
number_of_vectors = ev.get_vector_dataset_size(cur)
query_samples = ev.get_samples(con, cur, number_of_query_samples,
number_of_vectors)
target_samples = ev.get_samples(con, cur, number_of_target_samples,
number_of_vectors)
prediction = []
real = []
divergence = []
divergence_relative = []
for i in range(num_iters):
params = [(ivpqEv.add_escapes([x]), k, ivpqEv.add_escapes(target_samples))
for x in query_samples]
trackings, _ = ev.create_track_statistics(cur, con, query_template, params)
prediction += ([float(t['target_count'][0][0]) for t in trackings])
real += ([float(t['retrieved'][0][0]) for t in trackings])
print([float(t['target_count'][0][0]) for t in trackings])
# divergence +=
def time_and_precision_measurement_for_ivpq_batch(con,
cur,
search_parameters,
names,
query,
k,
param_query,
parameter_variables,
num_queries,
num_targets,
small_sample_size,
outlier_detect=0):
USE_MEDIAN = True
all_execution_times = defaultdict(list)
all_inner_times = defaultdict(list)
all_precision_values = defaultdict(list)
count = 0
data_size = data_size = ev.get_vector_dataset_size(cur)
# init phase
for i, search_params in enumerate(search_parameters):
all_execution_times[names[i]] = [
[] for j in range(len(parameter_variables[i]))
]
all_inner_times[names[i]] = [
[] for j in range(len(parameter_variables[i]))
]
all_precision_values[names[i]] = [
[] for j in range(len(parameter_variables[i]))
]
# measurement phase
for iteration in range(NUM_ITERATIONS):
print('Start Iteration', iteration)
for i, search_params in enumerate(search_parameters):
for j, elem in enumerate(parameter_variables[i]):
# TODO set parameter variable
cur.execute(param_query.format(elem))
con.commit()
times = all_execution_times[names[i]][j]
inner_times = all_inner_times[names[i]][j]
precision_values = all_precision_values[names[i]][j]
# big sample set
samples = ev.get_samples(con, cur, num_queries, data_size)
# create smaller sample set (bootstraping)
small_samples = [
samples[random.randint(0, num_queries - 1)]
for i in range(small_sample_size)
]
target_samples = ev.get_samples(con, cur, num_targets,
data_size)
# calculate exact results
start_time = time.time()
exact_results = get_exact_results(cur, con,
add_escapes(small_samples),
k,
add_escapes(target_samples))
print("--- %s seconds ---" % (time.time() - start_time))
set_search_params(con, cur, search_params)
cur.execute(param_query.format(elem))
con.commit()
params = [('iv' if names[i] != 'Baseline' else '',
add_escapes(samples), k,
add_escapes(target_samples))]
trackings, execution_times = ev.create_track_statistics(
cur, con, query, params, log=False)
times.append(execution_times)
print(names[i], search_params, elem, "arguments:",
len(samples), params[0][2], len(target_samples),
float(trackings[0]['total_time'][0][0]))
inner_times.append(float(trackings[0]['total_time'][0][0]))
# execute approximated query to obtain results
cur.execute(query.format(*(params[0])))
approximated_results = defaultdict(list)
for res in cur.fetchall():
approximated_results[res[0]].append(res[1])
precision_values.append(
calculate_precision(exact_results, approximated_results,
k))
count += 1
print(str(
round((count * 100) /
(NUM_ITERATIONS *
sum([len(p)
for p in parameter_variables])), 2)) + '%',
end='\r')
# evaluation phase
raw_data = {
'execution_times': copy.deepcopy(all_execution_times),
'inner_times': copy.deepcopy(all_inner_times),
'precision_values': copy.deepcopy(all_precision_values)
}
for i, search_params in enumerate(search_parameters):
for j, elem in enumerate(parameter_variables[i]):
if outlier_detect:
all_execution_times[names[i]][j] = mean([
v for v in all_execution_times[names[i]][j]
if not is_outlier(v, all_execution_times[names[i]][j])
])
all_inner_times[names[i]][j] = mean([
v for v in all_inner_times[names[i]][j]
if not is_outlier(v, all_inner_times[names[i]][j])
])
else:
if USE_MEDIAN:
all_execution_times[names[i]][j] = median(
all_execution_times[names[i]][j])
all_inner_times[names[i]][j] = median(
all_inner_times[names[i]][j])
else:
all_execution_times[names[i]][j] = mean(
all_execution_times[names[i]][j])
all_inner_times[names[i]][j] = mean(
all_inner_times[names[i]][j])
all_precision_values[names[i]][j] = mean(
all_precision_values[names[i]][j])
return all_execution_times, all_inner_times, all_precision_values, raw_data