def fit(self, X):
if self._method_name == 'vptree':
# To avoid this issue:
# terminate called after throwing an instance of 'std::runtime_error'
# what(): The data size is too small or the bucket size is too big. Select the parameters so that <total # of records> is NOT less than <bucket size> * 1000
# Aborted (core dumped)
self._index_param.append('bucketSize=%d' %
min(int(X.shape[0] * 0.0005), 1000))
self._index = nmslib.init(self._nmslib_metric, [], self._method_name,
nmslib.DataType.DENSE_VECTOR,
nmslib.DistType.FLOAT)
for i, x in enumerate(X):
nmslib.addDataPoint(self._index, i, x.tolist())
if os.path.exists(self._index_name):
print "Loading index from file"
nmslib.loadIndex(self._index, self._index_name)
else:
nmslib.createIndex(self._index, self._index_param)
if self._save_index:
nmslib.saveIndex(self._index, self._index_name)
nmslib.setQueryTimeParams(self._index, self._query_param)
def test_string_fresh(batch=True):
DATA_STRS = ["xyz", "beagcfa", "cea", "cb",
"d", "c", "bdaf", "ddcd",
"egbfa", "a", "fba", "bcccfe",
"ab", "bfgbfdc", "bcbbgf", "bfbb"
]
QUERY_STRS = ["abc", "def", "ghik"]
space_type = 'leven'
space_param = []
method_name = 'small_world_rand'
index_name = method_name + '.index'
index = nmslib.init(
space_type,
space_param,
method_name,
nmslib.DataType.OBJECT_AS_STRING,
nmslib.DistType.INT)
if batch:
print 'DATA_STRS', DATA_STRS
positions = nmslib.addDataPointBatch(index, np.arange(len(DATA_STRS), dtype=np.int32), DATA_STRS)
else:
for id, data in enumerate(DATA_STRS):
nmslib.addDataPoint(index, id, data)
print 'Let\'s print a few data entries'
print 'We have added %d data points' % nmslib.getDataPointQty(index)
for i in range(0,min(MAX_PRINT_QTY,nmslib.getDataPointQty(index))):
print nmslib.getDataPoint(index,i)
print 'Let\'s invoke the index-build process'
index_param = ['NN=17', 'initIndexAttempts=3', 'indexThreadQty=4']
query_time_param = ['initSearchAttempts=3']
nmslib.createIndex(index, index_param)
nmslib.setQueryTimeParams(index, query_time_param)
print 'Query time parameters are set'
print "Results for the freshly created index:"
k = 2
if batch:
num_threads = 10
res = nmslib.knnQueryBatch(index, num_threads, k, QUERY_STRS)
for idx, data in enumerate(QUERY_STRS):
res = nmslib.knnQuery(index, k, data)
print idx, data, res, [DATA_STRS[i] for i in res]
nmslib.saveIndex(index, index_name)
print "The index %s is saved" % index_name
nmslib.freeIndex(index)
def test_string_fresh(batch=True):
DATA_STRS = [
"xyz", "beagcfa", "cea", "cb", "d", "c", "bdaf", "ddcd", "egbfa", "a",
"fba", "bcccfe", "ab", "bfgbfdc", "bcbbgf", "bfbb"
]
QUERY_STRS = ["abc", "def", "ghik"]
space_type = 'leven'
space_param = []
method_name = 'small_world_rand'
index_name = method_name + '.index'
index = nmslib.init(space_type, space_param, method_name,
nmslib.DataType.OBJECT_AS_STRING, nmslib.DistType.INT)
if batch:
print 'DATA_STRS', DATA_STRS
positions = nmslib.addDataPointBatch(
index, np.arange(len(DATA_STRS), dtype=np.int32), DATA_STRS)
else:
for id, data in enumerate(DATA_STRS):
nmslib.addDataPoint(index, id, data)
print 'Let\'s print a few data entries'
print 'We have added %d data points' % nmslib.getDataPointQty(index)
for i in range(0, min(MAX_PRINT_QTY, nmslib.getDataPointQty(index))):
print nmslib.getDataPoint(index, i)
print 'Let\'s invoke the index-build process'
index_param = ['NN=17', 'initIndexAttempts=3', 'indexThreadQty=4']
query_time_param = ['initSearchAttempts=3']
nmslib.createIndex(index, index_param)
nmslib.setQueryTimeParams(index, query_time_param)
print 'Query time parameters are set'
print "Results for the freshly created index:"
k = 2
if batch:
num_threads = 10
res = nmslib.knnQueryBatch(index, num_threads, k, QUERY_STRS)
for idx, data in enumerate(QUERY_STRS):
res = nmslib.knnQuery(index, k, data)
print idx, data, res, [DATA_STRS[i] for i in res]
nmslib.saveIndex(index, index_name)
print "The index %s is saved" % index_name
nmslib.freeIndex(index)
def test_object_as_string_fresh(batch=True):
space_type = 'cosinesimil'
space_param = []
method_name = 'small_world_rand'
index_name = method_name + '.index'
if os.path.isfile(index_name):
os.remove(index_name)
index = nmslib.init(
space_type,
space_param,
method_name,
nmslib.DataType.OBJECT_AS_STRING,
nmslib.DistType.FLOAT)
if batch:
data = [s for s in read_data_as_string('sample_dataset.txt')]
positions = nmslib.addDataPointBatch(index, np.arange(len(data), dtype=np.int32), data)
else:
for id, data in enumerate(read_data_as_string('sample_dataset.txt')):
nmslib.addDataPoint(index, id, data)
print 'Let\'s print a few data entries'
print 'We have added %d data points' % nmslib.getDataPointQty(index)
for i in range(0,min(MAX_PRINT_QTY,nmslib.getDataPointQty(index))):
print nmslib.getDataPoint(index, i)
print 'Let\'s invoke the index-build process'
index_param = ['NN=17', 'initIndexAttempts=3', 'indexThreadQty=4']
query_time_param = ['initSearchAttempts=3']
nmslib.createIndex(index, index_param)
print 'The index is created'
nmslib.setQueryTimeParams(index,query_time_param)
print 'Query time parameters are set'
print "Results for the freshly created index:"
k = 3
for idx, data in enumerate(read_data_as_string('sample_queryset.txt')):
print idx, nmslib.knnQuery(index, k, data)
nmslib.saveIndex(index, index_name)
print "The index %s is saved" % index_name
nmslib.freeIndex(index)
def test_object_as_string_fresh(batch=True):
space_type = 'cosinesimil'
space_param = []
method_name = 'small_world_rand'
index_name = method_name + '.index'
if os.path.isfile(index_name):
os.remove(index_name)
index = nmslib.init(
space_type,
space_param,
method_name,
nmslib.DataType.OBJECT_AS_STRING,
nmslib.DistType.FLOAT)
if batch:
data = [s for s in read_data_as_string('sample_dataset.txt')]
positions = nmslib.addDataPointBatch(index, np.arange(len(data), dtype=np.int32), data)
else:
for id, data in enumerate(read_data_as_string('sample_dataset.txt')):
nmslib.addDataPoint(index, id, data)
print('Let\'s print a few data entries')
print('We have added %d data points' % nmslib.getDataPointQty(index))
for i in range(0,min(MAX_PRINT_QTY,nmslib.getDataPointQty(index))):
print(nmslib.getDataPoint(index, i))
print('Let\'s invoke the index-build process')
index_param = ['NN=17', 'efConstruction=50', 'indexThreadQty=4']
query_time_param = ['efSearch=50']
nmslib.createIndex(index, index_param)
print('The index is created')
nmslib.setQueryTimeParams(index,query_time_param)
print('Query time parameters are set')
print("Results for the freshly created index:")
k = 3
for idx, data in enumerate(read_data_as_string('sample_queryset.txt')):
print(idx, nmslib.knnQuery(index, k, data))
nmslib.saveIndex(index, index_name)
print("The index %s is saved" % index_name)
nmslib.freeIndex(index)
def test_sparse_vector_fresh():
space_type = 'cosinesimil_sparse'
space_param = []
method_name = 'small_world_rand'
index_name = method_name + '_sparse.index'
if os.path.isfile(index_name):
os.remove(index_name)
index = nmslib.init(
space_type,
space_param,
method_name,
nmslib.DataType.SPARSE_VECTOR,
nmslib.DistType.FLOAT)
for id, data in enumerate(read_sparse_data('sample_sparse_dataset.txt')):
nmslib.addDataPoint(index, id, data)
print 'We have added %d data points' % nmslib.getDataPointQty(index)
for i in range(0,min(MAX_PRINT_QTY,nmslib.getDataPointQty(index))):
print nmslib.getDataPoint(index,i)
print 'Let\'s invoke the index-build process'
index_param = ['NN=17', 'initIndexAttempts=3', 'indexThreadQty=4']
query_time_param = ['initSearchAttempts=3']
nmslib.createIndex(index, index_param)
print 'The index is created'
nmslib.setQueryTimeParams(index,query_time_param)
print 'Query time parameters are set'
print "Results for the freshly created index:"
k = 3
for idx, data in enumerate(read_sparse_data('sample_sparse_queryset.txt')):
print idx, nmslib.knnQuery(index, k, data)
nmslib.saveIndex(index, index_name)
print "The index %s is saved" % index_name
nmslib.freeIndex(index)
def test_sparse_vector_fresh():
space_type = 'cosinesimil_sparse_fast'
space_param = []
method_name = 'small_world_rand'
index_name = method_name + '_sparse.index'
if os.path.isfile(index_name):
os.remove(index_name)
index = nmslib.init(
space_type,
space_param,
method_name,
nmslib.DataType.SPARSE_VECTOR,
nmslib.DistType.FLOAT)
for id, data in enumerate(read_sparse_data('sample_sparse_dataset.txt')):
nmslib.addDataPoint(index, id, data)
print('We have added %d data points' % nmslib.getDataPointQty(index))
for i in range(0,min(MAX_PRINT_QTY,nmslib.getDataPointQty(index))):
print(nmslib.getDataPoint(index,i))
print('Let\'s invoke the index-build process')
index_param = ['NN=17', 'efConstruction=50', 'indexThreadQty=4']
query_time_param = ['efSearch=50']
nmslib.createIndex(index, index_param)
print('The index is created')
nmslib.setQueryTimeParams(index,query_time_param)
print('Query time parameters are set')
print("Results for the freshly created index:")
k = 3
for idx, data in enumerate(read_sparse_data('sample_sparse_queryset.txt')):
print(idx, nmslib.knnQuery(index, k, data))
nmslib.saveIndex(index, index_name)
print("The index %s is saved" % index_name)
nmslib.freeIndex(index)
def fit(self, X):
self._index = nmslib.init(self._metric, [], "hnsw",
nmslib.DataType.DENSE_VECTOR,
nmslib.DistType.FLOAT)
if os.path.exists(self._index_name):
logging.debug("Loading index from file")
nmslib.loadIndex(self._index, self._index_name)
else:
logging.debug("Create Index")
for i, x in enumerate(X):
self._index.addDataPoint(i, x)
nmslib.createIndex(self._index, self._index_param)
nmslib.saveIndex(self._index, self._index_name)
nmslib.setQueryTimeParams(self._index, self._query_param)
def fit(self, X):
import nmslib
self._index = nmslib.init(self._nmslib_metric, [], self._method_name,
nmslib.DataType.DENSE_VECTOR,
nmslib.DistType.FLOAT)
for i, x in enumerate(X):
nmslib.addDataPoint(self._index, i, x.tolist())
if os.path.exists(self._index_name):
logging.debug("Loading index from file")
nmslib.loadIndex(self._index, self._index_name)
else:
logging.debug("Create Index")
nmslib.createIndex(self._index, self._index_param)
if self._save_index:
nmslib.saveIndex(self._index, self._index_name)
nmslib.setQueryTimeParams(self._index, self._query_param)
def fit(self, X):
if self._method_name == 'vptree':
# To avoid this issue:
# terminate called after throwing an instance of 'std::runtime_error'
# what(): The data size is too small or the bucket size is too big. Select the parameters so that <total # of records> is NOT less than <bucket size> * 1000
# Aborted (core dumped)
self._index_param.append('bucketSize=%d' % min(int(X.shape[0] * 0.0005), 1000))
self._index = nmslib.init(self._nmslib_metric, [], self._method_name, nmslib.DataType.DENSE_VECTOR, nmslib.DistType.FLOAT)
for i, x in enumerate(X):
nmslib.addDataPoint(self._index, i, x.tolist())
if os.path.exists(self._index_name):
print "Loading index from file"
nmslib.loadIndex(self._index, self._index_name)
else:
nmslib.createIndex(self._index, self._index_param)
if self._save_index:
nmslib.saveIndex(self._index, self._index_name)
nmslib.setQueryTimeParams(self._index, self._query_param)
def bench_sparse_vector(batch=True):
dim = 20000
dataset = np.random.binomial(1, 0.01, size=(40000, dim))
queryset = np.random.binomial(1, 0.009, size=(1000, dim))
print 'dataset[0]:', [[i, v] for i, v in enumerate(dataset[0]) if v > 0]
k = 3
q0 = queryset[0]
res = []
for i in range(dataset.shape[0]):
res.append([i, distance.cosine(q0, dataset[i,:])])
res.sort(key=lambda x: x[1])
print 'q0 res', res[:k]
data_matrix = csr_matrix(dataset, dtype=np.float32)
query_matrix = csr_matrix(queryset, dtype=np.float32)
data_to_return = range(dataset.shape[0])
with TimeIt('building MultiClusterIndex'):
cp = snn.MultiClusterIndex(data_matrix, data_to_return)
with TimeIt('knn search'):
res = cp.search(query_matrix, k=k, return_distance=False)
print res[:5]
for i in res[0]:
print int(i), distance.cosine(q0, dataset[int(i),:])
#space_type = 'cosinesimil_sparse'
space_type = 'cosinesimil_sparse_fast'
space_param = []
method_name = 'small_world_rand'
index_name = method_name + '_sparse.index'
if os.path.isfile(index_name):
os.remove(index_name)
index = nmslib.init(space_type,
space_param,
method_name,
nmslib.DataType.SPARSE_VECTOR,
nmslib.DistType.FLOAT)
if batch:
with TimeIt('batch add'):
positions = nmslib.addDataPointBatch(index, np.arange(len(dataset), dtype=np.int32), data_matrix)
print 'positions', positions
else:
d = []
q = []
with TimeIt('preparing'):
for data in dataset:
d.append([[i, v] for i, v in enumerate(data) if v > 0])
for data in queryset:
q.append([[i, v] for i, v in enumerate(data) if v > 0])
with TimeIt('adding points'):
for id, data in enumerate(d):
nmslib.addDataPoint(index, id, data)
print 'Let\'s invoke the index-build process'
index_param = ['NN=17', 'initIndexAttempts=3', 'indexThreadQty=4']
query_time_param = ['initSearchAttempts=3']
with TimeIt('building index'):
nmslib.createIndex(index, index_param)
print 'The index is created'
nmslib.setQueryTimeParams(index,query_time_param)
print 'Query time parameters are set'
print "Results for the freshly created index:"
with TimeIt('knn query'):
if batch:
num_threads = 10
res = nmslib.knnQueryBatch(index, num_threads, k, query_matrix)
for idx, v in enumerate(res):
if idx < 5:
print idx, v
if idx == 0:
for i in v:
print 'q0', i, distance.cosine(q0, dataset[i,:])
else:
for idx, data in enumerate(q):
res = nmslib.knnQuery(index, k, data)
if idx < 5:
print idx, res
nmslib.saveIndex(index, index_name)
print "The index %s is saved" % index_name
nmslib.freeIndex(index)
def test_vector_fresh(fast=True):
space_type = 'cosinesimil'
space_param = []
method_name = 'small_world_rand'
index_name = method_name + '.index'
if os.path.isfile(index_name):
os.remove(index_name)
index = nmslib.init(
space_type,
space_param,
method_name,
nmslib.DataType.DENSE_VECTOR,
nmslib.DistType.FLOAT)
start = time.time()
if fast:
data = read_data_fast('sample_dataset.txt')
print 'data.shape', data.shape
positions = nmslib.addDataPointBatch(index, np.arange(len(data), dtype=np.int32), data)
else:
for id, data in enumerate(read_data('sample_dataset.txt')):
pos = nmslib.addDataPoint(index, id, data)
if id != pos:
print 'id %s != pos %s' % (id, pos)
sys.exit(1)
end = time.time()
print 'added data in %s secs' % (end - start)
print 'Let\'s print a few data entries'
print 'We have added %d data points' % nmslib.getDataPointQty(index)
for i in range(0,min(MAX_PRINT_QTY,nmslib.getDataPointQty(index))):
print nmslib.getDataPoint(index, i)
print 'Let\'s invoke the index-build process'
index_param = ['NN=17', 'initIndexAttempts=3', 'indexThreadQty=4']
query_time_param = ['initSearchAttempts=3']
nmslib.createIndex(index, index_param)
print 'The index is created'
nmslib.setQueryTimeParams(index,query_time_param)
print 'Query time parameters are set'
print "Results for the freshly created index:"
k = 3
start = time.time()
if fast:
num_threads = 10
query = read_data_fast('sample_queryset.txt')
res = nmslib.knnQueryBatch(index, num_threads, k, query)
for idx, v in enumerate(res):
print idx, v
else:
for idx, data in enumerate(read_data('sample_queryset.txt')):
print idx, nmslib.knnQuery(index, k, data)
end = time.time()
print 'querying done in %s secs' % (end - start)
nmslib.saveIndex(index, index_name)
print "The index %s is saved" % index_name
nmslib.freeIndex(index)
def bench_sparse_vector(batch=True):
# delay importing these so CI can import module
from scipy.sparse import csr_matrix
from scipy.spatial import distance
from pysparnn.cluster_index import MultiClusterIndex
dim = 20000
dataset = np.random.binomial(1, 0.01, size=(40000, dim))
queryset = np.random.binomial(1, 0.009, size=(1000, dim))
print('dataset[0]:', [[i, v] for i, v in enumerate(dataset[0]) if v > 0])
k = 3
q0 = queryset[0]
res = []
for i in range(dataset.shape[0]):
res.append([i, distance.cosine(q0, dataset[i, :])])
res.sort(key=lambda x: x[1])
print('q0 res', res[:k])
data_matrix = csr_matrix(dataset, dtype=np.float32)
query_matrix = csr_matrix(queryset, dtype=np.float32)
data_to_return = range(dataset.shape[0])
with TimeIt('building MultiClusterIndex'):
cp = MultiClusterIndex(data_matrix, data_to_return)
with TimeIt('knn search'):
res = cp.search(query_matrix, k=k, return_distance=False)
print(res[:5])
for i in res[0]:
print(int(i), distance.cosine(q0, dataset[int(i), :]))
#space_type = 'cosinesimil_sparse'
space_type = 'cosinesimil_sparse_fast'
space_param = []
method_name = 'small_world_rand'
index_name = method_name + '_sparse.index'
if os.path.isfile(index_name):
os.remove(index_name)
index = nmslib.init(space_type, space_param, method_name,
nmslib.DataType.SPARSE_VECTOR, nmslib.DistType.FLOAT)
if batch:
with TimeIt('batch add'):
positions = nmslib.addDataPointBatch(
index, np.arange(len(dataset), dtype=np.int32), data_matrix)
print('positions', positions)
else:
d = []
q = []
with TimeIt('preparing'):
for data in dataset:
d.append([[i, v] for i, v in enumerate(data) if v > 0])
for data in queryset:
q.append([[i, v] for i, v in enumerate(data) if v > 0])
with TimeIt('adding points'):
for id, data in enumerate(d):
nmslib.addDataPoint(index, id, data)
print('Let\'s invoke the index-build process')
index_param = ['NN=17', 'efConstruction=50', 'indexThreadQty=4']
query_time_param = ['efSearch=50']
with TimeIt('building index'):
nmslib.createIndex(index, index_param)
print('The index is created')
nmslib.setQueryTimeParams(index, query_time_param)
print('Query time parameters are set')
print("Results for the freshly created index:")
with TimeIt('knn query'):
if batch:
num_threads = 10
res = nmslib.knnQueryBatch(index, num_threads, k, query_matrix)
for idx, v in enumerate(res):
if idx < 5:
print(idx, v)
if idx == 0:
for i in v:
print('q0', i, distance.cosine(q0, dataset[i, :]))
else:
for idx, data in enumerate(q):
res = nmslib.knnQuery(index, k, data)
if idx < 5:
print(idx, res)
nmslib.saveIndex(index, index_name)
print("The index %s is saved" % index_name)
nmslib.freeIndex(index)
def save(self, fname):
nmslib.saveIndex(self.index, fname)
def test_vector_fresh(fast=True):
space_type = 'cosinesimil'
space_param = []
method_name = 'small_world_rand'
index_name = method_name + '.index'
if os.path.isfile(index_name):
os.remove(index_name)
index = nmslib.init(space_type, space_param, method_name,
nmslib.DataType.DENSE_VECTOR, nmslib.DistType.FLOAT)
start = time.time()
if fast:
data = read_data_fast('sample_dataset.txt')
print('data.shape', data.shape)
positions = nmslib.addDataPointBatch(
index, np.arange(len(data), dtype=np.int32), data)
else:
for id, data in enumerate(read_data('sample_dataset.txt')):
pos = nmslib.addDataPoint(index, id, data)
if id != pos:
print('id %s != pos %s' % (id, pos))
sys.exit(1)
end = time.time()
print('added data in %s secs' % (end - start))
print('Let\'s print a few data entries')
print('We have added %d data points' % nmslib.getDataPointQty(index))
print("Distance between points (0,0) " +
str(nmslib.getDistance(index, 0, 0)))
print("Distance between points (1,1) " +
str(nmslib.getDistance(index, 1, 1)))
print("Distance between points (0,1) " +
str(nmslib.getDistance(index, 0, 1)))
print("Distance between points (1,0) " +
str(nmslib.getDistance(index, 1, 0)))
for i in range(0, min(MAX_PRINT_QTY, nmslib.getDataPointQty(index))):
print(nmslib.getDataPoint(index, i))
print('Let\'s invoke the index-build process')
index_param = ['NN=17', 'initIndexAttempts=3', 'indexThreadQty=4']
query_time_param = ['initSearchAttempts=3']
nmslib.createIndex(index, index_param)
print('The index is created')
nmslib.setQueryTimeParams(index, query_time_param)
print('Query time parameters are set')
print("Results for the freshly created index:")
k = 3
start = time.time()
if fast:
num_threads = 10
query = read_data_fast('sample_queryset.txt')
res = nmslib.knnQueryBatch(index, num_threads, k, query)
for idx, v in enumerate(res):
print(idx, v)
else:
for idx, data in enumerate(read_data('sample_queryset.txt')):
print(idx, nmslib.knnQuery(index, k, data))
end = time.time()
print('querying done in %s secs' % (end - start))
nmslib.saveIndex(index, index_name)
print("The index %s is saved" % index_name)
nmslib.freeIndex(index)
import time
import math
import pdb
from xclib.data import data_utils
import hnswlib
lbl_ft_file = sys.argv[1]
model_file = sys.argv[2]
M = int(sys.argv[3])
efC = int(sys.argv[4])
num_threads = int(sys.argv[5])
num_ft = int(sys.argv[6])
metric_space = sys.argv[7]
start = time.time()
data = data_utils.read_sparse_file(lbl_ft_file)
end = time.time()
start = time.time()
index = nmslib.init(method='hnsw',
space='cosinesimil_sparse',
data_type=nmslib.DataType.SPARSE_VECTOR)
index.addDataPointBatch(data)
index.createIndex({
'M': M,
'indexThreadQty': num_threads,
'efConstruction': efC
})
end = time.time()
print('Training time of ANNS datastructure = %f' % (end - start))
nmslib.saveIndex(index, model_file)
