def _add_data_to_index(self, data, indexes, batch_size):
if type(indexes) is list:
offset = indexes[0]
elif isinstance(indexes, numbers.Number):
offset = indexes
else:
offset = 0
for data_batch in np.split(data, batch_size, axis=0):
indices = np.arange(len(data_batch), dtype=np.int32) + offset
nmslib.addDataPointBatch(self.index, indices, data_batch)
offset += data_batch.shape[0]
return offset
def test_vector_load(fast=True, fast_batch=True, seq=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)
f = '/tmp/foo.txt'
if not os.path.isfile(f):
print 'creating %s' % f
np.savetxt(f, np.random.rand(100000,1000), delimiter="\t")
print 'done'
if fast:
index = nmslib.init(
space_type,
space_param,
method_name,
nmslib.DataType.DENSE_VECTOR,
nmslib.DistType.FLOAT)
with TimeIt('fast add data point'):
data = read_data_fast(f)
nmslib.addDataPointBatch(index, np.arange(len(data), dtype=np.int32), data)
nmslib.freeIndex(index)
if fast_batch:
index = nmslib.init(
space_type,
space_param,
method_name,
nmslib.DataType.DENSE_VECTOR,
nmslib.DistType.FLOAT)
with TimeIt('fast_batch add data point'):
offset = 0
for data in read_data_fast_batch(f, 10000):
nmslib.addDataPointBatch(index, np.arange(len(data), dtype=np.int32) + offset, data)
offset += data.shape[0]
print 'offset', offset
nmslib.freeIndex(index)
if seq:
index = nmslib.init(
space_type,
space_param,
method_name,
nmslib.DataType.DENSE_VECTOR,
nmslib.DistType.FLOAT)
with TimeIt('seq add data point'):
for id, data in enumerate(read_data(f)):
nmslib.addDataPoint(index, id, data)
nmslib.freeIndex(index)
def test_vector_load(fast=True, fast_batch=True, seq=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)
f = '/tmp/foo.txt'
if not os.path.isfile(f):
print('creating %s' % f)
np.savetxt(f, np.random.rand(100000,1000), delimiter="\t")
print('done')
if fast:
index = nmslib.init(
space_type,
space_param,
method_name,
nmslib.DataType.DENSE_VECTOR,
nmslib.DistType.FLOAT)
with TimeIt('fast add data point'):
data = read_data_fast(f)
nmslib.addDataPointBatch(index, np.arange(len(data), dtype=np.int32), data)
nmslib.freeIndex(index)
if fast_batch:
index = nmslib.init(
space_type,
space_param,
method_name,
nmslib.DataType.DENSE_VECTOR,
nmslib.DistType.FLOAT)
with TimeIt('fast_batch add data point'):
offset = 0
for data in read_data_fast_batch(f, 10000):
nmslib.addDataPointBatch(index, np.arange(len(data), dtype=np.int32) + offset, data)
offset += data.shape[0]
print('offset', offset)
nmslib.freeIndex(index)
if seq:
index = nmslib.init(
space_type,
space_param,
method_name,
nmslib.DataType.DENSE_VECTOR,
nmslib.DistType.FLOAT)
with TimeIt('seq add data point'):
for id, data in enumerate(read_data(f)):
nmslib.addDataPoint(index, id, data)
nmslib.freeIndex(index)
def test_add_points_batch5(self):
positions = nmslib.addDataPointBatch(
self.index,
np.array([0, 1, 2], dtype=np.int32),
np.array([[0.34, 0.54], [0.55, 0.52], [0.21, 0.68]], dtype=np.float32),
)
nt.assert_array_equal(np.array([0, 1, 2], dtype=np.int32), positions)
def test_add_points_batch5(self):
row = np.array([0, 0, 1, 2, 2])
col = np.array([0, 2, 1, 1, 2])
data = np.array([0.3, 0.2, 0.4, 0.1, 0.6])
m = csr_matrix((data, (row, col)), dtype=np.float32, shape=(3, 3))
print m.toarray()
positions = nmslib.addDataPointBatch(self.index, np.array([0, 1, 2], dtype=np.int32), m)
nt.assert_array_equal(np.array([0, 1, 2], dtype=np.int32), positions)
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_add_points_batch5(self):
row = np.array([0, 0, 1, 2, 2])
col = np.array([0, 2, 1, 1, 2])
data = np.array([0.3, 0.2, 0.4, 0.1, 0.6])
m = csr_matrix((data, (row, col)), dtype=np.float32, shape=(3, 3))
print m.toarray()
positions = nmslib.addDataPointBatch(
self.index, np.array([0, 1, 2], dtype=np.int32), m)
nt.assert_array_equal(np.array([0, 1, 2], dtype=np.int32), positions)
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_add_points_batch5(self):
data = np.array([[0.34, 0.54], [0.55, 0.52], [0.21, 0.68]],
dtype=np.float32)
positions = nmslib.addDataPointBatch(
self.index, np.array([0, 1, 2], dtype=np.int32), data)
nt.assert_array_equal(np.array([0, 1, 2], dtype=np.int32), positions)
def test_add_points_batch5(self):
positions = nmslib.addDataPointBatch(
self.index, np.array([0, 1, 2], dtype=np.int32),
["string1", "string2", "string3"])
nt.assert_array_equal(np.array([0, 1, 2], dtype=np.int32), positions)
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 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)
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):
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_add_points_batch5(self):
positions = nmslib.addDataPointBatch(self.index,
np.array([0, 1, 2], dtype=np.int32),
["string1", "string2", "string3"])
nt.assert_array_equal(np.array([0, 1, 2], dtype=np.int32), positions)
