def test_reset(self):
"""test IndexNSG.reset()"""
d = self.xq.shape[1]
metrics = {faiss.METRIC_L2: 'L2', faiss.METRIC_INNER_PRODUCT: 'IP'}
metric = faiss.METRIC_L2
flat_index = faiss.IndexFlat(d, metric)
flat_index.add(self.xb)
Dref, Iref = flat_index.search(self.xq, 1)
index = faiss.IndexNSGFlat(d, 16)
index.verbose = True
index.GK = 32
index.add(self.xb)
Dnsg, Insg = index.search(self.xq, 1)
recalls = (Iref == Insg).sum()
print('metric: {}, nb equal: {}'.format(metrics[metric], recalls))
self.assertGreaterEqual(recalls, 475)
self.subtest_connectivity(index, self.xb.shape[0])
index.reset()
index.add(self.xb)
Dnsg, Insg = index.search(self.xq, 1)
recalls = (Iref == Insg).sum()
print('metric: {}, nb equal: {}'.format(metrics[metric], recalls))
self.assertGreaterEqual(recalls, 475)
self.subtest_connectivity(index, self.xb.shape[0])
def test_order(self):
"""make sure that output results are sorted"""
d = self.xq.shape[1]
index = faiss.IndexNSGFlat(d, 32)
index.train(self.xb)
index.add(self.xb)
k = 10
nq = self.xq.shape[0]
D, _ = index.search(self.xq, k)
indices = np.argsort(D, axis=1)
gt = np.arange(0, k)[np.newaxis, :] # [1, k]
gt = np.repeat(gt, nq, axis=0) # [nq, k]
assert np.array_equal(indices, gt)
def subtest_build(self, knn_graph, thresh, metric=faiss.METRIC_L2):
d = self.xq.shape[1]
metrics = {faiss.METRIC_L2: 'L2', faiss.METRIC_INNER_PRODUCT: 'IP'}
flat_index = faiss.IndexFlat(d, metric)
flat_index.add(self.xb)
Dref, Iref = flat_index.search(self.xq, 1)
index = faiss.IndexNSGFlat(d, 16, metric)
index.verbose = True
index.build(self.xb, knn_graph)
Dnsg, Insg = index.search(self.xq, 1)
recalls = (Iref == Insg).sum()
print('metric: {}, nb equal: {}'.format(metrics[metric], recalls))
self.assertGreaterEqual(recalls, thresh)
self.subtest_connectivity(index, self.xb.shape[0])
if 'kmeans_hnsw' in todo:
print("Performing kmeans on sift1M using HNSW assignment")
clus = faiss.Clustering(d, 16384)
clus.verbose = True
clus.niter = 10
index = faiss.IndexHNSWFlat(d, 32)
# increase the default efSearch, otherwise the number of empty
# clusters is too high.
index.hnsw.efSearch = 128
clus.train(xb, index)
if 'nsg' in todo:
print("Testing NSG Flat")
index = faiss.IndexNSGFlat(d, 32)
index.build_type = 1
# training is not needed
# this is the default, higher is more accurate and slower to
# construct
print("add")
# to see progress
index.verbose = True
index.add(xb)
print("search")
for search_L in -1, 16, 32, 64, 128, 256:
print("search_L", search_L, end=' ')
index.nsg.search_L = search_L