def __init__(self, metric: str, dimension: int):
self.name = f"eknn-exact-metric={metric}_dimension={dimension}"
self.metric = metric
self.dimension = dimension
self.model = ElastiknnModel("exact", dealias_metric(metric))
self.batch_res = None
es_wait()
def test_exact_jaccard_mnist(self):
# First run the query and make sure the results have the right form.
n_neighbors = 20
model = ElastiknnModel('exact', 'jaccard')
model.fit(digits_train)
inds1 = model.kneighbors(digits_validate, n_neighbors)
inds2, dists2 = model.kneighbors(digits_validate,
n_neighbors,
return_similarity=True)
assert np.all(inds1 == inds2)
assert inds1.shape == (digits_validate.shape[0], n_neighbors)
assert dists2.shape == inds2.shape
# Then compare against scikit-learn. Intentionally using fewer neighbors to make sure recall will be 1
# despite out-of-order indices due to equal distances.
ref = NearestNeighbors(n_neighbors=int(n_neighbors / 2),
algorithm='brute',
metric='jaccard',
n_jobs=1)
ref.fit(digits_train)
inds3 = ref.kneighbors(digits_validate, return_distance=False)
# Compute and check the recall.
rec = self.recall(inds3, inds2)
assert np.all(rec == 1)
def __init__(self, L: int, k: int, w: int):
self.name_prefix = f"eknn-l2lsh-L={L}-k={k}-w={w}"
self.name = None # set based on query args.
self.model = ElastiknnModel("lsh",
"l2",
mapping_params=dict(L=L, k=k, w=w))
self.X_max = 1.0
self.query_params = dict()
self.batch_res = None
self.sum_query_dur = 0
self.num_queries = 0
es_wait()
def lsh(dataset: Dataset,
bands: int = 165,
rows: int = 1,
candidates: float = 1.5):
n_neighbors = len(dataset.queries[0].indices)
eknn = ElastiknnModel(
algorithm='lsh',
metric='jaccard',
n_jobs=1,
index=f"{INDEX}-{int(time())}",
mapping_params={
"bands": bands,
"rows": rows
},
query_params={"candidates": int(candidates * n_neighbors)})
return evaluate(dataset, eknn)
def indexed(dataset: Dataset):
eknn = ElastiknnModel(algorithm='sparse_indexed',
metric='jaccard',
n_jobs=1,
index=f"{INDEX}-{int(time())}")
return evaluate(dataset, eknn)
def exact(dataset: Dataset):
eknn = ElastiknnModel(algorithm='exact',
metric='jaccard',
n_jobs=1,
index=f"{INDEX}-exact")
return evaluate(dataset, eknn)