🏐Lab 4
ANNOY vs. FAISS vs. HNSW
import numpy as np
import time
from annoy import AnnoyIndex
import faiss
import hnswlib
# ===============================
# Build 1mio rows with 5D
# ===============================
n_data = 1_000_000 # try 100_000 if you have limited memory
dim = 5
X = np.random.random((n_data, dim)).astype(np.float32)
# Query point
query = np.random.random((1, dim)).astype(np.float32)
k = 10
# ===============================
# 2. Annoy
# ===============================
print("=== Annoy ===")
ann_index = AnnoyIndex(dim, 'euclidean')
start = time.time()
for i in range(n_data):
ann_index.add_item(i, X[i])
ann_index.build(10) # 10 trees
build_time = time.time() - start
start = time.time()
neighbors = ann_index.get_nns_by_vector(query[0], k, include_distances=True)
query_time = time.time() - start
print("Build time:", build_time, "detik")
print("Query time:", query_time, "detik")
print("Neighbors:", neighbors[0][:5], "...")
# ===============================
# 3. FAISS (Flat Index)
# ===============================
print("\n=== FAISS (IndexFlatL2) ===")
faiss_index = faiss.IndexFlatL2(dim)
start = time.time()
faiss_index.add(X)
build_time = time.time() - start
start = time.time()
distances, indices = faiss_index.search(query, k)
query_time = time.time() - start
print("Build time:", build_time, "detik")
print("Query time:", query_time, "detik")
print("Neighbors:", indices[0][:5], "...")
# ===============================
# 4. HNSW (hnswlib)
# ===============================
print("\n=== HNSW (hnswlib) ===")
hnsw_index = hnswlib.Index(space='l2', dim=dim)
start = time.time()
hnsw_index.init_index(max_elements=n_data, ef_construction=200, M=16)
hnsw_index.add_items(X)
build_time = time.time() - start
hnsw_index.set_ef(50)
start = time.time()
labels, distances = hnsw_index.knn_query(query, k=k)
query_time = time.time() - start
print("Build time:", build_time, "detik")
print("Query time:", query_time, "detik")
print("Neighbors:", labels[0][:5], "...")Last updated