| """Optional dense ANN baseline helpers. |
| |
| These deliberately accept PRECOMPUTED embeddings. They never run a transformer. |
| Install with: pip install 'geomretrieval[ann]' |
| """ |
| from __future__ import annotations |
| import time |
| import numpy as np |
|
|
|
|
| def faiss_flat_ip(corpus: np.ndarray, queries: np.ndarray, k: int = 100): |
| import faiss |
| xb = np.ascontiguousarray(corpus.astype(np.float32)) |
| xq = np.ascontiguousarray(queries.astype(np.float32)) |
| index = faiss.IndexFlatIP(xb.shape[1]) |
| index.add(xb) |
| t0 = time.perf_counter() |
| D, I = index.search(xq, k) |
| ms = (time.perf_counter() - t0) * 1000.0 / len(xq) |
| return I, D, ms |
|
|
|
|
| def faiss_hnsw_ip(corpus: np.ndarray, queries: np.ndarray, k: int = 100, M: int = 32, ef_search: int = 128): |
| import faiss |
| xb = np.ascontiguousarray(corpus.astype(np.float32)) |
| xq = np.ascontiguousarray(queries.astype(np.float32)) |
| index = faiss.IndexHNSWFlat(xb.shape[1], M, faiss.METRIC_INNER_PRODUCT) |
| index.hnsw.efSearch = ef_search |
| index.add(xb) |
| t0 = time.perf_counter() |
| D, I = index.search(xq, k) |
| ms = (time.perf_counter() - t0) * 1000.0 / len(xq) |
| return I, D, ms |
|
|
|
|
| def faiss_ivf_flat_ip(corpus: np.ndarray, queries: np.ndarray, k: int = 100, nlist: int = 4096, nprobe: int = 64): |
| import faiss |
| xb = np.ascontiguousarray(corpus.astype(np.float32)) |
| xq = np.ascontiguousarray(queries.astype(np.float32)) |
| quant = faiss.IndexFlatIP(xb.shape[1]) |
| index = faiss.IndexIVFFlat(quant, xb.shape[1], nlist, faiss.METRIC_INNER_PRODUCT) |
| index.train(xb) |
| index.add(xb) |
| index.nprobe = nprobe |
| t0 = time.perf_counter() |
| D, I = index.search(xq, k) |
| ms = (time.perf_counter() - t0) * 1000.0 / len(xq) |
| return I, D, ms |
|
|
