"""Downstream use: the centroids as the coarse quantizer of an IVF index (the use case e-means targets). For each method and seed, train k centroids on DEEP* 1M, build IndexIVFFlat over the same 1M rows with those centroids, and measure on the 10K public DEEP1B queries (exact ground truth on the 1M rows): - recall@1 and recall@10 (10-NN found among the 10 returned) at nprobe in NPROBES, - distance computations per query (sum of probed list sizes), the real cost of an IVF search, - list imbalance factor k * sum(n_i^2) / n^2 (1 = perfectly balanced). Usage: python benchmarks/ivf_downstream.py DATA/deep1b OUT.jsonl --epochs 3 --seeds 10 11 12 --methods official pisces """ import argparse, json, os, sys, time import numpy as np import faiss ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) sys.path.insert(0, ROOT) sys.path.insert(0, os.path.join(ROOT, "benchmarks")) from paper_protocol import METHODS, make_variant # noqa: E402 NPROBES = [1, 2, 4, 8, 16, 32, 64, 128] def read_fbin(path): n, d = np.fromfile(path, dtype="int32", count=2) return np.fromfile(path, dtype="float32", offset=8).reshape(n, d) def main(): ap = argparse.ArgumentParser() ap.add_argument("data"); ap.add_argument("out") ap.add_argument("--k", type=int, default=16384) ap.add_argument("--epochs", type=float, nargs="+", default=[3]) ap.add_argument("--seeds", type=int, nargs="+", default=[10, 11, 12]) ap.add_argument("--methods", nargs="+", default=["official", "pisces"]) ap.add_argument("--threads", type=int, default=4) a = ap.parse_args() faiss.omp_set_num_threads(a.threads) x = np.ascontiguousarray(np.load(os.path.join(a.data, "train_1M.npy"))) q = np.ascontiguousarray(read_fbin(os.path.join(a.data, "query.public.10K.fbin"))) _, gt = faiss.knn(q, x, 10) from aries.pemeans import PISCES METHODS["pisces"] = make_variant(PISCES, len(x)) done = set() if os.path.exists(a.out): done = {(r["method"], r["epochs"], r["seed"]) for r in map(json.loads, open(a.out))} for T in a.epochs: for s in a.seeds: init = x[np.random.RandomState(s).permutation(len(x))[:a.k]].copy() for m in a.methods: if (m, T, s) in done: continue t0 = time.time() c, _ = METHODS[m](x.shape[1], T, s).train(x, init_centroids=init.copy()) sec = time.time() - t0 c = np.ascontiguousarray(c, dtype="float32") quant = faiss.IndexFlatL2(x.shape[1]); quant.add(c) ivf = faiss.IndexIVFFlat(quant, x.shape[1], a.k); ivf.is_trained = True; ivf.add(x) sizes = np.array([ivf.invlists.list_size(i) for i in range(a.k)], dtype="float64") imb = float(a.k * (sizes ** 2).sum() / sizes.sum() ** 2) r = dict(method=m, epochs=T, seed=s, k=a.k, train_seconds=sec, imbalance=imb, empty_lists=int((sizes == 0).sum()), curve=[]) for npb in NPROBES: ivf.nprobe = npb _, Dq = quant.search(q, npb) ndis = float(sizes[Dq].sum(1).mean()) t1 = time.time(); _, I = ivf.search(q, 10); qs = time.time() - t1 r1 = float((I[:, 0] == gt[:, 0]).mean()) r10 = float(np.mean([len(set(I[i]) & set(gt[i])) / 10 for i in range(len(q))])) r["curve"].append(dict(nprobe=npb, recall1=r1, recall10=r10, ndis=ndis, seconds=qs)) with open(a.out, "a") as f: f.write(json.dumps(r) + "\n") print(m, T, s, f"imb {imb:.3f}", " ".join(f"{c_['nprobe']}:{c_['recall1']:.3f}/{c_['ndis']:.0f}" for c_ in r["curve"]), flush=True) if __name__ == "__main__": main()