def run_gpu_inference(topk=100, index_file=INDEX_FILE, ef_search=300):
print("=" * BARRIER_SIZE)
data_path = pjoin(RES_DIR, DATA_FILE)
index_path = pjoin(RES_DIR, index_file)
LOGGER.info("gpu inference on %s with index %s", data_path, index_path)
ch0 = CuHNSW(OPT)
LOGGER.info("load model from %s by cuhnsw", index_path)
ch0.load_index(index_path)
h5f = h5py.File(data_path, "r")
queries = h5f["test"][:, :].astype(np.float32)
neighbors = h5f["neighbors"][:, :topk].astype(np.int32)
h5f.close()
num_queries = queries.shape[0]
queries /= np.linalg.norm(queries, axis=1)[:, None]
start = time.time()
pred_nn, _, _ = ch0.search_knn(queries, topk, ef_search)
el0 = time.time() - start
LOGGER.info("elapsed for inferencing %d queries of top@%d (ef_search: %d): "
"%.4e sec", num_queries, topk, ef_search, el0)
accs = []
for _pred_nn, _gt_nn in zip(pred_nn, neighbors):
intersection = set(_pred_nn) & set(_gt_nn)
acc = len(intersection) / float(topk)
accs.append(acc)
LOGGER.info("accuracy mean: %.4e, std: %.4e", np.mean(accs), np.std(accs))
return el0, np.mean(accs)
def run_gpu_inference2(topk=5, index_file="cuhnsw.index", ef_search=300):
print("=" * BARRIER_SIZE)
data_path = pjoin(RES_DIR, DATA_FILE)
index_path = pjoin(RES_DIR, index_file)
LOGGER.info("gpu inference on %s with index %s", data_path, index_path)
ch0 = CuHNSW(OPT)
LOGGER.info("load model from %s by cuhnsw", index_path)
ch0.load_index(index_path)
h5f = h5py.File(data_path, "r")
data = h5f["train"][:, :].astype(np.float32)
queries = h5f["test"][:5, :].astype(np.float32)
h5f.close()
if NRZ:
data /= np.linalg.norm(data, axis=1)[:, None]
nns, distances, found_cnt = ch0.search_knn(queries[:5], topk, ef_search)
for idx, (nn0, distance, cnt) in \
enumerate(zip(nns, distances, found_cnt)):
print("=" * BARRIER_SIZE)
print(f"query {idx + 1}")
print("-" * BARRIER_SIZE)
for _idx, (_nn, _dist) in enumerate(zip(nn0[:cnt], distance[:cnt])):
if DIST_TYPE == "l2":
real_dist = np.linalg.norm(data[_nn] - queries[idx])
_dist = np.sqrt(_dist)
elif DIST_TYPE == "dot":
real_dist = data[_nn].dot(queries[idx])
print(f"rank {_idx + 1}. neighbor: {_nn}, dist by lib: {_dist}, "
f"actual dist: {real_dist}")
def run_gpu_inference_large(topk=100, index_file=INDEX_FILE, ef_search=300,
num_queries=1000000, num_dims=50):
print("=" * BARRIER_SIZE)
index_path = pjoin(RES_DIR, index_file)
data_path = pjoin(RES_DIR, DATA_FILE)
LOGGER.info("gpu inference on %s with index %s", data_path, index_path)
ch0 = CuHNSW(OPT)
LOGGER.info("load model from %s by cuhnsw", index_path)
ch0.load_index(index_path)
queries = np.random.normal(size=(num_queries, num_dims)).astype(np.float32)
num_queries = queries.shape[0]
queries /= np.linalg.norm(queries, axis=1)[:, None]
start = time.time()
_, _, _ = ch0.search_knn(queries, topk, ef_search)
el0 = time.time() - start
LOGGER.info("elapsed for inferencing %d queries of top@%d (ef_search: %d): "
"%.4e sec", num_queries, topk, ef_search, el0)
return el0