issue regarding stats

[anupsingh15]

Hello,

I am implementing the HNSW index on a dummy dataset. I want to know how many exact distance computations are performed during the search. To do so, I use faiss.cvar.hnsw_stats.ndis; however, it always returns 0. Could you tell me how to fix the issue?

Code:

x = np.random.randn(10000,128)
M=32
index_hnsw = faiss.IndexHNSWFlat(128, M)            
index_hnsw.hnsw.efConstruction = 40        
index_hnsw.hnsw.efSearch = 16   
index_hnsw.add(x)

D, I = index_hnsw.search(x[:10], 2)
cands = faiss.cvar.hnsw_stats.ndis #total cands evaluated across all queries

[mlomeli1]

This was fixed by @mdouze in #3309 (thank you Matthijs for flagging!). @anupsingh15 you need use the nightly package (see INSTALL.md) so this fix is included in the version of faiss you have in your conda environment, for your reference:

conda create -n faiss_1_8_0_nightly python=3.10
conda activate faiss_1_8_0_nightly
conda install -c pytorch/label/nightly -c nvidia faiss-gpu=1.8.0

the output is then

import numpy as np
import faiss
x = np.random.randn(10000,128)
M=32
index_hnsw = faiss.IndexHNSWFlat(128, M)            
index_hnsw.hnsw.efConstruction = 40        
index_hnsw.hnsw.efSearch = 16   
index_hnsw.add(x)

D, I = index_hnsw.search(x[:10], 2)
cands = faiss.cvar.hnsw_stats.ndis #total cands evaluated across all queries
cands
10268