Dear all, does anyone know why the following code could return negative entries for D? I am calculating the L2-nearest neighbors of CIFAR images, for which I assume IndexFlatL2 should return non-negative distances (and 0 for exact match).

index = faiss.IndexFlatL2(d)
index.add(Data)
D, I = index.search(Data, 10)

Some notes:

1. The most problematic case seems to be when some feature dimensions have significantly larger variance than the others, in which case the negative entries can be quite large (say around -10^3).
2. I get the same problem using either CPU or GPU.
3. I get correct results running the tutorial code 1-Flat.py.

Thanks!