Python `IndexBinaryFlat.reconstruct_n()` broken

[Dobatymo]

Summary

reconstruct_n on a IndexBinaryFlat throws a TypeError about a additional argument for the reconstructed vectors. Other Indexes simply return them. Also providing a None value simply crashes Python.

>>> import faiss
>>> index = faiss.IndexBinaryFlat(64)
>>> index.reconstruct_n(0, 4)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
TypeError: reconstruct_n() missing 1 required positional argument: 'recons'
>>> index.reconstruct_n(0, 4, None)  # crash!

EDIT: It looks like the reconstruct_n definition is missing for handle_IndexBinary (?)

faiss/faiss/python/class_wrappers.py

Line 694 in 52dee5c

def handle_IndexBinary(the_class):

Platform

OS: Windows 10
Python version: 3.8.10
Faiss version: faiss-cpu==1.7.3
Installed from: pip

Running on:

• CPU
• GPU

Interface:

• C++
• Python

Reproduction instructions

See summary.

[mdouze]

This is because the reconstruct_n function is not wrapped in python. The proper way of calling it is with:

x  = np.empty((4, index.d // 8), dtype='uint8')
index.reconstruct_n(0, 4, faiss.swig_ptr(x)) 

But it would be better if the function was wrapped properly, hence the enhancement.

[Dobatymo]

@mdouze Thank you! But there is still something weird going on. Running this with index.d // 8 crashes Python also. Using simply index.d works, but then of course the array is to large for searches.

import numpy as np
import faiss
arr = np.random.randint(0, 256, size=(100, 10), dtype=np.uint8)
index = faiss.IndexBinaryFlat(arr.shape[1] * 8)
index.add(arr)
verify = np.empty((index.ntotal, index.d), dtype=np.uint8)
index.reconstruct_n(0, index.ntotal, faiss.swig_ptr(verify))
verify = verify[:,:index.d // 8]
assert np.array_equal(arr, verify)

verify[:,:index.d // 8] seems to do the trick...