Feature request: create_direct_map for GPU indexes and allow reconstruct

[urialon]

Summary

Would it be possible to make_direct_map() on GPU indexes, to allow calling index.reconstruct?
If not, is there another way to reconstruct quickly using the GPU?
Thanks!

Platform

OS: Ubuntu

Faiss version: 1.7.1

Installed from: pip install faiss-gpu

Running on:

• CPU
• GPU

Interface:

• C++
• Python

Reproduction instructions

Load any index, convert it to GPU:

import faiss

indexfile = 'my_knn.index'
cpu_index = faiss.read_index(indexfile, faiss.IO_FLAG_ONDISK_SAME_DIR)
co = faiss.GpuClonerOptions()
co.useFloat16 = True
gpu_index = faiss.index_cpu_to_gpu(faiss.StandardGpuResources(), 0, cpu_index, co)

Now, it would be great if I could run:

gpu_index.make_direct_map()

(which currently results in: AttributeError: 'GpuIndexIVFPQ' object has no attribute 'make_direct_map')

and then ideally I would like to run:

gpu_index.reconstruct(5)
gpu_index.reconstruct(8)
gpu_index.reconstruct(7)
...

(which currently results in: faiss/faiss/Index.cpp:48: reconstruct not implemented for this type of index)

[mdouze]

Right, GPU indexes do not support reconstruct.
Currently the best alternative is to keep a CPU copy of the index and reconstruct from that.