Matthijs Douze
Matthijs Douze
by default use_raft is set to false, see https://github.com/facebookresearch/faiss/blob/main/faiss/python/gpu_wrappers.py#L59 It should be set to `use_raft='auto'` and decide at runtime using `get_compile_options()`. https://github.com/facebookresearch/faiss/blob/main/faiss/gpu/GpuIndex.cu#L532
Add doc to the FAISS Wiki * New chapter in Faiss wiki * Add info about how CuVS is wrapped * Point back to CuVS doc where needed
See title. The removal should take place on the two indices. While we are there, check the behavior of filtered search. related: #3820
Summary: Allows factory strings like `IVF3k,Flat` as a shorthand for 3072 centroids. The main question is whether k or M should be metric (k=1000) or power of 2 (k=1024): *...
Summary: External modules should work fine as plug-ins to Faiss in the following cases: * additional objects that can be passed in as callbacks to Faiss * functions that use...
# Summary # Platform OS: Linux 64 intel Faiss version: 1.8.0 Installed from: conda Running on: CPU Interface: Python # Reproduction instructions The type mapping between c++ and Python integer...
GPU search supports `search_preassigned` but not `add_preassigned` which is required when an arbitrary quantizer is used. https://github.com/facebookresearch/faiss/blob/main/faiss/gpu/GpuIndexIVF.h#L101 It would be useful to implement it.
Faiss does not transparently convert numpy int64 and float64 / float32 to SWIG's int and float ```python import faiss import numpy as np a = np.array([4, 6]) d = a[0]...
For some use cases, Faiss serves only as a storage engine for flat arrays of compressed vectors. Maybe it would be useful to make a wrapper that makes this easier...
Summary: C++ Faiss is compiled with (at least) 3 compilers: - the one compiling the main CPU Faiss - the one called by nvcc to compile host code on GPU...