Unable to write IndexFlatIP in faiss-gpu

[gajghatenv]

Summary

I am using Faiss to retrieve similar products. My embedding size is 1024. I tried faiss-cpu but it was too slow. Hence, I am trying faiss-gpu. However, in my experiments, I am unable to write an IndexFlatIP index. I was able to use write_index() in faiss-cpu.

Platform

OS: Ubuntu 20.04.5 LTS

Faiss version: faiss-gpu: 1.7.2

Installed from: pip

Running on:

• [ ] CPU
• [x] GPU

Interface:

• [ ] C++
• [x] Python

Reproduction instructions

Code

import numpy as np
import faiss

def create_dummy_embeddings():
    # Create dummy embeddings
    embeddings = np.random.rand(100000, 1024).astype('float32')
    print('Dummy embeddings created:', embeddings.shape)
    return embeddings

def create_index(embeddings):
    # Create index
    EMBEDDING_SIZE = 1024
    print('Embedding size:', EMBEDDING_SIZE)

    # GPU
    res = faiss.StandardGpuResources()  # use a single GPU
    ngpus = faiss.get_num_gpus()
    print("Number of GPUs:", ngpus)
    cpu_index = faiss.IndexFlatIP(EMBEDDING_SIZE)
    gpu_index = faiss.index_cpu_to_gpu(res, 0, cpu_index)
    gpu_index.add(embeddings)
    print('Number of vectors in index:', gpu_index.ntotal)

    faiss.write_index(gpu_index, 'faiss_index_dummy.index')
    print('Index saved to file.')


def load_index():
    index = faiss.read_index('faiss_index_dummy.index')
    print('Index loaded from file.')
    print('Number of vectors in index:', index.ntotal)


def main():
    embeddings = create_dummy_embeddings()
    create_index(embeddings)
    load_index()

if __name__ == '__main__':
    main()

Error

RuntimeError: Error in void faiss::write_index(const faiss::Index*, faiss::IOWriter*) at /project/faiss/faiss/impl/index_write.cpp:590: don't know how to serialize this type of index

Hoping to get this fixed! Thank you!

[mdouze]

GPU indexes cannot be stored directly. Please convert to CPU first.

[gajghatenv]

So basically:

1. Create a cpu_index
2. Convert it to a gpu_index
3. Add vectors to the gpu_index
4. Convert gpu_index back to cpu_index
5. Write cpu_index

Am I correct?

And then when trying to use the index:

1. Read in the cpu_index
2. Convert to gpu_index
3. Perform searches

Did I get this right?