faiss_indexから学習に使用したembeddingを再構築する

> help(faiss_index.reconstruct_n)
Help on method replacement_reconstruct_n in module faiss.class_wrappers:

replacement_reconstruct_n(n0=0, ni=-1, x=None) method of faiss.swigfaiss.IndexFlatL2 instance
    Approximate reconstruction of vectors `n0` ... `n0 + ni - 1` from the index.
    Missing vectors trigger an exception.
    
    Parameters
    ----------
    n0 : int
        Id of the first vector to reconstruct (default 0)
    ni : int
        Number of vectors to reconstruct (-1 = default = ntotal)
    x : array_like, optional
        pre-allocated array to store the results
    
    Returns
    -------
    x : array_like
        Reconstructed vectors, size (`ni`, `self.d`), `dtype`=float32

batch_size = 4096

embeddings = np.zeros((faiss_index.ntotal, 384), dtype=np.float16)
for idx in tqdm(range(0, faiss_index.ntotal, batch_size), total=faiss_index.ntotal // batch_size):
    real_batch_size = min(idx + batch_size, faiss_index.ntotal) - idx
    embeddings[idx : idx + real_batch_size, :] = faiss_index.reconstruct_n(idx, real_batch_size)