Crate finalfusion

A library for reading, writing, and using word embeddings.

finalfusion allows you to read, write, and use word2vec/GloVe embeddings and read fastText embeddings. finalfusion uses finalfusion as its native data format, which has several benefits over the word2vec, GloVe, and fastText formats.

Reading finalfusion embeddings

finalfusion embeddings can be read with the read_embeddings method, which expects a reader that implements the BufRead trait.

Since finalfusion supports various types of vocabularies and embedding matrix (storage) formats, these should be specified as type parameters of the Embeddings type. However, typically one would want to read finalfusion embeddings with any type of vocabulary or embedding matrix. For this purpose, the VocabWrap and StorageWrap types are provided, which wrap any type of vocabulary and embeddung matrix.

We can thus load a finalfusion format and retrieve an embedding as follows:

use std::fs::File;
use std::io::BufReader;

use finalfusion::prelude::*;

let mut reader = BufReader::new(File::open("testdata/similarity.fifu").unwrap());

// Read the embeddings.
let embeddings: Embeddings<VocabWrap, StorageWrap> =
    Embeddings::read_embeddings(&mut reader)
    .unwrap();

// Look up an embedding.
let embedding = embeddings.embedding("Berlin");

For performing analogy/similarity queries on the embedding matrix, we need an embedding matrix which can act as a view. In that case one should use StorageViewWrap in place of StorageWrap. StorageViewWrap is only supported for a subset of embedding matrix types – in particular, quantized matrices cannot be used as a view.

Reading other embedding formats

Consult the documentation of the fasttext, text and word2vec modules for information on how to read fastText, GloVe, and word2vec embeddings.

Modules

• compat
  Readers/writers for other embedding formats.
• embeddings
  Word embeddings.
• error
  Error/result types
• io
  Traits and error types for I/O.
• metadata
  Metadata chunks
• norms
  Norms chunk
• prelude
  Prelude exports the most commonly-used types and traits.
• similarity
  Traits and trait implementations for similarity queries.
• storage
  Embedding matrix representations.
• subword
  Utilities for subword units.
• vocab
  Embedding vocabularies