finalfusion embeddings format and libraries

Follow me on GitHub

Python module


finalfusion-python is a Python module for reading, writing, and using finalfusion embeddings, but also offers methods to read and use fastText, word2vec and GloVe embeddings. This module is implemented in Rust as a wrapper around the finalfusion crate.

The Python module supports the same types of embeddings:

  • Vocabulary:
    • No subwords
    • Subwords
  • Embedding matrix:
    • Array
    • Memory-mapped
    • Quantized
  • Format:
    • finalfusion
    • fastText
    • word2vec
    • GloVe


The finalfusion module is available on PyPi for Linux, Mac and Windows. You can use pip to install the module:

$ pip install --upgrade finalfusion

Building from source

finalfusion can also be built from source. This requires a Rust toolchain that is installed through rustup. First, you need maturin:

$ cargo install maturin

finalfusion currently requires a nightly version of Rust. You can use rustup to switch to a nightly build:

# Use the nightly toolchain in the current directory.
$ rustup override set nightly

Now you can build finalfusion-python wheels for Python versions that are detected by maturin:

$ maturin build --release

The wheels are then in the target/wheels directory.


finalfusion embeddings can be loaded as follows:

import finalfusion
# Loading embeddings in finalfusion format
embeds = finalfusion.Embeddings("myembeddings.fifu")

# Or if you want to memory-map the embedding matrix:
embeds = finalfusion.Embeddings("myembeddings.fifu", mmap=True)

# fastText format
embeds = finalfusion.Embeddings.read_fasttext("myembeddings.bin")

# word2vec format
embeds = finalfusion.Embeddings.read_word2vec("myembeddings.w2v")

You can then compute an embedding, perform similarity queries, or analogy queries:

e = embeds.embedding("Tübingen")
# default similarity query for "Tübingen"

# similarity query based on a vector, returning the closest embedding to
# the input vector, skipping "Tübingen"
embeds.embeddings_similarity(e, skip={"Tübingen"})

# default analogy query
embeds.analogy("Berlin", "Deutschland", "Amsterdam")

# analogy query allowing "Deutschland" as answer
embeds.analogy("Berlin", "Deutschland", "Amsterdam", mask=(True,False,True))

If you want to operate directly on the full embedding matrix, you can get a copy of this matrix through:

# get copy of embedding matrix, changes to this won't touch the original matrix

Finally access to the vocabulary is provided through:

v = e.vocab()
# get a list of indices associated with "Tübingen"

# get a list of `(ngram, index)` tuples for "Tübingen"

# get a list of subword indices for "Tübingen"

More usage examples can be found in the examples directory.

Where to go from here