Python module


finalfusion is a Python package for reading, writing and using finalfusion embeddings, but also supports other commonly used embeddings like fastText, GloVe and word2vec.

The Python package supports the same types of embeddings as the finalfusion-rust crate:

  • Vocabulary:
    • No subwords
    • Subwords
  • Embedding matrix:
    • Array
    • Memory-mapped
    • Quantized
  • Norms
  • Metadata


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

$ pip install --upgrade finalfusion

Installing from source

Building from source depends on Cython. If you install the package using pip, you don’t need to explicitly install the dependency since it is specified in pyproject.toml.

$ git clone
$ cd finalfusion-python
$ pip install .

If you want to build wheels from source, wheel needs to be installed. It’s then possible to build wheels through:

$ python bdist_wheel

The wheels can be found in dist.

Package Usage

Basic usage

import finalfusion
# loading from different formats
w2v_embeds = finalfusion.load_word2vec("/path/to/w2v.bin")
text_embeds = finalfusion.load_text("/path/to/embeds.txt")
text_dims_embeds = finalfusion.load_text_dims("/path/to/embeds.dims.txt")
fasttext_embeds = finalfusion.load_fasttext("/path/to/fasttext.bin")
fifu_embeds = finalfusion.load_finalfusion("/path/to/embeddings.fifu")

# serialization to formats works similarly
finalfusion.compat.write_word2vec("to_word2vec.bin", fifu_embeds)

# embedding lookup
embedding = fifu_embeds["Test"]

# reading an embedding into a buffer
import numpy as np
buffer = np.zeros([1], dtype=np.float32)
fifu_embeds.embedding("Test", out=buffer)

# similarity and analogy query
sim_query = fifu_embeds.word_similarity("Test")
analogy_query = fifu_embeds.analogy("A", "B", "C")

# accessing the vocab and printing the first 10 words
vocab = fifu_embeds.vocab

# SubwordVocabs give access to the subword indexer:
subword_indexer = vocab.subword_indexer
print(subword_indexer.subword_indices("Test", with_ngrams=True))

# accessing the storage and calculate its dot product with an embedding
res =

# printing metadata

Beyond Embeddings

# load only a vocab from a finalfusion file
from finalfusion import load_vocab
vocab = load_vocab("/path/to/finalfusion_file.fifu")

# serialize vocab to single file

# more specific loading functions exist
from finalfusion.vocab import load_finalfusion_bucket_vocab
fifu_bucket_vocab = load_finalfusion_bucket_vocab("/path/to/vocab_file.fifu.voc")

The package supports loading and writing all finalfusion chunks this way. This is only supported by the Python package, reading will fail with e.g. the finalfusion-rust.


finalfusion also includes a conversion script ffp-convert to convert between the supported formats.

# convert from fastText format to finalfusion
$ ffp-convert -f fasttext fasttext.bin -t finalfusion embeddings.fifu

ffp-bucket-to-explicit can be used to convert bucket embeddings to embeddings with an explicit ngram lookup.

# convert finalfusion bucket embeddings to explicit
$ ffp-bucket-to-explicit -f finalfusion embeddings.fifu explicit.fifu

ffp-select generates new embedding files based on some embeddings and a word list. Using ffp-select with embeddings with a simple vocab results in a subset of the original embeddings. With subword embeddings, vectors for unknown words in the word list are computed and added to the new embeddings. The resulting embeddings cannot provide representations for OOV words anymore. The new vocabulary covers only the words in the word list.

$ ffp-select large-embeddings.fifu subset-embeddings.fifu words.txt

Finally, the package comes with ffp-similar and ffp-analogy to do analogy and similarity queries.

# get the 5 nearest neighbours of "Tübingen"
$ echo Tübingen | ffp-similar embeddings.fifu
# get the 5 top answers for "Tübingen" is to "Stuttgart" like "Heidelberg" to...
$ echo Tübingen Stuttgart Heidelberg | ffp-analogy embeddings.fifu

Where to go from here