Lightopic

This package addresses the specific use case of deploying a BERTopic model that you've trained, and now want to use for transforming new data, e.g. via an API.

This came up for me because I wanted to deploy such a model API but wanted to make the deployment smaller and faster. The BERTopic package is broad, which brings with it a load of dependencies (e.g. torch, a bunch of cuda libraries). So I wrote this as a way to do the transform step only, with a virtual environment that's about 95% smaller than one with the actual BERTopic package.

The main prerequisite is that you need to have trained a BERTopic model separately and have serialised it in a way that's compatible with lightopic. The lightopic package also offers you a way to do that: guidance on how is below. From that point you can instantiate a Lightopic object and use its transform method on new data.

Training and serialising your LightBERTopic model

This is a necessary step: you can't instantiate a Lightopic object without first having trained and serialised your model. To make this part easier the LightBERTopic class is available: this is a child class of bertopic.BERTopic, only with a method added to save_lightopic.

from lightopic.lightbertopic import LightBERTopic
docs = fetch_20newsgroups(subset='all',  remove=('headers', 'footers', 'quotes'))['data']

topic_model = LightBERTopic()
topics, probs = topic_model.fit_transform(docs)
topic_model.save_lightopic("model_directory")

NB. for this to work you must have bertopic installed, which you can do with pip install lightopic[bertopic].

NOTE: this package is still under development, so this required format may (and probably will) change!

Using a Lightopic model

Now the serialised model is ready to use.

from lightopic import Lightopic
topic_model = Lightopic()
topic_model.load("model_directory")
topic_model.transform(embeddings)

This transform step does not rely on BERTopic at all, so it can use the smaller installation you get from pip install lightopic.