Introduction

classy is a simple-to-use library for building high-performance Machine Learning models in NLP. It wraps the best libraries around (PyTorch, PyTorch Lightning, Transformers, Hydra, Streamlit, ...) and offers them to users with a simple CLI interface.

Installation

pip install classy-coreSuccessfully installed classy-core

Example Walkthrough

You have the following data at disposal:

User	Product-ID	Review	Overall Judgement
U1	P1	I really like this headphones	positive
U2	P1	Sound is terrible! I am returning these headphones right away.	negative
...	...	...	...

That is, a list of reviews made by users for a given product, along with their overall judgement. Now, say you want to train a classification model on this data so that, given a new review as input, it would yield whether its overall judgement is positive or negative.

info

This is a Sequence Classification problem, specifically Sentiment Classification.

In a nutshell, this is how you would do it with classy:

Data Organization

By default, classy is able to work with .tsv and .jsonl files. In the case of .tsv for Sequence Classification, classy needs the file to have only two columns, one with the sequence and the other with the label.

Say your review data is already stored in a .tsv file (with four columns, as we saw above), all you need to do is:

$ cut -f3,4 data/raw.tsv > data/classy-sentiment.tsv

tip

classy supports many tasks and input formats, and is independent of how you convert your data to a classy-compatible format (i.e., sed, awk, and a python script are all viable options).

Training

Now that our data is classy-compatible, we can train our model! Let us call it sequence-example:

classy train sequence data/classy-sentiment.tsv -n sequence-exampleTraining completed

info

sequence in the above command tells classy to train a Sequence Classification model. This is the only thing, beside organizing data, that classy expects you to do. We'll go back to this later on.

Using your trained model

Now that we have trained our model, classy offers a number of commands that let you use it:

• classy demo spawns a Streamlit-based interface that lets you query your model visually;
• classy predict lets you choose whether to use your model for file prediction or in bash-interactive mode (similarly to demo, but in the terminal);
• classy serve spawns a REST API (through FastAPI) that you can query using any REST-compatible client (requests, curl, you name it);
• classy evaluate runs your model on a given test set and reports metrics on the task.