README.md

Training utility

A utility application for training ML models for NACE code classification.

Its main purpose is to run in an automated process, automatically gathering and processing data to train a model without any necessary preparation.

The utility currently uses the FastText library for classification.

Prerequisites

You'll need Python 3. Configure a virtual env or whatever way of running you prefer.

Install requirements:

pip install -r requirements.txt

Running the utility

Run the utility with Python:

python training-util.py

The default behaviour should perform all steps required: gathering data, assembling and cleaning it, then training and testing a model.

More options

To see the full list of available commands and steps, run

python training-util.py --help

Features include setting log level, working directory and starting/stopping the processing at particular steps. This can be useful when e.g. experimenting with different training parameters while using the same training input.

Using the output

Output is written to the working directory (default: data).

The model is stored as model.ftz, and may be used with the FastText library for making predictions. Intermediary data is also stored, and can be re-used when "skipping" steps, e.g. to retrain model on same input:

python training-util.py --from-step=train