Repo for experimenting with building a T5 model of the English Dictionary.

Create a virtual environment with python 3.8. To install the requirements, run pip install -r requirements.txt.

If modifying the requirements.in file, you can generate a new requirements.txt by running pip install pip-tools && pip-compile requirements.in

The training script utilises wandb for tracking experiments. Create a free account and then use the command below to login from the command line. Once you've done this, experiments will be accessible and trackable in your online portal.

wandb login

Data is a direct copy of the data directory from the fantastic wordset dictionary.

Format and preprocess the data as a csv file:

python src/preprocess.py

These are defined in the .env file.

python src/train.py

Arguments can optionally be provided e.g. --model_name=t5-large

This is done using cortex.

First follow the instructions here to create new service, download the key and set it as an environment variable.

Then, spin up the cluster:

cd cortex && cortex cluster-gcp up --config=cluster.yaml

This will spin up a cluster. To create the API and run an example, run the consume.py file.

To spin down the cluster:

 cortex cluster-gcp down

cortex get <api-name>

or

cortex logs <api-name>