[Project Name] (a Temply project)

This project ... It uses the tommasobonomo/temply template for quick Deep Learning-based experimentation.

Install

If you just want to play around with the code, maybe run a quick evaluation, but you do not want to start building Docker images and running full training sessions, you should first install the requirements:

pip install -r requirements.txt

Otherwise, you can build a Docker image which takes care of running these steps in the build process:

docker build -t <project name> .

Usage

Usage happens through the fit_and_evaluate script, located at scripts/fit_and_evaluate.py. It is used both for training and evaluation and it can be easily configured through Hydra CLI options.

The Hydra configuration is stored as a Python object (more specifically @dataclass) named Config in the settings.py file. The root configuration object (Config) holds options to control the script behaviour and some integrations (for example Weights&Biases). It also holds two sub-configuration objects, ModelConfig and TrainerConfig. The first holds fields that modify the model that is being used and related settings, while the second one will be deconstructed as arguments to initialise a pytorch_lightning.Trainer, so one should take care not to add fields that are not compatible with the Trainer's initialisation arguments.

The options can be changed directly in the settings.py file or through Hydra's CLI syntax. For example, one could decide to fine-tune a different model than the default and not use Weights&Biases for training tracking with the following syntax:

export PYTHONPATH=. 
python scripts/fit_and_evaluate.py \
    model.transformer_model=bert-base-uncased \
    enable_wandb=false

The above script can be used also with the Docker image, just prepending docker run <project name>.

Training on Google Cloud Platform

It's also very easy to run a custom training job provisioned on Google Cloud Platform's Vertex AI service. If you have the gcloud utility installed and configured, you just need to run:

export JOB_NAME=<project name>_$(date +%Y%m%d_%H%M%S)

gcloud ai custom-jobs create $JOB_NAME \
    --config=gcp_training_config.yaml

Make sure you have built and uploaded to GCP's Container Registery the correct Docker image, then pass that image's URI as <DOCKER_IMAGE_URI> in gcp_training_config.yaml.

For more options and configs, see docs.