Test rig forecasting local training

The repository contains the source code for implementations of the local training on the test rig data. The implementations differ mainly in the orchestration and tracking stacks. The implemented orchestrators include dvc, prefect and kfp. All aspire to leverage mlflow for experiment tracking and optuna for tuning.

Set up environment

python -m venv .venv
source .venv/Scripts/activate
python -m pip install --upgrade pip setuptools
pip install -r conf/requirements-dev.txt

DVC orchestrated training

cd dvc/

Reproduce pipeline

dvc repro [OPTIONS]

See reference for details on the available dvc repro CLI options.

Run experiment

dvc exp run [OPTIONS] [PARAMS]

See reference for details on the available dvc exp run CLI options. The training parameters can be set with -S flag. To see the list of the available training parameters, see params.yaml.

Prefect orchestrated training

cd prefect/

Execute training run

python src/training.py [OPTIONS]
OPTIONS:
    --train_split TRAIN_SPLIT
    --lookback LOOKBACK
    --val_split VAL_SPLIT
    --epochs EPOCHS
    --batch_size BATCH_SIZE
    --patience PATIENCE
    --lstm_units LSTM_UNITS
    --learning_rate LEARNING_RATE
    --folds FOLDS
    --verbose VERBOSE
    --univariate

View runs dashboard

prefect orion start

See localhost:4200

View experiments dashboard

mlflow ui

See localhost:5000