COST-AWARE BAYESIAN HYPERPARAMETER OPTIMIZATION OF ML PIPELINES WITH BLACKBOX COST

Steps to run the T5 Pipeline Tuning Experiment

Option A

• Clone the repository git clone https://github.com/ridwan-salau/cost-aware-bo.git checkout to ridwan/t5-pipeline git checkout ridwan/t5-pipeline
• Download the datasets - tokenized_train_data.pt and tokenized_validation_data.pt. (N.B. you might need git-lfs to be able to pull the file from github as downloading larges files from git directly won't work.)
• Place the two files in the directory t5_fine_tuning/inputs/
• Create a conda environment by running conda create -f t5_fine_tuning/t5env.yml from the root.
• To setup WandB logging, run export WANDB_API_KEY=<<API KEY FROM WANDB>>
• Run an experiment using bash run.sh <<ACQF_Method>>, where ACQF_Method is one of {EI, EEIPU, EIPS, CArBO}

Option B (Docker)

• Clone the repository git clone https://github.com/ridwan-salau/cost-aware-bo.git checkout to ridwan/t5-pipeline git checkout ridwan/t5-pipeline
• Place