Clone the repository
git clone [email protected]:arthurviens/bayesrul.git
cd bayesrul
Use poetry to install dependencies
poetry install
Make sure to have the CMAPSS dataset files at data/ncmapss/N-CMAPSS_DS02-006.h5 (you can use other subsets by modifying arguments in bayesful/ncmapss/generate_files.py)
Launch the script
poetry run python -m bayesrul.ncmapss.generate_files
It will create the necessary parquet and lmdb files used later on. Parquet files are used to create lmdb files but are useless otherwise for now.
You now should have data/ncmapss/lmdb directory, with all that is needed inside.
You can now launch a training (for example a BNN)
poetry run python -m bayesrul.ncmapss.train_model --bayesian --archi inception --guide normal --GPU 0 --model-name My_Model
Or a frequentist model by removing --bayesian
poetry run python -m bayesrul.ncmapss.train_model --archi inception --GPU 0 --model-name My_Model
Or only launch the model on test set if it has already been trained
poetry run python -m bayesrul.ncmapss.train_model --bayesian --GPU 0 --model-name My_Model --test
It's possible to launch a hyperparameter search for LRT on GPU 0
poetry run python -m bayesrul.ncmapss.optimize_single --model lrt --study-name LRT --sampler TPE --GPU 0
In a JSON you can save the best parameter the search tried in the directory and file:
results/ncmapss/best_models/LRT/000.json
With such adictionary of parameters, it is possible to launch a training with these best found parameters. It will read the file and initialize a model accordingly before training.
poetry run python -m bayesrul.ncmapss.train_best_models --model LRT --GPU 0