lhj940825 / safe_ml_uncertainty Goto Github PK

This project is developed for Lab course: Safe ML Uncertainty offered by Fraunhofer IAIS.

Developer: Hojun Lim. Zhengjiang Hu

This project can be run to conduct experiments of methods of uncertainty estimation: MC-dropout, Deep Ensemble and Parametric Uncertainty

This project mainly contains the following runnable scripts:

safe_ml_lab_uncertainty/
    1. train_mc.py - training script for MC-dropout
    2. train_de.py - training script for Deep Ensemble
    3. train_pu.py - training script for Parametric Uncertainty
    4. train_ood.py - training script for OOD data study
    5. eval_mc.py - evaluation script for MC
    6. eval_de.py - evaluation script for DE
    7. eval_pu.py - evaluation script for PU
    8. eval_ood.py - evaluation script for OOD

    utils/
        9. dataset.py - script for generating datasets used for the experiments

Once pull the project, the directory should be as follows:

safe_ml_lab_uncertainty/
    configs - this folder and yml files in it should be in the working directory as they include chosen v-noise terms for respective datasets
    data - contain nine UCI datasets
    data_ood - contain nine UCI datasets for experiments with out-of-distribution data
    model 
    scripts
    utils
    .gitignore
    python and jupyter scripts	

To run experiment with a specific method, simply run the corresponding training script first, then evaluation script. Example: train_mc.py -> eval_mc.py for experiment of MC-dropout

Note that it is enough to run a training script only once as long as it remains unchanged. If a training script runs successfully, a corresponding output folder with loss figures as well as checkpoints will be generated. Example: run train_mc.py -> output_mc folder is created and store regarding figures inside. The corresponding evaluation script cannot be run if the training script has not been run for at least once(when the checkpoints are not stored in the corresponding output folder).

The experiment results will be saved in the following root output directories:

safe_ml_lab_uncertainty/
    1. output_mc - results of MC
    2. output_de - results of DE
    3. output_pu - results of PU
    4. output_ood - results of MC-Dropout, Deep Ensemble and Parametric Uncertainty with OOD datasets

Each root output directory contains results of all datasets grouped in individual sub-directories named by the datasets' names. Short summary of experiment results evaluated by RMSE and NLL can be found in err_summary folder in individual root output directory.

(Optional) Plots in Jupyter Notebook:

    1. plot_MC_DE_PU_figure.jpynb: Once run experiments with MC, DE and PU(train_*.py and eval_*.py) and have all figures stored in corresponding output directories, one can run this jupyter file then it will show some important figures for all datasets from respective methods at a glance.
    2. plot_OOD_figure.jpynb: plays the same role as the other jupyter file but this is for the experiemts with ood datasets. run this jupyter file after train_ood.py and eval_ood.py. 

(Optional) Data processing: The datasets have already been processed and saved in 'data', and 'data_ood' for normal datasets and OOD datasets. If one wants to generate the data himself, simply run "safe_ml_lab_uncertainty/utils/dataset.py". Please do not remove 'data' and 'data_ood' directories as they contain original datasets.