Active Surrogate Estimators: An Active Learning Approach to Label-Efficient Model Evaluation

Hi, good to see you here! 👋

This is code for `Active Surrogate Estimators: An Active Learning Approach to Label-Efficient Model Evaluation'.

Setup

We recommend you set up a conda environment like so:

conda-env update -f slurm/environment.yaml
conda activate ase

Reproducing the Experiments

Overview

• The `reproduce' folder contains scripts for running specific experiments.
• Execute a script as

sh reproduce/<script-name>.sh

• You can then create plots with the Jupyter Notebook at

notebooks/plots_paper.ipynb

• All scripts log continuously, so you should be able to create plots as the results are coming in.

Experiments

• To recreate the distribution shift experiments, run the script reproduce/Missing7.sh.
  • To get results in reasonable time, we recommend starting this multiple times in parallel, e.g. across different compute nodes of a cluster. We ran it on ~30 GPUs for ~1 day, where each GPU ran the script three times in parallel (so 100 processes total). Different runs will automatically be combined by the evaluation script.
• To recreate the ResNet experiments, run the scripts reproduce/ResNetCifar10.sh, reproduce/ResNetCifar100.sh, and reproduce/ResNetFMNIST.sh.

Details: Code Structure

• main.py is the main entry point into this code-base.

  • It executes a a total of n_runs experiments for a fixed setup.
  • Each experiment:
    • Trains (or loads) one main model.
    • This model can then be evaluated with a variety of acquisition strategies.
    • Risk estimates are then computed for all acquisition strategies and all risk estimators.

• This repository uses Hydra to manage configs.

  • Look at conf/config.yaml or one of the experiments in conf/... for configs and hyperparameters.
  • Experiments are autologged and results saved to the outputs/ directory.

• Different modules

  • main.py runs repeated experiments and orchestrates the whole shebang.
    • It iterates through all n_runs and acquisition strategies.
  • experiment.py handles a single experiment.
    • It combines the model, dataset, acquisition strategy, and risk estimators.
  • datasets.py, aquisition.py, loss.py, risk_estimators.py. Those should all contain more or less what you would expect.
  • hoover.py is a logging module.
  • models/ contains all models, scikit-learn and pyTorch.
    • In sk2torch.py we have some code that wraps torch models in a way that lets them be used as scikit-learn models from the outside.

And Finally

Thanks for stopping by!

If you find anything wrong with the code, please contact us.

We are happy to answer any questions related to the code and project.