amanpriyanshu / dp-hyperparamtuning

DP-HyperparamTuning offers an array of tools for fast and easy hypertuning of various hyperparameters for the DP-SGD algorithm.

DP-HyperparamTuning

The offical repository for all algorithms and code for the Efficient Hyperparameter Optimization for Differentially Private Deep Learning accepted at PPML Workshop @ ACM-CCS'2021.

A streamlined and basic implementation for all modules presented is available at:

• GitHub Repo
• GitHub Notebook

Note: Colab Demo for the same

Implementation:

Imports:

from DP_HyperparamTuning.experiment.train_single_model import Experiment
from DP_HyperparamTuning.algorithms.bayesian_optimization import Bayesian
from DP_HyperparamTuning.algorithms.grid_search_algorithm import GridSearch
from DP_HyperparamTuning.algorithms.evolutionary_optimization import EvolutionaryOptimization
from DP_HyperparamTuning.algorithms.reinforcement_learning_optimization import RLOptimization

Running Given Modules:

e = Experiment(get_model, criterion, train_dataset, test_dataset)
b = Bayesian(e.run_experiment, calculate_reward, num_limit, search_space_nm=search_space_nm, search_space_lr=search_space_nm)

Where, get_model, calculate_reward are functions, and criterion and train_dataset, test_dataset which are <class 'torch.nn.modules.loss.BCELoss'> and torch.utils.data.Dataset respectively.

Contributing

When contributing to this repository, please first discuss the change you wish to make via issue, email, or any other method with the owners of this repository before making a change. We also make available a CONTRIBUTING.md and CODE_OF_CONDUCT.md for easy communication and quick issue resolution.

Paper Citation:

@misc{priyanshu2021efficient,
      title={Efficient Hyperparameter Optimization for Differentially Private Deep Learning}, 
      author={Aman Priyanshu and Rakshit Naidu and Fatemehsadat Mireshghallah and Mohammad Malekzadeh},
      year={2021},
      eprint={2108.03888},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

dp-hyperparamtuning's Issues

Model architecture evaluated in the paper

I checked your source code closely and found there is a toy model given in your demo.

I want to reproduce your paper's results. Can you share me the model and search space you used to evaluate in the paper?
It would be very nice if you can provide the script used in the paper directly.

Thanks.