Differentially Private Paritioned Variational Inference

Paper

Please find the paper here.

Reproducibility

Requirements:

• Python 3.7
• We use Pipenv to handle Python package installations. Note: we use torch=1.1.0 for compatability with Pipenv. The Pipfile.lock can be used to ensure that you are running with the exact same dependencies as we have used.
• OSX users may need to install libomp e.g. with Homebrew brew install libomp
• A MongoDB is used to store results, and Sacred (Github) is used to place useful information in this database.

Please see the experiments/workshop directory for a detailed description of how to reproduce plots found in our NeurIPS 2019 Privacy in Machine Learning (PiML) Workshop Submission.

Datasets

preprocess_data.py expects the UCI datasets (here) in the data/abalone and the data/adult directories to be pre-processing. Each dataset then creates a labels file, y.csv and additionally several data files, x_ordinal.csv, x_scaled.csv and x_scaled_ordinal.csv

Logging

The Sacred (Github) logging functionality relies on the existence of a Mongo database, by default assumed to be in localhost:9001. If this is on another machine, ssh port forwarding is an easy way to get this working without changing the code. (E.g. for mlg people, if the database is on Hinton: ssh hinton -L 9001:localhost:9001 will connect port 9001 on Hinton to port 9001 on your machine. In order to bring up and bring down the database, suitable scripts are in /scratch/mjh252/mongo)

Omniboard (Github) is an excellent way to view results in a Sacred database.

Deleting database entries can be done as follows, in the mongo shell:

use sacred # use the correct database name
db.runs.remove({'config.experiment_tag':'....'}) # apply the correct filter
db.metrics.remove({run_id: {$gt: 333, $lt: 600}}) # remove metrics, linked by run id

Beware: whilst this removes runs from the omniboard view window, it will not actually delete the log files!!