PRDNN

PRDNN (pronounced "pardon") is a library for provable repair of Deep Neural Networks. DNN behavior involving either finitely-many or entire polytopes of points can be repaired using PRDNN.

The code in this repository is the latest artifact from our paper Provable Repair of Deep Neural Networks, to appear in PLDI 2021 and currently available on arXiv.

@inproceedings{PLDI2021,
  author = {Sotoudeh, Matthew and Thakur, Aditya V.},
  title = {Provable Repair of Deep Neural Networks},
  booktitle = {42nd {ACM} {SIGPLAN} International Conference on Programming Language Design and Implementation ({PLDI})},
  publisher = {ACM},
  year = {2021},
  note = {To appear}
}

Quickstart

Prerequisites

Using as a package

If you only wish to use PRDNN as a package in your own code, and not run any of the experiments in experiments, then the instructions are:

Install Gurobi (we tested 9.0.1 and 9.0.2)
In your project, install the prdnn package from PyPI: python3 -m pip install prdnn.
If you want to do polytope patching, then in another session clone SyReNN and run make start_server. This step is only necessary for polytope patching, not pointwise patching.

Reproducing our experiments

On the other hand, if you wish to reproduce the experiments in our paper, you will need the following prerequisites:

Follow the instructions in bazel_python to build a reproducible version of Python 3.7.4, which will be used by this project. It must be built with the relevant OpenSSL libraries installed.
Install Bazel (we have tested on a variety of versions, including 4.0.0).
Install Gurobi (we tested 9.0.1 and 9.0.2)
If you want to patch the ImageNet model, see ImageNet below.
In another session, clone SyReNN and run make start_server.
Run your desired tests or experiments (see below).

The only supported way to run our experiments is through Bazel and Bazel-Python, as described above. This ensures a reasonably reproducable environment.

However, PRDNN is written entirely in Python and it should be possible in many cases to run the experiments directly. However, in this scenario you will have to manage dependencies and downloading data on your own.

Hardware Requirements

Most of the experiments can be run on consumer-grade laptop hardware with no problems. When prompted for the number of rows to produce, note that the 4-row experiments generally require significantly more memory.

The paper experiments were run using 32 threads and maximum 300 GB of memory.

Running Tests

To run the library unit tests, use:

bazel test //prdnn/...

NOTE: Bazel does not pass the $HOME environment variable into tests. This means that if your Gurobi license file is stored in $HOME/gurobi.lic it will not be picked up by default, causing a failed test. To resolve this, you should explicitly set GRB_LICENSE_FILE before running the tests, e.g.:

GRB_LICENSE_FILE=$HOME/gurobi.lic bazel test //...

To get the coverage report after running the tests, use

bazel run coverage_report

Running Experiments

To run an experiment, use:

bazel run experiments:{experiment_name}

Where {experiment_name} is one of:

squeezenet_repair
mnist_repair
acas_repair

The baselines are experiments:

squeezenet_ft, squeezenet_mft
mnist_ft, mnist_mft
acas_ft, acas_mft

Results from the experiment will be printed, with detailed results placed in experiments/results/{experiment_name}.exp.tgz.

Known Issues

There are currently known issues on macOS. We have tested it successfully on Ubuntu 16.04, 18.04, and 20.04.

ImageNet

Currently, Bazel does not support archives with spaces in path names. This prevents us from using Bazel to manage downloading/unarchiving of the ImageNet-A and ImageNet datasets.

The below instructions are only necessary to run experiments:squeezenet_*.

For the ImageNet-A dataset, it can be downloaded as below:

URL: https://people.eecs.berkeley.edu/~hendrycks/imagenet-a.tar
SHA256: 3bb3632277e6ba6392ea64c02ddbf4dd2266c9caffd6bc09c9656d28f012589e

You should extract it to some place on disk and provide the path when requested by the ImageNet patching script.

In order to evaluate the patched network, you will also need to download and extract the original ImageNet validation set somewhere. AcademicTorrents should have it. We only need the validation set itself, not the surrounding devkit/etc.

95616ARG/PRDNN