Skip to content

NNV 3.0#295

Draft
sammsaski wants to merge 9 commits intoverivital:masterfrom
sammsaski:master
Draft

NNV 3.0#295
sammsaski wants to merge 9 commits intoverivital:masterfrom
sammsaski:master

Conversation

@sammsaski
Copy link
Contributor

Add NNV 3.0 submission folder with instructions on reproducing artifacts.

atumlin and others added 7 commits January 22, 2026 10:13
CodeOcean test runners for FairNNV, ProbVer, ModelStar, VideoStar,
and GNNV with environment setup, README documentation, and test
coverage matching original NNV 3.0 configurations.
MATLAB R2024b's importNetworkFromONNX creates nnet.onnx.layer.AddLayer
variants instead of nnet.cnn.layer.AdditionLayer for ONNX Add operations.
This caused test_regression_acasxu_vnnlib to fail with "Unsupported Class
of Layer". Add contains(class(L), 'AddLayer') check and use direct
constructor to handle both standard and ONNX-imported addition layers.
- Add generate_figures.py for publication figures and LaTeX tables
- Generate ModelStar, GNNV, FairNNV, VideoStar, ProbVer outputs
- Fix ProbVer Direction_trainer to save .mat instead of .npz
- Remove pyenv dependency from ProbReach_ImageStar and Prob_reach
- Make FairNNV and VideoStar paths portable (auto-detect nnvDir)
- Add ProbVer ONNX model, vnnlib properties, and results
- Add FairNNV results (counterfactual/individual fairness CSVs)
Restore matlab2nnv.m, Direction_trainer.py, ProbReach_ImageStar.m,
and Prob_reach.m to their original versions to fix CI test failures.
The codeocean capsule contains a standalone NNV copy with +nnet package
overrides that shadow MATLAB's built-in ONNX layer definitions when
genpath adds them to the path. This caused importNetworkFromONNX to
generate custom AddLayer classes instead of standard AdditionLayer,
breaking test_regression_acasxu_vnnlib and test_matlab2nnv.

Also update ModelStar figure to prevent timing labels from intersecting
the upper axis boundary.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants