search

Verified-Intelligence / auto_LiRPA

auto_LiRPA: An Automatic Linear Relaxation based Perturbation Analysis Library for Neural Networks and General Computational Graphs
https://arxiv.org/pdf/2002.12920
Other
282 stars 69 forks source link

Possible Bug? #35

Open j29scott opened 1 year ago

j29scott commented 1 year ago

I got a shape error here. This fix works for acas .

huanzhang12 commented 1 year ago

@j29scott Thank you for reporting the issue to us! Could you provide a full example that can reproduce the bug so we can take a look?

j29scott commented 1 year ago

I can prepare a short script to make it explicit in a few days.

But this error arose when applying: from onnx2pytorch import ConvertModel

On the net ACASXU_run2a_1_1_batch_2000.onnx

With perturbation PerturbationLpNorm(x_L=torch.Tensor(lower), x_U=torch.Tensor(upper))

where lower=array([ 0.26897843, 0.11140846, -0.4999999 , 0.22727273, -0. ]) and upper=array([ 0.67985777, 0.4999999 , -0.49840835, 0.5 , 0.5 ]) (property 10 (computed from dnnv's disjunction))

Edit: made few fixes

huanzhang12 commented 1 year ago

Thanks! This might be an issue in model conversion. We look forward to more details on reproducing this issue. In the meanwhile, it might be related to version problems of certain packages. You can also try to use the conda environment defined here: https://github.com/Verified-Intelligence/alpha-beta-CROWN/blob/main/complete_verifier/environment.yml