Parsing dubinsrejoin.onnx (VNN-Comp 2022)

[phK3]

When running VeriNet on the dubins_rejoin instances of VNN-Comp 2022 using
benchmark(result_path, benchmark_path, instances_csv_path, input_shape=(8,), transpose_fc_weights=False), I get the following error:

ValueError: Expected at least one connection to non-input node: VeriNetNNNode(idx: 1, op: Linear(in_features=256, out_features=8, bias=True), to: [2], from: [])

The error gets thrown at VeriNet/verinet/sip_torch/sip.py, line 337, in _process_node.

Parsing for other networks works fine.

Thanks in advance for your help.

[pat676]

Hi,

The ONNXParser uses strings to identify the input nodes. It provides some default input names which cover most cases; however, dubinsrejoin.onnx seems to have a non-standard name.

There are a few solutions for automatically inferring input nodes from the ONNX model. However, we do not actively maintain this repo anymore besides small bug fixes; thus a fix will most likely not be implemented in the near future.

My proposed solution is to change line 69 of benchmark.py from

model_onnx = ONNXParser(model_path)

to:

model_onnx = ONNXParser(model_path, input_names= ["dense_input"], transpose_matmul_weights=True)

Note that I also added "transpose_matmul_weights=True" which I believe is necessary for this benchmark.

Hope this helps!

[phK3]

It worked, when I changed line 74 in benchmark_vnnlib.py to

ONNXParser(model_path, input_names=["dense_input"], transpose_fc_weights=True)

(using transpose_fc_weights) if the model_path contains dubinsrejoin.

Thanks for your help!

[pat676]

Great! I'm closing this issue then