Split and verify an ONNX file into multiple subnets #697
Conversation
maraboupy/Marabou.py
Outdated
| """ | ||
| return MarabouNetworkONNX(filename, inputNames, outputNames, reindexOutputVars=reindexOutputVars) | ||
|
|
||
| def read_onnx_with_threshould(filename, inputNames=None, outputNames=None, reindexOutputVars=True, threshold=None): |
There was a problem hiding this comment.
Typo in first threshold
There was a problem hiding this comment.
Hi @MatthewDaggitt
Thanks for your proactive review!
However, this PR is WIP.
So don't worry about this so much at this point. I'm working on polishing :)
Anyway, thanks again your support, I will update as you commented.
maraboupy/MarabouNetworkONNX.py
Outdated
| for outputName in self.outputNames: | ||
| self.makeGraphEquations(outputName, True) | ||
| # if self.thresholdReached: | ||
| # return |
There was a problem hiding this comment.
Obsolete code to be removed?
maraboupy/Marabou.py
Outdated
| filename (str): Path to the ONNX file | ||
| inputNames (list of str, optional): List of node names corresponding to inputs | ||
| outputNames (list of str, optional): List of node names corresponding to outputs | ||
| threshold (int, optional): Threshold for the number of linear equations |
There was a problem hiding this comment.
A more readable name for this would be max_number_of_lin_equations or something similar. The current name is really not very meaningful, and the comment doesn't clear anything up. What happens when you cross the threshold?
It would also be good to provide in the docs why you would want to limit the number of linear equations?
| from maraboupy import MarabouNetworkONNX | ||
|
|
||
| class MarabouNetworkComposition(MarabouNetwork.MarabouNetwork): | ||
| """Constructs a MarabouNetworkComposition object from an ONNX file |
There was a problem hiding this comment.
This comment could be more useful. We know that the object constructs a MarabouNetworkComposition object. What do objects of this class actually represent? What do they do?
|
|
||
| network = MarabouNetworkONNX.MarabouNetworkONNX(filename, reindexOutputVars=reindexOutputVars, threshold=threshold) | ||
|
|
||
| network.saveQuery('q1.ipq') |
There was a problem hiding this comment.
Lots and lots of hard coded names here. At the very least they need to be heavily documented, but it would be better to make them configurable.
There was a problem hiding this comment.
reindexOutputVars was deleted since it is required
maraboupy/MarabouNetworkONNX.py
Outdated
|
|
||
| # Compute node's shape and create Marabou equations as needed | ||
| if self.thresholdReached: | ||
| return |
There was a problem hiding this comment.
Wouldn't doing the thresholding stuff on the C++ side with the final set of equations after the parsing of the network be a much better way of doing it? Then you would get this feature for all parsers and both Python and C++, rather than simply the ONNX network for the Python backend?
There was a problem hiding this comment.
This is out of scope of this PR. Maybe it will be supported in the future PR.
…xOutputVars from MarabouNetworkComposition
maraboupy/Marabou.py
Outdated
| :class:`~maraboupy.MarabouNetworkComposition.MarabouNetworkComposition` | ||
| """ | ||
| return MarabouNetworkComposition(filename, inputNames, outputNames, | ||
| maxNumberOfLinearEquations=maxNumberOfLinearEquations) |
maraboupy/MarabouNetworkONNX.py
Outdated
|
|
||
| if self.maxNumberOfLinearEquations is not None: | ||
| if not self.thresholdReached and len(self.equList) > self.maxNumberOfLinearEquations: | ||
| if self.splitNetworkAtNode(nodeName, networkNamePostSplit='post_split.onnx'): |
There was a problem hiding this comment.
I think one should probably use a tempfile name, otherwise two threads running in parallel would be interfering
3 participants
This PR is going to support a new feature to split an onnx file and verify and sequentially verify each subnet.