VeriNet
The VeriNet toolkit is a state-of-the-art sound and complete symbolic interval propagation based toolkit for verification of neural networks. VeriNet won second place overall and was the most performing among toolkits not using GPUs in the 2nd international verification of neural networks competition. VeriNet is devloped at the Verification of Autonomous Systems (VAS) group, Imperial College London.
Relevant Publications.
VeriNet is developed as part of the following publications:
Efficient Neural Network Verification via Adaptive Refinement and Adversarial Search
This version of VeriNet subsumes the VeriNet toolkit publised in the first paper and the DeepSplit toolkit published in the second paper.
Installation:
All dependencies can be installed via Pipenv.
VeriNet depends on the Xpress solver, which can solve smaller problems without a license; however, larger problems (networks with more than approximately 5000 nodes) require a license. Free academic licenses can be obtained at: https://content.fico.com/l/517101/2018-06-10/3fpbf
We recommend installing the developer dependencies for some extra optimisations during the loading of onnx models. These can be installed by running pipenv with the --dev option: $pipenv install --dev.
Usage:
Models:
VeriNet supports loading models in onnx format or custom models created with the VeriNetNN class, a subclass of torch.nn.module.
Loading onnx models:
Onnx models can be loaded as follows:
from verinet.parsers.onnx_parser import ONNXParser
onnx_parser = ONNXParser(onnx_model_path, input_names=("x",), transpose_fc_weights=False, use_64bit=False)
model = onnx_parser.to_pytorch()
model.eval()
The first argument is the path of the onnx file; input_names should be a tuple containing the input-variable name as stored in the onnx model; if transpose_fc_weights is true the weight matrices of fully-connected layers are transposed; if use_64bit is true the parameters of the model are stored as torch.DoubleTensors instead of torch.FloatTensors.
Custom models:
The following is a simple example of a VeriNetNN model with two inputs, one FC layer, one ReLU layer and 2 outputs:
import torch.nn as nn
from verinet.neural_networks.verinet_nn import VeriNetNN, VeriNetNNNode
nodes = [VeriNetNNNode(idx=0, op=nn.Identity(), connections_from=None, connections_to=[1]),
VeriNetNNNode(idx=1, op=nn.nn.Linear(2, 2)(), connections_from=[0], connections_to=[2]),
VeriNetNNNode(idx=2, op=nn.ReLU(), connections_from=[1], connections_to=[3]),
VeriNetNNNode(idx=3, op=nn.Identity(), connections_from=[2], connections_to=None)]
model = VeriNetNN(nodes)
VeriNetNN takes as input a list of nodes (note that 'nodes' here do not correspond to neurons, each node may have multiple neurons) where each node has the following parameters:
- idx: A unique node-index sorted topologically wrt the connections.
- op: The operation performed by the node, all operations defined in verinet/neural_networks/custom_layers.py as well as nn.ReLU, nn.Sigmoid, nn.Tanh, nn.Linear, nn.Conv2d, nn.AvgPool2d, nn.Identity, nn.Reshape, nn.Transpose and nn.Flatten are supported.
- connections_from: A list of which nodes' outputs are used as input in this node. Note that more than one output in a single node (residual connections) is only support for nodes with the AddDynamic op as defined in custom_layers.py.
- connections_to: A list of which nodes' input depend on this node's output corresponding to connections_from.
The first and last layer should be nn.Identity nodes. BatchNorm2d and MaxPool2d operations can be implemented by saving the model to onnx and reloading as the onnx parser automatically attempts to convert these to equivalent Conv2d and ReLU layers.
Verification Objectives:
VeriNet supports verification objectives in the VNN-COMP'21 vnnlib format and custom objectives.
Vnnlib:
VeriNet supports vnnlib files formated as described in the following discussion: https://github.com/stanleybak/vnncomp2021/issues/2. The files can be loaded as follows:
from verinet.parsers.vnnlib_parser import VNNLIBParser
vnnlib_parser = VNNLIBParser(vnnlib_path)
objectives = vnnlib_parser.get_objectives_from_vnnlib(model, input_shape)
The vnnlib_path parameter should be the path of the vnnlib file, model is the VeriNetNN model as discussed above while the input shape is a tuple describing the shape of the models input without batch-dimension (e.g. (784, ) for flattened MNIST (1, 28, 28) for MNIST images and (3, 32, 32) for CIFAR-10 Images).
Custom objectives:
The following is an example of how a custom verification objective for classification problems can be encoded (correct output larger than all other outputs):
from verinet.verification.objective import Objective
objective = Objective(input_bounds, output_size=10, model=model)
out_vars = objective.output_vars
for j in range(objective.output_size):
if j != correct_output:
# noinspection PyTypeChecker
objective.add_constraints(out_vars[j] <= out_vars[correct_output])
Here input bounds is an array of shape (*network_input_shape, 2) where network_input_shape is the input shape of the network (withut batch dimension) and the last dimension contains the lower bounds at position 0 and upper bounds at position 1.
Note that the verification objective encodes what it means for the network to be Safe/Robust. And-type constraints can be encoded by calling objective.add_constraints for each and-clause, while or-type constraints can be encoded with '|' (e.g. (out_vars[0] < 1) | (out_vars[0] < out_vars[1]))
.
Verification:
After defining the model and objective as described above, verification is performed by using the VeriNet class as follows:
from verinet.verification.verinet import VeriNet
solver = VeriNet(use_gpu=True, max_procs=None)
status = solver.verify(objective=objective, timeout=3600)
The parameters of VeriNet, use_gpu and max_procs, determines whether to use the GPU and the maximum number of processes (max_procs = None automatically determines the number of processes depending on the cores available).
The parameters in solver.verify correspond to the objective as discussed above and the timeout in seconds. Note that is recommended to keep solver alive instead of creating a new object every call to reduce overhead.
After each verification run the number of branches explored and maximum depth reached are stored in solver.branches_explored and solver.max_depth, respectively. If the objective is determined to be unsafe/not-robust, a counter example is stored in solver.counter_example.
At the end of each run, status will be either Status.Safe, Status.Unsafe, Status.Undecided or Status.Underflow. Safe means that the property is robust, Unsafe that a counter example was found, undecided that the solver timed-out and underflow that an error was encountered, most likely due to floating point precision.
Advanced usage:
Environment variables:
The .env file contains some environment variables that are automatically enabled if pipenv is used, if pipenv is not used make sure to export these variables.
Config file:
The config file in verinet/util/config.py contains several advanced settings. Of particular interest are the following:
- PRECISION: (32 or 64) The floating point precision used in SIP. Note that this does not affect the precision of the model itself, which can be adjusted in the ONNXParser as discussed above.
- MAX_ESTIMATED_MEM_USAGE: The maximum estimated memory usage acceptable in SIP. Can be reduced to reduce the memory requirements at the cost of computational performance.
- USE_SSIP and STORE_SSIP_BOUNDS: Performs a pre-processing using a lower cost SIP-variant. Should be enabled if the input-dimensionality is significantly smaller than the size of the network (e.g. less than 50 input nodes with more than 10k Relu nodes).
Authors:
Patrick Henriksen: [email protected]
Alessio Lomuscio.