def net_update(network: networks.NeuralNetwork) -> networks.NeuralNetwork:
if not network.up_to_date:
for alt_rep in network.alt_rep_cache:
if alt_rep.up_to_date:
if isinstance(alt_rep, cv.ONNXNetwork):
return cv.ONNXConverter().to_neural_network(alt_rep)
elif isinstance(alt_rep, cv.PyTorchNetwork):
return cv.PyTorchConverter().to_neural_network(alt_rep)
else:
raise NotImplementedError
else:
return network
in_pred_bias.append([-norm_input_lb[m]])
in_pred_bias.append([norm_input_ub[m]])
in_pred_bias = np.array(in_pred_bias)
in_pred_mat = np.array(in_pred_mat)
out_pred_mat = np.array(unsafe_mats[k])
out_pred_bias = np.array(unsafe_vecs[k])
prop = ver.NeVerProperty(in_pred_mat, in_pred_bias,
[out_pred_mat], [out_pred_bias])
# prop = ver.NeVerProperty(in_pred_mat, in_pred_bias, unsafe_mats, unsafe_vecs)
input_prefix = network.input_id
output_prefix = network.get_last_node().identifier
ver.temp_never2smt(prop, input_prefix, output_prefix,
f"smt_property/SMT_{p_id[k]}.smt2")
p2 = reading.SmtPropertyParser(
ver.SMTLIBProperty(f"smt_property/SMT_{p_id[k]}.smt2"),
input_prefix, output_prefix).parse_property()
print(p2)
prop_set = True
onnx_net = conv.ONNXConverter().from_neural_network(network)
pytorch_net = conv.PyTorchConverter().from_neural_network(network)
torch.save(pytorch_net.pytorch_network,
f"acas_pytorch/{network.identifier}.pt")
onnx.save(onnx_net.onnx_network,
f"acas_onnx/{network.identifier}.onnx")
lr_property = ver.NeVerProperty(in_pred_mat, in_pred_bias, out_pred_mat,
out_pred_bias)
master_log_file = "logs/MNISTExperimentLog.txt"
ver_param_sets = [[False, 0, False, 0]]
param_set_id = ["Over-Approx"]
with open(master_log_file, "a") as master_log:
master_log.write(f"Dataset,NetworkID,Methodology,Safety,Time\n")
for net_id in net_ids:
net_path = "mnist_nets/" + net_id + ".onnx"
net = conv.ONNXConverter().to_neural_network(
conv.ONNXNetwork(net_id, onnx.load(net_path)))
assert isinstance(net, networks.SequentialNetwork)
current_node = net.get_first_node()
while current_node is not None:
if isinstance(current_node, nodes.FullyConnectedNode):
current_node.weight = np.where(
np.abs(current_node.weight) < 0.000001, 0, current_node.weight)
current_node = net.get_next_node(current_node)
for p_set in ver_param_sets:
log_filepath = f"logs/MNIST_NET={net_id}_{param_set_id[0]}.txt"
print(f"MNIST_NET={net_id}_{param_set_id[0]}")
verifier = ver.NeverVerification(log_filepath,
neuron_relevance=p_set[0],
print("DROPOUT NODE TEST")
start_network = network.SequentialNetwork("NET_TEST", "X")
start_network.add_node(nodes.DropoutNode("Dropout_1", (3, 4, 5, 6)))
alt_network = converter.from_neural_network(start_network)
end_network = converter.to_neural_network(alt_network)
assert isinstance(end_network, network.SequentialNetwork)
start_node = start_network.get_first_node()
end_node = end_network.get_first_node()
assert isinstance(start_node, nodes.DropoutNode)
assert isinstance(end_node, nodes.DropoutNode)
assert start_node.p == end_node.p
assert start_node.identifier == end_node.identifier
converters = [conversion.ONNXConverter(), conversion.PyTorchConverter()]
for conv in converters:
print(f"Test for {conv.__class__.__name__}")
relu_node_test(conv)
sigmoid_node_test(conv)
fully_connected_node_test(conv, True)
fully_connected_node_test(conv, False)
batchnorm_node_test(conv)
conv_node_test(conv, True)
conv_node_test(conv, False)
averagepool_node_test(conv)
maxpool_node_test(conv)
lrn_node_test(conv)
softmax_node_test(conv)
unsqueeze_node_test(conv)
reshape_node_test(conv)