Need for help

[chenyi10]

未识别类 'Conv1DLayer' 的方法、属性或字段 'evaluate'。

出错 NN/evaluate_withConns (第 1123 行) y = obj.Layers{source_indx}.evaluate(x);

出错 NN/evaluate (第 130 行) y = obj.evaluate_withConns(x);

出错 NN/falsify (第 697 行) y = obj.evaluate(V(:, i));

出错 safe_uav (第 49 行) counter_inputs = F.falsify(I, U, n_samples);

whether Conv1DLayer can not do the test : falsify, isRobust, safety

[mldiego]

Can you please share the code you are using?

Thanks! Diego

[chenyi10]

net = importNetworkFromTensorFlow('C:\Users\Administrator\Desktop\safe_uav\uav\tf2_model\my_model'); dataArray1 = rand(8,10); dataArray2 = rand(8,10) dlX1 = dlarray(dataArray1,"TSC"); %disp(dlX1) %net = initialize(net, dlX1); summary(net) F = matlab2nnv(net);

%% test 1: falsify

% create NN W = [1 1; 0 1]; b = [0; 0.5]; L = LayerS(W, b, 'poslin'); Layers = {L};

%F = NN(Layers);

% create input set IM = rand(8,10) LB = rand(8,10) UB = rand(8,10) K = ImageStar(IM, LB, UB)

% compute reachability %reachMethod = 'approx-star'; % 使用近似星形集方法 %F.reachMethod = reachMethod; % 设置网络的可达集计算方法 R = F.reach(K); outputs = F.evaluate(K); % Create unsafe region % 定义半空间矩阵和向量 G = rand(8,64);
g = randn(8, 1);
U = HalfSpace(G, g);

% define number of samples to attempt falsification n_samples = 10; % find counter examples counter_inputs = F.falsify(K, U, n_samples); counter_outputs = F.evaluate(counter_inputs);

Here is my code,Thanks!do you have some idea @mldiego

[mldiego]

Could you share the tensorflow model with me?? Which MATLAB version are you using?

[chenyi10]

of course , The version of MATLAB is the latest one. my_model.zip

[mldiego]

I have added an example, based on your code, to evaluate (inference) and run reachability using your model here: https://github.com/verivital/nnv/tree/master/code/nnv/examples/other/issues/chenyi10

I will close the issue now, please let me know if there is anything else we can help with.

Example and fix can be found on #217

[chenyi10]

Thanks for your help,i have one last question,do this model is supported to generate counter inputs and outputs

[mldiego]

Not directly. But you could run random inferences (evaluateSequence), and then check if they meet the specification separately one time-step at a time.