05:25 < toshal> ShikharJ: How you used to test GAN models? I am using g++ to compile and run the ann model I have created. Let me know if there is a another way to run and test the models on savannah.

06:58 < ShikharJ> toshal: I merely changed our test cases to save the output towards the end, ran make, and did a bin/mlpack_test

07:48 < jenkins-mlpack2> Project docker mlpack nightly build build #350: STILL UNSTABLE in 3 hr 34 min: http://ci.mlpack.org/job/docker%20mlpack%20nightly%20build/350/

12:40 < jeffin143> zoq , rcurtin : for last 5 commits Travis build on xocde fails

12:40 < jeffin143> Users/travis/build/mlpack/mlpack/src/mlpack/tests/main_tests/radical_test.cpp:142: fatal error: in "RadicalMainTest/RadicalDiffNoiseStdDevTest": critical check arma::accu(Y == CLI::GetParam<arma::mat>("output_ic")) < Y.n_elem has failed [15 >= 15]

12:40 < jeffin143> Radical_test

13:22 < jeffin143> lozhnikov : while doing serialization , it throws up an error that boost::basic_string_view has no member serialize*

13:23 < jeffin143> The same error was thrown for deque as well , but I included boost/serialization/deque.hpp and it was fine then, but couldn't find how to do for the string_view

13:24 < lozhnikov> jeffin143: sure. There is no sense in that. string_view is just a link to a string.

13:24 < jeffin143> So I should just serialize , deque

13:24 < jeffin143> ?*

13:24 < lozhnikov> the deque and the labels

13:25 < jeffin143> Labels , map keys..??

13:26 < lozhnikov> So, if you are saving data, you should save the deque and the labels. If you are loading data, you should load the deque, the labels and initialize the map manually

13:26 < lozhnikov> The keys of the map are just string views. You can't serialize them. You should serialize the values.

13:27 < jeffin143> Ohh ok so I should save unordered_map, with keys as deque elements and their labels , and then while loading I, should manually create the the whole thing again

13:28 < lozhnikov> yes, sure

13:29 < jeffin143> Thanks :)

13:48 < jeffin143> lozhnikov : also can u tell me ur versio

13:48 < jeffin143> Version of boost , so that I can keep that as upper bound for if else for specialisation of boost::hash

13:51 < lozhnikov> jeffin143: I use boost 1.69. However the specialization of boost::hash could appear earlier. So the best choice is to look through the boost headers and find that out.

13:51 < jeffin143> Ok I ll do it*

15:39 < jeffin143> rcurtin : you there..??

15:46 < favre49> zoq: To prevent the generation of cycles in acyclic cases, I was planning to use the node depths we find when we create an AcyclicNet instance. Then we can choose to only create connections where the target node is less deep than the source node.

15:47 < favre49> But that would mean we would be creating an AcyclicNet instance during both evaluation and during mutation.

15:49 < favre49> Instead, i was thinking maybe we could split the Genome into CyclicGenome and AcyclicGenome classes, and in the AcyclicGenome class i could make the AcyclicNet a member variable which could be used during the mutation.

15:49 < favre49> That would be a departure from the user interface we discussed, though I don't think its too bad a change. What do you think?

15:51 < rcurtin> jeffin143: yeah, maybe the tolerance needs to be adjusted, I'll take care of it today

15:52 < jeffin143> rcurtin : No the doubt was not about that , I wanted to know whether param_model_in(), can take template name

15:53 < jeffin143> I mean suppose param_model_in(t) and then later t is assigned a class?

15:54 < rcurtin> I'm not sure what you mean

15:54 < rcurtin> when you write PARAM_MODEL_IN(ModelClassType, ...), we have to know what ModelClassType is

15:54 < rcurtin> if the root of the problem is that you have one of several classes you might want to serialize, then make a "wrapper" class that holds each of them

15:55 < rcurtin> as an example, you can take a look at AdaboostModel from src/mlpack/methods/adaboost/adaboost_model.hpp

15:55 < rcurtin> basically that class holds an Adaboost<Perceptron> and an Adaboost<DecisionStump> but either one can be used

15:56 < rcurtin> this allows the user to have an Adaboost class with either as a weak learner

15:56 < jeffin143> Oh yes , this was the one what I was looking for

15:56 < rcurtin> :)

15:56 < jeffin143> I have many class and then, would like to have one

15:56 < jeffin143> Thanks :)

17:04 < zoq> favre49: Agreed, the user interaction with the model remains the same, but this will definitely improve the runtime performance of the model.

17:38 < favre49> Okay thanks, I'll make the changes soon.