00:29 < sumedhghaisas_> rcurtin: Hey Ryan... there? I have this small doubt for you

09:41 < travis-ci> mlpack/mlpack#2679 (master - 0a0c51d : Yannis Mentekidis): The build was broken.

09:41 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/a911eddf936f...0a0c51dc6a6e

09:41 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/247852992

09:49 < shikhar> Looks like KRANNTest is failing reproducibly with CTest on Travis.

11:28 < rcurtin> shikhar: yeah, it seems to fail most of the time

11:29 < rcurtin> since it still hasn't been solved I think for now we can comment the failing part out; I don't think anything is wrong with the implementation... needs more investigation

12:16 < kris1> What is the offset value in. the gradient function incase of the minibatch sgd

12:20 < zoq> kris1: The begin/index of the current batch, in Minibatch SGD visitationOrder holds the number of batches and not the samples to iterate over e.g. visitationOrder[1, 2, 3, 4] and BatchSize = 4, the second batch starts with index 8.

12:57 < zoq> kris: Did you see my message?

13:00 < kris1> Yes

13:00 < kris1> I understood it now…thanks….. :)

13:00 < zoq> okay, great, here to help :)

13:07 < kris1> Also how can we test the serialise function

13:09 < kris1> I did not find the test for serialise function test for ffn

13:11 < zoq> Unfortunately there is no serialization test for the FFN class, which is something that should be fixed in the near future. About how to test the serialization function, you can take a look at the serialization_test.cpp.

13:13 < kris1> So i wrote the serialisation looking at ffn class. I did not understand how the internal layers are serialised meaning the serialisation function for the layer is not called

13:18 < zoq> kris1: Take a look at e.g. how serial decision_tree_impl.hpp

13:18 < zoq> serial

13:18 < zoq> is implemented

13:19 < zoq> especially "ar & CreateNVP(*children[i], name.str());"

14:34 < shikhar_> rcurtiv

14:34 < shikhar_> rcurtin: I am unable to replicate the issue here. I'll try and see if I can find something.

14:55 < kris1> I don’t understand the seralisation test particularly well. I think we define a couple of models and the we all the serializeobjectall and then test the models to see their parmaeters are equal or not.

14:55 < kris1> But shouldn’t we load the models and then test if the parmeters are equal

15:09 < lozhnikov> kris1: yes, we should create an object, then save the object to a file; load another object from the file and check that we got an equivalent copy

15:10 < kris1> Yes but that’s not how the test are written in the seralization_test.cpp file

15:10 < kris1> that’s what i don’t understand

15:10 < lozhnikov> for example?

15:12 < kris1> Look at the LogisticRegressionTest in the seralisation_test.cpp file

15:13 < kris1> the models are created but not saved anywhere

15:13 < lozhnikov> SerializeObjectAll

15:13 < lozhnikov> take a look at serialization.hpp

15:17 < kris1> got it thanks

22:38 < sumedhghaisas_> rcurtin: Hey Ryan... got some time?