09:00 < bldskspod> Is mlpack participating in GSOC this Year?

09:05 < BigChief> Accepted organisation will be released on 27th FEB

11:01 < kris> So, i am implementing the fanin visitor pattern.

11:02 < kris> I was going through some visitor patterns output width they have a condition that HasCheckModel and HasCheckoOutputWidth.

11:03 < kris> But when i try to run something like boost::apply_visitor(OutputWidthVisitor(), model); this results in a error.

11:03 < kris> I can't understand why. can anyone explain.

11:36 < kris> Okay i think the model layers have to implement the OutputWidth() function. Am i right.

11:45 < kris> At present does any layer implement the OutputWidth funtion.

12:33 < zoq> kris: Hello, if you call boost::apply_visitor(OutputWidthVisitor(), model); make sure model is of type LayerTypes.

12:34 < zoq> kris: Also the conv layer implements the output width function.

12:35 < kris> okay maybe this is a sklearn thing. But isn't model something like ffn. So the model would consists of layers with LayerTypes right.

12:37 < zoq> That is right, but the visitor expects a single layer. You can take a look at line 292 in ffn_impl.hpp.

12:39 < zoq> kris: What you like is a combination of OutputWidthVisitor and ParametersVisitor. If a layer implements the Parameters() function you can do return layer->Parameters().n_rows for the input size and layer->Parameters().n_cols for the output size.

12:42 < zoq> About the model, let's say model is of type ffn you can do:

12:42 < zoq> for (size_t i = 0; i < model.Model().size(); ++i)

12:42 < zoq> {

12:42 < zoq> boost::apply_visitor(OutputWidthVisitor(), model.Model()[i]);

12:43 < zoq> }

12:43 < kris> Okay so model.Model()[i] is equivalent to network[i]

12:44 < zoq> yes, model.Model() returns a the network parameter, which is vector of type std::vector<LayerTypes>

12:45 < kris> Aha thanks.....

12:45 < kris> :)

12:47 < kris> This is more of conceptual question. I have seen visitor patterns in which all the implementation for all the types is given in the overloaded operator method. They say it helps in removing the algorithm from the data structure. But in the case of lets say mlpack we implement all the functions not in the operator but in the layers itself. Is there any specific reason for that.

12:49 < kris> Do you understand the question.

12:49 < kris> zoq

12:57 < zoq> Not sure I complete understand your questions, but if you implement everything in the overloaded operator, doesn't that mean, that every data-structure works in the same way e.g. let's say we have a queue and a stack, I can call insert or remove on both datatypes since there is already an abstraction.

12:58 < zoq> In case of mlpack the abstraction is implemented in each layer, so that the vistor can call e.g. Forward or Backward on the datatype.

13:07 < kris> Actually you overload the implementation for each type in operator method. So we will have one implementation for the stack and another for queue. Same as given in the boost variant example.

13:16 < zoq> I guess in a small project it does make sense to implement some functions in a single class, but I'm not sure it does for a bigger. If someone likes to implement a new Layer outside of mlpack, extending the existing visitor class isn't that easy, but implementing a new class is straightforward.

13:40 < travis-ci> eddelbuettel/mlpack#3 (master - fa2d416 : Dirk Eddelbuettel): The build passed.

13:40 < travis-ci> Change view : https://github.com/eddelbuettel/mlpack/compare/c97358dcc08f...fa2d41624726

13:40 < travis-ci> Build details : https://travis-ci.org/eddelbuettel/mlpack/builds/205268261

13:47 < kris> zoq Yeah i get it now thanks. I will submit the Xavier init PR by tonight or tommrow.

14:06 < kris> zoq: model.Model() gives me error saying that ffn module has no member named Model()

14:08 < kris> *named Model

14:09 < zoq> kris: Ah, I haven't pushed it, you can add: std::vector<LayerTypes>& Model() { return network; } to ffn.hpp

17:07 < travis-ci> mlpack/mlpack#1868 (master - db088bd : Ryan Curtin): The build is still failing.

17:07 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/c97358dcc08f...db088bd156b2

17:07 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/205308161

18:07 < nish21> i noticed something in code for Adam optimizer. in the for loop at line 73 in adam_impl.hpp currentFunction is incremented along with the iteration count. Isn't it missing a modulo numFunctions?

18:13 < nish21> Suppose numFunctions is 3, visitationOrder will have 3 elements. In the 4th iteration of the loop, we get garbage when we try to access visitationOrder[4] right?

18:26 < zoq> nish21: In that case 'if ((currentFunction % numFunctions) == 0)' matches and resets currentFunction, so currentFunction can't be > numFunctions.

18:29 < nish21> ahhh, yes. missed that detail.

18:33 < zoq> nish21: Nevertheless nice that you looked over the code.

18:51 < travis-ci> mlpack/mlpack#1869 (master - 2e565e7 : Ryan Curtin): The build was fixed.

18:51 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/db088bd156b2...2e565e7a1ce8

18:51 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/205329608