01:03 < sumedhghaisas_> zoq: Hey Marus... are the gradients passing on your machine? cause I was going through my code yesterday ... but whatever changes I make the gradient are still passing so I dont know if I have fixed the issue of not

01:03 < sumedhghaisas_> *or

03:56 < ironstark> zoq: Yes that is the part where I was putting the labels together

03:58 < ironstark> I'll make that change

06:56 < kris1> lozhnikov: Did you look at my comments on Github.

06:58 < lozhnikov> kris1: I have woken up 5 minutes ago

06:59 < kris1> Ohhhh okay…….have a look when you can.......

07:50 < rcurtin> ironstark: zoq: I am finally back in stable, reliable internet zone, I will catch up on the benchmarking PRs tonight and tomorrow

14:59 < kartik_> hi <zoq> a little bit of documentation of cmaes remains. Will get completed in some hours. Can u take a look at the PR and tell your views?

15:01 < zoq> kartik_: Sure, I'll take a look once I get a chance, probably later today or tomorrow.

15:03 < zoq> kartik__: Looks like the nan issue still remains after switching from arma::sort.

15:03 < kartik__> <zoq> about CNE. can u give me some start pointers ..

15:04 < kartik__> i changed the uvec to vec and ran it several times .. i didnt saw one then.. also checked the output also of sort many times ..

15:05 < zoq> just checked the travis build error

15:07 < kartik__> yes after that switched to simple vector and the problem didnt persist. nan refers that the value it got is infinite ? . for CNE. using FFN and to put the weight values and running the model. then again retreiving and modify the values and running the model again. Is there a code where i can see this happening?

15:07 < zoq> About CNE, I guess a good start would be to take a closer look at the exiting CNE class: https://github.com/mlpack/mlpack/pull/753/files

15:10 < kartik__> here the network is made from scratch and ill be using the FFN class. is it possible to retrieve weights and modify and give them back and test the output of the neural network

15:12 < zoq> Yes, the same way as you used for the CMAES method, Parameters() returns the network weights and Evaluate() can be used to test the updated weights.

15:12 < kartik__> ill add into the CNE constructor all the params values .. or should i also go for the params class. Also default values for some parameters can also be set

15:13 < zoq> I would go with using the constructor to set the all values.

15:14 < kartik__> okae..

15:14 < zoq> So, bascially you can create a simple network first, and call the optimizer the same way as you did with the CMAES optimizer. Internally you can use Parameters() and Evaluate().

15:15 < kartik__> oh yes .. sry .. the evaluate gives all the weights one by one ..

15:16 < kartik__> so the model.train is called in FNN

15:16 < kartik__> then the evaluate is called with a function in FNN

15:18 < zoq> yes, something like this:

15:18 < zoq> CNE opt(CNE parameter);

15:18 < kartik__> what about the result of a feed forward pass. how would i get that... and having multiple weights of a population .. im getting a lil confused

15:18 < zoq> model.Train(trainData, trainLabels, opt);

15:20 < kartik__> but there wont be any training data and labels .. so i think i cant use the train method

15:21 < zoq> Inside the CNE optimizer, you can copy the network parameter to create a population. And as I said to get the results of the Forward pass (fitness) you can use the Evaluate function.

15:22 < zoq> In this case the train/label data is our environment.

15:23 < zoq> Just like before you optimized the logistic function on user defined data points.

15:24 < kartik__> that would be good. then to set the parameters back again etc. do we have a test code for this?

15:25 < zoq> kartik__: Not sure I get the point about the test code.

15:25 < zoq> I could sketch something up, if you think that would be helpful, but unfortunately not before tomorrow.

15:26 < kartik__> like after getting the parameters from a network. I modify it and want to test it again for the next fitness value .. how i will do that ?

15:26 < kartik__> it will be done in optimizer ..

15:26 < kartik__> where ill get the network as a function

15:27 < zoq> It's part of the Optimize function, it's the first parameter.

15:28 < kartik__> oh sry

15:28 < zoq> Take a look at the SGD optimizer

15:28 < kartik__> i totally get it now

15:28 < kartik__> just got a little confused

15:28 < kartik__> thats the evalute function

15:28 < zoq> no problem :)

15:28 < kartik__> so i think im good to go for that

15:29 < kartik__> ill start with the work then. Thanks.

15:29 < zoq> okay, I think you could use the SGD class as basis, remove the Gradient call and implement the CNE update routine instead of updatePolicy.Update

15:30 < zoq> in the SGD class case iterate contains the network parameter

15:31 < kartik__> ohkae .. ill ask question as it comes by .. thanks :)

15:31 < zoq> okay, here to help

20:01 < sumedhghaisas_> zoq: Hey Marcus... how I run copy task tests for NTM on the server?

20:02 < sumedhghaisas_> I am still not able to figure out the problem in gradients. Did you had any luck there?

20:02 < sumedhghaisas_> the individual componenets checks out fine. I think there might be small issue in NTM module.