00:42 < gmanlan> Hi rcurtin: I noticed that the sample app is showing how to get some metrics such as precision, F1, etc. from the cross-validation, but I'm not sure that's what is actually happening

00:42 < gmanlan> https://github.com/mlpack/mlpack/blob/master/doc/examples/sample-ml-app/sample-ml-app/sample-ml-app.cpp#L85

00:43 < gmanlan> accuracy is for sure coming from CV, but precision, recall and F1? that seems to be using the previously trained RF, not the CV ones

00:43 < gmanlan> is that accurate?

03:52 < Yashwants19> Hi gmanlan: I think so these are correct.

03:55 < Yashwants19> If you see the code of cross validation when we call Evaluate first it reset the modelPtr after passing the respectively parameter of the model

03:57 < Yashwants19> Then accuracy will also use the pre trained model(RF) only

03:58 < Yashwants19> Same as Precision, Recall and F1 score

04:03 < Yashwants19> I think so We can also use modelPtr to evaluate accuracy (but it requires cross validated trainingSet, labels)

04:03 < Yashwants19> But they require testSet and testLabels

05:08 < jeffin143> hey : Can anyone help me debugging this code : https://pastebin.com/GW5EG630

05:09 < jeffin143> How to pass Class function as argument, I ain't able to convert it the required form.

05:10 < jeffin143> UnCommenting the line in above code would throw an error

05:12 < jeffin143> Never Mind, I found the way out :) :-p

07:33 < karthik> Hello everyone!

07:37 < jenkins-mlpack2> Project docker mlpack nightly build build #374: STILL UNSTABLE in 3 hr 23 min: http://ci.mlpack.org/job/docker%20mlpack%20nightly%20build/374/

13:48 < zoq> karthik: Hello.

13:49 < zoq> jeffin143: Great :)

15:32 < sumedhghaisas> KimSangYeon-DGU: Hey :)

15:32 < sumedhghaisas> hows it going?

15:33 < KimSangYeon-DGU> Hi Sumedh

15:33 < KimSangYeon-DGU> Can you check the code? https://github.com/KimSangYeon-DGU/GSoC-2019/tree/master/Research/Optimization/Experiments

15:34 < KimSangYeon-DGU> I implemented QGMM with Adam optimizer for minimizing NLL + lambda * approx constant using tensorflow

15:34 < sumedhghaisas> Sure. Is that the EM updates one?

15:34 < KimSangYeon-DGU> https://docs.google.com/document/d/149GNgC2qVhXnrNPoT11XvlqTtrIGi1TWP3wSymuaqKE/edit

15:34 < KimSangYeon-DGU> Yeah

15:34 < KimSangYeon-DGU> Opps

15:34 < KimSangYeon-DGU> https://docs.google.com/document/d/149GNgC2qVhXnrNPoT11XvlqTtrIGi1TWP3wSymuaqKE/edit?usp=sharing

15:34 < KimSangYeon-DGU> Sorry it is exact link

15:36 < KimSangYeon-DGU> But I don't feel good about the result.

15:36 < sumedhghaisas> hmmm... So both the centers are moving towards each other

15:36 < sumedhghaisas> right?

15:36 < KimSangYeon-DGU> Yeah

15:37 < sumedhghaisas> just looking main.py

15:37 < sumedhghaisas> seems okay

15:37 < KimSangYeon-DGU> Yeah

15:38 < sumedhghaisas> but from next time could you do 80 character limit? :) Its little frustrating to scroll horizontally

15:38 < KimSangYeon-DGU> Ah

15:38 < KimSangYeon-DGU> Got it, sorry

15:38 < KimSangYeon-DGU> I'll edit it

15:39 < sumedhghaisas> you can do it next time no worries :)

15:39 < sumedhghaisas> the code looks correct

15:39 < sumedhghaisas> just for the clip_by_value

15:39 < sumedhghaisas> I agree with it

15:39 < sumedhghaisas> did you just in case try without?

15:39 < KimSangYeon-DGU> Yeah

15:39 < sumedhghaisas> I think you will get NAN

15:40 < KimSangYeon-DGU> Yeah

15:40 < KimSangYeon-DGU> exactly

15:40 < sumedhghaisas> seems fair

15:40 < sumedhghaisas> oky lets think on this for a sec

15:40 < sumedhghaisas> so we are doing MLE

15:40 < sumedhghaisas> with constraint optimization

15:41 < sumedhghaisas> we dont have exact constraints

15:41 < sumedhghaisas> hmm

15:41 < sumedhghaisas> NLL + lambda * approx contraint doesn't work

15:41 < KimSangYeon-DGU> Hmm

15:41 < KimSangYeon-DGU> So, I tried to find something to read

15:42 < KimSangYeon-DGU> How about https://arxiv.org/pdf/1711.05376.pdf?

15:42 < KimSangYeon-DGU> https://arxiv.org/pdf/1711.05376.pdf

15:42 < KimSangYeon-DGU> Sorry

15:42 < sumedhghaisas> We can formalize the problem in their formulation in the final document

15:42 < KimSangYeon-DGU> Ah yeah

15:43 < sumedhghaisas> looking at it now

15:43 < KimSangYeon-DGU> I'm not sure it can be applied

15:43 < KimSangYeon-DGU> to QGMM

15:44 < sumedhghaisas> very interesting indeed

15:44 < sumedhghaisas> I like their formulation

15:44 < sumedhghaisas> Did you go through the paper?

15:44 < KimSangYeon-DGU> Actually, I just read introduction and result

15:45 < KimSangYeon-DGU> I did focus on NLL optimization until now

15:45 < sumedhghaisas> ahh I think it could be worth to go over it a bit

15:45 < KimSangYeon-DGU> Really? Oh, I'll look into the paper more deeply

15:46 < sumedhghaisas> Their formulation does not look at the MLE estimate but

15:46 < sumedhghaisas> wait... but it won't still solve the normlization constant problem

15:46 < sumedhghaisas> hmm

15:47 < KimSangYeon-DGU> Yeah, and the sum of alphas of QGMM isn't one.

15:48 < sumedhghaisas> I am not still sure that should be the case

15:48 < sumedhghaisas> the weights of GMM should be summed to one

15:48 < KimSangYeon-DGU> Agreed

15:48 < sumedhghaisas> but for QGMM that is not the case

15:48 < KimSangYeon-DGU> Yes

15:48 < sumedhghaisas> ahh ou meant it that way

15:48 < sumedhghaisas> sorry sorry

15:48 < sumedhghaisas> I misunderstood

15:49 < KimSangYeon-DGU> I'm okay :)

15:50 < KimSangYeon-DGU> Hmm... Would it be a good way to reformulate the equations in the original paper? or should we try other approaches.

15:50 < KimSangYeon-DGU> ?

15:50 < sumedhghaisas> hmm

15:50 < sumedhghaisas> thinking

15:51 < sumedhghaisas> another way to do variational inferenece and maximize lower bound on MLE than actual MLE

15:51 < sumedhghaisas> but that would also involve problems with normalization

15:52 < sumedhghaisas> this normalization turning out to be a bigger problem than I anticipated

15:52 < KimSangYeon-DGU> Agreed

15:52 < sumedhghaisas> how many weeks are remaining of our project?

15:52 < KimSangYeon-DGU> About 6 weeks?

15:54 < KimSangYeon-DGU> 6~7 weeks left

15:54 < sumedhghaisas> I have couple of ideas about the stuff but that involves way more sampling and unconventional methods

15:55 < KimSangYeon-DGU> I'll try them

15:55 < sumedhghaisas> and would involve lot of reading on your side including variational inferenece

15:55 < KimSangYeon-DGU> I'm okay

15:55 < sumedhghaisas> I am thinking if its better if we approach a different problem and get some implementation

15:55 < KimSangYeon-DGU> I want to make this project complete.

15:56 < sumedhghaisas> me too. But the timeline seems very tight.

15:56 < sumedhghaisas> hm... thinking...

15:56 < gmanlan> Yashwants19: thanks for your response. Although I still believe we are missing something, if you take a look at the sample code in the app, it is only Accuracy that is using cv.Evaluate (therefore reusing the CV model/s) - but then Precision, Recall and F1 are not using cv.Evaluate but "rf" which is the single-trained model (not the CV one) - so Pr

15:56 < gmanlan> ecision, Recall and F1 are not Cross-Validation metrics, right?

15:57 < KimSangYeon-DGU> Approaching a different problem and implementation is good to me

15:58 < sumedhghaisas> Actually we can keep looking at solutions on the side. But What if we do small projects each week and get them done while we focus on research? What do you think about that?

15:59 < sumedhghaisas> I don't want to give up on this project yet but I think it involves little bit more reading.

15:59 < sumedhghaisas> that should give us some time to explore some small implementations

15:59 < sumedhghaisas> its up to you

15:59 < KimSangYeon-DGU> I'm good

15:59 < sumedhghaisas> We can do all reading or half half

16:00 < KimSangYeon-DGU> I'm ready to spend more time on this project :)

16:01 < sumedhghaisas> Ahh no I dont think you should spend more than alloted time daily. That is counter productive trust me. I do this all the time at work :P

16:01 < KimSangYeon-DGU> Yeah :)

16:01 < sumedhghaisas> hmmm... okay lets meet Thursday to discuss this? Till then can we go over the mentioned paper?

16:01 < sumedhghaisas> worse case we will try implementing that in MLPACK

16:01 < sumedhghaisas> what say?

16:01 < sumedhghaisas> I will also take some help from Ryan

16:02 < sumedhghaisas> the paper you mentioned I mean

16:03 < KimSangYeon-DGU> Yeah, let meet Thursday

16:03 < sumedhghaisas> I am sorry that I am little stuck on this normalization constant problem. :(

16:03 < KimSangYeon-DGU> Nope!

16:03 < KimSangYeon-DGU> It is our problem

16:03 < KimSangYeon-DGU> I'm also sorry

16:03 < sumedhghaisas> haha... thats a good attitude

16:04 < sumedhghaisas> okay lets go over the paper

16:04 < KimSangYeon-DGU> Yeah

16:04 < sumedhghaisas> and decide by Thursday?

16:04 < KimSangYeon-DGU> Yeah

17:09 < sumedhghaisas> KimSangYeon-DGU: Hey Kim

17:09 < sumedhghaisas> I think I found a bug in the code

17:10 < sumedhghaisas> there should a stop gradient on Q distribution