05:57 < jenkins-mlpack2> Project mlpack - git commit test build #216: UNSTABLE in 1 hr 4 min: http://ci.mlpack.org/job/mlpack%20-%20git%20commit%20test/216/

06:49 < jenkins-mlpack2> Project mlpack - git commit test build #217: STILL UNSTABLE in 51 min: http://ci.mlpack.org/job/mlpack%20-%20git%20commit%20test/217/

07:38 < jenkins-mlpack2> Project docker mlpack nightly build build #419: STILL UNSTABLE in 3 hr 24 min: http://ci.mlpack.org/job/docker%20mlpack%20nightly%20build/419/

13:39 < ShikharJ> sakshamB: Toshal: Are you guys there?

13:39 < sakshamB> ShikharJ: yes I am here

13:40 < ShikharJ> sakshamB: Great, let's begin. How's Spectral Norm coming up? Do you think you'd be able to wrap it up by next week?

13:42 < sakshamB> ShikharJ: currently I am working only on spectral norm as a wrapper for linear layer. I haven’t been able to generalize it as a wrapper for convolutional layer.

13:42 < sakshamB> if you look here https://github.com/pfnet-research/sngan_projection/tree/13c212a7f751c8f0cfd24bc5f35410a61ecb9a45/source/links

13:42 < sakshamB> ShikharJ they have different classes of spectral norm for different layers

13:43 < sakshamB> ShikharJ: I think I will be able to finish it over the weekend

13:44 < ShikharJ> sakshamB: Okay, cool no pressure. I think rest of your work is only pending review.

13:45 < ShikharJ> sakshamB: Toshal just cleared up the layer limit issue, so we're full steam ahead for merging your work.

13:45 < sakshamB> ShikharJ: yes I think we could merge it soon after review

13:45 < ShikharJ> sakshamB: Please update those PRs as soon as you can.

13:46 < sakshamB> ShikharJ: I have already done that

13:48 < ShikharJ> sakshamB: Perfect. Is there something you need help with from my side?

13:48 < sakshamB> ShikharJ: no nothing right now :)

13:49 < ShikharJ> sakshamB: I'm glad we could get the work done over the summer. I'm anticipating a full release, as soon as your and Toshal's work gets merged in :)

13:50 < sakshamB> ShikharJ: yes that would be great! Thanks for your constant reviews and time. :)

13:51 < ShikharJ> sakshamB: Alright, let's wrap this up. I'll see you on Monday then. Have a nice weekend :) I'll be off for now.

13:51 < sakshamB> ShikharJ: alright have a great weekend!

14:23 < mlutra> Hello. After reviewing the mlpack doc and some threads on github, I didn't find a way to obtain the performance value of each training epoch from the optimized when training a FNN. It would be nice if someone could help me with this. Thanks in advance and thank you for the work on this great library.

15:12 < rcurtin> multra: I don't have the best solution for you today, but I can say that we are about to add callbacks to the ensmallen optimization library, and then we will be able to easily add this support to mlpack's FFN code

15:12 < rcurtin> however, if you want to get the performance value at each epoch, currently the best way is a little bit clunky but it should work:

15:12 < rcurtin> when you call FFN::Train(), pass a custom optimizer that you have configured to only perform one epoch of training

15:13 < rcurtin> then, manually compute the error measure and print it

15:13 < rcurtin> you can do this in a for loop for the number of epochs you are hoping to perform total; each time you call Train(), it should not reset the parameters, so the net result is the same

15:14 < rcurtin> https://github.com/mlpack/models/blob/master/Kaggle/DigitRecognizer/src/DigitRecognizer.cpp#L136

15:14 < rcurtin> there's an example of the idea. like I said, it's just a little clunky, but it will get better soon :)

15:33 < mlutra> Thank you very much for your quick answer, rcurtin. I will try that. BTW, do you have plan to add Levenberg-Marquart method to ensmallen?

15:34 < jeffin143> rcurtin : is there any way to find the type of values store in arma::mat ?

15:35 < jeffin143> Suppose I want to declare a map with key as type of value store in arma::mat so how should I do that

15:35 < jeffin143> Something as map<typeof(arma::mat), size_t> name

15:56 < lozhnikov> jeffin143: arma::mat is an alias for arma::Mat<double>. So, it always stores doubles.

15:56 < sreenik[m]> zoq: I don't seem to be able to figure out what the variables inputParameter and outputParameter do. They are present in most ANN layers, but could you let me know its exact purpose?

15:57 < lozhnikov> jeffin143: I think each armadillo structure has the following typedef: typedef eT elem_type.

15:57 < jenkins-mlpack2> Project mlpack - git commit test build #218: STILL UNSTABLE in 50 min: http://ci.mlpack.org/job/mlpack%20-%20git%20commit%20test/218/

15:59 < lozhnikov> jeffin143: If the type of the matrix is a template parameter e.g. MatType then you could write typename MatType::elem_type

15:59 < jeffin143> Yes just found out that

16:00 < jeffin143> lozhnikov , yeh that was what I was looking for

16:00 < jeffin143> elem_type, thanks :) once again

16:00 < lozhnikov> jeffin143: I think it's not safe to use floating point values as the key due to various machine precision issues.

16:00 < jeffin143> Um , then what should I do..???

16:01 < lozhnikov> Probably it's better to find another way.

16:01 < jeffin143> Ok I will find a way* out

16:07 < zoq> sreenik[m]: They store intermediate results, like the output activation.

16:09 < zoq> mlutra2: Its on my list of optimizers I like to implement but it's not a priority.

16:11 < rcurtin> zoq: someone sent me a link to RAdam today, maybe that's interesting too. there is a FastAI blog post about it but honestly the blog post is of fairly low quality

16:11 < rcurtin> next time someone asks what they can write for ensmallen I'll have an idea though :)

16:12 < rcurtin> in my company we are doing a lot of work that focuses on proximal gradient algorithms, which could be another interesting way to handle constraints

16:12 < rcurtin> just something to think about though :)

16:15 < jeffin143> lozhnikov : you were correct of handling floating point , but if we used lower_bound search instead of find..??? Would it be ok

16:15 < jeffin143> It was for one hot encoding*

16:15 < jeffin143> If the labels are 1 , 1.5 , 2 , 2.5 then to support it we need to map it something

16:16 < jeffin143> Different values*

16:40 < zoq> rcurtin: Will take a look at the blogpost.

17:19 < rcurtin> zoq: the paper could be just fine (I only skimmed it) but I thought the blog post... needed some more clarity and correctness :)

17:25 < zoq> rcurtin: The medium post?

17:30 < lozhnikov> jeffin143: I think there are some cases when lower_bound() doesn't work very well. However you can call lower_bound(X - eps) in order to find X. This should work provided that you choose eps properly.

17:31 < jeffin143> Or may be I can write a function for comparator for map..??? Where I do x-eps

17:31 < jeffin143> Or x-key < eps..???

17:36 < rcurtin> zoq: yeah, I think that was it

17:40 < zoq> jeffin143: New picture, nice.

17:41 < jeffin143> :) thanks zoq

17:41 < jeffin143> Also I got placed in nutanix :) banglore india

17:41 < jeffin143> I will be joining them from January first week

17:42 < zoq> jeffin143: Congratulations!

17:42 < rcurtin> jeffin143: congratulations!

17:42 < rcurtin> :)

17:42 < jeffin143> :) Thanks

17:42 < rcurtin> I know Nutanix is a cloud infrastructure company but every time I hear the name I think that it must be something one spreads on toast...

17:43 < jeffin143> Haha :)

17:43 < zoq> :)

17:45 < rcurtin> I guess I am mixing it up with Nutella

17:46 < sreenik[m]> zoq: Oh, thanks.

17:47 < jeffin143> rcurtin : wasn't there any video conference planned..??

17:47 < rcurtin> jeffin143: I'd like to but I just haven't had the time to catch up and schedule it

17:47 < rcurtin> I wanted to come up with a list of things to address before a release first

17:48 < jeffin143> Ohh , no issues. I thought it was planned and missed the schedule, so just queried :)

17:48 < rcurtin> yeah, dont' worry, I'll send an email to the mlpack list and mention it here

17:48 < rcurtin> I hope to have a chance to get this figured out this weekend, but there is some family in town so they will take first priority :)

17:49 < jeffin143> May be After gsoc , we could finally release with all the work may be

17:51 < jeffin143> lozhnikov : sklearn has something know as CountVectorizer , should I implement it in Mlpack as encoding type..???

17:51 < jeffin143> Encoding ploicy*

17:51 < jeffin143> Policy*

17:56 < lozhnikov> jeffin143: I didn't look at the bindings yet. But I think it's better to spend the last week on the existing PRs.

17:57 < lozhnikov> jeffin143: Perhaps a comparator could work. It depends on the comparator.

17:58 < lozhnikov> I didn't hear about Nutanix. But, congratulations!

18:03 < jeffin143> :) thanks and yeah I will spend last week clearing up the existing PR's

18:05 < lozhnikov> jeffin143: Regarding floating point values as the key: I think lower_bound(X - eps) is the most reasonable way. However, it has its own disadvantages.

18:06 < lozhnikov> For example eps depends on the data. You can't use a static value.

18:07 < lozhnikov> I still think it's better to avoid FP keys.

18:08 < jeffin143> Ok , then we should avoid one hot encoding for double datatypes

18:08 < jeffin143> And only allow int "

18:14 < lozhnikov> jeffin143: Could you point out the exact place where you want to introduce the map?

18:21 < jeffin143> lozhnikov : https://github.com/mlpack/mlpack/blob/57af7320d79fd7e64468673e206bba267c73c269/src/mlpack/core/data/one_hot_encoding_impl.hpp#L41

18:21 < jeffin143> This was buggy , since it is already merged ,. I only noticed after you told me that

18:22 < lozhnikov> jeffin143: I'll look into it.

18:31 < lozhnikov> jeffin143: I don't see a quick solution. I need to think a bit.

18:33 < jeffin143> Sure , even I will think over it , may be we can come up with a solution tomorrow.

22:56 < jenkins-mlpack2> Project mlpack - git commit test build #219: STILL UNSTABLE in 50 min: http://ci.mlpack.org/job/mlpack%20-%20git%20commit%20test/219/