14:51 < alsc> zoq: I am writing a tiny decoder in vanilla C for models trained with mlpack ANN: I don't understand the negativeloglikelihood_impl file... shouldn't it be -log() of the input columns?

15:02 < zoq> alsc: It is expected that the input contains log-probabilities.

15:02 < alsc> ok..true sorry.. needs LogSoftMax first

15:03 < zoq> alsc: Right or you can now use The CrossEntropy function.

15:04 < zoq> https://github.com/mlpack/mlpack/blob/master/src/mlpack/methods/ann/layer/cross_entropy_error.hpp

15:08 < alsc> zoq: i trained with softmax, will keep it that way for now

15:09 < alsc> I have a half ready code generator that takes a FNN model and writes header files with the topology and coefficients, with no dependencies, just raw pointers

15:09 < zoq> alsc: Okay, that should work just fine, just if you don't like to introduce an extra layer you could switch to cross entropy

15:09 < alsc> kind of a portable tiny decoder maybe I'll upload it somewhere

15:13 < zoq> also, do you like me to open a PR for the layer modifcation or do you refork the repo?

15:14 < zoq> alsc: Would be cool to see the results :)

15:15 < alsc> zoq: yeah for the PRs, I have forked and waiting to hae a second to port the changes from the code you saw to this new fork so I can submit a PR from there.... give it some days...

15:15 < zoq> alsc: Okay sounds good, let me know if you encounter any issues.

15:16 < alsc> sure

15:35 < alsc> zoq: can you give me a little help to replicate that softmax function?

15:36 < alsc> I imagine having the activations, say 10, of the last layer, with the number of classes double activations[] = {0.123,0.213,0.151,0.153,0.21}

15:36 < alsc> I dn't understand what tanh has to do with the softmax, I'd expect just exp() in there

15:52 < alsc> anyay, from the algorithm, I'd say I'll have max_activation = 0.213; then normalize activations subtracting the max, then it looks like it's doing tanh of that, then other stuff..

15:52 < alsc> 😬

15:53 < alsc> zoq: so it's doing tanh first, then softmax

17:05 < manish7294> @rcurtin: Currently we are not building MVU and there is also a comment that it's not working. Could you give me some points as why it's not working. I couldn't get it seeing #189. And sorry for disturbing frequently, It must be hard managing so many things.

17:25 < rcurtin> manish7294: the issue is now #183 on github (#189 was the old number on trac)

17:25 < rcurtin> see also the MVU project for GSoC

17:42 < manish7294> rcurtin: Thanks for suggesting that. Do you think diving into MVU for some time will help getting more acquainted to LMNN ?

17:43 < rcurtin> at least being familiar with it might help, but in my view fixing MVU is a far harder task than LMNN+LRSDP

17:46 < manish7294> rcurtin: ya, I understand you have already mentioned "It's not for faint of heart" :-) . I just want to get familiar with SDP's.

17:50 < rcurtin> no worries, maybe the Vandenberghe and Boyd paper/tutorial is a good place to start?

17:50 < rcurtin> not sure I got the name right... at lunch now

21:15 < zoq> alsc: Sorry for the slow response, the comment is wrong, it's the exp approximation.

21:21 < zoq> alsc: Fixed with 4f664d208 thanks for the hint.

23:15 < rcurtin> zoq: I guess a quick fix for the static analysis is just to remove the -Werror from the cflags for pvs-studio, what do you think?

23:15 < rcurtin> if you think that's fine I'll go ahead and make the change on masterblaster

23:15 < zoq> or checkout the previous version?

23:15 < zoq> since it's a git repo

23:17 < rcurtin> ah, yeah, honestly that is the easier thing

23:18 < zoq> Are you going to make the adjustment?

23:18 < rcurtin> sure, I will do it

23:19 < zoq> okay

23:20 < rcurtin> I need to do a reboot test of masterblaster also; I think I will do it either later tonight or tomorrow morning unless there are any objections

23:20 < rcurtin> I'll basically just wait until no active jobs are running, then apt-get update && apt-get upgrade && apt-get dist-upgrade && reboot

23:21 < zoq> ah, yeah, no objections from my side