10:55 < partobs-mdp> zoq: Did you figure out the issue with parameters variable?

10:57 < zoq> partobs-mdp: Not sure, what you mean, can you help me?

10:58 < partobs-mdp> <partobs-mdp> zoq: rcurtin: Rolled back to the pre-LayerTypes version. Now trying to add Parameters() function (for gradients) What do you think about API? (in the latest commit)

10:58 < partobs-mdp> <partobs-mdp> Also, could you take a look at the strange issue with parameters being treated as a size_t? (Will be in a compiler error message)

11:00 < partobs-mdp> By the way, I feel fair to remind that I'm leaving to Moscow in a week's time, so I'm available for the project before Friday, 25 Aug (and maybe on 31 Aug, when I arrive back)

11:00 < zoq> Ah sorry, haven't seen the message, got disconnected from IRC ...

11:02 < zoq> When do you leave, before August 25 or after?

11:02 < zoq> Thanks for the reminder.

11:03 < partobs-mdp> On 25 Aug (7pm Moscow Time, 4pm GMT, if I remember the ticket correctly)

11:04 < zoq> Okay, we should make sure you submit the necessary code to the google platform before you leave.

11:07 < zoq> Do you see any other message that we missed? My last message is: "19:36 < kartik_> ohkay.. thanks :)"

11:07 < partobs-mdp> Well, I don't remember myself sending any more messages :) 19:36 - what time zone?

11:07 < zoq> UTC

11:09 < partobs-mdp> Nope, no missed messages from my side :)

11:09 < zoq> Okay, great thanks!

11:10 < partobs-mdp> There is one message, but it was some global anti-spam message :)

11:10 < partobs-mdp> -kline- [global notice] Hi there: freenode is experiencing a large quantity of spam. If you receive links from users you don't recognise, please do not open them. If you are affected, please set usermode +R to block messages from unidentified users.

11:10 < partobs-mdp> -kline- [global notice] Additionally, while working to minimise this spam, staff error led to the majority of the network receiving a temporary K-Line (network ban). The ban has since been lifted and those affected should be able to reconnect. Please accept our apologies, and contact us if you, or someone you know, still cannot connect.

11:10 < partobs-mdp> But I honestly don't find this relevant :)

11:12 < zoq> Yeah, not relevant for the mlpack channel :)

11:18 < zoq> about the parameters issue; I guess there is a missing ';' after writeCount = writeParams.n_elem,

11:18 < zoq> I mean replace the ',' with ';'

11:22 < partobs-mdp> zoq: Now waiting while it compiles; meanwhile, could you say do you think that it's OK to do the Parameters method as it is done right now?

11:24 < partobs-mdp> (It compiled - now waiting for linking and running corresponding unit test)

11:24 < partobs-mdp> Also, I took a closer look on the technology of the GSoC blog, and I should admit I like the static blog idea.

11:24 < zoq> I'll take a look at the changes right now.

11:25 < partobs-mdp> Could you tell me what was the rationale behind choosing Pelican engine? (I just don't know much yet about static blog engines)

11:26 < zoq> I used it before :) and back in the days Jekyll wasn't that famous.

11:28 < partobs-mdp> By the way, it successfully compiled, but there is a logic error (copying 27x1 matrix to 36x1 slot) - should be easily fixable :)

11:28 < zoq> I think something like this would also be nice to have: http://donw.io/post/github-comments/

11:28 < zoq> :)

11:28 < zoq> Btw. do you think it's a good idea to visit Moscow in December/January?

11:32 < partobs-mdp> Well, I was one there in February - it was moderately cold (-8 ~ -12 C) and there was some snow

11:32 < partobs-mdp> Extrapolating, I guess that in December it should be similarly cold with with more snow - which should also be more stereotypical of the Russian weather, by the way :)

11:33 < partobs-mdp> So yes, it should be a nice time if you go for a moderately cold wather

11:33 < partobs-mdp> But, of course, YMMV

11:36 < partobs-mdp> zoq: As a follow-up on the comment thing, do you know about any nice *static* comment engines? (As a bonus it would be nice to have a chance to use MathJax in the comments as well as in the blog itself)

11:37 < partobs-mdp> Follow-up 2 (on the compiler thing) - I have found the silly copy-paste error, now waiting for the results

11:39 < zoq> Not sure there is a nice solution for the comments, at the end you end up with some php code ...

11:39 < zoq> -8 is kinda cold, not sure if I could enjoy the day after spending 2 hours outside :)

11:39 < partobs-mdp> Follow-up 3 (on Moscow thing) - December get bonus points for a really nice pre-New Year environment (which is traditionally the thing taken seriously in our culture)

11:39 < zoq> But, Moscow or Saint Petersburg is so different from everything I've seen before

11:39 < zoq> hm, the problem with the RebuildParameters method is, that if you change parameters you don't change the parameters in the corresponding model e.g. embed.

11:40 < zoq> But. Kris encountered a similar problem, let me dig up the code

11:41 < partobs-mdp> zoq: It compiled and ran nicely, so the only problem to solve is the Parameters() method

11:43 < zoq> https://github.com/mlpack/mlpack/pull/1066/files#diff-73a3e2abae16e9a84a96856b191900e4R85 I think we can do something similair here.

11:44 < zoq> The only problem I see is that, we reinitialize the networks.

11:45 < partobs-mdp> zoq: By the way, did you manage to run MathJax with custom fonts? (e.g., Neo Euler --- the one used in "Concrete Mathematics")

11:46 < partobs-mdp> As much as I like MathJax project I don't really like standard Computer Modern fonts (Neo Euler is more appealing for me, for example)

11:47 < zoq> on the mlpack blog? haven't tried it

11:47 < zoq> I agree looks nice, I can take a look into it

11:50 < zoq> I think, for our case a reinitialization isn't a problem, we can adjust the network init class, later.

11:57 < partobs-mdp> zoq: I've also done a unit test for Parameters() method - not afraid to break anything now :)

12:00 < partobs-mdp> (I pushed it into my GitHub fork now)

12:01 < zoq> we have to adjust the test case, remove the Reset Parameters(); calls and do the parameter initialization after the hamUnit instantiation. But that should be straightforward, let me know if you need help with the initialization code.

12:02 < partobs-mdp> zoq: Sorry, didn't quite understand you - could you elaborate on that?

12:07 < zoq> The problem I see with the current buildParameters method is that we only copy the parameters, but if we change something in parameters we don't change the model parameter. So I propose to use the NetworkInit class, as used in the GAN PR, which reinitializes the parameters; means everything you did before the ham instantiation is removed. So the test would fail, but if we set the parameter after the

12:07 < zoq> instantiation we are good to go.

13:48 < rcurtin> knife.lugatgt.org banned from freenode for spamming? not sure what that is about... sending an email now...

13:48 < rcurtin> (that's the system I connect to #mlpack from)

13:48 < rcurtin> unfortunately this means logs will be down for a little while until this gets resolved

13:48 < zoq> same here, restart fixed the problem for me

13:49 < rcurtin> oh? ok, I will not send an email then

13:49 < rcurtin> let's see...

13:50 < rcurtin> well, ok, that works fine, I guess I should have tried that before composing the email to kline@freenode.net :)

22:41 < kris1> https://gist.github.com/kris-singh/9420d3e9a0afe12099826dd98222dce7

22:41 < kris1> Could somebody have a look at this

22:56 < kartik_> <zoq> to get random values from calling mlpack random function, the seed value should be made zero ?

22:57 < kartik_> i am getting the same set of values every time

23:10 < zoq> kartik_: Just commented on the PR.

23:10 < zoq> kris1: What do you mean with "the error is in dropout layer at i = 5"?

23:11 < kris1> network[5] has the error with dimensions

23:11 < kris1> More than that the error is gy % mask * scale line of the dropoout layer.

23:12 < zoq> Sounds like the conv setting aren't correct.

23:13 < kris1> But then why does the forward pass go through…..at all

23:14 < zoq> Did you build with DEBUG=OFF?

23:16 < kris1> nopes -g flag is on

23:17 < zoq> I guess either are the conv setting off my some value or there is a flaw in the conv layer backward step.

23:22 < zoq> Looks like another debug session, unfortunately I can't look into it before Saturday.

23:24 < kris1> That is okay i would in meanwhile try with another set of values….

23:40 < zoq> kris1: If you test other values, start with pad = 0 and stride = 1.