verne.freenode.net changed the topic of #mlpack to: http://www.mlpack.org/ -- We don't respond instantly... but we will respond. Give it a few minutes. Or hours. -- Channel logs: http://www.mlpack.org/irc/
manish7294 has joined #mlpack
< manish7294>
rcurtin: Finally I am able to reproduce those eval results. Earlier while doing those simulations I had commented out "if ( eval < -1 ) { bp = l; break; }" part, so that we do not loose any eval which is less than -1. But this time I wasn't doing that and that's the reason results were not same.
< rcurtin>
right, have you found any difference between the theory and what you implemented?
< rcurtin>
what I am understanding from your comment on github is that the pruning should be working much better
< rcurtin>
but it seems like the problem is in the implementation of the bounds, did I understand that right?
< manish7294>
rcurtin: Probably that could be the reason but I still have to find what's going wrong in there.
cjlcarvalho has joined #mlpack
< manish7294>
rcurtin: The error is while updating transformationOld matrix, It should be done in the last part once the for loops ends.
< manish7294>
But still runtime is not good.
< rcurtin>
are you caching the previous calculation correctly for each bound?
< jenkins-mlpack>
* manish887kr: Use PRINT_CALL() macro
< jenkins-mlpack>
* haritha1313: submat index error debug
< jenkins-mlpack>
* haritha1313: end parameter condition
< jenkins-mlpack>
* manish887kr: allow user to pass a initial learning point
< jenkins-mlpack>
* manish887kr: added some simple dimensionality tests for initial point
< jenkins-mlpack>
* manish887kr: Reviewed changes
< jenkins-mlpack>
* manish887kr: style fix
< jenkins-mlpack>
* manish887kr: comment out low rank accuracy test
< jenkins-mlpack>
* Ryan Curtin: Specify namespace to handle confused MSVC compiler.
< jenkins-mlpack>
* Ryan Curtin: Update history.
< jenkins-mlpack>
* akhandait45: make changes suggested in sampling PR
manish7294 has joined #mlpack
< manish7294>
rcurtin: Found a good speedup in the implementation of bounds and now things are way faster than current LMNN.
manish7294 has quit [Ping timeout: 252 seconds]
< rcurtin>
manish7294: (I am up late today!) if you want to go ahead and push it, I'll take a look tomorrow. glad to hear that you were able to get some acceleration :)
< Atharva>
Is there some utility function in mlpack or arma that will allow me to save matrices as images?
< zoq>
Atharva: For the mnist pgm or ppm should work: A.save("A.pgm", pgm_binary);
caiojcarvalho has quit [Ping timeout: 268 seconds]
cjlcarvalho has quit [Ping timeout: 245 seconds]
ImQ009 has joined #mlpack
caiocarvalho has quit [Ping timeout: 268 seconds]
< sumedhghaisas>
zoq, ryan: Hey guys, do you use any library to plot training curves?
< sumedhghaisas>
I mean in MLPack :)
< rcurtin>
sumedhghaisas: if I need to plot something I always end up just saving as CSV then importing into Python to use matplotlib...
< rcurtin>
I never bothered to learn gnuplot
< zoq>
same here :)
< zoq>
You could also pipe/process the output and plot the results afterwards as well.
< sumedhghaisas>
rcurtin: ahh Thats what I had in mind. :)
< sumedhghaisas>
Its not ideal though, for example VAE samples need to be stored before loading them in python... too much work :P
< rcurtin>
agreed---you could write a binding for VAE to Python (basically just write a _main.cpp file, then CMake it with 'add_python_binding(<name>)')
< rcurtin>
then you could do all the work in Python
< rcurtin>
but the needs of what you're doing might be too complex for that
< sumedhghaisas>
rcurtin: ahh no, Atharva and I were just checking the VAE training and seems to be working so YEAAAHHHH!
< sumedhghaisas>
we were thinking how to show the training curves
< sumedhghaisas>
and samples
< rcurtin>
great to hear it is working! :)
< Atharva>
Yess!
< rcurtin>
I guess the best idea I have is just write the points to be plotted to a CSV then use python... but you are right, it is not ideal
jenkins-mlpack2 has joined #mlpack
< Atharva>
zoq: The current evaluate functions don't let us calculate loss of the network given custom predictors and targets
< Atharva>
What do you think about an overload of Evaluate which takes in predictors and targets and returns the loss with the given parameters
< Atharva>
I think it's useful in cases where you are using mini batches to train and want to see the loss over some other input.
< Atharva>
The Evaluate functions we have now take the predictors matrix from the last forward pass.
vivekp has quit [Read error: Connection reset by peer]
vivekp has joined #mlpack
< zoq>
Atharva: What about using Predict(arma::mat predictors, arma::mat& results)?
< zoq>
Atharva: We would have to return the loss as well.
< Atharva>
Yeah, it doesn't return the loss
< zoq>
I guess, we could provide another Evaluate method or return the loss, your choice.
< Atharva>
To return the loss, the predict function will also need to take the targets as a parameter
< zoq>
Atharva: Ahh, you are right.
< Atharva>
So, is a Evaluate(predictors, targets) a good idea?
< Atharva>
an*
< zoq>
Atharva: Yeah sounds good to me.
< Atharva>
I will add it then
< zoq>
okay, great
< zoq>
The PR is going to be full of really neat features :)
< Atharva>
Yeah, seems like it. I am really glad to be of use. :)
< Atharva>
zoq: Whenever you are free, please give a look at #1366.
vivekp has quit [Read error: Connection reset by peer]
vivekp has joined #mlpack
< Atharva>
can someone give me an example of how to use the serialize function to save models?
vivekp has quit [Read error: Connection reset by peer]
vivekp has quit [Read error: Connection reset by peer]
vivekp has joined #mlpack
vivekp has quit [Read error: Connection reset by peer]
vivekp has joined #mlpack
< rcurtin>
ok, just an update on the build systems:
< rcurtin>
I've talked with my manager at Symantec and asked if he's willing to "donate" them for mlpack use, since they're already set up and honestly everything other than masterblaster is too obsolete to be useful to the company
< rcurtin>
he said he would think about it, and probably have an answer in two weeks; but until that time, I think the systems will remain up and untouched and we can continue to use them
< ShikharJ>
rcurtin: Yay!
< rcurtin>
nonetheless, I'm working now on the transition to using ci.mlpack.org (which is actually just mlpack.org) as the Jenkins master node
< rcurtin>
and I should have that done (or done enough) in the next couple of hours
< rcurtin>
Jenkins doesn't make export/import easy, so I'm just manually copying the jobs over... there may be some minor issues to work out
< rcurtin>
in the worst case, where Symantec decides not to donate the systems, I have two of the old benchmarking systems we used when I was at GT that I'll set up at my house
< rcurtin>
and then I can add a few other systems I have floating around. so we wouldn't have the kind of "firepower" we have now, but it would certainly be enough to get by
< rcurtin>
it's also not clear what my new company might be interested in providing for infrastructure, but I'll have to scope that out once I'm on the inside---that will probably be a month or two
< rcurtin>
(I don't start until early August... I am going to try and relax in July :))
< rcurtin>
I don't have any prior probability distribution on whether Symantec will let us keep using them, so I guess we will just have to wait and see
< zoq>
Good news for now, also I can easily provide access to two more machines, just let me know what I should install.
< rcurtin>
great, let me get Jenkins set up and then if you want to add them, we can do it then, or if you want to wait to see what Symantec will do, that's fine too
jenkins-mlpack2 has quit []
jenkins-mlpack2 has joined #mlpack
< ShikharJ>
zoq: I just tmux'd a build on the latest code to check if changing the move semantics doesn't break our design.
< zoq>
ShikharJ: For the standard GAN?
jenkins-mlpack2 has quit []
jenkins-mlpack2 has joined #mlpack
< rcurtin>
ok, now ci.mlpack.org/ is working
< rcurtin>
I think the PR builders should be working but it's not fully tested yet
< rcurtin>
I'll leave Jenkins running on masterblaster
< zoq>
rcurtin: So the PR builds are running on mlpack.org?
< rcurtin>
or any free node
< rcurtin>
but managed by Jenkins on ci.mlpack.org
< rcurtin>
or, at least, the next PR build should be
< rcurtin>
I think I'd like to set up those PR builds to run in an unprivileged docker container
< rcurtin>
but that should not be too hard
< zoq>
yeah, good idea
< zoq>
so, jenkins runs on mlpack.org and masterblaster is another build node?
< rcurtin>
right, that is how I have it for now
< rcurtin>
Jenkins is still running on masterblaster for now
< rcurtin>
but I transitioned all the plugins and all non-obsolete jobs
< rcurtin>
looks like at least the git commit webhook works
< rcurtin>
do you want to pish a simple change to a PR to test that? I don't think I have any open at the moment actually
< zoq>
sure, hold on
ImQ009 has quit [Quit: Leaving]
< rcurtin>
ha, you always have a new optimizer ready to go :)
< rcurtin>
hmm, no success. let me look into it a bit...
< zoq>
okay, let me know if I should trigger another build
< zoq>
We can also retrigger the webhoock in the github settings
< rcurtin>
ah, I see the jenkins credentials are not working right
< rcurtin>
I am on the train now, let me look into it once I get home
< jenkins-mlpack2>
Ryan Curtin: Update to ci.mlpack.org. Also hopefully trigger a build.
< rcurtin>
so, I think that I have copied over the credentials for the mlpack-jenkins github user wrong... but I have forgotten the credentials too!
< rcurtin>
so I am working with github to try and get access to that account again (since I'm actually not even sure what email was used to create it!)
< zoq>
hm, let me check
< zoq>
I created the account :)
< rcurtin>
oh! hm
< rcurtin>
well so clearly I have no idea what is going on :)
jenkins-mlpack2 has quit [Ping timeout: 256 seconds]