naywhayare 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/
Richard has joined #mlpack
Richard is now known as Guest59210
< Guest59210>
What do I need to do to Identify?
Guest59210 has quit [Client Quit]
Anand has joined #mlpack
sumedhghaisas has joined #mlpack
Anand has quit [Ping timeout: 246 seconds]
udit_s has joined #mlpack
sumedh_ has joined #mlpack
sumedhghaisas has quit [Ping timeout: 272 seconds]
govg has joined #mlpack
govg has quit [Changing host]
govg has joined #mlpack
govg has quit [Quit: leaving]
udit_s has quit [Read error: Connection reset by peer]
udit_s has joined #mlpack
udit_s has quit [Ping timeout: 272 seconds]
udit_s has joined #mlpack
udit_s has quit [Ping timeout: 244 seconds]
udit_s has joined #mlpack
udit_s has quit [Ping timeout: 245 seconds]
udit_s has joined #mlpack
udit_s has quit [Quit: Leaving]
andrewmw94 has joined #mlpack
udit_s has joined #mlpack
Anand has joined #mlpack
Anand has quit [Ping timeout: 246 seconds]
Anand has joined #mlpack
< Anand>
Marcus : I meant in the benchmarking system. Did you not do the timing analysis for logistic regression (no python file)?
< udit_s>
marcus_zoq: Hello ! I've uploaded an updated code for the perceptron. While testing it, I'm having some trouble with the Gradient Descent method. Could you have a look ?
Anand has quit [Ping timeout: 246 seconds]
Anand has joined #mlpack
Anand has quit [Ping timeout: 246 seconds]
< udit_s>
naywhayare: hey, are you free now ?
Anand has joined #mlpack
govg has joined #mlpack
govg has quit [Changing host]
govg has joined #mlpack
Anand has quit [Ping timeout: 246 seconds]
< marcus_zoq>
udit_s: Hello, I'll take a look in a few minutes, can you give me some details about your problems?
< udit_s>
I'm unable to get the gradient descent to converge or work properly.
< udit_s>
marcus_zoq: also, I'm having the compiler throw warnings of the unused variables while templatizing for gradient descent.
< udit_s>
marcus_zoq: but still, this makes the code messy, and unused variable warnings are being thrown. I was just going to mail you. I was thinking of leaving this method out.
< udit_s>
marcus_zoq: and probably implementing something like majority voting perceptron update
< udit_s>
or leaving it at only simple update.
< marcus_zoq>
udit_s: Okay, I think at this it's just fine to use the classic percetpron update rule. At the end we can think about implementing another update rule. However I'm going to look into the code, maybe it's just a small mistake.
< udit_s>
marcus_zoq: okay. Otherwise, I'm ready and complete with the Perceptron
< marcus_zoq>
udit_s: Okay, the tests are looking good :)
sumedh__ has joined #mlpack
< marcus_zoq>
udit_s: I think right now we can use three classifiers for the adaboost method. The decission tree, the perceptron and the naive bayes classifier.
sumedh_ has quit [Ping timeout: 255 seconds]
< marcus_zoq>
udit_s: You missed the squashing function (e.g. sigmoid function) to ensure that the values are within [0 1].
< udit_s>
marcus_zoq: Yeah, I came across that in : http://homepages.gold.ac.uk/nikolaev/311sperc.htm I realized I had missed that. I was ready to get into that - seems quite easy - but the compiler throwing warnings of unused variables made the design of the learn policy template parameter very messy, Any suggestions ?
< marcus_zoq>
udit_s: Hm, I need to download the code to look into the compiler warnings.
< udit_s>
marcus_zoq: no need. you see the updateweight method in the learn policy templates ?
< marcus_zoq>
udit_s: yes
< udit_s>
marcus_zoq: so the GradientDescent template does not use the row index while the simpleweightupdate does not use classlabels. also, maxval is another variable which is set but not used.
< marcus_zoq>
udit_s: I think there are two options, we can set a default value for the row index. Or we can move the whole update process into the learning rule and just passing the weights.
< marcus_zoq>
udit_s: But a default value as you did it with the mattype should fix the warnings.
< udit_s>
what about classLabels ?
< marcus_zoq>
udit_s: Is there a reason to pass the hole labels vector? Can you not jsut pass the actual class label for this instance?
< udit_s>
marcus_zoq: I could; and then by providing a default value can be a work around. Hmm. let me see if that works.
< jenkins-mlpack>
Starting build #1950 for job mlpack - svn checkin test (previous build: SUCCESS)
< udit_s>
marcus_zoq: nope, setting a default parameter does avoid a warning. I removed classlabels because vectorindex does exactly the same thing.
< marcus_zoq>
udit_s: can you push you changes?
< marcus_zoq>
udit_s: your changes
< udit_s>
marcus_zoq: done.
< udit_s>
marcus_zoq: and in both, classify and the constructor, the variable max val is thrown as a warning of unused variable.
< marcus_zoq>
udit_s: The wierd thing is, that my compiler (clang) doesn't complain about unused variables. Can you test if removing the name of the variable works? So just write size_t instead of size_t name.
< marcus_zoq>
udit_s: weird thing
< naywhayare>
marcus_zoq: looks like udit left the channel for now
< naywhayare>
I finally have stable internet again so I will have a chance to look through his code and answer some of his questions too
< marcus_zoq>
naywhayare: Okay, good. I wonder why clang doesn't complain about unused variables.
< naywhayare>
maybe CMake isn't using -Wall when clang is the compiler?
< naywhayare>
or -Wall for clang doesn't imply -Wunused (or something like that)?
< marcus_zoq>
naywhayare: Thanks, 'if(CMAKE_COMPILER_IS_GNUCC) set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall -Wextra")' I think I've missed the change.
< naywhayare>
ah, maybe we should set that for clang too
arcane has joined #mlpack
< arcane>
naywhayare, I was testing binary space partition tree with ball bound. There is a problem in tree construction. As each node splits itself, at one point it so happens that all the points in the current node are assigned to the right child, leaving the left child empty. Due to this the indices go out of bound.
< naywhayare>
arcane: I am sorry that I haven't had a chance to incorporate any of your work yet. I was away from the internet for about a week and have fallen very far behind...
< naywhayare>
why are all of the points being assigned to the right child?
andrewmw94 has quit [Read error: Connection reset by peer]
< naywhayare>
are you using the MeanSplit policy for splitting, or something else?
andrewmw94 has joined #mlpack
< arcane>
its ok. We can do the incorporation when you find time
< arcane>
I am using mean split. Though there is an issue with that too
< arcane>
the problem seems to be in the way ball bound expands when using the '|' operator
< arcane>
mean split happens to take such a value of splitval that all the points go to the right
< arcane>
in tree_test.cpp .... if you replace hrectbound by ball bound while constructing the BST ... you can reproduce the error
< arcane>
oh ... and you would need the updated ballbound.hpp and ballbound_impl.hpp. I had made some changes to make it work with BST. The new files are attached in the ticket
< naywhayare>
okay... let me finish a few other things and then I will take a look into this to figure out what is going on
< arcane>
alright
< naywhayare>
it sounds like it could be a bug in BallBound::operator|=()
< arcane>
There is a small bug (I think) .... the for loop in that function should start with 0 instead of 1. Thought this does not solve our problem, but it does incorporate one of the point that it initially was missing
< naywhayare>
I'm not surprised that it has bugs -- that code is not well-tested (which is, of course, a big part of why the ticket is open :))
< arcane>
right ... probably they will get removed as we attempt to make ball bound work with BST :)
< naywhayare>
marcus_zoq: fix for clang committed in r16695. it should use -Wall now when clang is used
< marcus_zoq>
naywhayare: Great, Thanks!
< jenkins-mlpack>
Starting build #1951 for job mlpack - svn checkin test (previous build: SUCCESS)
andrewmw94 has quit [Ping timeout: 245 seconds]
< marcus_zoq>
Okay, now I wonder why this is working: 'std::vector<double> scores = new std::vector<double>(referenceNode.NumChildren());'
< naywhayare>
it fails with clang when I build allknn
< naywhayare>
but if I just build mlpack, that particular bit of code isn't used in any instantiated template, so it doesn't actually get compiled
< naywhayare>
I imagine it would fail with gcc too... let me check
< naywhayare>
huh! it doesn't fail
< marcus_zoq>
gcc magic
< naywhayare>
for some definition of magic :)
< naywhayare>
ok, I see what is happening
< naywhayare>
the RTreeTraverser class is never actually being instantiated in allknn_main.cpp
< naywhayare>
but I guess clang is doing semantic analysis on it regardless, even though that's not technically required by the standard, while gcc does not do semantic analysis