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/
sumedhghaisas has quit [Ping timeout: 240 seconds]
sumedhghaisas has joined #mlpack
travis-ci has joined #mlpack
< travis-ci>
mlpack/mlpack#4523 (master - d0a5e6c : Ryan Curtin): The build passed.
< manish7294>
rcurtin: Sorry for disturbing! But Did you get a chance to view the proposal?
dmatt has quit [Quit: Page closed]
< rcurtin>
manish7294: I was not able to, but I will look tomorrow morning
< manish7294>
rcurtin: Thanks! Please try to make it before the deadline :)
govg has quit [Ping timeout: 260 seconds]
sumedhghaisas has quit [Read error: Connection reset by peer]
sumedhghaisas has joined #mlpack
Prabhat-IIT has quit [Ping timeout: 260 seconds]
csoni has quit [Ping timeout: 276 seconds]
manish7294 has quit [Remote host closed the connection]
sumedhghaisas has quit [Ping timeout: 268 seconds]
csoni has joined #mlpack
hassanmahmood has joined #mlpack
manthan has joined #mlpack
csoni has quit [Read error: Connection reset by peer]
govg has joined #mlpack
sumedhghaisas has joined #mlpack
csoni has joined #mlpack
sujith has joined #mlpack
hassanmahmood has quit [Ping timeout: 276 seconds]
Jackey-Huo has joined #mlpack
hassanmahmood has joined #mlpack
csoni has quit [Read error: Connection reset by peer]
addy__ has joined #mlpack
addy__ has quit [Client Quit]
hassanmahmood has quit [Ping timeout: 240 seconds]
hassanmahmood has joined #mlpack
hassanmahmood has quit [Ping timeout: 240 seconds]
< Jackey-Huo>
Hey, I plan to joint the mlpack's GSoC project in reinforcement learning. after read the source code, I find that some typical tabular MDP method donot implemented yet(e.g. value iteration...)
< Jackey-Huo>
but the project introduction said mlpack prefer implementations of recent ideas. So does mlpack have the plan to implement the fully set of RL method? or only recent, Neuron-Network related method is wanted?
sujith has quit [Ping timeout: 260 seconds]
dmatt has joined #mlpack
Jackey-Huo has quit [Quit: Page closed]
hassanmahmood has joined #mlpack
ckeshavabs has joined #mlpack
< ckeshavabs>
rcurtin: Hi, do we have the flexibility to use extra gpu-cores for mlpack algorithms? Like installing tensorflow-gpu provides great scale up in speeds if we have a gpu right?
Jackey-Huo has joined #mlpack
sumedhghaisas2 has joined #mlpack
sumedhghaisas has quit [Ping timeout: 256 seconds]
< ckeshavabs>
zoq: To perform image processing with OpenCV, we will need to install the OpenCV as a dependency on the local machine right? and then include only the relevant headers in preprocessing files? Any other way to achieve the same?
ckeshavabs has quit [Ping timeout: 260 seconds]
sumedhghaisas2 has quit [Read error: Connection reset by peer]
sumedhghaisas has joined #mlpack
sumedhghaisas2 has joined #mlpack
sumedhghaisas has quit [Read error: Connection reset by peer]
Atharva has joined #mlpack
< Atharva>
sumedhghaisas2: Hi Sumedh. It will be very helpful if you could answer the doubts I posted here yesterday, the deadline is in 6 hours and I want to make those changes.
< sumedhghaisas2>
@Atharva: Hi Atharva
< sumedhghaisas2>
We should use the distributions in MLPack to sample from and model the output.
< sumedhghaisas2>
We could implement the re-paramatrization trick inside Gaussian Distribution class, which provides us a sample from the distribution with mean and variance
< sumedhghaisas2>
we could implement it's backprob inside the class as well. This way any other Distribution could be easily converted to be used instead of Gaussian
< sumedhghaisas2>
I mean in case user does not want Gaussian but some other distribution as a prior over the latents.
< Atharva>
Okay yeah, I will mention this.
< sumedhghaisas2>
Another important use of distributions will be in modelling output.
< sumedhghaisas2>
if p(x | z) is the distribution over the output logits, then loss simply becomes
< sumedhghaisas2>
p.log_prob(reconstruction) + KL
< sumedhghaisas2>
this way user can use various distributions to model the output
< Atharva>
Also, when we pass encoder/decoder as LayerTypes, I still add layers with Add<>, right?
< Atharva>
I wasn’t able to reproduce the changes from your PR on my system
< sumedhghaisas2>
Yes. That's right.
< Atharva>
Okay thanks, I will try to make these changes in the PR, I have changed the timeline according to what you suggested.
< sumedhghaisas2>
as long as FFN is available to Boost variant type LayerTypes<>, the conversion will be implicit
hassanmahmood has quit [Ping timeout: 240 seconds]
< sumedhghaisas2>
great :)
hassanmahmood has joined #mlpack
sumedhghaisas2 has quit [Ping timeout: 256 seconds]
sumedhghaisas has joined #mlpack
sumedhghaisas has quit [Read error: Connection reset by peer]
sumedhghaisas has joined #mlpack
K4k has quit [Read error: Connection reset by peer]
hassanmahmood has quit [Ping timeout: 240 seconds]
Jackey-Huo has quit [Ping timeout: 260 seconds]
< zoq>
Jackey-Huo: Hello thanks for getting in touch, not all methods, but recent and interesting.
< zoq>
ckeshavabs: You could build against NVBLAS, to get GPU acceleration.
< zoq>
ckeshavabs: About OpenCV, I guess if someone needs OpenCV he could just link against the lib and use the functionality.
hassanmahmood has joined #mlpack
sujith has joined #mlpack
hassanmahmood has quit [Ping timeout: 265 seconds]
csoni has joined #mlpack
csoni has quit [Ping timeout: 264 seconds]
levi has joined #mlpack
levi has left #mlpack []
< manthan>
rcurtin : if you find time can you please have a look at my proposal (submitted as draft)
aditya has joined #mlpack
< rcurtin>
manthan: a couple quick comments---
< rcurtin>
I think that DatasetMapper<> already does the simplest tokenization operation (convert a string to a unique index), so you might be able to use that internally
< rcurtin>
also, I think it would help for clarity if you made a quick list of the exact algorithms you were planning to implement; extracting them from the list of tasks is a little time consuming and I am not sure if I missed something
Atharva has quit [Quit: Connection closed for inactivity]
< manthan>
ya , i missed DatasetMapper<> , this can get the unique tokens and i will build the algo on top of this
< manthan>
rcurtin : Each task is an algorithm as such(except testing task). But if you say so, i can make a compiled list after the deliverable section
< manthan>
is the deliverable and task list good enough given the timeframe? thats what I am focusing more on. Thanks for the comments.
< rcurtin>
I think that is fine, yes
< manthan>
rcurtin : Thanks a lot for the review, I will include your changes and submit the final proposal.
K4k has joined #mlpack
sumedhghaisas has quit [Read error: Connection reset by peer]
sumedhghaisas has joined #mlpack
ImQ009 has joined #mlpack
rajiv_ has joined #mlpack
sumedhghaisas has quit [Read error: Connection reset by peer]
sumedhghaisas has joined #mlpack
dmatt has quit [Quit: Page closed]
< rajiv_>
@rcurtin Thanks for the feedback! I have one small query, does auto_detect mode in load() work for formats like .png directly?
rajiv_ has quit [Quit: Page closed]
sumedhghaisas has quit [Read error: Connection reset by peer]
< dmatt>
Hello everyone! I am Thanasis Mattas and I study Physics in Aristotle University of Thessaloniki, Greece. I found out a bit late about GSoC, but I did upload a proposal about "Algorithm Optimization". Any review will be more than whelcome! Thank you!
dmatt has quit [Quit: Page closed]
dmatt has joined #mlpack
hassanmahmood has joined #mlpack
dmatt has quit [Ping timeout: 260 seconds]
sujith has quit [Quit: Page closed]
nikhilweee has quit [Remote host closed the connection]
sujith has joined #mlpack
sumedhghaisas has quit [Read error: Connection reset by peer]
sumedhghaisas has joined #mlpack
sourabhvarshney1 has joined #mlpack
< sourabhvarshney1>
@dmatt I think you are pretty much too late. But don't worry. GSoC is not the only think in which you can contribute to Open Source. You can contribute to mlpack freely too. Please feel free to ask your doubts here if there are any.
< sourabhvarshney1>
@rcurtin Does this build test whether mlpack will work with different versions of it's dependencies?
< rcurtin>
yes, exactly
< sourabhvarshney1>
@rcurtin I checked previous build #262. It was failing on NesterovMomentumSGDSpeedUp Test function. Was that due to less number of iterations?
< sourabhvarshney1>
I think I had committed for this test. Did you change it because now the build is passing?
< sourabhvarshney1>
Just curious about that
< rcurtin>
I loosened the tolerances somewhat, yes
< rcurtin>
it would occasionally fail with different random seeds
< sourabhvarshney1>
@rcurtin Does this imply that the test may not converge every time with different random seeds?
< rcurtin>
yeah, that is exactly what is happening; the nesterov momentum does not always get to a good solution with different random seeds
dmatt has quit [Quit: Page closed]
< sourabhvarshney1>
@rcurtin I think this is the drawback of the optimiser. What say? Thanks for providing me with such valuable information.
sumedhghaisas2 has joined #mlpack
sumedhghaisas has quit [Ping timeout: 276 seconds]
< rcurtin>
sourabhvarshney1: agreed, that is one of the issues with randomized methods. getting the tolerance right works for the tests; I don't think anything is wrong with the implementation
< sourabhvarshney1>
Thanks once again. Very grateful to have such information from you.
< manthan>
rcurtin : what would be better increasing tolerances or setting a fixed seed though?
sumedhghaisas has joined #mlpack
sumedhghaisas3 has joined #mlpack
< sourabhvarshney1>
@manthan I think both are necessary because setting a fixed seed might cause trouble if we need to check over various points, then increase tolerance would work. Not sure, so correct me if I am wrong.
sumedhghaisas2 has quit [Ping timeout: 246 seconds]
sumedhghaisas has quit [Ping timeout: 256 seconds]
csoni has joined #mlpack
< manthan>
ya i understand that both are necessary, i mean in a trade off what would be preferable to do
< rcurtin>
manthan: I prefer increasing tolerances
< rcurtin>
if we use a fixed seed, then it is possible that the algorithm is implemented completely incorrectly but just so happens to converge for one specific test seed
< rcurtin>
(this happens, by the way!)
< rcurtin>
if we don't use a fixed seed, it's possible users may be randomly hit by failing tests, but since we have a matrix build and do a lot of testing, we can have a lot more confidence that the implementation isn't broken
sourabhvarshney1 has quit [Ping timeout: 260 seconds]
sourabhvarshney1 has joined #mlpack
manthan has quit [Ping timeout: 260 seconds]
< sourabhvarshney1>
@rcurtin Upto which limit we can increase the tolerances? Within 1% or 2% at max. or it can be more?
< rcurtin>
sourabhvarshney1: there is no exact science for that
< sourabhvarshney1>
That means that depends on the test?
< rcurtin>
right
< sourabhvarshney1>
Ok got it. Thanks again
sourabhvarshney1 has quit [Ping timeout: 260 seconds]
sumedhghaisas3 has quit [Read error: Connection reset by peer]
sumedhghaisas has joined #mlpack
sourabhvarshney1 has joined #mlpack
sourabhvarshney1 has quit [Ping timeout: 260 seconds]
csoni has quit [Read error: Connection reset by peer]
ImQ009 has quit [Quit: Leaving]
aditya has quit [Ping timeout: 260 seconds]
hassanmahmood has quit [Ping timeout: 264 seconds]
sumedhghaisas has quit [Ping timeout: 240 seconds]