ChanServ changed the topic of #mlpack to: "mlpack: a fast, flexible machine learning library :: We don't always respond instantly, but we will respond; please be patient :: Logs at http://www.mlpack.org/irc/
< HimanshuPathakGi>
Hey @saksham189 are you there?? I want to ask some questions regarding gradients
< saksham189Gitter>
I am here now.
< saksham189Gitter>
We can have a meeting if you want
< saksham189Gitter>
Feel free to drop your questions here or on the PR and I will try to answer them as soon as possible
< HimanshuPathakGi>
Oh good can we meet today at 4pm
< HimanshuPathakGi>
My question is just about how should I calcute the values of gradients in linear layer we are doing this by by vectorising the error * input.t() but in the case of rbf what should I do I can't able to understand?? How should I find these values I am little confused about that just need some advice over it.
< nishantkr18[m]>
zoq: favre49 is it just me or does http://lab.mlpack.org/ take a longer time to run? It actually is causing a problem as i use agent.Step() inside the testing loop.
< nishantkr18[m]>
When i comment out agent.Step(), the env is able to run the simulation and give out the video..
< zoq>
nishantkr18[m]: the lab runs a docker image with limited resources, so it should be slower.
< zoq>
nishantkr18[m]: I could increase the number of CPU's and memory.
< nishantkr18[m]>
zoq: I think we'd have to do that, as any computation related tothe agent slows down the process, so rendering causes issues, thats what my guess is..
< zoq>
nishantkr18[m]: The think the problem is that the notebook drops the connections after some time, e.g. if agent.step takes a long time.
< nishantkr18[m]>
@zo
< nishantkr18[m]>
zoq: Yeah, could be
< saksham189Gitter>
Hey @zoq , there are several different approaches to learning RBF networks based on k-means clustering, gradient descent etc. Is there any specific learning approach that you want to implement?
< saksham189Gitter>
(edited) ... that you want to ... => ... that we should try to ...
< saksham189Gitter>
> My question is just about how should I calcute the values of gradients in linear layer we are doing this by by vectorising the error * input.t() but in the case of rbf what should I do I can't able to understand?? How should I find these values I am little confused about that just need some advice over it.
< saksham189Gitter>
I read some papers on RBFN and seems like there are numerous learning approaches for RBF. So, I will discuss with @zoq and let you know which approach we move ahead with.
< saksham189Gitter>
(edited) ... we move ... => ... we should move ...
< HimanshuPathakGi>
Yeah in some paper's they are using random data as centres and in some kmeans,
< HimanshuPathakGi>
So I was also confused due to this.
< saksham189Gitter>
yup, I saw that too :)
< zoq>
nishantkr18[m]: Done, but that doesn't solve the issue, I'm not sure this is a timeout issue, because I added: #include <boost/thread/thread.hpp> and boost::this_thread::sleep( boost::posix_time::seconds(5) ); instead of step and it still works.
< HimanshuPathakGi>
We have to wait for @zoq for this. Thanks for pointing out.
< zoq>
nishantkr18[m]: Did you test agent.State().Data() = on you local system?
< zoq>
nishantkr18[m]: Maybe that one already fails?
< zoq>
nishantkr18[m]: I mean in combination with agent.Step();
< nishantkr18[m]>
zoq: Yes, the code works with no errors on my machine
< nishantkr18[m]>
zoq: Yeah, I've tried the combination as well..
< zoq>
nishantkr18[m]: OKay, increased the timeout as well, can you test again?
< nishantkr18[m]>
zoq: yeah sure
< zoq>
saksham189Gitter: I think k-means would be interesting.
< zoq>
nishantkr18[m]: Hm, maybe I did something wrong ...
< nishantkr18[m]>
zoq: hmm, still not working âšī¸
< zoq>
nishantkr18[m]: Not sure, I set the timeout to 10 seconds, maybe sometimes we need more.
< nishantkr18[m]>
@zo
< nishantkr18[m]>
zoq: We'll need to try and find out
< zoq>
nishantkr18[m]: Does it work?
< nishantkr18[m]>
zoq: Now, it seems to be working.. I think the best way to test would be to get https://github.com/mlpack/mlpack/pull/2413 merged, and then test the trained agent
< zoq>
nishantkr18[m]: Right, if you can handle the comment on the PR I think it's ready.
< nishantkr18[m]>
zoq: Done đī¸
< zoq>
nishantkr18[m]: :)
< kartikdutt18Gitt>
Hey @rcurtin, Could we upload separate csv for mnist train and test (as done for iris) as well as one for bodyfat.tsv on mlpack.org?
< rcurtin>
kartikdutt18Gitt: I don't have an issue with it; we used the local processing idea because shrit[m] pointed out that using other servers can help save any bandwidth costs
< rcurtin>
(I'm not near the bandwidth limit for mlpack.org, but it doesn't hurt to try and save bandwidth :))
< rcurtin>
adding bodyfat.tsv seems just fine for me, let me know where the file is and I'll put it on mlpack.org/datasets/
ImQ009 has quit [Quit: Leaving]
< zoq>
rcurtin: If we ever run into any traffic concers let me know I have unlimited traffic.
< jeffin143[m]>
Unlimited traffic
< jeffin143[m]>
Woh
< rcurtin>
zoq: sounds good; the bandwidth limits for me are quite generous too, so it would probably be a pretty crazy amount of traffic before we had any issues :)
< rcurtin>
not as generous as "unlimited" though :-D
< jeffin143[m]>
rcurtin (@freenode_rcurtin:matrix.org) : is it ok to add protobuf as dependancy in mlpack
< jeffin143[m]>
Or do we need a separate dashboard ???
< zoq>
They restricted to 10 Mbit/s after 10TB.
< rcurtin>
jeffin143[m]: I would really strongly prefer not to add another dependency
< jeffin143[m]>
Repository*
< jeffin143[m]>
Ohk
< zoq>
jeffin143[m]: About the jupyter notebook service looks interesting, we should test it on the examples/models repo.
< jeffin143[m]>
<jeffin143[m] "Ohk"> zoq (@freenode_zoq:matrix.org): yes indeed , I am not sure if that works with xesus
< zoq>
I think they just render the notebook, just like github does
< jeffin143[m]>
Rcurtin is it ok to add one more repo for vis tool which has proto buf as dependancy ??