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/
govg has joined #mlpack
govg has quit [Ping timeout: 240 seconds]
govg has joined #mlpack
mikeling has joined #mlpack
govg has quit [Ping timeout: 264 seconds]
govg has joined #mlpack
mikeling has quit [Quit: Connection closed for inactivity]
< rcurtin>
zoq: I think some of the travis time issues will be fixed with the bindings commit; it will allow boost::program_options to not be included, which should help
< rcurtin>
at the same time it will now build a bunch of python targets... so any gains may be useless in the end if the python build takes too long...
< rcurtin>
(although, for the travis build, it'd be easy enough to just disable the python bindings build, we could do that on jenkins)
< zoq>
rcurtin: okay that sounds nice, just wanted to see if switching to trusty helps in some way.
< rcurtin>
yep, definitely if it does merge it in :)
palashahuja has joined #mlpack
< palashahuja>
I just wanted to know whether momentum is implemented for neural networks ?
< zoq>
palashahuja: I guess you mean momentum for SGD? If that's the case, the answer is no.
< palashahuja>
and what about adaptive learning rates ?
< zoq>
palashahuja: For SGD? Unfortunately no, but you could use RMSProp or Adam. Both method adjusts the learning rate while training.
< palashahuja>
I actually wanted to try rmsprop with momentum, is it possible in mlpack ?
< zoq>
hm, Adam already includes momentum and bias-correction, so yes.
< zoq>
But it would be great to see momentum for the SGD optimizer.
< palashahuja>
so, adam optimizer is a bit like rmsprop ?
< zoq>
Yes, the difference is mainly that Adam also keeps track of the past gradients, which is basically what momentum does.
< palashahuja>
ok, cool
< palashahuja>
also, what about nesterov momentum ?
< zoq>
Would be nice to have :)
palash123 has joined #mlpack
palash123 has quit [Client Quit]
< zoq>
rcurtin: 37 min 52 sec to 49 min 5 sec not bad ... let's restart and see if the time stays the same.
palashahuja has quit [Ping timeout: 240 seconds]
palashahuja has joined #mlpack
< rcurtin>
ah, I guess I did not see the finished time before you restarted
< zoq>
36 min 27 sec looks good
< rcurtin>
I agree, big improvement!
< zoq>
Now each build runs in an isolated Google Compute Engine virtual machine instead of a container on a shared host via Docker. The disadvantages: it takes longer to start the job after a commit.