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/
< chopper_inbound[>
Hi! Is it possible to multiply two cubes of shapes (t, e, b) and (e, s, b) having same number of slices and each slice is compatible to be multiplied, and without using for loop i.e. without iterating over all the slices? (In this case the resultant cube has to be of shape (t, s, b))
< zoq>
chopper_inbound[: I don't think so no.
< chopper_inbound[>
zoq: Oh yes. I just found -
< chopper_inbound[>
"* Matrix multiplication of two objects; not applicable to the Cube class unless multiplying a cube by a scalar"
favre49 has joined #mlpack
favre49 has quit [Ping timeout: 256 seconds]
favre49 has joined #mlpack
favre49 has quit [Quit: Lost terminal]
< HimanshuPathakGi>
Hey @saksham189 @zoq How much it should take to calculate 1000 centres in mnist dataset.
< HimanshuPathakGi>
*time
< HimanshuPathakGi>
Normally
< HimanshuPathakGi>
Aah forgot to mention using kmeans
< zoq>
HimanshuPathakGi: Over the entire dataset, that could take some time, also always depends on your machine.
< zoq>
HimanshuPathakGi: Do you see slow timings?
< HimanshuPathakGi>
Yes very slow actually from Approximately 6-7 hours and it is still calculating. I have low config machine i3 and 4gb Ram
< HimanshuPathakGi>
I think that's why
< zoq>
HimanshuPathakGi: Yeah, probabl not enough memory.
< zoq>
HimanshuPathakGi: You said it works on a subset right?
< HimanshuPathakGi>
Yes it was working on subset
< zoq>
HimanshuPathakGi: I can train it on another machine, just to make sure the results are resonable, if you think that is a good idea.
< zoq>
HimanshuPathakGi: But if it works on a subset, I think it will work on the full dataset as well.
< HimanshuPathakGi>
Yeah, I think it is a better idea
< HimanshuPathakGi>
It was giving 0.144 classification error
< HimanshuPathakGi>
With subset
< zoq>
right, that sounds good
< zoq>
HimanshuPathakGi: If you like you can also ssh into the machine, if that makes things easier for you.
< HimanshuPathakGi>
Sure can you details for that :)
< HimanshuPathakGi>
Share
< zoq>
It's a remote machine with more memory and a better CPU.
< HimanshuPathakGi>
So, can I share a cpp file to run RBFN:)
< HimanshuPathakGi>
Here
< zoq>
HimanshuPathakGi: Sure I can build the code as well.
< HimanshuPathakGi>
Hey zoq I have updated my PR and
< HimanshuPathakGi>
Now you can build the code and run the file thanks for helping
< HimanshuPathakGi>
:)
< zoq>
HimanshuPathakGi: Alright, will build the code later.
< HimanshuPathakGi>
Sure when you get time :)
< walragatver[m]>
birm: jeffin143: Are you here?
< saksham189Gitter>
@himanshupathak21061998 were you able to solve the issue with k-means?
< HimanshuPathakGi>
Yes when I tried it with the mnist_all.csv
< HimanshuPathakGi>
It was working I think
< HimanshuPathakGi>
There was problem when executing with mnist_train.csv
< saksham189Gitter>
Can you try to identify the exact problem and close the issue you created if it is not required?
< HimanshuPathakGi>
Also it is mentioned We have to find centres from all train and test
< saksham189Gitter>
where is it mentioned?
< HimanshuPathakGi>
In the paper
< saksham189Gitter>
I don't think so. Can you quote the exact line here?
< HimanshuPathakGi>
Wait
< saksham189Gitter>
(edited) I don't think so. Can ... => hmm. thats wierd. Can ...
< saksham189Gitter>
Anyways I think I found what you are talking about.
< saksham189Gitter>
Also you are getting a classification error of 0.144. Is this in percentage?
< HimanshuPathakGi>
The k-means are trained on the whole training+testing dataset 5.2
< saksham189Gitter>
> Also you are getting a classification error of 0.144. Is this in percentage?
< saksham189Gitter>
.
< HimanshuPathakGi>
> Also you are getting a classification error of 0.144. Is this in percentage?
< HimanshuPathakGi>
It is just same
< HimanshuPathakGi>
we are calculating in
< HimanshuPathakGi>
Test network function
< saksham189Gitter>
Hmm.. In the paper they achieve 3.3% so compared to that it is pretty high I would say. How many centres have you used?
< HimanshuPathakGi>
> Can you try to identify the exact problem and close the issue you created if it is not required?
< HimanshuPathakGi>
I am trying to find out it only occurs when k-means is used with mnist_train.csv
< HimanshuPathakGi>
> Hmm.. In the paper they achieve 3.3% so compared to that it is pretty high I would say. How many centres have you used?
< HimanshuPathakGi>
75 because it is small
< saksham189Gitter>
So, I guess check the difference between the csv files.
< HimanshuPathakGi>
> So, I guess check the difference between the csv files.
< HimanshuPathakGi>
Sure, once I find out I will close that issue
< saksham189Gitter>
Also in the output layer can you try having the same number of units as the number of labels and train that using MeanSquared Error (same as before). And then the prediction would be the node that outputs the highest value.
< jeffin143[m]>
@walragatver:matrix.org :
< jeffin143[m]>
Sorry I forgot
< jeffin143[m]>
About our meeting
< jeffin143[m]>
Sorry
< jeffin143[m]>
Really
< walragatver[m]>
jeffin143: it's fine
< HimanshuPathakGi>
> Also in the output layer can you try having the same number of units as the number of labels and train that using MeanSquared Error (same as before). And then the prediction would be the node that outputs the highest value.
< HimanshuPathakGi>
Previously I was trying it with that I used threshold of 1.5 for that
< jeffin143[m]>
:)
< walragatver[m]>
I think saksham saksham is taking a meet. Let's wait for them to finish
< jeffin143[m]>
Yeah sure :)
< birm[m]>
hi walragatver
< walragatver[m]>
saksham189 (Gitter) : let us know when we are done.
< walragatver[m]>
birm: Hi birm
< HimanshuPathakGi>
and share the results
< HimanshuPathakGi>
So, I think @zoq should wait to test the network
< HimanshuPathakGi>
Before that we should try mean squared error
< walragatver[m]>
Okay it looks like they are done
< walragatver[m]>
jeffin143: I have went through your PR.
< HimanshuPathakGi>
Yes you guys can carry on +1
< HimanshuPathakGi>
:)
< walragatver[m]>
You have done a fascinating job. I am quite happy that we have started well.
< walragatver[m]>
It's quite a big PR, it will take sometime to review it.
< jeffin143[m]>
walragatver : the only problem with threading was sometimes not terminating and not locking
< jeffin143[m]>
But I managed to debug it and it works fine
< RyanBirminghamGi>
I have marked it as a draft, just so I don't have to ask if you want it reviewed :)
< jeffin143[m]>
walragatver: I have shared two snippets of code in readme file you can try both of them and view them in tensorboard
< jeffin143[m]>
Ryan Birmingham (Gitter): thanks , I should have done it , because I was still working on that :) , but thanks for doing it
< walragatver[m]>
jeffin143: So the PR is not yet complete?
< jeffin143[m]>
No , the shared queue needs some more function just like isempty ()
< jeffin143[m]>
I am assuming it may come in handy sometime
< jeffin143[m]>
Also I think we should add metadata to scaler plots so I was figuring it out if that's possible
< jeffin143[m]>
By tmrw the pr will be ready to be reviewed
< walragatver[m]>
> By tmrw the pr will be ready to be reviewed
< walragatver[m]>
Okay fine just take your time
< jeffin143[m]>
walragatver : but I am sure we are fast enough with the timeline
< walragatver[m]>
Hmm yeah
< jeffin143[m]>
If possible , can we enable style checks like main repo , if possible
< jeffin143[m]>
So that i can Rectify them in advance
< walragatver[m]>
> <@jeffin143:matrix.org> If possible , can we enable style checks like main repo , if possible
< walragatver[m]>
> So that i can Rectify them in advance
< walragatver[m]>
jeffin143:
< RyanBirminghamGi>
I assume that the basic CI functions would be part of the next PR?
< walragatver[m]>
Yeah I want to discuss that
< RyanBirminghamGi>
(That said, you can't do the jenkins style checks via a PR, so I'll need to bug rcurtin zoq)
< walragatver[m]>
> (That said, you can't do the jenkins style checks via a PR, so I'll need to bug rcurtin zoq)
< walragatver[m]>
Okay I see
< walragatver[m]>
jeffin143: Once you completely the PR. Can you refer to MLpack/models/.ci folder and write .yaml scripts to trigger Azure pipelines?
< RyanBirminghamGi>
You may be able to make travis or azure perform those checks, if that interests you.
ImQ009_ has joined #mlpack
ImQ009_ has quit [Read error: Connection reset by peer]
< jeffin143[m]>
But what should I check
< jeffin143[m]>
In Mlpack models repo their were excecutbales
< jeffin143[m]>
But here there are only header files
< RyanBirminghamGi>
Start with just checking if cmake and make work
< jeffin143[m]>
Ok, I will try and adding those
ImQ009 has quit [Ping timeout: 246 seconds]
< walragatver[m]>
> Start with just checking if cmake and make work
< walragatver[m]>
Yeah correct
< jeffin143[m]>
By any chance if any of you have windows machine ???
< walragatver[m]>
Basically those .yaml files are written in specific format which trigger build.
< walragatver[m]>
> By any chance if any of you have windows machine ???
< walragatver[m]>
I have it
< jeffin143[m]>
So did you build mlboard in windows machine ??
< walragatver[m]>
But I never use it
< jeffin143[m]>
I build in mac and it works fine ? , And since mac is close with Ubuntu it should work ok in Linux as well
< jeffin143[m]>
But I wanted to see how it builds on windows , I will try and get a vm set in mac if possible
< walragatver[m]>
jeffin143: Hmm, yeah it's fine.
< walragatver[m]>
Windows build is not that priority now
< jeffin143[m]>
> Basically those .yaml files are written in specific format which trigger build.
< jeffin143[m]>
I will take a look
< jeffin143[m]>
I don't have anything else on my list today
< walragatver[m]>
> You may be able to make travis or azure perform those checks, if that interests you.
< walragatver[m]>
I quite wonder how it works on Jenkins. Do they parse the code?
< walragatver[m]>
jeffin143: birm: Regarding style checks I will let you know. I am also quite new to it.
< walragatver[m]>
And that's it from my side
< walragatver[m]>
birm: Anything you want to discuss?
< jeffin143[m]>
> jeffin143: birm: Regarding style checks I will let you know. I am also quite new to it.
< jeffin143[m]>
No hurry it will make your review easier
< jeffin143[m]>
Because I have history of making many styling error 😅
< walragatver[m]>
> I assume that the basic CI functions would be part of the next PR?
< walragatver[m]>
Also please add those .yaml files in new PR. Only basic file structure would be fine. Without code in it
< walragatver[m]>
I think birm might have lost his connection.
< jeffin143[m]>
Ok
< jeffin143[m]>
New pr as in the existing one
< jeffin143[m]>
Or open a fresh ??
< walragatver[m]>
Fresh PR. Containing build files and those static proto files as they are not going to change that greatly if I am not wrong and would be needed for successful build.
< walragatver[m]>
Basically it would help us in reviewing. I hope you would not mind.
< jeffin143[m]>
Not at all
< jeffin143[m]>
It would help me to
< jeffin143[m]>
I will do it :)
< jeffin143[m]>
Too*
< RyanBirminghamGi>
Ah, sorry, I upgraded my connection speed and I think it made it worse
< walragatver[m]>
birm: Okay, anything you would like to discuss?
< RyanBirminghamGi>
See an example style build for how the lint checks work