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/
wasiq has quit [Ping timeout: 252 seconds]
mentekid has joined #mlpack
Mathnerd314 has quit [Ping timeout: 260 seconds]
mizari has joined #mlpack
mizari has quit [Quit: Leaving]
Mathnerd314 has joined #mlpack
< rcurtin>
mentekid: I'm running some benchmarks, but they are taking a while... I am trying with lots of cutoff values
< rcurtin>
on the corel dataset it seems like cutoff=0.9 was actually a good choice for both default parameters and L=10,K=10, but I need to try with more datasets to understand
< mentekid>
Ah cool. Maybe once every test is done we'll get a clear picture :)
< mentekid>
by the way, I wanted to discuss my project abstract and milestones: a) Do you have any comments/suggestions about the abstract? GSoC said we can change them, do you think I should add/remove anything?
< mentekid>
b) regarding the milestones, I think I could move some things around... For example I have the "action plan describing proposed changes" scheduled for the end of the first week, but given that I more or less have a good idea of what has to be done, I believe I could actually deliver it on day 1
< rcurtin>
last I saw it, the abstract looked fine, but let me take a look again
< rcurtin>
this week, my entire team has been in atlanta for an onsite meeting, so I have had very little time to do what I want this week
< rcurtin>
mostly just sitting in meetings
< rcurtin>
but tomorrow everyone is gone again so I can get back to work...
< mentekid>
ok then I can gather my thoughts (I have other questions too, for example about the blog) and we can talk tomorrow or during the weekend whenever you have time
< rcurtin>
sure, tomorrow is fine
< rcurtin>
I'll let you know when I am online
< rcurtin>
I should have a good chunk of free time in the morning (like 1400-1600 UTC)
marcosirc has joined #mlpack
< mentekid>
that's good for me too, I'll be in the lab all day tomorrow
sumedhghaisas has joined #mlpack
< marcosirc>
Hi Sumedh.
< sumedhghaisas>
Hey Marcos...
< sumedhghaisas>
How are you??
< marcosirc>
Fine! Thanks.
< sumedhghaisas>
is your new computer setup??
< sumedhghaisas>
any problems??
< marcosirc>
No problems :) Everything is OK.
< marcosirc>
Running on lastest Ubuntu 16
< marcosirc>
Builds and tests.
< sumedhghaisas>
ohh is that stable? I saw lot of complaints online...
< sumedhghaisas>
yet the new features are awesome...
< marcosirc>
Yes? Fortunately, I did not have any problem!
< sumedhghaisas>
good to know that...
< marcosirc>
Just to know your time zone. Are you living in India?
< sumedhghaisas>
yeah... its 9:00 PM here...
< sumedhghaisas>
at your place??
< marcosirc>
Ok! Great to know.
< marcosirc>
I am in Rosario, Argentina.
< marcosirc>
its 12:30 here.
< sumedhghaisas>
ohh lunch time... So is this time works best for you??
< sumedhghaisas>
So I was thinking I will be online 2 days in a week atleast...
< sumedhghaisas>
I will try to be more involved...
< marcosirc>
No problem, thanks! Generally, I have lunch around 13.30 hs.
< sumedhghaisas>
Thursday will be one...
< sumedhghaisas>
Do you have any preference for the other??
< marcosirc>
Ok, Thursday is Ok for me.
< marcosirc>
Is Monday ok?
< marcosirc>
I don't have a strong preference.
< sumedhghaisas>
Sure... Monday is okay... Although sometimes I have meeting on monday night... in that case I will inform you by mail...
< sumedhghaisas>
then maybe we can shift Tuesday... we can decide that on mail...
< sumedhghaisas>
Anyways you can reach me anytime through mail...
< marcosirc>
Ok for me.
< marcosirc>
Yes, I will consider emails too!
< sumedhghaisas>
Okay So I will give you little background about my work in mlpack... and also in whole... just as an introduction...
< sumedhghaisas>
I did GSOC once under mlpack...
< sumedhghaisas>
In essence...
< sumedhghaisas>
I implemented many SVD algorithms to support the recommender framework...
< sumedhghaisas>
Before that I did many bug fixes in other modules...
< sumedhghaisas>
So have you seen the CF framework of mlpack??
< marcosirc>
Interesting! No, I haven't, I am reading your gsoc blog right now.
< sumedhghaisas>
Ohh great... its not that good though :)
< sumedhghaisas>
So In CF ... after decomposing the user and rating matrix ... KNN is used to find recommendations...
< sumedhghaisas>
anyways... just some extra information...
< marcosirc>
Cool, a clear real world application.
< sumedhghaisas>
yup... We did lot of tests with real world datasets ...
< sumedhghaisas>
you can find the result in the current benchmarking...
< sumedhghaisas>
The other thing I love working on are game players...
< sumedhghaisas>
I had a research opportunity in Reykjavik... where I was introduced to a concept called General Game Player...
< marcosirc>
Nice, I will go in depth next days!
< sumedhghaisas>
basically rather than playing a specific game it acts based on a runtime game input...
< sumedhghaisas>
basically it can play any game with game logic as input...
< sumedhghaisas>
its a pretty nice idea
< sumedhghaisas>
read about it if you get some spare time...
< sumedhghaisas>
I think AI will be my major when I will go for my M.Sc. :)
< sumedhghaisas>
Thats it... so in this busy corporate schedule... This is all I get to do... I envy student life now :)
< sumedhghaisas>
So when did you started working in machine learning??
< sumedhghaisas>
I know that you have done GSOC 3 times... thats great...
< marcosirc>
Sounds interesting! Another topic to read about!
< marcosirc>
Ahh ok. So you plan to continue studing or continue working in the computer industry?
< marcosirc>
Yes! I have participated 2 times before and were great experiences!
< sumedhghaisas>
No I am definitely getting out of this e-commerce field... first not much to research here... and very bad work life balance :(
< sumedhghaisas>
I am planning to pursue my M.Sc. next year at University of Edinburgh :)
< sumedhghaisas>
in AI
< sumedhghaisas>
So our last projects were in which field??
< sumedhghaisas>
*your
< marcosirc>
Ouch. I could imagine...
< marcosirc>
Sounds great. Some friends have done a M.Sc. Vibot
< sumedhghaisas>
Now I envy my good old university days :)
< marcosirc>
whiches includes a semester in Edinburgh
< marcosirc>
My past GSoC experiences were in Haskell related fields. On real world applications of Haskell.
< sumedhghaisas>
vision and robotics ahh...
< sumedhghaisas>
Ohh haskell??
< sumedhghaisas>
How is that language?? I never got any intro to that language...
< marcosirc>
You know, my degree is very general, so you learn a wide range of things, and at the end you are a bit confused where to go!
< sumedhghaisas>
Isn't your degree in computer science??
< sumedhghaisas>
I am bachelor of engineering ??
< sumedhghaisas>
*mean
< marcosirc>
Yes. What I mean is that we learn about many areas in computer science.
< marcosirc>
It is a 5 years long degree...
< marcosirc>
In Europe, some countries validate it as a M.Sc.
< sumedhghaisas>
ohh 5 years... mine was just 4 years... feel good that you re getting 1 extra year of fun :)
< marcosirc>
Haha yes! I really enjoys it.
< sumedhghaisas>
So what your your favorite area?? any particular subject you want to pursue??
< marcosirc>
I have done a course in AI and another in ML, as part of my degree.
< marcosirc>
As I got really interesting in this area, so I decided to focus on this.
< sumedhghaisas>
we will get along pretty well then :P
< marcosirc>
Well, I think I am interest in data structures / algorithms, specially applied to real world problems related to ML and AI.
< sumedhghaisas>
ohh thats a looooot fun...
< sumedhghaisas>
those tree algorithm for AI and dynamic programming...
< sumedhghaisas>
I loove them...
< sumedhghaisas>
I sort of use the same stuff in my game player :)
< marcosirc>
Thanks! :)
< marcosirc>
It was great fun!
< marcosirc>
I have done an internship for iRobot in 2014, I also was an amazing experience for me.
< sumedhghaisas>
This is just n introductory chat so feel free to run for your lunch any time you want :) but I would love to learn haskell from you sometime later ... I presume its similar to lisp... ??
< marcosirc>
that made me move from "formal" areas of computer science into more "real world" applied problems...
< sumedhghaisas>
ohh great...
< sumedhghaisas>
summer internship??
< marcosirc>
Thanks! I don't have problem.
< sumedhghaisas>
What type of work were you assigned?? Any interesting project??
< marcosirc>
Well, it is a functional programming language such as lisp. But is far more advanced, with many features.
< marcosirc>
Yes, a 4 months internship.
< marcosirc>
I worked with the Roomba product, and many external tools.
< sumedhghaisas>
Roomba ... sounds cool ....
< sumedhghaisas>
well need to do a google search for that :)
< sumedhghaisas>
Yeah I had lisp in my bachelor course... cant imagine anything advance than that :)
< marcosirc>
:) Yes. You know... when you work some time in a program and then you see a robot moving ... it is a great experience!
< sumedhghaisas>
I can imagine... Sometime real world related to AI...
< sumedhghaisas>
thrilling...
< marcosirc>
Yes! Thanks! So, you are now working for ebay, is it right?!
< sumedhghaisas>
yup... mostly working on barcode and optical character recognition... ]
< sumedhghaisas>
like the ones on products...
< sumedhghaisas>
I didn't have any experience in CV before coming here... tough it is...
< marcosirc>
Sounds intereseting!! I don't have experience in CV. I hope I get a chance to learn more in the near future!
< sumedhghaisas>
think before you hope :P
< sumedhghaisas>
anyways ... I also wanted to know you familiarity with C++...
< marcosirc>
Haha :) Well... My friends that made the Vibot M.Sc. I mentioned before, were fascinated by CV.
< sumedhghaisas>
cause we need to do some restructuring in KNN ...
< marcosirc>
Ok.
< sumedhghaisas>
and that would involve some design and C++11 things...
< sumedhghaisas>
at least I can't imagine being fascinated by CV... although I like maths its the huge data that bothers be sometimes :)
< marcosirc>
I have some past experience in C++. But, you know... C++ is a very complex language. So I am not an expert.
< marcosirc>
What I really enjoy learning on my own.
< marcosirc>
Yes, I understand!
< sumedhghaisas>
No one is expert in C++ :P its too powerful to wield elegantly :P
< sumedhghaisas>
yes thats the best way to understand C++... :)
< sumedhghaisas>
there is this book which was referred to me by Ryan...
< marcosirc>
Haha, yes! And every new version includes new things.
< sumedhghaisas>
when I was a student...
< marcosirc>
I have developed a Go player in C++.
govg has joined #mlpack
< sumedhghaisas>
thats explains new patterns ... might be helpful... but I am unable to remember the name :)
< marcosirc>
Then I used that template MCTS library to implement some game players:
< sumedhghaisas>
ohh thats great... I will definitely take a look...
< sumedhghaisas>
this is my fav field :)
< marcosirc>
For Connect4,Hex,Tateti and finally for Go.
< marcosirc>
Thanks! :)
< marcosirc>
I should translate documentation to English.
< sumedhghaisas>
yeah thats what I was going to say :P
< sumedhghaisas>
I understood nada...
< sumedhghaisas>
but the code looks great...
< marcosirc>
Haha. It was a copy paste, from a presentation to a conference here in Argentina, because of that it is in Spanish...
< marcosirc>
I will work on that!
< sumedhghaisas>
We also use some variant of MCTS with many other techniques to create the General Game Player...
< sumedhghaisas>
I remember the day it beat me in Connect 4 :)
< marcosirc>
hahahaa
< marcosirc>
Yes!! my connect4 player beats me too!!
< sumedhghaisas>
behold the time to singularity :P
< sumedhghaisas>
I hate to break this amazing conversation but I have go for dinner now...
< sumedhghaisas>
Before that
< sumedhghaisas>
So you were going to work on your references next week...
< marcosirc>
It is ok!
< sumedhghaisas>
So maybe we ca discuss about the project from next week... assuming I also will read some to get on the page :)
< marcosirc>
Yes. I have read the design guidelines and documentation provided by Ryan. Maybe I should focus on the papers related to this project now.
< sumedhghaisas>
If you get stuck or need my help... feel free to ask me :)
< marcosirc>
Thanks!
< sumedhghaisas>
It will be great working with you :) now that I know we share almost same interests...
< marcosirc>
I was going to review "An Investigation of Practical Approximate Nearest Neighbor Algorithms", where spill trees are explained in depth.
< sumedhghaisas>
yeah thats the paper I went through... very quickly...
< marcosirc>
Thanks, I agree! I hope we have a great summer working together!
< sumedhghaisas>
maybe we can discuss about that next week...
< sumedhghaisas>
got to go now... :)
< marcosirc>
Ok, sounds great!
< marcosirc>
See you. Good dinner!
marcosirc has quit [Quit: WeeChat 1.4]
nilay has joined #mlpack
nilay has quit [Quit: Page closed]
nilay has joined #mlpack
nilay has quit [Quit: Page closed]
nilay has joined #mlpack
< nilay>
Hi zoq
< nilay>
i have been reading about the localization algorithm for the project
< zoq>
nilay: Hello, how are things going? All exams over?
< zoq>
nilay: Sounds good.
< nilay>
yes.. exams over..
< nilay>
both edge boxes and selective search seem to be equal in performance
< nilay>
and both are very complex algorithms, if we try to implement them from scratch
< nilay>
so how do you recommend we choose one over the other.?
< nilay>
and does the implementation of these CV algorithms also have to be part of mlpack??
< nilay>
or we can just take any implementation and adapt it in any way we want?
< nilay>
edge boxes seems to outperform selective search in running time though.
< zoq>
Yeah, I guess, both algorithm are equally good when it comes the number of possible regions of interests. I, think it would go with edge boxes, mainly for two reason: 1. It's way faster 2. We could reuse the Hoeffding tree tree for the strcutured tree. But it's your project, to it's your decision.
< zoq>
I guess, we could use the OpenCV (beta) implementation as reference, and see if we get almost the same results.
< nilay>
opencv 3?
< zoq>
The code isn't really polished, so I'm not sure, it's helpful to be used as reference implementation.
< zoq>
So, I guess if you say, you like to implement the edges boxes method, we could go from there and Tham and I can probably tell you more about the way we could move on
< nilay>
yes i guess, if it is easier to implement then we should.
< zoq>
For selective search we have to implement the Graph-Based image method, which isn't that complicated either.
< zoq>
I've you like to implement another localization method it's also fine, the Bing method is also neat
< zoq>
*If
< nilay>
The core part of edge boxes (structured random forests) and selective search (hierarchical grouping using graph method) seem understandable to me.
< nilay>
it is when they start getting too technical about details of images and computer vision i have a bit difficulty
< zoq>
No problem, we don't expect that anyone is an expert ... the structured random forests is basically an decision tree with PCA or KMeans on top ... we can go trough the details step by step
< nilay>
and it is very easy to read all the implementations, understand little, get lost in details and not choose early, so i guess i should take some advice on this..
< nilay>
yes i read about structured random forests in a paper.
< zoq>
yeah, so my advice is "go with edge boxes" :)
< zoq>
or selective search or Bing :)
< nilay>
do we want to have a neat api for implementation algorithm or do we adapt it from some library?
< nilay>
haha
< zoq>
mlpack, already comes, with a neat API ... but we should dicuss the interface, before we start coding. We should keep it as simple as possible. At the end we just need a method that takes an image and returns possible locations.
< nilay>
ok
< zoq>
If we implement structured random forests, we can reuse the decision stump API.
< nilay>
how do we use hoeffding tree for structured tree
< zoq>
So, the Hoeffding tree is basically a decision tree for streaming data. I'm not quite sure, if we should use the Hoeffding tree or the decision stump ... I'll have to think about that.
< nilay>
and in that random forest class we could use normal decision tree with decision stump, the cloud's code that you mentioned?
< zoq>
Cloud's code is basically a good start we could use to implement the structured tree because it's an decision tree. The forest, uses the decision trees and votes over all trees.
< nilay>
ok, i understand.
< nilay>
and regarding the api: we mention the input(image) and output(locations). but what about the more specific details of first edge detection and then edge boxes method..
< nilay>
i sent you the public key for ssh login
< zoq>
We have to think about that, do you think you can propose something, that we could use as basis for our discussion, probably in the next days? I think we could use the opencv as basis.
< zoq>
Let me know if the login works for you or if you need any package etc.
< nilay>
i get a permission denied (publickey) error.
< nilay>
have you copied the public key or should i wait
< zoq>
hm, it should work, you used 8088 as port right?
< nilay>
yes i typed in the command as you mentioned.
< nilay>
do i have to do anything else also?
< zoq>
hm, I don't think so, can you send me the output if you add '-v'?
< nilay>
should i paste it here?
< zoq>
email, would be good
< nilay>
ok
< zoq>
Does the key you send me match with the pub key in /home/nilay/.ssh/id_ed25519.pub?
< nilay>
no file named id_ed25519.pub
< nilay>
in .ssh dir
< zoq>
oh, okay, but there is an private key called /home/nilay/.ssh/id_ed25519 ?
< nilay>
in the .ssh folder ls -a outputs: . .. authorized_keys id_rsa id_rsa.pub known_hosts
< nilay>
the private key is in the directory where i created the public key.
< zoq>
I'm confused, it's says 'debug1: Trying private key: /home/nilay/.ssh/id_ed25519'. Can you put the keys you created in /home/nilay/.ssh/
< zoq>
and try again
< nilay>
should i rename the file name also? from key.txt to id_ed25519.txt?
< zoq>
yes, good idea
< nilay>
*from key.txt to id_ed25519?
< zoq>
and key.txt.pub to id_ed25519.pub
< nilay>
yes it works now.
< zoq>
great :)
< nilay>
any directions on how do i use this
< zoq>
hm, first of all you can't really break things, if you do, no problem ... I'm not sure how your workflow normally looks like (e.g. what editor or IDE you normally use), but you could clone mlpack and see if you could build it.
< nilay>
ok i'll try building mlpack. and i use sublime and vim mostly.
< nilay>
and how do i break things? so that i won't do them.. :)
< nilay>
is this an ubuntu machine?
< zoq>
hm, I guess we find out once that happens :)
< zoq>
freebsd
< nilay>
oh i haven't used this before
< zoq>
I guess, from your perspective there shouldn't be much of a difference
< nilay>
there is no apt-get
< zoq>
yeah, right, if you need any package just tell me
< zoq>
so, I guess, you could use vim or use sublimetext in combination with the rsync plugin?
< zoq>
like if you prefer to have htop or screen over tmux, etc.
< nilay>
i don't use htop or tmux.. but can't i install any package on this.?
< nilay>
suppose i need to install python.?.
< zoq>
If it's not already installed, send me a message.
< nilay>
ok.
< nilay>
its not. or maybe path isn't updated. i cannot use the pkg command it seems.
< zoq>
python pkgutil?
< nilay>
pkgutil?
< zoq>
do you mean freebsd pkg?
< nilay>
yes
< zoq>
ah, you have to be root to be able to use pkg
< nilay>
zoq do you think maybe i should use my machine only, and use this remote machine when it is really needed..
< nilay>
or would it make much difference and calculations would be fast on remote machine..
< zoq>
nilay: that's up to you, use the way, you most comfortable with. I think, if you need more CPU performance or you need to run method xyz for hours, you can just sync your local code to the remote machine.
< zoq>
You can definitely use your local machine for coding and testing.
< nilay>
yes that is a good idea. :)
< zoq>
and switch to the remote whenever you think it's necessary
mentekid has quit [Ping timeout: 260 seconds]
< nilay>
i get this error when i build on remote machine after cloning. [ 30%] Linking CXX executable ../../../../bin/mlpack_fastmks CMakeFiles/mlpack_fastmks.dir/fastmks_main.cpp.o: file not recognized: File format not recognized c++: error: linker command failed with exit code 1 (use -v to see invocation) *** Error code 1 Stop. make[2]: stopped in /usr/home/nilay/mlpack/build *** Error code 1
< zoq>
nilay: ah right, I'll need to figure out why that happens, for now just comment fastmks in src/mlpack/methods/CMakeLists.txt
< zoq>
and the coressponding test src/mlpack/tests/CMakeLists.txt '#serialization_test.cpp and #fastmks_test.cpp'
< nilay>
build successful now. i'll study things discussed today and come up with something in next days.. thank you for your time zoq :)
< zoq>
nilay: sounds good, see you in the next days.