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/
< jeffin143[m]>
@shrit:matrix.org : great work
< jeffin143[m]>
If we are able to remove boost , it will be great
shashank has joined #mlpack
< shashank>
please provide me a small bug to fix
shashank has quit [Remote host closed the connection]
favre49 has joined #mlpack
< birm[m]>
shashank: you may like the "good first issue" tag on the github issues
ImQ009 has joined #mlpack
favre49 has quit [Remote host closed the connection]
< shrit[m]>
jeffin143 for boost program option it is easy. However, boost serialization requires more effort
desaidj has joined #mlpack
desaidj has quit [Ping timeout: 245 seconds]
desaidj has joined #mlpack
< desaidj>
i am a newbie
< desaidj>
and I am intrested to contribute to your organisation. How should i start
< zoq>
desaidj: Most of us some linux distribution, about the editor use whatever you like.
< zoq>
desaidj: I use vim, but I know a lot use Visual Studio Code.
< desaidj>
I've just downloadedthe ML pack... will you please help me out with the next step?
< zoq>
desaidj: As pointed out on the involve page, the next step is to build mlpack from source.
< zoq>
desaidj: After that you can test some tutorials, and maybe you can find some intersting issue on GitHub you like to solve.
< zoq>
desaidj: Or maybe you have some method in mind you like to see in mlpack?
< desaidj>
I'll check it out
desaidj has quit [Ping timeout: 245 seconds]
rcurtin_ has joined #mlpack
rcurtin has quit [Ping timeout: 246 seconds]
< saksham189Gitter>
@himanshupathak21061998 can you try finding the classification error with `mnist_first250_training_4s_and_9s.arm` for centers ranging from 10 to 1000 in increments of 5, 10 or 20 and paste the results on the PR.
< saksham189Gitter>
Also, did you find the issue with k-means?
< saksham189Gitter>
I think we have a meet right now, right?
< KimSangYeon-DGU>
Yes
< kartikdutt18Gitt>
Yes, we do.
< saksham189Gitter>
Is there anything specific you would like to discuss?
< KimSangYeon-DGU>
Ok, let's check the timeline and progress we've done
< kartikdutt18Gitt>
Hey @saksham189, I don't think we can see the messages from riot on hitter.
< kartikdutt18Gitt>
*gitter.
< saksham189Gitter>
Is Sangyeon using gitter?
< kartikdutt18Gitt>
Sure, as proposed in the timeline NMS and IoU metrics are now merged.
< KimSangYeon-DGU>
I'm on IRC
< kartikdutt18Gitt>
I think he is also using Riot.
< KimSangYeon-DGU>
Oh
< KimSangYeon-DGU[>
Can you see my message?
< saksham189Gitter>
hmm you mean IRC?
< KimSangYeon-DGU[>
Yes, now I'm Riot
< KimSangYeon-DGU[>
Both of them...
< abernauer[m]>
KimSangYeon-DGU: Yout messages are coming through on my end.
< kartikdutt18Gitt>
Yes. @KimSangYeon-DGU (@kimsangyeon-dgu:matrix.org) Messages from the IRC aren't visible on gitter.
< KimSangYeon-DGU[>
Oh... yes
< KimSangYeon-DGU[>
abernauer (@abernauer:matrix.org): thanks for letting me know
< saksham189Gitter>
Is SangYeon here yet?
< saksham189Gitter>
(edited) Is SangYeon here ... => Is Sangyeon here ...
< KimSangYeon-DGU[>
<saksham189Gitter "Is SangYeon here yet?"> saksham189 (Gitter): can you see my message?
< kartikdutt18Gitt>
He is here, Could you use the IRC or @KimSangYeon-DGU (@kimsangyeon-dgu:matrix.org) could use Gitter.
< abernauer[m]>
saksham189 (Gitter): He is in the chat, but I would recommend using riot if it's not inconvenient.
< KimSangYeon-DGU[>
I'll use Gitter
< abernauer[m]>
He said he will use gitter
< kartikdutt18Gitt>
@saksham189, @KimSangYeon-DGU (@kimsangyeon-dgu:matrix.org) is making the switch to Gitter.
< abernauer[m]>
Oh they probably can't seee my messages or saksham.
< saksham189Gitter>
yes thanks :)
< kartikdutt18Gitt>
About the progress, The PR for NMS and IoU have been merged. The PR for simple dataloader and it's documentation have also been merged.
saksham189 has joined #mlpack
< kartikdutt18Gitt>
Currently I'm working on Image Dataloader which adds the following functionalities :
< kartikdutt18Gitt>
Yes it's almost done. There are two things that need to be discussed. One of them is use of field vector, it would prevent time and memory when preprocessor for YOLO is added.
< saksham189Gitter>
This is confusing messages coming from everywhere :sweat_smile:
< kartikdutt18Gitt>
The other one is data split. mlpack doesn't template it split. So labels need to be row and input needs to be matrix. Since the labels in this case are also matrices, it would be templated or I could split them in dataloader itself.
< kartikdutt18[m]>
I also added Cifar 10 and 100 dataset as well and flow from directory functionality.
< KimSangYeon-DGU>
Nice
< KimSangYeon-DGU>
Some checks are failed, can you think you could fix them?
< KimSangYeon-DGU>
*do you think you could fix them?
< kartikdutt18[m]>
Yes it's ready for review. If we decide to go with field type then I'll add one more commit. Other than that once everything is done, I'll have to trouble rcurtin for uploading test set (small subset of data set on mlpack.org)
< KimSangYeon-DGU>
Ok, and I think it would be better to another PR handling templatization of Split
< KimSangYeon-DGU>
*to open
< kartikdutt18[m]>
About the data split, Kindly let me know what you think. It won't work if labels are of the field type or the matrix type. So should I template split or implement datasplit internally?
< KimSangYeon-DGU>
Do you mean `datasplit` as another split method?
< KimSangYeon-DGU>
I think it would be better to extend the existing split method
< KimSangYeon-DGU>
But, if it's only used internally we could do implement datasplit.
< kartikdutt18[m]>
> But, if it's only used internally we could do implement datasplit.
< kartikdutt18[m]>
Yes, it will be used internally only.
< saksham189>
Do you think it would be useful anywhere else in the codebase or users? If not then we could just implement locally for now.
< kartikdutt18[m]>
I am not sure, Most users could concat the features and labels and perform train test (data)split provided their arma data type is compatible.
< kartikdutt18[m]>
That would not be the case if we use feild type for labels.
< kartikdutt18[m]>
* That would not be the case for us if we use feild type for labels.
< kartikdutt18[m]>
I can implement it locally and open an issue for templating it if that makes sense?
< KimSangYeon-DGU>
Yes :)
< saksham189>
kartikdutt18 after this we would start with the darknet model?
< kartikdutt18[m]>
Yes.
< KimSangYeon-DGU>
For summary
< KimSangYeon-DGU>
Oops
< kartikdutt18[m]>
Since we have the image data loader for classification and detection ready we can start with the implementation. I'll also write the image dataloader tutorial this week.
< kartikdutt18[m]>
I had one more question.
< KimSangYeon-DGU>
Great, go ahead
< kartikdutt18[m]>
Which dataset should we use for training Darknet model, Cifar 10 has 60 k images but images are simpler and smaller. Pascal VOC has lesser images but larger. Kindly let me know which one should I use. I don't think imagenet is an option since it has around 15m images.
< saksham189>
I think we should go with CIFAR since it is very commonly used and should give us a good measure of the performance
< saksham189>
KimSangYeon-DGU what do you think?
< KimSangYeon-DGU>
I think so
< KimSangYeon-DGU>
First, let's try with the simple and small dataset
< kartikdutt18[m]>
Sure, makes sense. We'll use Cifar then.
< kartikdutt18[m]>
> First, let's try with the simple and small dataset
< kartikdutt18[m]>
Any dataset that you suggest?
< KimSangYeon-DGU>
Ahh, I mean CIFAR-10 :) And then, let's try with the more complex dataset later on
< kartikdutt18[m]>
Great.
< KimSangYeon-DGU>
Ahh, I mean CIFAR-10 :) And then, let's try with the more complex dataset later onFor summary, there are two blockersOne is about field vector and Kartik is working on that and will open a PR.The other is about mlpack’s Split() and you’ll implement datasplit method for internal use. Also, we’ll open a PR for mlpack’s Split() needs to be
< KimSangYeon-DGU>
templatized.
< KimSangYeon-DGU>
Oops
< KimSangYeon-DGU>
For summary, there are two blockersOne is about field vector and Kartik is working on that and will open a PR.The other is about mlpack’s Split() and you’ll implement datasplit method for internal use. Also, we’ll open a PR for mlpack’s Split() needs to be templatized.
< saksham189>
I think we'll just be implementing the split locally for now and later see if we need to extend mlpack's Split method.
< KimSangYeon-DGU>
Yes, agreed
< KimSangYeon-DGU>
kartikdutt18[m]: Can you open an issue about the mlpack's Split method?
< saksham189>
kartikdutt18 Is there anything else you want to discuss?
< kartikdutt18[m]>
Sounds good, I'll update the rest of the community with the new functionalities and the all the work done in my weekly blog.
< kartikdutt18[m]>
> kartikdutt18: Can you open an issue about the mlpack's Split method?
< kartikdutt18[m]>
That PR was ready. Last week I trained LeNet 1,3 and 5 and created weights as well as wrote tests for model. The tests include loading weights for training and predicting on validation dataset such that they have an error less than the threshold.
< kartikdutt18[m]>
However the issue arrises of invalid memory access most probably due to lack of copy constructors. There is a PR open for that as well as two issue that might be resolved by it 1. Invalid access in FFN and 2. Invalid read in GANs and Simple DQN.
< KimSangYeon-DGU>
Ok, then the PR has nothing to do with the project, right?
< kartikdutt18[m]>
> Ok, then the PR has nothing to do with the project, right?
< kartikdutt18[m]>
Directly no, If the invalid memory access is resolved and the PR is merged, The testing part for all models would already exist. Otherwise I can simply copy it for my other PRs.
< HimanshuPathakGi>
Also I think because the subset only contain 500 samples we don't have to go for 1000 centres k-means will not work for greater than 500.