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/
< kartikdutt18Gitt>
Hi @tejasvi, I have left some comments, Have a look when you get the chance.
volhard[m] has joined #mlpack
< volhard[m]>
Should I feed a recurrent net (GRU) audio in the time or frequency domain?
< kartikdutt18Gitt>
Hi @volhard, you can work in time domain also but generally data is converted into frequency domain using techniques such as stft.
< kartikdutt18Gitt>
I think in frequency patterns become more apparent, if you have heard about tdft, it is a discrete version of fourier transform which is periodic.
< metahost>
volhard: you may! Tasks like wake word detection (like Hey Siri), use CNNs (+ sliding window) to predict when a phrase is detected. If you need contextual information, you should probably use RNNs.
< kartikdutt18Gitt>
Hi @zoq, could you have a look at #2195, I wanted to know how I should proceed with it.
pranay2 has joined #mlpack
pranay2 has left #mlpack []
< PrinceGuptaGitte>
Hi everyone, I was trying to get better understanding of codebase of ann. I have a doubt.
< PrinceGuptaGitte>
Why is everything templated instead of using inheritance approach? For example BaseLayer class acts as a template for activation layers where we can use any type of activation function. We could have also done it so that BaseLayer's takes in a ActivationFunction type and all activation functions inherit from it. Could it be because of virtual functions being slow? Or is there some other reason.
< PrinceGuptaGitte>
I'm sorry if it's a silly doubt but I don't understand why would we wanna template everything.
pranay2 has joined #mlpack
< PrinceGuptaGitte>
(edited) ... we wanna template ... => ... we want to template ...
pranay2 has quit [Remote host closed the connection]
pranay2 has joined #mlpack
< GauravSinghGitte>
Hi, @zoq kindly have a look at #2191 I have incorporated the changed that you suggested. Thank you.
pranay has quit [Remote host closed the connection]
pranay2 has quit [Remote host closed the connection]
< GauravSinghGitte>
(edited) ... the changed that ... => ... the changes that ...
lozhnikov_ has quit [Ping timeout: 268 seconds]
lozhnikov has joined #mlpack
< tejasvi[m]>
What should be the ideal workflow of development? The way I debug is tragic. After making changes in the file, I build and run mlpack_test and use BOOST_TEST_MESSAGE as a cout incarnate. Given the ~45m build time this isn't helpful enough. I tried gdb but it refuses to drill beyond the code of test file. Should I use gdb with handmade code like examples in https://www.mlpack.org/doc/mlpack-3.2.2/doxygen/sample.html? I'm
< tejasvi[m]>
bit swamped here.
< SriramSKGitter[m>
@tstomar[m] I've noticed much faster build times on subsequent builds. It's only the first time that its ~45 min. Building with -jN and building only specific targets ought to bring compile time down to reasonable levels.
< PrinceGuptaGitte>
@tstomar[m] Also whenever you run "make" command it only builds the files that changed (and the connected files)
< PrinceGuptaGitte>
@kartikdutt18 I've noticed MLpack doesn't have regular softmax, it only have log softmax. Should I add regular softmax layer?
< zoq>
PrinceGupta: There is an open PR that implements the regular SoftMax layer.
< sreenik[m]>
Yes, I remember having started it but didn't finish it
< sreenik[m]>
If I remember correctly there is probably some mistake in that PR but it would be a lot easier to finish it up rather than take up the work from scratch
< zoq>
agreed
< sreenik[m]>
In case anyone is interested, feel free to take it up
EL-SHREIFGitter[ has joined #mlpack
< EL-SHREIFGitter[>
why their is no mentor in Visualization Tool project for GSoC 2020?
< zoq>
EL-SHREIF: Nice catch, will update it later.
< PrinceGuptaGitte>
I don't understand what's the problem here
< zoq>
PrinceGupta: My first guess is the input size isn't correct.
< PrinceGuptaGitte>
Input size is (42000,784) where 42000 is no.. of data samples
< PrinceGuptaGitte>
and ouput is one ho encoded matrix of (42000,10)
< PrinceGuptaGitte>
(edited) ... one ho encoded ... => ... one hot encoded ...
< zoq>
PrinceGupta: Note armadillo column major, so the matrix size should be (784,42000).
< PrinceGuptaGitte>
ok I'll try to transpose them
< PrinceGuptaGitte>
@zoq apparently the program works when I only take 1 sample (from the 42000 available) and then feed that to `.Train()` function.