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/
< rcurtin[m]>
hmm, that function is just a crappy preprocessor implementation that substitutes the right types for fpeT1, eT1, fpeT2, eT2, etc.
< rcurtin[m]>
all that should be needed to add a new kernel is to add it to kernelsrc.hpp, and then *also to kernelid.hpp (the ordering in the src file should match the ordering in kernelid.hpp)
< zoq>
Yeah, wondering if I just add some crappy code on top for now :)
< rcurtin[m]>
but maybe I misunderstood the issue?
< zoq>
I mean I can't add __global__ void COOT_FN(PREFIX,accu)(const float* in, const UWORD n_elem, float* out)
< zoq>
if I already have __global__ void COOT_FN(PREFIX,accu)(const eT1* in, const UWORD n_elem, eT1* out)
< zoq>
since the method will substitute the type
< zoq>
so I end up with more than one instance of an overloaded
< zoq>
__global__ void COOT_FN(PREFIX,accu)(const float* in, const UWORD n_elem, float* out) is not 100% correct
< zoq>
So if I add COOT_FN(f_,accu) to the kernel_src, substitute_types will add another one COOT_FN(f_,accu).
< zoq>
So my understanding is that I can only add a kernel that matches for all types.
< rcurtin[m]>
oh, yeah, I see; that is how I wrote it, yeah
< rcurtin[m]>
feel free to change how the functionality works, maybe adding a blacklist for kernels to not generate or something?
< zoq>
Maybe, yes; also I think I can make the kernel generic enough and build upon the subsitute type method.
< rcurtin[m]>
awesome, feel free to modify it in whatever way is needed
< rcurtin[m]>
what I wrote was just the first thing that seemed to fit the needs I had at that particular moment :)
Samyak has joined #mlpack
Samyak has quit [Remote host closed the connection]
Anwaar has joined #mlpack
Anwaar has quit [Ping timeout: 245 seconds]
melonred[m] has quit [Quit: Idle for 30+ days]
< AakashkaushikGit>
Hey @zoq, I am curious about what you needed to do to make it work ? Also the comparisons look very similar in the context of timing for the small network.
< AniketKhandagale>
Hey everyone! I'm Aniket Khandagale 2nd year computer science undergraduate from MU, India. I'm kind of new but really interested to start contributing. I have knowledge of C++ and would love to learn other stuff in the way. A little guidance would help. Can i get link for beginner issues which i can fix
< AyushSingh[m]>
I would now like to work on Addition of Transformer Encoder Layer.
< AyushSingh[m]>
I guess that everything required for it has already been implemented, except a layer normalization layer.
< AyushSingh[m]>
So, should I start with implementation of a layer normalization layer?
< AyushSingh[m]>
Just to confirm - There isn't anything left to work in making a seq2seq architecture as we already have implemented the LSTM, right? The user can create the architecture him/herself.
< rcurtin[m]>
jeffin143: the sun hasn't gone down here yet, but it looks like a clear night, so I'm hoping to at least wander outside and take a look :)
< shrit[m]>
It rained the entire day here, storms, winds. I do not have that chance :(