rcurtin_irc changed the topic of #mlpack to: mlpack: a scalable machine learning library (https://www.mlpack.org/) -- channel logs: https://libera.irclog.whitequark.org/mlpack -- NOTE: messages sent here might not be seen by bridged users on matrix, gitter, or slack
ShubhamAgrawal[m has quit [Quit: You have been kicked for being idle]
SimeetNayan[m] has quit [Quit: You have been kicked for being idle]
PavanKalyanThota has joined #mlpack
gegen07GermanoBa has joined #mlpack
SuvarshaChennare has joined #mlpack
progs has joined #mlpack
progs has left #mlpack [#mlpack]
PragatoBhattacha has joined #mlpack
PragatoBhattacha has left #mlpack [#mlpack]
PragatoBhattach4 has joined #mlpack
<PragatoBhattach4> Greetings,
PragatoBhattach4 is now known as progs[m]
<progs[m]> Greetings everyone! I'm Pragato Bhattacharjee a student of Maulana Abul Kalam Azad University of Technology, India currently in my 2nd year of pursuing a bachelor's degree in Computer science and engineering. I have a keen interest in the field of Deep Learning and have made small contributions to a few open source projects in the past. I am eager to apply my skills and work collaboratively with others to make a meaningful contribution to the
<progs[m]> open source community. I have been looking forward to pursue the the 'ready to use models' contribution mentioned in your GSoC idea list.
<IWNMWEIWNMWE[m]> Hey is there any way I can contact the mentor or ask for a proposal review
<vaibhavp[m]> rcurtin : Hey, I had some questions about
<rcurtin[m]> akhunti1: you need to extract the filename as a `std::string`; the `SeldonMessage` class must contain the filename parameter somehow (but I am not sure exactly how, I'm not familiar with Seldon)
<rcurtin[m]> SUKRUTA JENA: sorry for the slow response, you should be able to use the Slack inviter... wait, I see that it is down. let me look into that...
<rcurtin[m]> (but, the Slack inviter just takes you to a Slack channel that is bridged to this, so if you're already here through gitter/matrix, Slack will show you the same stuff)
<rcurtin[m]> * into that... *edit*: actually, it's up, sorry for the confusion
<rcurtin[m]> IWNMWE (IWNMWE): probably the best way to ask for a proposal review is here or through email, but as you can probably tell folks are pretty busy :) so it could be a while
<vaibhavp[m]> rcurtin: Hi! How are you doing? So, I am stuck with a problem regarding the Layer class with respect to the proposed DAG network. Can you help me out here or share what are your thoughts on this?
<vaibhavp[m]> solutions seem very elegant or efficient in my opinion. So what do you think we can do here? Is there a efficient solution to this problem?
<vaibhavp[m]> So, the current implementation was made with consideration for a stack-like network in mind and you cannot pass multiple matrices/inputs into the Forward/Backward or Gradient methods directly. One of the solutions is to join the input matrices before passing, which will be very slow and perform unnecessary copies. Another solution is to pass pointers to the matrices beforehand and pass nothing in the input matrices. But none of the
<vaibhavp[m]> Thank you!
<rcurtin[m]> vaibhavp: this is not a great answer and I realize that (sorry, I am short on time), but you might consider scrolling way back in the history here; I am pretty sure the DAG network API and implementation was discussed at length probably in April or May of last year. without digging too deeply into what you have proposed, the DAG network can't realistically be built using a `MultiLayer` or even the `FFN` class, it probably requires its own
<rcurtin[m]> `DAGNetwork` class or similar that internally does all the management to pass a single input matrix to a single input layer
<vaibhavp[m]> <rcurtin[m]> "vaibhavp: this is not a great..." <- Thank you for the answer. I will try to find it. And also I have already created a DAGNetwork class(I have shared the code previously) that manually stores and traverses the graph to perform the forward and backward passes, the bottleneck to the traversing the graph was performing the joining of the input matrices at the moment making the code very slow to execute for a simpler ResNet
<vaibhavp[m]> model.
<rcurtin[m]> you can almost certainly avoid copies of data in between layers by using aliases to some preallocated memory; see how FFN (and MultiLayer) deal with this issue
<vaibhavp[m]> rcurtin[m]: That came into my mind, but how will a single matrix hold data from two or more different matrices?
<rcurtin[m]> look at the code for FFN and MultiLayer, the answer is there
<rcurtin[m]> this only works if you are concatenating the matrices; if you want to add the results of two layers (or do some other reduction), there is no way around doing that computation
<vaibhavp[m]> Ohhh.... I get it now.
<vaibhavp[m]> I might have a solution now.
<vaibhavp[m]> Let me try it out. Maybe it works.
<vaibhavp[m]> I think it will setting the output of each layer and input of layer such that inputs of a layer are continuously stored. This will involve cleverly traversing the graph from the last layer back to the first layer... uhmm... let's see.
<vaibhavp[m]> * it will involve setting the
<zoq[m]> <progs[m]> "Greetings everyone! I'm Pragato..." <- Hello, let us know if there is anything we should clarify.
<zoq[m]> > <@sukrutajena:gitter.im> Hello Everyone ,My name is SUKRUTA JENA. I am from Bhubaneswar, Odisha. I am a 3rd-year undergraduate pursuing Btech in Computer science and technology from Parala Maharaja Engineering College, India.... (full message at <https://libera.ems.host/_matrix/media/v3/download/libera.chat/3c408193a89205b84e433629b209fd0c532bfd08>)
<zoq[m]> <Janan[m]> "Hello everyone! My name is Janan..." <- Hello, thanks for getting in touch :)
Guest3 has joined #mlpack
Guest3 has quit [Client Quit]
krushia has joined #mlpack