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/
vivekp has joined #mlpack
petris has quit [Remote host closed the connection]
petris has joined #mlpack
manish7294 has joined #mlpack
ImQ009 has joined #mlpack
manish7294 has quit [Ping timeout: 260 seconds]
govg has quit [Ping timeout: 264 seconds]
ShikharJ_ has joined #mlpack
sulan_ has joined #mlpack
vivekp has quit [Ping timeout: 248 seconds]
vivekp has joined #mlpack
govg has joined #mlpack
< ShikharJ_>
zoq: You there? I wanted to discuss the implementation of the Transposed Convolution Layer a bit.
ShikharJ___ has joined #mlpack
ShikharJ_ has quit [Ping timeout: 260 seconds]
ShikharJ___ is now known as ShikharJ_
< zoq>
ShikharJ_: Hello, please go ahead.
< ShikharJ_>
Referrig from this paper (https://arxiv.org/pdf/1603.07285.pdf), Relation 14, I believe we can let the convolution layer template be as it is, and just change the input to the required size and pass on to the existing routines. But I'm not sure whether it'd be the right thing to do.
< ShikharJ_>
*Referring
vivekp has quit [Ping timeout: 260 seconds]
< zoq>
ShikharJ_: Yeah, that way we could simply use the backward pass from the existing conv layer.
< zoq>
we also have to set outputWidth and outputHeight
< zoq>
and we also have to set the output paramter which is set inside the forward pass, but I guess it might be a good idea to just duplicate the code and modify the backward pass?
< ShikharJ_>
I guess that should work, basically, we're swapping backward and forward passes right? From the template?
< zoq>
yes
< zoq>
It might be a good idea to implement a simple test that compares the result of the Forward pass against some manually computed results.
< zoq>
Same for the backward pass.
< ShikharJ_>
That clears up things for me. Thanks, actually the visual 4.7 in the paper had a erroneous description (we would get a 3 x 3 input in whch we have to insert zeroes and not a 2x2).
< ShikharJ_>
Checkout the third line on the description (page 27)
< zoq>
I wonder if the same issue exsists in the coressponding github project
< ShikharJ_>
Seemingly the github project doesn't mention this description, only the technical report does.
< ShikharJ_>
I'll open an issue, let's see.
< zoq>
yeah, good idea
vivekp has joined #mlpack
sulan_ has quit [Quit: Leaving]
vivekp has quit [Ping timeout: 264 seconds]
govg has quit [Ping timeout: 240 seconds]
govg has joined #mlpack
vivekp has joined #mlpack
ImQ009 has quit [Ping timeout: 256 seconds]
ImQ009 has joined #mlpack
ImQ009 has quit [Ping timeout: 256 seconds]
sulan_ has joined #mlpack
ImQ009 has joined #mlpack
ShikharJ_ has quit [Quit: Page closed]
ShikharJ has joined #mlpack
ImQ009 has quit [Ping timeout: 255 seconds]
ImQ009 has joined #mlpack
ImQ009_ has joined #mlpack
ImQ009 has quit [Ping timeout: 256 seconds]
sulan_ has quit [Quit: Leaving]
ShikharJ has quit [Quit: Leaving]
ShikharJ has joined #mlpack
ShikharJ has quit [Client Quit]
govg has quit [Ping timeout: 240 seconds]
sulan_ has joined #mlpack
< rcurtin>
ok, so since #1382 (OpenMP warning/bugfix) is merged, I think it would be a good idea to release mlpack 3.0.1
< rcurtin>
does anyone have any other open PRs or changes that they would like to see incorporated before the release? (of course, it is easy to just release another 3.0.2 shortly after this too)
< zoq>
I guess you could include Big Batch SGD as well.
sulan_ has quit [Quit: Leaving]
ImQ009_ has quit [Quit: Leaving]
< rcurtin>
right, I'll note that in the history file also