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/
yassoudw has joined #mlpack
yassoudw has quit [Client Quit]
travis-ci has joined #mlpack
< travis-ci>
ShikharJ/mlpack#125 (Deconv - 1a3f924 : Shikhar Jaiswal): The build has errored.
< ShikharJ>
zoq: I'm struggling a bit in understanding the present layout of the `Gradients` code for the convolution layer. Why is the size of the `mappedError` reduced in the presence of padding?
< zoq>
ShikharJ: Good point, that's an issue in the implementation.
sulan_ has joined #mlpack
< ShikharJ>
zoq: The way I see it, for a simple non padded case, we are taking input A, convolving with kernel B to get the output C. Then were, taking A again and convolving with the differences (which have the same shape as the output) and getting the output the shape of B, and updating it?
< zoq>
ShikharJ: Yeah, that's the pipline.
< ShikharJ>
Then shouldn't it be the same for the padded case?
< zoq>
ShikharJ: Yes, can't think of any reason to resize the error, so I think you found a bug.
< zoq>
I think we should write another test case and fix this, I can do this, just let me know.
< ShikharJ>
Sure, I'll be busy with the transposed convolution layer testing (which I wish to finish today). Thanks for the help :)
manish7294 has quit [Remote host closed the connection]
sumedhghaisas has quit [Ping timeout: 260 seconds]
< zoq>
On a second thought, it makes sense to reduce the mapped error, since we are updating against the non-padded input, but I think this can be simplified.
travis-ci has joined #mlpack
< travis-ci>
ShikharJ/mlpack#126 (Deconv - 5ff2cd0 : Shikhar Jaiswal): The build has errored.