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/
ImQ009 has joined #mlpack
sds has joined #mlpack
sds is now known as Guest65307
Guest65307 has quit [Client Quit]
vivekp has joined #mlpack
wenhao has quit [Quit: Page closed]
< ShikharJ>
zoq: I received some good results on the full dataset as well :) Posting in the PR now.
< ShikharJ>
zoq: I think we can cut down a lot on the time elapsed, once the support for batches is implemented. We are all set to merge the GAN PR now!
< ShikharJ>
zoq: Do you mean the `std::cout << "Loading Parameters" << std::endl;` statement?
< zoq>
yeah
< zoq>
If you like keep the output, but we should use Log::Debug.
< zoq>
That way a user can disable the output.
< ShikharJ>
zoq: I'll make use of Log::Info?
< zoq>
yeah, that's fine as well.
< ShikharJ>
zoq: Done!
< zoq>
ShikharJ: Okay, once the travis build finished, I'll hit the merge button :)
< ShikharJ>
zoq: Great. COuld you also tell of a source where I can find the CelebA dataset in the csv format? The original author's link seems to have gone dead, and I'm only able to find the jpg or png format files.
< manish7294>
zoq: I was using BigBatchSGD with adaptive stepsize and it turns out that effectiveBatchSize is surpassing permissible value leading to errors, whereas backtracking line search works fine. Here is the backtrace https://pastebin.com/q0LWjM4V
< manish7294>
initially I kept batch size of 50
manish7294 has quit [Ping timeout: 260 seconds]
< Atharva>
zoq: In ffn_impl.hpp line 188 : res += Evaluate(parameters, i, true); is their some reason to writing the third parameter as a boolean, because there is no definition of Evaluate with the third parameter as a boolean and I think it's just getting converted to size_t = 1? Am I missing something?
< ShikharJ>
zoq: One build runs the 10,000 image subset, should be done in 8 hours from now.
< zoq>
manish7294: Ahh, can I use the code from the PR to reproduce the issue?
< zoq>
ShikharJ: Nice, let's see if we get some reasonable results.
< zoq>
manish7294: Another optimizer you could test is SGRD with CyclicalDecay.
< Atharva>
zoq: That's true but that function has four parameters, the one I pointed to has three.
< Atharva>
The evaluate function with 3 parameters has size_t as the last parameter
< zoq>
ohh, you are right, nice catch
< zoq>
this should be Evaluate(parameters, i, 1, true);
< zoq>
Do you like to open a PR or should I fix this?
< Atharva>
It's just one line, if you can directly push it, it will be better.
< Atharva>
Is it okay?
sulan_ has joined #mlpack
manish7294 has joined #mlpack
< manish7294>
zoq: I haven't pushed the change regarding optimizer on the PR yet because of the issue but you can reproduce it by just passing the BigBatchSGD through lmnn_main.cpp --- you will just have to make a very short change in the last part of lmnn_main.cpp
< zoq>
Atharva: Yeah, no problem.
< manish7294>
zoq: I tried SGDR. The results were similar to SGD and comes up with the same problem of coordinates matrix divergence we are facing with SGD. So, I guess BigBatchSGD and L-BFGS are the best for us.
< zoq>
manish7294: Okay, I'll take a look into the issue.
< ShikharJ>
zoq: For the next one week, I would like to focus on getting the support for batches and optimizer separation (and debug and test the DCGAN tests on the side). Would that be fine?
< zoq>
ShikharJ: Absolutely, let us put some time into tuning the existing code.