01:43 < tham> Hi, I do some refactoring on sparse_autoencoder

01:45 < tham> the speed improve about 28.795%

01:46 < tham> I compare the results with the original implementation, the result of the output are same(the final output optimize by the L-BFGS optimizer)

01:47 < tham> What I did is reduce the temporary objects and recalculation of the activation of hidden layers, output layers and the rhoCap

01:51 < tham> I intent to add two more template parameters to SparseAutoEncoderFunction, try to make it more versatile

01:51 < tham> Things like "template<typename HiddenLayer, typename OutputLayer> class SparseAutoencoderFunction"

01:56 < tham> The last thing is I notice that current implementation of SparseAutoEncoderFunction do not check on "division by zero"

01:57 < tham> and prevent the log function exist any zero member(ex : log(0))

01:57 < tham> Forget to check or in the real world, we do not need to worry about them?

02:00 < tham> Should I do the checking?Or I could safely omit them?Thanks

02:01 < tham> I would like to commit it back to the community if the results are reasonable

02:02 < naywhayare> hi tham, I'd love to include your contributions

02:03 < naywhayare> you can submit a pull request on github or send an email to the mailing list with the patches and I'll look over them tomorrow and commit them

02:03 < naywhayare> as for the division by zero, often mlpack methods will assume that the data is not going to cause problems like this, in order to be faster

02:19 < tham> ok, I will learn how to submit a pull request on today, and thanks for your reply

02:22 < tham> I also send an email to the mailing list yesterday, it mention about the compilation issue of current codes

02:22 < tham> I hope this could save some troubles for the other users when they try to compile the next release of mlpack on windows

03:07 < naywhayare> tham: I didn't see the email... I will see if the list is having any issues

03:08 < naywhayare> (tomorrow though. I need to sleep now...)

04:13 < tham> bye, I do not know why you haven't see the email.Whatever, i post the contents at here(http://pastebin.com/3xFhzEeZ)

04:13 < tham> Have a nice dream

14:06 < naywhayare> zoq: I didn't realize he was talking about writing a sparse autoencoder with the ann code, I thought he meant the one Siddharth wrote in methods/sparse_autoencoder/

14:06 < naywhayare> I can still take a look through the pull request, but presumably you have a plan for the ann code so the decision is yours :)

14:26 < zoq> I just throw a glance at the code. I guess I can look deeper into the pull request in the next few days. It would be nice to see a test for the code. Would save me time to write my own test suite. Maybe tham is kind enough to write some test cases.

14:49 < naywhayare> we could also probably adapt the sparse autoencoder tests that siddharth wrote... I'll look into that

15:50 < zoq> ah right, that would be the best idea, ideally it should work without any modification

15:53 < naywhayare> yeah it will probably need some minor API changes but not anything else