00:22 < tham> I finished the function exportFiltersToPGMGrid, which could show the encoder matrix of sparseAutoencoder

00:22 < tham> which module would you suggest me to place it?

00:23 < tham> exportFiltersToPGMGrid(std::string const &fileName, arma::mat const &input)

00:24 < tham> I try to add zero check to SparseAutoencoder, and find out this do not have obvious impact to the performance

00:25 < tham> try on 256 features and 10000 samples, with and without nan/inf checking, both of them finish at the same time(63 seconds on my machine)

00:42 < tham> ah, sorry, it is 64 features and 10000 samples

01:43 < tham> finished bilinearInterpolation, now exportFiltersToPGMGrid could specify the size of each block

01:43 < tham> The api looks like

01:43 < tham> arma::mat exportFiltersToPGMGrid(std::string const &fileName, arma::mat const &input, size_t height_per_block, size_t width_per_block);

02:29 < zoq> tham: Right now, there isn't a tutorial or something like that that shows how to use the code to do image classification. You can look at the cnn test suite, which uses the modules to classify a subset of the mnist dataset. I'm working on a tutorial including some demos that shows the potential of the existing code: https://urgs.org/mnist/index.html

02:29 < zoq> If you are planing to do speech recognition maybe you like to read soemthing about recurrent neural networks.

02:58 < tham> thank you very much

02:59 < tham> By the way, when I studying the test cases, I find out the cnn and fnn cannot compile by vc2015

03:00 < tham> I think it maybe a bug of the compiler, you can check the details on the pull request

03:01 < tham> Thanks for your cool ann modules

04:02 < naywhayare> tham: I have a bunch of other stuff to pop off the stack tomorrow, but I'm really hoping to have a chance to reproduce your vs2015 bug and commit the fix

04:02 < naywhayare> by the way, you mentioned that you sent a mail to the mailing list, but I never saw it... are you subscribed to the list? you need to be subscribed to post, so maybe that is the issue

04:07 < tham> I already subscribed to the list, the email I send to is "mlpack-git-request@cc.gatech.edu"

04:08 < tham> Maybe I should open an issue on github?

04:19 < tham> I also send to "mlpack-git@cc.gatech.edu", anyway, thanks for your works

15:30 < zoq> tham: You have to subscribe to the mlpack discussion list available at: https://lists.cc.gatech.edu/mailman/listinfo/mlpack (mlpack@cc.gatech.edu) the mlpack-git@cc.gatech.edu is the commit notification list and is read only.

20:04 < tham> After some experiment, link to openBLAS is much faster than BLAS even on windows, it is atleast three times faster on my laptop

20:05 < tham> If you plan to provide builded version for version, please link to openBLAS

20:07 < tham> I will update the build instruction later, my first build of mlpack is based on the blas, but not openBLAS

20:07 < tham> hope this help out

20:07 < tham> thank you

20:11 < naywhayare> nice! I should test openblas timings on linux too

20:11 < naywhayare> then we can say mlpack is not singlethreaded in many cases, too :)