13:27 < danR_> hi everyone! Installed mlpack3 from pip and I am trying to pass a numpy array of features to hmm_train (which I think I read somewhere should be possible), but it is not working. With the current python bindings, is it only possible to pass a string with a file location? Thanks

14:08 < rcurtin> danR_: hi there, unfortunately you are right, it is currently only possible to pass a string with a file location

14:09 < rcurtin> the file can contain data, or a list of other files which are each training sequences to train on

14:09 < rcurtin> I agree this isn't an optimal interface, but it's what we currently have. it could be improved, it would just take a little bit of implementation to do it...

15:07 < danR_> rcurtin: Okay, I see. Thanks! :)

18:29 < luffy1996> zoq: I was recently going through the implementation of Arcade Learning Environment https://github.com/mgbellemare/Arcade-Learning-Environment . What I saw that there exists an easy interface to import the ALE on C++ . You might want to visit Page 6 of https://github.com/mgbellemare/Arcade-Learning-Environment/blob/master/doc/manual/manual.pdf . Do you think, we can add ALE support to mlpack?

20:43 < daivik> rcurtin: zoq: I wrote this code --> https://thepasteb.in/p/xGhmg4zVgJyFM

20:44 < daivik> and it gave this output --> https://thepasteb.in/p/wjh04z2jkXrtv

20:45 < daivik> What i'm trying to do, is to minimize a simple function y = x*x using the Adam Optimizer.

20:47 < daivik> I've written the same code using tensorflow --> https://thepasteb.in/p/Q1hBKEYyomNC8

20:49 < daivik> Here is the tensorflow output --> https://thepasteb.in/p/pghQJN78rjqSR

20:50 < rcurtin> daivik: thanks, I'll look shortly. I'm traveling at the moment and also under a bunch of deadlines... :-\

20:52 < daivik> rcurtin: Okay, sure. No worries. -- I'll just post some thoughts here anyway, for future reference.

20:55 < daivik> It seems that the mlpack optimizer stops after some 3000 odd iterations (and doesn't reach the optimum), quite a bit earlier than the specified maxIterations. I'm looking into this, but any leads would be greatly appreciated.

20:57 < daivik> While the tensorflow optimizer -- under the same parameters and initial value of x, runs through the all of maxIterations iterations and reaches the optimal value of y.

21:01 < zoq> daivik: Note the optimizer will stop if the tolerance between two iterations is reached, the default tolerance is 1e-5. So either you can set the tolrance at creation time or afterwards with optimizer.Tolerance() = -1;

21:02 < zoq> daivik: Let me know if that solves the issue.

21:08 < daivik> zoq: Yes, you're right -- that is why the mlpack optimizer stops. I actually wanted to see where the CNN model I wrote (models repo PR #12), differs from a tensorflow implementation - because I've tried hyperparameter tuning for a while now, and I'm not able to get any performance improvement.

21:08 < daivik> Let me check what the terminating criteria for TFs AdamOptimizer is -- they don't seem to have a tolerance option.

21:19 < daivik> zoq: yes, optimizer.Tolerance() = -1 did solve the issue. Thanks a lot. Any hints on what could be the issue with why the CNN model is not performing as well as expected?

21:26 < zoq> daivik: Not sure, have to take a closer look into the code, will do that tomorrow.