10:58 < manthan> zoq : you there?

11:00 < manthan> i explicitly printed costPlus, costMinus, orgGradient and estGradient for the test and found that for both cases (full module and specific module), costPlus = costMinus, but still in full module test, estGradient comes to be equal to orgGradient but in specific test it comes to be 0.

11:00 < manthan> might be due to any precision loss issues?

11:05 < zoq> possible

11:07 < zoq> Checked the backward and gradient step and I don't think there is an issue.

11:10 < manthan> ya then why could the checkgradient be failing?

14:16 < daivik> rcurtin: zoq: I opened PR #12 on the models repo which implements a CNN for the digit recognition Kaggle challenge. Unfortunately, there was already an open PR for implementing LeNet for the same - I apologize, I completely overlooked this. As of now, my model performs slightly better than the one on PR #9 (the LeNet PR) - giving a 94.485% accuracy

14:16 < daivik> on the hidden test data on Kaggle in 40 epochs of training. Also, I think that PR#9 is going to use the original MNIST dataset now (instead of the Kaggle dataset)? I'm hesitant to close my PR right away because I think that I can push the performance a bit more with parameter tuning.. What do you suggest?

14:17 < daivik> *hyperparameter tuning

14:22 < rcurtin> daivik: I agree, I'd prefer to take the model that does better. I hadn't looked into that #9 was going to use the MNIST dataset, I agree that we should stick wtih the Kaggle dataset

14:23 < rcurtin> go ahead and leave it open, I'll try and comment in the next few days to figure out how to reconcile the two PRs, maybe we can take a little from each (but I haven't looked closely at yours, and it's been a few days since I looked at #9)

14:39 < daivik> okay, sorry again for not checking the open PRs beforehand

14:40 < rcurtin> no problem at all, there are a lot of open PRs :)

21:19 < rcurtin> alright, I have tested and working PyPI packages for linux that I'll upload once the final rebuild is done

21:19 < rcurtin> nothing for Windows though

22:49 < xa0> hi all, i'm wondering if mlpack supports multi layer perceptrons?

23:16 < rcurtin> xa0: yeah, you can use the deep learning framework to build one

23:17 < xa0> i've just discovered this :D now to upgrade to version 3.0.0 so the deep learning framework actually even exists

23:17 < xa0> thanks rcurtin

23:17 < rcurtin> you could look in src/mlpack/tests/feedforward_network_test.cpp for an example

23:17 < rcurtin> and also http://mlpack.org/docs/mlpack-3.0.0/doxygen/anntutorial.html

23:17 < rcurtin> sure, happy to help :)

23:18 < xa0> gotcha!