02:38 < ironstark> rcurtin: Sure, I'll run them for an algorithm and see how we can display the results better. Also, I ran the command make run METHODBLOCK=NBC LOG=True on slake but the reports.db file did not get generated. How do I generate that?

10:39 < zoq> ironstark: Did you use driver: 'sqlite' and database: 'reports.db' in the config file? Also you can benchmark specifc libs: make run METHODBLOCK=NBC LOG=True BLOCK=mlpack,shogun CONFIG=config.yaml

13:05 < kris___> lozhnikov: https://gist.github.com/kris-singh/ee57aee957c9e1780e982342d1e5bd96

13:05 < kris___> Here are the pre-processing scripts changed according to the paper.

13:06 < kris___> Though it seems incorrect to me to compare 2 algorithms with different feature vectors.

13:06 < kris___> Let me know if the feature extraction scripts are correct i would then run the ssRBM on top of it.

13:59 < lozhnikov> kris___: the paper doesn't describe how many patches we should get from each image

13:59 < lozhnikov> right now you get only one patch from each image

13:59 < lozhnikov> that could be incorrect

14:00 < kris___> no if you look at the file random_patch.py

14:00 < kris___> for j in range(0, random_samples):

14:01 < kris___> where random_samples is the number of random samples you want to take

14:01 < lozhnikov> ah, I see

14:02 < lozhnikov> the paper doesn't describe the preprocessing part. I mean "return preprocess(allpatch[0:5000][:])"

14:02 < lozhnikov> so, I think it isn't needed

14:03 < kris___> https://www.irccloud.com/pastebin/1LJiqOeh/

14:03 < kris___> In all of the experiments, "images were pre-processed by PCA whitening retaining 99% of the variance"

14:03 < kris___> This is from the paper.

14:03 < lozhnikov> let me check the paper

14:04 < kris___> This is in paragraph 4

14:18 < lozhnikov> hmm, I see. I think the transform is incorrect. The paper states the images were pre-processed PCA whitening i.e. the PCA matrix is equal to sqrt(S) * U' * X in your notation

14:19 < lozhnikov> * by PCA whitening

14:19 < kris___> Yes i used the zca whitening. Because i don't want another parameters k in my calculations

14:19 < kris___> For pca whiteing you would have to define the number of components.

14:20 < lozhnikov> could you provide the link?

14:20 < kris___> To what ??

14:21 < lozhnikov> https://stats.stackexchange.com/questions/95806/how-to-whiten-the-data-using-principal-component-analysis

14:21 < lozhnikov> there are no extra parameters

14:22 < rcurtin> lozhnikov: kris___: I know I have not been a part of the discussion, but maybe WhitenUsingSVD() or WhitenUsingEig() in src/mlpack/core/math/lin_alg.hpp is helpful?

14:22 < rcurtin> sorry if my comment is off base :)

14:23 < kris___> lozhnikov: http://scikit-learn.org/stable/modules/generated/sklearn.decomposition.PCA.html

14:23 < lozhnikov> rcurtin: nice idea:)

14:26 < lozhnikov> kris___: I see. Maybe ZCA whitening gives the same results, maybe not. The only thing I know: the paper uses PCA whitening.

14:27 < kris___> hmmmm sure i could pca whitening from skelarn and leave the n_components parameter empty...

14:28 < lozhnikov> I think it is reasonable to verify that you use reshape() properly. Are you sure that the function reshapes matrices in correct order?

14:29 < lozhnikov> and the paper states "In all of the experiments, images were pre-processed". Actually, you pre-precess patches rather than images

14:30 < lozhnikov> so, I think it is reasonable to sample patches from pre-processed images. How do you think?

14:30 < kris___> Ahhh sorry i did do that in patches.py file

14:31 < kris___> but i think i did not change that in random_patches.py file

14:32 < kris___> I will make the changes sometime later today. And send you the gist. I will also zca --> pca.

14:34 < lozhnikov> sounds good. what do you think about Ryan's idea ("maybe WhitenUsingSVD() or WhitenUsingEig() in src/mlpack/core/math/lin_alg.hpp is helpful")?

14:35 < rcurtin> I guess, I only suggested that in the case that it might save you the time it takes to implement a whitening method yourself

14:35 < kris___> Ahhh well i would like to handle this in python. Mainly because i can code it quickly and i want to keep the image processing part in python.......

14:35 < rcurtin> I haven't looked through it too specifically, but I would imagine that WhitenUsingSVD() is almost identical to PCA whitening

14:35 < kris___> rcurtin: Well sklearn already has support for pca whitening.

14:36 < rcurtin> ah, if you are just doing some offline preprocessing and you are already using python, maybe this is the better idea, but it is up to you; I didn't know if you needed to implement the whitening in code you were already using or what

14:46 < kris___> lozhnikov: Any updates on the conv - gan model ??

14:47 < kris___> I tried to test the discriminator alone like you said but i would need some more time....

14:49 < lozhnikov> kris___: I'll dig into the on Saturday

15:21 < lozhnikov> * into that

15:51 < rcurtin> kris1: I read your blog post, very cool to see that the RBM implementation is 1.5x faster than scikit

20:38 < kris___> rcurtin: Thanks...