Hello Prasanna, thanks for getting back.
> I was going through application guideline given at mlpack wiki page and I came > across testing the project section. I never actually thought about testing up > until now but now that I think about it, it is hard to test generative models. > One direct way to do it is give it a noisy input image and perform CD to get > sample from model distribution and make sure that reconstruction error is > less. > However I think this much is not sufficient to test generative models. Also > this > is applicable to RBM, DBN and DBM only. I haven't yet gone through GAN (I am > planning on finishing it in next couple of days). So I don't have any concrete > strategy for testing these models. I am searching for appropriate testing > methods. May be you can help me for this? For each model there are a couple of tests that I can think of: RBM: - Train the model on a subset of the MNIST dataset and make sure the trained filters lead to Gabor-like filters. - Reconstruct the negative samples obtained as image and check its correlation with the input. Take a look at: http://deeplearning.net/tutorial/rbm.html#tracking-progress <http://deeplearning.net/tutorial/rbm.html#tracking-progress> for more information about how to test RBM's. DBN: - Could be tested on the validation error with respect to the classification of the MNIST dataset, by stacking a logistic regression layer at the top layer. - Compare results with other existing implementations. RBFN: - Test on a subset of the MNIST dataset. - Test on the NETtalk task. - Compare results with other existing implementations. I'm sure there are a couple more simple tests that we can come up with. Also, it looks like I missed the previous message, I'm not sure I see a benefit of implementing the Hopfield model, correct me if I'm wrong but isn't the RBM similar and the superior/sucessor model? I hope this is helpful. Thanks, Marcus > On 15 Mar 2017, at 06:38, Prasanna Patil <[email protected]> wrote: > > Hi Marcus, > > I was going through application guideline given at mlpack wiki page and I > came across testing the project section. I never actually thought about > testing up until now but now that I think about it, it is hard to test > generative models. One direct way to do it is give it a noisy input image and > perform CD to get sample from model distribution and make sure that > reconstruction error is less. However I think this much is not sufficient to > test generative models. Also this is applicable to RBM, DBN and DBM only. I > haven't yet gone through GAN (I am planning on finishing it in next couple of > days). So I don't have any concrete strategy for testing these models. I am > searching for appropriate testing methods. May be you can help me for this? > > I know you are very busy and ignore this email if you don't have time for it. > > Thanks, > Prasanna > > On Wed, Mar 1, 2017 at 7:25 AM, Prasanna Patil <[email protected] > <mailto:[email protected]>> wrote: > Hi Marcus, > > I guess, if you find some model you think is interesting and is somewhat > manageable to implement, you can do that. We are always open for new > interesting > models/methods. That's also a great way to work with the codebase, but don't > feel obligated. > > I was thinking of implementing Hopfield network if it is not part of GSoC > project (?). > > I have implemented it in Python here > <https://github.com/prasanna08/MachineLearning/blob/master/hopfield.py>. So > can you help me with the interface of hopfield model (functions that should > be visible to user. Such as get_output(input will be corrupted image), > compute_energy(energy associated with particular input state) etc...) > > The version I have implemented is quite basic and works good for binary > images. I have not tried it with grayscale images (my implementation does use > tanh outputs rather than binary outputs, however). I will go though this > paper <http://page.mi.fu-berlin.de/rojas/neural/chapter/K13.pdf> you > mentioned for more details. Also I am having exams this week so I will > implement this in next week, if that's okay ? > > Initially I was thinking of implementing batch normalization layer. But I > found that in mlpack only one input train case is processed per iteration (am > I correct?) . Batch norm depends on minibatch of inputs so I don't know how > to do it with current scenario. > > Thanks, > Prasanna >
_______________________________________________ mlpack mailing list [email protected] http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack
