15:17 < kris__> lozhnikov: I made most of the changes that you asked for ResizeLayer.

15:18 < kris__> But i can't get the backward pass test right.

15:18 < kris__> Though by visual inspection of image from the backward pass it looks very close to the input.

15:30 < kris__> zoq: Regarding the ssRBM could you explain your comments a little bit more. I asked a doubt on the github if you could explain it. That would be great.

15:40 < lozhnikov> kris__: It seems you misunderstood my suggestion.

15:40 < lozhnikov> I suggested to fill the matrix using a linear function e.g. f(x) = ax + by

15:40 < lozhnikov> i.e. matrix(i, j) = a * i / nx + b * j / ny,

15:40 < lozhnikov> where (nx, ny) is the shape of the matrix

15:40 < lozhnikov> moreover, you didn't take into account the computation accuracy

15:41 < lozhnikov> the precision of double is equal to 1e-16 around 1

17:01 < kris__> Okay so the test you are suggesting is scale(matrix(i,j) and compare it with what?

17:50 < lozhnikov> kris__: with the same linear function

17:52 < lozhnikov> i.e. output(i, j) should be equal to a * i / outNx + b * j / outNy, where (outNx, outNy) is the shape of the output matrix

17:57 < kris__> why i / outNx and not i % outNx

17:59 < kris__> okay i think i get it.

18:01 < kris__> Let me first test out the present implementation with cnn architecture and then i would update the test.

18:10 < kris__> Hi, lozhnikov i have a concern can you look at this..https://arxiv.org/pdf/1701.07875.pdf

18:11 < kris__> page 13 figure 6

18:11 < kris__> The paper states that without the batch normalisation the results of the present gan are pretty poor.

18:26 < kris__> Should i implement the BatchNormalisation before moving ahead.

18:32 < lozhnikov> That's quite interesting, but in that case we have to do the same tests again. I think you haven't got enough time for that (until the final evaluation). I suggest to finish the oreilly test first since we spent a lot of time to prepare for this test. Anyway, you can do that after GSoC

18:33 < kris__> Okay sure..... i am done with orilley implementation. I will run the test and see the results....:)

19:43 < kris__> lozhnikov is this test okay for the resize layer.

19:43 < kris__> https://gist.github.com/kris-singh/b331c6815beb1e762b030d5dafbac3bf

19:51 < lozhnikov> kris__: As for me it is better to define all constants in the beginning

19:52 < lozhnikov> I mean the shapes

19:52 < kris__> Hmmm sure.... i am still not getting the backward pass correctly though..

19:56 < lozhnikov> yeah, I replied at github. Looks like the slide is incorrect. Different scales look weird.

23:01 < kris__> lozhnikov: I have got the test working now....https://gist.github.com/kris-singh/4b355418edd9c69ede11c4af18086438

23:02 < kris__> For the orilley example. The problem is the Forward pass itself is taking a lot of time. I see that it is stuck in valid convolution function.

23:03 < kris__> Can you look at the code once ..... i am not very comfortable with convolution code so i am not sure if the code is correct.

23:39 < kris__> zoq: Could you have a look at the above code.

23:40 < kris__> Is the architecture defined correctly?