ilkhem / icebeem Goto Github PK

Hi,

I use the simulations.py for running the simulated experiment using imca dataset but I am not able to get the plot of the true v/s latent dimension using the arguments to be passed.
Could you help shed light on the way to get the plot of true v/s latent dimensions.

--Prashant

Thanks for the excellent work. I just wanted to know if you have worked on some popular datasets as well, such as 3dshapes, smallnorb and cars3d. If so, that could be helpful to merge the benchmark with disentanglement_lib.

Besides, as the conditional case requires the factor label to train the model, how could that still be "unsupervised"?

@ilkhem @piomonti

1. In fce.py, the value for the noise segment label is set as 1/number_of_segments:
   self.contrastSegments = (np.ones(self.segments.shape) / self.segments.shape[1]).astype(np.float32)
   I could not find an explanation for this choice in the paper. Could you clarify?
   Also, would you have a different recommendation if the number_of_observations were different for each segment?

2. In compute_ebm_logpdf, the augmented evaluation of f(x).T * g(y) does not seem to work for noise samples. My understanding is that here both f(x) and g(y) are augmented with f^2 and g^2. If that is the case, then g^2 = (W*y)^2, where * means matrix multiplication, squaring is element-wise and W = self.ebm_finalLayer. When y = one-hot encoded labels of data, g^2 = W[:,k]^2 where one-hot selects the k-th col of W which is then squared. This is equivalent to the implementation in line 63: self.ebm_finalLayer * self.ebm_finalLayer (this is just element-wise squaring of W) followed by * seg in line 66 for column selection (roughly speaking).
   The problem is with noise samples, for which y = self.contrastSegments instead of one-hot encoded labels.
   Then g^2 = mean(W,1)^2, the mean of W cols... and the MEANS (not the elements of W) are squared.
   So the current implementation in line 63: self.ebm_finalLayer * self.ebm_finalLayer does not seem to work for noise samples.
   Could you provide a clarification on this?

Thank you for your time and for the outstanding work!

Rogers

Great repo!
My name is Yerzhan, I am a masters student at Skoltech, Moscow. I am doing research at a neuroimaging lab here, and we are looking for the ways to perform rs-fMRI component analysis using deep learning methods.

Can you please comment on application of iVAE and TCL to fMRI component analysis. It is briefly mentioned in the paper, yet I was wondering if you can elaborate more on results that you obtained. How good they were, and perhaps there are some reproducible code of model being trained on HCP data. That would be of tremendous help.

Thank you!

It seems this part was not mentioned in the AISTATS'20 paper, is this trick important in terms of performance? could you briefly clarify what this part does?