gibbs sampling and contrastive divergence

[Ahmed Mazari]

Hello,

l want to have 1000 configurations at the end of the process of gibbs sampling rather than one configuration. How can l do that properly

Thank you

[Michael Krabbe Borregaard]

What function are you using?

--
You received this message because you are subscribed to the Google Groups "julia-stats" group.
To unsubscribe from this group and stop receiving emails from it, send an email to julia-stats+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Ahmed Mazari]

function gibbs(rbm::HTRBM,bp::BPTree,vis; n_times=1)
  v_pos = vis
  h_pos = sample_hid(rbm,bp,v_pos)


  v_neg = sample_vis(rbm,hid)
  h_neg = sample_hid(rbm,bp,v_neg)

  for i=1:n_times-1
    v_neg = sample_vis(rbm,h_neg)
    h_neg = sample_hid(rbm,bp,v_neg)
  end

  return v_pos,h_pos,v_neg,h_neg
end

CD : contrastive divergence

function CD(rbm::HTRBM,bp::BPTree,vis; n_times=1)
  v_pos = vis
 
  h_pos,h_pos_cov,h_samples = get_hid(rbm,bp,v_pos)

  # CD1
  v_neg = sample_vis(rbm,h_samples)
 
  h_neg,h_neg_cov,h_samples = get_hid(rbm,bp,v_neg)

  # CDn
  for i=1:n_times-1
    v_neg = sample_vis(rbm,h_samples)
    h_neg,h_neg_cov,h_samples = get_hid(rbm,v_neg)
  end


  return v_pos,h_pos,h_pos_cov,v_neg,h_neg,h_neg_cov
end

On Friday, August 19, 2016 at 4:30:50 PM UTC+2, Michael Borregaard wrote:

What function are you using?

On Fri, Aug 19, 2016 at 4:23 PM, Ahmed Mazari <ahmedma...@gmail.com> wrote:

Hello,

l want to have 1000 configurations at the end of the process of gibbs sampling rather than one configuration. How can l do that properly

Thank you

--
You received this message because you are subscribed to the Google Groups "julia-stats" group.

To unsubscribe from this group and stop receiving emails from it, send an email to julia-stats...@googlegroups.com.

[Ahmed Mazari]

i have another question related to the topic. When we would like to draw a histogram for our gibbs samplaing what should be the  axis of abscissa and Ordinate ?

[Andrei Zh]

l want to have 1000 configurations at the end of the process of gibbs sampling rather than one configuration.

So just take 1000 samples, i.e. run `gibbs(...)` 1000 times, preferably with different input data points (`vis` parameter).

Note that the idea of contrastive divergence is to reduce number of Gibbs sampling iterations per datapoint. Normally, only 1 pass is used (i.e. CD-1).

 

i have another question related to the topic. When we would like to draw a histogram for our gibbs samplaing what should be the  axis of abscissa and Ordinate ?

I think the question is what you are trying to achieve. Gibbs sampling is used to produce samples from a highly multivariate distribution, so visualizing it using histograms doesn't sound very reasonable. Can you give some context of your work and reference to a library/project you are using if it's available.