Hi everyone,
I am generating samples from a DBN which has been pretrained by stacking k
RBMs one on top of another, in an unsupervised way.
After pretraining, to generate the samples from the whole networks I think
I should do a foward pass, sampling each RBM hidden unit given the visibles
and then I should do a backward pass, sampling visible units given the
hidden values.
I start the chain either by feeding a random visible configuration or a
zero vector to the first RBM.
This is like one step of my Gibbs chain; I would then perform first several
burn-in steps and then actually take samples only after them.
Is this the correct way to generate N samples from a DBN?
Mimicking the code in DeepLearningTutorials/code/DBN.py, I would do
something like this:
def gibbs_vhv(self, v0_sample):
v_sample = v0_sample
for rbm in self.rbm_layers:
pre_sigmoid_h, h_mean, h_sample = rbm.sample_h_given_v(v_sample)
v_sample = h_sample
h_sample = v_sample
for rbm in reversed(self.rbm_layers):
pre_sigmoid_v, v_mean, v_sample = rbm.sample_v_given_h(h_sample)
h_sample = v_sample
return h_sample
My main issue is that in this case my samples are very similar one to
another. For instance, if I am training my DBN on MNIST the generated digit
samples are slightly variations of a same digit class.
Am I doing something not appropriate? How could I generate 'more different'
samples?
--
---
You received this message because you are subscribed to the Google Groups
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
For more options, visit https://groups.google.com/d/optout.
