Bug fix for #412

[dustinvtran]

Consider the unit test in this pull request. It performs HMC or SGLD on Bayesian NN classification. It does not run (see #412); with this PR, it does.

In HMC, we assume the order of the dictionaries z_old and z_new align. They do not: to see this, run the test while adding the following above line 167 of hmc.py:

print(list(six.iterkeys(z_old)))
print(list(six.iterkeys(z_new)))

This alignment is necessary because in line 102, sample always uses the key order in new_sample. I resolved this by enforcing OrderedDicts.

In SGLD, we assume the order of the dictionaries old_sample and self.latent_vars align. They do not: to see this, run the test while adding the following below line 61 of sgld.py:

print(list(six.iterkeys(old_sample)))
print(list(six.iterkeys(self.latent_vars)))

This alignment is necessary because in the loop of line 70, we assume grad_log_joint (which
uses old_sample's order) has the same order as self.latent_vars. I resolved this by changing the looping variables.

todo's

• make unit test actually a unit test
• check if this is also a problem in other monte carlo methods