Get logp from SGMCMC

[rzu512]

The HMC class allows accessing the log-joint probability before and after the update step through the log_prob and orig_log_prob attributes of the HMCInfo class. However, the stochastic gradient MCMC classes in zhusuan don't expose these two log-joint probabilities.

Is it possible to get a tensorflow operation for the log-joint probability at the end of the update step in SGMCMC? Or alternatively Is it possible to get a tensorflow operation for the value of the parameters before the update step?

[csy530216]

Hello, I think it is not difficult to get the value of the parameters as long as you perform something like sess.run(latent). About the log-joint probability, since in SGMCMC algorithm we do not need to compute it (unlike in HMC, since we need to perform the Metropolis-Hastings step there), by default the log_prob is not returned (to reduce the size of computation graph).

If you want to get log_prob given the value of parameters stored in a dictionary latent, you could add some codes in your script as follows:

bn = model.observe(**merge_dicts(latent, observed))
log_prob = bn.log_joint()

[rzu512]

In the end, I modify my joint_prob function and the SGHMC class to store the logp and parameters at the half step. For my case, I only need to do one additional sum to get the logp after calculating the gradient. So it make sense to calculate the gradient and logp at the same time t.

I use SGHMC to fit 1000 chains with the approximate logp and gradient. For every n steps, I calculate the accurate logp for the approximately best chain.