Home | Trees | Indices | Help |
|
---|
|
|
|||
KeyedList |
|
|||
|
|||
|
|||
|
|||
|
|||
|
|||
|
|||
|
|||
|
|||
|
|||
|
|||
|
|||
|
|||
|
|||
|
|||
|
|||
|
|
|||
logger = logging.getLogger('Ensembles')
|
Imports: logging, copy, shutil, sys, time, scipy, KeyedList_mod, Utility, HAVE_PYPAR, my_rank, my_host, num_procs, pypar
|
Generate a Bayesian ensemble of parameter sets consistent with the data in the model. The sampling is done in terms of the bare parameters. Inputs: (All not listed here are identical to those of ensemble_log_params.) recalc_func --- Function used to calculate the hessian matrix. It should take only a parameters argument and return the matrix. If this is None, default is to use m.GetJandJtJ. |
Generate a Bayesian ensemble of parameter sets consistent with the data in the model. The sampling is done in terms of the logarithm of the parameters. Inputs: m -- Model to generate the ensemble for params -- Initial parameter KeyedList to start from hess -- Hessian of the model steps -- Maximum number of Monte Carlo steps to attempt max_run_hours -- Maximum number of hours to run temperature -- Temperature of the ensemble step_scale -- Additional scale applied to each step taken. step_scale < 1 results in steps shorter than those dictated by the quadratic approximation and may be useful if acceptance is low. sing_val_cutoff -- Truncate the quadratic approximation at eigenvalues smaller than this fraction of the largest. seeds -- A tuple of two integers to seed the random number generator recalc_hess_alg --- If True, the Monte-Carlo is done by recalculating the hessian matrix every timestep. This signficantly increases the computation requirements for each step, but it may be worth it if it improves convergence. recalc_func --- Function used to calculate the hessian matrix. It should take only a log parameters argument and return the matrix. If this is None, default is to use m.GetJandJtJInLogParameteters save_hours --- If save_to is not None, the ensemble will be saved to that file every 'save_hours' hours. save_to --- Filename to save ensemble to. skip_elems --- If non-zero, skip_elems are skipped between each included step in the returned ensemble. For example, skip_elems=1 will return every other member. Using this option can reduce memory consumption. Outputs: ens, ens_fes, ratio ens -- List of KeyedList parameter sets in the ensemble ens_fes -- List of free energies for each parameter set ratio -- Fraction of attempted moves that were accepted The sampling is done by Markov Chain Monte Carlo, with a Metropolis-Hasting update scheme. The canidate-generating density is a gaussian centered on the current point, with axes determined by the hessian. For a useful introduction see: Chib and Greenberg. "Understanding the Metropolis-Hastings Algorithm" _The_American_Statistician_ 49(4), 327-335 |
The cost from the quadratic approximation of a trialMove, given the hessian. (Note: the hessian here is assumed to be the second derivative matrix of the cost, without an additional factor of 1/2.) |
Return the mean and standard deviation trajectory objects for the given input list of trajectories. (All must be evaluated at the same points.) |
Return the Principle Component Analysis eigenvalues and eigenvectors (in log parameters) of an ensemble. (This function takes the logs for you.) |
Home | Trees | Indices | Help |
|
---|
Generated by Epydoc 3.0.1 on Wed Jun 3 10:15:00 2009 | http://epydoc.sourceforge.net |