Improve ptemcee

This is a total rewrite of the ptemcee implementation. Originally, we had based ptemcee on emcee, but emcee has changed quite significantly and I anticipate it will be easier to maintain the two uncoupled from hereon. This adds sufficient functionality to ptemcee for it to be usable on gravitational wave events. Here is a summary of the changes

1. Add automatic convergence to the user interface: users now specify nsamples as a the primary input. The sampler will run until it has nsamples independent samples from the posterior. The user now has access to a reduced set of the ptemcee full functionality (e.g. iterations), but this isn't required given the convergence checking.
2. Add diagnostic plots for the autocorrelation time, e.g.:
3. Improve the traceplots to show the convergence and the current posterior samples
4. Add dumping of posterior samples to a .dat file
5. Add multi-processing support, tested on a 16 core machine to give a significant reduction in compute time (exact timings yet to be calculated).
6. Add regular 10-minute checkpointing

Note, I have intentionally put much of the primary logic into functions rather than class methods. This is because eventually we should port changes to parallel bilby and potentially other samplers and this is most easily done from a functional interface (at least for parallel bilby where inheritance causes issues with the parallel aspects).

bilby/core/sampler/ptemcee.py

@@ -24,36 +24,124 @@ class Ptemcee(MCMCSampler):
     documentation for that class for further help. Under Other Parameters, we
     list commonly used kwargs and the bilby defaults.
 
+    Parameters
+    ----------
+    nsamples: int, (5000)
+        The requested number of samples. Note, in cases where the
+        autocorrelation parameter is difficult to measure, it is possible to
+        end up with more than nsamples.
+    burn_in_act, thin_by_nact: int, (50, 1)
+        The number of burn-in autocorrelation times to discard and the thin-by
+        factor. Increasing burn_in_act increases the time required for burn-in.
+        Increasing thin_by_nact increases the time required to obtain nsamples.
+    autocorr_tol: int, (50)
+        The minimum number of autocorrelation times needed to trust the
+        estimate of the autocorrelation time.
+    autocorr_c: int, (5)
+        The step size for the window search used by emcee.autocorr.integrated_time
+    safety: int, (1)
+        A multiplicitive factor for the estimated autocorrelation. Useful for
+        cases where non-convergence can be observed by eye but the automated
+        tools are failing.
+    autocorr_tau:
+        The number of autocorrelation times to use in assessing if the
+        autocorrelation time is stable.
+    frac_threshold: float, (0.01)
+        The maximum fractional change in the autocorrelation for the last
+        autocorr_tau steps. If the fractional change exceeds this value,
+        sampling will continue until the estimate of the autocorrelation time
+        can be trusted.
+    min_tau: int, (1)
+        A minimum tau (autocorrelation time) to accept.
+    check_point_deltaT: float, (600)
+        The period with which to checkpoint (in seconds).
+    threads: int, (1)
+        If threads > 1, a MultiPool object is setup and used.
+    exit_code: int, (77)
+        The code on which the sampler exits.
+    store_walkers: bool (False)
+        If true, store the unthinned, unburnt chaines in the result. Note, this
+        is not recommended for cases where tau is large.
+    ignore_keys_for_tau: str
+        A pattern used to ignore keys in estimating the autocorrelation time.
+    pos0: str, list ("prior")
+        If a string, one of "prior" or "minimize". For "prior", the initial
+        positions of the sampler are drawn from the sampler. If "minimize",
+        a scipy.optimize step is applied to all parameters a number of times.
+        The walkers are then initialized from the range of values obtained.
+        If a list, for the keys in the list the optimization step is applied,
+        otherwise the initial points are drawn from the prior.
+
+
     Other Parameters
     ----------------
-    nwalkers: int, (100)
+    nwalkers: int, (200)
         The number of walkers
     nsteps: int, (100)
         The number of steps to take
-    nburn: int (50)
-        The fixed number of steps to discard as burn-in
     ntemps: int (2)
         The number of temperatures used by ptemcee
+    Tmax: float
+        The maximum temperature
 
     """
+
     # Arguments used by ptemcee
     default_kwargs = dict(
-        ntemps=20, nwalkers=200, Tmax=None, betas=None,
-        a=2.0, loglargs=[], logpargs=[], loglkwargs={},
-        logpkwargs={}, adaptation_lag=10000, adaptation_time=100, random=None,
-        adapt=False, swap_ratios=False)
-
-    def __init__(self, likelihood, priors, outdir='outdir', label='label',
-                 use_ratio=False, check_point_plot=True, skip_import_verification=False,
-                 resume=True, nsamples=5000, burn_in_nact=50, thin_by_nact=1,
-                 autocorr_c=5, safety=1, frac_threshold=0.01,
-                 autocorr_tol=50, autocorr_tau=5, min_tau=1, check_point_deltaT=600,
-                 threads=1, exit_code=77, plot=False, store_walkers=False,
-                 ignore_keys_for_tau="recalib", pos0="prior", **kwargs):
+        ntemps=20,
+        nwalkers=200,
+        Tmax=None,
+        betas=None,
+        a=2.0,
+        loglargs=[],
+        logpargs=[],
+        loglkwargs={},
+        logpkwargs={},
+        adaptation_lag=10000,
+        adaptation_time=100,
+        random=None,
+        adapt=False,
+        swap_ratios=False,
+    )
+
+    def __init__(
+        self,
+        likelihood,
+        priors,
+        outdir="outdir",
+        label="label",
+        use_ratio=False,
+        check_point_plot=True,
+        skip_import_verification=False,
+        resume=True,
+        nsamples=5000,
+        burn_in_nact=50,
+        thin_by_nact=1,
+        autocorr_tol=50,
+        autocorr_c=5,
+        safety=1,
+        autocorr_tau=5,
+        frac_threshold=0.01,
+        min_tau=1,
+        check_point_deltaT=600,
+        threads=1,
+        exit_code=77,
+        plot=False,
+        store_walkers=False,
+        ignore_keys_for_tau=None,
+        pos0="prior",
+        **kwargs
+    ):
         super(Ptemcee, self).__init__(
-            likelihood=likelihood, priors=priors, outdir=outdir,
-            label=label, use_ratio=use_ratio, plot=plot,
-            skip_import_verification=skip_import_verification, **kwargs)
+            likelihood=likelihood,
+            priors=priors,
+            outdir=outdir,
+            label=label,
+            use_ratio=use_ratio,
+            plot=plot,
+            skip_import_verification=skip_import_verification,
+            **kwargs
+        )
 
         signal.signal(signal.SIGTERM, self.write_current_state_and_exit)
         signal.signal(signal.SIGINT, self.write_current_state_and_exit)
@@ -77,26 +165,34 @@ class Ptemcee(MCMCSampler):
         self.pos0 = pos0
 
         self.check_point_plot = check_point_plot
-        self.resume_file = "{}/{}_checkpoint_resume.pickle".format(self.outdir, self.label)
+        self.resume_file = "{}/{}_checkpoint_resume.pickle".format(
+            self.outdir, self.label
+        )
         self.exit_code = exit_code
 
     @property
     def sampler_function_kwargs(self):
-        keys = ['adapt', 'swap_ratios']
+        keys = ["adapt", "swap_ratios"]
         return {key: self.kwargs[key] for key in keys}
 
     @property
     def sampler_init_kwargs(self):
-        return {key: value
-                for key, value in self.kwargs.items()
-                if key not in self.sampler_function_kwargs}
+        return {
+            key: value
+            for key, value in self.kwargs.items()
+            if key not in self.sampler_function_kwargs
+        }
 
     def get_pos0_from_prior(self):
         """ for ptemcee, the pos0 has the shape ntemps, nwalkers, ndim """
         logger.info("Generating pos0 samples")
-        return [[self.get_random_draw_from_prior()
-                 for _ in range(self.sampler_init_kwargs["nwalkers"])]
-                for _ in range(self.kwargs['ntemps'])]
+        return [
+            [
+                self.get_random_draw_from_prior()
+                for _ in range(self.sampler_init_kwargs["nwalkers"])
+            ]
+            for _ in range(self.kwargs["ntemps"])
+        ]
 
     def get_pos0_from_minimize(self, minimize_list=None):
         logger.info("Attempting to set pos0 from minimize")
@@ -112,13 +208,17 @@ class Ptemcee(MCMCSampler):
 
         def neg_log_like(params):
             likelihood_copy.parameters.update(
-                {key: val for key, val in zip(minimize_list, params)})
+                {key: val for key, val in zip(minimize_list, params)}
+            )
             try:
                 return -likelihood_copy.log_likelihood()
             except RuntimeError:
                 return +np.inf
-        bounds = [(self.priors[key].minimum, self.priors[key].maximum)
-                  for key in minimize_list]
+
+        bounds = [
+            (self.priors[key].minimum, self.priors[key].maximum)
+            for key in minimize_list
+        ]
         trials = 0
         success = []
         while True:
@@ -126,7 +226,8 @@ class Ptemcee(MCMCSampler):
             likelihood_copy.parameters.update(draw)
             x0 = [draw[key] for key in minimize_list]
             res = minimize(
-                neg_log_like, x0, bounds=bounds, method='L-BFGS-B', tol=1e-15)
+                neg_log_like, x0, bounds=bounds, method="L-BFGS-B", tol=1e-15
+            )
             if res.success:
                 success.append(res.x)
             if trials > 100:
@@ -138,16 +239,20 @@ class Ptemcee(MCMCSampler):
         for i, key in enumerate(minimize_list):
             pos0_min = np.min(success[:, i])
             pos0_max = np.max(success[:, i])
-            logger.info("Initialize {} walkers from {}->{}"
-                        .format(key, pos0_min, pos0_max))
+            logger.info(
+                "Initialize {} walkers from {}->{}".format(key, pos0_min, pos0_max)
+            )
             j = self.search_parameter_keys.index(key)
             pos0[:, :, j] = np.random.uniform(
-                pos0_min, pos0_max,
-                size=(self.kwargs["ntemps"], self.kwargs["nwalkers"]))
+                pos0_min,
+                pos0_max,
+                size=(self.kwargs["ntemps"], self.kwargs["nwalkers"]),
+            )
         return pos0
 
     def setup_sampler(self):
         import ptemcee
+
         if os.path.isfile(self.resume_file) and self.resume is True:
             logger.info("Resume data {} found".format(self.resume_file))
             with open(self.resume_file, "rb") as file:
@@ -163,21 +268,32 @@ class Ptemcee(MCMCSampler):
 
             pos0 = None
 
-            logger.info("Resuming from previous run with time={}".format(self.sampler.time))
+            logger.info(
+                "Resuming from previous run with time={}".format(self.sampler.time)
+            )
 
         else:
             # Initialize the PTSampler
             if self.threads == 1:
                 self.sampler = ptemcee.Sampler(
-                    dim=self.ndim, logl=self.log_likelihood, logp=self.log_prior,
-                    **self.sampler_init_kwargs)
+                    dim=self.ndim,
+                    logl=self.log_likelihood,
+                    logp=self.log_prior,
+                    **self.sampler_init_kwargs
+                )
             else:
                 self.sampler = ptemcee.Sampler(
-                    dim=self.ndim, logl=do_nothing_function, logp=do_nothing_function,
-                    pool=self.pool, threads=self.threads, **self.sampler_init_kwargs)
+                    dim=self.ndim,
+                    logl=do_nothing_function,
+                    logp=do_nothing_function,
+                    pool=self.pool,
+                    threads=self.threads,
+                    **self.sampler_init_kwargs
+                )
 
                 self.sampler._likeprior = LikePriorEvaluator(
-                    self.search_parameter_keys, use_ratio=self.use_ratio)
+                    self.search_parameter_keys, use_ratio=self.use_ratio
+                )
 
             # Set up empty lists
             self.tau_list = []
@@ -202,11 +318,11 @@ class Ptemcee(MCMCSampler):
     def setup_pool(self):
         if self.threads > 1:
             import schwimmbad
+
             logger.info("Creating MultiPool with {} processes".format(self.threads))
             self.pool = schwimmbad.MultiPool(
-                self.threads,
-                initializer=init,
-                initargs=(self.likelihood, self.priors))
+                self.threads, initializer=init, initargs=(self.likelihood, self.priors)
+            )
         else:
             self.pool = None
 
@@ -219,6 +335,7 @@ class Ptemcee(MCMCSampler):
 
     def run_sampler_internal(self):
         import emcee
+
         sampler, pos0 = self.setup_sampler()
 
         t0 = datetime.datetime.now()
@@ -235,27 +352,32 @@ class Ptemcee(MCMCSampler):
             samples = sampler.chain[0, :, : sampler.time, :]
             taus = []
             for ii in range(sampler.nwalkers):
+                tau_ii = []
                 for jj, key in enumerate(self.search_parameter_keys):
                     if self.ignore_keys_for_tau and self.ignore_keys_for_tau in key:
                         continue
                     try:
-                        taus.append(
+                        tau_ii.append(
                             emcee.autocorr.integrated_time(
                                 samples[ii, :, jj], c=self.autocorr_c, tol=0
                             )[0]
                         )
                     except emcee.autocorr.AutocorrError:
                         taus.append(np.inf)
+                taus.append(tau_ii)
+
+            tau = np.max(np.mean(taus, axis=0))
 
             # Apply multiplicitive safety factor
-            tau = self.safety * np.mean(taus)
+            tau = self.safety * tau
 
             # Store for convergence checking and plotting
-            self.tau_list.append(tau)
+            self.tau_list.append(np.mean(taus, axis=0))
             self.tau_list_n.append(sampler.time)
 
             # Convert to an integer
             tau_int = int(np.ceil(tau)) if not np.isnan(tau) else tau
+            self.tau_int = tau_int
 
             if np.isnan(tau_int) or np.isinf(tau_int):
                 print_progress(
@@ -266,24 +388,30 @@ class Ptemcee(MCMCSampler):
                     np.nan,
                     np.nan,
                     [np.nan],
-                    False)
+                    False,
+                )
                 continue
 
             # Calculate the effective number of samples available
             self.nburn = int(self.burn_in_nact * tau_int)
             self.thin = int(np.max([1, self.thin_by_nact * tau_int]))
             samples_per_check = sampler.nwalkers / self.thin
-            self.nsamples_effective = int(sampler.nwalkers * (sampler.time - self.nburn) / self.thin)
+            self.nsamples_effective = int(
+                sampler.nwalkers * (sampler.time - self.nburn) / self.thin
+            )
 
             # Calculate convergence boolean
             converged = self.nsamples < self.nsamples_effective
 
             # Calculate fractional change in tau from previous iterations
             check_taus = np.array(self.tau_list[-tau_int * self.autocorr_tau:])
+            taus_per_parameter = np.array(taus)[-1, :]
             if not np.any(np.isnan(check_taus)):
-                frac = (tau - check_taus) / tau
+                frac = (taus_per_parameter - check_taus) / taus_per_parameter
+                max_frac = np.max(frac)
                 tau_usable = np.all(frac < self.frac_threshold)
             else:
+                max_frac = np.nan
                 tau_usable = False
 
             if sampler.time < tau_int * self.autocorr_tol or tau_int < self.min_tau:
@@ -297,7 +425,7 @@ class Ptemcee(MCMCSampler):
                 self.nsamples_effective,
                 samples_per_check,
                 tau_int,
-                check_taus,
+                max_frac,
                 tau_usable,
             )
 
@@ -319,10 +447,11 @@ class Ptemcee(MCMCSampler):
 
         # Get 0-likelihood samples and store in the result
         samples = sampler.chain[0, :, :, :]  # nwalkers, nsteps, ndim
-        self.result.samples = (
-            samples[:, self.nburn: sampler.time:self.thin, :].reshape((-1, self.ndim)))
+        self.result.samples = samples[
+            :, self.nburn : sampler.time : self.thin, :
+        ].reshape((-1, self.ndim))
         loglikelihood = sampler.loglikelihood[
-            0, :, self.nburn:sampler.time:self.thin
+            0, :, self.nburn : sampler.time : self.thin
         ]  # nwalkers, nsteps
         self.result.log_likelihood_evaluations = loglikelihood.reshape((-1))
 
@@ -336,35 +465,57 @@ class Ptemcee(MCMCSampler):
         self.result.log_evidence = log_evidence
         self.result.log_evidence_err = log_evidence_err
 
-        self.result.sampling_time = datetime.timedelta(seconds=np.sum(self.time_per_check))
+        self.result.sampling_time = datetime.timedelta(
+            seconds=np.sum(self.time_per_check)
+        )
 
         return self.result
 
     def write_current_state_and_exit(self, signum=None, frame=None):
         logger.warning("Run terminated with signal {}".format(signum))
-        if getattr(self, 'pool', None):
+        if getattr(self, "pool", None) or self.threads == 1:
             self.write_current_state(plot=False)
-            logger.warning("Closing pool")
+        if getattr(self, "pool", None):
+            logger.info("Closing pool")
             self.pool.close()
         sys.exit(self.exit_code)
 
     def write_current_state(self, plot=True):
-        checkpoint(self.outdir, self.label, self.nsamples_effective,
-                   self.sampler, self.nburn, self.thin,
-                   self.search_parameter_keys, self.resume_file, self.tau_list,
-                   self.tau_list_n, self.time_per_check)
+        checkpoint(
+            self.outdir,
+            self.label,
+            self.nsamples_effective,
+            self.sampler,
+            self.nburn,
+            self.thin,
+            self.search_parameter_keys,
+            self.resume_file,
+            self.tau_list,
+            self.tau_list_n,
+            self.time_per_check,
+        )
 
         if plot and not np.isnan(self.nburn):
             # Generate the walkers plot diagnostic
             plot_walkers(
                 self.sampler.chain[0, :, : self.sampler.time, :],
-                self.nburn, self.thin, self.search_parameter_keys, self.outdir,
-                self.label
+                self.nburn,
+                self.thin,
+                self.search_parameter_keys,
+                self.outdir,
+                self.label,
             )
 
             # Generate the tau plot diagnostic
-            plot_tau(self.tau_list_n, self.tau_list, self.outdir, self.label,
-                     self.autocorr_tau)
+            plot_tau(
+                self.tau_list_n,
+                self.tau_list,
+                self.search_parameter_keys,
+                self.outdir,
+                self.label,
+                self.tau_int,
+                self.autocorr_tau,
+            )
 
 
 def print_progress(
@@ -374,7 +525,7 @@ def print_progress(
     nsamples_effective,
     samples_per_check,
     tau_int,
-    tau_list,
+    max_frac,
     tau_usable,
 ):
     # Setup acceptance string
@@ -388,9 +539,7 @@ def print_progress(
     )
 
     ave_time_per_check = np.mean(time_per_check[-3:])
-    time_left = (
-        (nsamples - nsamples_effective) * ave_time_per_check / samples_per_check
-    )
+    time_left = (nsamples - nsamples_effective) * ave_time_per_check / samples_per_check
     if time_left > 0:
         time_left = str(datetime.timedelta(seconds=int(time_left)))
     else:
@@ -398,7 +547,10 @@ def print_progress(
 
     sampling_time = datetime.timedelta(seconds=np.sum(time_per_check))
 
-    tau_str = "{}:{:0.1f}->{:0.1f}".format(tau_int, np.min(tau_list), np.max(tau_list))
+    if max_frac >= 0:
+        tau_str = "{}(+{:0.1f})".format(tau_int, max_frac)
+    else:
+        tau_str = "{}({:0.1f})".format(tau_int, max_frac)
     if tau_usable:
         tau_str = "={}".format(tau_str)
     else:
@@ -427,16 +579,26 @@ def print_progress(
     )
 
 
-def checkpoint(outdir, label, nsamples_effective, sampler, nburn, thin,
-               search_parameter_keys, resume_file, tau_list, tau_list_n,
-               time_per_check):
+def checkpoint(
+    outdir,
+    label,
+    nsamples_effective,
+    sampler,
+    nburn,
+    thin,
+    search_parameter_keys,
+    resume_file,
+    tau_list,
+    tau_list_n,
+    time_per_check,
+):
     logger.info("Writing checkpoint and diagnostics")
     ndim = sampler.dim
 
     # Store the samples if possible
     if nsamples_effective > 0:
         filename = "{}/{}_samples.txt".format(outdir, label)
-        samples = sampler.chain[0, :, nburn:sampler.time:thin, :].reshape(
+        samples = sampler.chain[0, :, nburn : sampler.time : thin, :].reshape(
             (-1, ndim)
         )
         df = pd.DataFrame(samples, columns=search_parameter_keys)
@@ -451,8 +613,11 @@ def checkpoint(outdir, label, nsamples_effective, sampler, nburn, thin,
     sampler_copy._beta_history = sampler._beta_history[:, : sampler.time]
 
     data = dict(
-        sampler=sampler_copy, tau_list=tau_list, tau_list_n=tau_list_n,
-        time_per_check=time_per_check)
+        sampler=sampler_copy,
+        tau_list=tau_list,
+        tau_list_n=tau_list_n,
+        time_per_check=time_per_check,
+    )
 
     with open(resume_file, "wb") as file:
         dill.dump(data, file, protocol=4)
@@ -465,16 +630,24 @@ def plot_walkers(walkers, nburn, thin, parameter_labels, outdir, label):
     nwalkers, nsteps, ndim = walkers.shape
     idxs = np.arange(nsteps)
     fig, axes = plt.subplots(nrows=ndim, ncols=2, figsize=(8, 3 * ndim))
-    scatter_kwargs = dict(lw=0, marker="o", markersize=1)
+    scatter_kwargs = dict(lw=0, marker="o", markersize=1, alpha=0.05,)
     # Plot the burn-in
     for i, (ax, axh) in enumerate(axes):
         ax.plot(
-            idxs[: nburn + 1], walkers[:, : nburn + 1, i].T, color="C1", **scatter_kwargs
+            idxs[: nburn + 1],
+            walkers[:, : nburn + 1, i].T,
+            color="C1",
+            **scatter_kwargs
         )
 
     # Plot the thinned posterior samples
     for i, (ax, axh) in enumerate(axes):
-        ax.plot(idxs[nburn::thin], walkers[:, nburn::thin, i].T, color="C0", **scatter_kwargs)
+        ax.plot(
+            idxs[nburn::thin],
+            walkers[:, nburn::thin, i].T,
+            color="C0",
+            **scatter_kwargs
+        )
         axh.hist(walkers[:, nburn::thin, i].reshape((-1)), bins=50, alpha=0.8)
         axh.set_xlabel(parameter_labels[i])
         ax.set_ylabel(parameter_labels[i])
@@ -485,15 +658,14 @@ def plot_walkers(walkers, nburn, thin, parameter_labels, outdir, label):
     plt.close(fig)
 
 
-def plot_tau(tau_list_n, tau_list, outdir, label, autocorr_tau):
+def plot_tau(tau_list_n, tau_list, search_parameter_keys, outdir, label, tau, autocorr_tau):
     fig, ax = plt.subplots()
-    ax.plot(tau_list_n, tau_list, "-", color='C1')
-    check_tau_idx = -int(tau_list[-1] * autocorr_tau)
-    check_taus = tau_list[check_tau_idx:]
-    check_taus_n = tau_list_n[check_tau_idx:]
-    ax.plot(check_taus_n, check_taus, "-", color='C0')
+    for i, key in enumerate(search_parameter_keys):
+        ax.plot(tau_list_n, np.array(tau_list)[:, i], label=key)
+    ax.axvline(tau_list_n[-1] - tau * autocorr_tau)
     ax.set_xlabel("Iteration")
     ax.set_ylabel(r"$\langle \tau \rangle$")
+    ax.legend()
     fig.savefig("{}/{}_checkpoint_tau.png".format(outdir, label))
     plt.close(fig)
 
@@ -502,7 +674,7 @@ def compute_evidence(sampler, outdir, label, nburn, thin, make_plots=True):
     """ Computes the evidence using thermodynamic integration """
     betas = sampler.betas
     # We compute the evidence without the burnin samples, but we do not thin
-    lnlike = sampler.loglikelihood[:, :, nburn:sampler.time]
+    lnlike = sampler.loglikelihood[:, :, nburn : sampler.time]
     mean_lnlikes = np.mean(np.mean(lnlike, axis=1), axis=1)
 
     mean_lnlikes = mean_lnlikes[::-1]
@@ -590,14 +762,14 @@ class LikePriorEvaluator(object):
     def __call__(self, x):
         lp = self.logp(x)
         if np.isnan(lp):
-            raise ValueError('Prior function returned NaN.')
+            raise ValueError("Prior function returned NaN.")
 
-        if lp == float('-inf'):
+        if lp == float("-inf"):
             # Can't return -inf, since this messes with beta=0 behaviour.
             ll = 0
         else:
             ll = self.logl(x)
             if np.isnan(ll).any():
-                raise ValueError('Log likelihood function returned NaN.')
+                raise ValueError("Log likelihood function returned NaN.")
 
         return ll, lp