"""
Sampling history for MCMC.
MCMC keeps track of a number of things during sampling.
The results may be queried as follows::
draws, generation, thinning, total_generations
sample(condition) returns draws, points, logp
gen_logp() returns genid, logp
acceptance_rate() returns draws, AR
traces() returns genid, chains
CR_weight() returns draws, CR_weight
best() returns best_x, best_logp
outliers() returns outliers
show()/save(file)/load(file)
Data is stored in circular arrays, which keeps the last N generations and
throws the rest away.
draws is the total number of draws from the sampler.
generation is the total number of generations. Due to tests for partially
full circular buffers throughout the code (state.generation < state.Ngen)
we are resetting generation to the size of the stored history on resume,
and setting the generation_offset to the start of the history. If you need the
number of generations across resume then use total_generations.
thinning is the number of generations per stored sample.
draws[i] is the number of draws including those required to produce the
information in the corresponding return vector. Note that draw numbers
need not be linearly spaced, since techniques like delayed rejection
will result in a varying number of samples per generation.
logp[i] is the set of log likelihoods, one for each member of the population.
The logp() method returns the complete set, and the sample() method returns
a thinned set, with on element of logp[i] for each vector point[i, :].
AR[i] is the acceptance rate at generation i, showing the proportion of
proposed points which are accepted into the population.
chains[i, :, :] is the set of points in the differential evolution population
at thinned generation i. Ideally, the thinning rate of the MCMC process
is chosen so that thinned generations i and i+1 are independent samples
from the posterior distribution, though there is a chance that this may
not be the case, and indeed, some points in generation i+1 may be identical
to those in generation i. Actual generation number is i*thinning.
points[i, :] is the ith point in a returned sample. The i is just a place
holder; there is no inherent ordering to the sample once they have been
extracted from the chains. Note that the sample may be from a marginal
distribution.
R[i] is the Gelman R statistic measuring convergence of the Markov chain.
CR_weight[i] is the set of weights used for selecting between the crossover
ratios available to the candidate generation process of differential
evolution. These will be fixed early in the sampling, even when adaptive
differential evolution is selected.
outliers[i] is a vector containing the thinned generation number at which
an outlier chain was removed, the id of the chain that was removed and
the id of the chain that replaced it. We leave it to the reader to decide
if the cloned samples, point[:generation, :, removed_id], should be included
in further analysis.
best_logp is the highest log likelihood observed during the analysis and
best_x is the corresponding point at which it was observed.
generation is the last generation number
"""
# TODO: state should be collected in files as we go
__all__ = ["MCMCDraw", "Draw", "dream_load", "h5load", "h5dump"]
import os.path
import re
import gzip
from typing import List, Dict, TextIO, Tuple, Union, Optional, Callable
from pathlib import Path
import warnings
from fnmatch import fnmatch
import numpy as np
from numpy import empty, sum, asarray, inf, argmax, hstack, dstack
from numpy import loadtxt, savetxt, reshape
from numpy.typing import NDArray
from .convergence import burn_point
from .outliers import identify_outliers
from .util import rng
from .gelman import gelman
# Don't load hdf5 if you don't need it
from typing import TYPE_CHECKING
if TYPE_CHECKING:
from h5py import Group
EXT = ".mc.gz"
CREATE = gzip.open
# EXT = ".mc"
# CREATE = open
SelectionType = Optional[Dict[Union[int, str], Tuple[float, float]]]
UNCERTAINTY_DTYPE = "d"
MAX_LABEL_LENGTH = 1024
LABEL_DTYPE = f"|S{MAX_LABEL_LENGTH}"
H5_COMPRESSION = 5
# TODO: Unused code. Using webview.server.cli.reload_export instead
[docs]
def dream_load(store):
"""
Load the saved DREAM state.
*store* is the path to the stored state, either as an HDF5 history file
or as a bumps export directory. If the directory contains multiple exports
then use the path to the .par file within the directory.
See also: h5load, load_state
"""
from pathlib import Path
import h5py
store = Path(store).expanduser().resolve()
modelname = None
if not store.exists():
raise RuntimeError(f"Path {store} does not exist.")
# First try the HDF reader since the user can save the session to any file.
if store.is_file():
h5 = None
try:
with h5py.File(store) as h5:
if "fitting/fit_state" in h5:
return h5load(h5["fitting/fit_state"])
else:
raise RuntimeError(f"fitting/fit_state not found in {store}")
except Exception:
if h5 is not None:
raise
# Fall through to the non-hdf5 loader
if store.is_dir():
# It's a directory: look for model name in store/*.par
pars = list(store.glob("*.par"))
if not pars:
raise RuntimeError("No .par file found in %s" % store)
if len(pars) > 1:
print(pars)
raise RuntimeError("Multiple models found in %s: %s" % (store, ", ".join(p.name for p in pars)))
modelname = pars[0].stem
else:
# It's a filename: parse store=path/model.par
if store.suffix != ".par":
raise RuntimeError(f"Expected model.par file, got {store}")
modelname = store.stem
store = store.parent
# Reload the MCMC data
basename = str(store / modelname)
state = load_state(basename)
return state
def _h5_write_field(group: "Group", field: str, data: Union[NDArray, str]):
import h5py
# print(f"h5 writing {field} in {group} with {data}")
if isinstance(data, str):
dtype = h5py.string_dtype(encoding="utf-8", length=len(data))
return group.create_dataset(field, data=data, dtype=dtype)
elif isinstance(data, (int, np.integer)):
return group.create_dataset(field, data=data, dtype=np.int64)
elif isinstance(data, (float, np.floating)):
return group.create_dataset(field, data=data, dtype=np.double)
elif np.isscalar(data):
return group.create_dataset(field, data=data, dtype=data.dtype)
else:
return group.create_dataset(field, data=data, dtype=data.dtype, compression=H5_COMPRESSION)
def _h5_read_field(group: "Group", field: str):
raw_data = group[field][()]
size = raw_data.size
if isinstance(raw_data, np.integer):
return int(raw_data)
elif isinstance(raw_data, np.floating):
return float(raw_data)
elif size is not None and size > 0:
return raw_data
else:
return None
[docs]
def h5dump(group: "Group", state: "MCMCDraw"):
class Fields:
total_generations = state.total_generations
thinning = state.thinning
gen_draws, gen_logp = state.logp(full=True)
_, AR = state.acceptance_rate()
thin_draws, thin_point, thin_logp = state.chains()
update_draws, update_CR_weight = state.CR_weight()
labels = np.array(state.labels, dtype=LABEL_DTYPE)
# These are marked outliers. They should still be marked when reloading the
# state, but they need to be cleared when updates are given. Make sure the
# request for the population from dream includes the outliers.
good_chains = None if isinstance(state._good_chains, slice) else state._good_chains
# In case the MCMC chains are stored as 32-bit, save the current population
# as 64-bit so that resume is consistent.
current_population, current_logp = state._last_gen()
best_x, best_logp = state.best()
best_gen = state._best_gen
portion = state.portion
# print(f"wrote {Fields.total_generations} generations")
# print(f"{state._gen_offset=} {state.Ngen=}")
# print("wrote", Fields.__dict__)
for field, value in Fields.__dict__.items():
if field[0] != "_" and value is not None:
try:
_h5_write_field(group, field, value)
except Exception as exc:
exc.args = (f"{exc.args[0]} while writing {field}",) + exc.args[1:]
raise
[docs]
def h5load(group: "Group"):
# print(f"dream.state.h5load {group}")
class Fields: ...
for field in group:
setattr(Fields, field, _h5_read_field(group, field))
# print("read", Fields.__dict__)
# Guess dimensions
Ngen = Fields.gen_draws.shape[0]
Nthin, Npop, Nvar = Fields.thin_point.shape
Nupdate, Ncr = Fields.update_CR_weight.shape
# Nthin -= skip
good_chains = getattr(Fields, "good_chains", None)
total_generations = getattr(Fields, "total_generations", Ngen)
thinning = getattr(Fields, "thinning", 1)
current_population = getattr(Fields, "current_population", None)
if current_population is None:
current_population = Fields.thin_point[-1].astype("d")
current_logp = getattr(Fields, "current_logp", None)
if current_logp is None:
current_logp = Fields.thin_logp[-1].astype("d")
best_x = getattr(Fields, "best_x", None)
best_logp = getattr(Fields, "best_logp", 0.0)
best_gen = getattr(Fields, "best_gen", 0)
portion = getattr(Fields, "portion", 1.0)
# Create empty draw and fill it with loaded data
state = MCMCDraw(0, 0, 0, 0, 0, 0, thinning)
state.portion = portion
state.draws = Ngen * Npop
state.labels = [label.decode() for label in Fields.labels]
state.generation = Ngen
# print(f"loading {total_generations=} {Ngen=}")
state._gen_offset = total_generations - Ngen
state._gen_index = 0
state._gen_draws = Fields.gen_draws.astype(np.int64)
state._gen_acceptance_rate = Fields.AR
state._gen_logp = Fields.gen_logp
state._thin_count = Nthin
state._thin_index = 0
state._thin_draws = Fields.thin_draws.astype(np.int64)
state._thin_logp = Fields.thin_logp
state._thin_point = Fields.thin_point
state._gen_current = current_population
state._gen_current_logp = current_logp
state._update_count = Nupdate
state._update_index = 0
state._update_draws = Fields.update_draws.astype(np.int64)
state._update_CR_weight = Fields.update_CR_weight
state._outliers = []
if best_x is not None:
state._best_x = best_x
state._best_logp = best_logp
state._best_gen = best_gen
else:
bestidx = np.unravel_index(np.argmax(Fields.thin_logp), Fields.thin_logp.shape)
state._best_logp = Fields.thin_logp[bestidx]
state._best_x = Fields.thin_point[bestidx]
# We are not multiplying the following by thinning because thinning and best_gen
# were added at the same time, so the reloaded thinning=1. Even if it were not
# one this could still be wrong since thinning may have changed on resume.
state._best_gen = bestidx[0]
state._good_chains = slice(None, None) if good_chains is None else good_chains.astype(int)
return state
def save_state(state: "MCMCDraw", filename: str):
# import sys; trace = sys.stdout
# trace = open(filename+"-trace.mc", "wt")
# trace.write("starting trace\n")
# Build 2-D data structures
# trace.write("extracting draws, logp\n")
draws, logp = state.logp(full=True)
# trace.write("extracting acceptance rate\n")
_, AR = state.acceptance_rate(portion=1)
# trace.write("building chain from draws, AR and logp\n")
chain = hstack((draws[:, None], AR[:, None], logp))
# trace.write("extracting point, logp\n")
_, point, logp = state.chains()
Nthin, Npop, Nvar = point.shape
# trace.write("shape is %d,%d,%d\n" % (Nthin, Npop, Nvar))
# trace.write("adding logp to point\n")
point = dstack((logp[:, :, None], point))
# trace.write("collapsing to draws x point\n")
point = reshape(point, (point.shape[0] * point.shape[1], point.shape[2]))
# trace.write("extracting CR_weight\n")
draws, CR_weight = state.CR_weight()
Nupdate, Ncr = CR_weight.shape
# trace.write("building stats\n")
stats = hstack((draws[:, None], CR_weight))
# TODO: missing _outliers from save_state
# Write convergence info
# trace.write("writing chain\n")
fid = CREATE(filename + "-chain" + EXT, "wb")
fid.write(f"# draws acceptance_rate {Npop}*logp\n".encode())
savetxt(fid, chain)
fid.close()
# Write point info
# trace.write("writing point\n")
fid = CREATE(filename + "-point" + EXT, "wb")
fid.write(f"# logp point (Nthin x Npop x Nvar = [{Nthin},{Npop},{Nvar}])\n".encode())
savetxt(fid, point)
fid.close()
# Write stats
# trace.write("writing stats\n")
fid = CREATE(filename + "-stats" + EXT, "wb")
fid.write(f"# draws {Ncr}*CR_weight\n".encode())
savetxt(fid, stats)
fid.close()
# trace.write("done state save\n")
# trace.close()
IND_PAT = re.compile("-1#IND")
INF_PAT = re.compile("1#INF")
def loadtxt_with_fallback(file: TextIO, report=0):
"""
Try to load the file with numpy.loadtxt, but if it fails then
fall back to loadtxt_MSVC, which is adapted for windows non-finite numbers.
"""
# record current position in file so that we can reset it if loadtxt fails
pos = file.tell()
try:
return loadtxt(file)
except Exception:
file.seek(pos)
return loadtxt_MSVC(file, report=report)
def loadtxt_MSVC(fh: TextIO, report=0):
"""
Like numpy loadtxt, but adapted for windows non-finite numbers.
"""
res = []
section = 0
lineno = 0
for line in fh:
lineno += 1
if report and lineno % report == 0:
print("read", section * report)
section += 1
IND_PAT.sub("nan", line)
INF_PAT.sub("inf", line)
line = line.split("#")[0].strip()
values = line.split()
if len(values) > 0:
try:
res.append([float(v) for v in values])
except ValueError:
print("Parse error:", values)
return asarray(res)
def path_contains_saved_state(filename: str) -> bool:
chain_file = filename + "-chain" + EXT
return os.path.exists(chain_file)
def openmc(filename: str) -> TextIO:
# If filename ends in .mc.gz, also check for a .mc file.
# If filename ends in .mc, also check for a .mc.gz file.
if filename.endswith(".gz"):
if os.path.exists(filename):
# print("opening with gzip")
fh = gzip.open(filename, "rt")
elif os.path.exists(filename[:-3]):
fh = open(filename[:-3], "rt")
else:
raise RuntimeError("file %s does not exist" % filename)
else:
if os.path.exists(filename):
fh = open(filename, "rt")
elif os.path.exists(filename + ".gz"):
# print("opening with gzip")
fh = gzip.open(filename + ".gz", "rt")
else:
raise RuntimeError("file %s does not exist" % filename)
return fh
def load_state(filename: str, skip=0, report=0, derived_vars=0):
"""
*filename* is the path to the saved MCMC state up to the final -chain.mc, etc.
Any extension will be removed before using.
*derived_vars* is the number of columns added to each point, derived
from other columns in that point. The newer set_derived_vars interface generates
the derived variables on demand rather than storing them in the state object and
so it will be zero always.
"""
# Trim extension from filename
filename = str(Path(filename).with_suffix(""))
# Read chain file
with openmc(filename + "-chain" + EXT) as fid:
chain = loadtxt_with_fallback(fid)
# Read point file
with openmc(filename + "-point" + EXT) as fid:
line = fid.readline()
point_dims = line[line.find("[") + 1 : line.find("]")]
Nthin, Npop, Nvar = eval(point_dims)
for _ in range(skip * Npop):
fid.readline()
point = loadtxt_with_fallback(fid, report=report * Npop)
# Read stats file
with openmc(filename + "-stats" + EXT) as fd:
stats_header = fd.readline()
stats = loadtxt_with_fallback(fd)
# Determine number of R-stat stored in the stats file
if "R-stat" in stats_header:
# Old header looks like:
# # draws {Nvar}*R-stat {Ncr}*CR_weight
# however, number of R-stat stored in stats file is the number of
# variables stored each generation, not including the derived variables
# calculated after the MCMC has completed.
num_r = int(stats_header.split("*")[0].split()[-1]) - derived_vars
else:
num_r = 0
# Guess dimensions
Ngen = chain.shape[0]
thinning = 1
Nthin -= skip
Nupdate = stats.shape[0]
# Create empty draw and fill it with loaded data
state = MCMCDraw(0, 0, 0, 0, 0, 0, thinning)
# print("gen, var, pop", Ngen, Nvar, Npop)
state.portion = 1.0
state.draws = Ngen * Npop
state.generation = Ngen
state._gen_offset = 0
state._gen_index = 0
state._gen_draws = chain[:, 0].astype(np.int64)
state._gen_acceptance_rate = chain[:, 1]
state._gen_logp = chain[:, 2:]
state._thin_count = Ngen // thinning
state._thin_index = 0
state._thin_draws = state._gen_draws[(skip + 1) * thinning - 1 :: thinning]
state._thin_logp = point[:, 0].reshape((Nthin, Npop))
state._thin_point = reshape(point[:, 1 : Nvar + 1 - derived_vars], (Nthin, Npop, -1))
state._gen_current = state._thin_point[-1].copy()
state._gen_current_logp = state._thin_logp[-1].copy()
state._update_count = Nupdate
state._update_index = 0
state._update_draws = stats[:, 0].astype(np.int64)
state._update_CR_weight = stats[:, 1 + num_r :]
state._outliers = []
bestidx = np.argmax(point[:, 0])
state._best_logp = point[bestidx, 0]
state._best_x = point[bestidx, 1 : Nvar + 1 - derived_vars]
state._best_gen = 0
# Clean up _thin dimensions.
# It seems that some 0.x versions of bumps are yielding thin draws one shorter
# than logp and points, so throw away the first logp and points.
if state._thin_draws.shape[0] == state._thin_logp.shape[0] - 1:
# print(f"shapes differ {state._gen_draws.shape=}, {state._thin_draws.shape=}, {state._thin_logp.shape=}")
state._thin_logp = state._thin_logp[1:]
state._thin_point = state._thin_point[1:]
# Load labels from the .par file if it exists.
parfile = Path(f"{filename}.par")
if parfile.exists():
try:
text = parfile.read_text().strip()
lines = text.split("\n")
labels = [line.rsplit(None, 1)[0].strip() for line in lines]
state.labels = labels
except Exception as exc:
print(f"Couldn't load labels from {parfile}: {exc}")
return state
# TODO: Unused code. Using webview.server.cli.reload_export instead
def reload_state_from_export(parfile):
"""
For the simplified fit with an export path we can reload
the bumps dream state from the exported MCMC files.
*parfile* is the name of the model.par file saved in the
export directory.
Use *state.show()* for the returned state to produce the plots.
"""
assert parfile.endswith(".par"), "need model.par file"
path = str(parfile)[:-4]
state = load_state(path)
state.mark_outliers()
state.portion = state.trim_portion()
return state
[docs]
class MCMCDraw(object):
"""
Initialization parameters:
*Ngen* is the number of generations to store. These are retrievable through
*gen_logp()* and *acceptance_rate()* methods. Note that this is not the same
as the length of the saved chains.
*Nthin* is the number of thinned generations to store. These are retrievable
through the *draw()* and *traces()* methods.
*Nupdate* is the number of crossover ratio sets to save. The DREAM algorithm
runs in batches of steps, with various checks such as convergence and
crossover ratio updates after each batch. The results are retrievable
through the *CR_weight()* method.
*Nvar* is the number of parameters stored in each point.
*Npop* is the number of running chains. Note that unlike the --pop command
line option, this is not the number of chains per parameter.
*Ncr* is the number of crossover ratios. DREAM maintains a fixed set of
crossover ratios, choosing amongst them at each DE step. Depending on the
success rate of the steps, it will adapt the weights after every batch of
updates.
*thinning* is the number of update generations to skip between saves to the
MC chains. Additional thinning can be done when drawing samples from the
saved points.
Attributes and properties:
*Ngen*, *Nthin*, *Nupdate*, *Nvar*, *Npop*, *Ncr* gives the size of the
buffers stored in state. If the buffers are not yet full, the returned sizes
for the *traces()*, *gen_logp()*, *acceptance_rate()*, *CR_weight()* and
*draw()* methods will be smaller.
*generation* is the number of generations in the current fit run. This is
reset to zero each time the fit is resumed.
*draws* is the number of draws in the current fit run. This is reset to zero
each time the fit is resumed. Use *total_generations* time *Npop* if you
want the number of draws including all previous runs.
*total_generations* is the total number of generations for the fit across
all fit runs.
*title* a title for the plot
*labels* a list of names for each parameter
*thinning* amount of thinning between the gen_logp buffer and the traces
buffer. DREAM stores every nth generation starting at generation n
(1-origin). It is possible to change thinning on resume, which will change
the short range correlation in the traces, however these changes are not
recorded. Generation number on the trace plot will be incorrect until DREAM
burns through the entire buffer.
*portion* gives the fraction of each chain to use for statistical plots. A
value can be automatically determined by calling *state.portion =
state.trim_portion()*, or supplied by the user to various functions that
work with the chains. The value should be between 0.0 and 1.0, where 1.0
means the entire chain is used for statistical plots.
"""
_derived_fn: Callable[[NDArray], NDArray] = None
_derived_labels: List[str] = None
_labels = None
_integer_vars = None # boolean array of integer variables, or None
title = None
def __init__(
self,
Ngen: int,
Nthin: int,
Nupdate: int,
Nvar: int,
Npop: int,
Ncr: int,
thinning: int,
):
# self.generation and self.draws are used to control the number of iterations in
# the dream loop, and must therefore be set to the current size of the state on
# resume rather than the cumulative across all runs. To retrieve the totals use
# generations=self.total_generations and draws=self.total_generations*self.Npop.
# Total number of draws so far
self.draws = 0
self.portion = 1.0
# Maximum observed likelihood
# Note: with thinning and burn the best may not be in the set of samples
self._best_x = None
self._best_logp = -inf
self._best_gen = 0
# TODO: Scrap the per-generation data. It takes memory and disk, but doesn't add value.
# Per generation iteration
self.generation = 0
self._gen_offset = 0
self._gen_index = 0
self._gen_draws = empty(Ngen, np.int64)
self._gen_logp = empty((Ngen, Npop), dtype=UNCERTAINTY_DTYPE)
self._gen_acceptance_rate = empty(Ngen, dtype=UNCERTAINTY_DTYPE)
# TODO: should current generation be single precision if the model is single precision?
# The current generation, stored as double precision so that nothing
# is lost on resume.
self._gen_current = empty((Npop, Nvar), dtype="d")
self._gen_current_logp = empty((Npop,), dtype="d")
# Per thinned generation iteration
self.thinning = thinning
self._thin_index = 0
self._thin_count = 0
self._thin_timer = 0
self._thin_draws = empty(Nthin, np.int64)
self._thin_point = empty((Nthin, Npop, Nvar), dtype=UNCERTAINTY_DTYPE)
self._thin_logp = empty((Nthin, Npop), dtype=UNCERTAINTY_DTYPE)
# Per update iteration
self._update_index = 0
self._update_count = 0
self._update_draws = empty(Nupdate, np.int64)
self._update_CR_weight = empty((Nupdate, Ncr), dtype=UNCERTAINTY_DTYPE)
self._outliers = []
# Query functions will not return outlier chains; initially, all
# chains are marked as good. Call mark_outliers to remove
# outlier chains from the set.
self._good_chains = slice(None, None)
@property
def Ngen(self):
return self._gen_draws.shape[0]
@property
def total_generations(self):
return self._gen_offset + self.generation
@property
def Nsamples(self):
return self._gen_logp.size
@property
def Nthin(self):
return self._thin_draws.shape[0]
@property
def Nupdate(self):
return self._update_draws.shape[0]
@property
def Nvar(self):
"""Number of parameters in the fit"""
return self._thin_point.shape[2]
@property
def Npop(self):
return self._gen_logp.shape[1]
@property
def Ncr(self):
return self._update_CR_weight.shape[1]
[docs]
def resize(self, Ngen: int, Nthin: int, Nupdate: int, Nvar: int, Npop: int, Ncr: int, thinning: int):
if self.Nvar != Nvar or self.Npop != Npop or self.Ncr != Ncr:
raise ValueError("Cannot change Nvar, Npop or Ncr on resize")
# print("resize")
# print(self.generation, self.Ngen, Ngen)
# print(self._update_count, self.Nupdate, Nupdate)
# print(self._thin_count, self.Nthin, Nthin)
# When resuming from live state, make sure we unroll the state before resizing.
# This is probably a no-op since plotting will already have forced an unroll.
self._unroll()
self.thinning = thinning
def buf_resize(v, new_size, position):
"""
Resize a circular buffer to *new_size*.
*position* is the number of entries that have been added to the buffer. This
may be less than the existing size if the buffer is not yet filled with one
complete cycle. Assume the buffer has been unrolled so that the next location
index is zero. Don't assume that the next location index is *position%size*.
When extending, we add space to the end of the buffer and set the next index
after the previous end of the buffer. When contracting, we keep the last N entries
in the buffer. If there are fewer than N entries, keep all of them and include
enough empty entries to make the buffer size N.
Need to set the next index to min(position, old_size, new_size)%new_size
"""
if v.shape[0] < new_size: # expand
v = np.resize(v, (new_size, *v.shape[1:]))
elif position < new_size: # contract when not enough
v = v[:new_size]
else: # contract when have enough
v = v[-new_size:]
return v
self._gen_offset += max(self.generation - Ngen, 0)
self.generation = min(self.generation, Ngen, self.Ngen)
self.draws = self.generation * self.Npop
self._gen_index = 0
# Reset sample size to current size on resume.
self.generation = min(self.generation, self.Ngen, Ngen)
self.draws = self.generation * self.Npop
self._gen_index = self.generation % Ngen
self._gen_draws = buf_resize(self._gen_draws, Ngen, self.generation)
self._gen_logp = buf_resize(self._gen_logp, Ngen, self.generation)
self._gen_acceptance_rate = buf_resize(self._gen_acceptance_rate, Ngen, self.generation)
self._thin_count = min(self._thin_count, self.Nthin, Nthin)
self._thin_index = self._thin_count % Nthin
self._thin_draws = buf_resize(self._thin_draws, Nthin, self._thin_count)
self._thin_point = buf_resize(self._thin_point, Nthin, self._thin_count)
self._thin_logp = buf_resize(self._thin_logp, Nthin, self._thin_count)
self._update_count = min(self._update_count, self.Nupdate, Nupdate)
self._update_index = self._update_count % Nupdate
self._update_draws = buf_resize(self._update_draws, Nupdate, self._update_count)
self._update_CR_weight = buf_resize(self._update_CR_weight, Nupdate, self._update_count)
[docs]
def save(self, filename: Union[Path, str]):
save_state(self, filename)
[docs]
def trim_portion(self):
"""
Estimate the point at which the Markov chains have reached equilibrium,
returning it as a portion of the currently stored length.
If not converged, then trim the first half of the samples.
"""
index = burn_point(self)
saved_gens = min(self.Ngen, self.generation)
portion = 1 - (index / saved_gens) if index >= 0 else 0.5
return portion
[docs]
def trim_index(self, portion: Optional[float] = None, generation: Optional[int] = None):
"""
Returns the generation index corresponding to the trim portion. The returned
generation is relative to the start of the sampling, even if resume has been
called multiple times to extend the burn or the sample size.
The optional *generation* parameter is needed in case we have a state object
that is out of date with respect to the number of generations seen so far.
This can happen when the state is transferred to the user interface thread
much less frequently than the step monitor.
"""
# Note: self.total_generations = self.generation + self._gen_offset
# If generation is provided, it represents self.total_generations, so
# we can reconstruct the number of saved generations we would see if the
# state object were up to date.
if generation is None:
total_gens = self.total_generations
saved_gens = min(self.Ngen, self.generation)
else:
total_gens = generation
saved_gens = min(self.Ngen, generation - self._gen_offset)
portion = self.portion if portion is None else portion
return total_gens - int(portion * saved_gens)
[docs]
def show(self, portion: Optional[float] = None, figfile: Union[str, Path, None] = None):
from .views import plot_all
plot_all(self, portion=portion, figfile=figfile)
# TODO: _last_gen and _draw_pop do similar things. Can they be merged?
def _last_gen(self):
"""
Returns x, logp for most recently saved generation. This may be several
generations ago if thinning is enabled.
"""
return self._gen_current, self._gen_current_logp
def _generation(self, new_draws: int, x: NDArray, logp: NDArray, accept: NDArray, force_keep: bool = False):
"""
Called from dream.py after each generation is completed with
a set of accepted points and their values.
"""
# Keep track of the total number of draws
# Note: this is first so that we tag the record with the number of
# draws taken so far, including the current draw.
self.draws += new_draws
self.generation += 1
# Record if this is the best so far
maxid = argmax(logp)
if logp[maxid] > self._best_logp:
self._best_logp = logp[maxid]
self._best_x = x[maxid, :] + 0 # Force a copy
self._best_gen = self.total_generations
# print("new best", logp[maxid], self.generation)
# Record acceptance rate and cost
i = self._gen_index
# print("generation", i, self.draws, "\n x", x, "\n logp", logp, "\n accept", accept)
self._gen_draws[i] = self.draws
self._gen_acceptance_rate[i] = 100 * sum(accept) / new_draws
self._gen_logp[i] = logp
i = i + 1
if i == len(self._gen_draws):
i = 0
self._gen_index = i
# Keep every nth iteration
self._thin_timer += 1
if self._thin_timer == self.thinning or force_keep:
self._thin_timer = 0
self._thin_count += 1
i = self._thin_index
self._thin_draws[i] = self.draws
self._thin_point[i] = x
self._thin_logp[i] = logp
i = i + 1
if i == len(self._thin_draws):
i = 0
self._thin_index = i
self._gen_current[:] = x
self._gen_current_logp[:] = logp
def _update(self, CR_weight: NDArray):
"""
Called from dream.py when a series of DE steps is completed and
summary statistics/adaptations are ready to be stored.
"""
self._update_count += 1
i = self._update_index
# print("update", i, self.draws, "\n CR weight", CR_weight)
self._update_draws[i] = self.draws
self._update_CR_weight[i] = CR_weight
i = i + 1
if i == len(self._update_draws):
i = 0
self._update_index = i
@property
def labels(self):
"""
Labels associated with each parameter in a point. This doesn't
include the labels for derived quantities. For these you will
need to query state.draw().labels.
"""
if self._labels is None:
result = ["P%d" % i for i in range(self._thin_point.shape[2])]
else:
result = self._labels
return result
@labels.setter
def labels(self, v: List[str]):
self._labels = v
def _draw_pop(self):
"""
Return the current population.
"""
return self._gen_current
def _unroll(self):
"""
Unroll the circular queue so that data access can be done inplace.
Call this when done stepping, and before plotting. Calls to
logp, sample, etc. assume the data is already unrolled.
"""
if self.generation > self._gen_index > 0:
self._gen_draws[:] = np.roll(self._gen_draws, -self._gen_index, axis=0)
self._gen_logp[:] = np.roll(self._gen_logp, -self._gen_index, axis=0)
self._gen_acceptance_rate[:] = np.roll(self._gen_acceptance_rate, -self._gen_index, axis=0)
self._gen_index = 0
if self._thin_count > self._thin_index > 0:
self._thin_draws[:] = np.roll(self._thin_draws, -self._thin_index, axis=0)
self._thin_point[:] = np.roll(self._thin_point, -self._thin_index, axis=0)
self._thin_logp[:] = np.roll(self._thin_logp, -self._thin_index, axis=0)
self._thin_index = 0
if self._update_count > self._update_index > 0:
self._update_draws[:] = np.roll(self._update_draws, -self._update_index, axis=0)
self._update_CR_weight[:] = np.roll(self._update_CR_weight, -self._update_index, axis=0)
self._update_index = 0
[docs]
def remove_outliers(self, x: NDArray, logp: NDArray, test: str = "IQR"):
"""
Replace outlier chains with clones of good ones. This should happen
early in the sampling processes so the clones have an opportunity
to evolve their own identity. Only the head of the chain is modified.
*state* contains the chains, with log likelihood for each point.
*x*, *logp* are the current population and the corresponding
log likelihoods; these are updated with cloned chain values.
*test* is the name of the test to use (one of IQR, Grubbs, Mahal
or none). See :func:`.outliers.identify_outliers` for details.
Updates *state*, *x* and *logp* to reflect the changes.
Returns a list of the outliers that were removed.
"""
# Grab the last part of the chain histories
_, chains = self.logp()
chain_len, Nchains = chains.shape
outliers = identify_outliers(test, chains, x)
# if len(outliers): print("old llf", logp[outliers])
# Loop over each outlier chain, replacing each with another
for old in outliers:
# Draw another chain at random, with replacement
# TODO: consider using relative likelihood as a weight factor
while True:
new = rng.randint(Nchains)
if new not in outliers:
break
# Update the saved state and current population
self._replace_outlier(old=old, new=new)
x[old, :] = x[new, :]
logp[old] = logp[new]
# if len(outliers): print("new llf", logp[outliers])
return outliers
def _replace_outlier(self, old: int, new: int):
"""
Called from outliers.py when a chain is replaced by the
clone of another.
"""
self._outliers.append((self._thin_index, old, new))
# 2017-10-06 [PAK] only replace the head, not the full chain
index = self._gen_index
self._gen_current[old] = self._gen_current[new]
self._gen_logp[index, old] = self._gen_logp[index, new]
self._thin_logp[index, old] = self._thin_logp[index, new]
self._thin_point[index, old, :] = self._thin_point[index, new, :]
[docs]
def mark_outliers(self, test: str = "IQR", portion: Optional[float] = None):
"""
Mark some chains as outliers but don't remove them. This can happen
after drawing is complete, so that chains that did not converge are
not included in the statistics.
*test* is 'IQR', 'mahal', 'grubbs', or 'none'.
*portion* indicates what portion of the samples should be included
in the outlier test. If None, then the stored portion is used.
"""
_, chains, logp = self.chains()
if test == "none":
self._good_chains = slice(None, None)
else:
thin_saved_gens = chains.shape[0]
portion = self.portion if portion is None else portion
start = int(thin_saved_gens * (1 - portion))
outliers = identify_outliers(test, logp[start:], chains[-1])
# print("outliers", outliers, logp.shape, chains.shape)
if len(outliers) > 0:
self._good_chains = np.array([i for i in range(logp.shape[1]) if i not in outliers])
else:
self._good_chains = slice(None, None)
# print(self._good_chains)
[docs]
def logp(self, full: bool = False):
"""
Return the cumulative number of draws and the log likelihoods for each
chain.
Note that draw[i] represents the total number of samples taken,
including those for the samples in logp[i].
If full is True, then return all chains, not just good chains.
** Deprecated ** Use state.gen_logp() instead.
"""
# self._unroll()
# draws, logp = self._gen_draws, self._gen_logp
# if self.generation == self._gen_index:
# draws, logp = [v[:self.generation] for v in (draws, logp)]
# Don't do a full unroll here
if self.generation == self._gen_index:
draws = self._gen_draws[: self._gen_index]
logp = self._gen_logp[: self._gen_index]
elif self._gen_index > 0:
draws = np.roll(self._gen_draws, -self._gen_index, axis=0)
logp = np.roll(self._gen_logp, -self._gen_index, axis=0)
else:
draws = self._gen_draws
logp = self._gen_logp
return draws, (logp if full else logp[:, self._good_chains])
[docs]
def gen_logp(self, portion: Optional[float] = None, outliers: bool = False):
"""
Return the iteration number and the log likelihood for each point in
the individual sequences in that iteration.
For example, to plot the convergence of each sequence::
genid, logp = state.gen_logp()
plot(genid, logp)
*portion* is the amount to trim from the head of the chain, or None
to use the stored burn point.
If *outliers* is True, then return all chains, not just good chains.
"""
# Don't do a full unroll here
if self.generation == self._gen_index:
logp = self._gen_logp[: self._gen_index]
elif self._gen_index > 0:
logp = np.roll(self._gen_logp, -self._gen_index, axis=0)
else:
logp = self._gen_logp
stop, step = self.total_generations, 1 # No thinning on _gen_logp, _gen_draws
start = stop - (len(logp) - 1) * step
genid = np.arange(start, stop + 1, step) # 1-origin, end-inclusive
assert len(logp) == len(genid)
portion = self.portion if portion is None else portion
start = int((1 - portion) * len(logp))
gen_index = slice(start, None)
chain_index = slice(None) if outliers else self._good_chains
return genid[gen_index], logp[gen_index, chain_index]
[docs]
def logp_slice(self, n: int):
"""
Return a slice of the logp chains, either the first n if n > 0
or the last n if n < 0. Avoids unrolling the circular buffer if
possible.
"""
if n < 0: # tail
if self._gen_index >= -n:
return self._gen_logp[self._gen_index + n : self._gen_index]
elif self._gen_index == 0:
return self._gen_logp[n:]
else: # unroll across boundary
return np.vstack((self._gen_logp[n + self._gen_index :], self._gen_logp[: self._gen_index]))
else: # head
if self.generation < self.Ngen:
return self._gen_logp[:n]
elif self._gen_index + n <= self.Ngen:
return self._gen_logp[self._gen_index : self._gen_index + n]
else:
return np.vstack((self._gen_logp[self._gen_index :], self._gen_logp[-n + self._gen_index :]))
[docs]
def min_slice(self, n: int):
"""
Return the minimum logp for n slices, from the head if positive
or the tail if negative.
This is a specialized function so it can be fast. Convergence
can be quickly rejected if the min in a short head is smaller
than the min in a long tail. Unfortunately, if the data is
wrapped, then the max function will cost extra.
"""
# Copy the logic of slice
if n < 0: # tail
if self._gen_index >= -n:
return np.min(self._gen_logp[self._gen_index + n : self._gen_index])
elif self._gen_index == 0:
return np.min(self._gen_logp[n:])
else: # max across boundary
return min(np.min(self._gen_logp[n + self._gen_index :]), np.min(self._gen_logp[: self._gen_index]))
else: # head
if self.generation < self.Ngen:
return np.min(self._gen_logp[:n])
elif self._gen_index + n <= self.Ngen:
return np.min(self._gen_logp[self._gen_index : self._gen_index + n])
else:
return min(np.min(self._gen_logp[self._gen_index :]), np.min(self._gen_logp[-n + self._gen_index :]))
[docs]
def acceptance_rate(self, portion: Optional[float] = None):
"""
Return the iteration number and the acceptance rate for that iteration.
For example, to plot the acceptance rate over time::
genid, AR = state.acceptance_rate()
plot(genid, AR)
V1.0.2: Now returns (genid, AR) rather than (num_draws, AR)
"""
AR = self._gen_acceptance_rate
if self.generation == self._gen_index:
AR = AR[: self.generation]
elif self._gen_index > 0:
AR = np.roll(AR, -self._gen_index, axis=0)
# Generation numbers for the buffered acceptance rates. The acceptance
# rate is stored for each generation regardless of thinning, so this is
# a simple increment.
last, step = self.total_generations, 1
first = last - (len(AR) - 1) * step
generation = np.arange(first, last + 1, step) # 1-origin, end-inclusive
portion = self.portion if portion is None else portion
index = int((1 - portion) * len(AR))
return generation[index:], AR[index:]
[docs]
def chains(self):
"""
Returns the observed Markov chains and the corresponding likelihoods.
The return value is a tuple (*draws*, *chains*, *logp*).
*draws* is the number of samples taken up to and including the samples
for the current generation.
*chains* is a three dimensional array of generations X chains X vars
giving the set of points observed for each chain in every generation.
Only the thinned samples are returned.
*logp* is a two dimensional array of generation X population giving
the log likelihood of observing the set of variable values given in
chains.
"""
self._unroll()
retval = self._thin_draws, self._thin_point, self._thin_logp
if self._thin_count == self._thin_index:
retval = [v[: self._thin_count] for v in retval]
return retval
[docs]
def traces(self, portion: Optional[float] = None, thin: int = 1, outliers: bool = False):
"""
Grab traces for trace plot.
*portion* gives the starting point of the trace, skipping over the initial frames
that may not have converged.
*thin* is the amount of additional thinning to apply on the traces.
*outliers* (default False) is True if bad chains should be included in the trace.
Returns *generation* [Ngen] and *chains* [Ngen x Nchain x Nvar]
"""
portion = self.portion if portion is None else portion
# TODO: Chains code is duplicated in Draw
# Draw samples from the state according to portion, thinning and outliers.
# If we also want derived variables and value range limits then we should
# switch to the code in Draw.
_, chains, _ = self.chains()
start = int((1 - portion) * len(chains))
gen_index = slice(start, None, thin)
chain_index = slice(None) if outliers else self._good_chains
chains = chains[gen_index, chain_index, :]
num_steps = chains.shape[0]
last, step = self.total_generations, self.thinning * thin
first = last - (num_steps - 1) * step
generation = np.arange(first + 1, last + 2, step) # 1-origin inclusive
return generation, chains
[docs]
def gelman(self, portion: Optional[float] = None):
"""
Compute the Gelman and Rubin R-statistic for the Markov chains.
"""
# Note: Not allowing a portion parameter here because that would
# require unrolling the chains or otherwise making the logic in
# gelman() too complicated. Keep it simple so that we can show
# a simple R statistic over time without thrashing memory.
if self.generation < self.Ngen:
return gelman(self._thin_point[: self.generation], portion=1.0)
else:
return gelman(self._thin_point, portion=1.0)
[docs]
def CR_weight(self):
"""
Return the crossover ratio weights to be used in the next generation.
For example, to see if the adaptive CR is stable use::
draw, weight = state.CR_weight()
plot(draw, weight)
See :mod:`.crossover` for details.
"""
self._unroll()
retval = self._update_draws, self._update_CR_weight
if self._update_count == self._update_index:
retval = [v[: self._update_count] for v in retval]
return retval
[docs]
def outliers(self):
"""
Return a list of outlier removal operations.
Each outlier operation is a tuple giving the thinned generation
in which it occurred, the old chain id and the new chain id.
The chains themselves have already been updated to reflect the
removal.
Curiously, it is possible for the maximum likelihood seen so far
to be removed by this operation.
"""
return asarray(self._outliers, "i")
[docs]
def best(self):
"""
Return the best point seen and its log likelihood.
"""
return self._best_x, self._best_logp
[docs]
def stable_best(self):
"""
Return True if the at least one full cycle of the circular
buffer has passed since the best logp was first observed.
"""
# print(f"stable_best {self._best_gen} + {self.Ngen} <= {self.total_generations}")
return self._best_gen + self.Ngen <= self.total_generations
[docs]
def keep_best(self):
"""
Place the best point at the end of the last good chain.
Good chains are defined by mark_outliers.
Because the Markov chain is designed to wander the parameter
space, the best individual seen during the random walk may have
been observed during the burn-in period, and may no longer be
present in the chain. If this is the case, replace the final
point with the best, otherwise swap the positions of the final
and the best.
"""
# Get state as a 1D array
_, chains, logp = self.chains()
Ngen, Npop, Nvar = chains.shape
points = reshape(chains, (Ngen * Npop, Nvar))
logp = reshape(logp, Ngen * Npop)
# Set the final position to the end of the last good chain. If
# mark_outliers has not been called, then _good_chains will
# just be slice(None, None)
if isinstance(self._good_chains, slice):
final = -1
else:
final = self._good_chains[-1] - Npop
# Find the location of the best point if it exists and swap with
# the final position, or overwrite the final position if it does
# not exist.
idx = np.where(logp == self._best_logp)[0]
if len(idx) == 0:
logp[final] = self._best_logp
points[final, :] = self._best_x
else:
# If there are multiple best values, arbitrarily choose one of them.
# The chances are very high that it is the same point that was kept
# after a proposed step was rejected.
idx = idx[0]
logp[final], logp[idx] = logp[idx], logp[final]
points[final, :], points[idx, :] = points[idx, :], points[final, :]
[docs]
def sample(self, **kw):
"""
Return a sample from the posterior distribution.
**Deprecated** use :meth:`draw` instead.
"""
drawn = self.draw(**kw)
return drawn.points, drawn.logp
[docs]
def entropy(
self,
vars: Optional[List[int]] = None,
portion: Optional[float] = None,
selection: SelectionType = None,
n_est: int = 10000,
thin: Optional[int] = None,
method: Optional[str] = None,
):
r"""
Return entropy estimate and uncertainty from an MCMC draw.
*portion* is the portion of each chain to use (uses self.portion if None).
*vars* is the set of variables to marginalize over. It is None for
the visible variables, or a list of variables.
*vars* is the list of variables to use for marginalization.
*selection* sets the range each parameter in the returned distribution,
using {variable: (low, high)}. Missing variables use the full range.
*n_est* is the number of points to use from the draw when estimating
the entropy (default=10000).
*thin* is the amount of thinning to use when selecting points from the
draw.
*method* determines which entropy calculation to use:
* gmm: fit sample to a gaussian mixture model (GMM) with $5 \sqrt{d}$
components where $d$ is the number fitted parameters and estimate
entropy by sampling from the GMM.
* llf: estimates likelihood scale factor from ratio of density
estimate to model likelihood, then computes Monte Carlo entropy
from sample; this does not work for marginal likelihood estimates.
DOI:10.1109/CCA.2010.5611198
* mvn: fit sample to a multi-variate Gaussian and return the entropy
of the best fit gaussian; uses bootstrap to estimate uncertainty.
* wnn: estimate entropy from nearest-neighbor distances in sample.
DOI:10.1214/18-AOS1688
"""
from . import entropy
# Get the sample from the state.
# set default thinning to max((steps * samples/step) // n_est, 1)
if thin is None:
Nsteps = min(self.Nthin, self._thin_count)
thin = max(Nsteps * self.Npop // n_est, 1)
# print("thin", thin, Nsteps, self.Npop, self.Nthin, self._thin_count)
drawn = self.draw(portion=portion, vars=vars, selection=selection, thin=thin)
# TODO: don't print within a library function!
M = entropy.MVNEntropy(drawn.points)
print("Entropy from MVN: %s" % str(M))
if method is None:
# TODO: change default to gmm
method = "llf"
if method == "llf":
S, Serr = entropy.entropy(drawn.points, drawn.logp, N_entropy=n_est)
# print("Entropy from llf (Kramer): %s"%str(S))
elif method == "gmm":
# Try pure gmm ... pretty good
S, Serr = entropy.gmm_entropy(drawn.points, n_est=n_est)
# print("Entropy from gmm: %g +/- %g"% (S, Serr))
elif method == "wnn":
# Try pure wnn ... no good
S, Serr = entropy.wnn_entropy(drawn.points, n_est=n_est)
# print("Entropy from wnn: %s"%str(S))
elif method == "mvn":
S, Serr = entropy.mvn_entropy_bootstrap(drawn.points)
# print("Entropy from mvn: %s"%str(S))
else:
raise ValueError("unknown method %r" % method)
# Always return entropy estimate from draw, even if it is normal
return S, Serr
[docs]
def show_labels(self):
"""
List of parameters in the state, with the parameter number for each. Use this
to help with the inputs to state.draw().
"""
print("\n".join(f"p[{k}] = {v}" for k, v in enumerate(self.labels)))
[docs]
def draw(
self,
portion: Optional[float] = None,
vars: Optional[List[int]] = None,
exclude: Optional[List[int]] = None,
selection: SelectionType = None,
thin: int = 1,
outliers: bool = False,
derived: Optional[Callable] = None,
):
"""
Return a sample from the posterior distribution.
*portion* is the portion of each chain to use
*vars* is the list of vars that are active. Variables can be
0-origin integers or strings matching a variable label. Match uses
standard glob rules, with `*` matching any sequence, `?` matching any
character, `[abc]` matching one of `abc`, and `[!abc]` not matching
one of `abc`.
*selection* sets the range each parameter in the returned distribution,
using {var: (low, high)}. If *var* matches multiple labels, then use the
first restriction only. Missing variables use the full range.
*thin* takes every nth item.
*outliers* is True if outlier chains should be included (default False).
To plot the distribution for parameter p1::
draw = state.draw()
hist(draw.points[:, 0])
To plot the interdependence of p1 and p2::
draw = state.sample()
plot(draw.points[:, 0], draw.points[:, 1], '.')
"""
# Fill in defaults
vars = vars if vars is not None else getattr(self, "_shown", None)
portion = self.portion if portion is None else portion
return Draw(
state=self,
portion=portion,
vars=vars,
exclude=exclude,
selection=selection,
thin=thin,
outliers=outliers,
derived=derived,
)
# TODO: Move processing of visible/integer/derived out of state
[docs]
def set_visible_vars(self, labels: List[str]):
self._shown = [self.labels.index(v) for v in labels]
# print("\n".join(str(pair) for pair in enumerate(self.labels)))
# print(labels)
# print(self._shown)
[docs]
def set_integer_vars(self, labels: List[str]):
"""
Indicate tha variables should be considered integer variables when
computing statistics.
"""
self._integer_vars = np.array([var in labels for var in self.labels])
[docs]
def set_derived_vars(self, fn: Callable[[NDArray], NDArray], labels: List[str]):
"""
Define derived variables from the sample. When calling draw() it will add
columns for the derived variables to each sample.
*fn* is a function taking points p[:, k] for k in 0 ... samples and
returning a set of derived variables pj[k] for each sample k. The
variables can be returned as any kind of sequence including an
array or a tuple with one entry per variable. The caller uses
asarray to convert the returned variables into a vars X samples array.
For convenience, a single variable can be returned by itself.
*labels* are the labels to use for the derived variables.
The following example adds the new variable x+y = P[0] + P[1]::
state.derive_vars(lambda p: p[0]+p[1], labels=["x+y"])
"""
self._derived_fn = fn
self._derived_labels = labels
[docs]
def derive_vars(self, fn: Callable[[NDArray], NDArray], labels: Optional[List[str]] = None):
"""
*** DEPRECATED ***
Like set_derived_vars but operating in place, modifying the points in the history.
"""
warnings.warn("Use state.set_derived_vars() instead of state.derive_vars()", DeprecationWarning, stacklevel=2)
# Grab all samples as a set of points
_, chains, _ = self.chains()
Ngen, Npop, Nvar = chains.shape
points = reshape(chains, (Ngen * Npop, Nvar))
# Compute new variables from the points
newvars = asarray(fn(points.T)).T
Nnew = newvars.shape[1] if len(newvars.shape) == 2 else 1
newvars.reshape((Ngen, Npop, Nnew))
# Extend new variables to be the same length as the stored selection
Nthin = self._thin_point.shape[0]
newvars = np.resize(newvars, (Nthin, Npop, Nnew))
# Add labels for the new variables, if available. This must
# occur before updating self._thin_point, otherwise a default
# label may be generated for the new column.
if labels is not None:
self.labels = self.labels + labels
elif self._labels is not None:
labels = ["P%d" % i for i in range(Nvar, Nvar + Nnew)]
self.labels = self.labels + labels
else: # no labels specified, old or new
pass
# Add new variables to the points
self._thin_point = dstack((self._thin_point, newvars))
[docs]
class Draw:
"""
A set of sample points from an MCMC draw with restrictions applied.
**Inputs**
*state* is the :class:`MCMCDraw` containing all points.
*portion* is the fraction of points to use for the sample, starting
from the ends of the Markov chains.
*vars* is the list of vars that are active. Variables can be
0-origin integers or strings matching a variable label. Match uses
standard glob rules, with `*` matching any sequence, `?` matching any
character, `[abc]` matching one of `abc`, and `[!abc]` not matching
one of `abc`.
*exclude* all labels that match a pattern in the list. Matching rules
are the same as for *vars*. If a variable matches both *vars* and
*exclude* then it is included.
*selection* restricts the points returned to a specific region of the
parameter space using {var: (low, high)}. The var can be a 0-origin
integer or label as for *vars*. It can also be *logp* if restricting by
log likelihood. If a label matches multiple patterns then choose the
first restriction (this may change to all that match in future).
*thin* is the amount of thinning to apply, using every nth point in
the chain. This reduces autocorrelation in the chains and reduces
artificial spikes in the marginal distributions. (The distributions
will still be spiky, but the spikiness will be appropriate for the
number of points in the bins.)
**Attributes**
*points[n,k]* are the set of sampled points, where n is the number of
points and k is the number of parameters. The earlier points come
from earlier generations, but this is an accident of the implementation
and should not be relied on. n is usually the number of chains times the
number of saved generations, but this will change depending on *portion*,
*selection*, *thin* and *outliers* parameters. k is usually the number
of fitted parameters, but the derived and active parameter lists will
adjust this.
*Nvar* is the number of dimensions per point in the draw after
including derived parameters and excluding nuisance parameters.
*weights* are the weights associated with each point. Since the chains
are run at a single temperature the weights will all be 1.
*logp[n]* is the negative log likelihuud for each sampled point.
*labels* are the labels for the variables remaining after applying derived
variables and restricting to the requested *vars* list.
*integers* is an array of flags, one per parameter* indicating whether the
parameter is float (0) or int (1=floor). The values in the chain are
managed as floats, and it is the responsibility of the likelihood function
to transform them to integers. Models which use round, trunc or ceil rather
than float are not yet supported.[1]
*state* is the original MCMCDraw.
[1] Note that discrete parameter handling is likely to change in the future.
The derivative based optimizers in particular will need to know that the
minimum step size for discrete parameters is one when computing the partial
derivative. If bumps converts the parameters to integers before calling the
likelihood function then we will only need a discrete flag since we will
always use the *floor()* function.
"""
state: MCMCDraw
vars: Optional[List[Union[int, str]]]
portion: float
selection: SelectionType
thin: int
Nvar: int
title: Optional[str]
labels: List[str]
integers: NDArray # boolean
logp: NDArray
points: NDArray
weights: NDArray
# # MC chains, with mask for included points.
# generation: NDArray
# chains: NDArray
# chains_logp: NDArray
# mask: Optional[NDArray]
def __init__(
self,
state: MCMCDraw,
portion: float = 1.0,
vars: Optional[List[Union[int, str]]] = None,
exclude: Optional[List[Union[int, str]]] = None,
selection: SelectionType = None,
thin: int = 1,
outliers: bool = False,
derived: Optional[Callable] = None,
):
self.state = state
self.portion = portion
self.thin = thin
self.selection = selection
self.labels = state.labels[:] # copy of labels
self.integers = (
state._integer_vars if state._integer_vars is not None else np.zeros(len(self.labels), dtype=bool)
)
self.title = state.title
# TODO: Make a copy?
# Draw samples from the state according to portion and thinning.
_, chains, logp = state.chains()
start = int((1 - portion) * len(chains))
gen_index = slice(start, None, thin)
chain_index = slice(None) if outliers else state._good_chains
chains = chains[gen_index, chain_index, :]
chains_logp = logp[gen_index, chain_index]
Ngen, Npop, Nvar = chains.shape
# Derived variables are created during the draw so that existing data isn't altered.
# This allows resume to work without extra effort. The code is also much simpler.
if derived is None:
derived = state._derived_fn
if derived is not None:
newvars = derived(chains.reshape(-1, Nvar).T)
# If derived returns a dict, then use its keys as parameter labels
if isinstance(newvars, dict):
newlabels, newvars = zip(*newvars.items())
else:
# Backward compatibility: old interface had derived returning a list
newlabels = state._derived_labels
if not newlabels:
newlabels = [f"F{k+1}" for k in range(len(derived))]
newvars = asarray(newvars).T
newvars = newvars.reshape(Ngen, Npop, -1)
chains = np.dstack((chains, newvars))
Nvar = chains.shape[2] # Update the number of variables
self.integers = np.append(self.integers, [False] * newvars.shape[2])
self.labels.extend(newlabels)
# Select points within range limits given by selection, if any.
# This happens before variable subsetting so that numerical indices
# are independent of the displayed parameters. It happens after
# calculating derived variables so that we can apply selection to derived values.
mask = True
selection = self._selection_pattern_to_index(selection)
for var, limits in selection.items():
if var == "logp":
mask = mask & (chains_logp >= limits[0]) & (chains_logp <= limits[1])
else:
mask = mask & (chains[..., var] >= limits[0]) & (chains[..., var] <= limits[1])
# TODO: with pattern matching we've lost the ability to reorder parameters
# Restrict draw to a subset of the variables. Labels matching include take
# precedence over those matching exclude.
include = self._var_pattern_to_index(vars)
exclude = self._var_pattern_to_index(exclude)
exclude = [k for k in exclude if k not in include]
include = [k for k in include if k not in exclude]
# if vars and not include:
# # Note: could be empty because everything is excluded
# logging.warning(f"No parameters match {vars} in {self.labels}")
if include: # if nothing matches then match everything
chains = chains[:, :, include]
# Mask doesn't include the final dimension, so doesn't need to be trimmed.
Nvar = chains.shape[-1]
self.labels = [self.labels[v] for v in include]
self.integers = self.integers[include]
self.vars = include # indices of selected items
else:
self.vars = None
# Convert to a vector of selected points.
if mask is True:
# All points selected, so simply adjust the view
self.points = chains.reshape(-1, Nvar)
self.logp = chains_logp.flatten()
else:
# Only some points selected, so grab those under the mask
self.points = chains[mask, :]
self.logp = chains_logp[mask]
# # Keep the thinned chains and the mask for trace plots
# self.chains = chains
# self.chains_logp = chains_logp
# self.mask = None if mask is True else mask
# stop, step = state.total_generations, thin*state.thinning
# self.generation = np.arange(stop - (Ngen-1)*step + 1, stop + 2, step) # 1-origin inclusive
self.weights = None
self.Nvar = Nvar
# Cache for the sorted argument indices.
self._argsort_indices = {}
[docs]
def get_argsort_indices(self, var: int):
"""
Sort var by value. Unlike argsort this caches the results for reuse.
"""
if var not in self._argsort_indices:
self._argsort_indices[var] = np.argsort(self.points[:, var].flatten())
return self._argsort_indices[var]
def _selection_pattern_to_index(self, selection: dict[int | str, tuple[float, float]] | None):
"""
For the selection dictionary {pattern: data}, return a dictionary of {index: data}
where data comes from the first pattern to matched by the label. Patterns are the
standard unix glob patterns, or integers for specific numbered variables
(see :meth:`State.show_labels`).
"""
if selection is None:
return {}
result = {}
if data := selection.get("logp", None):
result["logp"] = data
for index, label in enumerate(self.labels):
for var, data in selection.items():
if index == var or (isinstance(var, str) and fnmatch(label, var)):
result[index] = data
break
return result
def _var_pattern_to_index(self, vars: list[int | str] | None):
"""
Convert the include/exclude list of var patterns into a list of integer
parameter indices. Patterns are the standard unix glob patterns, or integers
for specific numbered variables (see :meth:`State.show_labels`)
"""
if vars is None:
return []
def var_match(var, label, index):
return index == var or (isinstance(var, str) and fnmatch(label, var))
return [k for k, p in enumerate(self.labels) if any(var_match(v, p, k) for v in vars)]
def test():
from numpy.linalg import norm
from numpy.random import rand
from numpy import arange
# Make some fake data
Nupdate, Nstep = 3, 5
Ngen = Nupdate * Nstep
Nvar, Npop, Ncr = 3, 6, 2
xin = rand(Ngen, Npop, Nvar)
xin[..., 2] *= 9 # make one parameter range large enough for integer tests
xin += 0.1 # force above 0.1 so that arrays print nicer
pin = rand(Ngen, Npop)
accept = rand(Ngen, Npop) < 0.8
CRin = rand(Nupdate, Ncr)
# thinning = 2
# Nthin = int(Ngen/thinning)
# Put it into a state
thinning = 2
Nthin = int(Ngen / thinning)
thinned = xin[thinning - 1 :: thinning, :, :]
state = MCMCDraw(Ngen=Ngen, Nthin=Nthin, Nupdate=Nupdate, Nvar=Nvar, Npop=Npop, Ncr=Ncr, thinning=thinning)
for i in range(Nupdate):
state._update(CR_weight=CRin[i])
for j in range(Nstep):
gen = i * Nstep + j
state._generation(new_draws=Npop, x=xin[gen], logp=pin[gen], accept=accept[gen])
# Check that it got there
assert (thinned == state.chains()[1]).all()
draws, logp = state.logp()
assert norm(draws - Npop * arange(1, Ngen + 1)) == 0
# Data is generated as double but may be stored as single,
# so there could be a loss of precision.
assert norm(logp - pin.astype(UNCERTAINTY_DTYPE)) == 0
genid, logp = state.gen_logp()
assert norm(genid - arange(1, Ngen + 1)) == 0
assert norm(logp - pin) == 0
genid, AR = state.acceptance_rate()
assert norm(genid - arange(1, Ngen + 1)) == 0
assert norm(AR - (100 * sum(accept, axis=1) / Npop).astype(UNCERTAINTY_DTYPE)) == 0
draws, logp = state.sample()
# assert norm(draws - thinning*Npop*arange(1, Nthin+1)) == 0
# assert norm(sample - xin[thinning-1::thinning]) == 0
# assert norm(logp - pin[thinning-1::thinning]) == 0
draws, CR = state.CR_weight()
assert norm(draws - Npop * Nstep * arange(Nupdate)) == 0
assert norm(CR - CRin.astype(UNCERTAINTY_DTYPE)) == 0
x, p = state.best()
bestid = argmax(pin)
i, j = bestid // Npop, bestid % Npop
assert pin[i, j] == p
assert norm(xin[i, j, :] - x) == 0
draw = state.draw(selection={"P1": (0.4, 1.1)}, portion=1, outliers=True)
assert ((draw.points[:, 1] >= 0.4) & (draw.points[:, 1] <= 1.1)).all()
# print(thinned[..., 1])
# print(thinned[..., 1] >= 0.4, thinned.shape)
# print(len(draw.points[:, 1]), np.sum(thinned[..., 1] >= 0.4))
# print("thinned", thinned[thinned[..., 1] >= 0.4][:, 1])
# print("points", draw.points[:, 1])
mask = (thinned[..., 1] >= 0.4) & (thinned[..., 1] <= 1.1)
assert len(draw.points[:, 1]) == np.sum(mask)
if 0: # test of deprecated interface
# Test derived variables (inplace update! ick!)
state.derive_vars(lambda p: p[0] + p[1], labels=["inplace x+y"])
assert state.traces()[1].shape[2] == Nvar + 1
draw = state.draw()
assert draw.labels[-1] == "inplace x+y"
assert (draw.points[:, -1] == draw.points[:, 0] + draw.points[:, 1]).all()
assert draw.points.shape[1] == Nvar + 1
assert len(draw.labels) == Nvar + 1
# Test derived variables (ephemeral)
state.set_derived_vars(lambda p: p[0] + p[1], labels=["x+y"])
draw = state.draw()
assert draw.labels[-1] == "x+y"
assert (draw.points[:, -1] == draw.points[:, 0] + draw.points[:, 1]).all()
assert draw.points.shape[1] == Nvar + 1
assert len(draw.labels) == Nvar + 1
from .stats import var_stats, format_vars
vstats = var_stats(state.draw(vars=[2, "x+y"]))
print(format_vars(vstats))
# Test integer vars
state.set_integer_vars(labels=["P2"])
draw = state.draw()
vstats_int = var_stats(draw)
# print("integers", draw.integers)
# print(format_vars(vstats_int))
assert vstats_int[2].mean == np.mean(np.floor(draw.points[:, 2]))
draw = state.draw(selection={"P1": (0.4, 1.1)})
# print(format_vars(var_stats(draw)))
assert (draw.points[:, 1] >= 0.4).all()
# print(thinned[..., 1])
# print(thinned[..., 1] >= 0.4, thinned.shape)
# print(len(draw.points[:, 1]), np.sum(thinned[..., 1] >= 0.4))
# print("thinned", thinned[thinned[..., 1] >= 0.4][:, 1])
# print("points", draw.points[:, 1])
assert len(draw.points[:, 1]) == np.sum(thinned[..., 1] >= 0.4)
# Note: check mods to the state data structures last
# Check that outlier updates properly
state._replace_outlier(1, 2)
outliers = state.outliers()
draws, logp = state.sample()
assert norm(outliers - asarray([[state._thin_index, 1, 2]])) == 0
# assert norm(sample[:, 1, :] - xin[thinning-1::thinning, 2, :]) == 0
# assert norm(sample[:, 2, :] - xin[thinning-1::thinning, 2, :]) == 0
# assert norm(logp[:, 1] - pin[thinning-1::thinning, 2]) == 0
# assert norm(logp[:, 2] - pin[thinning-1::thinning, 2]) == 0
if __name__ == "__main__":
test()