def __init__(self, parent=None):
Frame.__init__(self, parent, bg="yellow")
self.pack(expand=YES, fill=BOTH)
# number of points to sample when manySteps function called
self.manyStepsN = 1000
# initialize the density from which to sample
self.d = MyBeta()
# initialize data members related to MCMC proposal
# delta is the width of the proposal window centered
# around current value
self.delta = window_width
self.x = 0.5 # will be reset in rescalePlot
# initialize SAMC-related data members
self.using_samc = True
self.biased_pi = True
self.m = None
self.levels = None
self.loglevels = None
self.theta = None
self.true_pi = None
self.est_pi = None
self.sample_size = None
self.gain = 1.0
self.t0 = 10.0
self.limit_log_theta = 100.0 * math.log(10.0)
# initialize miscellaneous data members
self._step_num_within_step = 0
self._just_point = False
self.plotter = None
self.n_func_evals = 0
# set up pseuodorandom number generator
self.r = Lot()
# self.r.setSeed(13579)
print "seed is =", self.r.getSeed()
# create a frame to hold the menu buttons
menuf = Frame(self, bg="magenta")
menuf.pack(expand=NO, fill=X)
# create the File menu button
self.filemb = Menubutton(menuf, text="File", relief=RAISED, anchor=W, borderwidth=0)
self.filemb.pack(expand=NO, fill=X, side=LEFT)
self.filemb.menu = Menu(self.filemb, tearoff=0)
self.filemb["menu"] = self.filemb.menu
self.filemb.menu.add_command(label="Quit", command=self.quit)
# create the Sample menu button
self.samplemb = Menubutton(menuf, text="Sample", relief=RAISED, anchor=W, borderwidth=0)
self.samplemb.pack(expand=YES, fill=X, side=LEFT)
self.samplemb.menu = Menu(self.samplemb, tearoff=1)
self.samplemb["menu"] = self.samplemb.menu
self.samplemb.menu.add_command(label="Toggle between SAMC and MCMC (s)", command=self.toggleSAMC)
self.samplemb.menu.add_command(label="One step (n)", command=self.oneStep)
self.samplemb.menu.add_command(label="Many steps (m)", command=self.manySteps)
self.samplemb.menu.add_command(label="Many steps using slice sampler (M)", command=self.manyStepsSliceSampler)
self.samplemb.menu.add_command(label="Biased SAMC (b)", command=self.biasedSAMC)
self.samplemb.menu.add_command(label="Unbiased SAMC (u)", command=self.unbiasedSAMC)
self.samplemb.menu.add_command(label="Reset (r)", command=self.reset)
self.samplemb.menu.add_command(label="Use Beta (B)", command=self.switchToBeta)
self.samplemb.menu.add_command(label="Use Bimodal (I)", command=self.switchToBimodal)
self.samplemb.menu.add_command(label="Use Exponential (E)", command=self.switchToExpon)
# bind some keys to the application (doesn't matter which widget has focus
# when you use bind_all)
self.bind_all("<KeyPress-s>", self.keybdToggleSAMC)
self.bind_all("<KeyPress-n>", self.keybdOneStep)
self.bind_all("<KeyPress-m>", self.keybdManySteps)
self.bind_all("<Shift-KeyPress-M>", self.keybdManyStepsSliceSampler)
self.bind_all("<KeyPress-b>", self.keybdBiasedSAMC)
self.bind_all("<KeyPress-u>", self.keybdUnbiasedSAMC)
self.bind_all("<KeyPress-r>", self.keybdReset)
self.bind_all("<KeyPress-q>", self.keybdQuit)
self.bind_all("<Shift-KeyPress-B>", self.keybdSwitchToBeta)
self.bind_all("<Shift-KeyPress-I>", self.keybdSwitchToBimodal)
self.bind_all("<Shift-KeyPress-E>", self.keybdSwitchToExpon)
# create the canvas
# func_to_plot should return density, not log density
self.plotter = Plotter(parent=self, func_to_plot=self.d.getPDF)
self.plotter.pack(side=TOP, expand=YES, fill=BOTH)
# create a status bar
self.status_label = Label(self, justify=LEFT, relief=SUNKEN, height=1, anchor=W, text="Ready")
self.status_label.pack(side=TOP, expand=NO, fill=X)
self.switchToBimodal()
class SAMCDemo(Frame):
def __init__(self, parent=None):
Frame.__init__(self, parent, bg="yellow")
self.pack(expand=YES, fill=BOTH)
# number of points to sample when manySteps function called
self.manyStepsN = 1000
# initialize the density from which to sample
self.d = MyBeta()
# initialize data members related to MCMC proposal
# delta is the width of the proposal window centered
# around current value
self.delta = window_width
self.x = 0.5 # will be reset in rescalePlot
# initialize SAMC-related data members
self.using_samc = True
self.biased_pi = True
self.m = None
self.levels = None
self.loglevels = None
self.theta = None
self.true_pi = None
self.est_pi = None
self.sample_size = None
self.gain = 1.0
self.t0 = 10.0
self.limit_log_theta = 100.0 * math.log(10.0)
# initialize miscellaneous data members
self._step_num_within_step = 0
self._just_point = False
self.plotter = None
self.n_func_evals = 0
# set up pseuodorandom number generator
self.r = Lot()
# self.r.setSeed(13579)
print "seed is =", self.r.getSeed()
# create a frame to hold the menu buttons
menuf = Frame(self, bg="magenta")
menuf.pack(expand=NO, fill=X)
# create the File menu button
self.filemb = Menubutton(menuf, text="File", relief=RAISED, anchor=W, borderwidth=0)
self.filemb.pack(expand=NO, fill=X, side=LEFT)
self.filemb.menu = Menu(self.filemb, tearoff=0)
self.filemb["menu"] = self.filemb.menu
self.filemb.menu.add_command(label="Quit", command=self.quit)
# create the Sample menu button
self.samplemb = Menubutton(menuf, text="Sample", relief=RAISED, anchor=W, borderwidth=0)
self.samplemb.pack(expand=YES, fill=X, side=LEFT)
self.samplemb.menu = Menu(self.samplemb, tearoff=1)
self.samplemb["menu"] = self.samplemb.menu
self.samplemb.menu.add_command(label="Toggle between SAMC and MCMC (s)", command=self.toggleSAMC)
self.samplemb.menu.add_command(label="One step (n)", command=self.oneStep)
self.samplemb.menu.add_command(label="Many steps (m)", command=self.manySteps)
self.samplemb.menu.add_command(label="Many steps using slice sampler (M)", command=self.manyStepsSliceSampler)
self.samplemb.menu.add_command(label="Biased SAMC (b)", command=self.biasedSAMC)
self.samplemb.menu.add_command(label="Unbiased SAMC (u)", command=self.unbiasedSAMC)
self.samplemb.menu.add_command(label="Reset (r)", command=self.reset)
self.samplemb.menu.add_command(label="Use Beta (B)", command=self.switchToBeta)
self.samplemb.menu.add_command(label="Use Bimodal (I)", command=self.switchToBimodal)
self.samplemb.menu.add_command(label="Use Exponential (E)", command=self.switchToExpon)
# bind some keys to the application (doesn't matter which widget has focus
# when you use bind_all)
self.bind_all("<KeyPress-s>", self.keybdToggleSAMC)
self.bind_all("<KeyPress-n>", self.keybdOneStep)
self.bind_all("<KeyPress-m>", self.keybdManySteps)
self.bind_all("<Shift-KeyPress-M>", self.keybdManyStepsSliceSampler)
self.bind_all("<KeyPress-b>", self.keybdBiasedSAMC)
self.bind_all("<KeyPress-u>", self.keybdUnbiasedSAMC)
self.bind_all("<KeyPress-r>", self.keybdReset)
self.bind_all("<KeyPress-q>", self.keybdQuit)
self.bind_all("<Shift-KeyPress-B>", self.keybdSwitchToBeta)
self.bind_all("<Shift-KeyPress-I>", self.keybdSwitchToBimodal)
self.bind_all("<Shift-KeyPress-E>", self.keybdSwitchToExpon)
# create the canvas
# func_to_plot should return density, not log density
self.plotter = Plotter(parent=self, func_to_plot=self.d.getPDF)
self.plotter.pack(side=TOP, expand=YES, fill=BOTH)
# create a status bar
self.status_label = Label(self, justify=LEFT, relief=SUNKEN, height=1, anchor=W, text="Ready")
self.status_label.pack(side=TOP, expand=NO, fill=X)
self.switchToBimodal()
def getLevelIndex(self, logf):
"""
Suppose f = 0.6 (but note that log(f), not f, should be passed into this function), and
supposing also that these are the levels:
level 4 (= m - 1)
-------- 1.00
level 3
-------- 0.75
level 2 <-- 0.6
-------- 0.50
level 1
-------- 0.25
level 0
then this function would return 2.
"""
for i in range(self.m - 1):
if logf < self.loglevels[i]:
return i
return self.m - 1
def takeStep(self, suppress_repaint=False):
"""
This function proposes a new step and shows the results as a point on the plot.
"""
# propose a new value for self.x
u = self.r.uniform()
x0 = (self.x - self.delta / 2.0) + self.delta * u
if x0 < 0.0:
x0 = -x0
logf0 = self.d.getLnPDF(x0)
logf = self.d.getLnPDF(self.x)
logR = logf0 - logf
# adjust logR if using SAMC
x_index = self.getLevelIndex(logf)
x0_index = self.getLevelIndex(logf0)
if self.using_samc:
logR += self.theta[x_index] - self.theta[x0_index]
logu = math.log(self.r.uniform())
if logu < logR:
self.x = x0
logf = logf0
x_index = x0_index
self.est_pi[x0_index] += 1
else:
self.est_pi[x_index] += 1
self.sample_size += 1
# update weights if using SAMC
for i in range(self.m):
if x_index == i:
self.theta[i] += self.gain * (1.0 - self.true_pi[i])
else:
self.theta[i] += self.gain * (0.0 - self.true_pi[i])
# point (x,y) is the point ultimately chosen along the horizontal slice
x = self.x
fx = self.d.getPDF(self.x)
u = self.r.uniform()
if self.using_samc:
if (x_index == self.m - 1) or (self.levels[x_index] > fx):
y_hi = fx
else:
y_hi = self.levels[x_index]
if x_index > 0:
y_lo = self.levels[x_index - 1]
else:
y_lo = 0.0
y = y_lo + (y_hi - y_lo) * u
else:
y = fx * u
self.plotter.points.append((x, y))
if not suppress_repaint:
self.plotter.repaint()
def _getSAMCLnPDF(self, x):
"""
Computes the log of the weighted PDF. For use when using slice sampler to
sample from the SAMC target distribution.
"""
logf = self.d.getLnPDF(x)
if self.using_samc:
x_index = self.getLevelIndex(logf)
q = self.theta[x_index]
else:
q = 0.0
return logf - q
def takeStepSliceSampler(self, suppress_repaint=False):
"""
This function uses the slice sampler to draw a new value and shows the
results as a point on the plot.
"""
# propose a new value for self.x
self.x = self.s.sample()
logf = self.d.getLnPDF(self.x)
# update weights and counts if using SAMC
if self.using_samc:
x_index = self.getLevelIndex(logf)
# update counts
self.est_pi[x_index] += 1
self.sample_size += 1
# update weights
for i in range(self.m):
if x_index == i:
self.theta[i] += self.gain * (1.0 - self.true_pi[i])
else:
self.theta[i] += self.gain * (0.0 - self.true_pi[i])
# point (x,y) is the point ultimately chosen along the horizontal slice
fx = self.d.getPDF(self.x)
u = self.r.uniform()
if self.using_samc:
if (x_index == self.m - 1) or (self.levels[x_index] > fx):
y_hi = fx
else:
y_hi = self.levels[x_index]
if x_index > 0:
y_lo = self.levels[x_index - 1]
else:
y_lo = 0.0
y = y_lo + (y_hi - y_lo) * u
else:
y = fx * u
self.plotter.points.append((self.x, y))
if not suppress_repaint:
self.plotter.repaint()
def toggleSAMC(self):
if self.using_samc:
self.modifyStatus("Switched to MCMC")
self.using_samc = False
else:
self.modifyStatus("Switched to SAMC")
self.using_samc = True
self.reset()
def keybdToggleSAMC(self, event):
self.toggleSAMC()
def oneStep(self):
self.takeStep()
self.modifyStatus("Took 1 step (%d total steps so far)" % self.sample_size)
def keybdOneStep(self, event):
self.oneStep()
def biasedSAMC(self):
self.modifyStatus("Lowest energy level will now be visited twice as often as other levels")
self.using_samc = True
self.biased_pi = True
self.reset()
def keybdBiasedSAMC(self, event):
self.biasedSAMC()
def unbiasedSAMC(self):
self.modifyStatus("Each energy level will now be visited equally")
self.using_samc = True
self.biased_pi = False
self.reset()
def keybdUnbiasedSAMC(self, event):
self.unbiasedSAMC()
def switchToBeta(self):
self.modifyStatus("Switching to Beta(1.5) distribution")
self.d = MyBeta()
self.reset()
def switchToExpon(self):
self.modifyStatus("Switching to Exponential(2.0) distribution")
self.d = MyExpon()
self.reset()
def switchToBimodal(self):
self.modifyStatus("Switching to a distribution that is an equal mixture of Beta(2,19) and Beta(19,2)")
self.d = MyBimodal()
self.reset()
def keybdSwitchToBeta(self, event):
self.switchToBeta()
def keybdSwitchToBimodal(self, event):
self.switchToBimodal()
def keybdSwitchToExpon(self, event):
self.switchToExpon()
def manySteps(self):
self.modifyStatus("Patience...")
for x in range(self.manyStepsN):
self.takeStep(suppress_repaint=True)
self.plotter.repaint()
self.modifyStatus("Took %d steps (%d total steps so far)" % (self.manyStepsN, self.sample_size))
def keybdManySteps(self, event):
self.manySteps()
def manyStepsSliceSampler(self):
self.modifyStatus("Patience...")
for x in range(self.manyStepsN):
self.takeStepSliceSampler(suppress_repaint=True)
self.plotter.repaint()
self.modifyStatus(
"Took %d steps using slice sampler (%d total steps so far)" % (self.manyStepsN, self.sample_size)
)
def keybdManyStepsSliceSampler(self, event):
self.manyStepsSliceSampler()
# def reset(self):
# self.x = starting_x
# self.modifyStatus('Ready')
# self.initLevels()
# if self.plotter:
# try:
# pw, ph, plotm = self.plotter.plotw, self.plotter.ploth, self.plotter.plotm
# except:
# pw, ph, plotm = 800, 600, 40
# self.plotter.resize(pw, ph, plotm)
# self.plotter.reset()
def keybdReset(self, event):
self.reset()
# next two functions cause the application to terminate
def quit(self):
Frame.quit(self)
def keybdQuit(self, event):
self.quit()
def modifyStatus(self, msg):
self.status_label.config(text=msg)
self.status_label.update_idletasks()
def initLevels(self, ymin, ymax, nlevels):
"""
If self.biased_pi is True, lowest energy level is made twice as probable
as all other energy levels. If self.biased_pi is False, all energy levels
have the same probability.
"""
# set up energy levels for this density
# boundary between level 0 and 1 is log(ymin)
# other boundaries divided evenly between log(ymin) and log(ymax)
log_ymax = math.log(ymax)
log_ymin = math.log(ymin)
yincr = (log_ymax - log_ymin) / float(num_levels - 1)
self.d.energy_levels = []
for i in range(num_levels - 1):
self.d.energy_levels.append(log_ymin + yincr * float(i))
self.loglevels = self.d.energy_levels
self.m = len(self.loglevels) + 1
self.levels = [math.exp(v) for v in self.loglevels]
self.theta = [0.0] * self.m
if self.biased_pi:
self.true_pi = [1.0] * self.m
self.true_pi[0] = 2.0
denom = sum(self.true_pi)
self.true_pi = [p / denom for p in self.true_pi]
else:
self.true_pi = [1.0 / float(self.m)] * self.m
self.est_pi = [0] * self.m
self.sample_size = 0
def reset(self):
"""
Called after switching the probabilty density to explore.
"""
# self.plotter.func should be function that returns a probability density,
# not a log density
self.plotter.func = self.d.getPDF
# self.f and self.ff should be functions that return the natural log of a
# probability density, or the natural log of a function proportional to a
# probability density
# self.f = AdHocDensity(self.d.getLnPDF)
self.f = AdHocDensity(self._getSAMCLnPDF)
# create a slice sampler
self.s = SliceSampler(self.r, self.f)
# create the status label
self.modifyStatus("Ready")
# grab a few samples to use in scaling the plot (number to generate is passed
# to the rescalePlot function)
self.rescalePlot(1000)
# reset statistics such as average slice width, number of function evalutations, etc.
w = self.s.getSliceUnitWidth()
self.s.setSliceUnitWidth(w / 3.0)
self.s.resetDiagnostics()
def rescalePlot(self, n):
"""
The rescalePlot function draws n samples from the target distribution to determine
the approximate limits of the main part of the density function. It then sets
the plotter's x and y axis ranges accordingly
"""
# temporarily turn off samc for purposes of rescaling the plot
prev_using_samc = self.using_samc
self.using_samc = False
smallest = getEffectiveLnZero()
largest = -smallest
ymin = 0.0 # always zero
ymin_not_zero = None # smallest y value seen that is greater than zero
ymax = smallest
xmin = largest
xmax = smallest
for i in range(n):
x = self.s.sample()
if x < xmin:
xmin = x
if x > xmax:
xmax = x
y = math.exp(self.f(x))
if y > ymax:
ymax = y
if (ymin_not_zero is None) or (y > 0.0 and y < ymin_not_zero):
ymin_not_zero = y
# restore using_samc setting
self.using_samc = prev_using_samc
# start x off at a random point between xmin and xmax
self.x = xmin + (xmax - xmin) * self.r.uniform()
# store the energy levels for this distribution
self.initLevels(ymin_not_zero, ymax, num_levels)
# make sure plot is wide enough to accommodate the worst-case scenario
# in terms of positioning the initial interval of width w
w = self.s.getSliceUnitWidth()
self.plotter.setXRange(xmin - w, xmax + w, (xmax - xmin + 2 * w) / 5.0)
self.plotter.setYRange(ymin, ymax, (ymax - ymin) / 5.0)
try:
pw, ph, plotm = self.plotter.plotw, self.plotter.ploth, self.plotter.plotm
except:
pw, ph, plotm = 800, 600, 40
self.plotter.resize(pw, ph, plotm)
self.plotter.reset()
