# This program is in the public domain
# Author: Paul Kienzle
"""
Parameter bounds and prior probabilities.
Parameter bounds encompass several features of our optimizers.
First and most trivially they allow for bounded constraints on
parameter values.
Secondly, for parameter values known to follow some distribution,
the bounds encodes a penalty function as the value strays from
its nominal value. Using a negative log likelihood cost function
on the fit, then this value naturally contributes to the overall
likelihood measure.
Predefined bounds are::
Unbounded
range (-inf, inf)
BoundedBelow
range (base, inf)
BoundedAbove
range (-inf, base)
Bounded
range (low, high)
Normal
range (-inf, inf) with gaussian probability
BoundedNormal
range (low, high) with gaussian probability within
SoftBounded
range (low, high) with gaussian probability outside
New bounds can be defined following the abstract base class
interface defined in :class:`Bounds`, or using Distribution(rv)
where rv is a scipy.stats continuous distribution.
For generating bounds given a value, we provide a few helper
functions::
v +/- d: pm(x,dx) or pm(x,-dm,+dp) or pm(x,+dp,-dm)
return (x-dm,x+dm) limited to 2 significant digits
v +/- p%: pmp(x,p) or pmp(x,-pm,+pp) or pmp(x,+pp,-pm)
return (x-pm*x/100, x+pp*x/100) limited to 2 sig. digits
pm_raw(x,dx) or raw_pm(x,-dm,+dp) or raw_pm(x,+dp,-dm)
return (x-dm,x+dm)
pmp_raw(x,p) or raw_pmp(x,-pm,+pp) or raw_pmp(x,+pp,-pm)
return (x-pm*x/100, x+pp*x/100)
nice_range(lo,hi)
return (lo,hi) limited to 2 significant digits
"""
from __future__ import division
__all__ = ['pm', 'pmp', 'pm_raw', 'pmp_raw', 'nice_range', 'init_bounds',
'Bounds', 'Unbounded', 'Bounded', 'BoundedAbove', 'BoundedBelow',
'Distribution', 'Normal', 'BoundedNormal', 'SoftBounded']
import math
from math import log, log10, sqrt, pi, ceil, floor
from numpy import inf, isinf, isfinite, clip
import numpy.random as RNG
try:
from scipy.stats import norm as normal_distribution
except ImportError:
# Normal distribution is an optional dependency. Leave it as a runtime
# failure if it doesn't exist.
pass
[docs]def pm(v, *args):
"""
Return the tuple (~v-dv,~v+dv), where ~expr is a 'nice' number near to
to the value of expr. For example::
>>> r = pm(0.78421, 0.0023145)
>>> print("%g - %g"%r)
0.7818 - 0.7866
If called as pm(value, +dp, -dm) or pm(value, -dm, +dp),
return (~v-dm, ~v+dp).
"""
return nice_range(pm_raw(v, *args))
[docs]def pmp(v, *args):
"""
Return the tuple (~v-%v,~v+%v), where ~expr is a 'nice' number near to
the value of expr. For example::
>>> r = pmp(0.78421, 10)
>>> print("%g - %g"%r)
0.7 - 0.87
>>> r = pmp(0.78421, 0.1)
>>> print("%g - %g"%r)
0.7834 - 0.785
If called as pmp(value, +pp, -pm) or pmp(value, -pm, +pp),
return (~v-pm%v, ~v+pp%v).
"""
return nice_range(pmp_raw(v, *args))
# Generate ranges using x +/- dx or x +/- p%*x
[docs]def pm_raw(v, *args):
"""
Return the tuple [v-dv,v+dv].
If called as pm_raw(value, +dp, -dm) or pm_raw(value, -dm, +dp),
return (v-dm, v+dp).
"""
if len(args) == 1:
dv = args[0]
return v - dv, v + dv
elif len(args) == 2:
plus, minus = args
if plus < minus:
plus, minus = minus, plus
# if minus > 0 or plus < 0:
# raise TypeError("pm(value, p1, p2) requires both + and - values")
return v + minus, v + plus
else:
raise TypeError("pm(value, delta) or pm(value, -p1, +p2)")
[docs]def pmp_raw(v, *args):
"""
Return the tuple [v-%v,v+%v]
If called as pmp_raw(value, +pp, -pm) or pmp_raw(value, -pm, +pp),
return (v-pm%v, v+pp%v).
"""
if len(args) == 1:
percent = args[0]
b1, b2 = v * (1 - 0.01 * percent), v * (1 + 0.01 * percent)
elif len(args) == 2:
plus, minus = args
if plus < minus:
plus, minus = minus, plus
# if minus > 0 or plus < 0:
# raise TypeError("pmp(value, p1, p2) requires both + and - values")
b1, b2 = v * (1 + 0.01 * minus), v * (1 + 0.01 * plus)
else:
raise TypeError("pmp(value, delta) or pmp(value, -p1, +p2)")
return (b1, b2) if v > 0 else (b2, b1)
[docs]def nice_range(bounds):
"""
Given a range, return an enclosing range accurate to two digits.
"""
step = bounds[1] - bounds[0]
if step > 0:
d = 10 ** (floor(log10(step)) - 1)
return floor(bounds[0]/d)*d, ceil(bounds[1]/d)*d
else:
return bounds
[docs]def init_bounds(v):
"""
Returns a bounds object of the appropriate type given the arguments.
This is a helper factory to simplify the user interface to parameter
objects.
"""
# if it is none, then it is unbounded
if v is None:
return Unbounded()
# if it isn't a tuple, assume it is a bounds type.
try:
lo, hi = v
except TypeError:
return v
# if it is a tuple, then determine what kind of bounds we have
if lo is None:
lo = -inf
if hi is None:
hi = inf
# TODO: consider issuing a warning instead of correcting reversed bounds
if lo >= hi:
lo, hi = hi, lo
if isinf(lo) and isinf(hi):
return Unbounded()
elif isinf(lo):
return BoundedAbove(hi)
elif isinf(hi):
return BoundedBelow(lo)
else:
return Bounded(lo, hi)
[docs]class Bounds(object):
"""
Bounds abstract base class.
A range is used for several purposes. One is that it transforms parameters
between unbounded and bounded forms depending on the needs of the optimizer.
Another is that it generates random values in the range for stochastic
optimizers, and for initialization.
A third is that it returns the likelihood of seeing that particular value
for optimizers which use soft constraints. Assuming the cost function that
is being optimized is also a probability, then this is an easy way to
incorporate information from other sorts of measurements into the model.
"""
limits = (-inf, inf)
# TODO: need derivatives wrt bounds transforms
[docs] def get01(self, x):
"""
Convert value into [0,1] for optimizers which are bounds constrained.
This can also be used as a scale bar to show approximately how close to
the end of the range the value is.
"""
[docs] def put01(self, v):
"""
Convert [0,1] into value for optimizers which are bounds constrained.
"""
[docs] def getfull(self, x):
"""
Convert value into (-inf,inf) for optimizers which are unconstrained.
"""
[docs] def putfull(self, v):
"""
Convert (-inf,inf) into value for optimizers which are unconstrained.
"""
[docs] def random(self, n=1, target=1.0):
"""
Return a randomly generated valid value.
*target* gives some scale independence to the random number
generator, allowing the initial value of the parameter to influence
the randomly generated value. Otherwise fits without bounds have
too large a space to search through.
"""
[docs] def nllf(self, value):
"""
Return the negative log likelihood of seeing this value, with
likelihood scaled so that the maximum probability is one.
For uniform bounds, this either returns zero or inf. For bounds
based on a probability distribution, this returns values between
zero and inf. The scaling is necessary so that indefinite and
semi-definite ranges return a sensible value. The scaling does
not affect the likelihood maximization process, though the resulting
likelihood is not easily interpreted.
"""
[docs] def residual(self, value):
"""
Return the parameter 'residual' in a way that is consistent with
residuals in the normal distribution. The primary purpose is to
graphically display exceptional values in a way that is familiar
to the user. For fitting, the scaled likelihood should be used.
To do this, we will match the cumulative density function value
with that for N(0,1) and find the corresponding percent point
function from the N(0,1) distribution. In this way, for example,
a value to the right of 2.275% of the distribution would correspond
to a residual of -2, or 2 standard deviations below the mean.
For uniform distributions, with all values equally probable, we
use a value of +/-4 for values outside the range, and 0 for values
inside the range.
"""
[docs] def start_value(self):
"""
Return a default starting value if none given.
"""
return self.put01(0.5)
def __contains__(self, v):
return self.limits[0] <= v <= self.limits[1]
def __str__(self):
limits = tuple(num_format(v) for v in self.limits)
return "(%s,%s)" % limits
# CRUFT: python 2.5 doesn't format indefinite numbers properly on windows
def num_format(v):
"""
Number formating which supports inf/nan on windows.
"""
if isfinite(v):
return "%g" % v
elif isinf(v):
return "inf" if v > 0 else "-inf"
else:
return "NaN"
[docs]class Unbounded(Bounds):
"""
Unbounded parameter.
The random initial condition is assumed to be between 0 and 1
The probability is uniformly 1/inf everywhere, which means the negative
log likelihood of P is inf everywhere. A value inf will interfere
with optimization routines, and so we instead choose P == 1 everywhere.
"""
[docs] def random(self, n=1, target=1.0):
scale = target + (target == 0.)
return RNG.randn(n)*scale
[docs] def nllf(self, value):
return 0
[docs] def residual(self, value):
return 0
[docs] def get01(self, x):
return _get01_inf(x)
[docs] def put01(self, v):
return _put01_inf(v)
[docs] def getfull(self, x):
return x
[docs] def putfull(self, v):
return v
[docs]class BoundedBelow(Bounds):
"""
Semidefinite range bounded below.
The random initial condition is assumed to be within 1 of the maximum.
[base,inf] <-> (-inf,inf) is direct above base+1, -1/(x-base) below
[base,inf] <-> [0,1] uses logarithmic compression.
Logarithmic compression works by converting sign*m*2^e+base to
sign*(e+1023+m), yielding a value in [0,2048]. This can then be
converted to a value in [0,1].
Note that the likelihood function is problematic: the true probability
of seeing any particular value in the range is infinitesimal, and that
is indistinguishable from values outside the range. Instead we say
that P = 1 in range, and 0 outside.
"""
def __init__(self, base):
self.limits = (base, inf)
self._base = base
[docs] def start_value(self):
return self._base + 1
[docs] def random(self, n=1, target=1.):
target = max(abs(target), abs(self._base))
scale = target + (target == 0.)
return self._base + abs(RNG.randn(n)*scale)
[docs] def nllf(self, value):
return 0 if value >= self._base else inf
[docs] def residual(self, value):
return 0 if value >= self._base else -4
[docs] def get01(self, x):
m, e = math.frexp(x - self._base)
if m >= 0 and e <= _E_MAX:
v = (e + m) / (2. * _E_MAX)
return v
else:
return 0 if m < 0 else 1
[docs] def put01(self, v):
v = v * 2 * _E_MAX
e = int(v)
m = v - e
x = math.ldexp(m, e) + self._base
return x
[docs] def getfull(self, x):
v = x - self._base
return v if v >= 1 else 2 - 1. / v
[docs] def putfull(self, v):
x = v if v >= 1 else 1. / (2 - v)
return x + self._base
[docs]class BoundedAbove(Bounds):
"""
Semidefinite range bounded above.
[-inf,base] <-> [0,1] uses logarithmic compression
[-inf,base] <-> (-inf,inf) is direct below base-1, 1/(base-x) above
Logarithmic compression works by converting sign*m*2^e+base to
sign*(e+1023+m), yielding a value in [0,2048]. This can then be
converted to a value in [0,1].
Note that the likelihood function is problematic: the true probability
of seeing any particular value in the range is infinitesimal, and that
is indistinguishable from values outside the range. Instead we say
that P = 1 in range, and 0 outside.
"""
def __init__(self, base):
self.limits = (-inf, base)
self._base = base
[docs] def start_value(self):
return self._base - 1
[docs] def random(self, n=1, target=1.0):
target = max(abs(self._base), abs(target))
scale = target + (target == 0.)
return self._base - abs(RNG.randn(n)*scale)
[docs] def nllf(self, value):
return 0 if value <= self._base else inf
[docs] def residual(self, value):
return 0 if value <= self._base else 4
[docs] def get01(self, x):
m, e = math.frexp(self._base - x)
if m >= 0 and e <= _E_MAX:
v = (e + m) / (2. * _E_MAX)
return 1 - v
else:
return 1 if m < 0 else 0
[docs] def put01(self, v):
v = (1 - v) * 2 * _E_MAX
e = int(v)
m = v - e
x = -(math.ldexp(m, e) - self._base)
return x
[docs] def getfull(self, x):
v = x - self._base
return v if v <= -1 else -2 - 1. / v
[docs] def putfull(self, v):
x = v if v <= -1 else -1. / (v + 2)
return x + self._base
[docs]class Bounded(Bounds):
"""
Bounded range.
[lo,hi] <-> [0,1] scale is simple linear
[lo,hi] <-> (-inf,inf) scale uses exponential expansion
While technically the probability of seeing any value within the
range is 1/range, for consistency with the semi-infinite ranges
and for a more natural mapping between nllf and chisq, we instead
set the probability to 0. This choice will not affect the fits.
"""
def __init__(self, lo, hi):
self.limits = (lo, hi)
self._nllf_scale = log(hi - lo)
[docs] def random(self, n=1, target=1.0):
lo, hi = self.limits
#print("= uniform",lo,hi)
return RNG.uniform(lo, hi, size=n)
[docs] def nllf(self, value):
lo, hi = self.limits
return 0 if lo <= value <= hi else inf
# return self._nllf_scale if lo<=value<=hi else inf
[docs] def residual(self, value):
lo, hi = self.limits
return -4 if lo > value else (4 if hi < value else 0)
[docs] def get01(self, x):
lo, hi = self.limits
return float(x - lo) / (hi - lo) if hi - lo > 0 else 0
[docs] def put01(self, v):
lo, hi = self.limits
return (hi - lo) * v + lo
[docs] def getfull(self, x):
return _put01_inf(self.get01(x))
[docs] def putfull(self, v):
return self.put01(_get01_inf(v))
[docs]class Distribution(Bounds):
"""
Parameter is pulled from a distribution.
*dist* must implement the distribution interface from scipy.stats.
In particular, it should define methods rvs, nnlf, cdf and ppf and
attributes args and dist.name.
"""
def __init__(self, dist):
self.dist = dist
[docs] def random(self, n=1, target=1.0):
return self.dist.rvs(n)
[docs] def nllf(self, value):
return -log(self.dist.pdf(value))
[docs] def residual(self, value):
return normal_distribution.ppf(self.dist.cdf(value))
[docs] def get01(self, x):
return self.dist.cdf(x)
[docs] def put01(self, v):
return self.dist.ppf(v)
[docs] def getfull(self, x):
return x
[docs] def putfull(self, v):
return v
def __getstate__(self):
# WARNING: does not preserve and restore seed
return self.dist.__class__, self.dist.args, self.dist.kwds
def __setstate__(self, state):
cls, args, kwds = state
self.dist = cls(*args, **kwds)
def __str__(self):
return "%s(%s)" % (self.dist.dist.name,
",".join(str(s) for s in self.dist.args))
[docs]class Normal(Distribution):
"""
Parameter is pulled from a normal distribution.
If you have measured a parameter value with some uncertainty (e.g., the
film thickness is 35+/-5 according to TEM), then you can use this
measurement to restrict the values given to the search, and to penalize
choices of this fitting parameter which are different from this value.
*mean* is the expected value of the parameter and *std* is the 1-sigma
standard deviation.
"""
def __init__(self, mean=0, std=1):
Distribution.__init__(self, normal_distribution(mean, std))
self._nllf_scale = log(sqrt(2 * pi * std ** 2))
[docs] def nllf(self, value):
# P(v) = exp(-0.5*(v-mean)**2/std**2)/sqrt(2*pi*std**2)
# -log(P(v)) = -(-0.5*(v-mean)**2/std**2 - log( (2*pi*std**2) ** 0.5))
# = 0.5*(v-mean)**2/std**2 + log(2*pi*std**2)/2
mean, std = self.dist.args
return 0.5 * ((value-mean)/std)**2 + self._nllf_scale
[docs] def residual(self, value):
mean, std = self.dist.args
return (value-mean)/std
def __getstate__(self):
return self.dist.args # args is mean,std
def __setstate__(self, state):
mean, std = state
self.__init__(mean=mean, std=std)
[docs]class BoundedNormal(Bounds):
"""
truncated normal bounds
"""
def __init__(self, sigma=1, mu=0, limits=(-inf, inf)):
self.limits = limits
self.sigma, self.mu = sigma, mu
self._left = normal_distribution.cdf((limits[0]-mu)/sigma)
self._delta = normal_distribution.cdf((limits[1]-mu)/sigma) - self._left
self._nllf_scale = log(sqrt(2 * pi * sigma ** 2)) + log(self._delta)
[docs] def get01(self, x):
"""
Convert value into [0,1] for optimizers which are bounds constrained.
This can also be used as a scale bar to show approximately how close to
the end of the range the value is.
"""
v = ((normal_distribution.cdf((x-self.mu)/self.sigma) - self._left)
/ self._delta)
return clip(v, 0, 1)
[docs] def put01(self, v):
"""
Convert [0,1] into value for optimizers which are bounds constrained.
"""
x = v * self._delta + self._left
return normal_distribution.ppf(x) * self.sigma + self.mu
[docs] def getfull(self, x):
"""
Convert value into (-inf,inf) for optimizers which are unconstrained.
"""
raise NotImplementedError
[docs] def putfull(self, v):
"""
Convert (-inf,inf) into value for optimizers which are unconstrained.
"""
raise NotImplementedError
[docs] def random(self, n=1, target=1.0):
"""
Return a randomly generated valid value, or an array of values
"""
return self.get01(RNG.rand(n))
[docs] def nllf(self, value):
"""
Return the negative log likelihood of seeing this value, with
likelihood scaled so that the maximum probability is one.
"""
if value in self:
return 0.5 * ((value-self.mu)/self.sigma)**2 + self._nllf_scale
else:
return inf
[docs] def residual(self, value):
"""
Return the parameter 'residual' in a way that is consistent with
residuals in the normal distribution. The primary purpose is to
graphically display exceptional values in a way that is familiar
to the user. For fitting, the scaled likelihood should be used.
For the truncated normal distribution, we can just use the normal
residuals.
"""
return (value - self.mu) / self.sigma
[docs] def start_value(self):
"""
Return a default starting value if none given.
"""
return self.put01(0.5)
def __contains__(self, v):
return self.limits[0] <= v <= self.limits[1]
def __str__(self):
vals = (
self.limits[0], self.limits[1],
self.mu, self.sigma,
)
return "(%s,%s), norm(%s,%s)" % tuple(num_format(v) for v in vals)
[docs]class SoftBounded(Bounds):
"""
Parameter is pulled from a stretched normal distribution.
This is like a rectangular distribution, but with gaussian tails.
The intent of this distribution is for soft constraints on the values.
As such, the random generator will return values like the rectangular
distribution, but the likelihood will return finite values based on
the distance from the from the bounds rather than returning infinity.
Note that for bounds constrained optimizers which force the value
into the range [0,1] for each parameter we don't need to use soft
constraints, and this acts just like the rectangular distribution.
"""
def __init__(self, lo, hi, std=None):
self._lo, self._hi, self._std = lo, hi, std
self._nllf_scale = log(hi - lo + sqrt(2 * pi * std))
[docs] def random(self, n=1, target=1.0):
return RNG.uniform(self._lo, self._hi, size=n)
[docs] def nllf(self, value):
# To turn f(x) = 1 if x in [lo,hi] else G(tail)
# into a probability p, we need to normalize by \int{f(x)dx},
# which is just hi-lo + sqrt(2*pi*std**2).
if value < self._lo:
z = self._lo - value
elif value > self._hi:
z = value - self._hi
else:
z = 0
return (z / self._std) ** 2 / 2 + self._nllf_scale
[docs] def residual(self, value):
if value < self._lo:
z = self._lo - value
elif value > self._hi:
z = value - self._hi
else:
z = 0
return z / self._std
[docs] def get01(self, x):
v = float(x - self._lo) / (self._hi - self._lo)
return v if 0 <= v <= 1 else (0 if v < 0 else 1)
[docs] def put01(self, v):
return v * (self._hi - self._lo) + self._lo
[docs] def getfull(self, x):
return x
[docs] def putfull(self, v):
return v
def __str__(self):
return "box_norm(%g,%g,sigma=%g)" % (self._lo, self._hi, self._std)
_E_MIN = -1023
_E_MAX = 1024
def _get01_inf(x):
"""
Convert a floating point number to a value in [0,1].
The value sign*m*2^e to sign*(e+1023+m), yielding a value in [-2048,2048].
This can then be converted to a value in [0,1].
Sort order is preserved. At least 14 bits of precision are lost from
the 53 bit mantissa.
"""
# Arctan alternative
# Arctan is approximately linear in (-0.5, 0.5), but the
# transform is only useful up to (-10**15,10**15).
# return atan(x)/pi + 0.5
m, e = math.frexp(x)
s = math.copysign(1.0, m)
v = (e - _E_MIN + m * s) * s
v = v / (4 * _E_MAX) + 0.5
v = 0 if _E_MIN > e else (1 if _E_MAX < e else v)
return v
def _put01_inf(v):
"""
Convert a value in [0,1] to a full floating point number.
Sort order is preserved. Reverses :func:`_get01_inf`, but with fewer
bits of precision.
"""
# Arctan alternative
# return tan(pi*(v-0.5))
v = (v - 0.5) * 4 * _E_MAX
s = math.copysign(1., v)
v *= s
e = int(v)
m = v - e
x = math.ldexp(s * m, e + _E_MIN)
# print "< x,e,m,s,v",x,e+_e_min,s*m,s,v
return x