You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

456 lines
14 KiB
Python

from __future__ import division, print_function, absolute_import
import os
import sys
import time
import numpy as np
from numpy.testing import assert_
import pytest
from scipy._lib.six import reraise
from scipy.special._testutils import assert_func_equal
try:
import mpmath
except ImportError:
pass
# ------------------------------------------------------------------------------
# Machinery for systematic tests with mpmath
# ------------------------------------------------------------------------------
class Arg(object):
"""Generate a set of numbers on the real axis, concentrating on
'interesting' regions and covering all orders of magnitude.
"""
def __init__(self, a=-np.inf, b=np.inf, inclusive_a=True, inclusive_b=True):
if a > b:
raise ValueError("a should be less than or equal to b")
if a == -np.inf:
a = -0.5*np.finfo(float).max
if b == np.inf:
b = 0.5*np.finfo(float).max
self.a, self.b = a, b
self.inclusive_a, self.inclusive_b = inclusive_a, inclusive_b
def _positive_values(self, a, b, n):
if a < 0:
raise ValueError("a should be positive")
# Try to put half of the points into a linspace between a and
# 10 the other half in a logspace.
if n % 2 == 0:
nlogpts = n//2
nlinpts = nlogpts
else:
nlogpts = n//2
nlinpts = nlogpts + 1
if a >= 10:
# Outside of linspace range; just return a logspace.
pts = np.logspace(np.log10(a), np.log10(b), n)
elif a > 0 and b < 10:
# Outside of logspace range; just return a linspace
pts = np.linspace(a, b, n)
elif a > 0:
# Linspace between a and 10 and a logspace between 10 and
# b.
linpts = np.linspace(a, 10, nlinpts, endpoint=False)
logpts = np.logspace(1, np.log10(b), nlogpts)
pts = np.hstack((linpts, logpts))
elif a == 0 and b <= 10:
# Linspace between 0 and b and a logspace between 0 and
# the smallest positive point of the linspace
linpts = np.linspace(0, b, nlinpts)
if linpts.size > 1:
right = np.log10(linpts[1])
else:
right = -30
logpts = np.logspace(-30, right, nlogpts, endpoint=False)
pts = np.hstack((logpts, linpts))
else:
# Linspace between 0 and 10, logspace between 0 and the
# smallest positive point of the linspace, and a logspace
# between 10 and b.
if nlogpts % 2 == 0:
nlogpts1 = nlogpts//2
nlogpts2 = nlogpts1
else:
nlogpts1 = nlogpts//2
nlogpts2 = nlogpts1 + 1
linpts = np.linspace(0, 10, nlinpts, endpoint=False)
if linpts.size > 1:
right = np.log10(linpts[1])
else:
right = -30
logpts1 = np.logspace(-30, right, nlogpts1, endpoint=False)
logpts2 = np.logspace(1, np.log10(b), nlogpts2)
pts = np.hstack((logpts1, linpts, logpts2))
return np.sort(pts)
def values(self, n):
"""Return an array containing n numbers."""
a, b = self.a, self.b
if a == b:
return np.zeros(n)
if not self.inclusive_a:
n += 1
if not self.inclusive_b:
n += 1
if n % 2 == 0:
n1 = n//2
n2 = n1
else:
n1 = n//2
n2 = n1 + 1
if a >= 0:
pospts = self._positive_values(a, b, n)
negpts = []
elif b <= 0:
pospts = []
negpts = -self._positive_values(-b, -a, n)
else:
pospts = self._positive_values(0, b, n1)
negpts = -self._positive_values(0, -a, n2 + 1)
# Don't want to get zero twice
negpts = negpts[1:]
pts = np.hstack((negpts[::-1], pospts))
if not self.inclusive_a:
pts = pts[1:]
if not self.inclusive_b:
pts = pts[:-1]
return pts
class FixedArg(object):
def __init__(self, values):
self._values = np.asarray(values)
def values(self, n):
return self._values
class ComplexArg(object):
def __init__(self, a=complex(-np.inf, -np.inf), b=complex(np.inf, np.inf)):
self.real = Arg(a.real, b.real)
self.imag = Arg(a.imag, b.imag)
def values(self, n):
m = int(np.floor(np.sqrt(n)))
x = self.real.values(m)
y = self.imag.values(m + 1)
return (x[:,None] + 1j*y[None,:]).ravel()
class IntArg(object):
def __init__(self, a=-1000, b=1000):
self.a = a
self.b = b
def values(self, n):
v1 = Arg(self.a, self.b).values(max(1 + n//2, n-5)).astype(int)
v2 = np.arange(-5, 5)
v = np.unique(np.r_[v1, v2])
v = v[(v >= self.a) & (v < self.b)]
return v
def get_args(argspec, n):
if isinstance(argspec, np.ndarray):
args = argspec.copy()
else:
nargs = len(argspec)
ms = np.asarray([1.5 if isinstance(spec, ComplexArg) else 1.0 for spec in argspec])
ms = (n**(ms/sum(ms))).astype(int) + 1
args = []
for spec, m in zip(argspec, ms):
args.append(spec.values(m))
args = np.array(np.broadcast_arrays(*np.ix_(*args))).reshape(nargs, -1).T
return args
class MpmathData(object):
def __init__(self, scipy_func, mpmath_func, arg_spec, name=None,
dps=None, prec=None, n=None, rtol=1e-7, atol=1e-300,
ignore_inf_sign=False, distinguish_nan_and_inf=True,
nan_ok=True, param_filter=None):
# mpmath tests are really slow (see gh-6989). Use a small number of
# points by default, increase back to 5000 (old default) if XSLOW is
# set
if n is None:
try:
is_xslow = int(os.environ.get('SCIPY_XSLOW', '0'))
except ValueError:
is_xslow = False
n = 5000 if is_xslow else 500
self.scipy_func = scipy_func
self.mpmath_func = mpmath_func
self.arg_spec = arg_spec
self.dps = dps
self.prec = prec
self.n = n
self.rtol = rtol
self.atol = atol
self.ignore_inf_sign = ignore_inf_sign
self.nan_ok = nan_ok
if isinstance(self.arg_spec, np.ndarray):
self.is_complex = np.issubdtype(self.arg_spec.dtype, np.complexfloating)
else:
self.is_complex = any([isinstance(arg, ComplexArg) for arg in self.arg_spec])
self.ignore_inf_sign = ignore_inf_sign
self.distinguish_nan_and_inf = distinguish_nan_and_inf
if not name or name == '<lambda>':
name = getattr(scipy_func, '__name__', None)
if not name or name == '<lambda>':
name = getattr(mpmath_func, '__name__', None)
self.name = name
self.param_filter = param_filter
def check(self):
np.random.seed(1234)
# Generate values for the arguments
argarr = get_args(self.arg_spec, self.n)
# Check
old_dps, old_prec = mpmath.mp.dps, mpmath.mp.prec
try:
if self.dps is not None:
dps_list = [self.dps]
else:
dps_list = [20]
if self.prec is not None:
mpmath.mp.prec = self.prec
# Proper casting of mpmath input and output types. Using
# native mpmath types as inputs gives improved precision
# in some cases.
if np.issubdtype(argarr.dtype, np.complexfloating):
pytype = mpc2complex
def mptype(x):
return mpmath.mpc(complex(x))
else:
def mptype(x):
return mpmath.mpf(float(x))
def pytype(x):
if abs(x.imag) > 1e-16*(1 + abs(x.real)):
return np.nan
else:
return mpf2float(x.real)
# Try out different dps until one (or none) works
for j, dps in enumerate(dps_list):
mpmath.mp.dps = dps
try:
assert_func_equal(self.scipy_func,
lambda *a: pytype(self.mpmath_func(*map(mptype, a))),
argarr,
vectorized=False,
rtol=self.rtol, atol=self.atol,
ignore_inf_sign=self.ignore_inf_sign,
distinguish_nan_and_inf=self.distinguish_nan_and_inf,
nan_ok=self.nan_ok,
param_filter=self.param_filter)
break
except AssertionError:
if j >= len(dps_list)-1:
reraise(*sys.exc_info())
finally:
mpmath.mp.dps, mpmath.mp.prec = old_dps, old_prec
def __repr__(self):
if self.is_complex:
return "<MpmathData: %s (complex)>" % (self.name,)
else:
return "<MpmathData: %s>" % (self.name,)
def assert_mpmath_equal(*a, **kw):
d = MpmathData(*a, **kw)
d.check()
def nonfunctional_tooslow(func):
return pytest.mark.skip(reason=" Test not yet functional (too slow), needs more work.")(func)
# ------------------------------------------------------------------------------
# Tools for dealing with mpmath quirks
# ------------------------------------------------------------------------------
def mpf2float(x):
"""
Convert an mpf to the nearest floating point number. Just using
float directly doesn't work because of results like this:
with mp.workdps(50):
float(mpf("0.99999999999999999")) = 0.9999999999999999
"""
return float(mpmath.nstr(x, 17, min_fixed=0, max_fixed=0))
def mpc2complex(x):
return complex(mpf2float(x.real), mpf2float(x.imag))
def trace_args(func):
def tofloat(x):
if isinstance(x, mpmath.mpc):
return complex(x)
else:
return float(x)
def wrap(*a, **kw):
sys.stderr.write("%r: " % (tuple(map(tofloat, a)),))
sys.stderr.flush()
try:
r = func(*a, **kw)
sys.stderr.write("-> %r" % r)
finally:
sys.stderr.write("\n")
sys.stderr.flush()
return r
return wrap
try:
import posix
import signal
POSIX = ('setitimer' in dir(signal))
except ImportError:
POSIX = False
class TimeoutError(Exception):
pass
def time_limited(timeout=0.5, return_val=np.nan, use_sigalrm=True):
"""
Decorator for setting a timeout for pure-Python functions.
If the function does not return within `timeout` seconds, the
value `return_val` is returned instead.
On POSIX this uses SIGALRM by default. On non-POSIX, settrace is
used. Do not use this with threads: the SIGALRM implementation
does probably not work well. The settrace implementation only
traces the current thread.
The settrace implementation slows down execution speed. Slowdown
by a factor around 10 is probably typical.
"""
if POSIX and use_sigalrm:
def sigalrm_handler(signum, frame):
raise TimeoutError()
def deco(func):
def wrap(*a, **kw):
old_handler = signal.signal(signal.SIGALRM, sigalrm_handler)
signal.setitimer(signal.ITIMER_REAL, timeout)
try:
return func(*a, **kw)
except TimeoutError:
return return_val
finally:
signal.setitimer(signal.ITIMER_REAL, 0)
signal.signal(signal.SIGALRM, old_handler)
return wrap
else:
def deco(func):
def wrap(*a, **kw):
start_time = time.time()
def trace(frame, event, arg):
if time.time() - start_time > timeout:
raise TimeoutError()
return trace
sys.settrace(trace)
try:
return func(*a, **kw)
except TimeoutError:
sys.settrace(None)
return return_val
finally:
sys.settrace(None)
return wrap
return deco
def exception_to_nan(func):
"""Decorate function to return nan if it raises an exception"""
def wrap(*a, **kw):
try:
return func(*a, **kw)
except Exception:
return np.nan
return wrap
def inf_to_nan(func):
"""Decorate function to return nan if it returns inf"""
def wrap(*a, **kw):
v = func(*a, **kw)
if not np.isfinite(v):
return np.nan
return v
return wrap
def mp_assert_allclose(res, std, atol=0, rtol=1e-17):
"""
Compare lists of mpmath.mpf's or mpmath.mpc's directly so that it
can be done to higher precision than double.
"""
try:
len(res)
except TypeError:
res = list(res)
n = len(std)
if len(res) != n:
raise AssertionError("Lengths of inputs not equal.")
failures = []
for k in range(n):
try:
assert_(mpmath.fabs(res[k] - std[k]) <= atol + rtol*mpmath.fabs(std[k]))
except AssertionError:
failures.append(k)
ndigits = int(abs(np.log10(rtol)))
msg = [""]
msg.append("Bad results ({} out of {}) for the following points:"
.format(len(failures), n))
for k in failures:
resrep = mpmath.nstr(res[k], ndigits, min_fixed=0, max_fixed=0)
stdrep = mpmath.nstr(std[k], ndigits, min_fixed=0, max_fixed=0)
if std[k] == 0:
rdiff = "inf"
else:
rdiff = mpmath.fabs((res[k] - std[k])/std[k])
rdiff = mpmath.nstr(rdiff, 3)
msg.append("{}: {} != {} (rdiff {})".format(k, resrep, stdrep, rdiff))
if failures:
assert_(False, "\n".join(msg))