class HistoTuple(EquivUnary):
'''Histogram returned as list of (value,count) tuples
'''
name = 'histo_tuple'
ranking = ('iter_groupby', 'dict1')
tests = (
Test([1, 2, 3, 4, 1, 2, 3, 1, 1, 1, 9]) == [(1, 5), (2, 2), (3, 2),
(4, 1), (9, 1)],
Test([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]) == [(1, 11)],
)
@staticmethod
def iter_groupby(arg):
arg = as_any_array(arg)
return sorted([(k, len(list(g))) for k, g in groupby(sorted(arg))])
@staticmethod
def dict1(arg, out=None):
arg = as_any_array(arg)
out = {}
for val in arg:
out[val] = out.get(val, 0) + 1
return sorted(out.items())
class DictMapDefault(EquivUnary):
'''apply dictionary mapping to each element in array, using default for unmatched
>>> func = DictMapDefault()
>>> mapping = {1:100,2:200,3:300,'a':'aaa','b':'bbb'}
>>> assert func([1,3,6,7,'a',9],mapping=mapping).tolist() == [100,300,6,7,'aaa',9]
'''
name = 'dict_map_default'
ranking = ('naive_comp', 'naive_loop',)
tests = (
Test([1,3,6,7,7,9],mapping={1:100,2:200,3:300}) == [100,300,0,0,0,0],
Test([1,3,6,7,'a',9],mapping={'1':100,'a':'aaa','3':300}) == ['100','300','0','0','aaa','0'],
)
@staticmethod
def naive_loop(arg,mapping={},default=0):
arg = as_any_array(arg)
out = []
for value in arg:
out.append(mapping.get(value,default))
return as_any_array(out)
@staticmethod
def naive_comp(arg,mapping={},default=0):
arg = as_any_array(arg)
out = [mapping.get(value,default) for value in arg]
return as_any_array(out)
class WindowRange(EquivUnary):
"""Get value range within sliding window on single data vector
"""
name = 'window_range'
ranking = ('naive1', 'naive2')
ranking = ('naive1', )
bench_sizes = (1, 2, 3, 4, 5, 10, 20, 50, 100, 200, 300, 500, 1000, 5000)
tests = (
Test([0, 1, 2, 3]) == [0, 0, 0],
Test([0, 1, 2, 3], step=2) == [0, 1, 1],
Test(arange(20), sel=[0, 3, 5, 9, 20]) == [2, 1, 3, 9],
)
@staticmethod
def naive1(arg, sel=None, step=1):
arg = as_num_array(arg)
idx = arg_sel_step_to_idx(arg, sel, step)
jj = idx[:-step]
kk = idx[step:]
segments = [arg[j + 1:k + 1] for j, k in izip(jj, kk) if j != k]
omin = [seg.min() for seg in segments if len(seg)]
omax = [seg.max() for seg in segments if len(seg)]
return as_num_array(omax) - omin
def naive2(arg, sel=None, step=1):
arg = as_num_array(arg)
idx = arg_sel_step_to_idx(arg, sel, step)
jj = idx[:-step]
kk = idx[step:]
omin = [arg[j + 1:k + 1].min() for j, k in izip(jj, kk)]
omax = [arg[j + 1:k + 1].max() for j, k in izip(jj, kk)]
return as_num_array(omax) - omin
class RoundAll(EquivUnary):
'''apply python round function (there is no equivalent ufunc)
>>> func = RoundAll()
>>> assert func([12.3456,30,456.12,0],ndigits=2).tolist() == [12.35,30.,456.12,0.]
'''
name = 'round_all'
ranking = ('roundint','naive')
tests = (
Test([12.3456,30,456.12,0],ndigits=-1) == [10,30,460,0],
Test([12.3456,30,456.12,0],ndigits=0) == [12,30,456,0],
Test([12.3456,30,456.12,0],ndigits=1) == [12.3,30,456.1,0],
Test([12.3456,30,456.12,0],ndigits=2) == [12.35,30,456.12,0],
Test([-12.3456,-30,-456.12,0],ndigits=2) == [-12.35,-30,-456.12,0],
)
@staticmethod
def roundint(arg,ndigits=0):
arg = as_num_array(arg)
factor = 10**ndigits
adjust = where(arg>=0,0.5,-0.5)
out = ((arg*factor)+adjust).astype('Int') / float(factor)
return out
@staticmethod
def naive(arg,ndigits=0):
arg = as_num_array(arg)
out = [round(val,ndigits) for val in arg]
return as_num_array(out)
class MidpointsFloat(EquivUnary):
"""Return list of midpoints for all unique values in given vector.
One typical use is to choose possible cutpoints when tree building.
Note: returned vector has length of len(unique_values)-1.
Note: this method applies to any numeric vector type and returns
float vector.
"""
ranking = ('vector','naive')
tests = (
Test([1.0]) == [],
Test([2.0]) == [],
Test([1,3,5]) == [2.0,4.0],
Test([1,3,3]) == [2.0],
Test([1.2,1.3,1.5]) == [1.25,1.4],
)
@staticmethod
def naive(arg):
uniq = sorted(set(arg))
out = []
for i in range(len(uniq)-1):
a,b = uniq[i:i+2]
out.append((a+b)/2.0)
return out
@staticmethod
def vector(arg):
uniq = as_num_array(sorted(set(arg)))
if len(uniq) <= 1:
return []
return (uniq[1:]+uniq[:-1])/2.0
class Gini2(EquivBinary):
"""GINI on a pair of vectors
"""
itypes = ('i', 'i')
name = 'gini2'
ranking = ('simple2', 'simple1')
tests = (
Test([1, 1, 1], [1, 1, 1]) == 0.0,
Test([1, 1, 1], [2, 2, 2]) == 0.0,
Test([1, 1, 1], [1, 1, 2])**0.222222222222,
Test([1, 2, 3], [1, 2, 3])**0.666666666667,
Test([1, 2, 3], [1, 2, 3, 4, 5])**0.75,
)
@staticmethod
def simple2(arg1, arg2):
args = [as_num_array(arg) for arg in (arg1, arg2) if len(arg)]
n = float(sum(len(arg) for arg in args))
return sum((gini(arg) * len(arg) / n) for arg in args)
@staticmethod
def simple1(arg1, arg2):
gini = Gini()
args = [as_num_array(arg) for arg in (arg1, arg2) if len(arg)]
n = float(sum(len(arg) for arg in args))
return sum((gini(arg) * len(arg) / n) for arg in args)
class DeltaPrev(EquivUnary):
'''difference from previous value
>>> func = DeltaPrev()
>>> assert func([1,3,6,7,7,9]).tolist() == [0,2,3,1,0,2]
'''
name = 'delta_prev'
ranking = ('smart','fast', 'naive_comp', 'naive_loop',)
tests = (
Test([1,3,6,7,7,9]) == [0,2,3,1,0,2],
Test([0,3,0,7,0,0]) == [0,3,-3,7,-7,0],
Test([0]) == [0],
)
@staticmethod
def naive_loop(arg,out=None):
arg = as_num_array(arg)
if not out:
out = arg.new()
out[0] = 0
for i in xrange(1,len(arg)):
out[i] = arg[i] - arg[i-1]
return out
@staticmethod
def naive_comp(arg,out=None):
arg = as_num_array(arg)
if not out:
out = arg.new()
out[0] = 0
out[1:] = [arg[i]-arg[i-1] for i in xrange(1,len(arg))]
return out
@staticmethod
def fast(arg,out=None):
arg = as_num_array(arg)
if not out:
out = arg.new()
out[0] = 0
subtract(arg[1:],arg[:-1],out[1:])
return out
@classmethod
def smart(self,arg,out=None):
if len(arg) < 10:
return self.naive_loop(arg,out)
return self.fast(arg,out)
@staticmethod
def _check_result(out,arg,**kwargs):
assert out[0] == 0
for o,a1,a2 in izip(out[1:],arg[1:],arg[:-1]):
assert o == a1-a2
return True
def run():
loader = Loader()
loader.load()
# AMQP-0-9-1
adapter = loader.catalog['amqp-0-9-1']
test = Test(URL, adapter)
test()
# amqp
adapter = loader.catalog['amqp']
test = Test(URL, adapter)
test()
def run():
loader = Loader()
loader.load()
# amqp-1-0
adapter = loader.catalog['amqp-1-0']
test = Test(URL, adapter)
test.test_crud()
test.test_no_exchange()
# proton
adapter = loader.catalog['proton']
test = Test(URL, adapter)
test.test_crud()
test.test_no_exchange()
def run():
# AMQP-0-10
loader = Loader()
loader.load()
adapter = loader.catalog['amqp-0-10']
test = Test(URL, adapter)
test()
# qpid
adapter = loader.catalog['qpid']
test = Test(URL, adapter)
test()
# qpid-messaging
adapter = loader.catalog['qpid.messaging']
test = Test(URL, adapter)
test()
class FuncMap(EquivUnary):
'''apply function to each element in array
>>> func = FuncMap()
>>> mapping = {1:100,2:200,3:300,'a':'aaa','b':'bbb'}
>>> assert func([1,3,6,7,7,9],func=mapping.get).tolist() == [100,300,6,7,7,9]
'''
name = 'func_map'
ranking = ('naive_comp', 'naive_loop',)
tests = (
Test([1,3,6,7,7,9],func={1:100,2:200,3:300,6:6,7:7,9:9}.get) == [100,300,6,7,7,9],
)
@staticmethod
def _default_func(arg): return arg
@staticmethod
def naive_loop(arg,func=None):
func = func or FuncMap._default_func
arg = as_any_array(arg)
out = []
for value in arg:
out.append(func(value))
return as_any_array(out)
@staticmethod
def naive_comp(arg,func=None):
func = func or FuncMap._default_func
arg = as_any_array(arg)
out = [func(value) for value in arg]
return as_any_array(out)
class GaussianPdf(EquivUnary):
'''Gaussian probability distribution function
>>> func = GaussianPdf().gsl
>>> assert allclose(func([1.2,0.1,0.5],variance=1.0),[0.19418605,0.39695255,0.35206533])
'''
name = 'gaussian_pdf'
ranking = ('gsl',)
tests = (
Test([1.2,0.1,0.5],variance=1.0) ** [0.19418605,0.39695255,0.35206533],
)
@staticmethod
def gsl(arg,mean=0.0,variance=0.0,out=None):
arg = as_num_array(arg)
if not out:
out = arg.new()
if mean == 0.0:
out[:] = gaussian_pdf(arg-mean,variance)
else:
out[:] = gaussian_pdf(arg,variance)
return out
class UniqueMask(EquivUnary):
'''given sorted data, return mask selecting unique values.
Typically used to reduce data with identical timestamps.
By default, last value from a run of equal values is taken.
'''
name = 'unique_mask'
ranking = (
'fast',
'naive_loop',
)
tests = (
Test([0]) == [1],
Test([0], first=True) == [1],
Test([1, 2, 3, 7, 8, 9]) == [1, 1, 1, 1, 1, 1],
Test([1, 2, 3, 7, 8, 9], first=True) == [1, 1, 1, 1, 1, 1],
Test([1, 3, 3, 3, 4, 5]) == [1, 0, 0, 1, 1, 1],
Test([1, 3, 3, 3, 4, 5], first=True) == [1, 1, 0, 0, 1, 1],
Test([1, 1, 3, 3, 5, 5], first=False) == [0, 1, 0, 1, 0, 1],
Test([1, 1, 3, 3, 5, 5], first=True) == [1, 0, 1, 0, 1, 0],
)
@staticmethod
def naive_loop(arg, first=False):
arg = as_num_array(arg)
if first:
out = [1]
for i in xrange(len(arg) - 1):
if arg[i] != arg[i + 1]:
out.append(1)
else:
out.append(0)
else:
out = []
for i in xrange(1, len(arg)):
if arg[i] != arg[i - 1]:
out.append(1)
else:
out.append(0)
out.append(1)
return as_num_array(out, type='Bool')
@staticmethod
def fast(arg, first=False):
arg = as_num_array(arg)
out = ones(len(arg), type='Bool')
if first:
reject = arg[1:] == arg[:-1]
out[1:] -= reject
else:
reject = arg[:-1] == arg[1:]
out[:-1] -= reject
return out
class MidpointsInteger(EquivUnary):
"""Return list of midpoints for all unique values in given vector.
One typical use is to choose possible cutpoints when tree building.
Note: returned vector has length of len(unique_values)-1.
Note: this method applies only to integer vectors and returns
integer vector
"""
ranking = ('vector','naive')
tests = (
Test([1]) == [],
Test([2]) == [],
Test([1,1,2]) == [2],
Test([2,2,3]) == [3],
Test([1,3,5]) == [2,4],
Test([-5,-1,0,1,8,2,2]) == [-3,0,1,2,5],
)
@staticmethod
def naive(arg):
uniq = sorted(set(arg))
out = []
for i in range(len(uniq)-1):
a,b = uniq[i:i+2]
out.append(((a+b)+1)//2)
return out
@staticmethod
def vector(arg):
uniq = as_num_array(sorted(set(arg)))
if len(uniq) <= 1:
return []
# assuming integer values and '<' op, the +1 causes round-up
return (uniq[1:]+uniq[:-1]+1)//2
class RangeCap(EquivUnary):
'''cap extreme values at specified limits
>>> func = RangeCap()
>>> assert func([1,3,-9,-3,6,7,9],lower=-5,upper=6).tolist() == [1,3,-5,-3,6,6,6]
'''
name = 'range_cap'
ranking = ('naive', 'clip')
tests = (
Test([1, 3, -9, -3, 6, 7, 9, 0], lower=None,
upper=None) == [1, 3, -9, -3, 6, 7, 9, 0],
Test([1, 3, -9, -3, 6, 7, 9, 0], lower=5,
upper=None) == [5, 5, 5, 5, 6, 7, 9, 5],
Test([1, 3, -9, -3, 6, 7, 9, 0], lower=-5,
upper=6) == [1, 3, -5, -3, 6, 6, 6, 0],
)
@staticmethod
def clip(arg, lower=None, upper=None):
arg = as_num_array(arg)
if lower is not None and upper is not None:
arg = clip(arg, lower, upper)
else:
if lower is not None:
arg = maximum(arg, lower)
if upper is not None:
arg = minimum(arg, upper)
return arg
@staticmethod
def naive(arg, lower=None, upper=None):
arg = as_num_array(arg)
if lower is not None:
arg = maximum(arg, lower)
if upper is not None:
arg = minimum(arg, upper)
return arg
class Discretize(EquivUnary):
"""Discretize according to sorted segment table
>>> func = Discretize()
>>> b = as_num_array([1,10,50])
>>> v = as_num_array([0,5,35,100])
>>> data = as_num_array([100,75,50,25,2,1,0])
>>> print func(data,boundaries=b)
[3 3 2 2 1 0 0]
>>> print func(data,boundaries=b,values=v)
[100 100 35 35 5 0 0]
"""
ranking = ('fast','loop')
tests = (
Test([0],boundaries=[0]) == [0],
Test([1,2,3,4,5],boundaries=[3]) == [0,0,0,1,1],
Test([1,2,1.3,1.4,1.33],boundaries=[1.33]) == [0,1,0,1,0],
Test([100,75,50,25,2,1,0],boundaries=[1,10,50]) == [3,3,2,2,1,0,0],
Test([100,75,50,25,2,1,0],boundaries=[1,10,50],values=[0,5,35,100]) == [100,100,35,35,5,0,0],
Test([100,-25,50,75,2,0,1],boundaries=[1,10,50],values=[0,5,35,100])== [100,0,35,100,5,0,0],
Test([1,2,3,4,5],boundaries=[3,2,1]) == [0,0,0,3,3], # nonsense
)
@staticmethod
def fast(arg,boundaries=[0,100,1000],values=None):
assert len(boundaries), "at least one boundary is required"
if values is not None:
assert len(boundaries)+1 == len(values), "len(values) must be len(boundaries)+1, (%s,%s)" % (len(values),len(boundaries))
idx = searchsorted(boundaries,arg)
if values is None:
return idx
return as_num_array(values).take(idx)
@staticmethod
def loop(arg,boundaries=[0,100,1000],values=None):
assert len(boundaries), "at least one boundary is required"
if values is not None:
assert len(boundaries)+1 == len(values), "len(values) must be len(boundaries)+1, (%s,%s)" % (len(values),len(boundaries))
idx = []
for value in arg:
for i,x in it.izip(it.count(),boundaries):
if value <= x:
idx.append(i)
break
else:
idx.append(len(boundaries))
if values is None:
return idx
return [values[i] for i in idx]
class UniformPdf(EquivUnary):
'''Uniform probability distribution function
>>> func = UniformPdf().gsl
>>> assert allclose(func([1.2,0.1,0.5],b=1.0),[0.0,1.0,1.0])
'''
name = 'uniform_pdf'
ranking = ('gsl', )
tests = (Test([1.2, 0.1, 0.5], b=1.0)**[0.0, 1.0, 1.0], )
@staticmethod
def gsl(arg, a=0.0, b=0.0, out=None):
arg = as_num_array(arg)
if not out:
out = arg.new()
out[:] = flat_pdf(arg, a, b)
return out
class ExponentialPdf(EquivUnary):
'''Exponential probability distribution function
>>> func = ExponentialPdf().gsl
>>> assert allclose(func([1.2,0.1,0.5],mu=1.0),[0.30119421,0.90483742,0.60653066])
'''
name = 'exponential_pdf'
ranking = ('gsl', )
tests = (Test([1.2, 0.1, 0.5],
mu=1.0)**[0.30119421, 0.90483742, 0.60653066], )
@staticmethod
def gsl(arg, mu=0.0, out=None):
arg = as_num_array(arg)
if not out:
out = arg.new()
out[:] = exponential_pdf(arg, mu)
return out
class CusumReset(EquivUnary):
'''CUSUM with reset algorithm
>>> func = CusumReset().iterfunc
>>> assert func([1,-3,6,7,-7,-9]).tolist() == [1,0,6,13,6,0]
'''
name = 'cusum_reset'
ranking = ('iterfunc', )
#ranking = ('iterfunc','iterloop')
tests = (Test([1, -3, 6, 7, -7, -9]) == [1, 0, 6, 13, 6, 0], )
@staticmethod
def iterfunc(arg, reset_value=0.0, out=None):
def gen_cusum(data, reset_value=0.0):
# no obvious way to vectorize this
out = 0.0
for value in data:
out = max(reset_value, out + value)
yield out
arg = as_num_array(arg)
if out is None:
out = arg.new()
out[:] = list(gen_cusum(arg, reset_value))
return out
@staticmethod
def iterloop(arg, reset_value=0.0, out=None):
arg = as_num_array(arg)
if not out:
out = arg.new()
last = 0.0
for i, value in it.izip(it.count(), arg):
out[i] = max(reset_value, last + value)
return out
class CarryForward(EquivUnary):
'''fill empty values with previous non-empty value
>>> func = CarryForward().fast
>>> assert func([0,3,0,0,4,0]).tolist() == [0,3,3,3,4,4]
'''
name = 'carry_forward'
ranking = ('naive_loop', 'naive_iter', 'array_idx')
tests = (
Test([0,3,0,0,4,0]) == [0,3,3,3,4,4],
Test([0,0,0,0,0,0]) == [0,0,0,0,0,0],
Test([0,0,0,0,0,5]) == [0,0,0,0,0,5],
Test([9,0,0,0,0,0]) == [9,9,9,9,9,9],
)
@staticmethod
def array_idx(arg,out=None):
arg = as_num_array(arg)
if not out:
out = arg.new()
idx = arg.nonzero()[0]
try:
first = idx[0]
except IndexError:
first = len(out)
out[:first] = 0
if not len(idx):
return out
a,b = it.tee(idx)
b.next()
for start,stop in izip(a,b):
out[start:stop] = arg[start]
last = idx[-1]
out[last:] = arg[last]
return out
@staticmethod
def naive_iter(arg,out=None):
arg = as_num_array(arg)
if not out:
out = arg.new()
last = 0
for i,value in izip(it.count(),arg):
if value != 0:
out[i] = last = value
else:
out[i] = last
return out
@staticmethod
def naive_loop(arg,out=None):
arg = as_num_array(arg)
if not out:
out = arg.new()
last = 0
for i in xrange(len(arg)):
if arg[i] != 0:
last = arg[i]
out[i] = last
return out
def run():
print URL
test = Test(URL)
test.test_crud()
test.test_no_exchange()
class PartitionIntegerGini(EquivBinary):
"""Given a pair of dependent+independent integer vectors,
consider all possible cutpoints in the independent vector, and
return list of Gini scores for resulting splits in dependent.
"""
itypes = ('i', 'i')
ranking = ('isort', 'deltacnt2', 'deltacnt1', 'deltacnt0', 'presort',
'idxsel', 'masksel', 'naive')
bench_sizes = (
1,
2,
3,
4,
5,
10,
20,
50,
100,
200,
300,
500,
1000,
5000,
10000,
# the following sizes take too long to include in the default benchmarking run
#50000,
#100000,
#500000,
#1000000,
)
tests = (
Test([1, 2]) == [0.0],
Test([1, 2, 3])**([0.3333333] * 2),
Test([1, 2, 3, 4])**([0.5] * 3),
Test([1, 2, 3, 4, 5])**([0.5999999] * 4),
Test([0, 1, 2, 3, 4])**([0.5999999] * 4),
Test([1, 2, 3, 2, 1])**[0.2666666, 0.4],
Test([3, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1])**[0.071428571428571425, 0.13186813186813176],
Test([1, 2, 3, 4], [6, 7, 6, 8])**[0.5, 0.5],
Test([1, 2, 3, 4, 5, 6], [6, 7, 6, 8, 9, 6])**([0.6666666] * 3),
Test([1, 2, 1, 2, 1, 2], [6, 7, 8, 8, 9, 9])**[0.3999999, 0.5, 0.5],
Test([1, 2, 1, 2, 1, 2],
[4, 5, 6, 7, 8, 9])**[0.3999999, 0.5, 0.4444444, 0.5, 0.3999999],
Test([1, 2, 1, 2, 1, 2], [4, 5, 6, 7, 8, 9],
cutpoints=[7, 8])**[0.4444444, 0.5],
)
def _prep_testdata(self, *args, **kwargs):
# benchmark for inputs that are already vectors
# simplification for tests: dep == indep
out = [as_num_array(arg) for arg in args]
if len(out) == 1:
out.append(out[0].copy())
if not kwargs.get('dep_sorted'):
idx = argsort(out[0])
out = [take(vec, idx) for vec in out]
kwargs['dep_sorted'] = True
return (out, kwargs)
@staticmethod
def smart(dep, indep, cutpoints=None, **kwargs):
# not needed unless high penalty for small datasets
dep = as_num_array(dep)
indep = as_num_array(indep)
if cutpoints is None:
cutpoints = midpoints_integer(indep)
if len(dep) < 100:
return PartitionIntegerGini.naive(dep,
indep,
cutpoints=cutpoints,
**kwargs)
return PartitionIntegerGini.isort(dep,
indep,
cutpoints=cutpoints,
**kwargs)
@staticmethod
def isort(dep, indep, cutpoints=None, **kwargs):
dep = as_num_array(dep)
indep = as_num_array(indep)
if cutpoints is None:
cutpoints = midpoints_integer(indep)
if not len(cutpoints):
return []
# sort both vectors by *indep*
idx = argsort(indep)
dep = take(dep, idx)
indep = take(indep, idx)
#print len(dep),len(indep),len(cutpoints)
cutidx = [0, 0]
for ival, isub in it.groupby(indep):
ilen = len(list(isub))
if ival < cutpoints[len(cutidx) - 2]:
cutidx[-1] += ilen
else:
if len(cutidx) > len(cutpoints):
break
cutidx.append(cutidx[-1] + ilen)
assert len(cutidx) - 1 == len(
cutpoints), '%s != %s' % (len(cutidx) - 1, len(cutpoints))
out = []
cnt2 = dict(histo_tuple(dep))
cnt1 = dict.fromkeys(cnt2.keys(), 0)
for i1, i2 in izip(cutidx[:-1], cutidx[1:]):
# update the counts from the last cut
for d, cnt in histo_tuple(dep[i1:i2]):
cnt1[d] += cnt
cnt2[d] -= cnt
# calculate results based on counts
a1 = as_num_array([val for val in cnt1.itervalues() if val != 0])
a2 = as_num_array([val for val in cnt2.itervalues() if val != 0])
out.append(gini2_counts(a1, a2))
assert len(out) == len(cutpoints), '%s != %s' % (len(out),
len(cutpoints))
return out
@staticmethod
def deltacnt2(dep, indep, cutpoints=None, **kwargs):
dep = as_num_array(dep)
indep = as_num_array(indep)
if cutpoints is None:
cutpoints = midpoints_integer(indep)
out = []
# dictionary of counts in each dataset
cnt2 = dict(histo_tuple(dep))
cnt1 = dict.fromkeys(cnt2.keys(), 0)
lastmask = (indep != indep)
for cut in cutpoints:
mask = indep < cut
# examine only the new values from the last cut
maskdelta = mask & ~lastmask
lastmask |= mask
idxdelta = nonzero(maskdelta)[0]
# update the counts from the last cut
for d, cnt in histo_tuple(dep[idxdelta]):
cnt1[d] += cnt
cnt2[d] -= cnt
# calculate results based on counts
a1 = as_num_array([val for val in cnt1.itervalues() if val != 0])
a2 = as_num_array([val for val in cnt2.itervalues() if val != 0])
out.append(gini2_counts(a1, a2))
return out
@staticmethod
def deltacnt1(dep, indep, cutpoints=None, **kwargs):
dep = as_num_array(dep)
indep = as_num_array(indep)
if cutpoints is None:
cutpoints = midpoints_integer(indep)
out = []
# dictionary of counts in each dataset
cnt2 = dict(histo_tuple(dep))
cnt1 = dict.fromkeys(cnt2.keys(), 0)
lastmask = (indep != indep)
for cut in cutpoints:
mask = indep < cut
# examine only the new values from the last cut
maskdelta = mask & ~lastmask
lastmask |= mask
idxdelta = nonzero(maskdelta)[0]
# update the counts from the last cut
for d in dep[idxdelta]:
cnt1[d] += 1
cnt2[d] -= 1
# calculate results based on counts
a1 = as_num_array([val for val in cnt1.itervalues() if val != 0])
a2 = as_num_array([val for val in cnt2.itervalues() if val != 0])
out.append(gini2_counts(a1, a2))
return out
@staticmethod
def deltacnt0(dep, indep, cutpoints=None, **kwargs):
dep = as_num_array(dep)
indep = as_num_array(indep)
if cutpoints is None:
cutpoints = midpoints_integer(indep)
out = []
# get vector of counts in each dataset
tmp = histo_tuple(dep)
dep_keys = dict((x[0], i) for i, x in enumerate(tmp))
cnt2 = as_num_array([x[1] for x in tmp])
cnt1 = cnt2 - cnt2
lastmask = (indep != indep)
for cut in cutpoints:
mask = indep < cut
maskdelta = mask & ~lastmask
lastmask |= mask
for d in dep[maskdelta]:
key = dep_keys[d]
cnt1[key] += 1
cnt2[key] -= 1
a1 = cnt1[cnt1 != 0]
a2 = cnt2[cnt2 != 0]
out.append(gini2_counts(a1, a2))
return out
@staticmethod
def presort(dep, indep, cutpoints=None, dep_sorted=False):
dep = as_num_array(dep)
indep = as_num_array(indep)
if not dep_sorted:
idx = argsort(dep)
dep = take(dep, idx)
indep = take(indep, idx)
if cutpoints is None:
cutpoints = midpoints_integer(indep)
out = []
for cut in cutpoints:
mask = indep < cut
i1 = nonzero(mask)[0]
i2 = nonzero(~mask)[0]
a1 = dep[i1]
a2 = dep[i2]
out.append(gini2_presorted(a1, a2))
return out
@staticmethod
def idxsel(dep, indep, cutpoints=None, **kwargs):
dep = as_num_array(dep)
indep = as_num_array(indep)
if cutpoints is None:
cutpoints = midpoints_integer(indep)
out = []
for cut in cutpoints:
mask = indep < cut
i1 = nonzero(mask)[0]
i2 = nonzero(~mask)[0]
a1 = dep[i1]
a2 = dep[i2]
out.append(gini2(a1, a2))
return out
@staticmethod
def masksel(dep, indep, cutpoints=None, **kwargs):
dep = as_num_array(dep)
indep = as_num_array(indep)
if cutpoints is None:
cutpoints = midpoints_integer(indep)
out = []
for cut in cutpoints:
mask = indep < cut
a1 = dep[mask]
a2 = dep[~mask]
out.append(gini2(a1, a2))
return out
@staticmethod
def naive(dep, indep, cutpoints=None, **kwargs):
if cutpoints is None:
cutpoints = midpoints_integer(indep)
out = []
for cut in cutpoints:
a1 = [d for d, i in izip(dep, indep) if i < cut]
a2 = [d for d, i in izip(dep, indep) if i >= cut]
out.append(gini2(a1, a2))
return out
def run():
loader = Loader()
loader.load()
# amqp-1-0
adapter = loader.catalog['amqp-1-0']
test = Test(URL, adapter)
test.test_crud()
test.test_no_exchange()
# proton
adapter = loader.catalog['proton']
test = Test(URL, adapter)
test.test_crud()
test.test_no_exchange()
class WindowApply(EquivUnary):
"""Apply function to sliding window within single data vector
"""
name = 'window_apply'
ranking = ('naive1', 'naive2')
bench_sizes = (1, 2, 3, 4, 5, 10, 20, 50, 100, 200, 300, 500, 1000, 5000)
tests = (
Test(arange(20),sel=[3,5,9,19]) == [6,9,30,145],
Test(arange(10),sel=zeros(10,type='Bool'),step=1) == [],
Test(arange(10),sel=ones(10,type='Bool'),step=1) \
== [0,1,2,3,4,5,6,7,8,9],
Test(arange(10),sel=(arange(10)%2)==0,step=1) == [0,3,7,11,15],
Test(arange(10),sel=(arange(10)%2)!=0,step=1) == [1,5,9,13,17],
Test(arange(10),step=2) == [1,3,5,7,9,11,13,15,17],
Test(arange(1,11),step=2) == [2,5,7,9,11,13,15,17,19],
Test(arange(10),step=3) == [3,6,9,12,15,18,21,24],
Test(arange(1,11),step=3) == [5,9,12,15,18,21,24,27],
Test(arange(2,12),step=3) == [7,12,15,18,21,24,27,30],
Test([0,1,2,3],func=sum) == [0,1,2,3],
Test([0,1,2,3],step=2) == [1,3,5],
#Test([0,1,2,3],func=minimum) == [0,1,2,3],
)
def _prep_testdata_broken(self, *args, **kwargs):
out = [as_num_array(arg) for arg in args]
if not kwargs:
# automatic test cases
kwargs['sel'] = as_num_array(args[0]) == 0
return (out, kwargs)
@staticmethod
def naive1(arg, sel=None, step=1, func=sum):
arg = as_num_array(arg)
idx = arg_sel_step_to_idx(arg, sel, step)
out = []
for i in xrange(len(idx) - step):
j = idx[i]
k = idx[i + step]
chunk = arg[j + 1:k + 1]
out.append(func(chunk))
return as_num_array(out)
@staticmethod
def naive2(arg, sel=None, step=1, func=sum):
arg = as_num_array(arg)
idx = arg_sel_step_to_idx(arg, sel, step)
jj = idx[:-step]
kk = idx[step:]
out = []
for j, k in izip(jj, kk):
chunk = arg[j + 1:k + 1]
out.append(func(chunk))
return as_num_array(out)
class Gini(EquivUnary):
"""GINI on a single vector
"""
itypes = 'i'
name = 'gini'
ranking = ('groupby2', 'smart', 'loop3', 'groupby3', 'groupby1', 'loop2',
'loop1')
tests = (
Test([0, 0, 0]) == 0.0,
Test([1, 1, 1]) == 0.0,
Test([0, 0, 1])**0.444444444444,
Test([0, 1, 1])**0.444444444444,
Test([1, 1, 2])**0.444444444444,
Test([1, 2, 2])**0.444444444444,
Test([1, 2, 3])**0.666666666667,
Test([1, 2, 3, 4]) == 0.75,
Test([4, 3, 2, 1]) == 0.75,
Test([1, 2, 3, 2, 3, 3])**0.611111111111,
Test([2, 1, 3, 200, 3000, 30000])**0.833333333333,
Test(range(100))**0.99,
Test(range(1000))**0.999,
)
@staticmethod
def smart(arg):
arg = as_num_array(arg)
if len(arg) < 50:
return Gini.loop3(arg)
return Gini.groupby2(arg)
@staticmethod
def groupby3(arg):
arg = as_num_array(arg)
n = float(len(arg))
gfx = as_num_array([len(list(g)) for k, g in groupby(sort(arg))]) / n
gfx *= gfx
out = 1.0 - gfx.sum()
return out
@staticmethod
def groupby2(arg):
arg = as_num_array(arg)
n = float(len(arg))
gfx = [len(list(g)) / n for k, g in groupby(sorted(arg))]
out = 1.0
for gf in gfx:
out -= gf * gf
return out
@staticmethod
def groupby1(arg):
arg = as_num_array(arg)
histo = [(k, len(list(g))) for k, g in groupby(sorted(arg))]
n = float(len(arg))
out = 1.0
for (val, cnt) in histo:
gf = cnt / n
out -= gf * gf
return out
@staticmethod
def loop3(arg):
arg = as_num_array(arg)
n = float(len(arg))
enum = {}
for val in arg:
enum[val] = 1 + enum.setdefault(val, 0)
out = 1.0
for cnt in enum.itervalues():
gf = (cnt * 1.0) / n
out -= gf * gf
return out
@staticmethod
def loop2(arg):
arg = as_num_array(arg)
n = float(len(arg))
enum = {}
for val in arg:
cnt = enum.get(val, 0)
enum[val] = cnt + 1
out = 1.0
for (val, cnt) in enum.iteritems():
gf = (cnt * 1.0) / n
out -= gf * gf
return out
@staticmethod
def loop1(arg):
arg = as_num_array(arg)
n = float(len(arg))
enum = {}
for val in arg:
cnt = enum.get(val, 0)
enum[val] = cnt + 1
out = 1.0
for (val, cnt) in enum.items():
gf = (cnt * 1.0) / n
out -= gf * gf
return out
class GiniGain(EquivBinary):
"""GINI gain for a given split
Note: this is not likely to be the most efficient method for tree
building since it would result in repeated calculation of the gini()
of the combined nodes. Better to have the application calculate that
once and then loop over the possible splits.
This method is intended more as documentation and as an additional
test of the combination of the above methods.
"""
itypes = ('i', 'i')
name = 'gini_gain'
ranking = ('simple1', )
tests = (
Test([], [1, 1, 1, 1, 1, 1]) == 0.0,
Test([1], [1, 1, 1, 1, 1]) == 0.0,
Test([1, 1, 1], [1, 1, 1]) == 0.0,
Test([1], [1, 1, 2, 2, 2])**0.1,
Test([1, 1], [1, 2, 2, 2]) == 0.25,
Test([1, 1, 1], [2, 2, 2]) == 0.5,
Test([1, 1, 1, 2], [2, 2]) == 0.25,
Test([1, 1, 1], [1, 1, 2])**0.055555555555,
Test([1, 2, 3], [1, 2, 3]) == 0.0,
Test([1, 2, 3], [1, 2, 3, 4, 5])**0.03125,
Test([1, 3], [2, 999]) == 0.25,
)
@staticmethod
def simple1(arg1, arg2):
gini1 = Gini()
gini2 = Gini2()
args = [as_num_array(arg) for arg in (arg1, arg2) if len(arg)]
if len(args) != 2:
return 0.0
return gini1(concatenate(args)) - gini2(*args)
class GetShift(EquivUnary):
'''get value of other column shifted by some offset
(typically useful for previous value (or arbitrary offset into column)
>>> func = GetShift()
>>> assert func([1,3,6,7,7,9],filler=99).tolist() == [3,6,7,7,9,99]
'''
name = 'get_shift'
ranking = ('smart','fast', 'naive_comp', 'naive_loop',)
tests = (
Test([0]) == [0],
Test([1,3,6,7,7,9]) == [0,1,3,6,7,7],
Test([1,3,6,7,7,9],offset=3) == [0,0,0,1,3,6,],
Test([1,3,6,7,7,9],offset=-2) == [6,7,7,9,0,0],
Test([1,3,6,7,7,9],filler=99) == [99,1,3,6,7,7],
)
@staticmethod
def naive_loop(arg,offset=1,filler=0,out=None):
arg = as_num_array(arg)
if not out:
out = arg.new()
if offset < 0:
cut = len(arg)+offset
for i in xrange(cut):
out[i] = arg[i-offset]
for i in xrange(cut,len(arg)):
out[i] = filler
else:
cut = offset
for i in xrange(cut):
out[i] = filler
for i in xrange(cut,len(arg)):
out[i] = arg[i-offset]
return out
@staticmethod
def naive_comp(arg,offset=1,filler=0,out=None):
arg = as_num_array(arg)
if not out:
out = arg.new()
if offset < 0:
cut = len(arg)+offset
out[:cut] = [arg[i] for i in xrange(-offset,len(arg))]
out[cut:] = [filler]*(-offset)
else:
cut = offset
out[:cut] = [filler]*(offset)
out[cut:] = [arg[i] for i in xrange(len(arg)-cut)]
return out
@staticmethod
def fast(arg,offset=1,filler=0,out=None):
arg = as_num_array(arg)
if offset < 0:
cut = len(arg)+offset
out1 = arg[-offset:len(arg)]
out2 = (-offset)*[filler]
else:
cut = offset
out1 = offset*[filler]
out2 = arg[:len(arg)-cut]
if not out:
out = arg.new()
out[:cut] = out1
out[cut:] = out2
return out
@classmethod
def smart(self,arg,offset=1,filler=0,out=None):
if len(arg) < 100:
return self.naive_loop(arg,offset=offset,filler=filler,out=out)
return self.fast(arg,offset=offset,filler=filler,out=out)
@staticmethod
def _check_result(out,arg,**kwargs):
offset = kwargs.get('offset',1)
filler = kwargs.get('filler',0)
if offset < 0:
cut = len(arg)+offset
for i in xrange(cut):
assert out[i] == arg[i-offset]
for i in xrange(cut,len(arg)):
assert out[i] == filler
else:
cut = offset
for i in xrange(cut):
assert out[i] == filler
for i in xrange(cut,len(arg)):
assert out[i] == arg[i-offset]
return True
class LinearBins(EquivUnary):
"""Discretize to given number of equal sized bins, returning either
1) if idx==True, the bin number (0 to bins-1) for each value.
2) otherwise returning midpoint value for the range in each bin.
This is mainly a convenience wrapper to Discretize().
"""
ranking = ('wrapper',)
tests = (
Test([0]) == [0],
Test([0,2],bins=1) == [1.0,1.0],
Test([0,2],bins=2) == [0.5,1.5],
Test([0,2],bins=3) ** [0.333333,1.666666],
Test([0,1,2],bins=3) ** [0.333333,1.0,1.666666],
Test([1,2,3,4,5],bins=3) ** [1.666666,1.666666,3.0,4.333333,4.333333],
Test([1,2,1.3,1.4,1.33],bins=2) == [1.25,1.75,1.25,1.25,1.25],
Test([1,2,1.3,1.4,1.33],bins=2,min=0.5) == [0.875,1.625,1.625,1.625,1.625],
Test([1,2,1.3,1.4,1.33],bins=2,min=-1) == [1.25,1.25,1.25,1.25,1.25],
Test([0],idx=True) == [0],
Test([0,2],bins=1,idx=True) == [0,0],
Test([0,2],bins=2,idx=True) == [0,1],
Test([0,2],bins=3,idx=True) == [0,2],
Test([0,1,2],bins=3,idx=True) == [0,1,2],
Test([1,2,3,4,5],bins=3,idx=True) == [0,0,1,2,2],
Test([1,2,1.3,1.4,1.33],bins=2,idx=True) == [0,1,0,0,0],
Test([1,2,1.3,1.4,1.33],bins=2,min=0.5,idx=True) == [0,1,1,1,1],
Test([1,2,1.3,1.4,1.33],bins=2,min=-1,idx=True) == [1,1,1,1,1],
Test([100,75,50,25,2,1,0],bins=35,max=500,idx=True) == [6,5,3,1,0,0,0],
)
@staticmethod
def wrapper(arg,bins=10,min=None,max=None,idx=None):
arg = as_num_array(arg)
if min is None:
min = arg.min()
if max is None:
max = arg.max()
step = (max - min) / bins
if not step:
return zeros(len(arg))
boundaries = arange(min+step,max+step,step)
if idx:
values = None
else:
values = arange(min+(step/2),max+step+step,step)[:len(boundaries)+1]
return discretize(arg,boundaries=boundaries,values=values)
def run():
print(URL)
test = Test(URL)
test()
class LookAheadIndex(EquivUnary):
"""Given
1) a vector of sorted values that may contain duplicates and
irregular spacing (like event timestamps, for example), and
2) a numeric delta to add to the vector (to identify the
future timestamp some fixed interval away),
return an index vector into the original data vector such that
for each current value, the index points to either
1) the first occurrence of current value+delta, or
2) if the data doesn't contain such value, the first occurrence
of the lext lower value that does occur in the data.
If delta==0, then the index points to the first/last occurrence of the
current value in the data.
"""
itypes = 'i'
ranking = ('naive_loop',)
ranking = ('missing1','naive_loop')
bench_sizes = (1,2,3,4,5,10,20,50,100,200,300,500,1000,5000,
#10000,
#50000,
#100000,
#500000,
#1000000,
)
tests = (
Test([0,0,1,1,1,2,3,5,9],delta=0,first=True) == [0,0,2,2,2,5,6,7,8],
Test([0,0,1,1,1,2,3,5,9],delta=1,first=True) == [2,2,5,5,5,6,6,7,8],
Test([0,0,1,1,1,2,3,5,9],delta=2,first=True) == [5,5,6,6,6,6,7,7,8],
Test([0,0,1,1,1,2,3,5,9],delta=3,first=True) == [6,6,6,6,6,7,7,7,8],
Test([0,0,1,1,1,2,3,5,9],delta=4,first=True) == [6,6,7,7,7,7,7,8,8],
Test([0,0,1,1,1,2,3,5,9],delta=5,first=True) == [7,7,7,7,7,7,7,8,8],
Test([0,0,1,1,1,2,3,5,9],delta=0,first=False) == [1,1,4,4,4,5,6,7,8],
Test([0,0,1,1,1,2,3,5,9],delta=1,first=False) == [4,4,5,5,5,6,6,7,8],
Test([0,0,1,1,1,2,3,5,9],delta=2,first=False) == [5,5,6,6,6,6,7,7,8],
Test([0,0,1,1,1,2,3,5,9],delta=3,first=False) == [6,6,6,6,6,7,7,7,8],
Test([0,0,1,1,1,2,3,5,9],delta=4,first=False) == [6,6,7,7,7,7,7,8,8],
Test([0,0,1,1,1,2,3,5,9],delta=5,first=False) == [7,7,7,7,7,7,7,8,8],
Test([0,0,1,1,1,2,3,5,9],delta=-0,first=True) == [0,0,2,2,2,5,6,7,8],
Test([0,0,1,1,1,2,3,5,9],delta=-1,first=True) == [0,0,0,0,0,2,5,6,7],
Test([0,0,1,1,1,2,3,5,9],delta=-2,first=True) == [0,0,0,0,0,0,2,6,7],
Test([0,0,1,1,1,2,3,5,9],delta=-3,first=True) == [0,0,0,0,0,0,0,5,7],
Test([0,0,1,1,1,2,3,5,9],delta=-4,first=True) == [0,0,0,0,0,0,0,2,7],
Test([0,0,1,1,1,2,3,5,9],delta=-5,first=True) == [0,0,0,0,0,0,0,0,6],
Test([0,0,1,1,1,2,3,5,9],delta=-0,first=False) == [1,1,4,4,4,5,6,7,8],
Test([0,0,1,1,1,2,3,5,9],delta=-1,first=False) == [1,1,1,1,1,4,5,6,7],
Test([0,0,1,1,1,2,3,5,9],delta=-2,first=False) == [1,1,1,1,1,1,4,6,7],
Test([0,0,1,1,1,2,3,5,9],delta=-3,first=False) == [1,1,1,1,1,1,1,5,7],
Test([0,0,1,1,1,2,3,5,9],delta=-4,first=False) == [1,1,1,1,1,1,1,4,7],
Test([0,0,1,1,1,2,3,5,9],delta=-5,first=False) == [1,1,1,1,1,1,1,1,6],
Test([0],delta=-1) == [],
Test([1,5],delta=-3) == [0,0],
Test([0]) == [],
Test([1,5],delta=3) == [0,1],
)
def _prep_testdata(self,*args,**kwargs):
# benchmark for inputs that are already vectors
# simplification for tests: dep == indep
return [as_num_array(sorted(arg)) for arg in args]
@staticmethod
def missing1(arg,delta=1,first=False):
# build answer lookup mapping each arg value to first index
run_lens = [(k,len(list(g))) for k,g in groupby(arg)]
keys = as_num_array([k for k,l in run_lens])
lens = as_num_array([l for k,l in run_lens])
ends = cumsum(lens)
starts = ends - lens
if first:
answer = dict(izip(keys,starts))
else:
answer = dict(izip(keys,ends-1))
# identify missing keys
need = keys + delta
needset = set(need)
haveset = set(answer)
fillset = needset.difference(haveset)
fill = as_num_array(sorted(fillset))
#
#print
#print 'haveset:', haveset
#print 'need:', need
#print 'fill:', fill
#print 'answer1:', answer
#
minkey,maxkey = arg[0],arg[-1]
#
have_iter = iter(keys[-1::-1])
fill_iter = iter(fill[-1::-1])
thiskey = maxkey
thisval = answer[thiskey]
for fillkey in fill_iter:
# print 'fillkey:', fillkey
if thiskey >= fillkey:
try:
thiskey = dropwhile(lambda x:x>=fillkey,have_iter).next()
except StopIteration:
thiskey = minkey
thisval = answer[thiskey]
answer[fillkey] = thisval
#print 'answer2:', answer
out = [answer[val+delta] for val in arg]
return out
@staticmethod
def naive_loop(arg,delta=1,first=False):
out = []
for i,val in enumerate(arg):
# find answer range
target = val + delta
jj = i
if target > val:
# look forward
for j in xrange(i+1,len(arg)):
if arg[j] > target:
break
else:
j = len(arg)
jj = j-1
target = arg[jj]
elif target < val:
# look backward
for j in xrange(i-1,-1,-1):
if arg[j] <= target:
break
else:
j = 0
jj = j
target = arg[jj]
# find first or last answer within range
if first:
kk = 0
for k in xrange(jj,-1,-1):
if arg[k] != target:
kk = k+1
break
else:
kk = len(arg)-1
for k in xrange(jj,len(arg)):
if arg[k] != target:
kk = k-1
break
out.append(kk)
return out