https://github.com/annotation/text-fabric
Tip revision: 5eb3b9543b893105dde14862856d55e5cbda13cd authored by Dirk Roorda on 02 May 2018, 20:49:14 UTC
New minor release 3.4.12
New minor release 3.4.12
Tip revision: 5eb3b95
search.py
import re
import types
import collections
from functools import reduce
from random import randrange
from inspect import signature
from .data import WARP
STRATEGY = '''
small_choice_first
spread_1_first
big_choice_first
'''.strip().split()
escapes = (
'\\\\',
'\\ ',
'\\t',
'\\n',
'\\|',
'\\=',
)
rePat = '^([a-zA-Z0-9-_]+)~(.*)$'
reRe = re.compile(rePat)
reTp = type(reRe)
def esc(x):
for (i, c) in enumerate(escapes):
x = x.replace(c, chr(i))
return x
def unesc(x):
for (i, c) in enumerate(escapes):
x = x.replace(chr(i), c[1])
return x
class Search(object):
def __init__(self, api, silent):
self.api = api
self.good = True
self.silent = silent
self._basicRelations()
# API METHODS ###
def search(self, searchTemplate, limit=None):
self.silent = True
self.study(searchTemplate, silent=True)
return self.fetch(limit=limit)
def study(self, searchTemplate, strategy=None, silent=False):
info = self.api.info
self.silent = silent
self.api.indent(level=0, reset=True)
self.good = True
self._setStrategy(strategy)
if not self.good:
return
if not self.silent:
info('Checking search template ...')
self.searchTemplate = searchTemplate
self._parse()
self._prepare()
if not self.good:
return
if not self.silent:
info(
'Setting up search space for {} objects ...'.format(
len(self.qnodes)
)
)
self._spinAtoms()
if not self.silent:
info(
'Constraining search space with {} relations ...'.format(
len(self.qedges)
)
)
self._spinEdges()
if not self.silent:
info('Setting up retrieval plan ...')
self._stitch()
if self.good:
if not self.silent:
info(
'Ready to deliver results from {} nodes'.format(
sum(len(y) for y in self.yarns.values()),
)
)
if not self.silent:
info('Iterate over S.fetch() to get the results', tm=False)
if not self.silent:
info('See S.showPlan() to interpret the results', tm=False)
def fetch(self, limit=None):
if not self.good:
return []
if limit is None:
return self.results()
results = []
for r in self.results():
results.append(r)
if len(results) == limit:
break
return tuple(results)
def glean(self, r):
T = self.api.T
F = self.api.F
L = self.api.L
slotType = F.otype.slotType
lR = len(r)
if lR == 0:
return ''
fields = []
for (i, n) in enumerate(r):
otype = F.otype.v(n)
words = [n] if otype == slotType else L.d(n, otype=slotType)
if otype == T.sectionTypes[2]:
field = '{} {}:{}'.format(*T.sectionFromNode(n))
elif otype == slotType:
field = T.text(words)
elif otype in T.sectionTypes[0:2]:
field = ''
else:
field = '{}[{}{}]'.format(
otype,
T.text(words[0:5]),
'...' if len(words) > 5 else '',
)
fields.append(field)
return ' '.join(fields)
def count(self, progress=None, limit=None):
info = self.api.info
indent = self.api.indent
indent(level=0, reset=True)
if not self.good:
self.error(
'This search has problems. No results to count.', tm=False
)
return
PROGRESS = 100
LIMIT = 1000
if progress is None:
progress = PROGRESS
if limit is None:
limit = LIMIT
info(
'Counting results per {} up to {} ...'.format(
progress,
limit if limit > 0 else ' the end of the results',
)
)
indent(level=1, reset=True)
i = 0
j = 0
for r in self.results(remap=False):
i += 1
j += 1
if j == progress:
j = 0
info(i)
if limit > 0 and i >= limit:
break
indent(level=0)
info('Done: {} results'.format(i))
def showPlan(self, details=False):
if not self.good:
return
info = self.api.info
nodeLine = self.nodeLine
qedges = self.qedges
(qs, es) = self.stitchPlan
if details:
info(
'Search with {} objects and {} relations'.format(
len(qs), len(es)
),
tm=False
)
info(
'Results are instantiations of the following objects:',
tm=False
)
for q in qs:
self._showNode(q)
if len(es) != 0:
info(
'Instantiations are computed'
' along the following relations:',
tm=False
)
(firstE, firstDir) = es[0]
(f, rela, t) = qedges[firstE]
if firstDir == -1:
(f, t) = (t, f)
self._showNode(f, pos2=True)
for e in es:
self._showEdge(*e)
info(
'The results are connected to the'
' original search template as follows:'
)
resultNode = {}
for q in qs:
resultNode[nodeLine[q]] = q
for (i, line) in enumerate(self.searchLines):
rNode = resultNode.get(i, '')
info(
'{:>2} {:<1}{:<2} {}'.format(
i,
'R' if rNode != '' else '',
rNode,
line,
),
tm=False
)
# LOW-LEVEL NODE RELATIONS SEMANTICS ###
def _basicRelations(self):
C = self.api.C
F = self.api.F
E = self.api.E
Crank = C.rank.data
ClevDown = C.levDown.data
ClevUp = C.levUp.data
(CfirstSlots, ClastSlots) = C.boundary.data
Eoslots = E.oslots.data
slotType = F.otype.slotType
maxSlot = F.otype.maxSlot
def equalR(fTp, tTp):
return lambda n: (n, )
def spinEqual(fTp, tTp):
def doyarns(yF, yT):
x = set(yF) & set(yT)
return (x, x)
return doyarns
def unequalR(fTp, tTp):
return lambda n, m: n != m
def canonicalBeforeR(fTp, tTp):
return lambda n, m: Crank[n - 1] < Crank[m - 1]
def canonicalAfterR(fTp, tTp):
return lambda n, m: Crank[n - 1] > Crank[m - 1]
def spinSameSlots(fTp, tTp):
if fTp == slotType and tTp == slotType:
def doyarns(yF, yT):
x = set(yF) & set(yT)
return (x, x)
return doyarns
elif fTp == slotType or tTp == slotType:
def doyarns(yS, y2):
sindex = {}
for m in y2:
ss = Eoslots[m - maxSlot - 1]
if len(ss) == 1:
sindex.setdefault(ss[0], set()).add(m)
nyS = yS & set(sindex.keys())
ny2 = reduce(
set.union,
(sindex[s] for s in nyS),
set(),
)
return (nyS, ny2)
if fTp == slotType:
return doyarns
else:
def xx(yF, yT):
(nyT, nyF) = doyarns(yT, yF)
return (nyF, nyT)
return xx
else:
def doyarns(yF, yT):
sindexF = {}
for n in yF:
s = frozenset(Eoslots[n - maxSlot - 1])
sindexF.setdefault(s, set()).add(n)
sindexT = {}
for m in yT:
s = frozenset(Eoslots[m - maxSlot - 1])
sindexT.setdefault(s, set()).add(m)
nyS = set(sindexF.keys()) & set(sindexT.keys())
nyF = reduce(
set.union,
(sindexF[s] for s in nyS),
set(),
)
nyT = reduce(
set.union,
(sindexT[s] for s in nyS),
set(),
)
return (nyF, nyT)
return doyarns
def sameSlotsR(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n: (n, )
elif tTp == slotType:
return lambda n, m: Eoslots[n - maxSlot - 1] == (m, )
elif fTp == slotType:
return lambda n, m: Eoslots[m - maxSlot - 1] == (n, )
else:
return (
lambda n, m: (
frozenset(Eoslots[n-maxSlot-1]) ==
frozenset(Eoslots[m-maxSlot-1])
)
)
def spinOverlap(fTp, tTp):
if fTp == slotType and tTp == slotType:
def doyarns(yF, yT):
x = set(yF) & set(yT)
return (x, x)
return doyarns
elif fTp == slotType or tTp == slotType:
def doyarns(yS, y2):
sindex = {}
for m in y2:
for s in Eoslots[m - maxSlot - 1]:
sindex.setdefault(s, set()).add(m)
nyS = yS & set(sindex.keys())
ny2 = reduce(
set.union,
(sindex[s] for s in nyS),
set(),
)
return (nyS, ny2)
if fTp == slotType:
return doyarns
else:
def xx(yF, yT):
(nyT, nyF) = doyarns(yT, yF)
return (nyF, nyT)
return xx
else:
def doyarns(yF, yT):
REDUCE_FACTOR = 0.4
SIZE_LIMIT = 10000
sindexF = {}
for n in yF:
for s in Eoslots[n - maxSlot - 1]:
sindexF.setdefault(s, set()).add(n)
sindexT = {}
for m in yT:
for s in Eoslots[m - maxSlot - 1]:
sindexT.setdefault(s, set()).add(m)
nyS = set(sindexF.keys()) & set(sindexT.keys())
lsF = len(sindexF)
lsT = len(sindexT)
lsI = len(nyS)
if lsF == lsT: # spinning is completely useless
return (yF, yT)
if lsI > REDUCE_FACTOR * lsT and lsT > SIZE_LIMIT:
# spinning is not worth it
return (yF, yT)
if not self.silent:
self.api.info(
'1. reducing over {} elements'.format(len(nyS))
)
nyF = reduce(
set.union,
(sindexF[s] for s in nyS),
set(),
)
if not self.silent:
self.api.info(
'2. reducing over {} elements'.format(len(nyS))
)
nyT = reduce(
set.union,
(sindexT[s] for s in nyS),
set(),
)
return (nyF, nyT)
return doyarns
def overlapR(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n: (n, )
elif tTp == slotType:
return lambda n: Eoslots[n - maxSlot - 1]
elif fTp == slotType:
return lambda n, m: n in frozenset(Eoslots[m - maxSlot - 1])
else:
return (
lambda n, m: (
len(frozenset(Eoslots[n-maxSlot-1]) &
frozenset(Eoslots[m-maxSlot-1])) != 0
)
)
def diffSlotsR(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n, m: m != n
elif tTp == slotType:
return lambda n, m: Eoslots[m - maxSlot - 1] != (n, )
elif fTp == slotType:
return lambda n, m: Eoslots[n - maxSlot - 1] != (m, )
else:
return (
lambda n, m: (
frozenset(Eoslots[n-maxSlot-1]) !=
frozenset(Eoslots[m-maxSlot-1])
)
)
def disjointSlotsR(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n, m: m != n
elif tTp == slotType:
return lambda n, m: m not in frozenset(Eoslots[n-maxSlot-1])
elif fTp == slotType:
return lambda n, m: n not in frozenset(Eoslots[m-maxSlot-1])
else:
return (
lambda n, m: (
len(frozenset(Eoslots[n-maxSlot-1]) &
frozenset(Eoslots[m-maxSlot-1])) == 0
)
)
def inR(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n: ()
elif tTp == slotType:
return lambda n: ()
elif fTp == slotType:
return lambda n: ClevUp[n - 1]
else:
return lambda n: ClevUp[n - 1]
def hasR(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n: ()
elif fTp == slotType:
return lambda n: ()
elif tTp == slotType:
return lambda n: Eoslots[n - maxSlot - 1]
else:
return lambda n: ClevDown[n - maxSlot - 1]
def slotBeforeR(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n, m: n < m
elif fTp == slotType:
return lambda n, m: n < Eoslots[m - maxSlot - 1][0]
elif tTp == slotType:
return lambda n, m: Eoslots[n - maxSlot - 1][-1] < m
else:
return (
lambda n, m: (
Eoslots[n - maxSlot - 1][-1] <
Eoslots[m - maxSlot - 1][0]
)
)
def slotAfterR(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n, m: n > m
elif fTp == slotType:
return lambda n, m: n > Eoslots[m - maxSlot - 1][-1]
elif tTp == slotType:
return lambda n, m: Eoslots[n - maxSlot - 1][0] > m
else:
return (
lambda n, m: (
Eoslots[n - maxSlot - 1][0] >
Eoslots[m - maxSlot - 1][-1]
)
)
def sameFirstSlotR(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n: (n, )
elif fTp == slotType:
return lambda n: CfirstSlots[n - 1]
elif tTp == slotType:
return lambda n: (Eoslots[n - maxSlot - 1][0], )
else:
def xx(n):
fn = Eoslots[n - maxSlot - 1][0]
return CfirstSlots[fn - 1]
return xx
def sameLastSlotR(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n: (n, )
elif fTp == slotType:
return lambda n: ClastSlots[n - 1]
elif tTp == slotType:
return lambda n: (Eoslots[n - maxSlot - 1][-1], )
else:
def xx(n):
ln = Eoslots[n - maxSlot - 1][-1]
return ClastSlots[ln - 1]
return xx
def sameBoundaryR(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n: (n, )
elif fTp == slotType:
def xx(n):
fok = set(CfirstSlots[n - 1])
lok = set(ClastSlots[n - 1])
return tuple(fok & lok)
return xx
elif tTp == slotType:
def xx(n):
slots = Eoslots[n - maxSlot - 1]
fs = slots[0]
ls = slots[-1]
return (fs, ) if fs == ls else ()
return xx
else:
def xx(n):
fn = Eoslots[n - maxSlot - 1][0]
ln = Eoslots[n - maxSlot - 1][-1]
fok = set(CfirstSlots[fn - 1])
lok = set(ClastSlots[ln - 1])
return tuple(fok & lok)
return xx
def nearFirstSlotR(k):
def zz(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n: tuple(
m for m in range(max((1, n-k)), min((maxSlot, n+k+1)))
)
elif fTp == slotType:
def xx(n):
near = set(
l
for l in
range(max((1, n - k)), min((maxSlot, n + k + 1)))
)
return tuple(
reduce(
set.union,
(set(CfirstSlots[l - 1]) for l in near),
set(),
)
)
return xx
elif tTp == slotType:
def xx(n):
fn = Eoslots[n - maxSlot - 1][0]
return tuple(
m
for m in range(
max((1, fn - k)), min((maxSlot, fn + k + 1))
)
)
return xx
else:
def xx(n):
fn = Eoslots[n - maxSlot - 1][0]
near = set(
l
for l in range(
max((1, fn - k)), min((maxSlot, fn + k + 1))
)
)
return tuple(
reduce(
set.union,
(set(CfirstSlots[l - 1]) for l in near),
set(),
)
)
return xx
return zz
def nearLastSlotR(k):
def zz(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n: tuple(
m for m in range(max((1, n-k)), min((maxSlot, n+k+1)))
)
elif fTp == slotType:
def xx(n):
near = set(
l
for l in
range(max((1, n - k)), min((maxSlot, n + k + 1)))
)
return tuple(
reduce(
set.union,
(set(ClastSlots[l - 1]) for l in near),
set(),
)
)
return xx
elif tTp == slotType:
def xx(n):
ln = Eoslots[n - maxSlot - 1][-1]
return tuple(
m
for m in range(
max((1, ln - k)), min((maxSlot, ln + k + 1))
)
)
return xx
else:
def xx(n):
ln = Eoslots[n - maxSlot - 1][-1]
near = set(
l
for l in range(
max((1, ln - k)), min((maxSlot, ln + k + 1))
)
)
return tuple(
reduce(
set.union,
(set(ClastSlots[l - 1]) for l in near),
set(),
)
)
return xx
return zz
def nearBoundaryR(k):
def zz(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n: tuple(
m for m in range(max((1, n-k)), min((maxSlot, n+k+1)))
)
elif fTp == slotType:
def xx(n):
near = set(
l
for l in
range(max((1, n - k)), min((maxSlot, n + k + 1)))
)
fok = set(
reduce(
set.union,
(set(CfirstSlots[l - 1]) for l in near),
set(),
)
)
lok = set(
reduce(
set.union,
(set(ClastSlots[l - 1]) for l in near),
set(),
)
)
return tuple(fok & lok)
return xx
elif tTp == slotType:
def xx(n):
slots = Eoslots[n - maxSlot - 1]
fs = slots[0]
ls = slots[-1]
fok = set(
m
for m in range(
max((1, fs - k)), min((maxSlot, fs + k + 1))
)
)
lok = set(
m
for m in range(
max((1, ls - k)), min((maxSlot, ls + k + 1))
)
)
return tuple(fok & lok)
return xx
else:
def xx(n):
fn = Eoslots[n - maxSlot - 1][0]
ln = Eoslots[n - maxSlot - 1][-1]
nearf = set(
ls
for ls in range(
max((1, fn - k)), min((maxSlot, fn + k + 1))
)
)
nearl = set(
ls
for ls in range(
max((1, ln - k)), min((maxSlot, ln + k + 1))
)
)
fok = set(CfirstSlots[fn - 1])
lok = set(ClastSlots[ln - 1])
fok = set(
reduce(
set.union,
(set(CfirstSlots[ls - 1]) for ls in nearf),
set(),
)
)
lok = set(
reduce(
set.union,
(set(ClastSlots[ls - 1]) for ls in nearl),
set(),
)
)
return tuple(fok & lok)
return xx
return zz
def adjBeforeR(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n: (n + 1, ) if n < maxSlot else ()
else:
def xx(n):
if n == maxSlot:
return ()
myNext = n + 1 if n < maxSlot else Eoslots[n - maxSlot -
1][-1] + 1
if myNext > maxSlot:
return ()
return CfirstSlots[myNext - 1] + (myNext, )
return xx
def adjAfterR(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n: (n - 1, ) if n > 1 else ()
else:
def xx(n):
if n <= 1:
return ()
myPrev = n - 1 if n < maxSlot else Eoslots[n - maxSlot -
1][0] - 1
if myPrev <= 1:
return ()
return (myPrev, ) + ClastSlots[myPrev - 1]
return xx
def nearBeforeR(k):
def zz(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n: tuple(
m for m in
range(max((1, n+1-k)), min((maxSlot, n+1+k+1)))
)
else:
def xx(n):
myNext = n + 1 if n < maxSlot else Eoslots[n - maxSlot
- 1][-1] + 1
myNextNear = tuple(
ls
for ls in range(
max((1, myNext - k
)), min((maxSlot, myNext + k + 1))
)
)
nextSet = set(
reduce(
set.union,
(
set(CfirstSlots[ls - 1])
for ls in myNextNear
),
set(),
)
)
return tuple(nextSet) + myNextNear
return xx
return zz
def nearAfterR(k):
def zz(fTp, tTp):
if fTp == slotType and tTp == slotType:
return lambda n: tuple(
m for m in
range(max((1, n-1-k)), min((maxSlot, n-1+k+1)))
)
else:
def xx(n):
myPrev = n - 1 if n < maxSlot else Eoslots[n - maxSlot
- 1][0] - 1
myPrevNear = tuple(
l
for l in range(
max((1, myPrev - k
)), min((maxSlot, myPrev + k + 1))
)
)
prevSet = set(
reduce(
set.union,
(set(ClastSlots[l - 1]) for l in myPrevNear),
set(),
)
)
return tuple(prevSet) + myPrevNear
return xx
return zz
def makeEdgeMaps(efName):
def edgeR(ftP, tTp):
Es = self.api.Es
return lambda n: Es(efName).f(n)
def edgeIR(ftP, tTp):
Es = self.api.Es
return lambda n: Es(efName).t(n)
return (edgeR, edgeIR)
relations = [
(
('=', spinEqual, equalR, 'left equal to right (as node)'),
('=', spinEqual, equalR, None),
),
(
('#', 0.999, unequalR, 'left unequal to right (as node)'),
('#', 0.999, unequalR, None),
),
(
(
'<', 0.500, canonicalBeforeR,
'left before right (in canonical node ordering)'
),
(
'>', 0.500, canonicalAfterR,
'left after right (in canonical node ordering)'
),
),
(
(
'==', spinSameSlots, sameSlotsR,
'left occupies same slots as right'
),
('==', spinSameSlots, sameSlotsR, None),
),
(
(
'&&', spinOverlap, overlapR,
'left has overlapping slots with right'
),
('&&', spinOverlap, overlapR, None),
),
(
(
'##', 0.990, diffSlotsR,
'left and right do not have the same slot set'
),
('##', 0.990, diffSlotsR, None),
),
(
(
'||', 0.900, disjointSlotsR,
'left and right do not have common slots'
),
('||', 0.900, disjointSlotsR, None),
),
(
('[[', True, hasR, 'left embeds right'),
(']]', True, inR, 'left embedded in right'),
),
(
('<<', 0.490, slotBeforeR, 'left completely before right'),
('>>', 0.490, slotAfterR, 'left completely after right'),
),
(
(
'=:', True, sameFirstSlotR,
'left and right start at the same slot'
),
('=:', True, sameFirstSlotR, None),
),
(
(
':=', True, sameLastSlotR,
'left and right end at the same slot'
),
(':=', True, sameLastSlotR, None),
),
(
(
'::', True, sameBoundaryR,
'left and right start and end at the same slot'
),
('::', True, sameBoundaryR, None),
),
(
('<:', True, adjBeforeR, 'left immediately before right'),
(':>', True, adjAfterR, 'left immediately after right'),
),
(
(
'=k:', True, nearFirstSlotR,
'left and right start at k-nearly the same slot'
),
('=k:', True, nearFirstSlotR, None),
),
(
(
':k=', True, nearLastSlotR,
'left and right end at k-nearly the same slot'
),
(':k=', True, nearLastSlotR, None),
),
(
(
':k:', True, nearBoundaryR,
'left and right start and end at k-near slots'
),
(':k:', True, nearBoundaryR, None),
),
(
('<k:', True, nearBeforeR, 'left k-nearly before right'),
(':k>', True, nearAfterR, 'left k-nearly after right'),
),
]
self.api.TF.explore(silent=self.silent)
edgeMap = {}
for efName in sorted(self.api.TF.featureSets['edges']):
if efName == WARP[1]:
continue
r = len(relations)
(edgeR, edgeIR) = makeEdgeMaps(efName)
relations.append((
(
'-{}>'.format(efName), True, edgeR,
'edge feature "{}"'.format(efName)
),
(
'<{}-'.format(efName), True, edgeIR,
'edge feature "{}" (opposite direction)'.format(efName)
),
))
edgeMap[2 * r] = (efName, 1)
edgeMap[2 * r + 1] = (efName, -1)
lr = len(relations)
relationsAll = []
for (r, rc) in relations:
relationsAll.extend([r, rc])
self.relations = [
dict(
acro=r[0],
spin=r[1],
func=r[2],
desc=r[3],
) for r in relationsAll
]
self.relationFromName = dict(((r['acro'], i)
for (i, r) in enumerate(self.relations)))
self.relationLegend = '\n'.join(
'{:>23} {}'.format(r['acro'], r['desc']) for r in self.relations
if r['desc'] is not None
)
self.relationLegend += '''
The warp feature "{}" cannot be used in searches.
One of the above relations on nodes and/or slots will suit you better!
'''.format(WARP[1])
self.converse = dict(
tuple((2 * i, 2 * i + 1)
for i in range(lr)) + tuple((2 * i + 1, 2 * i)
for i in range(lr))
)
self.edgeMap = edgeMap
def _addRelations(self, varRels):
relations = self.relations
tasks = collections.defaultdict(set)
for (acro, ks) in varRels.items():
j = self.relationFromName[acro]
ji = self.converse[j]
if ji < j:
(j, ji) = (ji, j)
acro = relations[j]['acro']
acroi = relations[ji]['acro']
tasks[(j, acro, ji, acroi)] |= ks
for ((j, acro, ji, acroi), ks) in tasks.items():
for k in ks:
newAcro = acro.replace('k', str(k))
newAcroi = acroi.replace('k', str(k))
r = relations[j]
ri = relations[ji]
lr = len(relations)
relations.extend([
dict(
acro=newAcro,
spin=r['spin'],
func=r['func'](k),
desc=r['desc'],
),
dict(
acro=newAcroi,
spin=ri['spin'],
func=ri['func'](k),
desc=ri['desc'],
),
])
self.relationFromName[newAcro] = lr
self.relationFromName[newAcroi] = lr + 1
self.converse[lr] = lr + 1
self.converse[lr + 1] = lr
# SYNTACTIC ANALYSIS OF SEARCH TEMPLATE ###
def _tokenize(self):
if not self.good:
return
opPat = '(?:[#&|\[\]<>:=-]+\S*)'
atomPat = '(\s*)([^ \t=!~]+)(?:(?:\s*\Z)|(?:\s+(.*)))$'
atomOpPat = (
'(\s*)({op})\s+([^ \t=!~]+)(?:(?:\s*\Z)|(?:\s+(.*)))$'.format(
op=opPat
)
)
opLinePat = '(\s*)({op})\s*$'.format(op=opPat)
namePat = '[A-Za-z0-9_-]+'
relPat = '^\s*({nm})\s+({op})\s+({nm})\s*$'.format(
nm=namePat, op=opPat
)
atomOpRe = re.compile(atomOpPat)
atomRe = re.compile(atomPat)
opLineRe = re.compile(opLinePat)
nameRe = re.compile('^{}$'.format(namePat))
relRe = re.compile(relPat)
whiteRe = re.compile('^\s*$')
nonePat = '^([a-zA-Z0-9-_]+)!\s*$'
noneRe = re.compile(nonePat)
def getFeatures(x, i):
features = {}
featureString = x.replace('\\ ', chr(1)) if x is not None else ''
featureList = featureString.split()
good = True
for featStr in featureList:
feat = featStr.replace(chr(1), ' ')
match = noneRe.match(feat)
if match:
(featN,) = match.groups()
featName = unesc(featN)
featVals = None
else:
match = reRe.match(feat)
if match:
(featN, valRe) = match.groups()
featName = unesc(featN)
try:
featVals = re.compile(valRe)
except Exception() as err:
self.badSyntax.append(
'Wrong regular expression in line'
f'{i}: "{err}"'
)
good = False
featVals = None
else:
featComps = feat.split('=', 1)
featName = unesc(featComps[0])
if len(featComps) == 1:
featVals = True
else:
featValList = featComps[1]
featVals = set(
unesc(featVal)
for featVal in featValList.split('|')
)
features[featName] = featVals
return features if good else None
searchLines = self.searchLines
tokens = []
allGood = True
for (i, line) in enumerate(searchLines):
if line.startswith('#') or whiteRe.match(line):
continue
good = False
for x in [True]:
match = opLineRe.match(line)
if match:
(indent, op) = match.groups()
tokens.append((i, 'atom', len(indent), op))
good = True
break
match = relRe.match(line)
if match:
tokens.append((
i, 'rel', match.group(1), match.group(2),
match.group(3)
))
good = True
break
matchOp = atomOpRe.match(esc(line))
if not matchOp:
match = atomRe.match(esc(line))
if matchOp:
(indent, op, atom, features) = matchOp.groups()
elif match:
op = None
(indent, atom, features) = match.groups()
if matchOp or match:
atomComps = atom.split(':', 1)
if len(atomComps) == 1:
name = ''
else:
name = unesc(atomComps[0])
atom = unesc(atomComps[1])
mt = nameRe.match(name)
if not mt:
self.badSyntax.append(
'Illegal name at line {}: "{}"'.format(
i, name
)
)
good = False
features = getFeatures(features, i)
if features is None:
good = False
else:
tokens.append(
(i, 'atom', len(indent), op, name, atom, features)
)
good = True
break
features = getFeatures(esc(line), i)
if features is None:
good = False
else:
tokens.append((i, 'feat', features))
good = True
break
if not good:
allGood = False
if allGood:
self.tokens = tokens
else:
self.good = False
def _syntax(self):
error = self.api.error
self.good = True
self.badSyntax = []
self.searchLines = self.searchTemplate.split('\n')
self._tokenize()
if not self.good:
for (i, line) in enumerate(self.searchLines):
error('{:>2} {}'.format(i, line), tm=False)
for eline in self.badSyntax:
error(eline, tm=False)
# SEMANTIC ANALYSIS OF SEARCH TEMPLATE ###
def _grammar(self):
if not self.good:
return
prevKind = None
good = True
qnames = {}
qnodes = []
qedges = []
edgeLine = {}
nodeLine = {}
tokens = sorted(
self.tokens,
key=lambda t: (len(self.tokens) + t[0]) if t[1] == 'rel' else t[0]
)
# atomStack is a stack of qnodes with their indent levels
# such that every next member is one level deeper
# and every member is the last qnode encountered at that level
# The stack is implemented as a dict,
# keyed by the indent, and valued by the qnode
atomStack = {}
for (i, kind, *fields) in tokens:
if kind == 'atom':
if len(fields) == 2:
(indent, op) = fields
else:
(indent, op, name, otype, features) = fields
qnodes.append((otype, features))
q = len(qnodes) - 1
nodeLine[q] = i
name = ':{}'.format(i) if name == '' else name
qnames[name] = q
if len(atomStack) == 0:
if indent > 0:
self.badSemantics.append(
'Unexpected indent at line {}: {}, expected {}'.
format(i, indent, 0)
)
good = False
if op is not None:
self.badSemantics.append((
'Lonely relation at line {}:'
' not allowed at outermost level'
).format(i))
good = False
if len(fields) > 2:
atomStack[0] = q
else:
atomNest = sorted(atomStack.items(), key=lambda x: x[0])
top = atomNest[-1]
if indent == top[0]:
# sibling of previous atom
if len(atomNest) > 1:
if len(fields) == 2:
# lonely operator:
# left is previous atom, right is parent atom
qedges.append((top[1], op, atomNest[-2][1]))
edgeLine[len(qedges) - 1] = i
else:
# take the qnode of the subtop of the
# atomStack, if there is one
qedges.append((q, ']]', atomNest[-2][1]))
edgeLine[len(qedges) - 1] = i
if op is not None:
qedges.append((top[1], op, q))
edgeLine[len(qedges) - 1] = i
elif indent > top[0]:
if len(fields) == 2:
self.badSemantics.append((
'Lonely relation at line {}:'
' not allowed as first child'
).format(i))
good = False
else:
# child of previous atom
qedges.append((q, ']]', top[1]))
edgeLine[len(qedges) - 1] = i
if op is not None:
qedges.append((top[1], op, q))
edgeLine[len(qedges) - 1] = i
else:
# outdent action:
# look up the proper parent in the stack
if indent not in atomStack:
# parent cannot be found: indentation error
self.badSemantics.append((
'Unexpected indent at line {}:'
' {}, expected one of {}'
).format(
i,
indent,
','.join(
str(at[0]) for at in atomNest
if at[0] < indent
),
))
good = False
else:
parents = [
at[1] for at in atomNest if at[0] < indent
]
if len(
parents
) != 0: # if not already at outermost level
if len(
fields
) == 2: # connect previous sibling to parent
qedges.append(
(atomStack[indent], op, parents[-1])
)
edgeLine[len(qedges) - 1] = i
else:
qedges.append((q, ']]', parents[-1]))
edgeLine[len(qedges) - 1] = i
if op is not None:
qedges.append(
(atomStack[indent], op, q)
)
edgeLine[len(qedges) - 1] = i
removeKeys = [
at[0] for at in atomNest if at[0] > indent
]
for rk in removeKeys:
del atomStack[rk]
atomStack[indent] = q
elif kind == 'feat':
features = fields[0]
if prevKind is not None and prevKind != 'atom':
self.badSemantics.append(
'Features without atom at line {}: "{}"'.format(
i, features
)
)
good = False
else:
qnodes[-1][1].update(features)
elif kind == 'rel':
(fName, opName, tName) = fields
f = qnames.get(fName, None)
t = qnames.get(tName, None)
namesGood = True
for (q, n) in ((f, fName), (t, tName)):
if q is None:
self.badSemantics.append(
'Relation with undefined name at line {}: "{}"'.
format(i, n)
)
namesGood = False
if not namesGood:
good = False
else:
qedges.append((f, opName, t))
edgeLine[len(qedges) - 1] = i
prevKind = kind
if good:
self.qnames = qnames
self.qnodes = qnodes
self.qedgesRaw = qedges
self.nodeLine = nodeLine
self.edgeLine = edgeLine
else:
self.good = False
def _validation(self):
if not self.good:
return
levels = self.api.C.levels.data
otypes = set(x[0] for x in levels)
qnodes = self.qnodes
nodeLine = self.nodeLine
edgeMap = self.edgeMap
edgeLine = self.edgeLine
relationFromName = self.relationFromName
# check the object types of atoms
otypesGood = True
for (q, (otype, xx)) in enumerate(qnodes):
if otype not in otypes:
self.badSemantics.append(
'Unknown object type in line {}: "{}"'.format(
nodeLine[q], otype
)
)
otypesGood = False
if not otypesGood:
self.badSemantics.append(
'Valid object types are: {}'.format(
','.join(x[0] for x in levels),
)
)
self.good = False
# check the feature names of feature specs
# and check the types of their values
missingFeatures = {}
wrongValues = {}
for (q, (xx, features)) in enumerate(qnodes):
for (fName, values) in features.items():
if fName not in self.api.TF.featureSets['nodes']:
missingFeatures.setdefault(fName, []).append(q)
else:
requiredType = self.api.TF.features[fName].dataType
if values is True:
continue
elif values is None:
continue
elif isinstance(values, reTp):
if requiredType == 'int':
wrongValues.setdefault(fName, {}).setdefault(
values, []
).append(q)
else:
valuesCast = set()
if requiredType == 'int':
for val in values:
try:
valCast = int(val)
except Exception:
valCast = val
wrongValues.setdefault(
fName, {}).setdefault(
val, []
).append(q)
valuesCast.add(valCast)
features[fName] = valuesCast
if len(missingFeatures):
for (fName, qs) in sorted(missingFeatures.items()):
self.badSemantics.append(
'Missing feature "{}" in line(s) {}'.format(
fName,
','.join(str(nodeLine[q]) for q in qs),
)
)
self.good = False
if len(wrongValues):
for (fName, wrongs) in sorted(wrongValues.items()):
self.badSemantics.append(
'Feature "{}" has wrong values:'.format(fName)
)
for (val, qs) in sorted(wrongs.items()):
self.badSemantics.append(
' "{}" is not a number: line(s) {}'.format(
val,
','.join(str(nodeLine[q]) for q in qs),
)
)
self.good = False
# check the relational operator token in edges
# and replace them by an index
# in the relations array of known relations
qedges = []
edgesGood = True
kRe = re.compile('^([^0-9]*)([0-9]+)([^0-9]*)$')
addRels = {}
for (e, (f, opName, t)) in enumerate(self.qedgesRaw):
match = kRe.findall(opName)
if len(match):
(pre, k, post) = match[0]
opNameK = '{}{}{}'.format(pre, 'k', post)
addRels.setdefault(opNameK, set()).add(int(k))
self._addRelations(addRels)
for (e, (f, opName, t)) in enumerate(self.qedgesRaw):
match = kRe.findall(opName)
if len(match):
(pre, k, post) = match[0]
opNameK = '{}{}{}'.format(pre, 'k', post)
rela = relationFromName.get(opName, None)
if rela is None:
self.badSemantics.append(
'Unknown relation in line {}: "{}"'.format(
edgeLine[e], opName
)
)
edgesGood = False
qedges.append((f, rela, t))
if not edgesGood:
self.badSemantics.append(
'Allowed relations:\n{}'.format(self.relationLegend)
)
self.good = False
self.qedges = qedges
# determine which node and edge features are not yet loaded,
# and load them
eFeatsUsed = set()
for (f, rela, t) in qedges:
efName = edgeMap.get(rela, (None, ))[0]
if efName is not None:
eFeatsUsed.add(efName)
nFeatsUsed = set()
for (xx, features) in qnodes:
for nfName in features:
nFeatsUsed.add(nfName)
if self.good:
needToLoad = []
for fName in sorted(eFeatsUsed | nFeatsUsed):
fObj = self.api.TF.features[fName]
if not fObj.dataLoaded or not hasattr(self.api.F, fName):
needToLoad.append((fName, fObj))
if len(needToLoad):
self.api.TF.load(
[x[0] for x in needToLoad], add=True, silent=True,
)
def _semantics(self):
self.badSemantics = []
if not self.good:
return
error = self.api.error
self._grammar()
if not self.good:
for (i, line) in enumerate(self.searchLines):
error('{:>2} {}'.format(i, line), tm=False)
for eline in self.badSemantics:
error(eline, tm=False)
return
self._validation()
if not self.good:
for (i, line) in enumerate(self.searchLines):
error('{:>2} {}'.format(i, line), tm=False)
for eline in self.badSemantics:
error(eline, tm=False)
# WORKING WITH THE SEARCH GRAPH ###
def _connectedness(self):
error = self.api.error
qnodes = self.qnodes
qedges = self.qedges
componentIndex = dict(((q, {q}) for q in range(len(qnodes))))
for (f, rela, t) in qedges:
if f != t:
componentIndex[f] |= componentIndex[t]
for u in componentIndex[f] - {f}:
componentIndex[u] = componentIndex[f]
components = sorted(set(frozenset(c) for c in componentIndex.values()))
componentIndex = {}
for c in components:
for q in c:
componentIndex[q] = c
componentEdges = {}
for (e, (f, rela, t)) in enumerate(qedges):
c = componentIndex[f]
componentEdges.setdefault(c, []).append(e)
self.components = []
for c in components:
self.components.append([sorted(c), componentEdges.get(c, [])])
lComps = len(self.components)
if lComps == 0:
error('Search without instructions. Tell me what to look for.')
self.good = False
elif lComps > 1:
error(
'More than one connected components ({}):'.format(
len(self.components)
)
)
error(
'Either run the subqueries one by one,'
' or connect the components by a relation',
tm=False
)
self.good = False
def _showNode(self, q, pos2=False):
qnodes = self.qnodes
yarns = self.yarns
space = ' ' * 19
nodeInfo = 'node {} {:>2}-{:<13} ({:>6} choices)'.format(
space,
q,
qnodes[q][0],
len(yarns[q]),
) if pos2 else 'node {:>2}-{:<13} {} ({:>6} choices)'.format(
q,
qnodes[q][0],
space,
len(yarns[q]),
)
self.api.info(nodeInfo, tm=False)
def _showEdge(self, e, dir):
qnodes = self.qnodes
qedges = self.qedges
converse = self.converse
relations = self.relations
spreads = self.spreads
spreadsC = self.spreadsC
(f, rela, t) = qedges[e]
if dir == -1:
(f, rela, t) = (t, converse[rela], f)
self.api.info(
'edge {:>2}-{:<13} {:^2} {:>2}-{:<13} ({:8.1f} choices)'.format(
f,
qnodes[f][0],
relations[rela]['acro'],
t,
qnodes[t][0],
spreads.get(e, -1) if dir == 1 else spreadsC.get(e, -1),
),
tm=False
)
def _showYarns(self):
for q in range(len(self.qnodes)):
self._showNode(q)
# SPINNING ###
def _spinAtom(self, q):
F = self.api.F
Fs = self.api.Fs
qnodes = self.qnodes
(otype, features) = qnodes[q]
featureList = sorted(features.items())
yarn = set()
for n in F.otype.s(otype):
good = True
for (ft, val) in featureList:
fval = Fs(ft).v(n)
if val is None:
if fval is not None:
good = False
break
elif val is True:
if fval is None:
good = False
break
elif type(val) is str or type(val) is int:
if fval != val:
good = False
break
elif isinstance(val, reTp):
if fval is None or not val.search(fval):
good = False
break
else:
if fval not in val:
good = False
break
if good:
yarn.add(n)
self.yarns[q] = yarn
def _spinAtoms(self):
qnodes = self.qnodes
for q in range(len(qnodes)):
self._spinAtom(q)
def _estimateSpreads(self, both=False):
TRY_LIMIT_F = 10
TRY_LIMIT_T = 10
qnodes = self.qnodes
relations = self.relations
converse = self.converse
qedges = self.qedges
yarns = self.yarns
self.spreadsC = {}
self.spreads = {}
for (e, (f, rela, t)) in enumerate(qedges):
tasks = [(f, rela, t, 1)]
if both:
tasks.append((t, converse[rela], f, -1))
for (tf, trela, tt, dir) in tasks:
s = relations[trela]['spin']
yarnF = yarns[tf]
yarnT = yarns[tt]
dest = self.spreads if dir == 1 else self.spreadsC
if type(s) is float:
# fixed estimates
dest[e] = len(yarnT) * s
continue
yarnF = list(yarnF)
yarnT = yarns[tt]
totalSpread = 0
yarnFl = len(yarnF)
if yarnFl < TRY_LIMIT_F:
triesn = yarnF
else:
triesn = set(
yarnF[randrange(yarnFl)] for n in range(TRY_LIMIT_F)
)
if len(triesn) == 0:
dest[e] = 0
else:
r = relations[trela]['func'](qnodes[tf][0], qnodes[tt][0])
nparams = len(signature(r).parameters)
if nparams == 1:
for n in triesn:
mFromN = set(r(n)) & yarnT
totalSpread += len(mFromN)
else:
for n in triesn:
thisSpread = 0
yarnTl = len(yarnT)
if yarnTl < TRY_LIMIT_T:
triesm = yarnT
else:
triesm = set(
list(yarnT)[randrange(yarnTl)]
for m in range(TRY_LIMIT_T)
)
if len(triesm) == 0:
thisSpread = 0
else:
for m in triesm:
if r(n, m):
thisSpread += 1
totalSpread += yarnTl * thisSpread / len(triesm)
dest[e] = totalSpread / len(triesn)
def _chooseEdge(self):
F = self.api.F
yarnFractionNode = {}
qnodes = self.qnodes
qedges = self.qedges
spreads = self.spreads
for (q, (otype, xx)) in enumerate(qnodes):
(begin, end) = F.otype.sInterval(otype)
nOtype = 1 + end - begin
nYarn = len(self.yarns[q])
yf = nYarn / nOtype
yarnFractionNode[q] = yf * yf
yarnFractionEdge = {}
for (e, (f, rela, t)) in enumerate(qedges):
if self.uptodate[e]:
continue
yarnFractionEdge[
e
] = yarnFractionNode[f] + yarnFractionNode[t] + spreads[e]
firstEdge = sorted(yarnFractionEdge.items(), key=lambda x: x[1])[0][0]
return firstEdge
def _spinEdge(self, e):
SPIN_LIMIT = 1000
qnodes = self.qnodes
relations = self.relations
yarns = self.yarns
spreads = self.spreads
qedges = self.qedges
uptodate = self.uptodate
(f, rela, t) = qedges[e]
yarnF = yarns[f]
yarnT = yarns[t]
uptodate[e] = True
# if the yarns around an edge are big,
# and the spread of the relation is
# also big, spinning costs an enormous amount of time,
# and will not help in reducing the search space.
# condition for skipping: spread times length from-yarn >= SPIN_LIMIT
if spreads[e] * len(yarnF) >= SPIN_LIMIT:
return
# for some basic relations we know that spinning is useless
s = relations[rela]['spin']
if type(s) is float:
return
# for other basic realtions we have an optimized spin function
# if type(s) is types.FunctionType:
if isinstance(s, types.FunctionType):
(newYarnF, newYarnT) = s(qnodes[f][0], qnodes[t][0])(yarnF, yarnT)
else:
r = relations[rela]['func'](qnodes[f][0], qnodes[t][0])
nparams = len(signature(r).parameters)
newYarnF = set()
newYarnT = set()
if nparams == 1:
for n in yarnF:
mFromN = set(r(n)) & yarnT
if len(mFromN):
newYarnT |= mFromN
newYarnF.add(n)
else:
for n in yarnF:
mFromN = set(m for m in yarnT if r(n, m))
if len(mFromN):
newYarnT |= mFromN
newYarnF.add(n)
affectedF = len(newYarnF) != len(yarns[f])
affectedT = len(newYarnT) != len(yarns[t])
uptodate[e] = True
for (oe, (of, orela, ot)) in enumerate(qedges):
if oe == e:
continue
if (affectedF and f in {of, ot}) or (affectedT and t in {of, ot}):
self.uptodate[oe] = False
self.yarns[f] = newYarnF
self.yarns[t] = newYarnT
def _spinEdges(self):
qnodes = self.qnodes
qedges = self.qedges
yarns = self.yarns
uptodate = self.uptodate
self._estimateSpreads()
for e in range(len(qedges)):
uptodate[e] = False
it = 0
while 1:
if min(len(yarns[q]) for q in range(len(qnodes))) == 0:
break
# if reduce(
# lambda y,z: y and z,
# (uptodate[e] for e in range(len(qedges))),
# True,
# ): break
if all(uptodate[e] for e in range(len(qedges))):
break
e = self._chooseEdge()
(f, rela, t) = qedges[e]
self._spinEdge(e)
it += 1
# STITCHING: STRATEGIES ###
def _spread_1_first(self):
qedges = self.qedges
qnodes = self.qnodes
s1Edges = []
for (e, (f, rela, t)) in enumerate(qedges):
if self.spreads[e] <= 1:
s1Edges.append((e, 1))
if self.spreadsC[e] <= 1:
s1Edges.append((e, -1))
# s1Edges contain all edges with spread <= 1, or whose converse has spread <= 1
# now we want to build the largest graph
# with the original nodes and these edges,
# such that you can walk from a starting point
# over directed s1 edges to every other point
# we initialize candidate graphs: for each node: singletons graph, no edges.
candidates = []
# add s1 edges and nodes to all candidates
for q in range(len(qnodes)):
cnodes = {q}
cedges = set()
cedgesOrder = []
while 1:
added = False
for (e, dir) in s1Edges:
(f, rela, t) = qedges[e]
if dir == -1:
(t, f) = (f, t)
if f in cnodes:
if t not in cnodes:
cnodes.add(t)
added = True
if (e, dir) not in cedges:
cedges.add((e, dir))
cedgesOrder.append((e, dir))
added = True
if not added:
break
candidates.append((cnodes, cedgesOrder))
# pick the biggest graph (nodes and edges count for 1)
startS1 = sorted(candidates, key=lambda x: len(x[0]) + len(x[1]))[-1]
newNodes = startS1[0]
newEdges = startS1[1]
doneEdges = set(e[0] for e in newEdges)
# we add all edges that are not yet in our startS1.
# we add them two-fold: also with converse,
# and we sort the result by spread
# then we start a big loop:
# in every iteration, we take the edge with smallest spread
# that can be connected
# to the graph under construction
# then we start a new iteration, because the graph has grown,
# and and new edges might
# have become connectable by that
# in order to fail early, we can also add edges
# if their from-nodes and to-nodes both have been
# targeted.
# That means: an earlier edge went to f,
# an other earlier edge went to t, and if we
# have an edge from f to t, we'd better add it now,
# since it is an extra constraint
# and by testing it here we can avoid a lot of work.
remainingEdges = set()
for e in range(len(qedges)):
if e not in doneEdges:
remainingEdges.add((e, 1))
remainingEdges.add((e, -1))
remainingEdgesO = sorted(
remainingEdges,
key=lambda e: (
self.spreads[e[0]] if e[1] == 1 else self.spreadsC[e[0]]
),
)
while 1:
added = False
for (e, dir) in remainingEdgesO:
if e in doneEdges:
continue
(f, rela, t) = qedges[e]
if dir == -1:
(f, t) = (t, f)
if f in newNodes and t in newNodes:
newEdges.append((e, dir))
doneEdges.add(e)
added = True
for (e, dir) in remainingEdgesO:
if e in doneEdges:
continue
(f, rela, t) = qedges[e]
if dir == -1:
(f, t) = (t, f)
if f in newNodes:
newNodes.add(t)
newEdges.append((e, dir))
doneEdges.add(e)
added = True
break
if not added:
break
self.newNodes = newNodes
self.newEdges = newEdges
def _small_choice_first(self):
# This strategy does not try to make a big subgraph of
# edges with spread 1.
# The problem is that before the edges work,
# the initial yarn may have an enormous spread.
# Here we try out the strategy of postponing
# broad choices as long as possible.
# The inituition is that while we are making smaller choices,
# constraints are encountered,
# severely limiting the broader choices later on.
# So, we pick the yarn with the least amount of nodes
# as our starting point.
# The corresponding node is our singleton start set.
# In every iteration we do the following:
# - we pick all edges of which from- and to-nodes
# are already in the node set
# - we pick the edge with least spread
# that has a starting point in the set
# Until nothing changes anymore
qedges = self.qedges
qnodes = self.qnodes
newNodes = {
sorted(range(len(qnodes)), key=lambda x: len(self.yarns[x]))[0]
}
newEdges = []
doneEdges = set()
remainingEdges = set()
for e in range(len(qedges)):
remainingEdges.add((e, 1))
remainingEdges.add((e, -1))
remainingEdgesO = sorted(
remainingEdges,
key=lambda e: (
self.spreads[e[0]] if e[1] == 1 else self.spreadsC[e[0]]
),
)
while 1:
added = False
for (e, dir) in remainingEdgesO:
if e in doneEdges:
continue
(f, rela, t) = qedges[e]
if dir == -1:
(f, t) = (t, f)
if f in newNodes and t in newNodes:
newEdges.append((e, dir))
doneEdges.add(e)
added = True
for (e, dir) in remainingEdgesO:
if e in doneEdges:
continue
(f, rela, t) = qedges[e]
if dir == -1:
(f, t) = (t, f)
if f in newNodes:
newNodes.add(t)
newEdges.append((e, dir))
doneEdges.add(e)
added = True
break
if not added:
break
self.newNodes = newNodes
self.newEdges = newEdges
def _big_choice_first(self):
# For comparison: the opposite of _small_choice_first.
# Just to see what the performance difference is.
qedges = self.qedges
qnodes = self.qnodes
newNodes = {
sorted(range(len(qnodes)), key=lambda x: -len(self.yarns[x]))[0]
}
newEdges = []
doneEdges = set()
remainingEdges = set()
for e in range(len(qedges)):
remainingEdges.add((e, 1))
remainingEdges.add((e, -1))
remainingEdgesO = sorted(
remainingEdges,
key=lambda e: (
-self.spreads[e[0]] if e[1] == 1 else -self.spreadsC[e[0]]
),
)
while 1:
added = False
for (e, dir) in remainingEdgesO:
if e in doneEdges:
continue
(f, rela, t) = qedges[e]
if dir == -1:
(f, t) = (t, f)
if f in newNodes and t in newNodes:
newEdges.append((e, dir))
doneEdges.add(e)
added = True
for (e, dir) in remainingEdgesO:
if e in doneEdges:
continue
(f, rela, t) = qedges[e]
if dir == -1:
(f, t) = (t, f)
if f in newNodes:
newNodes.add(t)
newEdges.append((e, dir))
doneEdges.add(e)
added = True
break
if not added:
break
self.newNodes = newNodes
self.newEdges = newEdges
# STITCHING ###
def _stitch(self):
self._estimateSpreads(both=True)
self._stitchPlan()
if not self.good:
return
self._stitchResults()
# STITCHING: PLANNING ###
def _stitchPlan(self, strategy=None):
qnodes = self.qnodes
qedges = self.qedges
error = self.api.error
self._setStrategy(strategy, keep=True)
if not self.good:
return
# Apply the chosen strategy
self.strategy()
# remove spurious edges:
# if we have both the 1 and -1 version of an edge,
# we can leave out the one that we encounter in the second place
newNodes = self.newNodes
newEdges = self.newEdges
newCedges = set()
newCedgesOrder = []
for (e, dir) in newEdges:
if e not in newCedges:
newCedgesOrder.append((e, dir))
newCedges.add(e)
# conjecture: we have all edges and all nodes now
# reason: we work in a connected component, so all nodes are reachable
# by edges or inverses
# we check nevertheless
qnodesO = tuple(range(len(qnodes)))
newNodesO = tuple(sorted(newNodes))
if newNodesO != qnodesO:
error(
'''Object mismatch in plan:
In template: {}
In plan : {}'''.format(qnodesO, newNodesO),
tm=False
)
self.good = False
qedgesO = tuple(range(len(qedges)))
newCedgesO = tuple(sorted(newCedges))
if newCedgesO != qedgesO:
error(
'''Relation mismatch in plan:
In template: {}
In plan : {}'''.format(qedgesO, newCedgesO),
tm=False
)
self.good = False
self.stitchPlan = (newNodes, newCedgesOrder)
# STITCHING: DELIVERING ###
def _stitchResults(self):
qnodes = self.qnodes
qedges = self.qedges
plan = self.stitchPlan
relations = self.relations
converse = self.converse
yarns = self.yarns
planEdges = plan[1]
if len(planEdges) == 0:
# no edges, hence a single node (because of connectedness,
# hence we must deliver everything of its yarn
yarn = yarns[0]
def deliver(remap=True):
for n in yarn:
yield (n, )
self.results = deliver
return
# The next function must be optimized, and the lookup of functions and data
# should be as direct as possible.
# Because deliver() below fetches the results,
# of wich there are unpredictably many.
# We are going to build-up and deliver stitches,
# which are instantiations of all the query nodes
# by text nodes in a specific sequence
# which is the same for all stitches.
# We can compile stitching in such a way, that the stitcher thinks it is
# instantiating q node 0, then 1, and so on.
# I.e. we are going to permute every thing that the stitching process sees,
# so that it happens in this order.
# We build up the stitch in a recursive process.
# When there is choice between a and b, we essentially say
#
# def build(stitch)
# if there is choice
# build(stitch+a)
# build(stitch+b)
#
# But we do not have to pass on the stitch as an immutable data structure.
# We can just keep it as one single mutable datastructure, provided we
# do something between the two recursive calls above.
# Suppose stitch is an array, and in the outer build n elements are filled
# (the rest contains -1)
#
# Then we say
# if there is choice
# build(stitch+a)
# for k in range(n, len(stitch)): stitch[k] = -1
# build(stitch+b)
#
# It turns out that the data in stitch that is shared between calls
# is not modified by them.
# The only thing that happens, is that -1 values get new values.
# So coming out calls only requires us to restore -1's.
# And if the stitch is ordered in the right way,
# the -1's are always at the end.
# We start compiling and permuting
edgesCompiled = []
qPermuted = [
] # row of nodes in the order as will be created during stitching
qPermutedInv = {
} # mapping from original q node number to index in the permuted order
for (i, (e, dir)) in enumerate(planEdges):
(f, rela, t) = qedges[e]
if dir == -1:
(f, rela, t) = (t, converse[rela], f)
r = relations[rela]['func'](qnodes[f][0], qnodes[t][0])
nparams = len(signature(r).parameters)
if i == 0:
qPermuted.append(f)
qPermutedInv[f] = len(qPermuted) - 1
if t not in qPermuted:
qPermuted.append(t)
qPermutedInv[t] = len(qPermuted) - 1
edgesCompiled.append(
(qPermutedInv[f], qPermutedInv[t], r, nparams)
)
# now permute the yarns
yarnsPermuted = [yarns[q] for q in qPermuted]
def deliver(remap=True):
stitch = [None for q in range(len(qPermuted))]
lStitch = len(stitch)
edgesC = edgesCompiled
yarnsP = yarnsPermuted
def stitchOn(e):
if e >= len(edgesC):
if remap:
yield tuple(
stitch[qPermutedInv[q]] for q in range(lStitch)
)
else:
yield tuple(stitch)
return
(f, t, r, nparams) = edgesC[e]
yarnT = yarnsP[t]
if e == 0 and stitch[f] is None:
yarnF = yarnsP[f]
for sN in yarnF:
stitch[f] = sN
for s in stitchOn(e):
yield s
return
sN = stitch[f]
sM = stitch[t]
if sM is not None:
if nparams == 1:
if sM in set(r(sN)): # & yarnT:
for s in stitchOn(e + 1):
yield s
else:
if r(sN, sM):
for s in stitchOn(e + 1):
yield s
return
mFromN = tuple(set(r(sN)) & yarnT) if nparams == 1 else tuple(
m for m in yarnT if r(sN, m)
)
for m in mFromN:
stitch[t] = m
for s in stitchOn(e + 1):
yield s
stitch[t] = None
for s in stitchOn(0):
yield s
self.results = deliver
# TOP-LEVEL IMPLEMENTATION METHODS
def _parse(self):
self._syntax()
self._semantics()
def _prepare(self):
if not self.good:
return
self.yarns = {}
self.spreads = {}
self.spreadsC = {}
self.uptodate = {}
self.results = None
self._connectedness()
def _setStrategy(self, strategy, keep=False):
error = self.api.error
if strategy is None:
if keep:
return
strategy = STRATEGY[0]
if strategy not in STRATEGY:
error('Strategy not defined: "{}"'.format(strategy))
error(
'Allowed strategies:\n{}'.format(
'\n'.join(' {}'.format(s) for s in STRATEGY)
),
tm=False
)
self.good = False
return
method = '_{}'.format(strategy)
if not hasattr(self, method):
error(
'Strategy is defined, but not implemented: "{}"'.
format(strategy)
)
self.good = False
return
self.strategy = getattr(self, method)