def test_annotations(self):
x = Symbol('x')
x_next = Symbol('x.next')
f = Iff(x, Not(x_next))
ann = Annotations()
ann.add(x, 'next', x_next.symbol_name())
ann.add(f, 'trans', 'true')
ann.add(x, 'init', 'true')
tree_buf = StringIO()
dag_buf = StringIO()
tree_printer = SmtPrinter(tree_buf, annotations=ann)
dag_printer = SmtDagPrinter(dag_buf, annotations=ann)
dag_printer.printer(f)
tree_printer.printer(f)
self.assertEqual(
tree_buf.getvalue(),
"(! (= (! x :next x.next :init true) (not x.next)) :trans true)")
self.assertEqual(
dag_buf.getvalue(),
"(let ((.def_0 (not x.next))) (let ((.def_1 (= (! x :next x.next :init true) .def_0))) (! .def_1 :trans true)))"
)
def extend_ts(ts, modifier):
affect_init = False
if ts.ftrans is None:
return (ts, [])
new_ftrans = {}
vars = []
for (assign, cond_assign_list) in ts.ftrans.items():
fv = get_free_variables(assign)
assert len(fv) == 1
var = fv.pop()
is_next = TS.has_next(var)
refvar = TS.get_ref_var(var)
nomvar = Symbol(NOMIN % refvar.symbol_name(), var.symbol_type())
fvar = Symbol(FAULT % refvar.symbol_name(), BOOL)
vars.append(nomvar)
vars.append(fvar)
repldic = dict([(refvar.symbol_name(), nomvar.symbol_name()), \
(TS.get_prime(refvar).symbol_name(), TS.get_prime(nomvar).symbol_name())])
# Remapping nominal behavior to new variable
new_ftrans[substitute(assign,
repldic)] = [(substitute(c[0], repldic),
substitute(c[1], repldic))
for c in cond_assign_list]
# Definition of the nominal behavior
new_ftrans[refvar] = [(Not(fvar), nomvar)]
# Application of the faulty behavior
new_ftrans[refvar].append(
(fvar, modifier.get_behavior(nomvar, refvar)))
ts.trans = And(ts.trans, Implies(fvar, TS.get_prime(fvar)))
if affect_init:
ts.init = substitute(ts.init, repldic)
else:
ts.init = And(ts.init, Not(fvar))
# add the vars to the transition system
for var in vars:
ts.add_var(var)
ts.ftrans = new_ftrans
return (ts, vars)
class SMTExplainer(object):
"""
An SMT-inspired minimal explanation extractor for XGBoost models.
"""
def __init__(self, formula, intvs, imaps, ivars, feats, nof_classes,
options, xgb):
"""
Constructor.
"""
self.feats = feats
self.intvs = intvs
self.imaps = imaps
self.ivars = ivars
self.nofcl = nof_classes
self.optns = options
self.idmgr = IDPool()
# saving XGBooster
self.xgb = xgb
self.verbose = self.optns.verb
self.oracle = Solver(name=options.solver)
self.inps = [] # input (feature value) variables
for f in self.xgb.extended_feature_names_as_array_strings:
if '_' not in f:
self.inps.append(Symbol(f, typename=REAL))
else:
self.inps.append(Symbol(f, typename=BOOL))
self.outs = [] # output (class score) variables
for c in range(self.nofcl):
self.outs.append(Symbol('class{0}_score'.format(c), typename=REAL))
# theory
self.oracle.add_assertion(formula)
# current selector
self.selv = None
# save and use dual explanations whenever needed
self.dualx = []
# number of oracle calls involved
self.calls = 0
def prepare(self, sample):
"""
Prepare the oracle for computing an explanation.
"""
if self.selv:
# disable the previous assumption if any
self.oracle.add_assertion(Not(self.selv))
# creating a fresh selector for a new sample
sname = ','.join([str(v).strip() for v in sample])
# the samples should not repeat; otherwise, they will be
# inconsistent with the previously introduced selectors
assert sname not in self.idmgr.obj2id, 'this sample has been considered before (sample {0})'.format(
self.idmgr.id(sname))
self.selv = Symbol('sample{0}_selv'.format(self.idmgr.id(sname)),
typename=BOOL)
self.rhypos = [] # relaxed hypotheses
# transformed sample
self.sample = list(self.xgb.transform(sample)[0])
self.sel2fid = {} # selectors to original feature ids
self.sel2vid = {} # selectors to categorical feature ids
# preparing the selectors
for i, (inp, val) in enumerate(zip(self.inps, self.sample), 1):
feat = inp.symbol_name().split('_')[0]
selv = Symbol('selv_{0}'.format(feat))
val = float(val)
self.rhypos.append(selv)
if selv not in self.sel2fid:
self.sel2fid[selv] = int(feat[1:])
self.sel2vid[selv] = [i - 1]
else:
self.sel2vid[selv].append(i - 1)
# adding relaxed hypotheses to the oracle
if not self.intvs:
for inp, val, sel in zip(self.inps, self.sample, self.rhypos):
if '_' not in inp.symbol_name():
hypo = Implies(self.selv,
Implies(sel, Equals(inp, Real(float(val)))))
else:
hypo = Implies(self.selv,
Implies(sel, inp if val else Not(inp)))
self.oracle.add_assertion(hypo)
else:
for inp, val, sel in zip(self.inps, self.sample, self.rhypos):
inp = inp.symbol_name()
# determining the right interval and the corresponding variable
for ub, fvar in zip(self.intvs[inp], self.ivars[inp]):
if ub == '+' or val < ub:
hypo = Implies(self.selv, Implies(sel, fvar))
break
self.oracle.add_assertion(hypo)
# in case of categorical data, there are selector duplicates
# and we need to remove them
self.rhypos = sorted(set(self.rhypos),
key=lambda x: int(x.symbol_name()[6:]))
# propagating the true observation
if self.oracle.solve([self.selv] + self.rhypos):
model = self.oracle.get_model()
else:
assert 0, 'Formula is unsatisfiable under given assumptions'
# choosing the maximum
outvals = [float(model.get_py_value(o)) for o in self.outs]
maxoval = max(zip(outvals, range(len(outvals))))
# correct class id (corresponds to the maximum computed)
self.out_id = maxoval[1]
self.output = self.xgb.target_name[self.out_id]
# forcing a misclassification, i.e. a wrong observation
disj = []
for i in range(len(self.outs)):
if i != self.out_id:
disj.append(GT(self.outs[i], self.outs[self.out_id]))
self.oracle.add_assertion(Implies(self.selv, Or(disj)))
if self.verbose:
inpvals = self.xgb.readable_sample(sample)
self.preamble = []
for f, v in zip(self.xgb.feature_names, inpvals):
if f not in v:
self.preamble.append('{0} = {1}'.format(f, v))
else:
self.preamble.append(v)
print(' explaining: "IF {0} THEN {1}"'.format(
' AND '.join(self.preamble), self.output))
def explain(self, sample, smallest):
"""
Hypotheses minimization.
"""
# reinitializing the number of used oracle calls
# 1 because of the initial call checking the entailment
self.calls = 1
# adapt the solver to deal with the current sample
self.prepare(sample)
# saving external explanation to be minimized further
self.to_consider = [True for h in self.rhypos]
self.time = resource.getrusage(resource.RUSAGE_CHILDREN).ru_utime + \
resource.getrusage(resource.RUSAGE_SELF).ru_utime
# if satisfiable, then the observation is not implied by the hypotheses
if self.oracle.solve(
[self.selv] +
[h for h, c in zip(self.rhypos, self.to_consider) if c]):
print(' no implication!')
print(self.oracle.get_model())
sys.exit(1)
if self.optns.xtype == 'abductive':
# abductive explanations => MUS computation and enumeration
if not smallest and self.optns.xnum == 1:
expls = [self.compute_minimal_abductive()]
else:
expls = self.enumerate_abductive(smallest=smallest)
else: # contrastive explanations => MCS enumeration
if self.optns.usemhs:
expls = self.enumerate_contrastive()
else:
if not smallest:
expls = self.enumerate_minimal_contrastive()
else:
# expls = self.enumerate_smallest_contrastive()
expls = self.enumerate_contrastive()
self.time = resource.getrusage(resource.RUSAGE_CHILDREN).ru_utime + \
resource.getrusage(resource.RUSAGE_SELF).ru_utime - self.time
expls = list(
map(lambda expl: sorted([self.sel2fid[h] for h in expl]), expls))
if self.dualx:
self.dualx = list(
map(lambda expl: sorted([self.sel2fid[h] for h in expl]),
self.dualx))
if self.verbose:
if expls[0] != None:
for expl in expls:
preamble = [self.preamble[i] for i in expl]
if self.optns.xtype == 'abductive':
print(' explanation: "IF {0} THEN {1}"'.format(
' AND '.join(preamble),
self.xgb.target_name[self.out_id]))
else:
print(
' explanation: "IF NOT {0} THEN NOT {1}"'.format(
' AND NOT '.join(preamble),
self.xgb.target_name[self.out_id]))
print(' # hypos left:', len(expl))
print(' time: {0:.2f}'.format(self.time))
# here we return the last computed explanation
return expls
def compute_minimal_abductive(self):
"""
Compute any subset-minimal explanation.
"""
i = 0
# filtering out unnecessary features if external explanation is given
rhypos = [h for h, c in zip(self.rhypos, self.to_consider) if c]
# simple deletion-based linear search
while i < len(rhypos):
to_test = rhypos[:i] + rhypos[(i + 1):]
self.calls += 1
if self.oracle.solve([self.selv] + to_test):
i += 1
else:
rhypos = to_test
return rhypos
def enumerate_minimal_contrastive(self):
"""
Compute a subset-minimal contrastive explanation.
"""
def _overapprox():
model = self.oracle.get_model()
for sel in self.rhypos:
if int(model.get_py_value(sel)) > 0:
# soft clauses contain positive literals
# so if var is true then the clause is satisfied
self.ss_assumps.append(sel)
else:
self.setd.append(sel)
def _compute():
i = 0
while i < len(self.setd):
if self.optns.usecld:
_do_cld_check(self.setd[i:])
i = 0
if self.setd:
# it may be empty after the clause D check
self.calls += 1
self.ss_assumps.append(self.setd[i])
if not self.oracle.solve([self.selv] + self.ss_assumps +
self.bb_assumps):
self.ss_assumps.pop()
self.bb_assumps.append(Not(self.setd[i]))
i += 1
def _do_cld_check(cld):
self.cldid += 1
sel = Symbol('{0}_{1}'.format(self.selv.symbol_name(), self.cldid))
cld.append(Not(sel))
# adding clause D
self.oracle.add_assertion(Or(cld))
self.ss_assumps.append(sel)
self.setd = []
st = self.oracle.solve([self.selv] + self.ss_assumps +
self.bb_assumps)
self.ss_assumps.pop() # removing clause D assumption
if st == True:
model = self.oracle.get_model()
for l in cld[:-1]:
# filtering all satisfied literals
if int(model.get_py_value(l)) > 0:
self.ss_assumps.append(l)
else:
self.setd.append(l)
else:
# clause D is unsatisfiable => all literals are backbones
self.bb_assumps.extend([Not(l) for l in cld[:-1]])
# deactivating clause D
self.oracle.add_assertion(Not(sel))
# sets of selectors to work with
self.cldid = 0
expls = []
# detect and block unit-size MCSes immediately
if self.optns.unitmcs:
for i, hypo in enumerate(self.rhypos):
self.calls += 1
if self.oracle.solve([self.selv] + self.rhypos[:i] +
self.rhypos[(i + 1):]):
expls.append([hypo])
if len(expls) != self.optns.xnum:
self.oracle.add_assertion(Or([Not(self.selv), hypo]))
else:
break
self.calls += 1
while self.oracle.solve([self.selv]):
self.ss_assumps, self.bb_assumps, self.setd = [], [], []
_overapprox()
_compute()
expl = [list(f.get_free_variables())[0] for f in self.bb_assumps]
expls.append(expl)
if len(expls) == self.optns.xnum:
break
self.oracle.add_assertion(Or([Not(self.selv)] + expl))
self.calls += 1
self.calls += self.cldid
return expls if expls else [None]
def enumerate_abductive(self, smallest=True):
"""
Compute a cardinality-minimal explanation.
"""
# result
expls = []
# just in case, let's save dual (contrastive) explanations
self.dualx = []
with Hitman(bootstrap_with=[[
i for i in range(len(self.rhypos)) if self.to_consider[i]
]],
htype='sorted' if smallest else 'lbx') as hitman:
# computing unit-size MCSes
for i, hypo in enumerate(self.rhypos):
if self.to_consider[i] == False:
continue
self.calls += 1
if self.oracle.solve([self.selv] + self.rhypos[:i] +
self.rhypos[(i + 1):]):
hitman.hit([i])
self.dualx.append([self.rhypos[i]])
# main loop
iters = 0
while True:
hset = hitman.get()
iters += 1
if self.verbose > 1:
print('iter:', iters)
print('cand:', hset)
if hset == None:
break
self.calls += 1
if self.oracle.solve([self.selv] +
[self.rhypos[i] for i in hset]):
to_hit = []
satisfied, unsatisfied = [], []
removed = list(
set(range(len(self.rhypos))).difference(set(hset)))
model = self.oracle.get_model()
for h in removed:
i = self.sel2fid[self.rhypos[h]]
if '_' not in self.inps[i].symbol_name():
# feature variable and its expected value
var, exp = self.inps[i], self.sample[i]
# true value
true_val = float(model.get_py_value(var))
if not exp - 0.001 <= true_val <= exp + 0.001:
unsatisfied.append(h)
else:
hset.append(h)
else:
for vid in self.sel2vid[self.rhypos[h]]:
var, exp = self.inps[vid], int(
self.sample[vid])
# true value
true_val = int(model.get_py_value(var))
if exp != true_val:
unsatisfied.append(h)
break
else:
hset.append(h)
# computing an MCS (expensive)
for h in unsatisfied:
self.calls += 1
if self.oracle.solve([self.selv] +
[self.rhypos[i] for i in hset] +
[self.rhypos[h]]):
hset.append(h)
else:
to_hit.append(h)
if self.verbose > 1:
print('coex:', to_hit)
hitman.hit(to_hit)
self.dualx.append([self.rhypos[i] for i in to_hit])
else:
if self.verbose > 1:
print('expl:', hset)
expl = [self.rhypos[i] for i in hset]
expls.append(expl)
if len(expls) != self.optns.xnum:
hitman.block(hset)
else:
break
return expls
def enumerate_smallest_contrastive(self):
"""
Compute a cardinality-minimal contrastive explanation.
"""
# result
expls = []
# computing unit-size MUSes
muses = set([])
for hypo in self.rhypos:
self.calls += 1
if not self.oracle.solve([self.selv, hypo]):
muses.add(hypo)
# we are going to discard unit-size MUSes from consideration
rhypos = set(self.rhypos).difference(muses)
# introducing interer cost literals for rhypos
costlits = []
for i, hypo in enumerate(rhypos):
costlit = Symbol(name='costlit_{0}_{1}'.format(
self.selv.symbol_name(), i),
typename=INT)
costlits.append(costlit)
self.oracle.add_assertion(
Ite(hypo, Equals(costlit, Int(0)), Equals(costlit, Int(1))))
# main loop (linear search unsat-sat)
i = 0
while i < len(rhypos) and len(expls) != self.optns.xnum:
# fresh selector for the current iteration
sit = Symbol('iter_{0}_{1}'.format(self.selv.symbol_name(), i))
# adding cardinality constraint
self.oracle.add_assertion(Implies(sit, LE(Plus(costlits), Int(i))))
# extracting explanations from MaxSAT models
while self.oracle.solve([self.selv, sit]):
self.calls += 1
model = self.oracle.get_model()
expl = []
for hypo in rhypos:
if int(model.get_py_value(hypo)) == 0:
expl.append(hypo)
# each MCS contains all unit-size MUSes
expls.append(list(muses) + expl)
# either stop or add a blocking clause
if len(expls) != self.optns.xnum:
self.oracle.add_assertion(Implies(self.selv, Or(expl)))
else:
break
i += 1
self.calls += 1
return expls
def enumerate_contrastive(self, smallest=True):
"""
Compute a cardinality-minimal contrastive explanation.
"""
# core extraction is done via calling Z3's internal API
assert self.optns.solver == 'z3', 'This procedure requires Z3'
# result
expls = []
# just in case, let's save dual (abductive) explanations
self.dualx = []
# mapping from hypothesis variables to their indices
hmap = {h: i for i, h in enumerate(self.rhypos)}
# mapping from internal Z3 variable into variables of PySMT
vmap = {self.oracle.converter.convert(v): v for v in self.rhypos}
vmap[self.oracle.converter.convert(self.selv)] = None
def _get_core():
core = self.oracle.z3.unsat_core()
return sorted(filter(lambda x: x != None,
map(lambda x: vmap[x], core)),
key=lambda x: int(x.symbol_name()[6:]))
def _do_trimming(core):
for i in range(self.optns.trim):
self.calls += 1
self.oracle.solve([self.selv] + core)
new_core = _get_core()
if len(core) == len(new_core):
break
return new_core
def _reduce_lin(core):
def _assump_needed(a):
if len(to_test) > 1:
to_test.remove(a)
self.calls += 1
if not self.oracle.solve([self.selv] + list(to_test)):
return False
to_test.add(a)
return True
else:
return True
to_test = set(core)
return list(filter(lambda a: _assump_needed(a), core))
def _reduce_qxp(core):
coex = core[:]
filt_sz = len(coex) / 2.0
while filt_sz >= 1:
i = 0
while i < len(coex):
to_test = coex[:i] + coex[(i + int(filt_sz)):]
self.calls += 1
if to_test and not self.oracle.solve([self.selv] +
to_test):
# assumps are not needed
coex = to_test
else:
# assumps are needed => check the next chunk
i += int(filt_sz)
# decreasing size of the set to filter
filt_sz /= 2.0
if filt_sz > len(coex) / 2.0:
# next size is too large => make it smaller
filt_sz = len(coex) / 2.0
return coex
def _reduce_coex(core):
if self.optns.reduce == 'lin':
return _reduce_lin(core)
else: # qxp
return _reduce_qxp(core)
with Hitman(bootstrap_with=[[
i for i in range(len(self.rhypos)) if self.to_consider[i]
]],
htype='sorted' if smallest else 'lbx') as hitman:
# computing unit-size MUSes
for i, hypo in enumerate(self.rhypos):
if self.to_consider[i] == False:
continue
self.calls += 1
if not self.oracle.solve([self.selv, self.rhypos[i]]):
hitman.hit([i])
self.dualx.append([self.rhypos[i]])
elif self.optns.unitmcs:
self.calls += 1
if self.oracle.solve([self.selv] + self.rhypos[:i] +
self.rhypos[(i + 1):]):
# this is a unit-size MCS => block immediately
hitman.block([i])
expls.append([self.rhypos[i]])
# main loop
iters = 0
while True:
hset = hitman.get()
iters += 1
if self.verbose > 1:
print('iter:', iters)
print('cand:', hset)
if hset == None:
break
self.calls += 1
if not self.oracle.solve([self.selv] + [
self.rhypos[h] for h in list(
set(range(len(self.rhypos))).difference(set(hset)))
]):
to_hit = _get_core()
if len(to_hit) > 1 and self.optns.trim:
to_hit = _do_trimming(to_hit)
if len(to_hit) > 1 and self.optns.reduce != 'none':
to_hit = _reduce_coex(to_hit)
self.dualx.append(to_hit)
to_hit = [hmap[h] for h in to_hit]
if self.verbose > 1:
print('coex:', to_hit)
hitman.hit(to_hit)
else:
if self.verbose > 1:
print('expl:', hset)
expl = [self.rhypos[i] for i in hset]
expls.append(expl)
if len(expls) != self.optns.xnum:
hitman.block(hset)
else:
break
return expls
