def test_close_at_6_sig_figs_exhaustive(self, M=6):
for a in np.arange(-9.0, 9.0, 0.01):
for offset in np.arange(5.000001e-6, 2 * 5.000001e-6, 1e-6):
isEqual = closeAtMSigFigs(a, a + offset, M=M, tol=5)
assert not isEqual
for offset in np.arange(-5e-6, 5e-6 + EPS, 1e-6):
isEqual = closeAtMSigFigs(a, a + offset, M=M, tol=5)
if not isEqual:
print a
print a + offset
assert isEqual
def test_close_at_6_sig_figs_exhaustive(self, M=6):
for a in np.arange(-9., 9., 0.01):
for offset in np.arange(5.000001e-6, 2 * 5.000001e-6, 1e-6):
isEqual = closeAtMSigFigs(a, a + offset, M=M, tol=5)
assert not isEqual
for offset in np.arange(-5e-6, 5e-6 + EPS, 1e-6):
isEqual = closeAtMSigFigs(a, a + offset, M=M, tol=5)
if not isEqual:
print a
print a + offset
assert isEqual
def test_close_at_3_sig_figs_exhaustive(self, M=3):
for a in np.arange(-9., 9., 0.01):
for offset in np.arange(0.005000001, 0.01, 0.001):
isEqual = closeAtMSigFigs(a, a + offset, M=M, tol=5)
assert not isEqual
for offset in np.arange(-0.005000001, -0.01, -0.001):
isEqual = closeAtMSigFigs(a, a + offset, M=M, tol=5)
assert not isEqual
for offset in np.arange(-0.005, 0.005 + EPS, 0.001):
isEqual = closeAtMSigFigs(a, a + offset, M=M, tol=5)
if not isEqual:
print a
print a + offset
assert isEqual
def test_close_at_3_sig_figs_exhaustive(self, M=3):
for a in np.arange(-9.0, 9.0, 0.01):
for offset in np.arange(0.005000001, 0.01, 0.001):
isEqual = closeAtMSigFigs(a, a + offset, M=M, tol=5)
assert not isEqual
for offset in np.arange(-0.005000001, -0.01, -0.001):
isEqual = closeAtMSigFigs(a, a + offset, M=M, tol=5)
assert not isEqual
for offset in np.arange(-0.005, 0.005 + EPS, 0.001):
isEqual = closeAtMSigFigs(a, a + offset, M=M, tol=5)
if not isEqual:
print a
print a + offset
assert isEqual
def verify_evidence(self, evBound=0.00001, prevBound=0, lapFrac=None):
''' Compare current and previous evidence (ELBO) values,
verify that (within numerical tolerance) increases monotonically
'''
if np.isnan(evBound):
raise ValueError("Evidence should never be NaN")
if np.isinf(prevBound):
return False
isIncreasing = prevBound <= evBound
M = self.algParams['convergeSigFig']
isWithinTHR = closeAtMSigFigs(prevBound, evBound, M=M)
mLPkey = 'doMemoizeLocalParams'
if not isIncreasing and not isWithinTHR:
serious = True
if self.hasMove('birth') \
and (len(self.BirthCompIDs) > 0 or len(self.ModifiedCompIDs) > 0):
warnMsg = 'ev decreased during a birth'
warnMsg += ' (so monotonic increase not guaranteed)\n'
serious = False
elif mLPkey in self.algParams and not self.algParams[mLPkey]:
warnMsg = 'ev decreased when doMemoizeLocalParams=0'
warnMsg += ' (so monotonic increase not guaranteed)\n'
serious = False
else:
warnMsg = 'evidence decreased!\n'
warnMsg += ' prev = % .15e\n' % (prevBound)
warnMsg += ' cur = % .15e\n' % (evBound)
if lapFrac is None:
prefix = "WARNING: "
else:
prefix = "WARNING @ %.3f: " % (lapFrac)
if serious or not self.algParams['doShowSeriousWarningsOnly']:
Log.error(prefix + warnMsg)
return isWithinTHR
def verify_evidence(self, evBound=0.00001, prevBound=0, lapFrac=None):
''' Compare current and previous evidence (ELBO) values,
verify that (within numerical tolerance) increases monotonically
'''
if np.isnan(evBound):
raise ValueError("Evidence should never be NaN")
if np.isinf(prevBound):
return False
isIncreasing = prevBound <= evBound
M = self.algParams['convergeSigFig']
isWithinTHR = closeAtMSigFigs(prevBound, evBound, M=M)
mLPkey = 'doMemoizeLocalParams'
if not isIncreasing and not isWithinTHR:
serious = True
if self.hasMove('birth') \
and (len(self.BirthCompIDs) > 0 or len(self.ModifiedCompIDs) > 0):
warnMsg = 'ev decreased during a birth'
warnMsg += ' (so monotonic increase not guaranteed)\n'
serious = False
elif mLPkey in self.algParams and not self.algParams[mLPkey]:
warnMsg = 'ev decreased when doMemoizeLocalParams=0'
warnMsg += ' (so monotonic increase not guaranteed)\n'
serious = False
else:
warnMsg = 'evidence decreased!\n'
warnMsg += ' prev = % .15e\n' % (prevBound)
warnMsg += ' cur = % .15e\n' % (evBound)
if lapFrac is None:
prefix = "WARNING: "
else:
prefix = "WARNING @ %.3f: " % (lapFrac)
if serious or not self.algParams['doShowSeriousWarningsOnly']:
Log.error(prefix + warnMsg)
return isWithinTHR
def test_sovb_with_one_batch_equivalent_to_vb(self):
if 'soVB' not in self.learnAlgs:
raise SkipTest
kwargs = self.get_kwargs()
kwargs['nBatch'] = 1
kwargs['rhoexp'] = 0
__, __, vbInfo = run(self.Data, self.allocModelName,
self.obsModelName, 'VB', **kwargs)
__, __, sovbInfo = run(self.Data, self.allocModelName,
self.obsModelName, 'soVB', **kwargs)
vbEv = vbInfo['evTrace'][:-1]
sovbEv = sovbInfo['evTrace']
for ii in range(len(vbEv)):
if hasattr(self, 'mustRetainLPAcrossLapsForGuarantees'):
print vbEv[ii], sovbEv[ii]
assert closeAtMSigFigs(vbEv[ii], sovbEv[ii], M=2)
else:
assert closeAtMSigFigs(vbEv[ii], sovbEv[ii], M=8)
def test_sovb_with_one_batch_equivalent_to_vb(self):
if 'soVB' not in self.learnAlgs:
raise SkipTest
kwargs = self.get_kwargs()
kwargs['nBatch'] = 1
kwargs['rhoexp'] = 0
__, __, vbInfo = run(self.Data, self.allocModelName, self.obsModelName,
'VB', **kwargs)
__, __, sovbInfo = run(self.Data, self.allocModelName,
self.obsModelName, 'soVB', **kwargs)
vbEv = vbInfo['evTrace'][:-1]
sovbEv = sovbInfo['evTrace']
for ii in range(len(vbEv)):
if hasattr(self, 'mustRetainLPAcrossLapsForGuarantees'):
print vbEv[ii], sovbEv[ii]
assert closeAtMSigFigs(vbEv[ii], sovbEv[ii], M=2)
else:
assert closeAtMSigFigs(vbEv[ii], sovbEv[ii], M=8)
def test_vb_sovb_and_movb_all_estimate_evBound_in_same_ballpark(self):
if 'moVB' not in self.learnAlgs:
raise SkipTest
kwargs = self.get_kwargs()
kwargs['nBatch'] = 5
__, __, vbInfo = run(self.Data, self.allocModelName, self.obsModelName,
'VB', **kwargs)
__, __, movbInfo = run(self.Data, self.allocModelName,
self.obsModelName, 'moVB', **kwargs)
__, __, sovbInfo = run(self.Data, self.allocModelName,
self.obsModelName, 'soVB', **kwargs)
vbEv = vbInfo['evBound']
movbEv = movbInfo['evBound']
sovbEv = np.mean(sovbInfo['evTrace'][-10:])
print vbEv
print movbEv
print sovbEv
assert closeAtMSigFigs(vbEv, movbEv, M=2)
assert closeAtMSigFigs(vbEv, sovbEv, M=2)
def test_vb_sovb_and_movb_all_estimate_evBound_in_same_ballpark(self):
if 'moVB' not in self.learnAlgs:
raise SkipTest
kwargs = self.get_kwargs()
kwargs['nBatch'] = 5
__, __, vbInfo = run(self.Data, self.allocModelName,
self.obsModelName, 'VB', **kwargs)
__, __, movbInfo = run(self.Data, self.allocModelName,
self.obsModelName, 'moVB', **kwargs)
__, __, sovbInfo = run(self.Data, self.allocModelName,
self.obsModelName, 'soVB', **kwargs)
vbEv = vbInfo['evBound']
movbEv = movbInfo['evBound']
sovbEv = np.mean(sovbInfo['evTrace'][-10:])
print vbEv
print movbEv
print sovbEv
assert closeAtMSigFigs(vbEv, movbEv, M=2)
assert closeAtMSigFigs(vbEv, sovbEv, M=2)
def test_movb_with_one_batch_equivalent_to_vb(self):
if 'moVB' not in self.learnAlgs:
raise SkipTest
kwargs = self.get_kwargs()
kwargs['nBatch'] = 1
__, __, vbInfo = run(self.Data, self.allocModelName, self.obsModelName,
'VB', **kwargs)
__, __, movbInfo = run(self.Data, self.allocModelName,
self.obsModelName, 'moVB', **kwargs)
vbEv = vbInfo['evTrace'][:-1]
movbEv = movbInfo['evTrace']
print vbEv
print movbEv
assert len(vbEv) == len(movbEv)
for ii in range(len(vbEv)):
assert closeAtMSigFigs(vbEv[ii], movbEv[ii], M=8)
def test_movb_with_one_batch_equivalent_to_vb(self):
if 'moVB' not in self.learnAlgs:
raise SkipTest
kwargs = self.get_kwargs()
kwargs['nBatch'] = 1
__, __, vbInfo = run(self.Data, self.allocModelName,
self.obsModelName, 'VB', **kwargs)
__, __, movbInfo = run(self.Data, self.allocModelName,
self.obsModelName, 'moVB', **kwargs)
vbEv = vbInfo['evTrace'][:-1]
movbEv = movbInfo['evTrace']
print vbEv
print movbEv
assert len(vbEv) == len(movbEv)
for ii in range(len(vbEv)):
assert closeAtMSigFigs(vbEv[ii], movbEv[ii], M=8)
def test_vb_repeatable_when_continued(self):
if 'VB' not in self.learnAlgs:
raise SkipTest
kwargs = self.get_kwargs_do_save_disk()
kwargs['nLap'] = 10
hmodel1, LP1, Info1 = run(self.Data, self.allocModelName,
self.obsModelName, 'VB', **kwargs)
kwargs['nLap'] = 5
kwargs['startLap'] = 5
hmodel2, LP2, Info2 = continueRun(self.Data, self.allocModelName,
self.obsModelName, 'VB', **kwargs)
if hasattr(self, 'mustRetainLPAcrossLapsForGuarantees'):
print Info1['evTrace'][-1]
print Info2['evTrace'][-1]
assert closeAtMSigFigs(Info1['evTrace'][-1], Info2['evTrace'][-1], M=2)
else:
assert Info1['evTrace'][-1] == Info2['evTrace'][-1]
def test_vb_repeatable_when_continued(self):
if 'VB' not in self.learnAlgs:
raise SkipTest
kwargs = self.get_kwargs_do_save_disk()
kwargs['nLap'] = 10
hmodel1, LP1, Info1 = run(self.Data, self.allocModelName,
self.obsModelName, 'VB', **kwargs)
kwargs['nLap'] = 5
kwargs['startLap'] = 5
hmodel2, LP2, Info2 = continueRun(self.Data, self.allocModelName,
self.obsModelName, 'VB', **kwargs)
if hasattr(self, 'mustRetainLPAcrossLapsForGuarantees'):
print Info1['evTrace'][-1]
print Info2['evTrace'][-1]
assert closeAtMSigFigs(Info1['evTrace'][-1],
Info2['evTrace'][-1],
M=2)
else:
assert Info1['evTrace'][-1] == Info2['evTrace'][-1]
def test_close_at_2_sig_figs_boundary(self, M=2):
assert closeAtMSigFigs(9.96, 9.98, M=M, tol=5)
# once one arg goes about 10, the new threshold is 0.5, not 0.05
assert closeAtMSigFigs(9.96, 10.01, M=M, tol=5)
assert closeAtMSigFigs(9.96, 10.46, M=M, tol=5)
assert not closeAtMSigFigs(9.96, 10.4601, M=M, tol=5)
def test_close_at_2_sig_figs_readable(self, M=2):
assert closeAtMSigFigs(1.42, 1.42, M=M, tol=5)
assert closeAtMSigFigs(1.42, 1.47, M=M, tol=5)
assert not closeAtMSigFigs(1.42, 1.4701, M=M, tol=5)
assert closeAtMSigFigs(1.42, 1.37, M=M, tol=5)
assert not closeAtMSigFigs(1.42, 1.36999, M=M, tol=5)
assert closeAtMSigFigs(1.42e5, 1.47e5, M=M, tol=5)
assert not closeAtMSigFigs(1.42e5, 1.4701e5, M=M, tol=5)
assert closeAtMSigFigs(1.42e5, 1.37e5, M=M, tol=5)
assert not closeAtMSigFigs(1.42e5, 1.3699e5, M=M, tol=5)
assert closeAtMSigFigs(-1.42e2, -1.47e2, M=M, tol=5)
assert not closeAtMSigFigs(-1.42e2, -1.4701e2, M=M, tol=5)
assert closeAtMSigFigs(-1.42e2, -1.37e2, M=M, tol=5)
assert not closeAtMSigFigs(-1.42e2, -1.369999e2, M=M, tol=5)
def test_close_at_2_sig_figs_readable(self, M=2):
assert closeAtMSigFigs(1.42, 1.42, M=M, tol=5)
assert closeAtMSigFigs(1.42, 1.47, M=M, tol=5)
assert not closeAtMSigFigs(1.42, 1.4701, M=M, tol=5)
assert closeAtMSigFigs(1.42, 1.37, M=M, tol=5)
assert not closeAtMSigFigs(1.42, 1.36999, M=M, tol=5)
assert closeAtMSigFigs(1.42e5, 1.47e5, M=M, tol=5)
assert not closeAtMSigFigs(1.42e5, 1.4701e5, M=M, tol=5)
assert closeAtMSigFigs(1.42e5, 1.37e5, M=M, tol=5)
assert not closeAtMSigFigs(1.42e5, 1.3699e5, M=M, tol=5)
assert closeAtMSigFigs(-1.42e2, -1.47e2, M=M, tol=5)
assert not closeAtMSigFigs(-1.42e2, -1.4701e2, M=M, tol=5)
assert closeAtMSigFigs(-1.42e2, -1.37e2, M=M, tol=5)
assert not closeAtMSigFigs(-1.42e2, -1.369999e2, M=M, tol=5)
def test_close_at_2_sig_figs_boundary(self, M=2):
assert closeAtMSigFigs(9.96, 9.98, M=M, tol=5)
# once one arg goes about 10, the new threshold is 0.5, not 0.05
assert closeAtMSigFigs(9.96, 10.01, M=M, tol=5)
assert closeAtMSigFigs(9.96, 10.46, M=M, tol=5)
assert not closeAtMSigFigs(9.96, 10.4601, M=M, tol=5)
