def pprint(msg, level=None, prefix='', linewidth=80):
global Log
global DEFAULTLEVEL
if isinstance(msg, list):
msgs = list()
prefixes = list()
for ii, m_ii in enumerate(msg):
prefix_ii = prefix[ii]
msgs_ii = split_str_into_fixed_width_lines(m_ii,
linewidth=linewidth -
len(prefix_ii))
msgs.extend(msgs_ii)
prefixes.extend([prefix[ii] for i in range(len(msgs_ii))])
for ii in range(len(msgs)):
pprint(prefixes[ii] + msgs[ii], level=level)
return
if DEFAULTLEVEL == 'print':
print(msg)
if Log is None:
return
if level is None:
level = DEFAULTLEVEL
if isinstance(level, str):
if level.count('info'):
level = logging.INFO
elif level.count('debug'):
level = logging.DEBUG
Log.log(level, msg)
for h in Log.handlers:
h.flush()
# Small bit of code to track recent messages
# for debugging birth proposals. Used when dumping to HTML
global RecentMessages
if isinstance(RecentMessages, list):
RecentMessages.append(msg)
def countvec2str(curN_K):
if W is None:
str_sum_w = ' '.join(['%7.0f' % (x) for x in curN_K])
else:
assert np.allclose(curN_K.sum(), W.sum())
str_sum_w = ' '.join(['%7.2f' % (x) for x in curN_K])
return split_str_into_fixed_width_lines(str_sum_w, tostr=True)
def pprint(msg, level=None, prefix='', linewidth=80):
global Log
global DEFAULTLEVEL
global RecentMessages
if isinstance(msg, list):
msgs = list()
prefixes = list()
for ii, m_ii in enumerate(msg):
prefix_ii = prefix[ii]
msgs_ii = split_str_into_fixed_width_lines(m_ii,
linewidth=linewidth -
len(prefix_ii))
msgs.extend(msgs_ii)
prefixes.extend([prefix[ii] for i in range(len(msgs_ii))])
for ii in range(len(msgs)):
pprint(prefixes[ii] + msgs[ii], level=level)
return
if level == 'print':
print(msg)
if Log is None:
return
if level is None:
level = DEFAULTLEVEL
if isinstance(level, str):
if level.count('info'):
level = logging.INFO
elif level.count('debug'):
level = logging.DEBUG
Log.log(level, msg)
if isinstance(RecentMessages, list):
RecentMessages.append(msg)
def runKMeans_BregmanDiv(X,
K,
obsModel,
W=None,
Niter=100,
seed=0,
init='plusplus',
smoothFracInit=1.0,
smoothFrac=0,
logFunc=None,
eps=1e-10,
setOneToPriorMean=0,
distexp=1.0,
assert_monotonic=True,
**kwargs):
''' Run hard clustering algorithm to find K clusters.
Returns
-------
Z : 1D array, size N
Mu : 2D array, size K x D
Lscores : 1D array, size Niter
'''
chosenZ, Mu, _, _ = initKMeans_BregmanDiv(
X,
K,
obsModel,
W=W,
seed=seed,
smoothFrac=smoothFracInit,
distexp=distexp,
setOneToPriorMean=setOneToPriorMean)
# Make sure we update K to reflect the returned value.
# initKMeans_BregmanDiv will return fewer than K clusters
# in some edge cases, like when data matrix X has duplicate rows
# and specified K is larger than the number of unique rows.
K = len(Mu)
assert K > 0
assert Niter >= 0
if Niter == 0:
Z = -1 * np.ones(X.shape[0])
if chosenZ[0] == -1:
Z[chosenZ[1:]] = np.arange(chosenZ.size - 1)
else:
Z[chosenZ] = np.arange(chosenZ.size)
Lscores = list()
prevN = np.zeros(K)
for riter in range(Niter):
Div = obsModel.calcSmoothedBregDiv(X=X,
Mu=Mu,
W=W,
includeOnlyFastTerms=True,
smoothFrac=smoothFrac,
eps=eps)
Z = np.argmin(Div, axis=1)
Ldata = Div.min(axis=1).sum()
Lprior = obsModel.calcBregDivFromPrior(Mu=Mu,
smoothFrac=smoothFrac).sum()
Lscore = Ldata + Lprior
Lscores.append(Lscore)
# Verify objective is monotonically increasing
if assert_monotonic:
try:
# Test allows small positive increases that are
# numerically indistinguishable from zero. Don't care about these.
assert np.all(np.diff(Lscores) <= 1e-5)
except AssertionError:
msg = "iter %d: Lscore %.3e" % (riter, Lscore)
msg += '\nIn the kmeans update loop of FromScratchBregman.py'
msg += '\nLscores not monotonically decreasing...'
if logFunc:
logFunc(msg)
else:
print(msg)
assert np.all(np.diff(Lscores) <= 1e-5)
N = np.zeros(K)
for k in range(K):
if W is None:
W_k = None
N[k] = np.sum(Z == k)
else:
W_k = W[Z == k]
N[k] = np.sum(W_k)
if N[k] > 0:
Mu[k] = obsModel.calcSmoothedMu(X[Z == k], W_k)
else:
Mu[k] = obsModel.calcSmoothedMu(X=None)
if logFunc:
logFunc("iter %d: Lscore %.3e" % (riter, Lscore))
if W is None:
str_sum_w = ' '.join(['%7.0f' % (x) for x in N])
else:
assert np.allclose(N.sum(), W.sum())
str_sum_w = ' '.join(['%7.2f' % (x) for x in N])
str_sum_w = split_str_into_fixed_width_lines(str_sum_w, tostr=True)
logFunc(str_sum_w)
if np.max(np.abs(N - prevN)) == 0:
break
prevN[:] = N
uniqueZ = np.unique(Z)
if Niter > 0:
# In case a cluster was pushed to zero
if uniqueZ.size < len(Mu):
Mu = [Mu[k] for k in uniqueZ]
else:
# Without full pass through dataset, many items not assigned
# which we indicated with Z value of -1
# Should ignore this when counting states
uniqueZ = uniqueZ[uniqueZ >= 0]
assert len(Mu) == uniqueZ.size
return Z, Mu, np.asarray(Lscores)
