def calcSummaryStats(Data, SS, LP, **kwargs):
''' Calculate summary statistics for given dataset and local parameters
Returns
--------
SS : SuffStatBag object, with K components.
'''
X = Data.X
if 'resp' in LP:
resp = LP['resp']
K = resp.shape[1]
# 1/2: Compute mean statistic
S_x = dotATB(resp, X)
# 2/2: Compute expected outer-product statistic
S_xxT = np.zeros((K, Data.dim, Data.dim))
sqrtResp_k = np.sqrt(resp[:, 0])
sqrtRX_k = sqrtResp_k[:, np.newaxis] * Data.X
S_xxT[0] = dotATA(sqrtRX_k)
for k in range(1, K):
np.sqrt(resp[:, k], out=sqrtResp_k)
np.multiply(sqrtResp_k[:, np.newaxis], Data.X, out=sqrtRX_k)
S_xxT[k] = dotATA(sqrtRX_k)
else:
spR = LP['spR']
K = spR.shape[1]
# 1/2: Compute mean statistic
S_x = spR.T * X
# 2/2: Compute expected outer-product statistic
S_xxT = calcSpRXXT(X=X, spR_csr=spR)
if SS is None:
SS = SuffStatBag(K=K, D=Data.dim)
# Expected mean for each state k
SS.setField('x', S_x, dims=('K', 'D'))
# Expected outer-product for each state k
SS.setField('xxT', S_xxT, dims=('K', 'D', 'D'))
# Expected count for each k
# Usually computed by allocmodel. But just in case...
if not hasattr(SS, 'N'):
if 'resp' in LP:
SS.setField('N', LP['resp'].sum(axis=0), dims='K')
else:
SS.setField('N', as1D(toCArray(LP['spR'].sum(axis=0))), dims='K')
return SS
def get_global_suff_stats(self, Data, LP, doPrecompEntropy=None, **kwargs):
''' Create sufficient stats needed for global param updates
Args
-------
Data : bnpy data object
LP : Dictionary containing the local parameters. Expected to contain:
resp : Data.nObs x K array
respPair : Data.nObs x K x K array (from the def. of respPair, note
respPair[0,:,:] is undefined)
Returns
-------
SS : SuffStatBag with fields
StartStateCount : A vector of length K with entry i being
resp(z_{1k}) = resp[0,:]
TransStateCount : A K x K matrix where TransStateCount[i,j] =
sum_{n=2}^K respPair(z_{n-1,j}, z_{nk})
N : A vector of length K with entry k being
sum_{n=1}^Data.nobs resp(z_{nk})
The first two of these are used by FiniteHMM.update_global_params,
and the third is used by ObsModel.update_global_params.
(see the documentation for information about resp and respPair)
'''
resp = LP['resp']
respPair = LP['respPair']
K = resp.shape[1]
startLocIDs = Data.doc_range[:-1]
StartStateCount = np.sum(resp[startLocIDs], axis=0)
N = np.sum(resp, axis=0)
TransStateCount = np.sum(respPair, axis=0)
SS = SuffStatBag(K=K, D=Data.dim)
SS.setField('StartStateCount', StartStateCount, dims=('K'))
SS.setField('TransStateCount', TransStateCount, dims=('K', 'K'))
SS.setField('N', N, dims=('K'))
if doPrecompEntropy is not None:
entropy = self.elbo_entropy(Data, LP)
SS.setELBOTerm('Elogqz', entropy, dims=None)
return SS
def calcHardMergeGap(self, SS, kA, kB):
''' Calculate scalar improvement in ELBO for hard merge of comps kA, kB
Does *not* include any entropy.
Returns
---------
L : scalar
'''
m_K = SS.K - 1
m_SS = SuffStatBag(K=SS.K, D=0)
m_SS.setField('StartStateCount', SS.StartStateCount.copy(), dims='K')
m_SS.setField('TransStateCount', SS.TransStateCount.copy(),
dims=('K', 'K'))
m_SS.mergeComps(kA, kB)
# Create candidate beta vector
m_beta = StickBreakUtil.rho2beta(self.rho)
m_beta[kA] += m_beta[kB]
m_beta = np.delete(m_beta, kB, axis=0)
# Create candidate rho and omega vectors
m_rho = StickBreakUtil.beta2rho(m_beta, m_K)
m_omega = np.delete(self.omega, kB)
# Create candidate startTheta
m_startTheta = self.startAlpha * m_beta.copy()
m_startTheta[:m_K] += m_SS.StartStateCount
# Create candidate transTheta
m_transTheta = self.transAlpha * np.tile(m_beta, (m_K, 1))
if self.kappa > 0:
m_transTheta[:, :m_K] += self.kappa * np.eye(m_K)
m_transTheta[:, :m_K] += m_SS.TransStateCount
# Evaluate objective func. for both candidate and current model
Lcur = calcELBO_LinearTerms(
SS=SS, rho=self.rho, omega=self.omega,
startTheta=self.startTheta, transTheta=self.transTheta,
alpha=self.transAlpha, startAlpha=self.startAlpha,
gamma=self.gamma, kappa=self.kappa)
Lprop = calcELBO_LinearTerms(
SS=m_SS, rho=m_rho, omega=m_omega,
startTheta=m_startTheta, transTheta=m_transTheta,
alpha=self.transAlpha, startAlpha=self.startAlpha,
gamma=self.gamma, kappa=self.kappa)
# Note: This gap relies on fact that all nonlinear terms are entropies,
return Lprop - Lcur
def calcSummaryStatsForContigBlock(self, Data, SS=None, a=0, b=0):
''' Calculate sufficient stats for a single contiguous block of data
'''
if SS is None:
SS = SuffStatBag(K=1, D=Data.dim)
SS.setField('N', (b - a) * np.ones(1), dims='K')
SS.setField('x',
np.sum(Data.X[a:b], axis=0)[np.newaxis, :],
dims=('K', 'D'))
SS.setField('xxT',
dotATA(Data.X[a:b])[np.newaxis, :, :],
dims=('K', 'D', 'D'))
return SS
def calcSummaryStatsForContigBlock(self, Data, a=0, b=0, **kwargs):
''' Calculate summary stats for a contiguous block of the data.
Returns
--------
SS : SuffStatBag object, with 1 component.
'''
Xab = Data.X[a:b] # 2D array, Nab x D
CountON = np.sum(Xab, axis=0)[np.newaxis, :]
CountOFF = (b - a) - CountON
SS = SuffStatBag(K=1, D=Data.dim)
SS.setField('N', np.asarray([b - a], dtype=np.float64), dims='K')
SS.setField('Count1', CountON, dims=('K', 'D'))
SS.setField('Count0', CountOFF, dims=('K', 'D'))
return SS
def calcSummaryStats(Data, SS, LP, **kwargs):
''' Calculate summary statistics for given dataset and local parameters
Returns
--------
SS : SuffStatBag object, with K components.
'''
X = Data.X
if 'resp' in LP:
resp = LP['resp']
K = resp.shape[1]
# 1/2: Compute mean statistic
S_x = dotATB(resp, X)
# 2/2: Compute expected outer-product statistic
S_xx = calcRXX_withDenseResp(resp, X)
else:
spR = LP['spR']
K = spR.shape[1]
# 1/2: Compute mean statistic
S_x = spR.T * X
# 2/2: Compute expected outer-product statistic
S_xx = calcSpRXX(X=X, spR_csr=spR)
if SS is None:
SS = SuffStatBag(K=K, D=Data.dim)
# Expected mean for each state k
SS.setField('x', S_x, dims=('K', 'D'))
# Expected sum-of-squares for each state k
SS.setField('xx', S_xx, dims=('K', 'D'))
# Expected count for each k
# Usually computed by allocmodel. But sometimes not (eg TopicModel)
if not hasattr(SS, 'N'):
if 'resp' in LP:
SS.setField('N', LP['resp'].sum(axis=0), dims='K')
else:
SS.setField('N', as1D(toCArray(LP['spR'].sum(axis=0))), dims='K')
return SS
def get_global_suff_stats(self, Data, LP, doPrecompEntropy=0, **kwargs):
''' Compute sufficient stats for provided dataset and local params
Returns
-------
SS : SuffStatBag
Updated fields
* NodeStateCount : 2D array, nNodes x K
* N : 2D array, size K x K
'''
K = LP['resp'].shape[-1]
V = Data.nNodes
SS = SuffStatBag(K=K, D=Data.dim, V=V)
# NodeStateCount_src[i,k]
# Num edges assigned to topic k associated with node i as source
srcResp = LP['resp'].sum(axis=2)
NodeStateCount_src = Data.getSparseSrcNodeMat() * srcResp
# Equivalent but slower: for loop
# NodeStateCount_src = np.zeros((Data.nNodes, K))
# for i in xrange(Data.nNodes):
# mask_i = Data.edges[:,0] == i
# NodeStateCount_src[i,:] = srcResp[mask_i].sum(axis=0)
# NodeStateCount_rcv[i,k]
# Num edges assigned to topic k associated with node i as receiver
rcvResp = LP['resp'].sum(axis=1)
NodeStateCount_rcv = Data.getSparseRcvNodeMat() * rcvResp
# Summing src counts and rcv counts gives the total
SS.setField('NodeStateCount',
NodeStateCount_src + NodeStateCount_rcv,
dims=('V', 'K'))
# Compute total atoms assigned to each cluster pair
Nresp = np.sum(LP['resp'], axis=0)
SS.setField('N', Nresp, dims=('K', 'K'))
if doPrecompEntropy:
# Remember, resp has shape nEdges x K x K
# So, need to sum so we track scalar entropy, not K x K
Hresp = calcLentropyAsScalar(LP)
SS.setELBOTerm('Hresp', Hresp, dims=None)
return SS
def calcSummaryStats(Data, SS, LP, DataAtomType='doc', **kwargs):
''' Calculate summary statistics for given dataset and local parameters
Returns
--------
SS : SuffStatBag object, with K components.
'''
if 'resp' in LP:
K = LP['resp'].shape[1]
else:
K = LP['spR'].shape[1]
nnzPerRow = LP['nnzPerRow']
if SS is None:
SS = SuffStatBag(K=K, D=Data.vocab_size)
if DataAtomType == 'doc':
# X : 2D sparse matrix, size nDoc x vocab_size
X = Data.getSparseDocTypeCountMatrix()
# WordCounts : 2D array, size K x vocab_size
# obtained by sparse matrix multiply
# here, '*' operator does this because X is sparse matrix type
Nvec = None
if 'resp' in LP:
WordCounts = LP['resp'].T * X
if not hasattr(SS, 'N'):
Nvec = LP['resp'].sum(axis=0)
else:
WordCounts = (LP['spR'].T * X).toarray()
if not hasattr(SS, 'N'):
Nvec = as1D(toCArray(LP['spR'].sum(axis=0)))
if Nvec is not None:
SS.setField('N', Nvec, dims=('K'))
else:
# 2D sparse matrix, size V x N
X = Data.getSparseTokenTypeCountMatrix()
if 'resp' in LP:
WordCounts = (X * LP['resp']).T # matrix-matrix product
else:
WordCounts = (X * LP['spR']).T.toarray()
SS.setField('WordCounts', WordCounts, dims=('K', 'D'))
SS.setField('SumWordCounts', np.sum(WordCounts, axis=1), dims=('K'))
return SS
"""
def calcSummaryStatsForContigBlock(self, Data, SS=None, a=0, b=0):
''' Calculate sufficient stats for a single contiguous block of data
'''
D = Data.X.shape[1]
E = Data.Xprev.shape[1]
if SS is None:
SS = SuffStatBag(K=1, D=D, E=E)
elif not hasattr(SS, 'E'):
SS._Fields.E = E
ppT = dotATA(Data.Xprev[a:b])[np.newaxis, :, :]
xxT = dotATA(Data.X[a:b])[np.newaxis, :, :]
pxT = dotATB(Data.Xprev[a:b], Data.X[a:b])[np.newaxis, :, :]
SS.setField('N', (b - a) * np.ones(1), dims='K')
SS.setField('xxT', xxT, dims=('K', 'D', 'D'))
SS.setField('ppT', ppT, dims=('K', 'E', 'E'))
SS.setField('pxT', pxT, dims=('K', 'E', 'D'))
return SS
def init_global_params(self, Data, K=0, **initArgs):
''' Initialize rho, omega, and theta to reasonable values.
This is only called by "from scratch" init routines.
'''
self.K = K
self.rho = OptimizerRhoOmega.create_initrho(K)
self.omega = (1.0 + self.gamma) * np.ones(K)
# To initialize theta, perform standard update given rho, omega
# but with "empty" sufficient statistics.
SS = SuffStatBag(K=self.K, D=Data.dim)
SS.setField('StartStateCount', np.ones(K), dims=('K'))
SS.setField('TransStateCount', np.ones((K, K)), dims=('K', 'K'))
self.transTheta, self.startTheta = self._calcTheta(SS)
def calcSummaryStats(Data, SS, LP, **kwargs):
''' Calculate sufficient statistics for local params at data slice.
Returns
-------
SS
'''
X = Data.X
Xprev = Data.Xprev
resp = LP['resp']
K = resp.shape[1]
D = Data.X.shape[1]
E = Data.Xprev.shape[1]
if SS is None:
SS = SuffStatBag(K=K, D=D, E=E)
elif not hasattr(SS, 'E'):
SS._Fields.E = E
# Expected count for each k
# Usually computed by allocmodel. But just in case...
if not hasattr(SS, 'N'):
SS.setField('N', np.sum(resp, axis=0), dims='K')
# Expected outer products
sqrtResp = np.sqrt(resp)
xxT = np.empty((K, D, D))
ppT = np.empty((K, E, E))
pxT = np.empty((K, E, D))
for k in xrange(K):
sqrtResp_k = sqrtResp[:, k][:, np.newaxis]
xxT[k] = dotATA(sqrtResp_k * Data.X)
ppT[k] = dotATA(sqrtResp_k * Data.Xprev)
pxT[k] = np.dot(Data.Xprev.T, resp[:, k][:, np.newaxis] * Data.X)
SS.setField('xxT', xxT, dims=('K', 'D', 'D'))
SS.setField('ppT', ppT, dims=('K', 'E', 'E'))
SS.setField('pxT', pxT, dims=('K', 'E', 'D'))
return SS
def get_global_suff_stats(self, Data, LP, doPrecompEntropy=None, **kwargs):
''' Calculate the sufficient statistics for global parameter updates
Only adds stats relevant for this allocModel.
Other stats are added by the obsModel.
Args
-------
Data : bnpy data object
LP : local param dict with fields
resp : Data.nObs x K array,
where resp[n,k] = posterior resp of comp k
doPrecompEntropy : boolean flag
indicates whether to precompute ELBO terms in advance
used for memoized learning algorithms (moVB)
Returns
-------
SS : SuffStats for K components, with field
N : vector of dimension K,
effective number of observations assigned to each comp
Npair : matrix of dimensions K x K, where Npair[l,m] =
effective # of obs x_{ij} with z_{il} and z_{jm}
'''
Npair = np.sum(LP['resp'], axis=0)
self.Npair = Npair
N = np.sum(LP['respSingle'], axis=0)
SS = SuffStatBag(K=N.shape[0], D=Data.dim)
SS.setField('Npair', Npair, dims=('K', 'K'))
SS.setField('N', N, dims=('K',))
if doPrecompEntropy is not None:
ElogqZ_vec = self.E_logqZ(LP)
SS.setELBOTerm('ElogqZ', ElogqZ_vec, dims=('K',))
return SS
def calcSummaryStatsForContigBlock(self,
Data,
SS=None,
a=None,
b=None,
**kwargs):
''' Calculate summary statistics for specific block of dataset
Returns
--------
SS : SuffStatBag object, with K components.
'''
SS = SuffStatBag(K=1, D=Data.dim)
# Expected count
SS.setField('N', (b - a) * np.ones(1, dtype=np.float64), dims='K')
# Expected outer-product
xxT = dotATA(Data.X[a:b])[np.newaxis, :, :]
SS.setField('xxT', xxT, dims=('K', 'D', 'D'))
return SS
def get_global_suff_stats(self, Data, LP, doPrecompEntropy=0, **kwargs):
''' Compute sufficient stats for provided dataset and local params
Returns
-------
SS : SuffStatBag with K components and fields
* sumSource : nNodes x K
* sumReceiver : nNodes x K
'''
V = Data.nNodes
K = LP['resp'].shape[-1]
SS = SuffStatBag(K=K, D=Data.dim, V=V)
if 'NodeStateCount' not in LP:
assert 'resp' in LP
LP = self.initLPFromResp(Data, LP)
SS.setField('NodeStateCount', LP['NodeStateCount'], dims=('V', 'K'))
if np.allclose(LP['resp'].sum(axis=1).min(), 1.0):
# If the LP fully represents all present edges,
# then the NodeStateCount should as well.
assert np.allclose(SS.NodeStateCount, Data.nEdges * 2)
SS.setField('N', LP['N_fg'], dims=('K', ))
SS.setField('scaleFactor', Data.nEdges, dims=None)
if 'Ldata_bg' in LP:
SS.setELBOTerm('Ldata_bg', LP['Ldata_bg'], dims=None)
if doPrecompEntropy:
Hresp_fg = LP['Lentropy_fg'] # = -1 * calcRlogR(LP['resp'])
Hresp_bg = LP['Lentropy_bg']
SS.setELBOTerm('Hresp', Hresp_fg, dims='K')
SS.setELBOTerm('Hresp_bg', Hresp_bg, dims=None)
return SS
def init_global_params(obsModel,
Data,
K=0,
seed=0,
initname='randexamples',
initBlockLen=20,
**kwargs):
''' Initialize parameters for Bernoulli obsModel, in place.
Parameters
-------
obsModel : AbstractObsModel subclass
Observation model object to initialize.
Data : bnpy.data.DataObj
Dataset to use to drive initialization.
obsModel dimensions must match this dataset.
initname : str
name of routine used to do initialization
Options: 'randexamples', 'randexamplesbydist', 'kmeans'
Returns
-------
initLP : dict
Local parameters used for initialization
Post Condition
--------------
obsModel has valid global parameters.
Either its EstParams or Post attribute will be contain K components.
'''
PRNG = np.random.RandomState(seed)
AdjMat = Data.toAdjacencyMatrix()
if AdjMat.ndim == 3:
AdjMat = AdjMat[:, :, 0]
K = np.minimum(K, Data.nNodes)
if len(obsModel.CompDims) == 2:
CompDims = (
K,
K,
)
else:
CompDims = (K, )
if initname == 'randexamples':
# Pick K nodes at random in provided graph,
# and set all edges belonging to that node to one cluster.
nNodes = AdjMat.shape[0]
chosenNodes = PRNG.choice(nNodes, size=K, replace=False)
elif initname == 'randexamplesbydist':
# Choose K items from the Data,
# selecting the first at random,
# then subsequently proportional to euclidean distance to the closest
# item
objID = PRNG.choice(AdjMat.shape[0])
chosenNodes = list([objID])
minDistVec = np.inf * np.ones(AdjMat.shape[0])
for k in range(1, K):
curDistVec = np.sum((AdjMat - AdjMat[objID])**2, axis=1)
minDistVec = np.minimum(minDistVec, curDistVec)
sum_minDistVec = np.sum(minDistVec)
if sum_minDistVec > 0:
p = minDistVec / sum_minDistVec
else:
D = minDistVec.size
p = 1.0 / D * np.ones(D)
objID = PRNG.choice(Data.nNodes, p=p)
chosenNodes.append(objID)
elif initname == 'kmeans':
# Fill in resp matrix with hard-clustering from K-means
# using an initialization with K randomly selected points from X
np.random.seed(seed)
centroids, chosenNodes = kmeans2(data=AdjMat, k=K, minit='points')
else:
raise NotImplementedError('Unrecognized initname ' + initname)
# Convert chosen nodes into temporary local param dict
# tempLP will have proper resp (nEdges x K), etc.
tempLP = chosenNodes_to_LP(chosenNodes, CompDims, Data, PRNG=PRNG, K=K)
# Perform summary step and global step to get obsModel global params.
SS = SuffStatBag(K=K, D=Data.dim)
SS = obsModel.get_global_suff_stats(Data, SS, tempLP)
obsModel.update_global_params(SS)
return tempLP
def init_global_params(obsModel,
Data,
K=0,
seed=0,
initname='randexamples',
initBlockLen=20,
**kwargs):
''' Initialize parameters for Gaussian obsModel, in place.
Parameters
-------
obsModel : bnpy.obsModel subclass
Observation model object to initialize.
Data : bnpy.data.DataObj
Dataset to use to drive initialization.
obsModel dimensions must match this dataset.
initname : str
name of routine used to do initialization
Options: ['randexamples', 'randexamplesbydist', 'kmeans',
'randcontigblocks', 'randsoftpartition',
]
Post Condition
-------
obsModel has valid global parameters.
Either its EstParams or Post attribute will be contain K components.
'''
K = int(K)
PRNG = np.random.RandomState(seed)
X = Data.X
if initname == 'randexamples':
# Choose K items uniformly at random from the Data
# then component params by M-step given those single items
resp = np.zeros((Data.nObs, K))
permIDs = PRNG.permutation(Data.nObs).tolist()
for k in range(K):
resp[permIDs[k], k] = 1.0
elif initname == 'randexamplesbydist':
# Choose K items from the Data,
# selecting the first at random,
# then subsequently proportional to euclidean distance to the closest
# item
objID = PRNG.choice(Data.nObs)
chosenObjIDs = list([objID])
minDistVec = np.inf * np.ones(Data.nObs)
for k in range(1, K):
curDistVec = np.sum((Data.X - Data.X[objID])**2, axis=1)
minDistVec = np.minimum(minDistVec, curDistVec)
objID = PRNG.choice(Data.nObs, p=minDistVec / minDistVec.sum())
chosenObjIDs.append(objID)
resp = np.zeros((Data.nObs, K))
for k in range(K):
resp[chosenObjIDs[k], k] = 1.0
elif initname == 'randcontigblocks':
# Choose K contig blocks of provided size from the Data,
# selecting each block at random from a particular sequence
if hasattr(Data, 'doc_range'):
doc_range = Data.doc_range.copy()
else:
doc_range = np.asarray([0, Data.X.shape[0]])
nDoc = doc_range.size - 1
docIDs = np.arange(nDoc)
PRNG.shuffle(docIDs)
resp = np.zeros((Data.nObs, K))
for k in range(K):
n = docIDs[k % nDoc]
start = doc_range[n]
stop = doc_range[n + 1]
T = stop - start
if initBlockLen >= T:
a = start
b = stop
else:
a = start + PRNG.choice(T - initBlockLen)
b = a + initBlockLen
resp[a:b, k] = 1.0
elif initname == 'randsoftpartition':
# Randomly assign all data items some mass in each of K components
# then create component params by M-step given that soft partition
resp = PRNG.gamma(1.0 / (K * K), 1, size=(Data.nObs, K))
resp[resp < 1e-3] = 0
rsum = np.sum(resp, axis=1)
badIDs = rsum < 1e-8
# if any rows have no content, just set them to unif resp.
if np.any(badIDs):
resp[badIDs] = 1.0 / K
rsum[badIDs] = 1
resp = resp / rsum[:, np.newaxis]
assert np.allclose(np.sum(resp, axis=1), 1.0)
elif initname == 'kmeans':
# Fill in resp matrix with hard-clustering from K-means
# using an initialization with K randomly selected points from X
np.random.seed(seed)
centroids, labels = kmeans2(data=Data.X, k=K, minit='points')
resp = np.zeros((Data.nObs, K))
for t in range(Data.nObs):
resp[t, labels[t]] = 1
else:
raise NotImplementedError('Unrecognized initname ' + initname)
tempLP = dict(resp=resp)
SS = SuffStatBag(K=K, D=Data.dim)
SS = obsModel.get_global_suff_stats(Data, SS, tempLP)
obsModel.update_global_params(SS)
def calcSummaryStats(Dslice, SS, LP, DataAtomType='doc', **kwargs):
''' Calculate summary statistics for given dataset and local parameters
Returns
--------
SS : SuffStatBag object, with K components.
'''
if 'resp' in LP:
N = LP['resp'].shape[0]
K = LP['resp'].shape[1]
if LP['resp'].ndim == 2:
CompDims = ('K', ) # typical case
else:
assert LP['resp'].ndim == 3
CompDims = ('K', 'K') # relational data
else:
assert 'spR' in LP
N, K = LP['spR'].shape
CompDims = ('K', )
if SS is None:
SS = SuffStatBag(K=K, D=Dslice.dim)
if not hasattr(SS, 'N'):
if 'resp' in LP:
SS.setField('N', np.sum(LP['resp'], axis=0), dims=CompDims)
else:
SS.setField('N', LP['spR'].sum(axis=0), dims=CompDims)
if hasattr(Dslice, 'X'):
X = Dslice.X
if 'resp' in LP:
# Matrix-matrix product, result is K x D (or KxKxD if relational)
CountON = np.tensordot(LP['resp'].T, X, axes=1)
CountOFF = np.tensordot(LP['resp'].T, 1 - X, axes=1)
else:
CountON = LP['spR'].T * X
CountOFF = LP['spR'].T * (1 - X)
elif DataAtomType == 'doc' or Dslice.nDoc == N:
X = Dslice.getSparseDocTypeBinaryMatrix()
if 'resp' in LP:
# Sparse matrix product
CountON = LP['resp'].T * X
else:
CountON = (LP['spR'].T * X).toarray()
CountOFF = SS.N[:, np.newaxis] - CountON
else:
CountON = np.zeros((SS.K, Dslice.vocab_size))
CountOFF = np.zeros((SS.K, Dslice.vocab_size))
for d in xrange(Dslice.nDoc):
words_d = Dslice.word_id[Dslice.doc_range[d]:Dslice.doc_range[d +
1]]
rstart_d = d * Dslice.vocab_size
rstop_d = (d + 1) * Dslice.vocab_size
if 'resp' in LP:
Count_d = LP['resp'][rstart_d:rstop_d, :].T
else:
raise NotImplementedError("TODO")
CountOFF += Count_d
CountON[:, words_d] += Count_d[:, words_d]
CountOFF[:, words_d] -= Count_d[:, words_d]
SS.setField('Count1', CountON, dims=CompDims + ('D', ))
SS.setField('Count0', CountOFF, dims=CompDims + ('D', ))
return SS
def calcSummaryStats(Data, LP,
doPrecompEntropy=0,
doPrecompMergeEntropy=0,
mPairIDs=None,
trackDocUsage=0,
**kwargs):
''' Calculate summary statistics for given data slice and local params.
Returns
-------
SS : SuffStatBag
'''
if mPairIDs is None:
M = 0
else:
M = len(mPairIDs)
resp = LP['resp']
K = resp.shape[1]
startLocIDs = Data.doc_range[:-1]
StartStateCount = np.sum(resp[startLocIDs], axis=0)
N = np.sum(resp, axis=0)
if 'TransCount' in LP:
TransStateCount = np.sum(LP['TransCount'], axis=0)
else:
respPair = LP['respPair']
TransStateCount = np.sum(respPair, axis=0)
SS = SuffStatBag(K=K, D=Data.dim, M=M)
SS.setField('StartStateCount', StartStateCount, dims=('K'))
SS.setField('TransStateCount', TransStateCount, dims=('K', 'K'))
SS.setField('N', N, dims=('K'))
SS.setField('nDoc', Data.nDoc, dims=None)
if doPrecompEntropy or 'Htable' in LP:
# Compute entropy terms!
# 'Htable', 'Hstart' will both be in Mdict
Mdict = calcELBO_NonlinearTerms(Data=Data,
LP=LP, returnMemoizedDict=1)
SS.setELBOTerm('Htable', Mdict['Htable'], dims=('K', 'K'))
SS.setELBOTerm('Hstart', Mdict['Hstart'], dims=('K'))
if doPrecompMergeEntropy:
subHstart, subHtable = HMMUtil.PrecompMergeEntropy_SpecificPairs(
LP, Data, mPairIDs)
SS.setMergeTerm('Hstart', subHstart, dims=('M'))
SS.setMergeTerm('Htable', subHtable, dims=('M', 2, 'K'))
SS.mPairIDs = np.asarray(mPairIDs)
if trackDocUsage:
# Track how often topic appears in a seq. with mass > thresh.
DocUsage = np.zeros(K)
for n in xrange(Data.nDoc):
start = Data.doc_range[n]
stop = Data.doc_range[n + 1]
DocUsage += np.sum(LP['resp'][start:stop], axis=0) > 0.01
SS.setSelectionTerm('DocUsageCount', DocUsage, dims='K')
return SS
def calcSummaryStats(Dslice,
LP=None,
alpha=None,
alphaEbeta=None,
doTrackTruncationGrowth=0,
doPrecompEntropy=0,
doPrecompMergeEntropy=0,
mergePairSelection=None,
mPairIDs=None,
trackDocUsage=0,
**kwargs):
""" Calculate summary from local parameters for given data slice.
Parameters
-------
Data : bnpy data object
LP : local param dict with fields
resp : Data.nObs x K array,
where resp[n,k] = posterior resp of comp k
doPrecompEntropy : boolean flag
indicates whether to precompute ELBO terms in advance
used for memoized learning algorithms (moVB)
Returns
-------
SS : SuffStatBag with K components
Relevant fields
* nDoc : scalar float
Counts total documents available in provided data.
* sumLogPi : 1D array, size K
Entry k equals \sum_{d in docs} E[ \log \pi_{dk} ]
* sumLogPiRem : scalar float
Equals sum over docs of probability of inactive topics.
Also has optional ELBO field when precompELBO is True
* Hvec : 1D array, size K
Vector of entropy contributions from each comp.
Hvec[k] = \sum_{n=1}^N H[q(z_n)], a function of 'resp'
"""
if mPairIDs is None:
M = 0
else:
M = len(mPairIDs)
K = LP['DocTopicCount'].shape[1]
if 'digammaSumTheta' not in LP:
digammaSumTheta = digamma(LP['theta'].sum(axis=1) + LP['thetaRem'])
LP['digammaSumTheta'] = digammaSumTheta # Used for merges
if 'ElogPi' not in LP:
LP['ElogPiRem'] = digamma(LP['thetaRem']) - LP['digammaSumTheta']
LP['ElogPi'] = digamma(LP['theta']) - \
LP['digammaSumTheta'][:, np.newaxis]
SS = SuffStatBag(K=K, D=Dslice.dim, M=M)
SS.setField('nDoc', Dslice.nDoc, dims=None)
SS.setField('sumLogPi', np.sum(LP['ElogPi'], axis=0), dims='K')
if 'ElogPiEmptyComp' in LP:
sumLogPiEmptyComp = np.sum(LP['ElogPiEmptyComp']) - \
np.sum(LP['ElogPiOrigComp'])
SS.setField('sumLogPiEmptyComp', sumLogPiEmptyComp, dims=None)
if doTrackTruncationGrowth:
remvec = np.zeros(K)
remvec[K - 1] = np.sum(LP['ElogPiRem'])
SS.setField('sumLogPiRemVec', remvec, dims='K')
else:
SS.setField('sumLogPiRem', np.sum(LP['ElogPiRem']), dims=None)
if doPrecompEntropy:
Mdict = calcELBO_NonlinearTerms(Data=Dslice,
LP=LP,
returnMemoizedDict=1)
if type(Mdict['Hresp']) == float:
# SPARSE HARD ASSIGNMENTS
SS.setELBOTerm('Hresp', Mdict['Hresp'], dims=None)
else:
SS.setELBOTerm('Hresp', Mdict['Hresp'], dims=('K', ))
SS.setELBOTerm('slackTheta', Mdict['slackTheta'], dims='K')
SS.setELBOTerm('gammalnTheta', Mdict['gammalnTheta'], dims='K')
if 'ElogPiEmptyComp' in LP:
SS.setELBOTerm('slackThetaEmptyComp', Mdict['slackThetaEmptyComp'])
SS.setELBOTerm('gammalnThetaEmptyComp',
Mdict['gammalnThetaEmptyComp'])
SS.setELBOTerm('HrespEmptyComp', Mdict['HrespEmptyComp'])
else:
SS.setELBOTerm('gammalnSumTheta',
Mdict['gammalnSumTheta'],
dims=None)
SS.setELBOTerm('slackThetaRem', Mdict['slackThetaRem'], dims=None)
SS.setELBOTerm('gammalnThetaRem',
Mdict['gammalnThetaRem'].sum(),
dims=None)
if doPrecompMergeEntropy:
if mPairIDs is None:
raise NotImplementedError("TODO: all pairs for merges")
m_Hresp = calcHrespForSpecificMergePairs(LP, Dslice, mPairIDs)
if m_Hresp is not None:
SS.setMergeTerm('Hresp', m_Hresp, dims=('M'))
m_sumLogPi = np.zeros(M)
m_gammalnTheta = np.zeros(M)
m_slackTheta = np.zeros(M)
for m, (kA, kB) in enumerate(mPairIDs):
theta_vec = LP['theta'][:, kA] + LP['theta'][:, kB]
ElogPi_vec = digamma(theta_vec) - LP['digammaSumTheta']
m_gammalnTheta[m] = np.sum(gammaln(theta_vec))
m_sumLogPi[m] = np.sum(ElogPi_vec)
# slack = (Ndm - theta_dm) * E[log pi_dm]
slack_vec = ElogPi_vec
slack_vec *= -1 * (alphaEbeta[kA] + alphaEbeta[kB])
m_slackTheta[m] = np.sum(slack_vec)
SS.setMergeTerm('gammalnTheta', m_gammalnTheta, dims=('M'))
SS.setMergeTerm('sumLogPi', m_sumLogPi, dims=('M'))
SS.setMergeTerm('slackTheta', m_slackTheta, dims=('M'))
# Uncomment this for verification of merge calculations.
# for (kA, kB) in mPairIDs:
# self.verifySSForMergePair(Data, SS, LP, kA, kB)
# .... end merge computations
# Selection terms (computes doc-topic correlation)
if mergePairSelection is not None:
if mergePairSelection.count('corr') > 0:
Tmat = LP['DocTopicCount']
SS.setSelectionTerm('DocTopicPairMat',
np.dot(Tmat.T, Tmat),
dims=('K', 'K'))
SS.setSelectionTerm('DocTopicSum', np.sum(Tmat, axis=0), dims='K')
if trackDocUsage:
# Track num of times a topic appears nontrivially in a doc
DocUsage = np.sum(LP['DocTopicCount'] > 0.01, axis=0)
SS.setSelectionTerm('DocUsageCount', DocUsage, dims='K')
Pi = LP['theta'] / LP['theta'].sum(axis=1)[:, np.newaxis]
SumPi = np.sum(Pi, axis=0)
SS.setSelectionTerm('SumPi', SumPi, dims='K')
return SS
def calcSummaryStats(Dslice,
LP=None,
alpha=None,
doPrecompEntropy=False,
cslice=(0, None),
**kwargs):
""" Calculate summary from local parameters for given data slice.
Parameters
-------
Data : bnpy data object
LP : local param dict with fields
resp : Data.nObs x K array,
where resp[n,k] = posterior resp of comp k
doPrecompEntropy : boolean flag
indicates whether to precompute ELBO terms in advance
used for memoized learning algorithms (moVB)
Returns
-------
SS : SuffStatBag with K components
* nDoc : scalar float
Counts total documents available in provided data.
Also has optional ELBO field when precompELBO is True
* Hvec : 1D array, size K
Vector of entropy contributions from each comp.
Hvec[k] = \sum_{n=1}^N H[q(z_n)], a function of 'resp'
"""
K = LP['DocTopicCount'].shape[1]
SS = SuffStatBag(K=K, D=Dslice.dim)
if cslice[1] is None:
SS.setField('nDoc', Dslice.nDoc, dims=None)
else:
SS.setField('nDoc', cslice[1] - cslice[0], dims=None)
SS.setField('nDoc', Dslice.nDoc, dims=None)
if doPrecompEntropy:
assert 'theta' in LP
Lalloc = L_alloc(Dslice, LP, alpha=alpha)
SS.setELBOTerm('L_alloc', Lalloc, dims=None)
if 'nnzPerRow' in LP and LP['nnzPerRow'] == 1:
SS.setELBOTerm('Hvec', 0.0, dims=None)
else:
Hvec = L_entropy(Dslice, LP, returnVector=1)
SS.setELBOTerm('Hvec', Hvec, dims='K')
return SS
def calcSummaryStats(Data, SS, LP, **kwargs):
''' Calculate summary statistics for given dataset and local parameters
Returns
--------
SS : SuffStatBag object, with K components.
'''
if not hasattr(Data, 'X_NE'):
Data.X_NE = np.hstack([Data.X, np.ones(Data.nObs)[:, np.newaxis]])
Y_N = Data.Y
X_NE = Data.X_NE
E = X_NE.shape[1]
if 'resp' in LP:
# Dense responsibility calculations
resp = LP['resp']
K = resp.shape[1]
S_yy_K = dotATB(resp, np.square(Y_N)).flatten()
S_yx_KE = dotATB(resp, Y_N * X_NE)
# Expected outer product
S_xxT_KEE = np.zeros((K, E, E))
sqrtResp_k_N = np.sqrt(resp[:, 0])
sqrtR_X_k_NE = sqrtResp_k_N[:, np.newaxis] * X_NE
S_xxT_KEE[0] = dotATA(sqrtR_X_k_NE)
for k in xrange(1, K):
np.sqrt(resp[:, k], out=sqrtResp_k_N)
np.multiply(sqrtResp_k_N[:, np.newaxis], X_NE, out=sqrtR_X_k_NE)
S_xxT_KEE[k] = dotATA(sqrtR_X_k_NE)
else:
raise ValueError("TODO")
spR = LP['spR']
K = spR.shape[1]
if SS is None:
SS = SuffStatBag(K=K, D=Data.dim, E=E)
elif not hasattr(SS, 'E'):
SS._Fields.E = E
SS.setField('xxT_KEE', S_xxT_KEE, dims=('K', 'E', 'E'))
SS.setField('yx_KE', S_yx_KE, dims=('K', 'E'))
SS.setField('yy_K', S_yy_K, dims=('K'))
# Expected count for each k
# Usually computed by allocmodel. But just in case...
if not hasattr(SS, 'N'):
if 'resp' in LP:
SS.setField('N', LP['resp'].sum(axis=0), dims='K')
else:
SS.setField('N', as1D(toCArray(LP['spR'].sum(axis=0))), dims='K')
#SS.setField("N_K", SS.N, dims="K")
return SS
def calcSummaryStats(Data,
LP,
doPrecompEntropy=False,
doPrecompMergeEntropy=False,
mPairIDs=None,
mergePairSelection=None,
trackDocUsage=False,
**kwargs):
''' Calculate sufficient statistics for global updates.
Parameters
-------
Data : bnpy data object
LP : local param dict with fields
resp : Data.nObs x K array,
where resp[n,k] = posterior resp of comp k
doPrecompEntropy : boolean flag
indicates whether to precompute ELBO terms in advance
used for memoized learning algorithms (moVB)
doPrecompMergeEntropy : boolean flag
indicates whether to precompute ELBO terms in advance
for certain merge candidates.
Returns
-------
SS : SuffStatBag with K components
Summarizes for this mixture model, with fields
* N : 1D array, size K
N[k] = expected number of items assigned to comp k
Also has optional ELBO field when precompELBO is True
* ElogqZ : 1D array, size K
Vector of entropy contributions from each comp.
ElogqZ[k] = \sum_{n=1}^N resp[n,k] log resp[n,k]
Also has optional Merge field when precompMergeELBO is True
* ElogqZ : 2D array, size K x K
Each term is scalar entropy of merge candidate
'''
if mPairIDs is not None and len(mPairIDs) > 0:
M = len(mPairIDs)
else:
M = 0
if 'resp' in LP:
Nvec = np.sum(LP['resp'], axis=0)
K = Nvec.size
else:
# Sparse assignment case
Nvec = as1D(toCArray(LP['spR'].sum(axis=0)))
K = LP['spR'].shape[1]
if hasattr(Data, 'dim'):
SS = SuffStatBag(K=K, D=Data.dim, M=M)
else:
SS = SuffStatBag(K=K, D=Data.vocab_size, M=M)
SS.setField('N', Nvec, dims=('K'))
if doPrecompEntropy:
Mdict = calcELBO_NonlinearTerms(LP=LP, returnMemoizedDict=1)
if type(Mdict['Hresp']) == float:
# SPARSE HARD ASSIGNMENTS
SS.setELBOTerm('Hresp', Mdict['Hresp'], dims=None)
else:
SS.setELBOTerm('Hresp', Mdict['Hresp'], dims=('K', ))
if doPrecompMergeEntropy:
m_Hresp = None
if 'resp' in LP:
m_Hresp = -1 * NumericUtil.calcRlogR_specificpairs(
LP['resp'], mPairIDs)
elif 'spR' in LP:
if LP['nnzPerRow'] > 1:
m_Hresp = calcSparseMergeRlogR(spR_csr=LP['spR'],
nnzPerRow=LP['nnzPerRow'],
mPairIDs=mPairIDs)
else:
raise ValueError("Need resp or spR in LP")
if m_Hresp is not None:
assert m_Hresp.size == len(mPairIDs)
SS.setMergeTerm('Hresp', m_Hresp, dims=('M'))
if trackDocUsage:
Usage = np.sum(LP['resp'] > 0.01, axis=0)
SS.setSelectionTerm('DocUsageCount', Usage, dims='K')
return SS