def setPostFactors(self, obsModel=None, SS=None, LP=None, Data=None,
nu=0, B=0, m=0, kappa=0,
**kwargs):
''' Set attribute Post to provided values.
'''
self.ClearCache()
if obsModel is not None:
if hasattr(obsModel, 'Post'):
self.Post = obsModel.Post.copy()
self.K = self.Post.K
else:
self.setPostFromEstParams(obsModel.EstParams)
return
if LP is not None and Data is not None:
SS = self.calcSummaryStats(Data, None, LP)
if SS is not None:
self.updatePost(SS)
else:
m = as2D(m)
if m.shape[1] != self.D:
m = m.T.copy()
K, _ = m.shape
self.Post = ParamBag(K=K, D=self.D)
self.Post.setField('nu', as1D(nu), dims=('K'))
self.Post.setField('B', B, dims=('K', 'D', 'D'))
self.Post.setField('m', m, dims=('K', 'D'))
self.Post.setField('kappa', as1D(kappa), dims=('K'))
self.K = self.Post.K
def buildCostMatrix(zHat, zTrue):
''' Construct cost matrix for alignment of estimated and true sequences
Args
--------
zHat : 1D array
each entry is an integer label in {0, 1, ... Kest-1}
zTrue : 1D array
each entry is an integer label in {0, 1, ... Ktrue-1}
with optional negative state labels
Returns
--------
CostMatrix : 2D array, size Ktrue x Kest
CostMatrix[j,k] = count of events across all timesteps,
where j is assigned, but k is not.
'''
zHat = as1D(zHat)
zTrue = as1D(zTrue)
Ktrue = int(np.max(zTrue)) + 1
Kest = int(np.max(zHat)) + 1
K = np.maximum(Ktrue, Kest)
CostMatrix = np.zeros((K, K))
for ktrue in range(K):
for kest in range(K):
CostMatrix[ktrue, kest] = np.sum(
np.logical_and(zTrue == ktrue, zHat != kest))
return CostMatrix
def packParamBagForPost(nu=None,
beta=None,
m=None,
kappa=None,
D=None,
Post=None,
**kwargs):
'''
'''
m = as2D(m)
beta = as2D(beta)
if D is None:
D = m.shape[1]
if m.shape[1] != D:
m = m.T.copy()
if beta.shape[1] != D:
beta = beta.T.copy()
K, _ = m.shape
if Post is None:
Post = ParamBag(K=K, D=D)
else:
assert isinstance(Post, ParamBag)
assert Post.K == K
assert Post.D == D
Post.setField('nu', as1D(nu), dims=('K'))
Post.setField('beta', beta, dims=('K', 'D'))
Post.setField('m', m, dims=('K', 'D'))
Post.setField('kappa', as1D(kappa), dims=('K'))
return Post
def packParamBagForPost(pnu_K=None,
ptau_K=None,
w_KE=None,
P_KEE=None,
Post=None,
**kwargs):
''' Parse provided array args and pack into parameter bag
Returns
-------
Post : ParamBag, with K clusters
'''
pnu_K = as1D(pnu_K)
ptau_K = as1D(ptau_K)
w_KE = as2D(w_KE)
P_KEE = as3D(P_KEE)
K = pnu_K.size
E = w_KE.shape[1]
if Post is None:
Post = ParamBag(K=K, D=E - 1, E=E)
elif not hasattr(Post, 'E'):
Post.E = E
assert Post.K == K
assert Post.D == E - 1
assert Post.E == E
Post.setField('pnu_K', pnu_K, dims=('K'))
Post.setField('ptau_K', ptau_K, dims=('K'))
Post.setField('w_KE', w_KE, dims=('K', 'E'))
Post.setField('P_KEE', P_KEE, dims=('K', 'E', 'E'))
return Post
def __init__(self, word_id=None, word_count=None, doc_range=None,
vocab_size=0, vocabList=None, vocabfile=None,
summary=None,
nDocTotal=None, TrueParams=None, **kwargs):
''' Constructor for BagOfWordsData object.
Represents bag-of-words dataset via several 1D vectors,
where each entry gives the id/count of a single token.
Parameters
-------
word_id : 1D array, size U
word_id[i] gives the int type for distinct word i.
Range: [0, 1, 2, 3, ... vocab_size-1]
word_count : 1D array, size U
word_count[i] gives the number of times distinct word i
appears in its corresponding document.
Range: [1, 2, 3, ...]
doc_range : 1D array, size nDoc+1
Defines section of each corpus-wide vector which belongs
to each document d:
word_id_d = self.word_id[doc_range[d]:doc_range[d+1]]
word_ct_d = self.word_count[doc_range[d]:doc_range[d+1]]
vocab_size : int
size of set of possible vocabulary words
vocabList : list of strings
specifies the word associated with each vocab type id
nDoc : int
number of groups (documents) represented in memory
by this object.
nDocTotal : int
total size of the corpus
will differ from nDoc when this dataset represents a
small minibatch of some larger corpus.
TrueParams : None [default], or dict of true parameters.
'''
self.word_id = as1D(toCArray(word_id, dtype=np.int32))
self.word_count = as1D(toCArray(word_count, dtype=np.float64))
self.doc_range = as1D(toCArray(doc_range, dtype=np.int32))
self.vocab_size = int(vocab_size)
if summary is not None:
self.summary = summary
# Store "true" parameters that generated toy-data, if provided
if TrueParams is not None:
self.TrueParams = TrueParams
# Add dictionary of vocab words, if provided
if vocabList is not None:
self.vocabList = vocabList
elif vocabfile is not None:
with open(vocabfile, 'r') as f:
self.vocabList = [x.strip() for x in f.readlines()]
else:
self.vocabList = None
if vocab_size == 0 and self.vocabList is not None:
self.vocab_size = len(self.vocabList)
self._verify_attributes()
self._set_corpus_size_attributes(nDocTotal)
def __init__(self, X=None, doc_range=None, nDocTotal=None,
Xprev=None, TrueZ=None,
TrueParams=None, fileNames=None, summary=None, **kwargs):
''' Create an instance of GroupXData for provided array X
Post Condition
---------
self.X : 2D array, size N x D
with standardized dtype, alignment, byteorder.
self.Xprev : 2D array, size N x D
with standardized dtype, alignment, byteorder.
self.doc_range : 1D array, size nDoc+1
'''
self.X = as2D(toCArray(X, dtype=np.float64))
self.doc_range = as1D(toCArray(doc_range, dtype=np.int32))
if summary is not None:
self.summary = summary
if Xprev is not None:
self.Xprev = as2D(toCArray(Xprev, dtype=np.float64))
# Verify attributes are consistent
self._set_dependent_params(doc_range, nDocTotal)
self._check_dims()
# Add optional true parameters / true hard labels
if TrueParams is not None:
self.TrueParams = dict()
for key, arr in TrueParams.items():
self.TrueParams[key] = toCArray(arr)
if TrueZ is not None:
if not hasattr(self, 'TrueParams'):
self.TrueParams = dict()
self.TrueParams['Z'] = as1D(toCArray(TrueZ))
self.TrueParams['K'] = np.unique(self.TrueParams['Z']).size
# Add optional source files for each group/sequence
if fileNames is not None:
if hasattr(fileNames, 'shape') and fileNames.shape == (1, 1):
fileNames = fileNames[0, 0]
if len(fileNames) > 1:
self.fileNames = [str(x).strip()
for x in np.squeeze(fileNames)]
else:
self.fileNames = [str(fileNames[0])]
# Add extra data attributes custom for the dataset
for key in kwargs:
if hasattr(self, key):
continue
if not key.startswith("__"):
arr = np.squeeze(as1D(kwargs[key]))
if arr.shape == ():
try:
arr = float(arr)
except TypeError:
continue
setattr(self, key, arr)
def calcHammingDistance(zTrue, zHat, excludeNegLabels=1, verbose=0,
**kwargs):
''' Compute Hamming distance: sum of all timesteps with different labels.
Normalizes result to be within [0, 1].
Args
--------
zHat : 1D array
each entry is an integer label in {0, 1, ... Kest-1}
zTrue : 1D array
each entry is an integer label in {0, 1, ... Ktrue-1}
Returns
------
d : int
Hamming distance from zTrue to zHat.
Examples
------
>>> calcHammingDistance([0, 0, 1, 1], [0, 0, 1, 1])
0.0
>>> calcHammingDistance([0, 0, 1, 1], [0, 0, 1, 2])
0.25
>>> calcHammingDistance([0, 0, 1, 1], [1, 1, 0, 0])
1.0
>>> calcHammingDistance([1, 1, 0, -1], [1, 1, 0, 0])
0.0
>>> calcHammingDistance([-1, -1, -2, 3], [1, 2, 3, 3])
0.0
>>> calcHammingDistance([-1, -1, 0, 1], [1, 2, 0, 1], excludeNegLabels=1)
0.0
>>> calcHammingDistance([-1, -1, 0, 1], [1, 2, 0, 1], excludeNegLabels=0)
0.5
'''
zHat = as1D(zHat)
zTrue = as1D(zTrue)
if excludeNegLabels:
assert np.sum(zHat < 0) == 0
good_tstep_mask = zTrue >= 0
nGood = np.sum(good_tstep_mask)
if verbose and np.sum(good_tstep_mask) < zTrue.size:
print('EXCLUDED %d/%d timesteps' % (np.sum(zTrue < 0), zTrue.size))
dist = np.sum(zTrue[good_tstep_mask] != zHat[good_tstep_mask])
dist = dist/float(nGood)
else:
dist = np.sum(zTrue != zHat) / float(zHat.size)
return dist
def getPrefixForLapQuery(taskpath, lapQuery):
''' Search among checkpoint laps for one nearest to query.
Returns
--------
prefix : str
For lap 1, prefix = 'Lap0001.000'.
For lap 5.5, prefix = 'Lap0005.500'.
lap : int
lap checkpoint for saved params close to lapQuery
'''
try:
saveLaps = np.loadtxt(os.path.join(taskpath, 'snapshot_lap.txt'))
except IOError:
fileList = glob.glob(os.path.join(taskpath, 'Lap*Topic*'))
if len(fileList) == 0:
fileList = glob.glob(os.path.join(taskpath, 'Lap*.log_prob_w'))
assert len(fileList) > 0
saveLaps = list()
for fpath in sorted(fileList):
basename = fpath.split(os.path.sep)[-1]
lapstr = basename[3:11]
saveLaps.append(float(lapstr))
saveLaps = np.sort(np.asarray(saveLaps))
saveLaps = as1D(saveLaps)
if lapQuery is None:
bestLap = saveLaps[-1] # take final saved value
else:
distances = np.abs(lapQuery - saveLaps)
bestLap = saveLaps[np.argmin(distances)]
return makePrefixForLap(bestLap), bestLap
def eta2pi(eta_d):
eta_d = as1D(np.asarray(eta_d))
pi_d = np.ones(eta_d.size+1)
pi_d[:-1] = np.exp(eta_d)
pi_d[:-1] += 1e-100
pi_d /= (1.0 + np.sum(pi_d[:-1]))
return pi_d
def setPostFactors(self, obsModel=None, SS=None, LP=None, Data=None,
nu=0, B=0, M=0, V=0,
**kwargs):
''' Set Post attribute to provided values.
'''
self.ClearCache()
if obsModel is not None:
if hasattr(obsModel, 'Post'):
self.Post = obsModel.Post.copy()
else:
self.setPostFromEstParams(obsModel.EstParams)
self.K = self.Post.K
return
if LP is not None and Data is not None:
SS = self.calcSummaryStats(Data, None, LP)
if SS is not None:
self.updatePost(SS)
else:
M = as3D(M)
B = as3D(B)
V = as3D(V)
K, D, E = M.shape
assert D == self.D
assert E == self.E
self.Post = ParamBag(K=K, D=self.D, E=self.E)
self.Post.setField('nu', as1D(nu), dims=('K'))
self.Post.setField('B', B, dims=('K', 'D', 'D'))
self.Post.setField('M', M, dims=('K', 'D', 'E'))
self.Post.setField('V', V, dims=('K', 'E', 'E'))
self.K = self.Post.K
def generateRandomBinaryDataFromMixture(**kwargs):
for key in Defaults:
if key not in kwargs:
kwargs[key] = Defaults[key]
phi = makePhi(**kwargs)
nObsTotal = kwargs['nObsTotal']
PRNG = np.random.RandomState(kwargs['seed'])
# Select number of observations from each cluster
beta = 1.0 / K * np.ones(K)
if nObsTotal < 2 * K:
# force examples from every cluster
nPerCluster = np.ceil(nObsTotal / K) * np.ones(K)
else:
nPerCluster = as1D(PRNG.multinomial(nObsTotal, beta, size=1))
nPerCluster = np.int32(nPerCluster)
# Generate data from each cluster!
X = np.zeros((nObsTotal, D))
Z = np.zeros(nObsTotal, dtype=np.int32)
start = 0
for k in xrange(K):
stop = start + nPerCluster[k]
X[start:stop] = np.float64(
PRNG.rand(nPerCluster[k], D) < phi[k, :][np.newaxis, :])
Z[start:stop] = k
start = stop
TrueParams = dict()
TrueParams['beta'] = beta
TrueParams['phi'] = phi
TrueParams['Z'] = Z
return XData(X, TrueParams=TrueParams)
def createParamBagForPrior(
Data=None, D=0,
pnu=0, ptau=None, w_E=0,
P_EE=None, P_diag_E=None, P_diag_val=1.0,
Prior=None,
**kwargs):
''' Initialize Prior ParamBag attribute.
Returns
-------
Prior : ParamBag
with dimension attributes K, D, E
with parameter attributes pnu, ptau, w_E, P_EE
'''
if Data is None:
D = int(D)
else:
D = int(Data.dim)
E = D + 1
# Init parameters of 1D Wishart prior on delta
pnu = np.maximum(pnu, 1e-9)
ptau = np.maximum(ptau, 1e-9)
# Initialize precision matrix of the weight vector
if P_EE is not None:
P_EE = np.asarray(P_EE)
elif P_diag_E is not None:
P_EE = np.diag(np.asarray(P_diag_E))
else:
P_EE = np.diag(P_diag_val * np.ones(E))
assert P_EE.ndim == 2
assert P_EE.shape == (E,E)
# Initialize mean of the weight vector
w_E = as1D(np.asarray(w_E))
if w_E.size < E:
w_E = np.tile(w_E, E)[:E]
assert w_E.ndim == 1
assert w_E.size == E
if Prior is None:
Prior = ParamBag(K=0, D=D, E=E)
if not hasattr(Prior, 'E'):
Prior.E = E
assert Prior.D == D
assert Prior.E == E
Prior.setField('pnu', pnu, dims=None)
Prior.setField('ptau', ptau, dims=None)
Prior.setField('w_E', w_E, dims=('E'))
Prior.setField('P_EE', P_EE, dims=('E', 'E'))
Pw_E = np.dot(P_EE, w_E)
wPw_1 = np.dot(w_E, Pw_E)
Prior.setField('Pw_E', Pw_E, dims=('E'))
Prior.setField('wPw_1', wPw_1, dims=None)
return Prior
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 calcSmoothedMu(self, X, W=None):
''' Compute smoothed estimate of mean of statistic xxT.
Args
----
X : 2D array, size N x D
Returns
-------
Mu_1 : 2D array, size D x D
Expected value of Cov[ X[n] ]
Mu_2 : 1D array, size D
Expected value of Mean[ X[n] ]
'''
if X is None:
Mu1 = self.Prior.B / self.Prior.nu
Mu2 = self.Prior.m
return Mu1, Mu2
if X.ndim == 1:
X = X[np.newaxis, :]
N, D = X.shape
# Compute suff stats
if W is None:
sum_wxxT = np.dot(X.T, X)
sum_wx = np.sum(X, axis=0)
sum_w = X.shape[0]
else:
W = as1D(W)
sqrtWX = np.sqrt(W)[:, np.newaxis] * X
sum_wxxT = np.dot(sqrtWX.T, sqrtWX)
sum_wx = np.dot(W, X)
sum_w = np.sum(W)
kappa = self.Prior.kappa + sum_w
m = (self.Prior.m * self.Prior.kappa + sum_wx) / kappa
Mu_2 = m
prior_kmmT = self.Prior.kappa * np.outer(self.Prior.m, self.Prior.m)
post_kmmT = kappa * np.outer(m, m)
B = sum_wxxT + self.Prior.B + prior_kmmT - post_kmmT
Mu_1 = B / (self.Prior.nu + sum_w)
assert Mu_1.ndim == 2
assert Mu_1.shape == (
D,
D,
)
assert Mu_2.shape == (D, )
return Mu_1, Mu_2
def calcSmoothedMu(self, X, W=None):
''' Compute smoothed estimate of mean of statistic xxT.
Args
----
X : 2D array, size N x D
Returns
-------
Mu_1 : 2D array, size D
Expected value of Var[ X[n,d] ]
Mu_2 : 1D array, size D
Expected value of Mean[ X[n] ]
'''
if X is None:
Mu1 = self.Prior.beta / self.Prior.nu
Mu2 = self.Prior.m
return Mu1, Mu2
if X.ndim == 1:
X = X[np.newaxis, :]
N, D = X.shape
# Compute suff stats
if W is None:
sum_wxx = np.sum(np.square(X), axis=0)
sum_wx = np.sum(X, axis=0)
sum_w = X.shape[0]
else:
W = as1D(W)
sum_wxx = np.dot(W, np.square(X))
sum_wx = np.dot(W, X)
sum_w = np.sum(W)
post_kappa = self.Prior.kappa + sum_w
post_m = (self.Prior.m * self.Prior.kappa + sum_wx) / post_kappa
Mu_2 = post_m
prior_kmm = self.Prior.kappa * (self.Prior.m * self.Prior.m)
post_kmm = post_kappa * (post_m * post_m)
post_beta = sum_wxx + self.Prior.beta + prior_kmm - post_kmm
Mu_1 = post_beta / (self.Prior.nu + sum_w)
assert Mu_1.ndim == 1
assert Mu_1.shape == (D, )
assert Mu_2.shape == (D, )
return Mu_1, Mu_2
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
D = Data.dim
if 'resp' in LP:
resp = LP['resp']
K = resp.shape[1]
# 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 xrange(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)
sqrtResp = np.sqrt(resp)
xxT = np.zeros((K, D, D))
for k in xrange(K):
xxT[k] = dotATA(sqrtResp[:, k][:, np.newaxis] * Data.X)
assert np.allclose(xxT, S_xxT)
else:
spR = LP['spR']
K = spR.shape[1]
# Compute expected outer-product statistic
S_xxT = calcSpRXXT(X=X, spR_csr=spR)
if SS is None:
SS = SuffStatBag(K=K, D=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 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 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 pi2eta(pi_d):
''' Transform vector on simplex to unconstrained real vector
Returns
-------
eta : 1D array, size K-1
Examples
--------
>>> print float(pi2eta(eta2pi(0.42)))
0.42
>>> print float(pi2eta(eta2pi(-1.337)))
-1.337
>>> print pi2eta(eta2pi([-1, 0, 1]))
[-1. 0. 1.]
'''
pi_d = as1D(np.asarray(pi_d))
eta_d = pi_d[:-1] / pi_d[-1]
np.log(eta_d, out=eta_d)
return eta_d
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 calcSmoothedMu(self, X, W=None):
''' Compute smoothed estimate of mean of statistic xxT.
Args
----
X : 2D array, size N x D
Returns
-------
Mu : 2D array, size D x D
'''
Prior_nu = self.Prior.nu - self.D - 1
# Prior_nu = self.Prior.nu
if X is None:
Mu = self.Prior.B / (Prior_nu)
return Mu
if X.ndim == 1:
X = X[np.newaxis, :]
N, D = X.shape
# Compute suff stats
if W is None:
sum_wxxT = np.dot(X.T, X)
sum_w = X.shape[0]
else:
W = as1D(W)
wX = np.sqrt(W)[:, np.newaxis] * X
sum_wxxT = np.dot(wX.T, wX)
sum_w = np.sum(W)
Mu = (self.Prior.B + sum_wxxT) / (Prior_nu + sum_w)
assert Mu.ndim == 2
assert Mu.shape == (
D,
D,
)
return Mu
def plotGauss1DFromHModel(hmodel,
compListToPlot=None,
compsToHighlight=None,
activeCompIDs=None,
MaxKToDisplay=50,
proba_thr=0.0001,
ax_handle=None,
Colors=Colors,
dataset=None,
**kwargs):
''' Make line plot of pdf for each component (1D observations).
'''
if ax_handle is not None:
pylab.sca(ax_handle)
if compsToHighlight is not None:
compsToHighlight = as1D(np.asarray(compsToHighlight))
else:
compsToHighlight = list()
if compListToPlot is None:
compListToPlot = np.arange(0, hmodel.obsModel.K)
if activeCompIDs is None:
activeCompIDs = np.arange(0, hmodel.obsModel.K)
# Load appearance probabilities as single vector
if hmodel.allocModel.K == hmodel.obsModel.K:
w = hmodel.allocModel.get_active_comp_probs()
else:
w = np.ones(hmodel.obsModel.K)
if dataset is not None:
if hasattr(dataset, 'X'):
pylab.hist(dataset.X[:, 0], 50, density=1)
#Xtile = np.tile(Data.X[:, 0], (2, 1))
#ys = 0.1 * np.arange(2)
#pylab.plot(Xtile, ys, 'k-')
nSkip = 0
nGood = 0
for ii, compID in enumerate(compListToPlot):
if compID not in activeCompIDs:
continue
kk = np.flatnonzero(activeCompIDs == compID)
assert kk.size == 1
kk = kk[0]
if w[kk] < proba_thr and compID not in compsToHighlight:
nSkip += 1
continue
mu = hmodel.obsModel.get_mean_for_comp(kk)
sigma2 = hmodel.obsModel.get_covar_mat_for_comp(kk)
if len(compsToHighlight) == 0 or compID in compsToHighlight:
color = Colors[ii % len(Colors)]
plotGauss1D(mu, sigma2, color=color)
elif kk not in compsToHighlight:
plotGauss1D(mu, sigma2, color='k')
nGood += 1
if nGood >= MaxKToDisplay:
print('DISPLAY LIMIT EXCEEDED. Showing %d/%d components' \
% (nGood, len(activeCompIDs)))
break
if nSkip > 0:
print('SKIPPED %d comps with size below %.2f' % (nSkip, proba_thr))
def loadTopicModel(matfilepath,
queryLap=None,
prefix=None,
returnWordCounts=0,
returnTPA=0,
normalizeTopics=0,
normalizeProbs=0,
**kwargs):
''' Load saved topic model
Returns
-------
topics : 2D array, K x vocab_size (if returnTPA)
probs : 1D array, size K (if returnTPA)
alpha : scalar (if returnTPA)
hmodel : HModel
WordCounts : 2D array, size K x vocab_size (if returnWordCounts)
'''
if prefix is None:
prefix, lapQuery = getPrefixForLapQuery(matfilepath, queryLap)
# avoids circular import
from bnpy.HModel import HModel
if len(glob.glob(os.path.join(matfilepath, "*.log_prob_w"))) > 0:
return loadTopicModelFromMEDLDA(matfilepath,
prefix,
returnTPA=returnTPA)
snapshotList = glob.glob(os.path.join(matfilepath, 'Lap*TopicSnapshot'))
matfileList = glob.glob(os.path.join(matfilepath, 'Lap*TopicModel.mat'))
if len(snapshotList) > 0:
if prefix is None:
snapshotList.sort()
snapshotPath = snapshotList[-1]
else:
snapshotPath = None
for curPath in snapshotList:
if curPath.count(prefix):
snapshotPath = curPath
return loadTopicModelFromTxtFiles(snapshotPath,
normalizeTopics=normalizeTopics,
normalizeProbs=normalizeProbs,
returnWordCounts=returnWordCounts,
returnTPA=returnTPA)
if prefix is not None:
matfilepath = os.path.join(matfilepath, prefix + 'TopicModel.mat')
Mdict = loadDictFromMatfile(matfilepath)
if 'SparseWordCount_data' in Mdict:
data = np.asarray(Mdict['SparseWordCount_data'], dtype=np.float64)
K = int(Mdict['K'])
vocab_size = int(Mdict['vocab_size'])
try:
indices = Mdict['SparseWordCount_indices']
indptr = Mdict['SparseWordCount_indptr']
WordCounts = scipy.sparse.csr_matrix((data, indices, indptr),
shape=(K, vocab_size))
except KeyError:
rowIDs = Mdict['SparseWordCount_i'] - 1
colIDs = Mdict['SparseWordCount_j'] - 1
WordCounts = scipy.sparse.csr_matrix((data, (rowIDs, colIDs)),
shape=(K, vocab_size))
Mdict['WordCounts'] = WordCounts.toarray()
if returnTPA:
# Load topics : 2D array, K x vocab_size
if 'WordCounts' in Mdict:
topics = Mdict['WordCounts'] + Mdict['lam']
else:
topics = Mdict['topics']
topics = as2D(toCArray(topics, dtype=np.float64))
assert topics.ndim == 2
K = topics.shape[0]
if normalizeTopics:
topics /= topics.sum(axis=1)[:, np.newaxis]
# Load probs : 1D array, size K
try:
probs = Mdict['probs']
except KeyError:
probs = (1.0 / K) * np.ones(K)
probs = as1D(toCArray(probs, dtype=np.float64))
assert probs.ndim == 1
assert probs.size == K
if normalizeProbs:
probs = probs / np.sum(probs)
# Load alpha : scalar float > 0
try:
alpha = float(Mdict['alpha'])
except KeyError:
if 'alpha' in os.environ:
alpha = float(os.environ['alpha'])
else:
raise ValueError('Unknown parameter alpha')
if 'eta' in Mdict:
return topics, probs, alpha, as1D(toCArray(Mdict['eta']))
return topics, probs, alpha
infAlg = 'VB'
if 'gamma' in Mdict:
aPriorDict = dict(alpha=Mdict['alpha'], gamma=Mdict['gamma'])
HDPTopicModel = AllocModelConstructorsByName['HDPTopicModel']
amodel = HDPTopicModel(infAlg, aPriorDict)
else:
FiniteTopicModel = AllocModelConstructorsByName['FiniteTopicModel']
amodel = FiniteTopicModel(infAlg, dict(alpha=Mdict['alpha']))
omodel = ObsModelConstructorsByName['Mult'](infAlg, **Mdict)
hmodel = HModel(amodel, omodel)
hmodel.set_global_params(**Mdict)
if returnWordCounts:
return hmodel, Mdict['WordCounts']
return hmodel
def from_dict(self, Dict):
self.inferType = Dict['inferType']
self.K = Dict['K']
self.uhat = as1D(Dict['uhat'])
def alignEstimatedStateSeqToTruth(zHat, zTrue, useInfo=None, returnInfo=False):
''' Relabel the states in zHat to minimize the hamming-distance to zTrue
Args
--------
zHat : 1D array
each entry is an integer label in {0, 1, ... Kest-1}
zTrue : 1D array
each entry is an integer label in {0, 1, ... Ktrue-1}
Returns
--------
zHatAligned : 1D array
relabeled version of zHat that aligns to zTrue
AInfo : dict
information about the alignment
'''
zHat = as1D(zHat)
zTrue = as1D(zTrue)
Kest = zHat.max() + 1
Ktrue = zTrue.max() + 1
if useInfo is None:
if not hasMunkres:
raise ImportError(
"alignEstimatedStateSeqToTruth requires the munkres package." +
" Please install via 'pip install munkres'")
CostMatrix = buildCostMatrix(zHat, zTrue)
MunkresAlg = munkres.Munkres()
tmpAlignedRowColPairs = MunkresAlg.compute(CostMatrix)
AlignedRowColPairs = list()
OrigToAlignedMap = dict()
AlignedToOrigMap = dict()
for (ktrue, kest) in tmpAlignedRowColPairs:
if kest < Kest:
AlignedRowColPairs.append((ktrue, kest))
OrigToAlignedMap[kest] = ktrue
AlignedToOrigMap[ktrue] = kest
else:
# Unpack existing alignment info
AlignedRowColPairs = useInfo['AlignedRowColPairs']
CostMatrix = useInfo['CostMatrix']
AlignedToOrigMap = useInfo['AlignedToOrigMap']
OrigToAlignedMap = useInfo['OrigToAlignedMap']
Ktrue = useInfo['Ktrue']
Kest = useInfo['Kest']
assert np.allclose(Ktrue, zTrue.max() + 1)
Khat = zHat.max() + 1
# Account for extra states present in zHat
# that have never been aligned before.
# They should align to the next available UID in set
# [Ktrue, Ktrue+1, Ktrue+2, ...]
# so they don't get confused for a true label
ktrueextra = np.max([r for r, c in AlignedRowColPairs])
ktrueextra = int(np.maximum(ktrueextra + 1, Ktrue))
for khat in np.arange(Kest, Khat + 1):
if khat in OrigToAlignedMap:
continue
OrigToAlignedMap[khat] = ktrueextra
AlignedToOrigMap[ktrueextra] = khat
AlignedRowColPairs.append((ktrueextra, khat))
ktrueextra += 1
zHatA = -1 * np.ones_like(zHat)
for kest in np.unique(zHat):
mask = zHat == kest
zHatA[mask] = OrigToAlignedMap[kest]
assert np.all(zHatA >= 0)
if returnInfo:
return zHatA, dict(CostMatrix=CostMatrix,
AlignedRowColPairs=AlignedRowColPairs,
OrigToAlignedMap=OrigToAlignedMap,
AlignedToOrigMap=AlignedToOrigMap,
Ktrue=Ktrue,
Kest=Kest)
else:
return zHatA
def alignEstimatedStateSeqToTruth(zHat, zTrue, useInfo=None, returnInfo=False, standardize_order_of_extras=True):
''' Relabel the states in zHat to minimize the hamming-distance to zTrue
Args
--------
zHat : 1D array
each entry is an integer label in {0, 1, ... Kest-1}
zTrue : 1D array
each entry is an integer label in {0, 1, ... Ktrue-1}
Returns
--------
zHatAligned : 1D array
relabeled version of zHat that aligns to zTrue
AInfo : dict
information about the alignment
'''
zHat = as1D(zHat)
zTrue = as1D(zTrue)
Kest = zHat.max() + 1
Ktrue = zTrue.max() + 1
if useInfo is None:
if not hasMunkres:
raise ImportError(
"alignEstimatedStateSeqToTruth requires the munkres package."
+ " Please install via 'pip install munkres'")
CostMatrix = buildCostMatrix(zHat, zTrue)
MunkresAlg = munkres.Munkres()
tmpAlignedRowColPairs = MunkresAlg.compute(CostMatrix)
AlignedRowColPairs = list()
OrigToAlignedMap = dict()
AlignedToOrigMap = dict()
for (ktrue, kest) in tmpAlignedRowColPairs:
if kest < Kest:
AlignedRowColPairs.append((ktrue, kest))
OrigToAlignedMap[kest] = ktrue
AlignedToOrigMap[ktrue] = kest
if Ktrue < Kest and np.unique(zTrue).size < np.unique(zHat).size and standardize_order_of_extras:
# All extra states will have ids in Ktrue, Ktrue+1, Ktrue+2, ....
# Break ties by assigning these in ascending order by first appearance
extra_states = []
for (kt, ke) in AlignedRowColPairs:
if kt >= Ktrue and ke in zHat:
extra_states.append(ke)
rank = dict()
for x in sorted(extra_states):
match_ids = np.flatnonzero(zHat == x)
if match_ids.size == 0:
continue
rank[x] = match_ids[0]
## Renumber by rank from smallest to largest
old_ids, rank_vals = zip(*list(rank.items()))
old_ids = np.asarray(old_ids)
new_ids = Ktrue + np.argsort(np.asarray(rank_vals))
old2new = dict(zip(old_ids, new_ids))
OrigToAlignedMap = dict()
AlignedToOrigMap = dict()
for a, b in AlignedRowColPairs:
newa = old2new.get(b, a)
OrigToAlignedMap[b] = newa
AlignedToOrigMap[newa] = b
AlignedRowColPairs = list(AlignedToOrigMap.items())
else:
# Unpack existing alignment info
AlignedRowColPairs = useInfo['AlignedRowColPairs']
CostMatrix = useInfo['CostMatrix']
AlignedToOrigMap = useInfo['AlignedToOrigMap']
OrigToAlignedMap = useInfo['OrigToAlignedMap']
Ktrue = useInfo['Ktrue']
Kest = useInfo['Kest']
assert Ktrue >= zTrue.max() + 1
Khat = zHat.max() + 1
# Account for extra states present in zHat
# that have never been aligned before.
# They should align to the next available UID in set
# [Ktrue, Ktrue+1, Ktrue+2, ...]
# so they don't get confused for a true label
ktrueextra = np.max([r for r, c in AlignedRowColPairs])
ktrueextra = int(np.maximum(ktrueextra + 1, Ktrue))
for khat in np.arange(Kest, Khat + 1):
if khat in OrigToAlignedMap:
continue
OrigToAlignedMap[khat] = ktrueextra
AlignedToOrigMap[ktrueextra] = khat
AlignedRowColPairs.append((ktrueextra, khat))
ktrueextra += 1
zHatA = -1 * np.ones_like(zHat)
for kest in np.unique(zHat):
mask = zHat == kest
zHatA[mask] = OrigToAlignedMap[kest]
assert np.all(zHatA >= 0)
if returnInfo:
return zHatA, dict(CostMatrix=CostMatrix,
AlignedRowColPairs=AlignedRowColPairs,
OrigToAlignedMap=OrigToAlignedMap,
AlignedToOrigMap=AlignedToOrigMap,
Ktrue=Ktrue,
Kest=Kest)
else:
return zHatA
def loadTopicModelFromTxtFiles(snapshotPath,
returnTPA=False,
returnWordCounts=False,
normalizeProbs=True,
normalizeTopics=True,
**kwargs):
''' Load from snapshot text files.
Returns
-------
hmodel
'''
Mdict = dict()
possibleKeys = [
'K', 'probs', 'alpha', 'beta', 'lam', 'gamma', 'nTopics', 'nTypes',
'vocab_size'
]
keyMap = dict(beta='lam', nTopics='K', nTypes='vocab_size')
for key in possibleKeys:
try:
arr = np.loadtxt(snapshotPath + "/%s.txt" % (key))
if key in keyMap:
Mdict[keyMap[key]] = arr
else:
Mdict[key] = arr
except Exception:
pass
assert 'K' in Mdict
assert 'lam' in Mdict
K = int(Mdict['K'])
V = int(Mdict['vocab_size'])
if os.path.exists(snapshotPath + "/topics.txt"):
Mdict['topics'] = np.loadtxt(snapshotPath + "/topics.txt")
Mdict['topics'] = as2D(toCArray(Mdict['topics'], dtype=np.float64))
assert Mdict['topics'].ndim == 2
assert Mdict['topics'].shape == (K, V)
else:
TWC_data = np.loadtxt(snapshotPath + "/TopicWordCount_data.txt")
TWC_inds = np.loadtxt(snapshotPath + "/TopicWordCount_indices.txt",
dtype=np.int32)
if os.path.exists(snapshotPath + "/TopicWordCount_cscindptr.txt"):
TWC_cscindptr = np.loadtxt(snapshotPath +
"/TopicWordCount_cscindptr.txt",
dtype=np.int32)
TWC = scipy.sparse.csc_matrix((TWC_data, TWC_inds, TWC_cscindptr),
shape=(K, V))
else:
TWC_csrindptr = np.loadtxt(snapshotPath +
"/TopicWordCount_indptr.txt",
dtype=np.int32)
TWC = scipy.sparse.csr_matrix((TWC_data, TWC_inds, TWC_csrindptr),
shape=(K, V))
Mdict['WordCounts'] = TWC.toarray()
if returnTPA:
# Load topics : 2D array, K x vocab_size
if 'WordCounts' in Mdict:
topics = Mdict['WordCounts'] + Mdict['lam']
else:
topics = Mdict['topics']
topics = as2D(toCArray(topics, dtype=np.float64))
assert topics.ndim == 2
K = topics.shape[0]
if normalizeTopics:
topics /= topics.sum(axis=1)[:, np.newaxis]
# Load probs : 1D array, size K
try:
probs = Mdict['probs']
except KeyError:
probs = (1.0 / K) * np.ones(K)
probs = as1D(toCArray(probs, dtype=np.float64))
assert probs.ndim == 1
assert probs.size == K
if normalizeProbs:
probs = probs / np.sum(probs)
# Load alpha : scalar float > 0
try:
alpha = float(Mdict['alpha'])
except KeyError:
if 'alpha' in os.environ:
alpha = float(os.environ['alpha'])
else:
raise ValueError('Unknown parameter alpha')
return topics, probs, alpha
# BUILD HMODEL FROM LOADED TXT
infAlg = 'VB'
# avoids circular import
from bnpy.HModel import HModel
if 'gamma' in Mdict:
aPriorDict = dict(alpha=Mdict['alpha'], gamma=Mdict['gamma'])
HDPTopicModel = AllocModelConstructorsByName['HDPTopicModel']
amodel = HDPTopicModel(infAlg, aPriorDict)
else:
FiniteTopicModel = AllocModelConstructorsByName['FiniteTopicModel']
amodel = FiniteTopicModel(infAlg, dict(alpha=Mdict['alpha']))
omodel = ObsModelConstructorsByName['Mult'](infAlg, **Mdict)
hmodel = HModel(amodel, omodel)
hmodel.set_global_params(**Mdict)
if returnWordCounts:
return hmodel, Mdict['WordCounts']
return hmodel
def __init__(self,
X=None,
nObsTotal=None,
TrueZ=None,
Xprev=None,
Y=None,
TrueParams=None,
name=None,
summary=None,
dtype='auto',
row_names=None,
column_names=None,
y_column_names=None,
xprev_column_names=None,
do_copy=True,
**kwargs):
''' Constructor for XData instance given in-memory dense array X.
Post Condition
---------
self.X : 2D array, size N x D
with standardized dtype, alignment, byteorder.
'''
if dtype == 'auto':
dtype = X.dtype
if not do_copy and X.dtype == dtype:
self.X = as2D(X)
else:
self.X = as2D(toCArray(X, dtype=dtype))
if Xprev is not None:
self.Xprev = as2D(toCArray(Xprev, dtype=dtype))
if Y is not None:
self.Y = as2D(toCArray(Y, dtype=dtype))
# Verify attributes are consistent
self._set_dependent_params(nObsTotal=nObsTotal)
self._check_dims(do_copy=do_copy)
# Add optional true parameters / true hard labels
if TrueParams is not None:
self.TrueParams = TrueParams
if TrueZ is not None:
if not hasattr(self, 'TrueParams'):
self.TrueParams = dict()
self.TrueParams['Z'] = as1D(toCArray(TrueZ))
self.TrueParams['K'] = np.unique(self.TrueParams['Z']).size
if summary is not None:
self.summary = summary
if name is not None:
self.name = str(name)
# Add optional row names
if row_names is None:
self.row_names = map(str, range(self.nObs))
else:
assert len(row_names) == self.nObs
self.row_names = map(str, row_names)
# Add optional column names
if column_names is None:
self.column_names = map(lambda n: "dim_%d" % n, range(self.dim))
else:
assert len(column_names) == self.dim
self.column_names = map(str, column_names)
if __name__ == '__main__':
N = 7
muG = np.asarray([0.02, 0.1, 0.5, 0.9, 0.98])
muG = N * np.vstack([muG, 1.0-muG]).T.copy()
print muG
phiG = mu2phi(muG, N)
print phiG
print phi2mu(mu2phi(muG, N), N)
print 'BREGMAN Dist Mat:'
print bregmanDiv(muG, muG, N)
for row in xrange(phiG.shape[0]):
phi = as1D(phiG[row])
mu = phi2mu(phi,N)
print 'phi = %.3f' % (phi)
print 'mu = ', mu
print '----------'
cdf = 0.0
for xVec in generateAllXForFixedN(N):
pdf = pdf_Phi(xVec, phi, N)
cdf += pdf
print '%.3f %.3f %.3f %s' % (
cdf,
pdf,
pdf_Mu(xVec, mu, N),
xVec)
#print pdf_Phi(xVec, mu2phi(mu,N), N),
def from_dict(self, Dict):
self.inferType = Dict['inferType']
self.K = Dict['K']
self.rho = as1D(Dict['rho'])
self.omega = as1D(Dict['omega'])
numPi_d, numf, numInfo = estimatePi2(ids_d, cts_d, topics, alpha,
scale=1.0, #/np.sum(cts_d),
approx_grad=True)
print("Numerical Pi[%d]:\n %s" % (d, pi2str(numPi_d)))
estPi_d, f, Info = estimatePi2(ids_d, cts_d, topics, alpha,
scale=1.0, #/np.sum(cts_d))
approx_grad=False,
)
print("Estimated Pi[%d]:\n %s" % (d, pi2str(estPi_d)))
PRNG = np.random.RandomState(d)
nMatch = 0
nRep = 10
for rep in range(nRep):
initPi_d = as1D(PRNG.dirichlet(K*np.ones(K), size=1))
estPiFromRand, f2, I2 = estimatePi2(ids_d, cts_d, topics, alpha,
scale=1.0, #/np.sum(cts_d),
piInit=initPi_d,
approx_grad=False)
if np.allclose(estPi_d, estPiFromRand, rtol=0, atol=atol):
nMatch += 1
else:
print("initrandom Pi[%d]:\n %s" % (
d, pi2str(estPiFromRand)))
print(f)
print(f2)
print("%d/%d random inits within %s" % (nMatch, nRep, atol))