def _debug_with_bound (itr, var_value, var_name, W, K, topicMean, sigT, vocab, vocabPrior, dtype, means, varcs, A, n):
if np.isnan(var_value).any():
printStderr ("WARNING: " + var_name + " contains NaNs")
if np.isinf(var_value).any():
printStderr ("WARNING: " + var_name + " contains INFs")
if var_value.dtype != dtype:
printStderr ("WARNING: dtype(" + var_name + ") = " + str(var_value.dtype))
old_bound = _debug_with_bound.old_bound
bound = var_bound(DataSet(W), ModelState(K, topicMean, sigT, vocab, vocabPrior, A, dtype, MODEL_NAME), QueryState(means, means.copy(), varcs, n))
diff = "" if old_bound == 0 else "%15.4f" % (bound - old_bound)
_debug_with_bound.old_bound = bound
addendum = ""
if var_name == "sigT":
try:
addendum = "det(sigT) = %g" % (la.det(sigT))
except:
addendum = "det(sigT) = <undefined>"
if isnan(bound):
printStderr ("Bound is NaN")
elif int(bound - old_bound) < 0:
printStderr ("Iter %3d Update %-15s Bound %22f (%15s) %s" % (itr, var_name, bound, diff, addendum))
else:
print ("Iter %3d Update %-15s Bound %22f (%15s) %s" % (itr, var_name, bound, diff, addendum))
def _debug_with_bound(itr, var_value, var_name, W, X, XTX, F, P, K, A, R_A, fv,
Y, R_Y, lfv, V, sigT, vocab, vocabPrior, dtype, means,
varcs, Ab, n):
if np.isnan(var_value).any():
printStderr("WARNING: " + var_name + " contains NaNs")
if np.isinf(var_value).any():
printStderr("WARNING: " + var_name + " contains INFs")
if var_value.dtype != dtype:
printStderr("WARNING: dtype(" + var_name + ") = " +
str(var_value.dtype))
modelState = ModelState(F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT, vocab,
vocabPrior, Ab, dtype, MODEL_NAME)
queryState = QueryState(means, means.copy(), varcs, n)
old_bound = _debug_with_bound.old_bound
bound = var_bound(DataSet(W, feats=X), modelState, queryState, XTX)
likely = log_likelihood(DataSet(W, feats=X), modelState, queryState)
diff = "" if old_bound == 0 else "%11.2f" % (bound - old_bound)
_debug_with_bound.old_bound = bound
if isnan(bound) or int(bound - old_bound) < 0:
printStderr(
"Iter %3d Update %-10s Bound %15.2f (%11s ) Perplexity %4.2f" %
(itr, var_name, bound, diff, np.exp(-likely / W.sum())))
else:
print("Iter %3d Update %-10s Bound %15.2f (%11s) Perplexity %4.2f" %
(itr, var_name, bound, diff, np.exp(-likely / W.sum())))
def _debug_with_bound(itr, var_value, var_name, W, X, means, docLens, K, A, U,
Y, V, covA, tv, ltv, fv, lfv, vocab, vocabPrior):
if np.isnan(var_value).any():
printStderr("WARNING: " + var_name + " contains NaNs")
if np.isinf(var_value).any():
printStderr("WARNING: " + var_name + " contains INFs")
dtype = A.dtype
old_bound = _debug_with_bound.old_bound
data = DataSet(W, X)
model = ModelState(K, A, U, Y, V, covA, tv, ltv, fv, lfv, vocab,
vocabPrior, dtype, MODEL_NAME)
query = QueryState(means, docLens)
bound = var_bound(data, model, query)
diff = "" if old_bound == 0 else "%15.4f" % (bound - old_bound)
_debug_with_bound.old_bound = bound
addendum = ""
perp = np.exp(-log_likelihood(data, model, query) / data.word_count)
if isnan(bound):
printStderr("Bound is NaN")
elif int(bound - old_bound) < 0:
printStderr(
"Iter %3d Update %-15s Bound %22f (%15s) Perplexity %5.1f %s" %
(itr, var_name, bound, diff, perp, addendum))
else:
print(
"Iter %3d Update %-15s Bound %22f (%15s) Perplexity %5.1f %s" %
(itr, var_name, bound, diff, perp, addendum))
def _debug_with_bound(itr, var_value, var_name, data, K, topicMean, topicCov,
vocab, dtype, means, varcs, A, n):
if np.isnan(var_value).any():
printStderr("WARNING: " + var_name + " contains NaNs")
if np.isinf(var_value).any():
printStderr("WARNING: " + var_name + " contains INFs")
if var_value.dtype != dtype:
printStderr("WARNING: dtype(" + var_name + ") = " +
str(var_value.dtype))
model = ModelState(K, topicMean, topicCov, vocab, A, dtype, MODEL_NAME)
query = QueryState(means, varcs, n)
old_bound = _debug_with_bound.old_bound
bound = var_bound(data, model, query)
diff = "" if old_bound == 0 else "%15.4f" % (bound - old_bound)
_debug_with_bound.old_bound = bound
addendum = ""
if var_name == "topicCov":
try:
addendum = "det(topicCov) = %g" % (la.det(topicCov))
except:
addendum = "det(topicCov) = <undefined>"
if isnan(bound):
printStderr("Bound is NaN")
else:
perp = perplexity_from_like(log_likelihood(data, model, query),
data.word_count)
if int(bound - old_bound) < 0:
printStderr(
"Iter %3d Update %-15s Bound %22f (%15s) (%5.0f) %s" %
(itr, var_name, bound, diff, perp, addendum))
else:
print("Iter %3d Update %-15s Bound %22f (%15s) (%5.0f) %s" %
(itr, var_name, bound, diff, perp, addendum))
def train (data, modelState, queryState, trainPlan):
'''
Infers the topic distributions in general, and specifically for
each individual datapoint.
Params:
W - the DxT document-term matrix
X - The DxF document-feature matrix, which is IGNORED in this case
modelState - the actual CTM model
queryState - the query results - essentially all the "local" variables
matched to the given observations
trainPlan - how to execute the training process (e.g. iterations,
log-interval etc.)
Return:
A new model object with the updated model (note parameters are
updated in place, so make a defensive copy if you want itr)
A new query object with the update query parameters
'''
W = data.words
D,_ = W.shape
# Unpack the the structs, for ease of access and efficiency
iterations, epsilon, logFrequency, diagonalPriorCov, debug = trainPlan.iterations, trainPlan.epsilon, trainPlan.logFrequency, trainPlan.fastButInaccurate, trainPlan.debug
means, expMeans, varcs, docLens = queryState.means, queryState.expMeans, queryState.varcs, queryState.docLens
K, topicMean, sigT, vocab, vocabPrior, A, dtype = modelState.K, modelState.topicMean, modelState.sigT, modelState.vocab, modelState.vocabPrior, modelState.A, modelState.dtype
# Book-keeping for logs
boundIters, boundValues, likelyValues = [], [], []
debugFn = _debug_with_bound if debug else _debug_with_nothing
# Initialize some working variables
isigT = la.inv(sigT)
R = W.copy()
pseudoObsMeans = K + NIW_PSEUDO_OBS_MEAN
pseudoObsVar = K + NIW_PSEUDO_OBS_VAR
priorSigT_diag = np.ndarray(shape=(K,), dtype=dtype)
priorSigT_diag.fill (NIW_PSI)
# Iterate over parameters
for itr in range(iterations):
# We start with the M-Step, so the parameters are consistent with our
# initialisation of the RVs when we do the E-Step
# Update the mean and covariance of the prior
topicMean = means.sum(axis = 0) / (D + pseudoObsMeans) \
if USE_NIW_PRIOR \
else means.mean(axis=0)
debugFn (itr, topicMean, "topicMean", W, K, topicMean, sigT, vocab, vocabPrior, dtype, means, varcs, A, docLens)
if USE_NIW_PRIOR:
diff = means - topicMean[np.newaxis,:]
sigT = diff.T.dot(diff) \
+ pseudoObsVar * np.outer(topicMean, topicMean)
sigT += np.diag(varcs.mean(axis=0) + priorSigT_diag)
sigT /= (D + pseudoObsVar - K)
else:
sigT = np.cov(means.T) if sigT.dtype == np.float64 else np.cov(means.T).astype(dtype)
sigT += np.diag(varcs.mean(axis=0))
if diagonalPriorCov:
diag = np.diag(sigT)
sigT = np.diag(diag)
isigT = np.diag(1./ diag)
else:
isigT = la.inv(sigT)
# FIXME Undo debug
sigT = np.eye(K)
isigT = la.inv(sigT)
debugFn (itr, sigT, "sigT", W, K, topicMean, sigT, vocab, vocabPrior, dtype, means, varcs, A, docLens)
# print(" sigT.det = " + str(la.det(sigT)))
# Building Blocks - temporarily replaces means with exp(means)
expMeans = np.exp(means - means.max(axis=1)[:,np.newaxis], out=expMeans)
R = sparseScalarQuotientOfDot(W, expMeans, vocab, out=R)
# Update the vocabulary
vocab *= (R.T.dot(expMeans)).T # Awkward order to maintain sparsity (R is sparse, expMeans is dense)
vocab += vocabPrior
vocab = normalizerows_ip(vocab)
# Reset the means to their original form, and log effect of vocab update
R = sparseScalarQuotientOfDot(W, expMeans, vocab, out=R)
V = expMeans * R.dot(vocab.T)
debugFn (itr, vocab, "vocab", W, K, topicMean, sigT, vocab, vocabPrior, dtype, means, varcs, A, docLens)
# And now this is the E-Step, though itr's followed by updates for the
# parameters also that handle the log-sum-exp approximation.
# Update the Variances: var_d = (2 N_d * A + isigT)^{-1}
varcs = np.reciprocal(docLens[:,np.newaxis] * (K-1.)/K + np.diagonal(sigT))
debugFn (itr, varcs, "varcs", W, K, topicMean, sigT, vocab, vocabPrior, dtype, means, varcs, A, docLens)
# Update the Means
rhs = V.copy()
rhs += docLens[:,np.newaxis] * means.dot(A) + isigT.dot(topicMean)
rhs -= docLens[:,np.newaxis] * rowwise_softmax(means, out=means)
if diagonalPriorCov:
means = varcs * rhs
else:
for d in range(D):
means[d, :] = la.inv(isigT + docLens[d] * A).dot(rhs[d, :])
# means -= (means[:,0])[:,np.newaxis]
debugFn (itr, means, "means", W, K, topicMean, sigT, vocab, vocabPrior, dtype, means, varcs, A, docLens)
if logFrequency > 0 and itr % logFrequency == 0:
modelState = ModelState(K, topicMean, sigT, vocab, vocabPrior, A, dtype, MODEL_NAME)
queryState = QueryState(means, expMeans, varcs, docLens)
boundValues.append(var_bound(data, modelState, queryState))
likelyValues.append(log_likelihood(data, modelState, queryState))
boundIters.append(itr)
print (time.strftime('%X') + " : Iteration %d: bound %f \t Perplexity: %.2f" % (itr, boundValues[-1], perplexity_from_like(likelyValues[-1], docLens.sum())))
if len(boundValues) > 1:
if boundValues[-2] > boundValues[-1]:
if debug: printStderr ("ERROR: bound degradation: %f > %f" % (boundValues[-2], boundValues[-1]))
# Check to see if the improvement in the bound has fallen below the threshold
if itr > 100 and len(likelyValues) > 3 \
and abs(perplexity_from_like(likelyValues[-1], docLens.sum()) - perplexity_from_like(likelyValues[-2], docLens.sum())) < 1.0:
break
return \
ModelState(K, topicMean, sigT, vocab, vocabPrior, A, dtype, MODEL_NAME), \
QueryState(means, expMeans, varcs, docLens), \
(np.array(boundIters), np.array(boundValues), np.array(likelyValues))
def train(data, modelState, queryState, trainPlan):
'''
Infers the topic distributions in general, and specifically for
each individual datapoint.
Params:
W - the DxT document-term matrix
X - The DxF document-feature matrix, which is IGNORED in this case
modelState - the actual CTM model
queryState - the query results - essentially all the "local" variables
matched to the given observations
trainPlan - how to execute the training process (e.g. iterations,
log-interval etc.)
Return:
A new model object with the updated model (note parameters are
updated in place, so make a defensive copy if you want itr)
A new query object with the update query parameters
'''
W, X = data.words, data.feats
D, T = W.shape
F = X.shape[1]
# tmpNumDense = np.array([
# 4 , 8 , 2 , 0 , 0,
# 0 , 6 , 0 , 17, 0,
# 12 , 13 , 1 , 7 , 8,
# 0 , 5 , 0 , 0 , 0,
# 0 , 6 , 0 , 0 , 44,
# 0 , 7 , 2 , 0 , 0], dtype=np.float64).reshape((6,5))
# tmpNum = ssp.csr_matrix(tmpNumDense)
#
# tmpDenomleft = (rd.random((tmpNum.shape[0], 12)) * 5).astype(np.int32).astype(np.float64) / 10
# tmpDenomRight = (rd.random((12, tmpNum.shape[1])) * 5).astype(np.int32).astype(np.float64)
#
# tmpResult = tmpNum.copy()
# tmpResult = sparseScalarQuotientOfDot(tmpNum, tmpDenomleft, tmpDenomRight)
#
# print (str(tmpNum.todense()))
# print (str(tmpDenomleft.dot(tmpDenomRight)))
# print (str(tmpResult.todense()))
# Unpack the the structs, for ease of access and efficiency
iterations, epsilon, logFrequency, diagonalPriorCov, debug = trainPlan.iterations, trainPlan.epsilon, trainPlan.logFrequency, trainPlan.fastButInaccurate, trainPlan.debug
means, docLens = queryState.means, queryState.docLens
K, A, U, Y, V, covA, tv, ltv, fv, lfv, vocab, vocabPrior, dtype = \
modelState.K, modelState.A, modelState.U, modelState.Y, modelState.V, modelState.covA, modelState.tv, modelState.ltv, modelState.fv, modelState.lfv, modelState.vocab, modelState.vocabPrior, modelState.dtype
tp, fp, ltp, lfp = 1. / tv, 1. / fv, 1. / ltv, 1. / lfv # turn variances into precisions
# FIXME Use passed in hypers
print("tp = %f tv=%f" % (tp, tv))
vocabPrior = np.ones(shape=(T, ), dtype=modelState.dtype)
# FIXME undo truncation
F = 363
A = A[:F, :]
X = X[:, :F]
U = U[:F, :]
data = DataSet(words=W, feats=X)
# Book-keeping for logs
boundIters, boundValues, likelyValues = [], [], []
debugFn = _debug_with_bound if debug else _debug_with_nothing
# Initialize some working variables
if covA is None:
precA = (fp * ssp.eye(F) +
X.T.dot(X)).todense() # As the inverse is almost always dense
covA = la.inv(precA,
overwrite_a=True) # it's faster to densify in advance
uniqLens = np.unique(docLens)
debugFn(-1, covA, "covA", W, X, means, docLens, K, A, U, Y, V, covA, tv,
ltv, fv, lfv, vocab, vocabPrior)
H = 0.5 * (np.eye(K) - np.ones((K, K), dtype=dtype) / K)
expMeans = means.copy()
expMeans = np.exp(means - means.max(axis=1)[:, np.newaxis], out=expMeans)
R = sparseScalarQuotientOfDot(W, expMeans, vocab, out=W.copy())
lhs = H.copy()
rhs = expMeans.copy()
Y_rhs = Y.copy()
# Iterate over parameters
for itr in range(iterations):
# Update U, V given A
V = try_solve_sym_pos(Y.T.dot(U.T).dot(U).dot(Y),
A.T.dot(U).dot(Y).T).T
V /= V[0, 0]
U = try_solve_sym_pos(Y.dot(V.T).dot(V).dot(Y.T),
A.dot(V).dot(Y.T).T).T
# Update Y given U, V, A
Y_rhs[:, :] = U.T.dot(A).dot(V)
Sv, Uv = la.eigh(V.T.dot(V), overwrite_a=True)
Su, Uu = la.eigh(U.T.dot(U), overwrite_a=True)
s = np.outer(Sv, Su).flatten()
s += ltv * lfv
np.reciprocal(s, out=s)
M = Uu.T.dot(Y_rhs).dot(Uv)
M *= unvec(s, row_count=M.shape[0])
Y = Uu.dot(M).dot(Uv.T)
debugFn(itr, Y, "Y", W, X, means, docLens, K, A, U, Y, V, covA, tv,
ltv, fv, lfv, vocab, vocabPrior)
A = covA.dot(fp * U.dot(Y).dot(V.T) + X.T.dot(means))
debugFn(itr, A, "A", W, X, means, docLens, K, A, U, Y, V, covA, tv,
ltv, fv, lfv, vocab, vocabPrior)
# And now this is the E-Step, though itr's followed by updates for the
# parameters also that handle the log-sum-exp approximation.
# TODO One big sort by size, plus batch it.
# Update the Means
rhs[:, :] = expMeans
rhs *= R.dot(vocab.T)
rhs += X.dot(A) * tp
rhs += docLens[:, np.newaxis] * means.dot(H)
rhs -= docLens[:, np.newaxis] * rowwise_softmax(means, out=means)
for l in uniqLens:
inds = np.where(docLens == l)[0]
lhs[:, :] = l * H
lhs[np.diag_indices_from(lhs)] += tp
lhs[:, :] = la.inv(lhs)
means[inds, :] = rhs[inds, :].dot(
lhs
) # left and right got switched going from vectors to matrices :-/
debugFn(itr, means, "means", W, X, means, docLens, K, A, U, Y, V, covA,
tv, ltv, fv, lfv, vocab, vocabPrior)
# Standard deviation
# DK = means.shape[0] * means.shape[1]
# newTp = np.sum(means)
# newTp = (-newTp * newTp)
# rhs[:,:] = means
# rhs *= means
# newTp = DK * np.sum(rhs) - newTp
# newTp /= DK * (DK - 1)
# newTp = min(max(newTp, 1E-36), 1E+36)
# tp = 1 / newTp
# if itr % logFrequency == 0:
# print ("Iter %3d stdev = %f, prec = %f, np.std^2=%f, np.mean=%f" % (itr, sqrt(newTp), tp, np.std(means.reshape((D*K,))) ** 2, np.mean(means.reshape((D*K,)))))
# Update the vocabulary
expMeans = np.exp(means - means.max(axis=1)[:, np.newaxis],
out=expMeans)
R = sparseScalarQuotientOfDot(W, expMeans, vocab, out=R)
vocab *= (
R.T.dot(expMeans)
).T # Awkward order to maintain sparsity (R is sparse, expMeans is dense)
vocab += vocabPrior
vocab = normalizerows_ip(vocab)
debugFn(itr, vocab, "vocab", W, X, means, docLens, K, A, U, Y, V, covA,
tv, ltv, fv, lfv, vocab, vocabPrior)
# print ("Iter %3d Vocab.min = %f" % (itr, vocab.min()))
# Update the vocab prior
# vocabPrior = estimate_dirichlet_param (vocab, vocabPrior)
# print ("Iter %3d VocabPrior.(min, max) = (%f, %f) VocabPrior.mean=%f" % (itr, vocabPrior.min(), vocabPrior.max(), vocabPrior.mean()))
if logFrequency > 0 and itr % logFrequency == 0:
modelState = ModelState(K, A, U, Y, V, covA, tv, ltv, fv, lfv,
vocab, vocabPrior, dtype, modelState.name)
queryState = QueryState(means, docLens)
boundValues.append(var_bound(data, modelState, queryState))
likelyValues.append(log_likelihood(data, modelState, queryState))
boundIters.append(itr)
print(
time.strftime('%X') +
" : Iteration %d: bound %f \t Perplexity: %.2f" %
(itr, boundValues[-1],
perplexity_from_like(likelyValues[-1], docLens.sum())))
if len(boundValues) > 1:
if boundValues[-2] > boundValues[-1]:
if debug:
printStderr("ERROR: bound degradation: %f > %f" %
(boundValues[-2], boundValues[-1]))
# Check to see if the improvement in the bound has fallen below the threshold
if itr > 100 and len(likelyValues) > 3 \
and abs(perplexity_from_like(likelyValues[-1], docLens.sum()) - perplexity_from_like(likelyValues[-2], docLens.sum())) < 1.0:
break
return \
ModelState(K, A, U, Y, V, covA, tv, ltv, fv, lfv, vocab, vocabPrior, dtype, modelState.name), \
QueryState(means, expMeans, docLens), \
(np.array(boundIters), np.array(boundValues), np.array(likelyValues))
def train(data, modelState, queryState, trainPlan):
'''
Infers the topic distributions in general, and specifically for
each individual datapoint.
Params:
data - the dataset of words, features and links of which only words and
features are used in this model
modelState - the actual CTM model
queryState - the query results - essentially all the "local" variables
matched to the given observations
trainPlan - how to execute the training process (e.g. iterations,
log-interval etc.)
Return:
A new model object with the updated model (note parameters are
updated in place, so make a defensive copy if you want itr)
A new query object with the update query parameters
'''
W, X = data.words, data.feats
D, _ = W.shape
# Unpack the the structs, for ease of access and efficiency
iterations, epsilon, logFrequency, fastButInaccurate, debug = trainPlan.iterations, trainPlan.epsilon, trainPlan.logFrequency, trainPlan.fastButInaccurate, trainPlan.debug
means, expMeans, varcs, docLens = queryState.means, queryState.expMeans, queryState.varcs, queryState.docLens
F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT, vocab, vocabPrior, Ab, dtype = modelState.F, modelState.P, modelState.K, modelState.A, modelState.R_A, modelState.fv, modelState.Y, modelState.R_Y, modelState.lfv, modelState.V, modelState.sigT, modelState.vocab, modelState.vocabPrior, modelState.Ab, modelState.dtype
# Book-keeping for logs
boundIters, boundValues, boundLikes = [], [], []
debugFn = _debug_with_bound if debug else _debug_with_nothing
_debug_with_bound.old_bound = 0
# For efficient inference, we need a separate covariance for every unique
# document length. For products to execute quickly, the doc-term matrix
# therefore needs to be ordered in ascending terms of document length
originalDocLens = docLens
sortIdx = np.argsort(docLens, kind=STABLE_SORT_ALG
) # sort needs to be stable in order to be reversible
W = W[sortIdx, :] # deep sorted copy
X = X[sortIdx, :]
means, varcs = means[sortIdx, :], varcs[sortIdx, :]
docLens = originalDocLens[sortIdx]
lens, inds = np.unique(docLens, return_index=True)
inds = np.append(inds, [W.shape[0]])
# Initialize some working variables
R = W.copy()
aI_P = 1. / lfv * ssp.eye(P, dtype=dtype)
print("Creating posterior covariance of A, this will take some time...")
XTX = X.T.dot(X)
R_A = XTX
leastSquares = lambda feats, targets: la.lstsq(
feats, targets, lapack_driver="gelsy")[0].T
if ssp.issparse(
R_A): # dense inverse typically as fast or faster than sparse
R_A = to_dense_array(
R_A) # inverse and the result is usually dense in any case
leastSquares = lambda feats, targets: np.array(
[ssp.linalg.lsqr(feats, targets[:, k])[0] for k in range(K)])
R_A.flat[::F + 1] += 1. / fv
R_A = la.inv(R_A)
print("Covariance matrix calculated, launching inference")
priorSigt_diag = np.ndarray(shape=(K, ), dtype=dtype)
priorSigt_diag.fill(0.001)
# Iterate over parameters
for itr in range(iterations):
A = leastSquares(X, means)
diff_a_yv = (A - Y.dot(V))
for _ in range(10): #(50 if itr == 0 else 1):
# Update the covariance of the prior
diff_m_xa = (means - X.dot(A.T))
sigT = 1. / lfv * (Y.dot(Y.T))
sigT += 1. / fv * diff_a_yv.dot(diff_a_yv.T)
sigT += diff_m_xa.T.dot(diff_m_xa)
sigT.flat[::K + 1] += varcs.sum(axis=0)
# As small numbers lead to instable inverse estimates, we use the
# fact that for a scalar a, (a .* X)^-1 = 1/a * X^-1 and use these
# scales whenever we use the inverse of the unscaled covariance
sigScale = 1. / (P + D + F)
isigScale = 1. / sigScale
isigT = la.inv(sigT)
debugFn(itr, sigT, "sigT", W, X, XTX, F, P, K, A, R_A, fv, Y, R_Y,
lfv, V, sigT, vocab, vocabPrior, dtype, means, varcs, Ab,
docLens)
# Update the vocabulary
vocab *= (
R.T.dot(expMeans)
).T # Awkward order to maintain sparsity (R is sparse, expMeans is dense)
vocab += vocabPrior
vocab = normalizerows_ip(vocab)
# Reset the means to their original form, and log effect of vocab update
R = sparseScalarQuotientOfDot(W, expMeans, vocab, out=R)
S = expMeans * R.dot(vocab.T)
debugFn(itr, vocab, "vocab", W, X, XTX, F, P, K, A, R_A, fv, Y,
R_Y, lfv, V, sigT, vocab, vocabPrior, dtype, means, varcs,
Ab, docLens)
# Update the Variances
varcs = 1. / ((docLens * (K - 1.) / K)[:, np.newaxis] +
isigScale * isigT.flat[::K + 1])
debugFn(itr, varcs, "varcs", W, X, XTX, F, P, K, A, R_A, fv, Y,
R_Y, lfv, V, sigT, vocab, vocabPrior, dtype, means, varcs,
Ab, docLens)
# Update the Means
rhs = X.dot(A.T).dot(isigT) * isigScale
rhs += S
rhs += docLens[:, np.newaxis] * means.dot(Ab)
rhs -= docLens[:, np.newaxis] * rowwise_softmax(means, out=means)
# Faster version?
for lenIdx in range(len(lens)):
nd = lens[lenIdx]
start, end = inds[lenIdx], inds[lenIdx + 1]
lhs = la.inv(isigT + sigScale * nd * Ab) * sigScale
means[start:end, :] = rhs[start:end, :].dot(
lhs
) # huh?! Left and right refer to eqn for a single mean: once we're talking a DxK matrix it gets swapped
# print("Vec-Means: %f, %f, %f, %f" % (means.min(), means.mean(), means.std(), means.max()))
debugFn(itr, means, "means", W, X, XTX, F, P, K, A, R_A, fv, Y,
R_Y, lfv, V, sigT, vocab, vocabPrior, dtype, means, varcs,
Ab, docLens)
expMeans = np.exp(means - means.max(axis=1)[:, np.newaxis],
out=expMeans)
# for _ in range(150):
# # Finally update the parameter V
# V = la.inv(sigScale * R_Y + Y.T.dot(isigT).dot(Y)).dot(Y.T.dot(isigT).dot(A))
# debugFn(itr, V, "V", W, X, XTX, F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT, vocab, vocabPrior, dtype, means,
# varcs, Ab, docLens)
#
# # Update the distribution on the latent space
# R_Y_base = aI_P + 1 / fv * V.dot(V.T)
# R_Y = la.inv(R_Y_base)
# debugFn(itr, R_Y, "R_Y", W, X, XTX, F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT, vocab, vocabPrior, dtype,
# means, varcs, Ab, docLens)
#
# Y = 1. / fv * A.dot(V.T).dot(R_Y)
# debugFn(itr, Y, "Y", W, X, XTX, F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT, vocab, vocabPrior, dtype, means,
# varcs, Ab, docLens)
#
# # Update the mapping from the features to topics
# A = (1. / fv * Y.dot(V) + (X.T.dot(means)).T).dot(R_A)
# debugFn(itr, A, "A", W, X, XTX, F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT, vocab, vocabPrior, dtype, means,
# varcs, Ab, docLens)
if logFrequency > 0 and itr % logFrequency == 0:
modelState = ModelState(F, P, K, A, R_A, fv, Y, R_Y, lfv, V,
sigT * sigScale, vocab, vocabPrior, Ab,
dtype, MODEL_NAME)
queryState = QueryState(means, expMeans, varcs, docLens)
boundValues.append(
var_bound(DataSet(W, feats=X), modelState, queryState, XTX))
boundLikes.append(
log_likelihood(DataSet(W, feats=X), modelState, queryState))
boundIters.append(itr)
perp = perplexity_from_like(boundLikes[-1], docLens.sum())
print(
time.strftime('%X') +
" : Iteration %d: Perplexity %4.0f bound %f" %
(itr, perp, boundValues[-1]))
if len(boundIters) >= 2 and boundValues[-2] > boundValues[-1]:
printStderr("ERROR: bound degradation: %f > %f" %
(boundValues[-2], boundValues[-1]))
# print ("Means: min=%f, avg=%f, max=%f\n\n" % (means.min(), means.mean(), means.max()))
# Check to see if the improvement in the likelihood has fallen below the threshold
if len(boundIters) > 2 and boundIters[-1] > 20:
lastPerp = perplexity_from_like(boundLikes[-2], docLens.sum())
if lastPerp - perp < 1:
break
revert_sort = np.argsort(sortIdx, kind=STABLE_SORT_ALG)
means = means[revert_sort, :]
varcs = varcs[revert_sort, :]
docLens = docLens[revert_sort]
return \
ModelState(F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT * sigScale, vocab, vocabPrior, Ab, dtype, MODEL_NAME), \
QueryState(means, expMeans, varcs, docLens), \
(boundIters, boundValues, boundLikes)
def train (data, modelState, queryState, trainPlan):
'''
Infers the topic distributions in general, and specifically for
each individual datapoint.
Params:
data - the dataset of words, features and links of which only words and
features are used in this model
modelState - the actual CTM model
queryState - the query results - essentially all the "local" variables
matched to the given observations
trainPlan - how to execute the training process (e.g. iterations,
log-interval etc.)
Return:
A new model object with the updated model (note parameters are
updated in place, so make a defensive copy if you want itr)
A new query object with the update query parameters
'''
W, X = data.words, data.feats
D, _ = W.shape
# Unpack the the structs, for ease of access and efficiency
iterations, epsilon, logFrequency, fastButInaccurate, debug = trainPlan.iterations, trainPlan.epsilon, trainPlan.logFrequency, trainPlan.fastButInaccurate, trainPlan.debug
means, expMeans, varcs, docLens = queryState.means, queryState.expMeans, queryState.varcs, queryState.docLens
F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT, vocab, vocabPrior, Ab, dtype = modelState.F, modelState.P, modelState.K, modelState.A, modelState.R_A, modelState.fv, modelState.Y, modelState.R_Y, modelState.lfv, modelState.V, modelState.sigT, modelState.vocab, modelState.vocabPrior, modelState.Ab, modelState.dtype
# Book-keeping for logs
boundIters = np.zeros(shape=(iterations // logFrequency,))
boundValues = np.zeros(shape=(iterations // logFrequency,))
boundLikes = np.zeros(shape=(iterations // logFrequency,))
bvIdx = 0
debugFn = _debug_with_bound if debug else _debug_with_nothing
_debug_with_bound.old_bound = 0
# For efficient inference, we need a separate covariance for every unique
# document length. For products to execute quickly, the doc-term matrix
# therefore needs to be ordered in ascending terms of document length
originalDocLens = docLens
sortIdx = np.argsort(docLens, kind=STABLE_SORT_ALG) # sort needs to be stable in order to be reversible
W = W[sortIdx,:] # deep sorted copy
X = X[sortIdx,:]
means, varcs = means[sortIdx,:], varcs[sortIdx,:]
docLens = originalDocLens[sortIdx]
lens, inds = np.unique(docLens, return_index=True)
inds = np.append(inds, [W.shape[0]])
# Initialize some working variables
R = W.copy()
aI_P = 1./lfv * ssp.eye(P, dtype=dtype)
print("Creating posterior covariance of A, this will take some time...")
XTX = X.T.dot(X)
R_A = XTX
R_A = R_A.todense() # dense inverse typically as fast or faster than sparse inverse
R_A.flat[::F+1] += 1./fv # and the result is usually dense in any case
R_A = la.inv(R_A)
print("Covariance matrix calculated, launching inference")
diff_m_xa = (means-X.dot(A.T))
means_cov_with_x_a = diff_m_xa.T.dot(diff_m_xa)
expMeans = np.zeros((BatchSize, K), dtype=dtype)
R = np.zeros((BatchSize, K), dtype=dtype)
S = np.zeros((BatchSize, K), dtype=dtype)
vocabScale = np.ones(vocab.shape, dtype=dtype)
# Iterate over parameters
batchIter = 0
for itr in range(iterations):
# We start with the M-Step, so the parameters are consistent with our
# initialisation of the RVs when we do the E-Step
# Update the covariance of the prior
diff_a_yv = (A-Y.dot(V))
sigT = 1./lfv * (Y.dot(Y.T))
sigT += 1./fv * diff_a_yv.dot(diff_a_yv.T)
sigT += means_cov_with_x_a
sigT.flat[::K+1] += varcs.sum(axis=0)
# As small numbers lead to instable inverse estimates, we use the
# fact that for a scalar a, (a .* X)^-1 = 1/a * X^-1 and use these
# scales whenever we use the inverse of the unscaled covariance
sigScale = 1. / (P+D+F)
isigScale = 1. / sigScale
isigT = la.inv(sigT)
debugFn (itr, sigT, "sigT", W, X, XTX, F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT, vocab, vocabPrior, dtype, means, varcs, Ab, docLens)
# Update the vocabulary
vocab *= vocabScale
vocab += vocabPrior
vocab = normalizerows_ip(vocab)
debugFn (itr, vocab, "vocab", W, X, XTX, F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT, vocab, vocabPrior, dtype, means, varcs, Ab, docLens)
# Finally update the parameter V
V = la.inv(sigScale * R_Y + Y.T.dot(isigT).dot(Y)).dot(Y.T.dot(isigT).dot(A))
debugFn (itr, V, "V", W, X, XTX, F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT, vocab, vocabPrior, dtype, means, varcs, Ab, docLens)
#
# And now this is the E-Step
#
# Update the distribution on the latent space
R_Y_base = aI_P + 1/fv * V.dot(V.T)
R_Y = la.inv(R_Y_base)
debugFn (itr, R_Y, "R_Y", W, X, XTX, F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT, vocab, vocabPrior, dtype, means, varcs, Ab, docLens)
Y = 1./fv * A.dot(V.T).dot(R_Y)
debugFn (itr, Y, "Y", W, X, XTX, F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT, vocab, vocabPrior, dtype, means, varcs, Ab, docLens)
# Update the mapping from the features to topics
A = (1./fv * Y.dot(V) + (X.T.dot(means)).T).dot(R_A)
debugFn (itr, A, "A", W, X, XTX, F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT, vocab, vocabPrior, dtype, means, varcs, Ab, docLens)
# Update the Variances
varcs = 1./((docLens * (K-1.)/K)[:,np.newaxis] + isigScale * isigT.flat[::K+1])
debugFn (itr, varcs, "varcs", W, X, XTX, F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT, vocab, vocabPrior, dtype, means, varcs, Ab, docLens)
# Faster version?
vocabScale[:,:] = 0
means_cov_with_x_a[:,:] = 0
for lenIdx in range(len(lens)):
nd = lens[lenIdx]
start, end = inds[lenIdx], inds[lenIdx + 1]
lhs = la.inv(isigT + sigScale * nd * Ab) * sigScale
for d in range(start, end, BatchSize):
end_d = min(d + BatchSize, end)
span = end_d - d
expMeans[:span,:] = np.exp(means[d:end_d,:] - means[d:end_d,:].max(axis=1)[:span,np.newaxis], out=expMeans[:span,:])
R = sparseScalarQuotientOfDot(W[d:end_d,:], expMeans[d:end_d,:], vocab)
S[:span,:] = expMeans[:span, :] * R.dot(vocab.T)
# Convert expMeans to a softmax(means)
expMeans[:span,:] /= expMeans[:span,:].sum(axis=1)[:span,np.newaxis]
mu = X[d:end_d,:].dot(A.T)
rhs = mu.dot(isigT) * isigScale
rhs += S[:span,:]
rhs += docLens[d:end_d,np.newaxis] * means[d:end_d,:].dot(Ab)
rhs -= docLens[d:end_d,np.newaxis] * expMeans[:span,:] # here expMeans is actually softmax(means)
means[d:end_d,:] = rhs.dot(lhs) # huh?! Left and right refer to eqn for a single mean: once we're talking a DxK matrix it gets swapped
expMeans[:span,:] = np.exp(means[d:end_d,:] - means[d:end_d,:].max(axis=1)[:span,np.newaxis], out=expMeans[:span,:])
R = sparseScalarQuotientOfDot(W[d:end_d,:], expMeans[:span,:], vocab, out=R)
stepSize = (Tau + batchIter) ** -Kappa
batchIter += 1
# Do a gradient update of the vocab
vocabScale += (R.T.dot(expMeans[:span,:])).T
# vocabScale *= vocab
# normalizerows_ip(vocabScale)
# # vocabScale += vocabPrior
# vocabScale *= stepSize
# vocab *= (1 - stepSize)
# vocab += vocabScale
diff = (means[d:end_d,:] - mu)
means_cov_with_x_a += diff.T.dot(diff)
# print("Vec-Means: %f, %f, %f, %f" % (means.min(), means.mean(), means.std(), means.max()))
debugFn (itr, means, "means", W, X, XTX, F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT, vocab, vocabPrior, dtype, means, varcs, Ab, docLens)
if logFrequency > 0 and itr % logFrequency == 0:
modelState = ModelState(F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT * sigScale, vocab, vocabPrior, Ab, dtype, MODEL_NAME)
queryState = QueryState(means, expMeans, varcs, docLens)
boundValues[bvIdx] = var_bound(DataSet(W, feats=X), modelState, queryState, XTX)
boundLikes[bvIdx] = log_likelihood(DataSet(W, feats=X), modelState, queryState)
boundIters[bvIdx] = itr
perp = perplexity_from_like(boundLikes[bvIdx], docLens.sum())
print (time.strftime('%X') + " : Iteration %d: Perplexity %4.0f bound %f" % (itr, perp, boundValues[bvIdx]))
if bvIdx > 0 and boundValues[bvIdx - 1] > boundValues[bvIdx]:
printStderr ("ERROR: bound degradation: %f > %f" % (boundValues[bvIdx - 1], boundValues[bvIdx]))
# print ("Means: min=%f, avg=%f, max=%f\n\n" % (means.min(), means.mean(), means.max()))
# Check to see if the improvement in the likelihood has fallen below the threshold
if bvIdx > 1 and boundIters[bvIdx] > 20:
lastPerp = perplexity_from_like(boundLikes[bvIdx - 1], docLens.sum())
if lastPerp - perp < 1:
boundIters, boundValues, likelyValues = clamp (boundIters, boundValues, boundLikes, bvIdx)
break
bvIdx += 1
revert_sort = np.argsort(sortIdx, kind=STABLE_SORT_ALG)
means = means[revert_sort,:]
varcs = varcs[revert_sort,:]
docLens = docLens[revert_sort]
return \
ModelState(F, P, K, A, R_A, fv, Y, R_Y, lfv, V, sigT * sigScale, vocab, vocabPrior, Ab, dtype, MODEL_NAME), \
QueryState(means, expMeans, varcs, docLens), \
(boundIters, boundValues, boundLikes)
def train(data, modelState, queryState, trainPlan):
'''
Infers the topic distributions in general, and specifically for
each individual datapoint.
Params:
W - the DxT document-term matrix
X - The DxF document-feature matrix, which is IGNORED in this case
modelState - the actual CTM model
queryState - the query results - essentially all the "local" variables
matched to the given observations
trainPlan - how to execute the training process (e.g. iterations,
log-interval etc.)
Return:
A new model object with the updated model (note parameters are
updated in place, so make a defensive copy if you want itr)
A new query object with the update query parameters
'''
W, L, LT, X = data.words, data.links, ssp.csr_matrix(
data.links.T), data.feats
D, _ = W.shape
out_links = np.squeeze(np.asarray(data.links.sum(axis=1)))
# Unpack the the structs, for ease of access and efficiency
iterations, epsilon, logFrequency, diagonalPriorCov, debug = trainPlan.iterations, trainPlan.epsilon, trainPlan.logFrequency, trainPlan.fastButInaccurate, trainPlan.debug
means, varcs, docLens = queryState.means, queryState.varcs, queryState.docLens
K, topicMean, topicCov, vocab, A, dtype = modelState.K, modelState.topicMean, modelState.topicCov, modelState.vocab, modelState.A, modelState.dtype
emit_counts = docLens + out_links
# Book-keeping for logs
boundIters, boundValues, likelyValues = [], [], []
if debug:
debugFn = _debug_with_bound
initLikely = log_likelihood(data, modelState, queryState)
initPerp = perplexity_from_like(initLikely, data.word_count)
print("Initial perplexity is: %.2f" % initPerp)
else:
debugFn = _debug_with_nothing
# Initialize some working variables
W_weight = W.copy()
L_weight = L.copy()
LT_weight = LT.copy()
pseudoObsMeans = K + NIW_PSEUDO_OBS_MEAN
pseudoObsVar = K + NIW_PSEUDO_OBS_VAR
priorSigT_diag = np.ndarray(shape=(K, ), dtype=dtype)
priorSigT_diag.fill(NIW_PSI)
# Iterate over parameters
for itr in range(iterations):
# We start with the M-Step, so the parameters are consistent with our
# initialisation of the RVs when we do the E-Step
# Update the mean and covariance of the prior
topicMean = means.sum(axis = 0) / (D + pseudoObsMeans) \
if USE_NIW_PRIOR \
else means.mean(axis=0)
debugFn(itr, topicMean, "topicMean", data, K, topicMean, topicCov,
vocab, dtype, means, varcs, A, docLens)
if USE_NIW_PRIOR:
diff = means - topicMean[np.newaxis, :]
topicCov = diff.T.dot(diff) \
+ pseudoObsVar * np.outer(topicMean, topicMean)
topicCov += np.diag(varcs.mean(axis=0) + priorSigT_diag)
topicCov /= (D + pseudoObsVar - K)
else:
topicCov = np.cov(
means.T) if topicCov.dtype == np.float64 else np.cov(
means.T).astype(dtype)
topicCov += np.diag(varcs.mean(axis=0))
if diagonalPriorCov:
diag = np.diag(topicCov)
topicCov = np.diag(diag)
itopicCov = np.diag(1. / diag)
else:
itopicCov = la.inv(topicCov)
debugFn(itr, topicCov, "topicCov", data, K, topicMean, topicCov, vocab,
dtype, means, varcs, A, docLens)
# print(" topicCov.det = " + str(la.det(topicCov)))
# Building Blocks - temporarily replaces means with exp(means)
expMeansCol = np.exp(means - means.max(axis=0)[np.newaxis, :])
lse_at_k = np.sum(expMeansCol, axis=0)
F = 0.5 * means \
- (1. / (2*D + 2)) * means.sum(axis=0) \
- expMeansCol / lse_at_k[np.newaxis, :]
expMeansRow = np.exp(means - means.max(axis=1)[:, np.newaxis])
W_weight = sparseScalarQuotientOfDot(W,
expMeansRow,
vocab,
out=W_weight)
# Update the vocabularies
vocab *= (
W_weight.T.dot(expMeansRow)
).T # Awkward order to maintain sparsity (R is sparse, expMeans is dense)
vocab += VocabPrior
vocab = normalizerows_ip(vocab)
docVocab = (
expMeansCol /
lse_at_k[np.newaxis, :]).T # FIXME Dupes line in definitino of F
# Recalculate w_top_sums with the new vocab and log vocab improvement
W_weight = sparseScalarQuotientOfDot(W,
expMeansRow,
vocab,
out=W_weight)
w_top_sums = W_weight.dot(vocab.T) * expMeansRow
debugFn(itr, vocab, "vocab", data, K, topicMean, topicCov, vocab,
dtype, means, varcs, A, docLens)
# Now do likewise for the links, do it twice to model in-counts (first) and
# out-counts (Second). The difference is the transpose
LT_weight = sparseScalarQuotientOfDot(LT,
expMeansRow,
docVocab,
out=LT_weight)
l_intop_sums = LT_weight.dot(docVocab.T) * expMeansRow
in_counts = l_intop_sums.sum(axis=0)
L_weight = sparseScalarQuotientOfDot(L,
expMeansRow,
docVocab,
out=L_weight)
l_outtop_sums = L_weight.dot(docVocab.T) * expMeansRow
# Reset the means and use them to calculate the weighted sum of means
meanSum = means.sum(axis=0) * in_counts
# And now this is the E-Step, though itr's followed by updates for the
# parameters also that handle the log-sum-exp approximation.
# Update the Variances: var_d = (2 N_d * A + itopicCov)^{-1}
varcs = np.reciprocal(docLens[:, np.newaxis] * (0.5 - 1. / K) +
np.diagonal(topicCov))
debugFn(itr, varcs, "varcs", data, K, topicMean, topicCov, vocab,
dtype, means, varcs, A, docLens)
# Update the Means
rhs = w_top_sums.copy()
rhs += l_intop_sums
rhs += l_outtop_sums
rhs += itopicCov.dot(topicMean)
rhs += emit_counts[:, np.newaxis] * (means.dot(A) -
rowwise_softmax(means))
rhs += in_counts[np.newaxis, :] * F
if diagonalPriorCov:
raise ValueError("Not implemented")
else:
for d in range(D):
rhs_ = rhs[d, :] + (1. /
(4 * D + 4)) * (meanSum -
in_counts * means[d, :])
means[d, :] = la.inv(itopicCov + emit_counts[d] * A +
np.diag(D * in_counts /
(2 * D + 2))).dot(rhs_)
if np.any(np.isnan(means[d, :])) or np.any(
np.isinf(means[d, :])):
pass
if np.any(np.isnan(
np.exp(means[d, :] - means[d, :].max()))) or np.any(
np.isinf(np.exp(means[d, :] - means[d, :].max()))):
pass
debugFn(itr, means, "means", data, K, topicMean, topicCov, vocab,
dtype, means, varcs, A, docLens)
if logFrequency > 0 and itr % logFrequency == 0:
modelState = ModelState(K, topicMean, topicCov, vocab, A, dtype,
MODEL_NAME)
queryState = QueryState(means, varcs, docLens)
boundValues.append(var_bound(data, modelState, queryState))
likelyValues.append(log_likelihood(data, modelState, queryState))
boundIters.append(itr)
print(
time.strftime('%X') +
" : Iteration %d: bound %f \t Perplexity: %.2f" %
(itr, boundValues[-1],
perplexity_from_like(likelyValues[-1], docLens.sum())))
if len(boundValues) > 1:
if boundValues[-2] > boundValues[-1]:
printStderr("ERROR: bound degradation: %f > %f" %
(boundValues[-2], boundValues[-1]))
# Check to see if the improvement in the bound has fallen below the threshold
if False and itr > 100 and abs(
perplexity_from_like(likelyValues[-1], docLens.sum()) -
perplexity_from_like(likelyValues[-2], docLens.sum())
) < 1.0:
break
return \
ModelState(K, topicMean, topicCov, vocab, A, dtype, MODEL_NAME), \
QueryState(means, varcs, docLens), \
(np.array(boundIters), np.array(boundValues), np.array(likelyValues))