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, 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, 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 testPerplexityOnRealData(self):
dtype = np.float64 # DTYPE
rd.seed(0xBADB055)
data = DataSet.from_files(words_file=AclWordPath, links_file=AclCitePath)
with open(AclDictPath, "rb") as f:
d = pkl.load(f)
data.convert_to_dtype(dtype)
data.prune_and_shuffle(min_doc_len=MinDocLen, min_link_count=MinLinkCount)
# IDF frequency for when we print out the vocab later
freq = np.squeeze(np.asarray(data.words.sum(axis=0)))
scale = np.reciprocal(1 + freq)
# Initialise the model
K = 50
model = mtm.newModelAtRandom(data, K, K - 1, dtype=dtype)
queryState = mtm.newQueryState(data, model)
trainPlan = mtm.newTrainPlan(iterations=200, logFrequency=10, fastButInaccurate=False, debug=True)
# Train the model, and the immediately save the result to a file for subsequent inspection
model, query, (bndItrs, bndVals, bndLikes) = mtm.train (data, model, queryState, trainPlan)
# with open(newModelFileFromModel(model), "wb") as f:
# pkl.dump ((model, query, (bndItrs, bndVals, bndLikes)), f)
# Plot the evolution of the bound during training.
fig, ax1 = plt.subplots()
ax1.plot(bndItrs, bndVals, 'b-')
ax1.set_xlabel('Iterations')
ax1.set_ylabel('Bound', color='b')
ax2 = ax1.twinx()
ax2.plot(bndItrs, bndLikes, 'r-')
ax2.set_ylabel('Likelihood', color='r')
fig.show()
plt.show()
fig, ax1 = plt.subplots()
ax1.imshow(model.topicCov, interpolation="nearest", cmap=cm.Greys_r)
fig.show()
plt.show()
# Print out the most likely topic words
# scale = np.reciprocal(1 + np.squeeze(np.array(data.words.sum(axis=0))))
vocab = mtm.wordDists(model)
topWordCount = 10
kTopWordInds = [self.topWordInds(vocab[k,:], topWordCount) for k in range(K)]
like = mtm.log_likelihood(data, model, query)
perp = perplexity_from_like(like, data.word_count)
print ("Prior %s" % (str(model.topicPrior)))
print ("Perplexity: %f\n\n" % perp)
for k in range(model.K):
print("\nTopic %d\n=============================" % k)
print("\n".join("%-20s\t%0.4f" % (d[kTopWordInds[k][c]], vocab[k][kTopWordInds[k][c]]) for c in range(topWordCount)))
def testPerplexityOnRealDataWithLdaInc(self):
dtype = np.float64 # DTYPE
rd.seed(0xBADB055)
data = DataSet.from_files(words_file=AclWordPath, links_file=AclCitePath)
with open(AclDictPath, "rb") as f:
d = pkl.load(f)
data.convert_to_dtype(dtype)
data.prune_and_shuffle(min_doc_len=MinDocLen, min_link_count=MinLinkCount)
# IDF frequency for when we print out the vocab later
freq = np.squeeze(np.asarray(data.words.sum(axis=0)))
scale = np.reciprocal(1 + freq)
# Initialise the model
topicCounts = [5, 10, 15, 20, 25, 30, 35, 40, 45, 50]
perps = []
for K in topicCounts:
model = lda.newModelAtRandom(data, K, dtype=dtype)
queryState = lda.newQueryState(data, model)
trainPlan = lda.newTrainPlan(iterations=800, logFrequency=10, fastButInaccurate=False, debug=False)
# Train the model, and the immediately save the result to a file for subsequent inspection
model, query, (bndItrs, bndVals, bndLikes) = lda.train (data, model, queryState, trainPlan)
# with open(newModelFileFromModel(model), "wb") as f:
# pkl.dump ((model, query, (bndItrs, bndVals, bndLikes)), f)
# Print out the most likely topic words
# scale = np.reciprocal(1 + np.squeeze(np.array(data.words.sum(axis=0))))
# vocab = lda.wordDists(model)
# topWordCount = 10
# kTopWordInds = [self.topWordInds(vocab[k,:], topWordCount) for k in range(K)]
like = lda.log_likelihood(data, model, query)
perp = perplexity_from_like(like, data.word_count)
perps.append(perp)
print ("K = %2d : Perplexity = %f\n\n" % (K, perp))
#
# for k in range(model.K):
# print("\nTopic %d\n=============================" % k)
# print("\n".join("%-20s\t%0.4f" % (d[kTopWordInds[k][c]], vocab[k][kTopWordInds[k][c]]) for c in range(topWordCount)))
# Plot the evolution of the bound during training.
fig, ax1 = plt.subplots()
ax1.plot(topicCounts, perps, 'b-')
ax1.set_xlabel('Topic Count')
ax1.set_ylabel('Perplexity', color='b')
fig.show()
plt.show()
def testCrossValPerplexityOnRealDataWithLdaGibbsInc(self):
ActiveFolds = 3
dtype = np.float64 # DTYPE
rd.seed(0xBADB055)
data = DataSet.from_files(words_file=AclWordPath, links_file=AclCitePath)
data.convert_to_dtype(np.int32) # Gibbs expects integers as input, regardless of model dtype
data.prune_and_shuffle(min_doc_len=MinDocLen, min_link_count=MinLinkCount)
# Training setup
TrainSamplesPerTopic = 10
QuerySamplesPerTopic = 2
Thin = 2
Debug = False
# Start running experiments
topicCounts = [5, 10, 15, 20, 25, 30, 35, 40, 45, 50]
for K in topicCounts:
trainPlan = lda_gibbs.newTrainPlan(K * TrainSamplesPerTopic, thin=Thin, debug=Debug)
queryPlan = lda_gibbs.newTrainPlan(K * QuerySamplesPerTopic, thin=Thin, debug=Debug)
trainPerps = []
queryPerps = []
for fold in range(ActiveFolds): # range(NumFolds):
trainData, queryData = data.cross_valid_split(fold, NumFolds)
estData, evalData = queryData.doc_completion_split()
model = lda_gibbs.newModelAtRandom(trainData, K, dtype=dtype)
query = lda_gibbs.newQueryState(trainData, model)
# Train the model, and the immediately save the result to a file for subsequent inspection
model, trainResult, (_, _, _) = lda_gibbs.train (trainData, model, query, trainPlan)
like = lda_gibbs.log_likelihood(trainData, model, trainResult)
perp = perplexity_from_like(like, trainData.word_count)
trainPerps.append(perp)
query = lda_gibbs.newQueryState(estData, model)
_, queryResult = lda_gibbs.query(estData, model, query, queryPlan)
like = lda_gibbs.log_likelihood(evalData, model, queryResult)
perp = perplexity_from_like(like, evalData.word_count)
queryPerps.append(perp)
trainPerps.append(sum(trainPerps) / ActiveFolds)
queryPerps.append(sum(queryPerps) / ActiveFolds)
print("K=%d,Segment=Train,%s" % (K, ",".join([str(p) for p in trainPerps])))
print("K=%d,Segment=Query,%s" % (K, ",".join([str(p) for p in queryPerps])))
def testOnRealData(self):
dtype = np.float64 # DTYPE
rd.seed(0xBADB055)
data = DataSet.from_files(words_file=TNewsWordsPath)
with open(TNewsDictPath, "rb") as f:
d = pkl.load(f)
data.convert_to_dtype(dtype)
data.prune_and_shuffle(min_doc_len=50, min_link_count=0)
# Initialise the model
K = 6
model = mom.newModelAtRandom(data, K, dtype=dtype)
queryState = mom.newQueryState(data, model)
# trainPlan = mom.newTrainPlan(iterations=1000, logFrequency=10, debug=False, burnIn=1000, thinning=10)
trainPlan = mom.newTrainPlan(iterations=200,
logFrequency=10,
debug=False)
# Train the model, and the immediately save the result to a file for subsequent inspection
model, query, (bndItrs, bndVals,
bndLikes) = mom.train(data, model, queryState,
trainPlan)
# with open(newModelFileFromModel(model), "wb") as f:
# pkl.dump ((model, query, (bndItrs, bndVals, bndLikes)), f)
# Plot the evolution of the bound during training.
fig, ax1 = plt.subplots()
ax1.plot(bndItrs, bndVals, 'b-')
ax1.set_xlabel('Iterations')
ax1.set_ylabel('Bound', color='b')
ax2 = ax1.twinx()
ax2.plot(bndItrs, bndLikes, 'r-')
ax2.set_ylabel('Likelihood', color='r')
fig.show()
plt.show()
# Print out the most likely topic words
print("Prior %s" % (str(model.topicPrior)))
print("Perplexity: %f\n\n" %
word_perplexity(mom.log_likelihood, model, query, data))
print("")
printWordDists(K, mom.wordDists(model), d)
def testCrossValPerplexityOnRealDataWithLdaInc(self):
ActiveFolds = 3
dtype = np.float64 # DTYPE
rd.seed(0xBADB055)
data = DataSet.from_files(words_file=AclWordPath, links_file=AclCitePath)
data.convert_to_dtype(dtype)
data.prune_and_shuffle(min_doc_len=MinDocLen, min_link_count=MinLinkCount)
# Initialise the model
trainPlan = lda.newTrainPlan(iterations=800, logFrequency=10, fastButInaccurate=False, debug=False)
queryPlan = lda.newTrainPlan(iterations=50, logFrequency=5, fastButInaccurate=False, debug=False)
topicCounts = [30, 35, 40, 45, 50] # [5, 10, 15, 20, 25, 30, 35, 40, 45, 50]
for K in topicCounts:
trainPerps = []
queryPerps = []
for fold in range(ActiveFolds): # range(NumFolds):
trainData, queryData = data.cross_valid_split(fold, NumFolds)
model = lda.newModelAtRandom(trainData, K, dtype=dtype)
query = lda.newQueryState(trainData, model)
# Train the model, and the immediately save the result to a file for subsequent inspection
model, trainResult, (_, _, _) = lda.train (trainData, model, query, trainPlan)
like = lda.log_likelihood(trainData, model, trainResult)
perp = perplexity_from_like(like, trainData.word_count)
trainPerps.append(perp)
estData, evalData = queryData.doc_completion_split()
query = lda.newQueryState(estData, model)
model, queryResult = lda.query(estData, model, query, queryPlan)
like = lda.log_likelihood(evalData, model, queryResult)
perp = perplexity_from_like(like, evalData.word_count)
queryPerps.append(perp)
trainPerps.append(sum(trainPerps) / ActiveFolds)
queryPerps.append(sum(queryPerps) / ActiveFolds)
print("K=%d,Segment=Train,%s" % (K, ",".join([str(p) for p in trainPerps])))
print("K=%d,Segment=Query,%s" % (K, ",".join([str(p) for p in queryPerps])))
def testPerplexityOnRealDataWithMtm2(self):
dtype = np.float64 # DTYPE
rd.seed(0xBADB055)
data = DataSet.from_files(words_file=AclWordPath, links_file=AclCitePath)
with open(AclDictPath, "rb") as f:
d = pkl.load(f)
data.convert_to_dtype(dtype)
data.prune_and_shuffle(min_doc_len=MinDocLen, min_link_count=MinLinkCount)
# IDF frequency for when we print out the vocab later
freq = np.squeeze(np.asarray(data.words.sum(axis=0)))
scale = np.reciprocal(1 + freq)
# Initialise the model
K = 30 # TopicCount
model = mtm2.newModelAtRandom(data, K, dtype=dtype)
queryState = mtm2.newQueryState(data, model)
trainPlan = mtm2.newTrainPlan(iterations=200, logFrequency=10, fastButInaccurate=False, debug=False)
# Train the model, and the immediately save the result to a file for subsequent inspection
model, query, (bndItrs, bndVals, bndLikes) = mtm2.train(data, model, queryState, trainPlan)
# with open(newModelFileFromModel(model), "wb") as f:
# pkl.dump ((model, query, (bndItrs, bndVals, bndLikes)), f)
# Plot the evolution of the bound during training.
fig, ax1 = plt.subplots()
ax1.plot(bndItrs, bndVals, 'b-')
ax1.set_xlabel('Iterations')
ax1.set_ylabel('Bound', color='b')
ax2 = ax1.twinx()
ax2.plot(bndItrs, bndLikes, 'r-')
ax2.set_ylabel('Likelihood', color='r')
fig.show()
plt.show()
fig, ax1 = plt.subplots()
ax1.imshow(model.topicCov, interpolation="nearest", cmap=cm.Greys_r)
fig.show()
plt.show()
# Print out the most likely topic words
# scale = np.reciprocal(1 + np.squeeze(np.array(data.words.sum(axis=0))))
vocab = mtm2.wordDists(model)
topWordCount = 10
kTopWordInds = [self.topWordInds(vocab[k,:], topWordCount) for k in range(K)]
like = mtm2.log_likelihood(data, model, query)
perp = perplexity_from_like(like, data.word_count)
print("Perplexity: %f\n\n" % perp)
for k in range(model.K):
print("\nTopic %d\n=============================" % k)
print("\n".join("%-20s\t%0.4f" % (d[kTopWordInds[k][c]], vocab[k][kTopWordInds[k][c]]) for c in range(topWordCount)))
print ("Most likely documents for each topic")
print ("====================================")
with open ("/Users/bryanfeeney/iCloud/Datasets/ACL/ACL.100/doc_ids.pkl", 'rb') as f:
fileIds = pkl.load (f)
docs_dict = [fileIds[fi] for fi in data.order]
for k in range(model.K):
arg_max_prob = np.argmax(query.means[:, k])
print("K=%2d Document ID = %s (found at %d)" % (k, docs_dict[arg_max_prob], arg_max_prob))
print ("Done")
with open ("/Users/bryanfeeney/Desktop/mtm2-" + str(K) + ".pkl", "wb") as f:
pkl.dump((model, query), f)
def testOnRealData(self):
dtype = np.float64 # DTYPE
rd.seed(0xBADB055)
data = DataSet.from_files(words_file=NipsWordsPath,
links_file=NipsCitePath)
with open(NipsDictPath, "rb") as f:
d = pkl.load(f)
data.convert_to_dtype(dtype)
data.prune_and_shuffle(min_doc_len=50, min_link_count=0)
# IDF frequency for when we print out the vocab later
freq = np.squeeze(np.asarray(data.words.sum(axis=0)))
scale = np.reciprocal(1 + freq)
# Initialise the model
K = 10
model = lda.newModelAtRandom(data, K, dtype=dtype)
queryState = lda.newQueryState(data, model)
trainPlan = lda.newTrainPlan(iterations=30,
logFrequency=2,
debug=False,
batchSize=50,
rate_retardation=1,
forgetting_rate=0.75)
# Train the model, and the immediately save the result to a file for subsequent inspection
model, query, (bndItrs, bndVals,
bndLikes) = lda.train(data, model, queryState,
trainPlan)
# with open(newModelFileFromModel(model), "wb") as f:
# pkl.dump ((model, query, (bndItrs, bndVals, bndLikes)), f)
# Plot the evolution of the bound during training.
fig, ax1 = plt.subplots()
ax1.plot(bndItrs, bndVals, 'b-')
ax1.set_xlabel('Iterations')
ax1.set_ylabel('Bound', color='b')
ax2 = ax1.twinx()
ax2.plot(bndItrs, bndLikes, 'r-')
ax2.set_ylabel('Likelihood', color='r')
fig.show()
plt.show()
vocab = lda.wordDists(model)
plt.imshow(vocab, interpolation="nearest", cmap=cm.Greys_r)
plt.show()
# Print out the most likely topic words
topWordCount = 100
kTopWordInds = [topWordIndices(vocab[k, :] * scale, topWordCount) \
for k in range(K)]
# Print out the most likely topic words
print("Prior %s" % (str(model.topicPrior)))
print("Perplexity: %f\n\n" %
word_perplexity(lda.log_likelihood, model, query, data))
print("")
printWordDists(K, lda.wordDists(model), d)
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 testMapOnRealData(self):
dtype = np.float64 # DTYPE
rd.seed(0xBADB055)
data = DataSet.from_files(words_file=AclWordPath,
links_file=AclCitePath)
with open(AclDictPath, "rb") as f:
dic = pkl.load(f)
data.convert_to_dtype(dtype)
data.convert_to_undirected_graph()
data.convert_to_binary_link_matrix()
data.prune_and_shuffle(min_doc_len=MinDocLen,
min_link_count=MinLinkCount)
trainData, testData = data.doc_completion_split()
for pseudoNegCount in (5, 10, 25, 50, 100):
rd.seed(0xC0FFEE)
# Initialise the model
K = TopicCount
model = rtm.newModelAtRandom(trainData,
K,
dtype=dtype,
pseudoNegCount=data.doc_count *
pseudoNegCount)
queryState = rtm.newQueryState(trainData, model)
trainPlan = rtm.newTrainPlan(iterations=50,
logFrequency=LogFreq,
fastButInaccurate=False,
debug=True)
# Train the model, and the immediately save the result to a file for subsequent inspection
model, topics, (bndItrs, bndVals,
bndLikes) = rtm.train(trainData, model, queryState,
trainPlan)
# with open(newModelFileFromModel(model), "wb") as f:
# pkl.dump ((model, query, (bndItrs, bndVals, bndLikes)), f)
# Plot the evolution of the bound during training.
fig, ax1 = plt.subplots()
ax1.plot(bndItrs, bndVals, 'b-')
ax1.set_xlabel('Iterations')
ax1.set_ylabel('Bound', color='b')
ax2 = ax1.twinx()
ax2.plot(bndItrs, bndLikes, 'r-')
ax2.set_ylabel('Likelihood', color='r')
fig.show()
plt.show()
# Print out the most likely topic words
# scale = np.reciprocal(1 + np.squeeze(np.array(data.words.sum(axis=0))))
vocab = rtm.wordDists(model)
topWordCount = 10
kTopWordInds = [
self.topWordInds(vocab[k, :], topWordCount) for k in range(K)
]
like = rtm.log_likelihood(trainData, model, topics)
perp = perplexity_from_like(like, trainData.word_count)
# print ("Prior %s" % (str(model.topicPrior)))
print("Pseudo Neg-Count: %d " % pseudoNegCount)
print("\tTrain Perplexity: %f\n\n" % perp)
# for k in range(model.K):
# print ("\nTopic %d\n=============================" % k)
# print ("\n".join("%-20s\t%0.4f" % (dic[kTopWordInds[k][c]], vocab[k][kTopWordInds[k][c]]) for c in range(topWordCount)))
min_probs = rtm.min_link_probs(model, topics, testData.links)
link_probs = rtm.link_probs(model, topics, min_probs)
try:
map = mean_average_prec(testData.links, link_probs)
except:
print("Unexpected error")
print("\tThe Mean-Average-Precision is %.3f" % map)
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)