def train(data, model, query, plan, updateVocab=True):
'''
Infers the topic distributions in general, and specifically for
each individual datapoint, and additionally learns the weights
needed to predict new links.
Params:
W - the DxT document-term matrix
X - The DxD document-document matrix
model - the initial model configuration. This is MUTATED IN-PLACE
qyery - the query results - essentially all the "local" variables
matched to the given observations. Also MUTATED IN-PLACE
plan - how to execute the training process (e.g. iterations,
log-interval etc.)
Return:
The updated model object (note parameters are updated in place, so make a
defensive copy if you want it)
The query object with the update query parameters
'''
iterations, epsilon, logFrequency, fastButInaccurate, debug = \
plan.iterations, plan.epsilon, plan.logFrequency, plan.fastButInaccurate, plan.debug
docLens, topicMeans = \
query.docLens, query.topicDists
K, topicPrior, vocabPrior, wordDists, weights, negCount, reg, dtype = \
model.K, model.topicPrior, model.vocabPrior, model.wordDists, model.weights, model.pseudoNegCount, model.regularizer, model.dtype
# Quick sanity check
if np.any(docLens < 1):
raise ValueError(
"Input document-term matrix contains at least one document with no words"
)
assert dtype == np.float64, "Only implemented for 64-bit floats"
# Prepare the data for inference
topicMeans = _convertDirichletParamToMeans(docLens, topicMeans, topicPrior)
W = data.words
D, T = W.shape
X = data.links
iters, bnds, likes = [], [], []
# Instead of storing the full topic assignments for every individual word, we
# re-estimate from scratch. I.e for the memberships z which is DxNxT in dimension,
# we only store a 1xNxT = NxT part.
z = np.empty((K, ), dtype=dtype, order='F')
diWordDistSums = np.empty((K, ), dtype=dtype)
diWordDists = np.empty(wordDists.shape, dtype=dtype)
for itr in range(iterations):
if debug: printAndFlushNoNewLine("\n %4d: " % itr)
diWordDistSums[:] = wordDists.sum(axis=1)
fns.digamma(diWordDistSums, out=diWordDistSums)
fns.digamma(wordDists, out=diWordDists)
if updateVocab:
# Perform inference, updating the vocab
wordDists[:, :] = vocabPrior
for d in range(D):
if debug and d % 100 == 0: printAndFlushNoNewLine(".")
wordIdx, z = _update_topics_at_d(d, data, weights, docLens,
topicMeans, topicPrior,
diWordDists, diWordDistSums)
wordDists[:, wordIdx] += W[d, :].data[np.newaxis, :] * z
_infer_weights(data, weights, topicMeans, topicPrior, negCount,
reg)
# Log bound and the determine if we can stop early
if itr % logFrequency == 0:
iters.append(itr)
bnds.append(_var_bound_internal(data, model, query))
likes.append(_log_likelihood_internal(data, model, query))
if debug: print("%.3f < %.3f" % (bnds[-1], likes[-1]))
if converged(iters, bnds, len(bnds) - 1, minIters=5):
break
# Update hyperparameters (do this after bound, to make sure bound
# calculation is internally consistent)
if itr > 0 and itr % HyperParamUpdateInterval == 0:
if debug: print("Topic Prior was " + str(topicPrior))
_updateTopicHyperParamsFromMeans(model, query)
if debug: print("Topic Prior is now " + str(topicPrior))
else:
for d in range(D):
_ = _update_topics_at_d(d, W, docLens, topicMeans, topicPrior,
diWordDists, diWordDistSums)
topicMeans = _convertMeansToDirichletParam(docLens, topicMeans, topicPrior)
return ModelState(K, topicPrior, vocabPrior, wordDists, weights, negCount, reg, dtype, model.name), \
QueryState(docLens, topicMeans), \
(np.array(iters, dtype=np.int32), np.array(bnds), np.array(likes))
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 are used in this model
modelState - the actual LDA model. In a training run (query = False) this
will be mutated in place, and then returned.
queryState - the query results - essentially all the "local" variables
matched to the given observations. This will be mutated in-place
and then returned.
trainPlan - how to execute the training process (e.g. iterations,
log-interval etc.)
query -
Return:
The updated model object (note parameters are updated in place, so make a
defensive copy if you want it)
The query object with the update query parameters
'''
iterations, epsilon, logFrequency, fastButInaccurate, debug, batchSize, rate_retardation, forgetting_rate = \
trainPlan.iterations, trainPlan.epsilon, trainPlan.logFrequency, trainPlan.fastButInaccurate, trainPlan.debug, \
trainPlan.batchSize, trainPlan.rate_retardation, trainPlan.forgetting_rate
W_list, docLens, topicDists = \
queryState.W_list, queryState.docLens, queryState.topicDists
K, topicPrior, vocabPrior, wordDists, dtype = \
modelState.K, modelState.topicPrior, modelState.vocabPrior, modelState.wordDists, modelState.dtype
W = data.words
D, T = W.shape
# Quick sanity check
if np.any(docLens < 1):
raise ValueError(
"Input document-term matrix contains at least one document with no words"
)
# Book-keeping for logs
logPoints = 1 if logFrequency == 0 else iterations // logFrequency
boundIters = np.zeros(shape=(logPoints, ))
boundValues = np.zeros(shape=(logPoints, ))
likelyValues = np.zeros(shape=(logPoints, ))
bvIdx = 0
# Instead of storing the full topic assignments for every individual word, we
# re-estimate from scratch. I.e for the memberships z which is DxNxT in dimension,
# we only store a 1xNxT = NxT part.
z_dnk = np.empty((docLens.max(), K), dtype=dtype, order='F')
# Select the training iterations function appropriate for the dtype
current_micro_time = lambda: int(time.time())
do_iterations = compiled.iterate_f32 \
if modelState.dtype == np.float32 \
else compiled.iterate_f64
# do_iterations = iterate # pure Python
# Iterate in segments, pausing to take measures of the bound / likelihood
segIters = logFrequency
remainder = iterations - segIters * (logPoints - 1)
totalItrs = 0
for segment in range(logPoints - 1):
start = current_micro_time()
totalItrs += do_iterations (segIters, \
batchSize, segment * segIters, rate_retardation, forgetting_rate, \
D, K, T, \
W_list, docLens, \
topicPrior, vocabPrior, \
z_dnk, topicDists, wordDists)
duration = current_micro_time() - start
boundIters[bvIdx] = segment * segIters
boundValues[bvIdx] = var_bound(data, modelState, queryState)
likelyValues[bvIdx] = log_likelihood(data, modelState, queryState)
perp = perplexity_from_like(likelyValues[bvIdx], W.sum())
bvIdx += 1
if converged(boundIters, boundValues, bvIdx, epsilon, minIters=20):
boundIters, boundValues, likelyValues = clamp(
boundIters, boundValues, likelyValues, bvIdx)
return ModelState(K, topicPrior, vocabPrior, wordDists, modelState.dtype, modelState.name), \
QueryState(W_list, docLens, topicDists), \
(boundIters, boundValues, likelyValues)
print(
"Segment %d/%d Total Iterations %d Duration %d Perplexity %4.0f Bound %10.2f Likelihood %10.2f"
% (segment, logPoints, totalItrs, duration, perp,
boundValues[bvIdx - 1], likelyValues[bvIdx - 1]))
# Final batch of iterations.
do_iterations (remainder, D, K, T, \
W_list, docLens, \
topicPrior, vocabPrior, \
z_dnk, topicDists, wordDists)
boundIters[bvIdx] = iterations - 1
boundValues[bvIdx] = var_bound(data, modelState, queryState)
likelyValues[bvIdx] = log_likelihood(data, modelState, queryState)
return ModelState(K, topicPrior, vocabPrior, wordDists, modelState.dtype, modelState.name), \
QueryState(W_list, docLens, topicDists), \
(boundIters, boundValues, likelyValues)
def train(data, model, query, plan, updateVocab=True):
'''
Infers the topic distributions in general, and specifically for
each individual datapoint, and additionally learns the weights
needed to predict new links.
Params:
W - the DxT document-term matrix
X - The DxD document-document matrix
model - the initial model configuration. This is MUTATED IN-PLACE
qyery - the query results - essentially all the "local" variables
matched to the given observations. Also MUTATED IN-PLACE
plan - how to execute the training process (e.g. iterations,
log-interval etc.)
Return:
The updated model object (note parameters are updated in place, so make a
defensive copy if you want it)
The query object with the update query parameters
'''
iterations, epsilon, logFrequency, fastButInaccurate, debug = \
plan.iterations, plan.epsilon, plan.logFrequency, plan.fastButInaccurate, plan.debug
docLens, topics, postTopicCov, U, V, tsums_bydoc, tsums_bytop, exp_tsums_bydoc, exp_tsums_bytop, lse_at_k, out_counts, in_counts = \
query.docLens, query.topics, query.postTopicCov, query.U, query.V, query.tsums_bydoc, query.tsums_bytop, query.exp_tsums_bydoc, query.exp_tsums_bytop, query.lse_at_k, query.out_counts, query.in_counts
K, Q, topicPrior, vocabPrior, wordDists, topicCov, dtype, name = \
model.K, model.Q, model.topicPrior, model.vocabPrior, model.wordDists, model.topicCov, model.dtype, model.name
# Quick sanity check
if np.any(docLens < 1):
raise ValueError(
"Input document-term matrix contains at least one document with no words"
)
assert dtype == np.float64, "Only implemented for 64-bit floats"
# Prepare the data for inference
W, L, X_fixme = data.words, data.links, data.feats
D, T = W.shape
iters, bnds, likes = [], [], []
# Instead of storing the full topic assignments for every individual word, we
# re-estimate from scratch. I.e for the memberships z which is DxNxT in dimension,
# we only store a 1xNxT = NxT part.
diWordDistSums = np.empty((K, ), dtype=dtype)
diWordDists = np.empty(wordDists.shape, dtype=dtype)
new_in_counts = in_counts.copy()
newCov = np.ndarray(shape=(K, K), dtype=model.dtype)
newCov.fill(0)
invCov = la.inv(topicCov)
S = topicCov.copy()
rhs = np.ndarray(shape=(K, ), dtype=model.dtype)
b, f = rhs.copy(), rhs.copy(
) # linear coefficients of the Bohing quadratic bounds
new_maxes_bytop = np.ndarray(shape=(K, ), dtype=model.dtype)
maxes_bytop = topics.max(axis=0)
for itr in range(iterations):
if itr % logFrequency == 0:
iters.append(itr)
bnds.append(_var_bound_internal(data, model, query))
likes.append(_log_likelihood_internal(data, model, query))
if debug:
print(
"Bound : %f \t Likelihood %f \t Perplexity %.2f" %
(bnds[-1], likes[-1], np.exp(-likes[-1] / docLens.sum())))
if converged(iters, bnds, len(bnds) - 1, minIters=5):
break
if debug: printAndFlushNoNewLine("\n %4d: " % itr)
newCov[:, :] = 0
new_in_counts[:] = 0
# U and V FIXME DEBUG
# U[:, :] = la.lstsq(V.T, topics.T)[0].T
# V[:, :] = la.lstsq(U, topics)[0]
diWordDistSums[:] = wordDists.sum(axis=1)
fns.digamma(diWordDistSums, out=diWordDistSums)
fns.digamma(wordDists, out=diWordDists)
wordDists[:, :] = vocabPrior
new_maxes_bytop.fill(1E-300)
for d in range(D):
if d % 100 == 0:
printAndFlushNoNewLine(".")
wordIdx, z, linkIdx, y = _update_topics_at_d(
d, data, docLens, topics, topicPrior, lse_at_k, diWordDists,
diWordDistSums)
# Update the word distributions
wordDists[:, wordIdx] += W[d, :].data[np.newaxis, :] * z
# Determine the topic distribution
# Step 1, the covariance
S[:, :] = invCov
S[np.diag_indices_from(S)] += docLens[d] + out_counts[d]
S[:, :] -= 1. / (K + 1)
S[np.diag_indices_from(S)] += (K - 1.) / K * in_counts
S = la.inv(S)
# Topics Step 2, the actual right-hand side
rhs[:] = invCov.dot(U[d, :].dot(V))
rhs += (z * W[d, :].data[np.newaxis, :]).sum(axis=1)
ysum = (y * L[d, :].data[:, np.newaxis]).sum(axis=0)
#rhs += ysum
b[:] = topics[d, :] - 1. / (K + 1) * topics[d, :].sum() - softmax(
topics[d, :])
b *= docLens[d]
rhs += b
f[:] = topics[d, :] - 1. / (D + 1) * tsums_bytop - np.exp(
topics[d, :] - maxes_bytop) / exp_tsums_bytop
f *= in_counts
rhs += f
rhs[:] += (D - 1) / (2 * D + 2) * (in_counts *
(tsums_bytop - topics[d, :]))
# Topics Step 3: solve
new_topics = S.dot(rhs)
# Topics Step 4: update the running counts and covariance, then assign the new topics to "topics'
tsums_bytop -= topics[d, :]
tsums_bytop += new_topics
new_maxes_bytop = np.maximum(new_maxes_bytop, new_topics)
new_in_counts += ysum
vec = new_topics - U[d, :].dot(V)
newCov += np.outer(vec, vec)
newCov += np.diag(S)
topics[d, :] = new_topics
# Next step is the posterior covariance
postTopicCov[d, :] = np.diag(S)
# The covariance hyper-parameter
topicCov[:, :] = newCov
invCov[:, :] = la.inv(topicCov)
# The remaining running counts, and the column-wise softmax adjustment
maxes_bytop[:] = new_maxes_bytop
in_counts[:] = new_in_counts
exp_tsums_bytop[:] = np.sum(np.exp(topics -
maxes_bytop[np.newaxis, :]),
axis=0)
return ModelState(K, Q, topicPrior, vocabPrior, wordDists, topicCov, dtype, model.name), \
QueryState(docLens, topics, postTopicCov, U, V, tsums_bydoc, tsums_bytop, exp_tsums_bydoc, exp_tsums_bytop, lse_at_k, out_counts, in_counts), \
(np.array(iters, dtype=np.int32), np.array(bnds), np.array(likes))