-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathsparseSCVB0_lda.jl
176 lines (161 loc) · 9.17 KB
/
sparseSCVB0_lda.jl
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
include("definitions.jl")
include("updateFunctions.jl")
include("utility.jl")
function sparseSCVB0_lda(documents::Array{SparseDocument,1},
numWords::Int64,
numDocuments::Int64,
numTopics::Int64,
numDocumentIterations::Int,
burnInPerDoc::Int,
minibatchSize::Int,
Alpha::Float64,
Beta::Float64,
dict,
tau::Float64,
kappa::Float64,
tau2::Float64,
kappa2::Float64,
scale::Float64,
sampleSize::Int,
time_limit::Float64,
miniEpoch::Int,
numDocuments_miniEpoch::Int64,
sparsity::Float64)
# sparse stochastic collapsed variational Bayes (sparseSCVB0) for LDA
Alpha = ones(1,numTopics).*Alpha;
Beta = ones(numWords,1).*Beta;
# initialization of sparseSCVB0 specific model parameters
#documentTopicCounts, wordTopicCounts, topicCounts, topicMem, sparseCounts = initialize(documents, numWords,numTopics,numDocuments);
sparseVBparams=initialize(documents,Alpha, Beta, numTopics,numWords,numDocumentIterations,numDocuments);
miniBatchesPerCorpus = numDocuments ./ minibatchSize;
stepSize = scale ./ (tau^kappa); # step size for global expected counts
#stepSize2 = 0.0; # initializing step size for local expected counts to make it soft local scope
wordTopicCounts_hat = zeros(numWords,numTopics);
topicCounts_hat = zeros(1,numTopics);
iter = 0;
saved_topics = Array{Array{Float64,2}}(0)
saved_Totaltopics = Array{Array{Float64,2}}(0)
saved_time = Float64[]
saved_iters = Int64[]
#saved_docTopics = Array{Array{Float64,2}}(0)
prev_doc = 0; # to track miniEpoch docs
doc_counter = 0;
# main sparse stochastic collapsed variational Bayes (sparseSCVB0) loop
while (sparseVBparams.wallClockTime < time_limit) && (doc_counter<numDocuments)
prev_doc,documentTopicCounts,topicMem,sparseCounts = initialize_miniEpoch(documents,numTopics,numWords,numDocuments_miniEpoch,prev_doc);
for epoch = 1:miniEpoch
for epochDoc = 1:numDocuments_miniEpoch
iter += 1;
tic();
#docInd_mini = mod(iter - 1, numDocuments_miniEpoch) + 1;
#docInd = doc_counter + docInd_mini;
docInd = doc_counter + epochDoc;
docLength = documents[docInd].docLength;
if docLength == 0
continue;
end
# burn-in pass for a document
sparseVBparams,documentTopicCounts,sparseCounts,topicMem = burnInUpdate(sparseVBparams,documentTopicCounts,documents,burnInPerDoc,docInd,epochDoc,docLength,tau2,kappa2,sparseCounts,topicMem,sampleSize,sparsity);
# main loop update for a word
sparseVBparams,documentTopicCounts,wordTopicCounts_hat,topicCounts_hat,sparseCounts,topicMem = mainLoopUpdate(sparseVBparams,documentTopicCounts,documents,burnInPerDoc,docInd,epochDoc,docLength,tau2,kappa2,sparseCounts,
topicMem,wordTopicCounts_hat,topicCounts_hat,sampleSize);
# mini-batch update of expected global counts
if mod(iter, minibatchSize) == 0
# compute effective stepsize to account for minibatches
stepSize = 1 - (1-stepSize)^minibatchSize
# minibatch update of expected global counts
#wordTopicCounts,topicCounts,wordTopicCounts_hat,topicCounts_hat=minibatchUpdate(stepSize,miniBatchesPerCorpus,wordTopicCounts,topicCounts,wordTopicCounts_hat,topicCounts_hat);
topicCounts = sparseVBparams.topicCounts;
wordTopicCounts = sparseVBparams.wordTopicCounts;
for topic = 1:numTopics
for word = 1:numWords
wordTopicCounts[word,topic] = (1 - stepSize) .* wordTopicCounts[word,topic] + stepSize .* miniBatchesPerCorpus .* wordTopicCounts_hat[word,topic];
wordTopicCounts_hat[word,topic] = wordTopicCounts_hat[word,topic]*0.0;
end
topicCounts[topic] = (1 - stepSize) .* topicCounts[topic] + stepSize .* miniBatchesPerCorpus .* topicCounts_hat[topic];
topicCounts_hat[topic] = topicCounts_hat[topic]*0.0;
end
stepSize = scale ./ (iter + tau)^kappa;
end
sparseVBparams.wallClockTime += toq();
if iter == 1
push!(saved_topics,copy(sparseVBparams.wordTopicCounts))
push!(saved_Totaltopics,copy(sparseVBparams.topicCounts))
push!(saved_time,sparseVBparams.wallClockTime)
push!(saved_iters,iter)
else
if mod(iter,200)==0
push!(saved_topics,copy(sparseVBparams.wordTopicCounts))
push!(saved_Totaltopics,copy(sparseVBparams.topicCounts))
push!(saved_time,sparseVBparams.wallClockTime)
push!(saved_iters,iter);
end
end
end
end
doc_counter = doc_counter + numDocuments_miniEpoch;
#push!(saved_docTopics,copy(documentTopicCounts))
end
return sparseVBparams,saved_topics,saved_Totaltopics,saved_time,saved_iters,iter;
end
function initialize(documents,Alpha, Beta, numTopics,numWords,numDocumentIterations,numDocuments);
# initialize sparseSCVB0 model parameters
sparseVBparams = sparseSCVBO_params_type(Alpha, Beta, numTopics,numWords,numDocumentIterations,numDocuments);
totalWordsInCorpus = 0;
for i = 1:numDocuments
docLength = documents[i].docLength;
totalWordsInCorpus = totalWordsInCorpus + docLength;
end
# initialize expected global counts
sparseVBparams.topicCounts.*= totalWordsInCorpus ./ numTopics;
sparseVBparams.wordTopicCounts.*= totalWordsInCorpus ./ (numWords .* numTopics);
# initializing alias samples from alias table for each word
sparseVBparams.cachedSamples,sparseVBparams.cached_qw,sparseVBparams.cached_Qw = aliasMethod(sparseVBparams.numWords,sparseVBparams.Alpha,
sparseVBparams.wordTopicCounts,sparseVBparams.topicCounts,sparseVBparams.Beta,
sparseVBparams.BetaSum,sparseVBparams.numTopics,sparseVBparams.numCachedSamples,
sparseVBparams.cachedSamples,sparseVBparams.cached_qw,sparseVBparams.cached_Qw);
return sparseVBparams;
end
function initialize_miniEpoch(documents,numTopics,numWords,numDocuments_miniEpoch,prev_doc);
# initialization of topic storage
topicMem = Array{TopicArray}(0);
# initialization for keeping track on non-zero sparse counts
sparseCounts = Array{SparseCounts}(0);
# initialize sparseSCVB0 miniEpoch parameters
documentTopicCounts = zeros(numDocuments_miniEpoch,numTopics);
j=0;
for i = (prev_doc+1):(prev_doc+numDocuments_miniEpoch)
j=j+1;
docLength = documents[i].docLength;
tokenLength = length(documents[i].tokenIndex); # storing topic for clumped documents
Assigned = Array{Int64}(tokenLength); # array to store topic assignment for each token of clumped document
nonzeroDocIndex = Array{Int64}(0); # array to keep track of non-zero document-topic counts
documentTopicTmp = rand(1,numTopics);
sum_documentTopicTmp = sum(documentTopicTmp);
for topic = 1:numTopics
documentTopicCounts[j,topic] = (documentTopicTmp[topic]./sum_documentTopicTmp).*docLength;
push!(nonzeroDocIndex,topic);
end
Assigned = rand(1:numTopics,tokenLength);
push!(topicMem,TopicArray(Assigned));
push!(sparseCounts,SparseCounts(nonzeroDocIndex));
end
prev_doc = prev_doc+numDocuments_miniEpoch;
return prev_doc,documentTopicCounts,topicMem,sparseCounts;
end
function aliasMethod(numWords,Alpha,wordTopicCounts,topicCounts,Beta,BetaSum,numTopics,numCachedSamples,cachedSamples,cached_qw,cached_Qw)
qw = Array{Float64}(numTopics); # unnormalized discrete distribution of dense bucket
for word = 1:numWords
cached_Qw[word] = 0.0;
for topic = 1:numTopics
qw[topic] = Alpha[topic] .* (wordTopicCounts[word,topic] + Beta[word]) ./ (topicCounts[topic] + BetaSum);
cached_qw[word,topic] = qw[topic];
cached_Qw[word] += qw[topic];
end
aliasTable = generateAlias(qw,numTopics);
for sample = 1:numCachedSamples
cachedSamples[word,sample] = sampleAlias(aliasTable,numTopics);
end
end
return cachedSamples,cached_qw,cached_Qw;
end