-
Notifications
You must be signed in to change notification settings - Fork 27
/
sparseTAM.m
296 lines (277 loc) · 12.9 KB
/
sparseTAM.m
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
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
function [acc theta eta_t eta_a qa_te] = sparseTAM(x,K,varargin)
%function [acc theta eta_t eta_a] = sparseTAM(x,K,varargin)
%[sparse aspects alpha gamma_dirichlet te_x te_aspects vocab ...
% seed hyperprior_update_interval compute_perplexity ...
% topic_aspects verbose init_eta_t init_eta_a max_its max_mstep_its ...
% save_prefix eval_period] = ...
% process_options(varargin,'sparse',1,'aspects',ones(size(x,1),1),'alpha',1/K,...
% 'init-gamma-dirichlet',1,'te-x',[],'te-aspects',[],...
% 'vocab',[],'seed',[],'hyperprior-update-interval',1,...
% 'compute-perplexity',1,'topic-aspects',0,'verbose',1,'init-eta-t',[],...
% 'init-eta-a',[],'max-its',100,'max-mstep-its',100,'save-prefix',[],'eval-period',5);
[sparse aspects alpha gamma_dirichlet te_x te_aspects vocab ...
seed hyperprior_update_interval compute_perplexity ...
topic_aspects verbose init_eta_t init_eta_a max_its max_mstep_its ...
save_prefix eval_period] = ...
process_options(varargin,'sparse',1,'aspects',ones(size(x,1),1),'alpha',1/K,...
'init-gamma-dirichlet',1,'te-x',[],'te-aspects',[],...
'vocab',[],'seed',[],'hyperprior-update-interval',1,...
'compute-perplexity',1,'topic-aspects',0,'verbose',1,'init-eta-t',[],...
'init-eta-a',[],'max-its',100,'max-mstep-its',100,'save-prefix',[],'eval-period',5);
disp(varargin);
%screen words that have zero counts, because they mess stuff up
word_counts = sum(x); x = x(:,word_counts>0);
if ~isempty(te_x), te_x = te_x(:,word_counts>0); end
if ~isempty(vocab), vocab = vocab(word_counts > 0); end
alpha = alpha * ones(1,K); acc = 0;
[D W] = size(x); theta = zeros(D,K); A = max(aspects);
if A > 1, compute_perplexity = 0; end
%% corpus initialization
if ~isempty(seed), rand('seed',seed); end
%mean
m = makeLogProbs(sum(x))';
%eta_t (topics)
if sparse
if ~isempty(init_eta_t),eta_t = init_eta_t; else
eta_t = zeros(W,K);
%need enough words to get a good initialization
D_init = round(10000 / (full(sum(sum(x))) / size(x,1)))
if K>1,
for k = 1:K,
%consider initializing non-sparsely
dsamp = randsample(D,D_init);
%eta_t_nonsparse = computeBetaVariational(full(sum(x(dsamp,:))'),m,'precision',1);
eta_t(:,k) = computeBetaSparseVariational(full(sum(x(dsamp,:))'),m,'max-its',max_mstep_its,'verbose',0);
end
end
end
%eta_a (aspects)
if ~isempty(init_eta_a), eta_a = init_eta_a; else
eta_a = zeros(W,A);
if A>1
for j = 1:A
eta_a(:,j) = computeBetaSparseVariational(full(sum(x(aspects==j,:))'),m,'max-its',max_mstep_its);
end
end
end
%eta_ta (interactions)
eta_ta = zeros(W,K,A); if topic_aspects && A>1, for j = 1:A, for k = 1:K, eta_ta(:,k,j) = 0.1 * (eta_t(:,k) + eta_a(:,j)); end; end; end
%sums
eta_sum = makeEtaSum(m,eta_a,eta_t,eta_ta);
else
assert(A==1); %can't do non-sparse aspect models now
for k = 1:K
eta_sum(:,k) = computeBetaDirichlet(full(sum(x(randsample(D,10),:))),gamma_dirichlet);
end
gamma_dirichlet = gamma_dirichlet * ones(1,W);
end
sigma = repmat(sum(x,2)/K,1,K) + repmat(alpha,D,1);
q_a = zeros(D+1,max(aspects)); q_a(end,:) = 1/max(aspects); %prior
eta_lv_score = 0; eta_ta_lv_score = 0; prior_prob = 0;
%iter = newIterator(max_its,'debug',true,'thresh',1e-5);
iter = newDeltaIterator(max_its,'debug',true,'thresh',0);
ecounts = zeros(W,K,A);
while ~iter.done
%% E-step
word_score = 0; estep_lv_score = 0;
old_ecounts = ecounts;
ecounts = zeros(W,K,A);
for i = 1:D
if (rem(i,100)==0), fprintf('+'); end
old_sigma = sigma(i,:);
log_p_a = digamma(sum(q_a)) - digamma(sum(sum(q_a)));
%[theta(i,:) q_a(i,:) new_counts sigma(i,:) score doc_lv_score
%doc_word_score] = tamEStep(x(i,:),eta_sum,alpha,log_p_a,old_sigma);
[theta(i,:) q_a(i,:) new_counts sigma(i,:) score doc_lv_score doc_word_score] = tamEStep(x(i,:),eta_sum(:,:,aspects(i)),alpha,0,old_sigma);
q_a(i,:) = 0; q_a(i,aspects(i)) = 1;
ecounts(:,:,aspects(i)) = ecounts(:,:,aspects(i)) + full(new_counts);
doc_lv_score = doc_lv_score + digamma(sum(q_a(:,aspects(i)))) - digamma(sum(sum(q_a))); %E[log P(a_i)] - E[log Q(a_i)]
word_score = word_score + doc_word_score;
estep_lv_score = estep_lv_score + doc_lv_score;
end
fprintf('fro norm of counts: %.3f\n',norm(reshape(ecounts,W,K*A)-reshape(old_ecounts,W,K*A),'fro'));
estep_lv_score = estep_lv_score - kldirichlet(sum(q_a),q_a(end,:));
fprintf('\n');
computeScore(word_score,estep_lv_score,eta_lv_score,prior_prob,'print',1);
%% M-step.
if A>1 && K > 1, max_its = 10; else max_its = 0; end
mstep_iter = newIterator(max_its,'thresh',1e-5,'debug',true);
if sparse
eta_t_lv_score = zeros(K,1); eta_a_lv_score = zeros(A,1);
old_eta_t = eta_t;
while ~mstep_iter.done
if K > 1,
for k = 1:K
eq_m = logNormalizeRows(reshape(eta_sum(:,k,:),W,A)' - repmat(eta_t(:,k),1,A)');
eta_t(:,k) = computeBetaSparseVariational(reshape(ecounts(:,k,:),W,A),eq_m','max-its',max_mstep_its);
eta_sum = makeEtaSum(m,eta_a,eta_t,eta_ta);
end
end
fprintf(' ');
if A > 1,
for j = 1:A
if sparse
eq_m = logNormalizeRows(eta_sum(:,:,j)' - repmat(eta_a(:,j),1,K)');
eta_a(:,j) = computeBetaSparseVariational(ecounts(:,:,j),eq_m','max-its',max_mstep_its);
eta_sum = makeEtaSum(m,eta_a,eta_t,eta_ta);
else
assert(K==1);
eta_a(:,j) = computeBetaDirichlet(ecounts(:,1,j)');
end
end
end
fprintf(' ');
if topic_aspects, for k = 1:K, for j=1:A,
assert(sparse==1);
eq_m = logNormalizeRows(eta_sum(:,k,j)' - eta_ta(:,k,j)');
eta_ta(:,k,j) = computeBetaSparseVariational(ecounts(:,k,j),eq_m','max-its',max_mstep_its);
end; end;
eta_sum = makeEtaSum(m,eta_a,eta_t,eta_ta);
end
word_score = 0;
for j = 1:A, for k = 1:K
word_score = word_score + scoreWords(ecounts(:,k,j),eta_sum(:,k,j));
end; end
eta_lv_score = sum(eta_t_lv_score) + sum(eta_a_lv_score) + sum(sum(eta_ta_lv_score));
fprintf('\n');
total_score = computeScore(word_score,estep_lv_score,eta_lv_score,prior_prob,'print',0);
density_t = sum(sum(abs(eta_t) > 1e-4)) / numel(eta_t);
density_a = sum(sum(abs(eta_a) > 1e-4)) / numel(eta_a);
density_ta = sum(sum(sum(abs(eta_ta)>1e-4))) / numel(eta_ta);
fprintf('%.3f\tT=%.3f A=%.3f TA=%.3f norm diff=%.3f\n',total_score,density_t,density_a,density_ta,norm(eta_t-old_eta_t,'fro'));
%mstep_iter = updateIterator(mstep_iter,word_score + eta_lv_score);
mstep_iter = updateDeltaIterator(mstep_iter,reshape(eta_sum,W,A*K));
end
else
if A == 1
eta_sum(:,:,1) = computeBetaDirichlet(ecounts(:,:,1)',gamma_dirichlet)';
elseif K == 1
for a = 1:A
eta_sum(:,1,a) = computeBetaDirichlet(ecounts(:,1,a)',gamma_dirichlet);
end
else
error('in non-sparse model, either K or A must be 1');
end
end
%% status
%iter = updateIterator(iter,total_score);
iter = updateDeltaIterator(iter,[reshape(theta,D*K,1)]);% reshape(q_a(1:end-1),D*A,1)]);
if ~isempty(vocab)
if sparse
if K > 1
fprintf(' ----- topics ----- \n');
if topic_aspects %print interaction terms
for k = 1:K
for j = 1:A
fprintf('K%dA%d\t ',k,j);
%makeTopicReport(eta_sum(1:numel(vocab),k,j)',vocab,'N',10,'background',m'); %i'm not sure supplying the background is a good idea
makeTopicReport(eta_ta(1:numel(vocab),k,j)',vocab,'N',10); %i'm not sure supplying the background is a good idea
end
fprintf('\n');
end
end
%same as above (for topics)
makeTopicReport(tprod(mean(q_a)',[-3],eta_sum,[1 2 -3],'n')',vocab,'N',20);
end
if A > 1
fprintf(' ----- aspects ----- \n');
makeTopicReport(eta_a(1:numel(vocab),:)',vocab,'N',20,'background',zeros(1,W));
end
else
if A > K
for k = 1:K,
if K>1, fprintf(' ----- topic %d ----- ',k); end
makeTopicReport(reshape(eta_sum(1:numel(vocab),k,:),W,A)',vocab,'N',20);
end
else
for a = 1:A
if A > 1, fprintf(' ----- aspect %d ----- \n', a); end
makeTopicReport(eta_sum(1:numel(vocab),:,a)',vocab,'N',20);
end
end
end
end
%% hyperpriors
prior_prob = 0;
if rem(iter.its,hyperprior_update_interval) == 0
%topics might be too sparse? consider removing this.
if K > 1
e_log_theta = digamma(sigma) - repmat(digamma(sum(sigma,2)),1,K);
alpha = fitDirichletPrior(e_log_theta);
end
if verbose > 0, fprintf('new alpha: %s\n',sprintf('%.2f ',alpha)); end
if ~sparse
%gamma = fitDirichletPrior(reshape(eta_sum,W,K*A)');
%here is a newton step that will increase your bound...
for i = 1:10
g_gamma = K * A * W * (digamma(W*gamma_dirichlet(1)) - digamma(gamma_dirichlet(1))) + sum(sum(sum(eta_sum)));
g_gamma = g_gamma - 2/gamma_dirichlet(1) + 1/(gamma_dirichlet(1)^2); %from IG(1,1) prior
h_gamma = K * A * W * (W * trigamma(W*gamma_dirichlet(1)) - trigamma(gamma_dirichlet(1)));
h_gamma = h_gamma + 2/(gamma_dirichlet(1)^2) - 2/(gamma_dirichlet(1)^3); %from IG(1,1) prior
gamma_dirichlet = gamma_dirichlet - 0.2 * g_gamma / h_gamma;
end
prior_prob = prior_prob - 2 * log(gamma_dirichlet(1)) - 1 / gamma_dirichlet(1); %from IG(1,1) prior
if verbose > 0, fprintf('new mean gamma = %.5f\n',mean(gamma_dirichlet)); end
end
end
%% TEST SET
qa_te = [];
if ~isempty(te_x)
theta_te = zeros(size(te_x,1),K);
qa_te = zeros(size(te_x,1),max(aspects));
e_log_a = digamma(sum(q_a)) - digamma(sum(sum(q_a)));
%use the Chib estimator from Wallach et al
if rem(iter.its,compute_perplexity)==0
if sparse, pred_topics = logToSimplex2(eta_sum');
else, pred_topics = logToSimplex2(computeBetaDirichlet(ecounts',gamma_dirichlet)); end
fprintf('PERPLEX %d: %.5f\n',iter.its,computePerplexity(te_x,pred_topics,alpha));
end
if A > 1 && rem(iter.its,eval_period)==0
for i = 1:size(te_x,1)
[theta_te(i,:) qa_te(i,:)] = tamEStep(te_x(i,:),eta_sum(:,:,1:max(aspects)),alpha,e_log_a);
end
[ig preds] = max(qa_te');
acc = mean(preds==te_aspects);
fprintf('ACC %d: = %.3f\n',iter.its,acc);
end
end
if ~isempty(save_prefix) && rem(iter.its,2)==1
save(sprintf('%s.%d.mat',save_prefix,iter.its));
end
end
if ~isempty(te_x)
if compute_perplexity
if sparse, pred_topics = normalize_rows(exp(eta_sum'));
else, pred_topics = logToSimplex2(computeBetaDirichlet(ecounts',gamma_dirichlet)); end
fprintf('FINAL PERPLEX\t%.5f\t%.5f\n',computePerplexity(te_x,pred_topics,alpha),iter.prev(1));
end
if A>1 && ~isempty(te_aspects )
for i = 1:size(te_x,1)
[theta_te(i,:) qa_te(i,:)] = tamEStep(te_x(i,:),eta_sum(:,:,1:max(aspects)),alpha,e_log_a);
end
[ig preds] = max(qa_te');
acc = mean(preds==te_aspects);
fprintf('FINAL ACC\t%.5f\t%.5f\n',acc,iter.prev(1));
end
end
if ~sparse %can't remember why i was doing this...
eta_t = eta_sum(:,:,1) - repmat(m,1,K);
eta_a = eta_sum(:,1,:) - repmat(m,1,A);
end
end
function eta_sum = makeEtaSum(m,eta_a,eta_t,eta_ta)
[W K A] = size(eta_ta);
eta_sum = zeros(W,K,A);
for j = 1:A
for k = 1:K
eta_sum(:,k,j) = logNormalizeVec(m + eta_a(:,j) + eta_t(:,k) + eta_ta(:,k,j));
end
end
end
function total_score = computeScore(word_score, estep_lv_score, eta_lv_score, prior_score, varargin)
[print] = process_options(varargin,'print',0);
total_score = word_score + estep_lv_score + eta_lv_score + prior_score;
if print,
fprintf('%.3f\t=\t%.3f\t+\t%.3f\t+\t%.3f\t+\t%.3f\n',total_score,word_score,estep_lv_score,eta_lv_score,prior_score);
end
end