-
Notifications
You must be signed in to change notification settings - Fork 16
/
evaluate_recognition.m
191 lines (163 loc) · 6.3 KB
/
evaluate_recognition.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
function recognition = evaluate_recognition(opts,data,embedding)
% Small fix for versions of matlab older than 2012b ('8') that do not support stable intersection
if verLessThan('matlab', '8')
inters=@stableintersection;
else
inters=@intersect;
end
% Create/load the dictionary (lexicon)
if ~exist(opts.fileLexicon,'file')
lexicon = extract_lexicon(opts,data);
else
load(opts.fileLexicon,'lexicon')
end
lexicon_phocs = single(lexicon.phocs);
words = lexicon.words;
if opts.TestCCA
emb = embedding.cca;
% Embed the test representations
attReprTe_emb = bsxfun(@rdivide, data.attReprTe,sqrt(sum(data.attReprTe.*data.attReprTe)));
attReprTe_emb(isnan(attReprTe_emb)) = 0;
attReprTe_emb = bsxfun(@minus, attReprTe_emb,emb.matts);
attReprTe_emb = emb.Wx(:,1:emb.K)' * attReprTe_emb;
attReprTe_emb = (bsxfun(@rdivide, attReprTe_emb, sqrt(sum(attReprTe_emb.*attReprTe_emb))));
% Embed the lexicon dictionary
lexicon_phocs_emb = bsxfun(@rdivide, lexicon_phocs,sqrt(sum(lexicon_phocs.*lexicon_phocs)));
lexicon_phocs_emb= bsxfun(@minus, lexicon_phocs_emb,emb.mphocs);
lexicon_phocs_emb = emb.Wy(:,1:emb.K)' * lexicon_phocs_emb;
lexicon_phocs_emb = (bsxfun(@rdivide, lexicon_phocs_emb, sqrt(sum(lexicon_phocs_emb.*lexicon_phocs_emb))));
elseif opts.TestKCCA
emb = embedding.kcca;
% Embed the test representations
matx = emb.rndmatx(1:emb.M,:);
tmp = matx*data.attReprTe;
attReprTe_emb = 1/sqrt(emb.M) * [ cos(tmp); sin(tmp)];
attReprTe_emb=bsxfun(@minus, attReprTe_emb, emb.matts);
attReprTe_emb = emb.Wx(:,1:emb.K)' * attReprTe_emb;
attReprTe_emb = (bsxfun(@rdivide, attReprTe_emb, sqrt(sum(attReprTe_emb.*attReprTe_emb))));
% Embed the lexicon dictionary
maty = emb.rndmaty(1:emb.M,:);
tmp = maty*lexicon_phocs;
lexicon_phocs_emb = 1/sqrt(emb.M) * [ cos(tmp); sin(tmp)];
lexicon_phocs_emb=bsxfun(@minus, lexicon_phocs_emb, emb.mphocs);
lexicon_phocs_emb = emb.Wy(:,1:emb.K)' * lexicon_phocs_emb;
lexicon_phocs_emb = (bsxfun(@rdivide, lexicon_phocs_emb, sqrt(sum(lexicon_phocs_emb.*lexicon_phocs_emb))));
lexicon_phocs_emb(isnan(lexicon_phocs_emb)) = 0;
end
% Get all the valid queries. For most datasets, that is all of them.
qidx = find(~strcmp({data.wordsTe.gttext},'-'));
N = length(qidx);
% Get the scores: p1, cer, and wer
p1small = zeros(N,1);
cersmall = zeros(N,1);
wersmall = zeros(N,1);
p1medium = zeros(N,1);
cermedium = zeros(N,1);
wermedium = zeros(N,1);
p1full = zeros(N,1);
cerfull = zeros(N,1);
werfull = zeros(N,1);
for i=1:N
% Get actual idx, feature vector, and gt transcription
pos = qidx(i);
feat = attReprTe_emb(:,pos);
gt = data.wordsTe(pos).gttext;
% Small lexicon available
if isfield(data.wordsTe,'sLexi')
smallLexicon = unique(data.wordsTe(i).sLexi);
[~,~,ind] = inters(smallLexicon,words,'stable');
scores = feat'*lexicon_phocs_emb(:,ind);
randInd = randperm(length(scores));
scores = scores(randInd);
[scores,I] = sort(scores,'descend');
I = randInd(I);
if strcmpi(gt,smallLexicon{I(1)})
p1small(i) = 1;
else
p1small(i) = 0;
cersmall(i) = levenshtein_c(gt, smallLexicon{I(1)});
end
end
% Medium lexicon available
if isfield(data.wordsTe,'mLexi')
mediumLexicon = unique(data.wordsTe(pos).mLexi);
[~,~,ind] = inters(mediumLexicon,words,'stable');
scores = feat'*lexicon_phocs_emb(:,ind);
randInd = randperm(length(scores));
scores = scores(randInd);
[scores,I] = sort(scores,'descend');
I = randInd(I);
if strcmpi(gt,mediumLexicon(I(1)))
p1medium(i) = 1;
else
p1medium(i) = 0;
cermedium(i) = levenshtein_c(gt, mediumLexicon{I(1)});
end
end
% Full lexicon always available
scores = feat'*lexicon_phocs_emb;
randInd = randperm(length(scores));
scores = scores(randInd);
[scores,I] = sort(scores,'descend');
I = randInd(I);
if strcmpi(gt,words{I(1)})
p1full(i) = 1;
else
p1full(i) = 0;
cerfull(i) = levenshtein_c(gt, words{I(1)});
end
end
% Compute wer if there is line info available
if isfield(data.wordsTe,'lineId')
linesTe = {data.wordsTe.lineId}';
linesTe = linesTe(qidx);
if isfield(data.wordsTe,'sLexi')
recognition.wersmall = compute_wer(linesTe,p1small);
end
if isfield(data.wordsTe,'mLexi')
recognition.wermedium = compute_wer(linesTe,p1medium);
end
recognition.werfull = compute_wer(linesTe,p1full);
end
% Display stuff
fprintf('\n');
disp('**************************************');
disp('************ Recognition ***********');
disp('**************************************');
disp('------------------------------------');
if isfield(data.wordsTe,'sLexi')
recognition.p1small = 100*mean(p1small);
recognition.cersmall = 100*mean(cersmall);
if isfield(recognition,'wersmall')
fprintf('lexicon small -- p@1: %.2f. cer: %.2f. wer: %.2f\n', recognition.p1small, recognition.cersmall, recognition.wersmall);
else
fprintf('lexicon small -- p@1: %.2f. cer: %.2f. wer: N/A\n', recognition.p1small, recognition.cersmall);
end
end
if isfield(data.wordsTe,'mLexi')
recognition.p1medium = 100*mean(p1medium);
recognition.cermedium = 100*mean(cermedium);
if isfield(recognition,'wermedium')
fprintf('lexicon medium -- p@1: %.2f. cer: %.2f. wer: %.2f\n', recognition.p1medium, recognition.cermedium, recognition.wermedium);
else
fprintf('lexicon medium -- p@1: %.2f. cer: %.2f. wer: N/A\n', recognition.p1medium, recognition.cermedium);
end
end
recognition.p1full = 100*mean(p1full);
recognition.cerfull = 100*mean(cerfull);
if isfield(recognition,'werfull')
fprintf('lexicon full -- p@1: %.2f. cer: %.2f. wer: %.2f\n', recognition.p1full, recognition.cerfull, recognition.werfull);
else
fprintf('lexicon full -- p@1: %.2f. cer: %.2f. wer: N/A\n', recognition.p1full, recognition.cerfull);
end
disp('------------------------------------');
end
% Ugly hack to deal with the lack of stable intersection in old versions of
% matlab
function [empty1, empty2, ind] = stableintersection(a, b, varargin)
empty1=0;
empty2=0;
[~,ia,ib] = intersect(a,b);
[~, tmp2] = sort(ia);
ind = ib(tmp2);
end