-
Notifications
You must be signed in to change notification settings - Fork 21
/
sn_FA_5c_train_v52.m
200 lines (153 loc) · 8.39 KB
/
sn_FA_5c_train_v52.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
addpath core/
addpath utils/
addpath optimization/
addpath data/speaker-naming/processed_training_data/train_face/
addpath data/speaker-naming/processed_training_data/train_audio/
clearvars -global config;
global config mem;
gpuDevice(1);
sn_FA_configure();
lstm_init_v52();
solid_idx = [1,4,8,12,16,19,23,27,31,34,38,42,46];
%null_idx = [2:3,5:7,9:11,13:15,17:18,20:22,24:26,28:30,32:33,35:37,39:41,43:45,47:49];
count = 0;
cost_avg = 0;
epoc = 0;
points_seen = 0;
display_points = 1000;
save_points = 10000;
% training data
load('data/speaker-naming/processed_training_data/train_face/1.mat');
face_labels = labels;
tmp = reshape(samples, size(samples,1), []);
config.data_mean = mean(tmp,2);
config.one_over_data_std = 1 ./ std(tmp')'; clear tmp;
face_samples = zeros(53, 49, size(samples,3));
face_samples(:,solid_idx,:) = samples;
for t = 1:size(face_samples,2)
if(face_samples(1,t,1) == 0)
face_samples(:,t,:) = face_samples(:,t-1,:);
end
end
load('data/speaker-naming/processed_training_data/train_audio/1');
samples = reshape(samples, size(samples,1), []);
config.data_mean = cat(1, config.data_mean, mean(samples,2));
config.one_over_data_std = cat(1, config.one_over_data_std, 1 ./ std(samples')');
% test data
load('data/speaker-naming/raw_full/test/5classes/1'); % normal test data
test_samples = test_samples(:,:,1:2000);
test_labels = test_labels(:,1:2000);
test_labels = reshape(test_labels, size(test_labels,1), 1, size(test_labels,2));
test_labels = repmat(test_labels, [1 size(test_samples,2) 1]);
test_samples = config.NEW_MEM(test_samples);
test_labels = config.NEW_MEM(test_labels);
test_samples = bsxfun(@times, bsxfun(@minus, test_samples, config.data_mean), config.one_over_data_std);
load('data/speaker-naming/raw_full/test/5classes/6'); % ourliers (face and audio does not belong to the same person)
outlier_test_samples = outlier_test_samples(:,:,1:2000);
outlier_test_samples = config.NEW_MEM(outlier_test_samples);
outlier_test_samples = bsxfun(@times, bsxfun(@minus, outlier_test_samples, config.data_mean), config.one_over_data_std);
fprintf('%s\n', datestr(now, 'dd-mm-yyyy HH:MM:SS FFF'));
for p = 1:100
for m = 1:11
load(strcat('data/speaker-naming/processed_training_data/train_audio/', num2str(m)));
labels_ = labels;
labels = reshape(labels, size(labels,1), 1, size(labels,2));
labels = repmat(labels, [1 size(samples,2) 1]);
perm = randperm(size(labels, 3));
samples = samples(:,:,perm);
labels = labels(:,:,perm);
labels_ = labels_(:,perm);
samples = config.NEW_MEM(samples);
labels = config.NEW_MEM(labels);
samples = padarray(samples, [53 0 0], 'pre');
% match the face and audio training data
for c = 1:5
audio_idx = find(labels_(c,:) == 1);
face_idx = find(face_labels(c,:) == 1);
selected_face_samples = face_samples(:,:,face_idx);
if(length(face_idx) < length(audio_idx))
perm = randperm(length(audio_idx) - length(face_idx));
selected_face_samples = cat(3, selected_face_samples, selected_face_samples(:,:,perm));
end
perm = randperm(size(selected_face_samples, 3));
selected_face_samples = selected_face_samples(:,:,perm);
selected_face_samples = selected_face_samples(:,:,1:length(audio_idx));
samples(1:53,:,audio_idx) = selected_face_samples;
end
samples = bsxfun(@times, bsxfun(@minus, samples, config.data_mean), config.one_over_data_std);
for i = 1:size(samples, 3)/config.batch_size
points_seen = points_seen + config.batch_size;
start_idx = config.batch_size * (i-1) + 1;
end_idx = start_idx + config.batch_size - 1;
in = samples(:,:,start_idx:end_idx);
label = labels(:,:,start_idx:end_idx);
lstm_core_v52(in, label);
if(cost_avg == 0)
cost_avg = config.cost{1} + config.cost{2};
else
cost_avg = (cost_avg + config.cost{1} + config.cost{2}) / 2;
end
eta = config.learning_rate / (1 + points_seen*config.decay);
adagrad_update(eta);
% display point
if(mod(points_seen, display_points) == 0)
count = count + 1;
fprintf('%d ', count);
end
% save point
if(mod(points_seen, save_points) == 0)
fprintf('\n%s', datestr(now, 'dd-mm-yyyy HH:MM:SS FFF'));
epoc = epoc + 1;
correct_num = 0;
train_correct_num = 0;
outlier_currect_num = 0;
outlier_thres = 10;
for ii = 1:size(test_samples, 3)/config.batch_size
start_idx = config.batch_size * (ii-1) + 1;
end_idx = start_idx + config.batch_size - 1;
% outlier rejection acc
val_sample = outlier_test_samples(:,:,start_idx:end_idx);
lstm_core_v52(val_sample, 1);
[vv1, pos1] = max(mem.net_out(:,25:end,:,1));
[vv2, pos2] = max(mem.net_out(:,25:end,:,2));
tt = sum((pos1 == pos2));
tt = reshape(tt, 1, config.batch_size);
estimated_labels = zeros(1, config.batch_size);
estimated_labels(tt < outlier_thres) = -1; % if less than 'outlier_thres' outputs agrees with each other, an outlier
true_labels = zeros(1, config.batch_size);
true_labels = true_labels - 1;
outlier_currect_num = outlier_currect_num + length(find(estimated_labels == true_labels));
% test acc
val_sample = test_samples(:,:,start_idx:end_idx);
val_label = test_labels(:,:,start_idx:end_idx);
lstm_core_v52(val_sample, 1);
[value, estimated_labels] = max(mem.net_out(:,end,:,1)+mem.net_out(:,end,:,2));
[vv1, pos1] = max(mem.net_out(:,25:end,:,1));
[vv2, pos2] = max(mem.net_out(:,25:end,:,2));
tt = sum((pos1 == pos2));
estimated_labels(tt < outlier_thres) = -1; % to compute the real accuracy, apply outliear rejection first
[value, true_labels] = max(val_label(:,end,:));
correct_num = correct_num + length(find(estimated_labels == true_labels));
% training acc
val_sample = samples(:,:,start_idx:end_idx);
val_label = labels(:,:,start_idx:end_idx);
lstm_core_v52(val_sample, 1);
[value, estimated_labels] = max(mem.net_out(:,end,:,1)+mem.net_out(:,end,:,2));
[vv1, pos1] = max(mem.net_out(:,25:end,:,1));
[vv2, pos2] = max(mem.net_out(:,25:end,:,2));
tt = sum((pos1 == pos2));
estimated_labels(tt < outlier_thres) = -1; % to compute the real accuracy, apply outliear rejection first
[value, true_labels] = max(val_label(:,end,:));
train_correct_num = train_correct_num + length(find(estimated_labels == true_labels));
end
acc = correct_num / size(test_samples, 3);
train_acc = train_correct_num / size(test_samples, 3);
outlier_acc = outlier_currect_num / size(test_samples, 3);
fprintf('\nepoc %d, training avg cost: %f, train_acc: %.2f%%, val_acc: %.2f%%, outlier_acc: %.2f%%\n', epoc, cost_avg, train_acc*100, acc*100, outlier_acc*100);
model = config;
save(strcat('results/speaker-naming/face_audio/', num2str(epoc), '.mat'), '-v7.3', 'model');
cost_avg = 0;
end
end
end
end