-
Notifications
You must be signed in to change notification settings - Fork 3
/
sentiment_analysis.lua
160 lines (107 loc) · 4.27 KB
/
sentiment_analysis.lua
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
require 'mobdebug'.start()
require 'nn'
require 'nngraph'
require 'optim'
require 'Embedding'
local model_utils=require 'model_utils'
require 'table_utils'
nngraph.setDebug(true)
function calc_f1(prediction, target)
local f1_accum = 0
local precision_accum = 0
local recall_accum = 0
for c = 1, 5 do
local p = torch.eq(prediction, c):double()
local t = torch.eq(target, c):double()
local true_positives = torch.mm(t:t(),p)[1][1]
p = torch.eq(prediction, c):double()
t = torch.ne(target, c):double()
local false_positives = torch.mm(t:t(),p)[1][1]
p = torch.ne(prediction, c):double()
t = torch.eq(target, c):double()
local false_negatives = torch.mm(t:t(),p)[1][1]
local precision = true_positives / (true_positives + false_positives)
local recall = true_positives / (true_positives + false_negatives)
local f1_score = 2 * precision * recall / (precision + recall)
f1_accum = f1_accum + f1_score
precision_accum = precision_accum + precision
recall_accum = recall_accum + recall
end
return {f1_accum / 5, precision_accum / 5, recall_accum / 5}
end
inv_vocabulary_en = table.load('inv_vocabulary_en')
vocabulary_en = table.load('vocabulary_en')
features_train, labels_train, text_train = unpack(torch.load('sentiment_train.t7'))
features_dev, labels_dev, text_dev= unpack(torch.load('sentiment_dev.t7'))
assert (features_train:size(1) == labels_train:size(1))
assert (features_train:size(1) == #text_train)
batch_size = 3000
data_index = 1
n_data = features_train:size(1)
function gen_batch()
start_index = data_index
end_index = math.min(n_data, start_index + batch_size - 1)
if end_index == n_data then
data_index = 1
else
data_index = data_index + batch_size
end
features = features_train[{{start_index, end_index}, {}}]
labels = labels_train[{{start_index, end_index}}]
text = text_train[start_index]
text_readable = {}
for i, word in pairs(text) do
text_readable[#text_readable + 1] = vocabulary_en[word]
end
text_readable = table.concat(text_readable, ' ')
return features, labels, text_readable
end
x_raw = nn.Identity()()
x = nn.Linear(10, 20)(x_raw)
x = nn.Tanh()(x)
x = nn.Linear(20, 10)(x)
x = nn.Tanh()(x)
x = nn.Linear(10, 5)(x)
x = nn.Tanh()(x)
x = nn.LogSoftMax()(x)
m = nn.gModule({x_raw}, {x})
local params, grad_params = model_utils.combine_all_parameters(m)
params:uniform(-0.08, 0.08)
criterion = nn.ClassNLLCriterion()
function feval(x_arg)
if x_arg ~= params then
params:copy(x_arg)
end
grad_params:zero()
local loss = 0
features, labels, text_readable = gen_batch()
------------------- forward pass -------------------
prediction = m:forward(features)
loss_m = criterion:forward(prediction, labels)
loss = loss + loss_m
-- complete reverse order of the above
dprediction = criterion:backward(prediction, labels)
dfeatures = m:backward(features, dprediction)
return loss, grad_params
end
optim_state = {learningRate = 1e-2}
for i = 1, 1000000 do
local _, loss = optim.adam(feval, params, optim_state)
if i % 100 == 0 then
local loss_train = loss[1]
local _, predicted_class = prediction:max(2)
local f1_score_train, precision_train, recall_train = unpack(calc_f1(predicted_class, torch.reshape(labels, predicted_class:size(1), predicted_class:size(2))))
local features = features_dev[{{}, {}}]
local labels = labels_dev[{{}}]
local prediction = m:forward(features)
local _, predicted_class = prediction:max(2)
local loss_dev = criterion:forward(prediction, labels)
local f1_score_dev, precision_dev, recall_dev = unpack(calc_f1(predicted_class, torch.reshape(labels, predicted_class:size(1), predicted_class:size(2))))
print(string.format("train set: loss = %6.8f, f1_score = %6.8f, precision = %6.8f, recall = %6.8f, grad_params:norm() = %6.4e, params:norm() = %6.4e", loss_train, f1_score_train, precision_train, recall_train, grad_params:norm(), params:norm()))
print(string.format("dev set: loss = %6.8f, f1_score = %6.8f, precision = %6.8f, recall = %6.8f", loss_dev, f1_score_dev, precision_dev, recall_dev))
if precision_dev > 0.3 then
print('success')
end
end
end
pass_dummy = 1