/
ladder_network.js
205 lines (185 loc) · 6.99 KB
/
ladder_network.js
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
import NeuralNetwork from './neuralnetwork.js'
/**
* Ladder network
*/
export default class LadderNetwork {
// https://www.kabuku.co.jp/developers/semi-supervised_learning_with_ladder_networks
/**
* @param {number[]} hidden_sizes Sizes of hidden layers
* @param {number[]} lambdas Regularization parameters
* @param {string} activation Activation name
* @param {string} optimizer Optimizer of the network
*/
constructor(hidden_sizes, lambdas, activation, optimizer) {
this._hidden_sizes = hidden_sizes
this._lambdas = lambdas
this._activation = activation
this._optimizer = optimizer
this._noise_std = Array(this._hidden_sizes.length + 2).fill(0.001)
this._model = null
this._classes = null
this._optimizer = optimizer
this._epoch = 0
}
/**
* Epoch
*
* @type {number}
*/
get epoch() {
return this._epoch
}
_build() {
const hidden_sizes = [...this._hidden_sizes, this._classes.length]
this._layers = [
{ type: 'input', name: 'x' },
{ type: 'identity', name: 'z_0' },
{ type: 'random', size: 'x', variance: this._noise_std[0] ** 2, name: 'noize' },
{ type: 'add', input: ['x', 'noize'], name: 'zt_0' },
{ type: 'mean', input: 'x', name: 'mean_0' },
{ type: 'std', input: 'x', name: 'std_0' },
{ type: 'concat', input: ['z_0', 'zt_0'], axis: 0 },
]
for (let i = 0; i < hidden_sizes.length; i++) {
const k = i + 1
this._layers.push(
{ type: 'full', out_size: hidden_sizes[i] },
{ type: 'split', size: 2, axis: 0, name: `zpre_${k}` },
{ type: 'batch_normalization', input: `zpre_${k}[1]`, name: `ztbn_${k}` },
{ type: 'random', size: `ztbn_${k}`, variance: this._noise_std[k] ** 2, name: `noize_${k}` },
{ type: 'add', input: [`ztbn_${k}`, `noize_${k}`], name: `zt_${k}` },
{ type: 'mean', input: `zpre_${k}[0]`, name: `mean_${k}` },
{ type: 'std', input: `zpre_${k}[0]`, name: `std_${k}` },
{ type: 'batch_normalization', input: `zpre_${k}[0]`, name: `z_${k}` },
{ type: 'concat', input: [`z_${k}`, `zt_${k}`], axis: 0, name: `zcon_${k}` },
{ type: 'variable', size: [1, hidden_sizes[i]], name: `b_${k}` },
{ type: 'add', input: [`zcon_${k}`, `b_${k}`], name: `hadd_${k}` },
{ type: 'variable', size: [1, hidden_sizes[i]], name: `g_${k}` },
{ type: 'mult', input: [`hadd_${k}`, `g_${k}`] },
{ type: this._activation }
)
}
this._layers.push(
{ type: 'split', size: 2, axis: 0, name: 'reduced' },
{ type: 'softmax', input: 'reduced[0]', name: 'predict' },
{ type: 'identity', input: 'reduced[1]' }
)
for (let k = hidden_sizes.length; k >= 0; k--) {
const size = k === 0 ? 'x' : hidden_sizes[k - 1]
if (k === hidden_sizes.length) {
this._layers.push({ type: 'batch_normalization', name: `u_${k}` })
} else {
this._layers.push(
{ type: 'full', out_size: size, input: `zh_${k + 1}` },
{ type: 'batch_normalization', name: `u_${k}` }
)
}
this._layers.push(
{ type: 'variable', size: `mean_${k}`, name: `a1_${k}` },
{ type: 'variable', size: `mean_${k}`, name: `a2_${k}` },
{ type: 'variable', size: `mean_${k}`, name: `a3_${k}` },
{ type: 'variable', size: `mean_${k}`, name: `a4_${k}` },
{ type: 'variable', size: `mean_${k}`, name: `a5_${k}` },
{ type: 'variable', size: `mean_${k}`, name: `a6_${k}` },
{ type: 'variable', size: `mean_${k}`, name: `a7_${k}` },
{ type: 'variable', size: `mean_${k}`, name: `a8_${k}` },
{ type: 'variable', size: `mean_${k}`, name: `a9_${k}` },
{ type: 'variable', size: `mean_${k}`, name: `a0_${k}` },
{ type: 'mult', input: [`a2_${k}`, `u_${k}`], name: `m1_${k}` },
{ type: 'add', input: [`a3_${k}`, `m1_${k}`] },
{ type: 'sigmoid', name: `m2_${k}` },
{ type: 'mult', input: [`a1_${k}`, `m2_${k}`], name: `m3_${k}` },
{ type: 'mult', input: [`a4_${k}`, `u_${k}`], name: `m4_${k}` },
{ type: 'add', input: [`a5_${k}`, `m3_${k}`, `m4_${k}`], name: `m_${k}` },
{ type: 'mult', input: [`a7_${k}`, `u_${k}`], name: `v1_${k}` },
{ type: 'add', input: [`a8_${k}`, `v1_${k}`] },
{ type: 'sigmoid', name: `v2_${k}` },
{ type: 'mult', input: [`a6_${k}`, `v2_${k}`], name: `v3_${k}` },
{ type: 'mult', input: [`a9_${k}`, `u_${k}`], name: `v4_${k}` },
{ type: 'add', input: [`a0_${k}`, `v3_${k}`, `v4_${k}`], name: `v_${k}` },
{ type: 'sub', input: [`zt_${k}`, `m_${k}`], name: `zh1_${k}` },
{ type: 'mult', input: [`zh1_${k}`, `v_${k}`], name: `zh2_${k}` },
{ type: 'add', input: [`zh2_${k}`, `m_${k}`], name: `zh_${k}` },
{ type: 'sub', input: [`zh_${k}`, `mean_${k}`], name: `zhbn1_${k}` },
{ type: 'div', input: [`zhbn1_${k}`, `std_${k}`], name: `zhbn_${k}` }
)
}
this._layers.push(
{ type: 'output', name: 'reconstruct' },
{ type: 'supervisor', name: 't' },
{ type: 'sub', input: ['t', 'predict'] },
{ type: 'square' },
{ type: 'sum', axis: 1 },
{ type: 'mean', name: 'cc0' },
{ type: 'input', name: 'is_labeled' },
{ type: 'mult', input: ['cc0', 'is_labeled'], name: 'cc' }
)
const costs = ['cc']
for (let i = 0; i <= hidden_sizes.length; i++) {
this._layers.push(
{ type: 'sub', input: [`z_${i}`, `zhbn_${i}`] },
{ type: 'square' },
{ type: 'sum', axis: 1 },
{ type: 'mean', name: `cl_${i}` },
{ type: 'mult', input: [this._lambdas[i], `cl_${i}`], name: `wcl_${i}` }
)
costs.push(`wcl_${i}`)
}
this._layers.push({ type: 'add', input: costs })
return NeuralNetwork.fromObject(this._layers, null, 'adam')
}
/**
* Fit model.
*
* @param {Array<Array<number>>} train_x Training data
* @param {(* | null)[]} train_y Target values
* @param {number} iteration Iteration count
* @param {number} rate Learning rate
* @param {number} batch Batch size
* @returns {{labeledLoss: number, unlabeledLoss: number}} Loss value
*/
fit(train_x, train_y, iteration, rate, batch) {
if (!this._classes) {
const classes = new Set()
for (let i = 0; i < train_y.length; i++) {
if (train_y[i] != null) {
classes.add(train_y[i])
}
}
this._classes = [...classes]
}
if (!this._model) {
this._model = this._build()
}
let lLoss = null
let uLoss = null
const labeled_x = train_x.filter((v, i) => train_y[i] != null)
if (labeled_x.length > 0) {
const labeled_y = train_y.filter(v => v != null)
const y = labeled_y.map(v => {
const yi = Array(this._classes.length).fill(0)
yi[this._classes.indexOf(v)] = 1
return yi
})
lLoss = this._model.fit({ x: labeled_x, is_labeled: [[1]] }, y, iteration, rate, batch)
}
const unlabeled_x = train_x.filter((v, i) => train_y[i] == null)
if (unlabeled_x.length > 0) {
uLoss = this._model.fit({ x: unlabeled_x, is_labeled: [[0]] }, [[0]], iteration, rate, batch)
}
this._epoch += iteration
return { labeledLoss: lLoss?.[0], unlabeledLoss: uLoss?.[0] }
}
/**
* Returns predicted values.
*
* @param {Array<Array<number>>} x Sample data
* @returns {*[]} Predicted values
*/
predict(x) {
return this._model
.calc({ x, is_labeled: null }, null, ['predict'])
.predict.argmax(1)
.value.map(v => this._classes[v])
}
}