-
Notifications
You must be signed in to change notification settings - Fork 1
/
vae_test.go
208 lines (186 loc) · 5.23 KB
/
vae_test.go
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
package gortex
import (
"fmt"
"math/rand"
"testing"
"time"
"math"
"encoding/json"
"io/ioutil"
"os"
"github.com/vseledkin/gortex/assembler"
)
func TestCharVae(t *testing.T) {
// maintain random seed
rand.Seed(time.Now().UnixNano())
trainFile := "100000_32.txt"
modelName := "CharVAE"
dic, e := LoadDictionary(modelName + ".dic")
if e != nil {
dic, e = DictionaryFromFile(trainFile, CharSplitter{})
if e != nil {
t.Fatal(e)
}
e = SaveDictionary(modelName+".dic", dic)
if e != nil {
t.Fatal(e)
}
}
hidden_size := 128
embedding_size := 128
z_size := 128
fmt.Printf("%s\n", dic)
fmt.Printf("Dictionary has %d tokens\n", dic.Len())
//optimizer := NewOptimizer(OpOp{Method: WINDOWGRAD, LearningRate: 0.001, Momentum: DefaultMomentum, Clip: 5})
optimizer := NewOptimizer(OpOp{Method: WINDOWGRAD, LearningRate: 0.0003, Momentum: DefaultMomentum, Clip: 4})
LookupTable := RandMat(embedding_size, dic.Len()) // Lookup Table matrix
encoder := MakeOutputlessLSTM(embedding_size, hidden_size)
encoder.ForgetGateTrick(2.0)
vae := MakeVae(hidden_size, z_size)
decoder := MakeLSTM(z_size, hidden_size, dic.Len())
decoder.ForgetGateTrick(2.0)
model := make(map[string]*Matrix)
// define model parameters
for k, v := range encoder.GetParameters("Encoder") {
model[k] = v
}
model["LookupTable"] = LookupTable
for k, v := range vae.GetParameters("VAE") {
model[k] = v
}
for k, v := range decoder.GetParameters("Decoder") {
model[k] = v
}
if _, err := os.Stat(modelName); err == nil {
loadedModel, e := LoadModel(modelName)
if e != nil {
t.Fatal(e)
}
e = SetParameters(model, loadedModel)
if e != nil {
t.Fatal(e)
}
}
count := 0
ma_cost := NewMovingAverage(512)
ma_kld_cost := NewMovingAverage(512)
ma_mean := NewMovingAverage(512)
ma_dev := NewMovingAverage(512)
//batch_size := 8
var e_steps, d_steps float32
batch_size := 32
kld_scale := float32(0.0175)
threads := 4
license := make(chan struct{}, threads)
for i := 0; i < threads; i++ {
license <- struct{}{}
}
CharSampleVisitor(trainFile, 1, CharSplitter{}, dic, func(epoch int, x []uint) {
<-license
count++
go func(count int) {
// read sample
sample := ""
for i := range x {
sample += dic.TokenByID(x[i])
}
G := &Graph{NeedsBackprop: true}
ht := Mat(hidden_size, 1).OnesAs() // vector of Zeros
ct := Mat(hidden_size, 1).OnesAs() // vector of Zeros
// encode sequence into z
for i := range x {
e_steps++
embedding := G.Lookup(LookupTable, int(x[i]))
ht, ct = encoder.Step(G, embedding, ht, ct)
}
distribution, mean, logvar := vae.Step(G, ht)
// estimate KLD
kld := vae.KLD(G, kld_scale, mean, logvar)
// decode sequence from z
var logit *Matrix
cost := float32(0)
decoded := ""
ht = Mat(hidden_size, 1).OnesAs() // vector of Zeros
ct = Mat(hidden_size, 1).OnesAs() // vector of Zeros
for i := range x {
d_steps++
ht, ct, logit = decoder.Step(G, distribution, ht, ct)
c, _ := G.Crossentropy(logit, x[i])
cid, _ := MaxIV(Softmax(logit))
decoded += dic.TokenByID(cid)
cost += c
}
cost /= float32(len(x))
G.Backward()
if count%batch_size == 0 && count > 0 {
//ScaleGradient(encoderModel, 1/e_steps)
//ScaleGradient(decoderModel, 1/d_steps)
optimizer.Step(model)
d_steps = 0
e_steps = 0
}
count++
//if count > 0 && count%batch_size == 0 {
//d_cost /= d_steps
//g_cost /= g_steps
ma_kld_cost.Add(kld)
ma_cost.Add(cost)
m := assembler.Sum(mean.W) / float32(len(mean.W))
dev := Exp(logvar)
v := assembler.Sum(dev.W) / float32(len(dev.W))
ma_mean.Add(m)
ma_dev.Add(float32(math.Sqrt(float64(v))))
//if sample != decoded {
//}
avg_cost := ma_cost.Avg()
avg_mean := ma_mean.Avg()
avg_dev := ma_dev.Avg()
if count%10000 == 0 {
SaveModel(modelName, model)
}
if count%500 == 0 {
fmt.Printf("\ndecoded: [%s]\n", decoded)
fmt.Printf("encoded: [%s]\n", sample)
fmt.Printf("epoch: %d step: %d loss: %f lr: %f kld_scale: %f\n", epoch, count, avg_cost, optimizer.LearningRate, kld_scale)
fmt.Printf("mean: %f dev: %f kld: %f\n", avg_mean, avg_dev, ma_kld_cost.Avg())
fmt.Printf("dev: %#v\n", dev.W[:10])
if avg_cost < 1 && kld_scale < 1.0 {
f, e := os.Open("kld_ch.json")
if e != nil {
t.Error(e)
}
cob, e := ioutil.ReadAll(f)
if e != nil {
t.Error(e)
}
var co struct{ GateInc, Lr float32 }
json.Unmarshal(cob, &co)
f.Close()
kld_scale += co.GateInc
optimizer.LearningRate = co.Lr
}
// interpolate between two pints
z1 := RandMat(vae.z_size, 1)
z2 := RandMat(vae.z_size, 1)
gg := &Graph{NeedsBackprop: false}
fmt.Printf("Interpolation\n")
for a := float32(0.0); a <= 1; a += 0.1 {
z := gg.Add(gg.MulConstant(1.0-a, z1), gg.MulConstant(a, z2))
decoded := ""
ht = Mat(hidden_size, 1).OnesAs() // vector of Zeros
ct = Mat(hidden_size, 1).OnesAs() // vector of Zeros
for range make([]struct{}, 32) {
ht, ct, logit = decoder.Step(gg, z, ht, ct)
cid, _ := MaxIV(Softmax(logit))
decoded += dic.TokenByID(cid)
}
fmt.Printf("%0.2f sentence: %s\n", a, decoded)
}
}
license <- struct{}{}
}(count)
})
for i := 0; i < threads; i++ {
<-license
}
}