/
main.go
203 lines (172 loc) · 4.95 KB
/
main.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
package main
import "C"
import (
"flag"
"fmt"
"image/jpeg"
"io"
"io/ioutil"
"log"
"math/rand"
"os"
"runtime/pprof"
"gonum.org/v1/gonum/blas/gonum"
T "gorgonia.org/gorgonia"
"gorgonia.org/gorgonia/examples/mnist"
"gorgonia.org/tensor"
)
var cpuprofile = flag.String("cpuprofile", "", "write cpu profile to file")
var memprofile = flag.String("memprofile", "", "write memory profile to this file")
var dataset = flag.String("dataset", "dev", "which data set to train on? Valid options: \"train\" or \"dev\"")
var bs = flag.Int("bs", 1, "training batch size (doesn't affect sgd batch size)")
var ptEpoch = flag.Int("pt", 16, "pretraining epoch")
var ftEpoch = flag.Int("ft", 40, "finetuning epoch")
var viz = flag.Int("viz", 0, "Visualize which layer?")
var save = flag.String("save", "", "Save file as")
var verbose = flag.Bool("v", false, "Verbose?")
var trainingWriter io.Writer
var trainingLog *log.Logger
var dt tensor.Dtype = tensor.Float64
func init() {
var err error
if trainingWriter, err = os.OpenFile("training.viz", os.O_CREATE|os.O_TRUNC|os.O_WRONLY, 0644); err != nil {
log.Fatal(err)
}
trainingLog = log.New(trainingWriter, "", log.Ltime|log.Lmicroseconds)
}
func predictBatch(logprobs tensor.Tensor, batchSize int) (guesses []int, err error) {
var argmax tensor.Tensor
if batchSize == 1 {
argmax, err = tensor.Argmin(logprobs, 0)
} else {
argmax, err = tensor.Argmin(logprobs, 1)
}
if err != nil {
return nil, err
}
guesses = argmax.Data().([]int)
return
}
func makeTargets(targets tensor.Tensor) []int {
ys := make([]int, targets.Shape()[0])
ys = ys[:0]
for i := 0; i < targets.Shape()[0]; i++ {
ysl, _ := targets.Slice(T.S(i))
raw := ysl.Data().([]float64)
for i, v := range raw {
if v == 0.9 {
ys = append(ys, i)
break
}
}
}
return ys
}
func verboseLog(format string, attrs ...interface{}) {
if *verbose {
log.Printf(format, attrs...)
}
}
func main() {
flag.Parse()
rand.Seed(1337)
loc := "../testdata/mnist/"
trainOn := *dataset
inputs, targets, err := mnist.Load(trainOn, loc, dt)
if err != nil {
log.Fatal(err)
}
size := inputs.Shape()[0]
inputSize := 784
outputSize := 10
hiddenSizes := []int{1000, 1000, 1000}
layers := len(hiddenSizes)
corruptions := []float64{0.1, 0.2, 0.3}
batchSize := *bs
pretrainEpoch := *ptEpoch
finetuneEpoch := *ftEpoch
deets := `Stacked Denoising AutoEncoder
==============================
Train on: %v
Training Size: %v
Hidden Sizes: %v
Corruptions: %v
Batch Size: %v
Pretraining Epoch: %v
Finetuning Epoch: %v
BLAS used: %v
`
fmt.Printf(deets, trainOn, size, hiddenSizes, corruptions, batchSize, pretrainEpoch, finetuneEpoch, T.WhichBLAS())
g := T.NewGraph()
sda := NewStackedDA(g, batchSize, size, inputSize, outputSize, layers, hiddenSizes, corruptions)
// start CPU profiling before we start training
if *cpuprofile != "" {
f, err := os.Create(*cpuprofile)
if err != nil {
log.Fatal(err)
}
pprof.StartCPUProfile(f)
defer pprof.StopCPUProfile()
}
verboseLog("Pretraining...")
for i := 0; i < pretrainEpoch; i++ {
if err = sda.Pretrain(inputs, i); err != nil {
ioutil.WriteFile("fullGraph_err.dot", []byte(g.ToDot()), 0644)
log.Fatalf("i: %d err :%v", i, err)
}
}
// Because for now LispMachine doesn't support batched BLAS
verboseLog("Starting to finetune now")
T.Use(gonum.Implementation{})
ys := makeTargets(targets)
for i := 0; i < finetuneEpoch; i++ {
if err = sda.Finetune(inputs, ys, i); err != nil {
log.Fatal(err)
}
}
// save model
if *save != "" {
if err = sda.Save(*save); err != nil {
log.Fatal(err)
}
}
// Visualize
visualizeLayer := *viz
verboseLog("Visualizing %dth layer", visualizeLayer)
finalWeights := sda.autoencoders[visualizeLayer].w.Value().(tensor.Tensor).Clone().(tensor.Tensor) // TODO: rewrite this plz, it's hard to understand
finalWeights.T()
finalWeights.Transpose()
for i := 0; i < finalWeights.Shape()[0]; i++ {
rowT, _ := finalWeights.Slice(T.S(i))
row := rowT.Data().([]float64)
img := visualizeRow(row)
f, _ := os.OpenFile(fmt.Sprintf("images/%d.jpg", i), os.O_CREATE|os.O_WRONLY|os.O_TRUNC, 0644)
jpeg.Encode(f, img, &jpeg.Options{jpeg.DefaultQuality})
f.Close()
}
/* PREDICTION TIME */
// here I'm using the test dataset as prediction.
// in real life you should probably be doing crossvalidations and whatnots
// but in this demo, we're going to skip all those
verboseLog("pred")
testX, testY, err := mnist.Load("test", loc, dt)
if err != nil {
log.Fatal(err)
}
var one, correct, lp tensor.Tensor
if one, err = testX.Slice(T.S(0, batchSize)); err != nil {
log.Fatal(err)
}
if correct, err = testY.Slice(T.S(0, batchSize)); err != nil {
log.Fatal(err)
}
correctYs := makeTargets(correct)
var predictions []int
if lp, err = sda.Forwards(one); err != nil {
log.Fatal(err)
}
if predictions, err = predictBatch(lp, batchSize); err != nil {
log.Fatal(err)
}
fmt.Printf("Correct: \n%+v. \nPredicted: %v. \nLogprobs: \n%+#3.3s", correctYs, predictions, lp)
}