-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.go
212 lines (163 loc) · 5.19 KB
/
train.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
209
210
211
212
package main
import (
"fmt"
"io/ioutil"
"log"
"os"
"time"
"mpi"
"github.com/dcu/godl"
"github.com/fatih/color"
"gorgonia.org/gorgonia"
"gorgonia.org/tensor"
)
// TrainOpts are the options to train the model
type TrainOpts struct {
Epochs int
BatchSize int
// DevMode detects common issues like exploding and vanishing gradients at the cost of performance
DevMode bool
WriteGraphFileTo string
// WithLearnablesHeatmap writes images representing heatmaps for the weights. Use it to debug.
WithLearnablesHeatmap bool
// Solver defines the solver to use. It uses gorgonia.AdamSolver by default if none is passed
Solver gorgonia.Solver
// ValidateEvery indicates the number of epochs to run before running a validation. Defaults 1 (every epoch)
ValidateEvery int
CostObserver func(epoch int, totalEpoch, batch int, totalBatch int, cost float32)
ValidationObserver func(confMat ConfusionMatrix, cost float32)
MatchTypeFor func(predVal, targetVal []float32) MatchType
CostFn godl.CostFn
}
func (o *TrainOpts) setDefaults() {
if o.Epochs == 0 {
o.Epochs = 10
}
if o.BatchSize == 0 {
o.BatchSize = 1024
}
if o.ValidateEvery == 0 {
o.ValidateEvery = 1
}
if o.CostFn == nil {
panic("CostFN must be set")
}
}
// Train trains the model with the given data
func Train(newComm *mpi.Communicator, m *Model, module godl.Module, trainX, trainY, validateX, validateY tensor.Tensor, opts TrainOpts) error {
opts.setDefaults()
if opts.DevMode {
warn("Start training in dev mode")
defer func() {
if err := recover(); err != nil {
graphFileName := "graph.dot"
log.Printf("panic triggered, dumping the model graph to: %v", graphFileName)
_ = ioutil.WriteFile(graphFileName, []byte(m.trainGraph.ToDot()), 0644)
panic(err)
}
}()
}
if opts.WithLearnablesHeatmap {
warn("Heatmaps will be stored in: %s", heatmapPath)
_ = os.RemoveAll(heatmapPath)
}
dl := NewDataLoader(trainX, trainY, DataLoaderOpts{
BatchSize: opts.BatchSize,
Shuffle: false,
})
xShape := append(tensor.Shape{opts.BatchSize}, trainX.Shape()[1:]...)
x := gorgonia.NewTensor(m.trainGraph, trainX.Dtype(), trainX.Shape().Dims(), gorgonia.WithShape(xShape...), gorgonia.WithName("x"))
y := gorgonia.NewMatrix(m.trainGraph, trainY.Dtype(), gorgonia.WithShape(opts.BatchSize, trainY.Shape()[1]), gorgonia.WithName("y"))
result := module.Forward(x)
// if opts.WriteGraphFileTo != "" {
// m.WriteSVG(opts.WriteGraphFileTo)
// }
var (
costVal gorgonia.Value
predVal gorgonia.Value
)
{
cost := opts.CostFn(result, y)
gorgonia.Read(cost, &costVal)
gorgonia.Read(result[0], &predVal)
if _, err := gorgonia.Grad(cost, m.Learnables()...); err != nil {
return fmt.Errorf("error calculating gradient: %w", err)
}
}
validationGraph := m.trainGraph.SubgraphRoots(result[0])
validationGraph.RemoveNode(y)
m.evalGraph = validationGraph
vmOpts := []gorgonia.VMOpt{
gorgonia.BindDualValues(m.learnables...),
}
if opts.DevMode {
vmOpts = append(
vmOpts,
gorgonia.TraceExec(),
gorgonia.WithNaNWatch(),
gorgonia.WithInfWatch(),
)
}
vm := gorgonia.NewTapeMachine(m.trainGraph, vmOpts...)
if opts.Solver == nil {
info("defaulting to RMS solver")
opts.Solver = gorgonia.NewRMSPropSolver(gorgonia.WithBatchSize(float64(opts.BatchSize)))
}
defer vm.Close()
startTime := time.Now()
for i := 0; i < opts.Epochs; i++ {
for dl.HasNext() {
xVal, yVal := dl.Next()
err := gorgonia.Let(x, xVal)
if err != nil {
fatal("error assigning x: %v", err)
}
err = gorgonia.Let(y, yVal)
if err != nil {
fatal("error assigning y: %v", err)
}
if err = vm.RunAll(); err != nil {
fatal("Failed at epoch %d, batch %d. Error: %v", i, dl.CurrentBatch, err)
}
if opts.WithLearnablesHeatmap {
m.saveHeatmaps(i, dl.CurrentBatch, dl.opts.BatchSize, dl.FeaturesShape[0])
}
if err = opts.Solver.Step(gorgonia.NodesToValueGrads(m.learnables)); err != nil {
fatal("Failed to update nodes with gradients at epoch %d, batch %d. Error %v", i, dl.CurrentBatch, err)
}
if opts.CostObserver != nil {
opts.CostObserver(i, opts.Epochs, dl.CurrentBatch, dl.Batches, costVal.Data().(float32))
} else {
color.Yellow(" Epoch %d %d | cost %v (%v)\n", i, dl.CurrentBatch, costVal, time.Since(startTime))
}
m.PrintWatchables()
vm.Reset()
}
///////////////////////////////////////////////////////////
for index, item := range m.learnables {
weight := item.Value().Data().([]float64)
end := WeightsLengthArray[index]
if index == 0 {
for i := 0; i < end; i++ {
Weightscomposed[i] = weight[i]
}
} else {
for i := 0; i < end; i++ {
Weightscomposed[WeightsLengthAccu[index-1]+i] = weight[i]
}
}
}
/////////////////////////////////////////////////////////////
newComm.SendFloat64s(Weightscomposed, 0, newComm.Rank())
Weightscomposed, _ = newComm.RecvFloat64s(0, newComm.Rank())
dl.Reset()
if i%opts.ValidateEvery == 0 {
err := Validate(m, x, y, costVal, predVal, validateX, validateY, opts)
if err != nil {
color.Red("Failed to run validation on epoch %v: %v", i, err)
}
color.Yellow(" Epoch %d | cost %v (%v)\n", i, costVal, time.Since(startTime))
}
}
return nil
}