-
Notifications
You must be signed in to change notification settings - Fork 1
/
adam.go
423 lines (376 loc) · 8.98 KB
/
adam.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
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
package trainer
import (
"errors"
"fmt"
"github.com/dereklstinson/half"
"github.com/dereklstinson/gocunets/devices/gpu/nvidia"
"github.com/dereklstinson/gocunets/devices/gpu/nvidia/cudnn"
"github.com/dereklstinson/gocunets/layers"
gocudnn "github.com/dereklstinson/gocudnn"
"github.com/dereklstinson/gocudnn/gocu"
"github.com/dereklstinson/gocudnn/xtra"
)
var debuggingadam bool
//DebuggingAdam is for debugging purposes
func DebuggingAdam() {
debuggingadam = true
}
//Adam is a struct that does the holds the params for adam optimization
type Adam struct {
dtype gocudnn.DataType
loss1h []half.Float16
loss2h []half.Float16
loss1 []float32
loss2 []float32
goptr1 *gocu.Wrapper
goptr2 *gocu.Wrapper
gpuloss1 *nvidia.Malloced
gpuloss2 *nvidia.Malloced
gsum *nvidia.Malloced
xsum *nvidia.Malloced
trainer *xtra.TrainerD
params xtra.TrainingParams
regparams xtra.RegParams
dims []int32
counter uint64
}
const defaultadambeta1 = 0.9
const defaultadambeta2 = 0.999
const defaultadameps = float32(1e-8)
const defaultadamrate = .001
//SetTrainingMem creates the training mem for the adam trainer
func (a *Adam) SetTrainingMem(han *cudnn.Handler, w *layers.Tensor) error {
// /a.freememer()
_, dtype, dims, err := w.Properties()
if err != nil {
if debuggingadam {
panic(err)
}
return err
}
a.dims = dims
//DeFault := gocudnn.MemcpyKindFlag{}.Default()
var dflg gocudnn.DataType
switch dtype {
case dflg.Float():
a.dtype.Float()
a.loss1 = make([]float32, 1)
a.loss2 = make([]float32, 1)
a.goptr1, err = gocu.MakeGoMem(a.loss1)
if err != nil {
if debuggingadam {
panic(err)
}
return err
}
a.goptr2, err = gocu.MakeGoMem(a.loss2)
if err != nil {
if debuggingadam {
panic(err)
}
return err
}
//asize := dimsize()
sizet := gocudnn.FindSizeTfromVol(dims, dtype)
a.gsum, err = nvidia.MallocGlobal(han, sizet)
if err != nil {
if debuggingadam {
panic(err)
}
return err
}
a.xsum, err = nvidia.MallocGlobal(han, sizet)
if err != nil {
if debuggingadam {
panic(err)
}
return err
}
err = a.xsum.SetAll(0)
if err != nil {
if debuggingadam {
fmt.Println("Dims are", dims)
fmt.Println("Adress for a.xsum,and a.gsum", a.xsum, a.gsum)
fmt.Println("a.xsum Cudasize", a.gsum.SIB())
panic(err)
}
}
err = a.gsum.SetAll(0)
if err != nil {
if debuggingadam {
fmt.Println("a.gsum Cudasize", a.gsum.SIB())
panic(err)
}
}
a.gpuloss1, err = nvidia.MallocGlobal(han, 4)
if err != nil {
if debuggingadam {
panic(err)
}
return err
}
a.gpuloss2, err = nvidia.MallocGlobal(han, 4)
if err != nil {
if debuggingadam {
panic(err)
}
return err
}
case dflg.Half():
a.dtype.Half()
a.loss1h = make([]half.Float16, 1)
a.loss2h = make([]half.Float16, 1)
a.goptr1, err = gocu.MakeGoMem(a.loss1h)
if err != nil {
if debuggingadam {
panic(err)
}
return err
}
a.goptr2, err = gocu.MakeGoMem(a.loss2h)
if err != nil {
if debuggingadam {
panic(err)
}
return err
}
//asize := dimsize()
sizet := gocudnn.FindSizeTfromVol(dims, dtype)
a.gsum, err = nvidia.MallocGlobal(han, sizet)
if err != nil {
if debuggingadam {
panic(err)
}
return err
}
a.xsum, err = nvidia.MallocGlobal(han, sizet)
if err != nil {
if debuggingadam {
panic(err)
}
return err
}
err = a.xsum.SetAll(0)
if err != nil {
if debuggingadam {
fmt.Println("Dims are", dims)
fmt.Println("Adress for a.xsum,and a.gsum", a.xsum, a.gsum)
fmt.Println("a.xsum Cudasize", a.gsum.SIB())
panic(err)
}
}
err = a.gsum.SetAll(0)
if err != nil {
if debuggingadam {
fmt.Println("a.gsum Cudasize", a.gsum.SIB())
panic(err)
}
}
a.gpuloss1, err = nvidia.MallocGlobal(han, 2)
if err != nil {
if debuggingadam {
panic(err)
}
return err
}
a.gpuloss2, err = nvidia.MallocGlobal(han, 2)
if err != nil {
if debuggingadam {
panic(err)
}
return err
}
default:
return errors.New("Only Float datatype supported at the moment")
}
return nil
}
//Dims returns the dims of the training parameter holders
func (a *Adam) Dims() []int32 {
return a.dims
}
//UpdateWeights updates the weights
func (a *Adam) UpdateWeights(handle *cudnn.Handler, dw, w *layers.Tensor, batchsize, counter int) error {
var err error
err = handle.Sync()
if err != nil {
return err
}
flg := a.dtype
switch a.dtype {
case flg.Float():
a.SetBatch(float32(batchsize))
err = a.trainer.L1L2Regularization(handle.XHandle(), dw.TD(), dw, w, a.gpuloss1, a.gpuloss2, a.regparams)
if err != nil {
return err
}
err = handle.Sync()
if err != nil {
return err
}
if debuggingadam {
fmt.Println(a.params)
//gsum, err := gocudnn.GetStringer(w.TD(), a.gsum)
//if err != nil {
// panic(err)
//}
//fmt.Println("GsumBefore", gsum)
//xsum, err := gocudnn.GetStringer(w.TD(), a.xsum)
//if err != nil {
// panic(err)
//}
//fmt.Println("xsum before", xsum)
fmt.Println("Before Update")
dw.TogglePrintValueForStringer()
fmt.Println(dw)
w.TogglePrintValueForStringer()
fmt.Println("Weights", w)
}
err = a.trainer.TrainValues(handle.XHandle(), dw.TD(), dw, w, a.gsum, a.xsum, a.params, (int32)(counter))
if err != nil {
return err
}
if debuggingadam {
//gsum, err := gocudnn.GetStringer(w.TD(), a.gsum)
//if err != nil {
// panic(err)
//}
//fmt.Println("GsumAfter", gsum)
//xsum, err := gocudnn.GetStringer(w.TD(), a.xsum)
//if err != nil {
// panic(err)
//}
//fmt.Println("xsum after", xsum)
fmt.Println("After Update")
fmt.Println("Weights", w)
fmt.Println(dw)
fmt.Println("(a *Adam) UpdateWeights")
for {
}
}
err = handle.Sync()
if err != nil {
return err
}
err = a.l1l2loss()
if err != nil {
return err
}
case flg.Half():
a.SetBatch(float32(batchsize))
err = a.trainer.L1L2Regularization(handle.XHandle(), dw.TD(), dw, w, a.gpuloss1, a.gpuloss2, a.regparams)
if err != nil {
return err
}
err = handle.Sync()
if err != nil {
return err
}
err = a.trainer.TrainValues(handle.XHandle(), dw.TD(), dw, w, a.gsum, a.xsum, a.params, (int32)(counter))
if err != nil {
return err
}
err = handle.Sync()
if err != nil {
return err
}
err = a.l1l2loss()
if err != nil {
return err
}
}
return handle.Sync()
}
func (a *Adam) l1l2loss() error {
var err error
err = nvidia.Memcpy(a.goptr1, a.gpuloss1, a.goptr1.TotalBytes())
if err != nil {
return err
}
err = nvidia.Memcpy(a.goptr2, a.gpuloss2, a.goptr2.TotalBytes())
if err != nil {
return err
}
return nil
}
//L1L2Loss returns the l1l2 loss of the memory that adam was training
func (a *Adam) L1L2Loss() (float32, float32) {
return a.loss1[0], a.loss2[0]
}
func dimsize(dims []int32) int32 {
x := int32(1)
for i := 0; i < len(dims); i++ {
x *= dims[i]
}
return x
}
//SetupAdamWandB returns a trainer for both WandB
func SetupAdamWandB(tctx *xtra.Handle, decay1, decay2 float32, batch int32) (*Adam, *Adam, error) {
adam1, err := SetupAdam(tctx, decay1, decay2, batch)
if err != nil {
return nil, nil, err
}
adam2, err := SetupAdam(tctx, decay1, decay2, batch)
return adam1, adam2, err
}
//SetupAdamWandB2 returns a trainer for both WandB and includes the learning rate
func SetupAdamWandB2(handle *cudnn.Handler, rate, dwalpha, decay1, decay2 float32, batch int32) (*Adam, *Adam, error) {
adam1, err := SetupAdam(handle.XHandle(), decay1, decay2, batch)
adam1.SetRates(rate, dwalpha)
if err != nil {
return nil, nil, err
}
adam2, err := SetupAdam(handle.XHandle(), decay1, decay2, batch)
adam2.SetRates(rate, dwalpha)
return adam1, adam2, err
}
//SetupAdam sets up adam
func SetupAdam(tctx *xtra.Handle, decay1, decay2 float32, batch int32) (*Adam, error) {
adam := xtra.TrainingModeFlag{}.Adam()
var dflt gocudnn.DataType
t, err := xtra.NewTrainingDescriptor(tctx, adam, dflt.Float())
if err != nil {
return nil, err
}
reg := xtra.CreateRegParamsFloat32(decay1, decay2, float32(batch))
x := xtra.CreateParamsFloat32(defaultadameps, defaultadamrate, defaultadambeta1, defaultadambeta2, 0)
return &Adam{
trainer: t,
params: x,
regparams: reg,
}, nil
}
//SetDecays sets the decay rates for the trainer
func (a *Adam) SetDecays(l1, l2 float32) {
a.regparams.SetDecay1(l1)
a.regparams.SetDecay2(l2)
}
//SetDecay1 sets decay1
func (a *Adam) SetDecay1(decay1 float32) {
a.regparams.SetDecay1(decay1)
}
//SetDecay2 sets decay 2
func (a *Adam) SetDecay2(decay2 float32) {
a.regparams.SetDecay2(decay2)
}
//SetBeta1 sets beta1
func (a *Adam) SetBeta1(beta1 float32) {
a.params.SetBeta1(beta1)
}
//SetBeta2 sets beta2
func (a *Adam) SetBeta2(beta2 float32) {
a.params.SetBeta2(beta2)
}
//SetRates sets rate
func (a *Adam) SetRates(rate, dwalpha float32) {
a.params.SetRate(rate)
a.params.SetDWalpha(dwalpha)
}
//SetBatch sets batch
func (a *Adam) SetBatch(batch float32) {
a.regparams.SetBatch(batch)
}
//SetEps sets eps
func (a *Adam) SetEps(eps float32) {
a.params.SetEps(eps)
}