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
|
package tensor
import (
"github.com/pkg/errors"
"gorgonia.org/tensor/internal/storage"
"reflect"
)
// SelectByIndices selects the values given the in `indices` tensor.
//
// Currently SelectByIndices only supports Dense tensors that do not require the use of iterators.
// Please make a pull request to support tensors that require the use of an iterator to traverse data.
func (e StdEng) SelectByIndices(a, indices Tensor, axis int, opts ...FuncOpt) (retVal Tensor, err error) {
if !indices.Shape().IsVectorLike() {
return nil, errors.Errorf("Expected indices to be a vector. Got %v instead", indices.Shape())
}
if indices.Dtype() != Int {
return nil, errors.Errorf("Expected indices to be a vector of ints. Got %v instead", indices.Dtype())
}
// if b is a scalar, then use Slice
if a.Shape().IsScalarEquiv() {
slices := make([]Slice, a.Shape().Dims())
slices[axis] = ss(getInts(indices)[0])
return a.Slice(slices...)
}
expectedShape := a.Shape().Clone()
expectedShape[axis] = indices.Shape().TotalSize()
var reuse DenseTensor
var safe, toReuse, _ bool
if reuse, safe, toReuse, _, _, err = handleFuncOpts(expectedShape, a.Dtype(), a.DataOrder(), true, opts...); err != nil {
return nil, errors.Wrap(err, "Unable to handle funcOpts")
}
if safe || !toReuse && reuse == nil && safe {
// create reuse
reuse = New(WithShape(expectedShape...), Of(a.Dtype()))
}
if !safe {
if a.Shape()[axis] != indices.Shape().TotalSize() {
expected := a.Shape().Clone()
expected[axis] = indices.Shape().TotalSize()
return nil, errors.Errorf("Expected a safe resuse to have the same shape as the expected shape of the result: %v. The input a has %v ", expected, a.Shape())
}
reuse = a.(DenseTensor)
}
typ := a.Dtype().Type
var dataA, dataB, dataReuse *storage.Header
var ait, bit, iit Iterator
var useIter bool
if dataA, dataB, dataReuse, ait, bit, iit, useIter, _, err = prepDataVV(a, indices, reuse); err != nil {
return nil, errors.Wrapf(err, "StdEng.Add")
}
if useIter {
e.iterSelectByIdx(axis, dataA, dataB, dataReuse, ait, bit, iit)
//TODO
return
}
e.selectByIdx(axis, dataB.Ints(), typ, dataA, dataReuse, a.(*Dense).AP, reuse.(*Dense).AP)
return reuse, nil
}
func (e StdEng) iterSelectByIdx(axis int, dataA, dataB, dataReuse *storage.Header, ait, bit, iit Iterator) {
panic("iterSelectByIdx is not yet implemented")
}
func (e StdEng) selectByIdx(axis int, indices []int, typ reflect.Type, dataA, dataRetVal *storage.Header, apA, apRet AP) {
isInnermost := axis == apA.shape.Dims()-1
outer := ProdInts(apA.shape[:axis])
axStride := apA.strides[axis]
retStride := apRet.strides[axis]
var outerRetStride int
if axis == 0 {
// then it's the outermost
outerRetStride = apRet.strides[axis] * 2
} else {
outerRetStride = apRet.strides[axis-1]
}
srcCoord := make([]int, apA.shape.Dims())
dstCoord := make([]int, apRet.shape.Dims())
if isInnermost {
prevAxis := axis - 1
if prevAxis < 0 {
// this may be the case if input is a vector
prevAxis = 0
}
prevStride := apA.strides[prevAxis]
retPrevStride := apRet.strides[prevAxis]
for i, idx := range indices {
srcCoord[axis] = idx
dstCoord[axis] = i
start, _ := Ltoi(apA.shape, apA.strides, srcCoord...)
dstStart, _ := Ltoi(apRet.shape, apRet.strides, dstCoord...)
for o := 0; o < outer; o++ {
end := start + axStride
dstEnd := dstStart + retStride
storage.CopySliced(typ, dataRetVal, dstStart, dstEnd, dataA, start, end)
start += prevStride
dstStart += retPrevStride
}
}
return
}
for i, idx := range indices {
srcCoord[axis] = idx
dstCoord[axis] = i
start, _ := Ltoi(apA.shape, apA.strides, srcCoord...)
dstStart, _ := Ltoi(apRet.shape, apRet.strides, dstCoord...)
for o := 0; o < outer; o++ {
end := start + axStride
dstEnd := dstStart + retStride
storage.CopySliced(typ, dataRetVal, dstStart, dstEnd, dataA, start, end)
start = end + axStride
dstStart = dstEnd + (outerRetStride - retStride)
}
}
}
// SelectByIndicesB computes the gradient of the result of `SelectByIndices`.
//
// Currently SelectByIndicesB only supports Dense tensors that do not require the use of iterators.
// Please make a pull request to support tensors that require the use of an iterator to traverse data.
func (e StdEng) SelectByIndicesB(input, outGrad, indices Tensor, axis int, opts ...FuncOpt) (retVal Tensor, err error) {
if !indices.Shape().IsVectorLike() {
return nil, errors.Errorf("Expected indices to be a vector. Got %v instead", outGrad.Shape())
}
if indices.Dtype() != Int {
return nil, errors.Errorf("Expected indices to be a vector of ints. Got %v instead", outGrad.Dtype())
}
// if b is a scalar, then use Slice
if input.Shape().IsScalarEquiv() {
slices := make([]Slice, input.Shape().Dims())
slices[axis] = ss(outGrad.Data().([]int)[0])
return input.Slice(slices...)
}
expectedShape := input.Shape().Clone()
var reuse DenseTensor
var _, toReuse, _ bool
if reuse, _, toReuse, _, _, err = handleFuncOpts(input.Shape(), input.Dtype(), input.DataOrder(), true, opts...); err != nil {
return nil, errors.Wrap(err, "Unable to handle funcOpts")
}
if !toReuse && reuse == nil {
// create reuse
reuse = New(WithShape(expectedShape...), Of(input.Dtype()))
}
typ := input.Dtype().Type
var _, dataB, dataReuse *storage.Header
var _, bit, iit Iterator
var useIter bool
if _, dataB, dataReuse, _, bit, iit, useIter, _, err = prepDataVV(input, outGrad, reuse); err != nil {
return nil, errors.Wrapf(err, "StdEng.SelectByIndicesB")
}
if useIter {
e.iterSelectByIndicesB(axis, dataB, dataReuse, bit, iit)
//TODO
return
}
e.selectByIndicesB(axis, getInts(indices), typ, dataB, dataReuse, outGrad.(*Dense).AP, reuse.(*Dense).AP)
return reuse, nil
}
func (e StdEng) iterSelectByIndicesB(axis int, dataB, dataGradA *storage.Header, bit, iit Iterator) {
panic("iterSelectByIndicesB not implemented yet")
}
func (e StdEng) selectByIndicesB(axis int, indices []int, typ reflect.Type, dataB, dataGradA *storage.Header, apB, apRet AP) {
isInnermost := axis == apB.shape.Dims()-1
outer := ProdInts(apB.shape[:axis])
axStride := apB.strides[axis]
retStride := apRet.strides[axis]
var outerRetStride int
if axis == 0 {
outerRetStride = apRet.strides[axis] * 2
} else {
outerRetStride = apRet.strides[axis-1]
}
dstCoord := make([]int, apB.shape.Dims())
srcCoord := make([]int, apRet.shape.Dims())
if isInnermost {
prevAxis := axis - 1
if prevAxis < 0 {
// this may be the case if input is a vector
prevAxis = 0
}
retPrevStride := apB.strides[prevAxis]
prevStride := apRet.strides[prevAxis]
for i, idx := range indices {
dstCoord[axis] = idx
srcCoord[axis] = i
dstStart, _ := Ltoi(apB.shape, apB.strides, dstCoord...)
start, _ := Ltoi(apRet.shape, apRet.strides, srcCoord...)
for o := 0; o < outer; o++ {
dstEnd := dstStart + axStride
end := start + retStride
e.E.AddSliced(typ, dataGradA, dstStart, dstEnd, dataB, start, end)
dstStart += prevStride
start += retPrevStride
}
}
return
}
for i, idx := range indices {
dstCoord[axis] = idx
srcCoord[axis] = i
dstStart, _ := Ltoi(apRet.shape, apRet.strides, dstCoord...)
start, _ := Ltoi(apB.shape, apB.strides, srcCoord...)
for o := 0; o < outer; o++ {
dstEnd := dstStart + axStride
end := start + retStride
e.E.AddSliced(typ, dataGradA, dstStart, dstEnd, dataB, start, end)
dstStart = dstEnd + axStride
start = end + (outerRetStride - retStride)
}
}
}
|
