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
|
#pragma once
#include "caffe2/core/operator.h"
namespace caffe2 {
template <typename Context>
void storm_update(
const int N,
const float* paramIn,
const float* momentIn,
const float* gradSqSumIn,
const float* gradIn,
const float* lr,
float* paramOut,
float* momentOut,
float* gradSqSumOut,
const float momentum,
const float beta,
Context* /*context*/) {
float gradSqSumTmp = 0.0;
for (const auto i : c10::irange(N)) {
const float gi = gradIn[i];
gradSqSumTmp += gi * gi;
}
gradSqSumOut[0] = gradSqSumIn[0] + gradSqSumTmp;
const float nlr = lr[0] * std::pow(beta + gradSqSumOut[0], -1.0 / 3.0);
const float alpha = momentum * nlr * nlr;
for (const auto i : c10::irange(N)) {
const float gi = gradIn[i];
const float mi = momentIn[i];
float new_mi = momentOut[i] = gi + (1.0 - alpha) * (mi - gi);
paramOut[i] = paramIn[i] + nlr * new_mi;
}
}
template <class Context>
class StormOp final : public Operator<Context> {
public:
USE_OPERATOR_CONTEXT_FUNCTIONS;
StormOp(const OperatorDef& operator_def, Workspace* ws)
: Operator<Context>(operator_def, ws),
OP_SINGLE_ARG(float, "momentum", momentum_, 10.0),
OP_SINGLE_ARG(float, "beta", beta_, 0.1) {}
bool RunOnDevice() override {
// Enforce shapes
CAFFE_ENFORCE_EQ(Input(GRAD).numel(), Input(PARAM).numel());
CAFFE_ENFORCE_EQ(Input(GRAD).numel(), Input(MOMENT).numel());
CAFFE_ENFORCE_EQ(Input(GRADSQSUM).numel(), 1);
CAFFE_ENFORCE_EQ(Input(LR).numel(), 1);
// Resize [potentially] out-of-place blobs
Output(OUTPUT_PARAM)->ResizeLike(Input(PARAM));
Output(OUTPUT_MOMENT)->ResizeLike(Input(MOMENT));
Output(OUTPUT_GRAGSQSUM)->ResizeLike(Input(GRADSQSUM));
storm_update<Context>(
Input(GRAD).numel(),
Input(PARAM).template data<float>(),
Input(MOMENT).template data<float>(),
Input(GRADSQSUM).template data<float>(),
Input(GRAD).template data<float>(),
Input(LR).template data<float>(),
Output(OUTPUT_PARAM)->template mutable_data<float>(),
Output(OUTPUT_MOMENT)->template mutable_data<float>(),
Output(OUTPUT_GRAGSQSUM)->template mutable_data<float>(),
momentum_,
beta_,
&context_);
return true;
}
protected:
const float momentum_;
const float beta_;
INPUT_TAGS(PARAM, MOMENT, GRADSQSUM, GRAD, LR);
OUTPUT_TAGS(OUTPUT_PARAM, OUTPUT_MOMENT, OUTPUT_GRAGSQSUM);
};
template <class Context>
class SparseStormOp final : public Operator<Context> {
public:
USE_OPERATOR_CONTEXT_FUNCTIONS;
SparseStormOp(const OperatorDef& operator_def, Workspace* ws)
: Operator<Context>(operator_def, ws),
OP_SINGLE_ARG(float, "momentum", momentum_, 10.0),
OP_SINGLE_ARG(float, "beta", beta_, 0.1) {}
bool RunOnDevice() override {
// Enforce shapes
CAFFE_ENFORCE_EQ(Input(PARAM).numel(), Input(MOMENT).numel());
CAFFE_ENFORCE_EQ(Input(GRADSQSUM).numel(), 1);
CAFFE_ENFORCE_EQ(Input(LR).numel(), 1);
CAFFE_ENFORCE_EQ(
Input(PARAM).size_from_dim(1),
Input(GRAD).size_from_dim(Input(INDICES).dim()));
return DispatchHelper<TensorTypes<int32_t, int64_t>>::call(
this, Input(INDICES));
}
template <typename SIndex>
bool DoRunWithType() {
const auto* paramIn = Input(PARAM).template data<float>();
const auto* momentIn = Input(MOMENT).template data<float>();
const auto* gradSqSumIn = Input(GRADSQSUM).template data<float>();
const auto* gradIn = Input(GRAD).template data<float>();
const auto* indices = Input(INDICES).template data<SIndex>();
const auto* lr = Input(LR).template data<float>();
auto* paramOut = Output(OUTPUT_PARAM)->template mutable_data<float>();
auto* momentOut = Output(OUTPUT_MOMENT)->template mutable_data<float>();
auto* gradSqSumOut =
Output(OUTPUT_GRAGSQSUM)->template mutable_data<float>();
auto n = Input(INDICES).numel();
if (n == 0) {
return true;
}
float gradSqSumTmp = 0.0;
for (const auto i : c10::irange(Input(GRAD).numel())) {
const float gi = gradIn[i];
gradSqSumTmp += gi * gi;
}
gradSqSumOut[0] = gradSqSumIn[0] + gradSqSumTmp;
const float nlr = lr[0] * std::pow(beta_ + gradSqSumOut[0], -1.0 / 3.0);
const float alpha = momentum_ * nlr * nlr;
const auto block_size = Input(GRAD).numel() / n;
for (const auto i : c10::irange(n)) {
auto idx = indices[i];
if (block_size == 1) {
const float gi = gradIn[i];
const float mi = momentIn[idx];
float new_mi = momentOut[idx] = gi + (1.0 - alpha) * (mi - gi);
paramOut[idx] = paramIn[idx] + nlr * new_mi;
} else {
auto offsetI = i * block_size;
auto offsetIdx = idx * block_size;
#ifndef NDEBUG
CAFFE_ENFORCE_GE(
Input(PARAM).numel(),
block_size + offsetIdx,
this->debug_def().input(PARAM),
", out of bound, idx:",
idx,
" for input i:",
i,
" and block size:",
block_size);
CAFFE_ENFORCE_GE(
Input(GRAD).numel(),
block_size + offsetI,
this->debug_def().input(GRAD),
", out of bound idx, idx:",
idx,
" for input i:",
i);
#endif
for (const auto j : c10::irange(block_size)) {
const float gi = gradIn[offsetI + j];
const float mi = momentIn[offsetIdx + j];
float new_mi = momentOut[offsetIdx + j] =
gi + (1.0 - alpha) * (mi - gi);
paramOut[offsetIdx + j] = paramIn[offsetIdx + j] + nlr * new_mi;
}
}
}
return true;
}
protected:
const float momentum_;
const float beta_;
INPUT_TAGS(PARAM, MOMENT, GRADSQSUM, GRAD, INDICES, LR);
OUTPUT_TAGS(OUTPUT_PARAM, OUTPUT_MOMENT, OUTPUT_GRAGSQSUM);
};
} // namespace caffe2
|
