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
|
#include "caffe2/core/common_gpu.h"
#include "caffe2/core/context_gpu.h"
#include "caffe2/sgd/adadelta_op.h"
namespace caffe2 {
namespace {
__global__ void AdadeltaUpdateKernel(
int N,
const float* w,
const float* g,
const float* h,
const float* d,
const float epsilon,
const float decay,
const float* lr,
float* nw,
float* nh,
float* nd) {
CUDA_1D_KERNEL_LOOP(i, N) {
float gi = g[i];
float di = d[i];
float hi = nh[i] = decay * h[i] + (1.0f - decay) * gi * gi;
float ng = sqrtf(di + epsilon) * rsqrtf(hi + epsilon) * gi;
nw[i] = w[i] + lr[0] * ng;
nd[i] = decay * di + (1.0f - decay) * ng * ng;
}
}
template <>
void AdadeltaUpdate<CUDAContext>(
int N,
const float* w,
const float* g,
const float* h,
const float* d,
const float epsilon,
const float decay,
const float* lr,
float* nw,
float* nh,
float* nd,
CUDAContext* context) {
AdadeltaUpdateKernel<<<
CAFFE_GET_BLOCKS(N),
CAFFE_CUDA_NUM_THREADS,
0,
context->cuda_stream()>>>(N, w, g, h, d, epsilon, decay, lr, nw, nh, nd);
C10_CUDA_KERNEL_LAUNCH_CHECK();
}
} // namespace
template <typename SIndex, typename THalf>
__global__ void SparseAdadeltaKernel(
const size_t N,
const size_t grad_slice_sz,
const float epsilon,
const float decay,
const SIndex* indices,
const float* grad,
const float* lr,
THalf* param,
THalf* param_mom,
THalf* param_mom_delta) {
const float LR = lr[0];
CUDA_1D_KERNEL_LOOP(i, N) {
const size_t gradIdx = i;
const SIndex index = indices[i / grad_slice_sz];
const size_t paramIdx = index * grad_slice_sz + (i % grad_slice_sz);
float mom_new = decay * param_mom[paramIdx] +
(1.0f - decay) * grad[gradIdx] * grad[gradIdx];
param_mom[paramIdx] = mom_new;
float grad_new = sqrtf(epsilon + param_mom_delta[paramIdx]) *
rsqrtf(mom_new + epsilon) * grad[gradIdx];
float param_new = LR * grad_new + param[paramIdx];
param[paramIdx] = param_new;
float mom_delta_new = decay * param_mom_delta[paramIdx] +
(1.0f - decay) * grad_new * grad_new;
param_mom_delta[paramIdx] = mom_delta_new;
}
}
template <class Context>
class CUDASparseAdadeltaOp final : public Operator<Context> {
public:
USE_OPERATOR_CONTEXT_FUNCTIONS;
CUDASparseAdadeltaOp(const OperatorDef& operator_def, Workspace* ws)
: Operator<Context>(operator_def, ws),
OP_SINGLE_ARG(float, "epsilon", epsilon_, 1e-5f),
OP_SINGLE_ARG(float, "decay", decay_, 0.95f) {}
bool RunOnDevice() override {
// Enforce shapes
CAFFE_ENFORCE_EQ(Input(PARAM).size(), Input(MOMENT_GRAD).size());
CAFFE_ENFORCE_EQ(Input(PARAM).size(), Input(MOMENT_DELTA).size());
CAFFE_ENFORCE_EQ(Input(LR).size(), 1);
CAFFE_ENFORCE_EQ(
Input(PARAM).size_from_dim(1),
Input(GRAD).size_from_dim(Input(INDICES).ndim()));
// Enforce domain constraints on attributes
CAFFE_ENFORCE_GE(epsilon_, 0.0f);
CAFFE_ENFORCE_GT(decay_, 0.0f);
CAFFE_ENFORCE_LT(decay_, 1.0f);
return DispatchHelper<TensorTypes<int32_t, int64_t>>::call(
this, Input(INDICES));
}
template <typename IndexType>
bool DoRunWithType() {
auto n = Input(INDICES).size();
if (n == 0) {
return true;
}
return DispatchHelper<TensorTypes2<float, at::Half>, IndexType>::call(
this, Input(PARAM));
}
template <typename IndexType, typename THalf>
bool DoRunWithType2() {
const auto* lr = Input(LR).template data<float>();
const auto* indices = Input(INDICES).template data<IndexType>();
const auto* gradIn = Input(GRAD).template data<float>();
const auto* paramIn = Input(PARAM).template data<THalf>();
const auto* momentIn = Input(MOMENT_GRAD).template data<THalf>();
const auto* momentDeltaIn = Input(MOMENT_DELTA).template data<THalf>();
auto* paramOut = Output(OUTPUT_PARAM)->template mutable_data<THalf>();
auto* momentOut =
Output(OUTPUT_MOMENT_GRAD)->template mutable_data<THalf>();
auto* momentDeltaOut =
Output(OUTPUT_MOMENT_DELTA)->template mutable_data<THalf>();
auto N = Input(GRAD).size();
auto grad_slice_sz = Input(GRAD).size_from_dim(Input(INDICES).ndim());
if (N == 0) {
// empty grad, nothing to do here, not even launching the kernel
return true;
}
SparseAdadeltaKernel<IndexType, THalf>
<<<CAFFE_GET_BLOCKS(N),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(
N,
grad_slice_sz,
epsilon_,
decay_,
indices,
gradIn,
lr,
paramOut,
momentOut,
momentDeltaOut);
C10_CUDA_KERNEL_LAUNCH_CHECK();
return true;
}
protected:
const float epsilon_;
const float decay_;
INPUT_TAGS(PARAM, MOMENT_GRAD, MOMENT_DELTA, INDICES, GRAD, LR);
OUTPUT_TAGS(OUTPUT_PARAM, OUTPUT_MOMENT_GRAD, OUTPUT_MOMENT_DELTA);
};
REGISTER_CUDA_OPERATOR(Adadelta, AdadeltaOp<CUDAContext>);
REGISTER_CUDA_OPERATOR(SparseAdadelta, CUDASparseAdadeltaOp<CUDAContext>);
} // namespace caffe2
|
