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
|
#include "caffe2/core/context_gpu.h"
#include "caffe2/sgd/lars_op.h"
namespace caffe2 {
__global__ void ComputeLearningRateKernel(
const float* wd,
const float* trust,
const float* lr_max,
float offset,
float lr_min,
float* X_norm,
float* dX_norm,
float* lr_rescaled) {
float val = 1.0;
if (*X_norm > 0) {
val = (*trust) / (*dX_norm / *X_norm + (*wd) + offset);
}
*lr_rescaled = fmaxf(fminf(val, *lr_max), lr_min);
}
template <>
void LarsOp<float, CUDAContext>::ComputeLearningRate(
const float* wd,
const float* trust,
const float* lr_max,
float offset,
float lr_min,
float* X_norm,
float* dX_norm,
float* lr_rescaled) {
ComputeLearningRateKernel<<<1, 1, 0, context_.cuda_stream()>>>(
wd, trust, lr_max, offset, lr_min, X_norm, dX_norm, lr_rescaled);
C10_CUDA_KERNEL_LAUNCH_CHECK();
}
REGISTER_CUDA_OPERATOR(Lars, LarsOp<float, CUDAContext>);
} // namespace caffe2
|
