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
|
#include "caffe2/operators/logit_op.h"
#include "caffe2/core/context_gpu.h"
namespace caffe2 {
namespace {
template <typename T>
__global__ void LogitKernel(const int N, const T* X, const float eps, T* Y) {
CUDA_1D_KERNEL_LOOP(i, N) {
Y[i] = fminf(X[i], (T(1) - eps));
Y[i] = fmaxf(Y[i], eps);
Y[i] = logf(Y[i] / (T(1) - Y[i]));
}
}
template <typename T>
__global__ void LogitGradientKernel(
const int N,
const T* X,
const T* dY,
const float eps,
T* dX) {
CUDA_1D_KERNEL_LOOP(i, N) {
dX[i] = (X[i] < eps || X[i] > T(1) - eps) ? T(0)
: (dY[i] / X[i] / (T(1) - X[i]));
}
}
} // namespace
template <>
template <typename T>
bool LogitFunctor<CUDAContext>::
operator()(const int N, const T* X, T* Y, CUDAContext* context) const {
LogitKernel<T>
<<<CAFFE_GET_BLOCKS(N),
CAFFE_CUDA_NUM_THREADS,
0,
context->cuda_stream()>>>(N, X, eps_, Y);
return true;
}
template <>
bool LogitGradientOp<float, CUDAContext>::RunOnDevice() {
auto& X = Input(0);
auto& dY = Input(1);
auto* dX = Output(0);
dX->ResizeLike(X);
int n = X.size();
LogitGradientKernel<<<
CAFFE_GET_BLOCKS(n),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(
n,
X.data<float>(),
dY.data<float>(),
eps_,
dX->template mutable_data<float>());
return true;
}
REGISTER_CUDA_OPERATOR(
Logit,
UnaryElementwiseWithArgsOp<
TensorTypes<float>,
CUDAContext,
LogitFunctor<CUDAContext>>);
REGISTER_CUDA_OPERATOR(LogitGradient, LogitGradientOp<float, CUDAContext>);
} // namespace caffe2
|
