blob: 3653fc8fa536056ef6e3d86325b400a4621c8774 (
plain)
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
|
#include "THCUNN.h"
#include "common.h"
#include "THCHalf.h"
#include "THCHalfAutoNumerics.cuh"
#include <thrust/functional.h>
#include <thrust/device_ptr.h>
#include <thrust/iterator/zip_iterator.h>
#include <thrust/transform.h>
#include <thrust/transform_reduce.h>
template <typename T>
inline __device__ T eps();
template <>
inline __device__ float eps() { return 1e-12f; }
template <>
inline __device__ double eps() { return 1e-12; }
template <typename Dtype, typename Acctype>
struct bce_functor
{
template <class Tuple>
__host__ __device__
Acctype operator()(Tuple x)
{
Dtype o = thrust::get<0>(x);
Dtype t = thrust::get<1>(x);
return - (t * THCNumerics<Acctype>::log(o + eps<Acctype>()) + (Acctype(1)- t) * THCNumerics<Acctype>::log(Acctype(1) - o + eps<Acctype>()));
}
};
template <typename Dtype, typename Acctype>
struct bce_functor_weights
{
template <class Tuple>
__host__ __device__
Acctype operator()(Tuple x)
{
Dtype o = thrust::get<0>(x);
Dtype t = thrust::get<1>(x);
Dtype w = thrust::get<2>(x);
return - w * (t * THCNumerics<Acctype>::log(o + eps<Acctype>()) + (Acctype(1) - t) * THCNumerics<Acctype>::log(Acctype(1) - o + eps<Acctype>()));
}
};
template <typename Dtype, typename Acctype>
struct bce_updateGradInput_functor
{
const Dtype norm;
bce_updateGradInput_functor(Dtype norm_)
: norm(norm_)
{}
template <class Tuple>
__host__ __device__
Dtype operator()(Tuple x)
{
Dtype o = thrust::get<0>(x);
Dtype t = thrust::get<1>(x);
return ScalarConvert<Acctype,Dtype>::to(- (t - o) / ((Acctype(1) - o + eps<Acctype>()) * (o + eps<Acctype>())) * norm);
}
};
template <typename Dtype, typename Acctype>
struct bce_updateGradInput_functor_weights
{
const Dtype norm;
bce_updateGradInput_functor_weights(Dtype norm_)
: norm(norm_)
{}
template <class Tuple>
__host__ __device__
Dtype operator()(Tuple x)
{
Dtype o = thrust::get<0>(x);
Dtype t = thrust::get<1>(x);
Dtype w = thrust::get<2>(x);
return ScalarConvert<Acctype, Dtype>::to(- (t - o) / ((Acctype(1) - o + eps<Acctype>()) * (o + eps<Acctype>())) * norm * w);
}
};
#include "generic/BCECriterion.cu"
#include "THCGenerateFloatTypes.h"
|
