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
|
#ifndef TH_GENERIC_FILE
#define TH_GENERIC_FILE "generic/SpatialMaxSampling.c"
#else
#ifndef MAX
#define MAX(a,b) ( ((a)>(b)) ? (a) : (b) )
#endif
#ifndef MIN
#define MIN(a,b) ( ((a)<(b)) ? (a) : (b) )
#endif
static int nn_(SpatialMaxSampling_forward)(lua_State *L)
{
// get all params
THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id));
int owidth = luaT_getfieldcheckint(L, 1, "owidth");
int oheight = luaT_getfieldcheckint(L, 1, "oheight");
THTensor *output = luaT_getfieldcheckudata(L, 1, "output", torch_(Tensor_id));
THTensor *indices = luaT_getfieldcheckudata(L, 1, "indices", torch_(Tensor_id));
// check dims
luaL_argcheck(L, input->nDimension == 3, 2, "3D tensor expected");
// dims
int ichannels = input->size[0];
int iheight = input->size[1];
int iwidth = input->size[2];
int ochannels = ichannels;
float dW = (float)iwidth/owidth;
float dH = (float)iheight/oheight;
// get contiguous input
input = THTensor_(newContiguous)(input);
// resize output
THTensor_(resize3d)(output, ochannels, oheight, owidth);
// indices will contain i,j locations for each output point
THTensor_(resize4d)(indices, 2, ochannels, oheight, owidth);
// get raw pointers
real *input_data = THTensor_(data)(input);
real *output_data = THTensor_(data)(output);
real *indices_data = THTensor_(data)(indices);
// compute max pooling for each input slice
long k;
for (k = 0; k < ochannels; k++) {
// pointers to slices
real *input_p = input_data + k*iwidth*iheight;
real *output_p = output_data + k*owidth*oheight;
real *indy_p = indices_data + k*owidth*oheight;
real *indx_p = indices_data + (k+ochannels)*owidth*oheight;
// loop over output
int i,j;
for(i = 0; i < oheight; i++) {
for(j = 0; j < owidth; j++) {
// compute nearest offsets
long ixs = (long)(j*dW);
long iys = (long)(i*dH);
long ixe = MAX(ixs+1, (long)((j+1)*dW));
long iye = MAX(iys+1, (long)((i+1)*dH));
// local pointers
real *op = output_p + i*owidth + j;
real *indxp = indx_p + i*owidth + j;
real *indyp = indy_p + i*owidth + j;
// compute local max:
long maxindex = -1;
real maxval = -THInf;
long tcntr = 0;
int x,y;
for(y = iys; y < iye; y++) {
for(x = ixs; x < ixe; x++) {
real val = *(input_p + y*iwidth + x);
if (val > maxval) {
maxval = val;
maxindex = tcntr;
}
tcntr++;
}
}
// set output to local max
*op = maxval;
// store location of max (x,y)
long kW = ixe-ixs;
*indyp = (int)(maxindex / kW)+1;
*indxp = (maxindex % kW) +1;
}
}
}
// cleanup
THTensor_(free)(input);
return 1;
}
static int nn_(SpatialMaxSampling_backward)(lua_State *L)
{
// get all params
THTensor *input = luaT_checkudata(L, 2, torch_(Tensor_id));
THTensor *gradOutput = luaT_checkudata(L, 3, torch_(Tensor_id));
THTensor *gradInput = luaT_getfieldcheckudata(L, 1, "gradInput", torch_(Tensor_id));
THTensor *indices = luaT_getfieldcheckudata(L, 1, "indices", torch_(Tensor_id));
int owidth = luaT_getfieldcheckint(L, 1, "owidth");
int oheight = luaT_getfieldcheckint(L, 1, "oheight");
// sizes
int ichannels = input->size[0];
int iheight = input->size[1];
int iwidth = input->size[2];
int ochannels = ichannels;
float dW = (float)iwidth/owidth;
float dH = (float)iheight/oheight;
// get contiguous gradOutput
gradOutput = THTensor_(newContiguous)(gradOutput);
// resize input
THTensor_(resizeAs)(gradInput, input);
THTensor_(zero)(gradInput);
// get raw pointers
real *gradInput_data = THTensor_(data)(gradInput);
real *gradOutput_data = THTensor_(data)(gradOutput);
real *indices_data = THTensor_(data)(indices);
// backprop all
long k;
for (k = 0; k < ichannels; k++) {
// pointers to slices
real *gradOutput_p = gradOutput_data + k*owidth*oheight;
real *gradInput_p = gradInput_data + k*iwidth*iheight;
real *indy_p = indices_data + k*owidth*oheight;
real *indx_p = indices_data + (k+ochannels)*owidth*oheight;
// calculate max points
int i,j;
for(i = 0; i < oheight; i++) {
for(j = 0; j < owidth; j++) {
// compute nearest offsets
long iys = (long)(i*dH);
long ixs = (long)(j*dW);
// retrieve position of max
real *indyp = indy_p + i*owidth + j;
real *indxp = indx_p + i*owidth + j;
long maxi = (*indyp) - 1 + iys;
long maxj = (*indxp) - 1 + ixs;
// update gradient
*(gradInput_p + maxi*iwidth + maxj) += *(gradOutput_p + i*owidth + j);
}
}
}
// cleanup
THTensor_(free)(gradOutput);
return 1;
}
static const struct luaL_Reg nn_(SpatialMaxSampling__) [] = {
{"SpatialMaxSampling_forward", nn_(SpatialMaxSampling_forward)},
{"SpatialMaxSampling_backward", nn_(SpatialMaxSampling_backward)},
{NULL, NULL}
};
static void nn_(SpatialMaxSampling_init)(lua_State *L)
{
luaT_pushmetaclass(L, torch_(Tensor_id));
luaT_registeratname(L, nn_(SpatialMaxSampling__), "nn");
lua_pop(L,1);
}
#endif
|