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
|
#ifndef THC_GENERIC_FILE
#define THC_GENERIC_FILE "generic/SpatialDilatedMaxPooling.cu"
#else
#include "../common.h"
void THNN_(SpatialDilatedMaxPooling_updateOutput)(
THCState *state,
THCTensor *input,
THCTensor *output,
THCIndexTensor *indices,
int kW, int kH,
int dW, int dH,
int padW, int padH,
int dilationW, int dilationH,
bool ceil_mode)
{
THCUNN_assertSameGPU_generic(state, 3, input, output, indices);
THArgCheck(input->nDimension == 3 || input->nDimension == 4, 2, "3D or 4D (batch) tensor expected");
long nInputCols, nInputRows, nInputPlane, batchSize;
long nOutputCols, nOutputRows;
if (input->nDimension == 3) {
nInputCols = input->size[2];
nInputRows = input->size[1];
nInputPlane = input->size[0];
batchSize = 1;
}
else
{
nInputCols = input->size[3];
nInputRows = input->size[2];
nInputPlane = input->size[1];
batchSize = input->size[0];
}
THArgCheck(nInputCols >= kW - padW && nInputRows >= kH - padH, 2, "input image smaller than kernel size");
THArgCheck(kW/2 >= padW && kH/2 >= padH, 2, "pad should be smaller than half of kernel size");
if(ceil_mode) {
nOutputCols = ceil(float(nInputCols - (dilationW * (kW - 1) + 1) + 2*padW) / float(dW)) + 1;
nOutputRows = ceil(float(nInputRows - (dilationH * (kH - 1) + 1) + 2*padH) / float(dH)) + 1;
}
else {
nOutputCols = floor(float(nInputCols - (dilationW * (kW - 1) + 1) + 2*padW) / float(dW)) + 1;
nOutputRows = floor(float(nInputRows - (dilationH * (kH - 1) + 1) + 2*padH) / float(dH)) + 1;
}
if (nOutputCols < 1 || nOutputRows < 1)
THError("Given input size: (%dx%dx%d). Calculated output size: (%dx%dx%d). Output size is too small",
nInputPlane,nInputRows,nInputCols,nInputPlane,nOutputRows,nOutputCols);
if (padW || padH)
{
// ensure that the last pooling starts inside the image
if ((nOutputRows - 1)*dH >= nInputRows + padH)
--nOutputRows;
if ((nOutputCols - 1)*dW >= nInputCols + padW)
--nOutputCols;
}
input = THCTensor_(newContiguous)(state, input);
real* input_data = THCTensor_(data)(state, input);
THCTensor_(resize4d)(state, output, batchSize, nInputPlane, nOutputRows, nOutputCols);
THCUNN_resizeAs_indices(state, indices, output);
THCIndex_t* indices_data = THCIndexTensor_(data)(state, indices);
real* output_data = THCTensor_(data)(state, output);
int count = THCTensor_(nElement)(state, output);
MaxPoolForward<real, accreal> <<< GET_BLOCKS(count), CUDA_NUM_THREADS, 0, THCState_getCurrentStream(state) >>>
(count, input_data,
batchSize, nInputPlane, nInputRows, nInputCols, nOutputRows, nOutputCols,
kH, kW, dH, dW, padH, padW, dilationH, dilationW, output_data, indices_data);
THCudaCheck(cudaGetLastError());
if(input->nDimension == 3)
THCTensor_(resize3d)(state, output, nInputPlane, nOutputRows, nOutputCols);
THCTensor_(free)(state, input);
}
void THNN_(SpatialDilatedMaxPooling_updateGradInput)(
THCState *state,
THCTensor *input,
THCTensor *gradOutput,
THCTensor *gradInput,
THCIndexTensor *indices,
int kW, int kH,
int dW, int dH,
int padW, int padH,
int dilationW, int dilationH,
bool ceil_mode)
{
THCUNN_assertSameGPU_generic(state, 4, input, gradOutput, indices, gradInput);
input = THCTensor_(newContiguous)(state, input);
gradOutput = THCTensor_(newContiguous)(state, gradOutput);
long nInputCols, nInputRows, nInputPlane, batchSize;
long nOutputCols, nOutputRows;
if (input->nDimension == 3) {
nInputCols = input->size[2];
nInputRows = input->size[1];
nInputPlane = input->size[0];
batchSize = 1;
}
else
{
nInputCols = input->size[3];
nInputRows = input->size[2];
nInputPlane = input->size[1];
batchSize = input->size[0];
}
if(ceil_mode) {
nOutputCols = ceil(float(nInputCols - (dilationW * (kW - 1) + 1) + 2*padW) / float(dW)) + 1;
nOutputRows = ceil(float(nInputRows - (dilationH * (kH - 1) + 1) + 2*padH) / float(dH)) + 1;
}
else {
nOutputCols = floor(float(nInputCols - (dilationW * (kW - 1) + 1) + 2*padW) / float(dW)) + 1;
nOutputRows = floor(float(nInputRows - (dilationH * (kH - 1) + 1) + 2*padH) / float(dH)) + 1;
}
if (nOutputCols < 1 || nOutputRows < 1)
THError("Given input size: (%dx%dx%d). Calculated output size: (%dx%dx%d). Output size is too small",
nInputPlane,nInputRows,nInputCols,nInputPlane,nOutputRows,nOutputCols);
gradOutput = THCTensor_(newContiguous)(state, gradOutput);
THCTensor_(resizeAs)(state, gradInput, input);
int count = THCTensor_(nElement)(state, input);
MaxPoolBackward<real, accreal> <<< GET_BLOCKS(count), CUDA_NUM_THREADS, 0, THCState_getCurrentStream(state) >>>
(count,
THCTensor_(data)(state, gradOutput),
THCIndexTensor_(data)(state, indices),
batchSize, nInputPlane, nInputRows, nInputCols, nOutputRows, nOutputCols,
kH, kW, dH, dW, padH, padW, dilationH, dilationW,
THCTensor_(data)(state, gradInput));
THCudaCheck(cudaGetLastError());
THCTensor_(free)(state, gradOutput);
// clean
THCTensor_(free)(state, input);
THCTensor_(free)(state, gradOutput);
}
#endif
|
