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
|
/*
* Copyright (c) 2019 Samsung Electronics Co., Ltd. All Rights Reserved
* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#ifndef __NNFW_CKER_TRANSPOSE_CONV_H__
#define __NNFW_CKER_TRANSPOSE_CONV_H__
#include "cker/Shape.h"
#include "cker/Types.h"
#include "cker/Utils.h"
namespace nnfw
{
namespace cker
{
inline void TransposeConv(const TransposeConvParams ¶ms, const Shape &input_shape,
const float *input_data, const Shape &filter_shape,
const float *filter_data, const Shape &output_shape, float *output_data)
{
const int stride_width = params.stride_width;
const int stride_height = params.stride_height;
const int pad_width = params.padding_values.width;
const int pad_height = params.padding_values.height;
assert(input_shape.DimensionsCount() == 4);
assert(filter_shape.DimensionsCount() == 4);
assert(output_shape.DimensionsCount() == 4);
const int batches = MatchingDim(input_shape, 0, output_shape, 0);
const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
const int input_height = input_shape.Dims(1);
const int input_width = input_shape.Dims(2);
const int filter_height = filter_shape.Dims(1);
const int filter_width = filter_shape.Dims(2);
const int output_height = output_shape.Dims(1);
const int output_width = output_shape.Dims(2);
// Although transpose convolution simplifies to convolution with transposed
// weights for strides of 1, non-unitary striding complicates matters. To
// keep this reference implementation as clear as possible, we use a
// "scatter" access pattern, where we loop through all the input elements,
// computing their influence on the output, rather than looping through the
// output elements in the typical "gather" access pattern of a conv. We
// therefore must initialize the output array to zero.
const int num_elements = output_shape.FlatSize();
for (int i = 0; i < num_elements; i++)
{
output_data[i] = 0.0f;
}
// Loop through input elements one at a time.
for (int batch = 0; batch < batches; ++batch)
{
for (int in_y = 0; in_y < input_height; ++in_y)
{
for (int in_x = 0; in_x < input_width; ++in_x)
{
for (int in_channel = 0; in_channel < input_depth; ++in_channel)
{
// Loop through the output elements it will influence
const int out_x_origin = (in_x * stride_width) - pad_width;
const int out_y_origin = (in_y * stride_height) - pad_height;
for (int filter_y = 0; filter_y < filter_height; ++filter_y)
{
for (int filter_x = 0; filter_x < filter_width; ++filter_x)
{
for (int out_channel = 0; out_channel < output_depth; ++out_channel)
{
// Compute output element location
const int out_x = out_x_origin + filter_x;
const int out_y = out_y_origin + filter_y;
// We cannot accumulate out of bounds
if ((out_x >= 0) && (out_x < output_width) && (out_y >= 0) &&
(out_y < output_height))
{
float input_value =
input_data[Offset(input_shape, batch, in_y, in_x, in_channel)];
float filter_value = filter_data[Offset(filter_shape, out_channel, filter_y,
filter_x, in_channel)];
output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] +=
input_value * filter_value;
}
}
}
}
}
}
}
}
}
} // namespace cker
} // namespace nnfw
#endif // __NNFW_CKER_TRANSPOSE_CONV_H__
|