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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
|
/*
COPYRIGHT
All contributions by Taiga Nomi
Copyright (c) 2013, Taiga Nomi
All rights reserved.
All other contributions:
Copyright (c) 2013-2016, the respective contributors.
All rights reserved.
Each contributor holds copyright over their respective contributions.
The project versioning (Git) records all such contribution source information.
LICENSE
The BSD 3-Clause License
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
* Redistributions of source code must retain the above copyright notice, this
list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above copyright notice,
this list of conditions and the following disclaimer in the documentation
and/or other materials provided with the distribution.
* Neither the name of tiny-dnn nor the names of its
contributors may be used to endorse or promote products derived from
this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#pragma once
namespace tiny_dnn {
namespace kernels {
inline void
conv2d_op_internal(const tensor_t& in_data,
const vec_t& W,
const vec_t& bias,
tensor_t& out_data,
const core::conv_params& params,
const bool parallelize) {
for_i(parallelize, in_data.size(), [&](int sample) {
const vec_t& in = in_data[sample];
vec_t& a = out_data[sample];
for (serial_size_t o = 0; o < params.out.depth_; o++) {
for (serial_size_t inc = 0; inc < params.in.depth_; inc++) {
if (!params.tbl.is_connected(o, inc)) continue;
serial_size_t idx = 0;
idx = params.in.depth_ * o + inc;
idx = params.weight.get_index(0, 0, idx);
const float_t *pw = &W[idx];
idx = params.in_padded.get_index(0, 0, inc);
const float_t *pi = &in[idx];
idx = params.out.get_index(0, 0, o);
float_t *pa = &a[idx];
for (serial_size_t y = 0; y < params.out.height_; y++) {
for (serial_size_t x = 0; x < params.out.width_; x++) {
const float_t * ppw = pw;
const float_t * ppi = pi + params.in_padded.width_ *
(y * params.h_stride) +
x * params.w_stride;
float_t sum = float_t(0);
// should be optimized for small kernel(3x3,5x5)
for (serial_size_t wy = 0; wy < params.weight.height_; wy++) { // NOLINT
for (serial_size_t wx = 0; wx < params.weight.width_; wx++) { // NOLINT
idx = wy * params.in_padded.width_ + wx;
sum += *ppw++ * ppi[idx];
}
}
pa[y * params.out.width_ + x] += sum;
}
}
}
if (params.has_bias) {
float_t * pa = &a[params.out.get_index(0, 0, o)];
float_t * paa = pa + params.out.width_ * params.out.height_;
std::for_each(pa, paa, [&](float_t& f) { f += bias[o]; });
}
}
});
}
/******************************************************************/
template <typename tensor_t, typename vec_t>
void
conv2d_op_internal(const tensor_t& prev_out,
const vec_t& W,
tensor_t& dW,
tensor_t& db,
tensor_t& curr_delta,
tensor_t& prev_delta,
const core::conv_params& params,
const bool parallelize) {
typedef typename vec_t::value_type float_t;
for_i(parallelize, prev_out.size(), [&](int sample) {
// propagate delta to previous layer
for (serial_size_t inc = 0; inc < params.in.depth_; inc++) {
for (serial_size_t outc = 0; outc < params.out.depth_; outc++) {
if (!params.tbl.is_connected(outc, inc)) continue;
serial_size_t idx = 0;
idx = params.in.depth_ * outc + inc;
idx = params.weight.get_index(0, 0, idx);
const float_t *pw = &W[idx];
idx = params.out.get_index(0, 0, outc);
const float_t *pdelta_src = &curr_delta[sample][idx];
idx = params.in_padded.get_index(0, 0, inc);
//float_t *pdelta_dst = &(*prev_delta)[sample][idx];
float_t *pdelta_dst = &prev_delta[sample][idx];
for (serial_size_t y = 0; y < params.out.height_; y++) {
for (serial_size_t x = 0; x < params.out.width_; x++) {
const float_t * ppw = pw;
idx = y * params.out.width_ + x;
const float_t ppdelta_src = pdelta_src[idx];
float_t * ppdelta_dst = pdelta_dst +
y * params.h_stride * params.in_padded.width_ +
x * params.w_stride;
for (serial_size_t wy = 0; wy < params.weight.height_; wy++) { // NOLINT
for (serial_size_t wx = 0; wx < params.weight.width_; wx++) { // NOLINT
idx = wy * params.in_padded.width_ + wx;
ppdelta_dst[idx] += *ppw++ * ppdelta_src;
}
}
}
}
}
}
// accumulate dw
for (serial_size_t inc = 0; inc < params.in.depth_; inc++) {
for (serial_size_t outc = 0; outc < params.out.depth_; outc++) {
if (!params.tbl.is_connected(outc, inc)) continue;
for (serial_size_t wy = 0; wy < params.weight.height_; wy++) {
for (serial_size_t wx = 0; wx < params.weight.width_; wx++) {
float_t dst = float_t(0);
serial_size_t idx = 0;
idx = params.in_padded.get_index(wx, wy, inc);
const float_t * prevo = &prev_out[sample][idx];
idx = params.out.get_index(0, 0, outc);
const float_t * delta = &curr_delta[sample][idx];
if (params.w_stride > 1) {
for (serial_size_t y = 0; y < params.out.height_; y++) {
serial_size_t prevo_idx = y * params.in_padded.width_ * params.h_stride;
serial_size_t delta_idx = y * params.out.width_;
for (serial_size_t x = 0; x < params.out.width_; x++) {
dst += prevo[prevo_idx + x * params.w_stride] * delta[delta_idx + x];
}
}
} else {
for (serial_size_t y = 0; y < params.out.height_; y++) {
dst += vectorize::dot(
prevo + y * params.in_padded.width_ * params.h_stride,
delta + y * params.out.width_,
params.out.width_);
}
}
idx = params.in.depth_ * outc + inc;
dW[sample][params.weight.get_index(wx, wy, idx)] += dst;
}
}
}
}
// accumulate db
if (params.has_bias) {
for (serial_size_t outc = 0; outc < params.out.depth_; outc++) {
serial_size_t idx = params.out.get_index(0, 0, outc);
const float_t * delta = &curr_delta[sample][idx];
const float_t * deltaa = delta + params.out.width_ *
params.out.height_;
db[sample][outc] += std::accumulate(delta, deltaa, float_t(0));
}
}
});
}
} // namespace kernels
} // namespace tiny_dnn
|
