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
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
|
/*
Copyright (c) 2013, Taiga Nomi
All rights reserved.
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 the <organization> 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
#include "gtest/gtest.h"
#include "testhelper.h"
#include "tiny_dnn/tiny_dnn.h"
namespace tiny_dnn {
TEST(quantized_deconvolutional, setup_internal) {
quantized_deconvolutional_layer<sigmoid> l(2, 2, 3, 1, 2,
padding::valid, true, 1, 1, backend_t::internal);
EXPECT_EQ(l.parallelize(), true); // if layer can be parallelized
EXPECT_EQ(l.in_channels(), serial_size_t(3)); // num of input tensors
EXPECT_EQ(l.out_channels(), serial_size_t(2)); // num of output tensors
EXPECT_EQ(l.in_data_size(), serial_size_t(4)); // size of input tensors
EXPECT_EQ(l.out_data_size(), serial_size_t(32)); // size of output tensors
EXPECT_EQ(l.in_data_shape().size(), serial_size_t(1)); // number of inputs shapes
EXPECT_EQ(l.out_data_shape().size(), serial_size_t(1)); // num of output shapes
EXPECT_EQ(l.weights().size(), serial_size_t(2)); // the wieghts vector size
EXPECT_EQ(l.weights_grads().size(), serial_size_t(2)); // the wieghts vector size
EXPECT_EQ(l.inputs().size(), serial_size_t(3)); // num of input edges
EXPECT_EQ(l.outputs().size(), serial_size_t(2)); // num of outpus edges
EXPECT_EQ(l.in_types().size(), serial_size_t(3)); // num of input data types
EXPECT_EQ(l.out_types().size(), serial_size_t(2)); // num of output data types
EXPECT_EQ(l.fan_in_size(), serial_size_t(9)); // num of incoming connections
EXPECT_EQ(l.fan_out_size(), serial_size_t(18)); // num of outgoing connections
EXPECT_STREQ(l.layer_type().c_str(), "q_deconv"); // string with layer type
}
TEST(quantized_deconvolutional, fprop) {
typedef network<sequential> CNN;
CNN nn;
quantized_deconvolutional_layer<sigmoid> l(2, 2, 3, 1, 2);
// layer::forward_propagation expects tensors, even if we feed only one input at a time
auto create_simple_tensor = [](size_t vector_size) {
return tensor_t(1, vec_t(vector_size));
};
// create simple tensors that wrap the payload vectors of the correct size
tensor_t in_tensor = create_simple_tensor(4)
, out_tensor = create_simple_tensor(32)
, a_tensor = create_simple_tensor(32)
, weight_tensor = create_simple_tensor(18)
, bias_tensor = create_simple_tensor(2);
// short-hand references to the payload vectors
vec_t &in = in_tensor[0]
, &out = out_tensor[0]
, &weight = weight_tensor[0];
ASSERT_EQ(l.in_shape()[1].size(), serial_size_t(18)); // weight
uniform_rand(in.begin(), in.end(), -1.0, 1.0);
std::vector<tensor_t*> in_data, out_data;
in_data.push_back(&in_tensor);
in_data.push_back(&weight_tensor);
in_data.push_back(&bias_tensor);
out_data.push_back(&out_tensor);
out_data.push_back(&a_tensor);
l.setup(false);
{
l.forward_propagation(in_data, out_data);
for (auto o: out)
EXPECT_NEAR(0.5, o, 1E-3);
}
weight[0] = 0.3f; weight[1] = 0.1f; weight[2] = 0.2f;
weight[3] = 0.0f; weight[4] =-0.1f; weight[5] =-0.1f;
weight[6] = 0.05f; weight[7] =-0.2f; weight[8] = 0.05f;
weight[9] = 0.0f; weight[10] =-0.1f; weight[11] = 0.1f;
weight[12] = 0.1f; weight[13] =-0.2f; weight[14] = 0.3f;
weight[15] = 0.2f; weight[16] =-0.3f; weight[17] = 0.2f;
in[0] = 3; in[1] = 2;
in[2] = 3; in[3] = 0;
{
l.forward_propagation(in_data, out_data);
EXPECT_NEAR(0.7109495, out[0], 1E-2);
EXPECT_NEAR(0.7109495, out[1], 1E-2);
EXPECT_NEAR(0.6899745, out[2], 1E-2);
EXPECT_NEAR(0.5986877, out[3], 1E-2);
EXPECT_NEAR(0.7109495, out[4], 1E-2);
EXPECT_NEAR(0.5000000, out[5], 1E-2);
EXPECT_NEAR(0.5249792, out[6], 1E-2);
EXPECT_NEAR(0.4501660, out[7], 1E-2);
EXPECT_NEAR(0.5374298, out[8], 1E-2);
EXPECT_NEAR(0.3100255, out[9], 1E-2);
EXPECT_NEAR(0.3658644, out[10], 1E-2);
EXPECT_NEAR(0.5249791, out[11], 1E-2);
EXPECT_NEAR(0.5374298, out[12], 1E-2);
EXPECT_NEAR(0.3543437, out[13], 1E-2);
EXPECT_NEAR(0.5374298, out[14], 1E-2);
EXPECT_NEAR(0.5000000, out[15], 1E-2);
}
}
/*
TEST(quantized_deconvolutional, fprop2) {
typedef network<sequential> CNN;
CNN nn;
quantized_deconvolutional_layer<sigmoid> l(2, 2, 3, 1, 2, padding::same);
vec_t in(4), out(32), a(32), weight(18), bias(2);
ASSERT_EQ(l.in_shape()[1].size(), 18); // weight
uniform_rand(in.begin(), in.end(), -1.0, 1.0);
std::vector<vec_t*> in_data, out_data;
in_data.push_back(&in);
in_data.push_back(&weight);
in_data.push_back(&bias);
out_data.push_back(&out);
out_data.push_back(&a);
l.setup(false, 1);
{
l.forward_propagation(0, in_data, out_data);
for (auto o: out)
EXPECT_NEAR(0.5, o, 1E-3);
}
weight[0] = 0.3; weight[1] = 0.1; weight[2] = 0.2;
weight[3] = 0.0; weight[4] =-0.1; weight[5] =-0.1;
weight[6] = 0.05; weight[7] =-0.2; weight[8] = 0.05;
weight[9] = 0.0; weight[10] =-0.1; weight[11] = 0.1;
weight[12] = 0.1; weight[13] =-0.2; weight[14] = 0.3;
weight[15] = 0.2; weight[16] =-0.3; weight[17] = 0.2;
in[0] = 3; in[1] = 2;
in[2] = 3; in[3] = 0;
{
l.forward_propagation(0, in_data, out_data);
EXPECT_NEAR(0.5000000, out[0], 1E-2);
EXPECT_NEAR(0.5249792, out[1], 1E-2);
EXPECT_NEAR(0.3100255, out[2], 1E-2);
EXPECT_NEAR(0.3658644, out[3], 1E-2);
}
}
TEST(quantized_deconvolutional, gradient_check) { // tanh - mse
network<sequential> nn;
nn << quantized_deconvolutional_layer<tan_h>(5, 5, 3, 1, 1);
vec_t a(25, 0.0);
label_t t = 3;
uniform_rand(a.begin(), a.end(), -1, 1);
nn.init_weight();
EXPECT_TRUE(nn.gradient_check<mse>(&a, &t, 1, epsilon<float_t>(), GRAD_CHECK_ALL));
}
TEST(quantized_deconvolutional, gradient_check2) { // sigmoid - mse
network<sequential> nn;
nn << quantized_deconvolutional_layer<sigmoid>(5, 5, 3, 1, 1);
vec_t a(25, 0.0);
label_t t = 3;
uniform_rand(a.begin(), a.end(), -1, 1);
nn.init_weight();
EXPECT_TRUE(nn.gradient_check<mse>(&a, &t, 1, epsilon<float_t>(), GRAD_CHECK_ALL));
}
TEST(quantized_deconvolutional, gradient_check3) { // rectified - mse
network<sequential> nn;
nn << quantized_deconvolutional_layer<rectified_linear>(5, 5, 3, 1, 1);
vec_t a(25, 0.0);
label_t t = 3;
uniform_rand(a.begin(), a.end(), -1, 1);
nn.init_weight();
EXPECT_TRUE(nn.gradient_check<mse>(&a, &t, 1, epsilon<float_t>(), GRAD_CHECK_ALL));
}
TEST(quantized_deconvolutional, gradient_check4) { // identity - mse
network<sequential> nn;
nn << quantized_deconvolutional_layer<identity>(5, 5, 3, 1, 1);
vec_t a(25, 0.0);
label_t t = 3;
uniform_rand(a.begin(), a.end(), -1, 1);
nn.init_weight();
EXPECT_TRUE(nn.gradient_check<mse>(&a, &t, 1, epsilon<float_t>(), GRAD_CHECK_ALL));
}
TEST(quantized_deconvolutional, gradient_check5) { // sigmoid - cross-entropy
network<sequential> nn;
nn << quantized_deconvolutional_layer<sigmoid>(5, 5, 3, 1, 1);
vec_t a(25, 0.0);
label_t t = 3;
uniform_rand(a.begin(), a.end(), -1, 1);
nn.init_weight();
EXPECT_TRUE(nn.gradient_check<cross_entropy>(&a, &t, 1, epsilon<float_t>(), GRAD_CHECK_ALL));
}
TEST(quantized_deconvolutional, read_write)
{
quantized_deconvolutional_layer<tan_h> l1(5, 5, 3, 1, 1);
quantized_deconvolutional_layer<tan_h> l2(5, 5, 3, 1, 1);
l1.init_weight();
l2.init_weight();
serialization_test(l1, l2);
}
TEST(quantized_deconvolutional, read_write2) {
#define O true
#define X false
static const bool connection[] = {
O, X, X, X, O, O,
O, O, X, X, X, O,
O, O, O, X, X, X
};
#undef O
#undef X
quantized_deconvolutional_layer<tan_h> layer1(14, 14, 5, 3, 6, connection_table(connection, 3, 6));
quantized_deconvolutional_layer<tan_h> layer2(14, 14, 5, 3, 6, connection_table(connection, 3, 6));
layer1.init_weight();
layer2.init_weight();
serialization_test(layer1, layer2);
}*/
} // namespace tiny-dnn
|
