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
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
|
/* Copyright 2015 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.
==============================================================================*/
#include "tensorflow/cc/ops/const_op.h"
#include "tensorflow/cc/ops/image_ops.h"
#include "tensorflow/cc/ops/nn_ops.h"
#include "tensorflow/cc/ops/standard_ops.h"
#include "tensorflow/core/common_runtime/kernel_benchmark_testlib.h"
#include "tensorflow/core/framework/fake_input.h"
#include "tensorflow/core/framework/node_def_builder.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/framework/types.pb.h"
#include "tensorflow/core/kernels/ops_testutil.h"
#include "tensorflow/core/kernels/ops_util.h"
#include "tensorflow/core/platform/test.h"
#include "tensorflow/core/platform/test_benchmark.h"
#include "tensorflow/core/public/session.h"
namespace tensorflow {
class FusedResizePadConvOpTest : public OpsTestBase {
protected:
void HandwrittenConv() {
const int stride = 1;
TF_EXPECT_OK(NodeDefBuilder("fused_resize_op", "FusedResizeAndPadConv2D")
.Input(FakeInput(DT_FLOAT))
.Input(FakeInput(DT_INT32))
.Input(FakeInput(DT_INT32))
.Input(FakeInput(DT_FLOAT))
.Attr("T", DT_FLOAT)
.Attr("resize_align_corners", false)
.Attr("mode", "REFLECT")
.Attr("strides", {1, stride, stride, 1})
.Attr("padding", "SAME")
.Finalize(node_def()));
TF_EXPECT_OK(InitOp());
const int depth = 1;
const int image_width = 4;
const int image_height = 3;
const int image_batch_count = 1;
// The image matrix is:
// | 1 | 2 | 3 | 4 |
// | 5 | 6 | 7 | 8 |
// | 9 | 10 | 11 | 12 |
Tensor image(DT_FLOAT,
{image_batch_count, image_height, image_width, depth});
test::FillValues<float>(&image, {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12});
// The filter matrix is:
// | 1 | 4 | 7 |
// | 2 | 5 | 8 |
// | 3 | 6 | 9 |
const int filter_size = 3;
const int filter_count = 1;
Tensor filter(DT_FLOAT, {filter_size, filter_size, depth, filter_count});
test::FillValues<float>(&filter, {1, 4, 7, 2, 5, 8, 3, 6, 9});
const int resized_width = image_width;
const int resized_height = image_height;
const int top_padding = 0;
const int bottom_padding = 0;
const int left_padding = 0;
const int right_padding = 0;
AddInputFromArray<float>(image.shape(), image.flat<float>());
AddInputFromArray<int32>(TensorShape({2}), {resized_height, resized_width});
AddInputFromArray<int32>(
TensorShape({4, 2}),
{0, 0, top_padding, bottom_padding, left_padding, right_padding, 0, 0});
AddInputFromArray<float>(filter.shape(), filter.flat<float>());
TF_ASSERT_OK(RunOpKernel());
// We're sliding the 3x3 filter across the 3x4 image, with accesses outside
// the input set to zero because we're using the 'SAME' padding mode.
// The calculations behind the expected output are:
// (1*0)+(4*0)+(7*0)+(2*0)+(5*1)+(8*2)+(3*0)+(6*5)+(9*6)=105
// (1*0)+(4*0)+(7*0)+(2*1)+(5*2)+(8*3)+(3*5)+(6*6)+(9*7)=150
// (1*0)+(4*0)+(7*0)+(2*2)+(5*3)+(8*4)+(3*6)+(6*7)+(9*8)=183
// (1*0)+(4*0)+(7*0)+(2*3)+(5*4)+(8*0)+(3*7)+(6*8)+(9*0)=95
// (1*0)+(4*1)+(7*2)+(2*0)+(5*5)+(8*6)+(3*0)+(6*9)+(9*10)=235
// (1*1)+(4*2)+(7*3)+(2*5)+(5*6)+(8*7)+(3*9)+(6*10)+(9*11)=312
// (1*2)+(4*3)+(7*4)+(2*6)+(5*7)+(8*8)+(3*10)+(6*11)+(9*12)=357
// (1*3)+(4*4)+(7*0)+(2*7)+(5*8)+(8*0)+(3*11)+(6*12)+(9*0)=178
// (1*0)+(4*5)+(7*6)+(2*0)+(5*9)+(8*10)+(3*0)+(6*0)+(9*0)=187
// (1*5)+(4*6)+(7*7)+(2*9)+(5*10)+(8*11)+(3*0)+(6*0)+(9*0)=234
// (1*6)+(4*7)+(7*8)+(2*10)+(5*11)+(8*12)+(3*0)+(6*0)+(9*0)=261
// (1*7)+(4*11)+(7*0)+(2*8)+(5*12)+(8*0)+(3*0)+(6*0)+(9*0)=121
// This means we should end up with this matrix:
// | 105 | 150 | 183 | 95 |
// | 235 | 312 | 357 | 178 |
// | 187 | 234 | 261 | 121 |
const int expected_width = image_width;
const int expected_height = image_height * filter_count;
Tensor expected(DT_FLOAT, TensorShape({image_batch_count, expected_height,
expected_width, filter_count}));
test::FillValues<float>(
&expected, {105, 150, 183, 95, 235, 312, 357, 178, 187, 234, 261, 121});
const Tensor& output = *GetOutput(0);
test::ExpectTensorNear<float>(expected, output, 1e-5);
}
void CompareFusedAndSeparate(int input_width, int input_height,
int input_depth, int resize_width,
int resize_height, int y_padding, int x_padding,
int filter_size, int filter_count,
bool resize_align_corners, string pad_mode,
int stride, string padding) {
auto root = tensorflow::Scope::NewRootScope();
using namespace ::tensorflow::ops; // NOLINT(build/namespaces)
Tensor input_data(DT_FLOAT,
TensorShape({1, input_height, input_width, input_depth}));
test::FillIota<float>(&input_data, 1.0f);
Output input =
Const(root.WithOpName("input"), Input::Initializer(input_data));
Tensor filter_data(DT_FLOAT, TensorShape({filter_size, filter_size,
input_depth, filter_count}));
test::FillIota<float>(&filter_data, 1.0f);
Output filter =
Const(root.WithOpName("filter"), Input::Initializer(filter_data));
Output resize_size =
Const(root.WithOpName("resize_size"), {resize_height, resize_width});
Output resize =
ResizeBilinear(root.WithOpName("resize"), input, resize_size,
ResizeBilinear::AlignCorners(resize_align_corners));
Output paddings =
Const(root.WithOpName("paddings"),
{{0, 0}, {y_padding, y_padding}, {x_padding, x_padding}, {0, 0}});
Output mirror_pad =
MirrorPad(root.WithOpName("mirror_pad"), resize, paddings, pad_mode);
Output conv = Conv2D(root.WithOpName("conv"), mirror_pad, filter,
{1, stride, stride, 1}, padding);
Output fused_conv = FusedResizeAndPadConv2D(
root.WithOpName("fused_conv"), input, resize_size, paddings, filter,
pad_mode, {1, stride, stride, 1}, padding,
FusedResizeAndPadConv2D::ResizeAlignCorners(resize_align_corners));
tensorflow::GraphDef graph;
TF_ASSERT_OK(root.ToGraphDef(&graph));
std::unique_ptr<tensorflow::Session> session(
tensorflow::NewSession(tensorflow::SessionOptions()));
TF_ASSERT_OK(session->Create(graph));
std::vector<Tensor> unfused_tensors;
TF_ASSERT_OK(session->Run({}, {"conv"}, {}, &unfused_tensors));
std::vector<Tensor> fused_tensors;
TF_ASSERT_OK(session->Run({}, {"fused_conv"}, {}, &fused_tensors));
test::ExpectTensorNear<float>(unfused_tensors[0], fused_tensors[0], 1e-5);
}
void CompareFusedPadOnlyAndSeparate(int input_width, int input_height,
int input_depth, int y_padding,
int x_padding, int filter_size,
int filter_count, string pad_mode,
int stride, string padding) {
auto root = tensorflow::Scope::NewRootScope();
using namespace ::tensorflow::ops; // NOLINT(build/namespaces)
Tensor input_data(DT_FLOAT,
TensorShape({1, input_height, input_width, input_depth}));
test::FillIota<float>(&input_data, 1.0f);
Output input =
Const(root.WithOpName("input"), Input::Initializer(input_data));
Tensor filter_data(DT_FLOAT, TensorShape({filter_size, filter_size,
input_depth, filter_count}));
test::FillIota<float>(&filter_data, 1.0f);
Output filter =
Const(root.WithOpName("filter"), Input::Initializer(filter_data));
Output paddings =
Const(root.WithOpName("paddings"),
{{0, 0}, {y_padding, y_padding}, {x_padding, x_padding}, {0, 0}});
Output mirror_pad =
MirrorPad(root.WithOpName("mirror_pad"), input, paddings, pad_mode);
Output conv = Conv2D(root.WithOpName("conv"), mirror_pad, filter,
{1, stride, stride, 1}, padding);
Output fused_conv =
FusedPadConv2D(root.WithOpName("fused_conv"), input, paddings, filter,
pad_mode, {1, stride, stride, 1}, padding);
tensorflow::GraphDef graph;
TF_ASSERT_OK(root.ToGraphDef(&graph));
std::unique_ptr<tensorflow::Session> session(
tensorflow::NewSession(tensorflow::SessionOptions()));
TF_ASSERT_OK(session->Create(graph));
std::vector<Tensor> unfused_tensors;
TF_ASSERT_OK(session->Run({}, {"conv"}, {}, &unfused_tensors));
std::vector<Tensor> fused_tensors;
TF_ASSERT_OK(session->Run({}, {"fused_conv"}, {}, &fused_tensors));
test::ExpectTensorNear<float>(unfused_tensors[0], fused_tensors[0], 1e-5);
}
};
TEST_F(FusedResizePadConvOpTest, HandwrittenConv) { HandwrittenConv(); }
TEST_F(FusedResizePadConvOpTest, IdentityComparative) {
CompareFusedAndSeparate(10, 10, 1, 10, 10, 0, 0, 1, 1, false, "REFLECT", 1,
"SAME");
}
TEST_F(FusedResizePadConvOpTest, ConvOnlyComparative) {
CompareFusedAndSeparate(10, 10, 3, 10, 10, 0, 0, 4, 4, false, "REFLECT", 1,
"SAME");
}
TEST_F(FusedResizePadConvOpTest, ResizeOnlyComparative) {
CompareFusedAndSeparate(10, 10, 1, 20, 20, 0, 0, 1, 1, false, "REFLECT", 1,
"SAME");
}
TEST_F(FusedResizePadConvOpTest, ResizeAndConvComparative) {
CompareFusedAndSeparate(2, 2, 4, 4, 2, 0, 0, 2, 2, false, "REFLECT", 1,
"SAME");
}
TEST_F(FusedResizePadConvOpTest, ResizeAlignAndConvComparative) {
CompareFusedAndSeparate(2, 2, 4, 4, 2, 0, 0, 2, 2, true, "REFLECT", 1,
"SAME");
}
TEST_F(FusedResizePadConvOpTest, ResizeAndConvStridedComparative) {
CompareFusedAndSeparate(2, 2, 4, 4, 2, 0, 0, 2, 2, false, "REFLECT", 2,
"SAME");
}
TEST_F(FusedResizePadConvOpTest, ResizeAlignAndConvValidComparative) {
CompareFusedAndSeparate(2, 2, 4, 4, 2, 0, 0, 2, 2, true, "REFLECT", 1,
"VALID");
}
TEST_F(FusedResizePadConvOpTest, PadOnlyComparative) {
CompareFusedAndSeparate(4, 4, 1, 4, 4, 2, 2, 1, 1, false, "REFLECT", 1,
"SAME");
}
TEST_F(FusedResizePadConvOpTest, PadOnlyWithChannelsComparative) {
CompareFusedAndSeparate(4, 4, 3, 4, 4, 2, 2, 1, 1, false, "REFLECT", 1,
"SAME");
}
TEST_F(FusedResizePadConvOpTest, ResizeAndPadComparative) {
CompareFusedAndSeparate(4, 4, 1, 6, 6, 2, 2, 1, 1, false, "REFLECT", 1,
"SAME");
}
TEST_F(FusedResizePadConvOpTest, PadOnlySymmetricComparative) {
CompareFusedAndSeparate(4, 4, 1, 4, 4, 2, 2, 1, 1, false, "SYMMETRIC", 1,
"SAME");
}
TEST_F(FusedResizePadConvOpTest, ResizeAndPadSymmetricComparative) {
CompareFusedAndSeparate(4, 4, 3, 6, 6, 2, 2, 1, 1, false, "SYMMETRIC", 1,
"SAME");
}
TEST_F(FusedResizePadConvOpTest, NoResizeIdentityComparative) {
CompareFusedPadOnlyAndSeparate(10, 10, 1, 0, 0, 1, 1, "REFLECT", 1, "SAME");
}
TEST_F(FusedResizePadConvOpTest, NoResizeConvOnlyComparative) {
CompareFusedPadOnlyAndSeparate(10, 10, 3, 0, 0, 4, 4, "REFLECT", 1, "SAME");
}
TEST_F(FusedResizePadConvOpTest, NoResizePadOnlyComparative) {
CompareFusedPadOnlyAndSeparate(4, 4, 1, 2, 2, 1, 1, "REFLECT", 1, "SAME");
}
TEST_F(FusedResizePadConvOpTest, NoResizePadOnlyWithChannelsComparative) {
CompareFusedPadOnlyAndSeparate(4, 4, 3, 2, 2, 1, 1, "REFLECT", 1, "SAME");
}
TEST_F(FusedResizePadConvOpTest, NoResizePadOnlySymmetricComparative) {
CompareFusedPadOnlyAndSeparate(4, 4, 1, 2, 2, 1, 1, "SYMMETRIC", 1, "SAME");
}
} // namespace tensorflow
|
