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
|
/* Copyright 2015 Google Inc. 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.
==============================================================================*/
// See docs in ../ops/image_ops.cc
#define EIGEN_USE_THREADS
#include <memory>
#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/lib/core/status.h"
#include "tensorflow/core/platform/logging.h"
namespace tensorflow {
typedef Eigen::ThreadPoolDevice CPUDevice;
template <typename Device, typename T>
class ResizeNearestNeighborOp : public OpKernel {
public:
explicit ResizeNearestNeighborOp(OpKernelConstruction* context)
: OpKernel(context) {
OP_REQUIRES_OK(context, context->GetAttr("align_corners", &align_corners_));
}
void Compute(OpKernelContext* context) override {
const Tensor& input = context->input(0);
OP_REQUIRES(context, input.dims() == 4,
errors::InvalidArgument("input must be 4-dimensional",
input.shape().DebugString()));
const Tensor& shape_t = context->input(1);
OP_REQUIRES(context, shape_t.dims() == 1,
errors::InvalidArgument("shape_t must be 1-dimensional",
shape_t.shape().DebugString()));
OP_REQUIRES(context, shape_t.NumElements() == 2,
errors::InvalidArgument("shape_t must have two elements",
shape_t.shape().DebugString()));
auto sizes = shape_t.vec<int32>();
OP_REQUIRES(context, sizes(0) > 0 && sizes(1) > 0,
errors::InvalidArgument("shape_t's elements must be positive"));
// Initialize shape to the batch size of the input, then add
// the rest of the dimensions
Tensor* output = nullptr;
OP_REQUIRES_OK(context, context->allocate_output(
0, TensorShape({input.dim_size(0), sizes(0),
sizes(1), input.dim_size(3)}),
&output));
const int64 batch_size = input.dim_size(0);
const int64 in_height = input.dim_size(1);
const int64 in_width = input.dim_size(2);
const int64 channels = input.dim_size(3);
const int64 out_height = output->dim_size(1);
const int64 out_width = output->dim_size(2);
typename TTypes<T, 4>::ConstTensor input_data = input.tensor<T, 4>();
typename TTypes<T, 4>::Tensor output_data = output->tensor<T, 4>();
const float height_scale =
(align_corners_ && out_height > 1)
? (in_height - 1) / static_cast<float>(out_height - 1)
: in_height / static_cast<float>(out_height);
const float width_scale =
(align_corners_ && out_width > 1)
? (in_width - 1) / static_cast<float>(out_width - 1)
: in_width / static_cast<float>(out_width);
for (int b = 0; b < batch_size; ++b) {
for (int y = 0; y < out_height; ++y) {
const int in_y = std::min(static_cast<int64>(floorf(y * height_scale)),
(in_height - 1));
for (int x = 0; x < out_width; ++x) {
const int in_x = std::min(static_cast<int64>(floorf(x * width_scale)),
(in_width - 1));
for (int c = 0; c < channels; ++c) {
output_data(b, y, x, c) = input_data(b, in_y, in_x, c);
}
}
}
}
}
private:
bool align_corners_;
};
template <typename Device, typename T>
class ResizeNearestNeighborOpGrad : public OpKernel {
public:
explicit ResizeNearestNeighborOpGrad(OpKernelConstruction* context)
: OpKernel(context) {
OP_REQUIRES_OK(context, context->GetAttr("align_corners", &align_corners_));
}
void Compute(OpKernelContext* context) override {
// Grab and validate the input:
const Tensor& input = context->input(0);
OP_REQUIRES(context, input.dims() == 4,
errors::InvalidArgument("input must be 4-dimensional",
input.shape().DebugString()));
// Grab and validate the output shape:
const Tensor& shape_t = context->input(1);
OP_REQUIRES(context, shape_t.dims() == 1,
errors::InvalidArgument("shape_t must be 1-dimensional",
shape_t.shape().DebugString()));
OP_REQUIRES(context, shape_t.NumElements() == 2,
errors::InvalidArgument("shape_t must have two elements",
shape_t.shape().DebugString()));
auto sizes = shape_t.vec<int32>();
OP_REQUIRES(context, sizes(0) > 0 && sizes(1) > 0,
errors::InvalidArgument("shape_t's elements must be positive"));
// Initialize shape to the batch size of the input, then add
// the rest of the dimensions
Tensor* output = nullptr;
OP_REQUIRES_OK(context, context->allocate_output(
0, TensorShape({input.dim_size(0), sizes(0),
sizes(1), input.dim_size(3)}),
&output));
const int64 batch_size = input.dim_size(0);
const int64 in_height = input.dim_size(1);
const int64 in_width = input.dim_size(2);
const int64 channels = input.dim_size(3);
const int64 out_height = output->dim_size(1);
const int64 out_width = output->dim_size(2);
typename TTypes<T, 4>::ConstTensor input_data = input.tensor<T, 4>();
typename TTypes<T, 4>::Tensor output_data = output->tensor<T, 4>();
const float height_scale =
(align_corners_ && in_height > 1)
? (out_height - 1) / static_cast<float>(in_height - 1)
: out_height / static_cast<float>(in_height);
const float width_scale =
(align_corners_ && in_width > 1)
? (out_width - 1) / static_cast<float>(in_width - 1)
: out_width / static_cast<float>(in_width);
for (int c = 0; c < channels; ++c) {
for (int y = 0; y < out_height; ++y) {
for (int x = 0; x < out_width; ++x) {
for (int b = 0; b < batch_size; ++b) {
output_data(b, y, x, c) = 0;
}
}
}
}
for (int c = 0; c < channels; ++c) {
for (int y = 0; y < in_height; ++y) {
const int out_y = std::min(static_cast<int64>(floorf(y * height_scale)),
(out_height - 1));
for (int x = 0; x < in_width; ++x) {
const int out_x = std::min(
static_cast<int64>(floorf(x * width_scale)), (out_width - 1));
for (int b = 0; b < batch_size; ++b) {
output_data(b, out_y, out_x, c) += input_data(b, y, x, c);
}
}
}
}
}
private:
bool align_corners_;
};
#define REGISTER_KERNEL(T) \
REGISTER_KERNEL_BUILDER(Name("ResizeNearestNeighbor") \
.Device(DEVICE_CPU) \
.TypeConstraint<T>("T") \
.HostMemory("size"), \
ResizeNearestNeighborOp<CPUDevice, T>); \
REGISTER_KERNEL_BUILDER(Name("ResizeNearestNeighborGrad") \
.Device(DEVICE_CPU) \
.TypeConstraint<T>("T") \
.HostMemory("size"), \
ResizeNearestNeighborOpGrad<CPUDevice, T>);
TF_CALL_REAL_NUMBER_TYPES(REGISTER_KERNEL);
#undef REGISTER_KERNEL
} // namespace tensorflow
|
