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
|
/* Copyright 2016 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.
==============================================================================*/
#define EIGEN_USE_THREADS
#if GOOGLE_CUDA
#define EIGEN_USE_GPU
#endif // GOOGLE_CUDA
#include "tensorflow/contrib/image/kernels/image_ops.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/platform/types.h"
namespace tensorflow {
namespace functor {
// Explicit instantiation of the CPU functor.
typedef Eigen::ThreadPoolDevice CPUDevice;
template struct FillProjectiveTransform<CPUDevice, uint8>;
template struct FillProjectiveTransform<CPUDevice, int32>;
template struct FillProjectiveTransform<CPUDevice, int64>;
template struct FillProjectiveTransform<CPUDevice, Eigen::half>;
template struct FillProjectiveTransform<CPUDevice, float>;
template struct FillProjectiveTransform<CPUDevice, double>;
} // end namespace functor
typedef Eigen::ThreadPoolDevice CPUDevice;
using functor::FillProjectiveTransform;
using generator::Interpolation;
using generator::INTERPOLATION_BILINEAR;
using generator::INTERPOLATION_NEAREST;
using generator::ProjectiveGenerator;
template <typename Device, typename T>
class ImageProjectiveTransform : public OpKernel {
private:
Interpolation interpolation_;
public:
explicit ImageProjectiveTransform(OpKernelConstruction* ctx) : OpKernel(ctx) {
string interpolation_str;
OP_REQUIRES_OK(ctx, ctx->GetAttr("interpolation", &interpolation_str));
if (interpolation_str == "NEAREST") {
interpolation_ = INTERPOLATION_NEAREST;
} else if (interpolation_str == "BILINEAR") {
interpolation_ = INTERPOLATION_BILINEAR;
} else {
LOG(FATAL) << "Invalid interpolation " << interpolation_str
<< ". Supported types: NEAREST, BILINEAR";
}
}
void Compute(OpKernelContext* ctx) override {
const Tensor& images_t = ctx->input(0);
const Tensor& transform_t = ctx->input(1);
const Tensor& shape_t = ctx->input(2);
OP_REQUIRES(ctx, images_t.shape().dims() == 4,
errors::InvalidArgument("Input images must have rank 4"));
OP_REQUIRES(ctx,
(TensorShapeUtils::IsMatrix(transform_t.shape()) &&
(transform_t.dim_size(0) == images_t.dim_size(0) ||
transform_t.dim_size(0) == 1) &&
transform_t.dim_size(1) ==
ProjectiveGenerator<Device, T>::kNumParameters),
errors::InvalidArgument(
"Input transform should be num_images x 8 or 1 x 8"));
OP_REQUIRES(ctx, shape_t.dims() == 1,
errors::InvalidArgument("output shape must be 1-dimensional",
shape_t.shape().DebugString()));
OP_REQUIRES(ctx, shape_t.NumElements() == 2,
errors::InvalidArgument("output shape must have two elements",
shape_t.shape().DebugString()));
auto shape_vec = shape_t.vec<int32>();
int32 out_height = shape_vec(0);
int32 out_width = shape_vec(1);
OP_REQUIRES(ctx, out_height > 0 && out_width > 0,
errors::InvalidArgument("output dimensions must be positive"));
Tensor* output_t;
OP_REQUIRES_OK(ctx, ctx->allocate_output(
0,
TensorShape({images_t.dim_size(0), out_height,
out_width, images_t.dim_size(3)}),
&output_t));
auto output = output_t->tensor<T, 4>();
auto images = images_t.tensor<T, 4>();
auto transform = transform_t.matrix<float>();
(FillProjectiveTransform<Device, T>(interpolation_))(
ctx->eigen_device<Device>(), &output, images, transform);
}
};
#define REGISTER(TYPE) \
REGISTER_KERNEL_BUILDER(Name("ImageProjectiveTransform") \
.Device(DEVICE_CPU) \
.TypeConstraint<TYPE>("dtype"), \
ImageProjectiveTransform<CPUDevice, TYPE>)
TF_CALL_uint8(REGISTER);
TF_CALL_int32(REGISTER);
TF_CALL_int64(REGISTER);
TF_CALL_half(REGISTER);
TF_CALL_float(REGISTER);
TF_CALL_double(REGISTER);
#undef REGISTER
#if GOOGLE_CUDA
typedef Eigen::GpuDevice GPUDevice;
namespace functor {
// NOTE(ringwalt): We get an undefined symbol error if we don't explicitly
// instantiate the operator() in GCC'd code.
#define DECLARE_FUNCTOR(TYPE) \
template <> \
void FillProjectiveTransform<GPUDevice, TYPE>::operator()( \
const GPUDevice& device, OutputType* output, const InputType& images, \
const TransformsType& transform) const; \
extern template struct FillProjectiveTransform<GPUDevice, TYPE>
TF_CALL_uint8(DECLARE_FUNCTOR);
TF_CALL_int32(DECLARE_FUNCTOR);
TF_CALL_int64(DECLARE_FUNCTOR);
TF_CALL_float(DECLARE_FUNCTOR);
TF_CALL_double(DECLARE_FUNCTOR);
} // end namespace functor
#define REGISTER(TYPE) \
REGISTER_KERNEL_BUILDER(Name("ImageProjectiveTransform") \
.Device(DEVICE_GPU) \
.TypeConstraint<TYPE>("dtype") \
.HostMemory("output_shape"), \
ImageProjectiveTransform<GPUDevice, TYPE>)
TF_CALL_uint8(REGISTER);
TF_CALL_int32(REGISTER);
TF_CALL_int64(REGISTER);
TF_CALL_float(REGISTER);
TF_CALL_double(REGISTER);
#undef REGISTER
#endif // GOOGLE_CUDA
} // end namespace tensorflow
|
