1/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2
3Licensed under the Apache License, Version 2.0 (the "License");
4you may not use this file except in compliance with the License.
5You may obtain a copy of the License at
6
7 http://www.apache.org/licenses/LICENSE-2.0
8
9Unless required by applicable law or agreed to in writing, software
10distributed under the License is distributed on an "AS IS" BASIS,
11WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12See the License for the specific language governing permissions and
13limitations under the License.
14==============================================================================*/
15
16#include "tensorflow/core/common_runtime/graph_constructor.h"
17#include "tensorflow/core/graph/node_builder.h"
18#include "tensorflow/core/graph/subgraph.h"
19#include "tensorflow/core/platform/init_main.h"
20#include "tensorflow/core/public/session.h"
21#include "tensorflow/tools/graph_transforms/transform_utils.h"
22
23namespace tensorflow {
24namespace graph_transforms {
25
26Status FlattenAtrousConv(const GraphDef& input_graph_def,
27 const TransformFuncContext& context,
28 GraphDef* output_graph_def) {
29 GraphDef replaced_graph_def;
30 TF_RETURN_IF_ERROR(ReplaceMatchingOpTypes(
31 input_graph_def, // clang-format off
32 {"BatchToSpaceND",
33 {
34 {"Conv2D|DepthwiseConv2dNative",
35 {
36 {"SpaceToBatchND",
37 {
38 {"*"}, // Input to the flattened op.
39 {"*"}, // block_shape
40 {"*"} // paddings
41 }
42 },
43 {"*"} // filter
44 }
45 },
46 {"*"}, // block_shape
47 {"*"} // crops
48 }
49 }, // clang-format on
50 [](const NodeMatch& match, const std::set<string>& input_nodes,
51 const std::set<string>& output_nodes,
52 std::vector<NodeDef>* new_nodes) {
53 // Find all the nodes we expect in the subgraph.
54 const NodeDef& batch_to_space_node = match.node;
55 const NodeDef& conv_node = match.inputs[0].node;
56 const NodeDef& filter_node = match.inputs[0].inputs[1].node;
57 const NodeDef& input_node = match.inputs[0].inputs[0].inputs[0].node;
58 const NodeDef& space_to_batch_block_shape_node =
59 match.inputs[0].inputs[0].inputs[1].node;
60
61 // The atrous rate value is inferred from the block shape.
62 Tensor block_shape =
63 GetNodeTensorAttr(space_to_batch_block_shape_node, "value");
64 const int32_t block_height = block_shape.flat<int32>()(0);
65 const int32_t block_width = block_shape.flat<int32>()(1);
66
67 // Compute the upsampled filter.
68 const Tensor& filter = GetNodeTensorAttr(filter_node, "value");
69 const int32_t filter_height = filter.dim_size(0);
70 const int32_t filter_width = filter.dim_size(1);
71 const int32_t in_channels = filter.dim_size(2);
72 const int32_t out_channels = filter.dim_size(3);
73
74 const int32_t upsampled_filter_height =
75 (filter_height - 1) * block_height + 1;
76 const int32_t upsampled_filter_width =
77 (filter_width - 1) * block_width + 1;
78 Tensor upsampled_filter(
79 DT_FLOAT,
80 TensorShape({upsampled_filter_height, upsampled_filter_width,
81 in_channels, out_channels}));
82
83 auto filter_eigen = filter.tensor<float, 4>();
84 auto upsampled_filter_eigen = upsampled_filter.tensor<float, 4>();
85
86 upsampled_filter_eigen.setZero();
87 for (int h = 0; h < filter_height; ++h) {
88 for (int w = 0; w < filter_width; ++w) {
89 for (int c_in = 0; c_in < in_channels; ++c_in) {
90 for (int c_out = 0; c_out < out_channels; ++c_out) {
91 upsampled_filter_eigen(block_height * h, block_width * w, c_in,
92 c_out) = filter_eigen(h, w, c_in, c_out);
93 }
94 }
95 }
96 }
97
98 NodeDef upsampled_filter_node;
99 upsampled_filter_node.set_op("Const");
100 upsampled_filter_node.set_name(filter_node.name());
101 SetNodeAttr("dtype", DT_FLOAT, &upsampled_filter_node);
102 SetNodeTensorAttr<float>("value", upsampled_filter,
103 &upsampled_filter_node);
104
105 // Set up the new flattened version of the convolution op.
106 NodeDef flattened_conv_node;
107
108 flattened_conv_node.set_name(batch_to_space_node.name());
109 flattened_conv_node.set_op(conv_node.op());
110 flattened_conv_node.set_device(conv_node.device());
111
112 AddNodeInput(input_node.name(), &flattened_conv_node);
113 AddNodeInput(upsampled_filter_node.name(), &flattened_conv_node);
114
115 CopyNodeAttr(conv_node, "T", "T", &flattened_conv_node);
116 CopyNodeAttr(conv_node, "strides", "strides", &flattened_conv_node);
117 SetNodeAttr("padding", "SAME", &flattened_conv_node);
118 CopyNodeAttr(conv_node, "data_format", "data_format",
119 &flattened_conv_node);
120
121 if (conv_node.op() == "Conv2D") {
122 CopyNodeAttr(conv_node, "use_cudnn_on_gpu", "use_cudnn_on_gpu",
123 &flattened_conv_node);
124 }
125
126 new_nodes->push_back(input_node);
127 new_nodes->push_back(upsampled_filter_node);
128 new_nodes->push_back(flattened_conv_node);
129
130 return OkStatus();
131 },
132 {}, &replaced_graph_def));
133 *output_graph_def = replaced_graph_def;
134 return OkStatus();
135}
136
137REGISTER_GRAPH_TRANSFORM("flatten_atrous_conv", FlattenAtrousConv);
138
139} // namespace graph_transforms
140} // namespace tensorflow
141