1/* Copyright 2016 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/constant_folding.h"
17#include "tensorflow/core/common_runtime/graph_constructor.h"
18#include "tensorflow/core/graph/node_builder.h"
19#include "tensorflow/core/graph/subgraph.h"
20#include "tensorflow/core/platform/init_main.h"
21#include "tensorflow/core/public/session.h"
22#include "tensorflow/tools/graph_transforms/fold_constants_lib.h"
23#include "tensorflow/tools/graph_transforms/transform_utils.h"
24
25namespace tensorflow {
26namespace graph_transforms {
27
28// Converts Conv2D or MatMul ops followed by column-wise Muls into equivalent
29// ops with the Mul baked into the convolution weights, to save computation
30// during inference.
31Status FoldBatchNorms(const GraphDef& input_graph_def,
32 const TransformFuncContext& context,
33 GraphDef* output_graph_def) {
34 GraphDef replaced_graph_def;
35 TF_RETURN_IF_ERROR(ReplaceMatchingOpTypes(
36 input_graph_def, // clang-format off
37 {"Mul", // mul_node
38 {
39 {"Conv2D|MatMul|DepthwiseConv2dNative", // conv_node
40 {
41 {"*"}, // input_node
42 {"Const"}, // weights_node
43 }
44 },
45 {"Const"}, // mul_values_node
46 }
47 }, // clang-format on
48 [](const NodeMatch& match, const std::set<string>& input_nodes,
49 const std::set<string>& output_nodes,
50 std::vector<NodeDef>* new_nodes) {
51 // Find all the nodes we expect in the subgraph.
52 const NodeDef& mul_node = match.node;
53 const NodeDef& conv_node = match.inputs[0].node;
54 const NodeDef& input_node = match.inputs[0].inputs[0].node;
55 const NodeDef& weights_node = match.inputs[0].inputs[1].node;
56 const NodeDef& mul_values_node = match.inputs[1].node;
57
58 // Check that nodes that we use are not used somewhere else.
59 for (const auto& node : {conv_node, weights_node, mul_values_node}) {
60 if (output_nodes.count(node.name())) {
61 // Return original nodes.
62 new_nodes->insert(new_nodes->end(),
63 {mul_node, conv_node, input_node, weights_node,
64 mul_values_node});
65 return OkStatus();
66 }
67 }
68
69 Tensor weights = GetNodeTensorAttr(weights_node, "value");
70 Tensor mul_values = GetNodeTensorAttr(mul_values_node, "value");
71
72 // Make sure all the inputs really are vectors, with as many entries as
73 // there are columns in the weights.
74 int64_t weights_cols;
75 if (conv_node.op() == "Conv2D") {
76 weights_cols = weights.shape().dim_size(3);
77 } else if (conv_node.op() == "DepthwiseConv2dNative") {
78 weights_cols =
79 weights.shape().dim_size(2) * weights.shape().dim_size(3);
80 } else {
81 weights_cols = weights.shape().dim_size(1);
82 }
83 if ((mul_values.shape().dims() != 1) ||
84 (mul_values.shape().dim_size(0) != weights_cols)) {
85 return errors::InvalidArgument(
86 "Mul constant input to batch norm has bad shape: ",
87 mul_values.shape().DebugString());
88 }
89
90 // Multiply the original weights by the scale vector.
91 auto weights_vector = weights.flat<float>();
92 Tensor scaled_weights(DT_FLOAT, weights.shape());
93 auto scaled_weights_vector = scaled_weights.flat<float>();
94 for (int64_t row = 0; row < weights_vector.dimension(0); ++row) {
95 scaled_weights_vector(row) =
96 weights_vector(row) *
97 mul_values.flat<float>()(row % weights_cols);
98 }
99
100 // Construct the new nodes.
101 NodeDef scaled_weights_node;
102 scaled_weights_node.set_op("Const");
103 scaled_weights_node.set_name(weights_node.name());
104 SetNodeAttr("dtype", DT_FLOAT, &scaled_weights_node);
105 SetNodeTensorAttr<float>("value", scaled_weights, &scaled_weights_node);
106 new_nodes->push_back(scaled_weights_node);
107
108 new_nodes->push_back(input_node);
109
110 NodeDef new_conv_node;
111 new_conv_node = conv_node;
112 new_conv_node.set_name(mul_node.name());
113 new_nodes->push_back(new_conv_node);
114
115 return OkStatus();
116 },
117 {}, &replaced_graph_def));
118 *output_graph_def = replaced_graph_def;
119 return OkStatus();
120}
121
122REGISTER_GRAPH_TRANSFORM("fold_batch_norms", FoldBatchNorms);
123
124} // namespace graph_transforms
125} // namespace tensorflow
126