• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2016 The TensorFlow Authors. All Rights Reserved.
2 
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6 
7     http://www.apache.org/licenses/LICENSE-2.0
8 
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations 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 
25 namespace tensorflow {
26 namespace 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.
FoldBatchNorms(const GraphDef & input_graph_def,const TransformFuncContext & context,GraphDef * output_graph_def)31 Status 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 Status::OK();
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 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 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 Status::OK();
116       },
117       {}, &replaced_graph_def));
118   *output_graph_def = replaced_graph_def;
119   return Status::OK();
120 }
121 
122 REGISTER_GRAPH_TRANSFORM("fold_batch_norms", FoldBatchNorms);
123 
124 }  // namespace graph_transforms
125 }  // namespace tensorflow
126