1 /* Copyright 2017 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 #ifndef TENSORFLOW_LITE_TOCO_TFLITE_OPERATOR_H_ 16 #define TENSORFLOW_LITE_TOCO_TFLITE_OPERATOR_H_ 17 18 #include <string> 19 20 #include "flatbuffers/flatbuffers.h" 21 #include "flatbuffers/flexbuffers.h" 22 #include "tensorflow/lite/schema/schema_generated.h" 23 #include "tensorflow/lite/toco/model.h" 24 #include "tensorflow/lite/tools/versioning/op_version.h" 25 26 namespace toco { 27 28 namespace tflite { 29 30 class BaseOperator; 31 32 // Return a map contained all know TF Lite Operators, keyed by their names. 33 // TODO(ycling): The pattern to propagate parameters (e.g. enable_select_tf_ops) 34 // is ugly here. Consider refactoring. 35 std::map<std::string, std::unique_ptr<BaseOperator>> BuildOperatorByNameMap( 36 bool enable_select_tf_ops = false); 37 38 // Return a map contained all know TF Lite Operators, keyed by the type of 39 // their tf.mini counterparts. 40 std::map<OperatorType, std::unique_ptr<BaseOperator>> BuildOperatorByTypeMap( 41 bool enable_select_tf_ops = false); 42 43 // Write the custom option FlexBuffer with a serialized TensorFlow NodeDef 44 // for a Flex op. 45 std::unique_ptr<flexbuffers::Builder> WriteFlexOpOptions( 46 const std::string& tensorflow_node_def); 47 48 // These are the flatbuffer types for custom and builtin options. 49 using CustomOptions = flatbuffers::Vector<uint8_t>; 50 using BuiltinOptions = void; 51 52 // A simple wrapper around the flatbuffer objects used to describe options that 53 // configure operators. 54 struct Options { 55 // Build custom options. CustomOptions56 static Options Custom(flatbuffers::Offset<CustomOptions> offset) { 57 return {::tflite::BuiltinOptions_NONE, 0, offset}; 58 } 59 60 // Build builtin options of the given type. BuiltinOptions61 static Options Builtin(::tflite::BuiltinOptions type, 62 flatbuffers::Offset<BuiltinOptions> offset) { 63 return {type, offset, 0}; 64 } 65 66 ::tflite::BuiltinOptions type; 67 flatbuffers::Offset<BuiltinOptions> builtin; 68 flatbuffers::Offset<CustomOptions> custom; 69 }; 70 71 // A BaseOperator encapsulates the relationship between operators in tf.mini 72 // and TF lite, and provides methods for converting between those two formats. 73 class BaseOperator { 74 public: 75 // Build an operator with the given TF Lite name and tf.mini type. BaseOperator(const std::string & name,OperatorType type)76 BaseOperator(const std::string& name, OperatorType type) 77 : name_(name), type_(type) {} 78 virtual ~BaseOperator() = default; 79 name()80 std::string name() const { return name_; } type()81 OperatorType type() const { return type_; } 82 83 // Given a tf.mini operator, create the corresponding flatbuffer options and 84 // return their offsets. 85 virtual Options Serialize(const Operator& op, 86 flatbuffers::FlatBufferBuilder* builder) const = 0; 87 88 // Read TF Lite options and create the appropriate tf.mini operator. 89 virtual std::unique_ptr<Operator> Deserialize( 90 const BuiltinOptions* builtin_options, 91 const CustomOptions* custom_options) const = 0; 92 93 // Get the op version using the OperatorSignature. 94 // The function needs to be overridden to return the op version based on the 95 // parameters. Note: 96 // * The first version for each op should be 1 (to be consistent with the 97 // default value in Flatbuffer. `return 1;` is okay for newly implemented 98 // ops. 99 // * When multiple versions are defined for an op, this function could be 100 // overridden. (See example in `operator_test.cc` and 101 // 'tools/versioning/op_version.cc`) 102 virtual int GetVersion(const OperatorSignature& op_signature) const = 0; 103 104 // Given a Toco `Operator`, return a list of booleans indicating the op 105 // mutates which input variables. 106 // * If the op mutates any input variables, it should return a list of bool 107 // with the same length as inputs. 108 // * Otherwise, it will return an empty list. GetMutatingInputVariables(const Operator & op)109 virtual std::vector<bool> GetMutatingInputVariables( 110 const Operator& op) const { 111 // Most ops don't have variable tensors. This function can be overridden. 112 return std::vector<bool>(); 113 } 114 115 private: 116 std::string name_; 117 OperatorType type_; 118 }; 119 120 // Helper function to create ::tflite::OpSignature from the given 121 // ::tflite::BuiltinOperator and OperatorSignature. 122 ::tflite::OpSignature GetVersioningOpSig(const ::tflite::BuiltinOperator op, 123 const OperatorSignature& op_signature); 124 125 // Helper function to determine if a unsupported TensorFlow op should be 126 // exported as an Flex op or a regular custom op. 127 bool ShouldExportAsFlexOp(bool enable_select_tf_ops, 128 const std::string& tensorflow_op_name); 129 130 } // namespace tflite 131 132 } // namespace toco 133 134 #endif // TENSORFLOW_LITE_TOCO_TFLITE_OPERATOR_H_ 135