1 /* Copyright 2015 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 // Class and associated machinery for specifying an Op's OpDef and shape 17 // inference function for Op registration. 18 19 #ifndef TENSORFLOW_CORE_FRAMEWORK_OP_DEF_BUILDER_H_ 20 #define TENSORFLOW_CORE_FRAMEWORK_OP_DEF_BUILDER_H_ 21 22 #include <string> 23 #include <utility> 24 #include <vector> 25 26 #include "tensorflow/core/framework/full_type.pb.h" 27 #include "tensorflow/core/framework/op_def.pb.h" 28 #include "tensorflow/core/framework/types.h" 29 #include "tensorflow/core/lib/core/status.h" 30 #include "tensorflow/core/lib/core/stringpiece.h" 31 #include "tensorflow/core/platform/macros.h" 32 33 namespace tensorflow { 34 35 // TODO(b/62899350): Refactor without proto dependencies. 36 typedef std::function<Status(OpDef* c)> OpTypeConstructor; 37 38 typedef std::vector<std::reference_wrapper<const FullTypeDef>> TypeRefVector; 39 typedef std::map<std::string, std::reference_wrapper<const FullTypeDef>> 40 TypeRefMap; 41 42 // A type inference function, called for each node during type inference 43 // (possibly multiple times). 44 // The first argument (input_types) will hold the type of each of the node's 45 // inputs. The second argument (type_vars) will hold the return type of 46 // each function referred from any type variable (e.g. `FuncVar`) present 47 // in the node's corresponding op definition. 48 // 49 // TODO(mdan): Consider a vector-in, vector-out contract. 50 // TODO(mdan): Rename to just TypeInferenceFn (since it's not always "forward"). 51 typedef std::function<StatusOr<FullTypeDef>(const TypeRefVector&, 52 const TypeRefMap&)> 53 ForwardTypeInferenceFn; 54 55 class FunctionDefHelper; 56 57 namespace shape_inference { 58 class InferenceContext; 59 } 60 typedef std::function<Status(shape_inference::InferenceContext* c)> 61 OpShapeInferenceFn; 62 63 struct OpRegistrationData { 64 public: OpRegistrationDataOpRegistrationData65 OpRegistrationData() {} OpRegistrationDataOpRegistrationData66 OpRegistrationData(const OpDef& def) : op_def(def) {} 67 OpRegistrationData(const OpDef& def, const OpShapeInferenceFn& fn, 68 bool is_function = false) op_defOpRegistrationData69 : op_def(def), shape_inference_fn(fn), is_function_op(is_function) {} 70 71 OpDef op_def; 72 OpShapeInferenceFn shape_inference_fn; 73 74 // Type constructor. This callable initializes the type of this op. 75 // It is provided as a programmatic mechanism for defining an op's 76 // type, as part of its registration. It is to be eventually replaced by a 77 // textual language. 78 // 79 // Important: historically, op registrations only contained partial 80 // input/output type information in non-standardized attribute declarations 81 // (e.g. typically, input types were held in a `dtype` attribute). The type 82 // constructor currently duplicates such attribute information, with the aim 83 // of entirely subsuming it, and eventually deprecating all type-related 84 // attributes. 85 // 86 // Since ops are typically parametrized, the type created by this constructor 87 // is also parametric. 88 // 89 // Example: for an op `Foo(x: T) -> Bar[T]`: 90 // 91 // * typically, its op registration included a single attribute `T: type`; 92 // then the respective input was defined as `x: T`; the output type `Bar` 93 // was implied by the op name. 94 // * the type constructor creates a FullType object containing `Bar[T]`; this 95 // still relies on the `T` attribute which it references. 96 // * in the future, the type constructor will create a FullType containing 97 // `Callable[(x: T), Bar[T]]`, and the attribute `T` will be deprecated. 98 OpTypeConstructor type_ctor; 99 100 // Forward type inference function. This callable infers the return type of an 101 // op based on its input types. 102 // 103 // Note that the type constructor and forward inference functions need not be 104 // mutually exclusive: if there is some static information that can be set 105 // based on attributes, then that should be set in the constructor. If more 106 // information can be extracted from inputs, that should be done in the 107 // forward inference function. 108 // 109 // This is similar to the shape function, but is more general, and applied 110 // directly to NodeDefs, rather than working on the ShapeAndType structures. 111 // Note that the op input/output declarations may specify some implicit type 112 // constraints through attribute references (i.e. two inputs pointing to the 113 // same type attribute). Those constraints may duplicate what this function 114 // specifies in its body. That's intended, for a gradual transition to a more 115 // formal type system. 116 // 117 // These type inference functions are intermediate solutions as well: once the 118 // op registration has a complete, formal type definition, along with 119 // a solver-based type inference, it will replace these functions. 120 // 121 // TODO(mdan): Merge with shape inference. 122 // TODO(mdan): Replace with a union-based type inference algorithm. 123 ForwardTypeInferenceFn fwd_type_fn; 124 125 // Reverse type inference function. This callable infers some input types 126 // based on the return type. 127 // 128 // TODO(mdan): Replace with a union-based type inference algorithm. 129 ForwardTypeInferenceFn rev_type_fn; 130 131 // The input number affected by reverse type inference. Only one input may be 132 // updated in this manner. 133 // TODO(mdan): Encode in a manner more consistent with the forward version. 134 int rev_type_input; 135 136 bool is_function_op = false; 137 }; 138 139 // Builder class passed to the REGISTER_OP() macro. 140 class OpDefBuilder { 141 public: 142 // Constructs an OpDef with just the name field set. 143 explicit OpDefBuilder(std::string op_name); 144 145 // Adds an attr to this OpDefBuilder (and returns *this). The spec has 146 // format "<name>:<type>" or "<name>:<type>=<default>" 147 // where <name> matches regexp [a-zA-Z][a-zA-Z0-9_]* 148 // (by convention only using capital letters for attrs that can be inferred) 149 // <type> can be: 150 // "string", "int", "float", "bool", "type", "shape", or "tensor" 151 // "numbertype", "realnumbertype", "quantizedtype" 152 // (meaning "type" with a restriction on valid values) 153 // "{int32,int64}" or {realnumbertype,quantizedtype,string}" 154 // (meaning "type" with a restriction containing unions of value types) 155 // "{\"foo\", \"bar\n baz\"}", or "{'foo', 'bar\n baz'}" 156 // (meaning "string" with a restriction on valid values) 157 // "list(string)", ..., "list(tensor)", "list(numbertype)", ... 158 // (meaning lists of the above types) 159 // "int >= 2" (meaning "int" with a restriction on valid values) 160 // "list(string) >= 2", "list(int) >= 2" 161 // (meaning "list(string)" / "list(int)" with length at least 2) 162 // <default>, if included, should use the Proto text format 163 // of <type>. For lists use [a, b, c] format. 164 // 165 // Note that any attr specifying the length of an input or output will 166 // get a default minimum of 1 unless the >= # syntax is used. 167 // 168 // TODO(josh11b): Perhaps support restrictions and defaults as optional 169 // extra arguments to Attr() instead of encoding them in the spec string. 170 // TODO(josh11b): Would like to have better dtype handling for tensor attrs: 171 // * Ability to say the type of an input/output matches the type of 172 // the tensor. 173 // * Ability to restrict the type of the tensor like the existing 174 // restrictions for type attrs. 175 // Perhaps by linking the type of the tensor to a type attr? 176 OpDefBuilder& Attr(std::string spec); 177 178 // Adds an input or output to this OpDefBuilder (and returns *this). 179 // The spec has form "<name>:<type-expr>" or "<name>:Ref(<type-expr>)" 180 // where <name> matches regexp [a-z][a-z0-9_]* and <type-expr> can be: 181 // * For a single tensor: <type> 182 // * For a sequence of tensors with the same type: <number>*<type> 183 // * For a sequence of tensors with different types: <type-list> 184 // Where: 185 // <type> is either one of "float", "int32", "string", ... 186 // or the name of an attr (see above) with type "type". 187 // <number> is the name of an attr with type "int". 188 // <type-list> is the name of an attr with type "list(type)". 189 // TODO(josh11b): Indicate Ref() via an optional argument instead of 190 // in the spec? 191 // TODO(josh11b): SparseInput() and SparseOutput() matching the Python 192 // handling? 193 OpDefBuilder& Input(std::string spec); 194 OpDefBuilder& Output(std::string spec); 195 196 // Turns on the indicated boolean flag in this OpDefBuilder (and 197 // returns *this). 198 OpDefBuilder& SetIsCommutative(); 199 OpDefBuilder& SetIsAggregate(); 200 OpDefBuilder& SetIsStateful(); 201 OpDefBuilder& SetAllowsUninitializedInput(); 202 OpDefBuilder& SetIsDistributedCommunication(); 203 204 // Deprecate the op at a certain GraphDef version. 205 OpDefBuilder& Deprecated(int version, std::string explanation); 206 207 // Adds docs to this OpDefBuilder (and returns *this). 208 // Docs have the format: 209 // <1-line summary> 210 // <rest of the description> 211 // <name>: <description of name> 212 // <name>: <description of name> 213 // <if long, indent the description on subsequent lines> 214 // Where <name> is the name of an attr, input, or output. Please 215 // wrap docs at 72 columns so that it may be indented in the 216 // generated output. For tensor inputs or outputs (not attrs), you 217 // may start the description with an "=" (like name:= <description>) 218 // to suppress the automatically-generated type documentation in 219 // generated output. 220 OpDefBuilder& Doc(std::string text); 221 222 // Sets the function to be used as type constructor. 223 // See OpRegistrationData::type_ctor. 224 OpDefBuilder& SetTypeConstructor(OpTypeConstructor c); 225 226 // Sets the function to be used for forward type inference. 227 // See OpRegistrationData::fwd_type_fn. 228 OpDefBuilder& SetForwardTypeFn(ForwardTypeInferenceFn f); 229 230 // Sets the function to be used for reverse type inference. 231 // See OpRegistrationData::rew_type_fn. 232 OpDefBuilder& SetReverseTypeFn(int input_number, ForwardTypeInferenceFn f); 233 234 // Sets the shape function to be used for shape inference. 235 // 236 // Note that currently (October 2016), python code still requires a 237 // RegisterShape call to invoke this; see call_cpp_shape_fn in 238 // python/framework/common_shapes.py 239 OpDefBuilder& SetShapeFn(OpShapeInferenceFn fn); 240 241 // Allows the `<type>` in calls to `Attr()` to be "any". 242 // This is used by PythonAPIWrapper for pass-through parameters. 243 OpDefBuilder& AllowAttrTypeAny(); 244 245 // Sets op_reg_data->op_def to the requested OpDef and 246 // op_reg_data->shape_inference_fn to the requested shape inference function, 247 // or returns an error. 248 // Must be called after all of the above methods. 249 // 250 // Note that OpDefBuilder only reports parsing errors. You should also 251 // call ValidateOpDef() to detect other problems. 252 Status Finalize(OpRegistrationData* op_reg_data) const; 253 254 private: 255 friend class FunctionDefHelper; 256 257 // Adds control output to this OpDefBuilder (and returns *this). 258 // The <name> must be a valid node name (matches regexp 259 // [a-zA-Z][a-zA-Z0-9_]*). Named control output can only exist for functions. 260 OpDefBuilder& ControlOutput(std::string name); 261 op_def()262 OpDef* op_def() { return &op_reg_data_.op_def; } 263 264 OpRegistrationData op_reg_data_; 265 std::vector<string> attrs_; 266 std::vector<string> inputs_; 267 std::vector<string> outputs_; 268 std::vector<string> control_outputs_; 269 std::string doc_; 270 std::vector<string> errors_; 271 bool allow_attr_type_any_ = false; 272 }; 273 274 } // namespace tensorflow 275 276 #endif // TENSORFLOW_CORE_FRAMEWORK_OP_DEF_BUILDER_H_ 277