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 16 #ifndef TENSORFLOW_CORE_GRAPPLER_GRAPPLER_ITEM_H_ 17 #define TENSORFLOW_CORE_GRAPPLER_GRAPPLER_ITEM_H_ 18 19 #include <memory> 20 #include <string> 21 #include <unordered_map> 22 #include <unordered_set> 23 #include <utility> 24 #include <vector> 25 26 #include "tensorflow/core/framework/graph.pb.h" 27 #include "tensorflow/core/framework/tensor.h" 28 #include "tensorflow/core/framework/variable.pb.h" 29 #include "tensorflow/core/protobuf/queue_runner.pb.h" 30 31 namespace tensorflow { 32 namespace grappler { 33 34 // A TensorFlow model to optimize. 35 // Models are represented by the combination of a graph, one of more fetch 36 // nodes, and potentially a set of nodes to feed. 37 struct GrapplerItem { 38 GrapplerItem() = default; 39 GrapplerItem(const GrapplerItem& other) = default; 40 GrapplerItem(GrapplerItem&& other) = default; 41 GrapplerItem& operator=(const GrapplerItem& other) = default; 42 GrapplerItem& operator=(GrapplerItem&& other) = default; 43 virtual ~GrapplerItem() = default; 44 45 // Create a copy of this GrapplerItem with graph swapped with the argument. 46 GrapplerItem WithGraph(GraphDef&& graph) const; 47 48 string id; // A unique id for this item 49 50 // Inputs 51 GraphDef graph; 52 std::vector<std::pair<string, Tensor>> feed; 53 std::vector<string> fetch; 54 55 // Initialization op(s). 56 std::vector<string> init_ops; 57 // Expected initialization time in seconds, or 0 if unknown 58 int64 expected_init_time = 0; 59 60 // Save/restore ops (if any) 61 string save_op; 62 string restore_op; 63 string save_restore_loc_tensor; 64 65 // Queue runner(s) required to run the queue(s) of this model. 66 std::vector<QueueRunnerDef> queue_runners; 67 68 // List of op names to keep in the graph. This includes nodes that are 69 // referenced in various collections, and therefore must be preserved to 70 // ensure that the optimized metagraph can still be loaded. 71 std::vector<string> keep_ops; 72 73 // Return the set of node evaluated during a regular train/inference step. 74 std::vector<const NodeDef*> MainOpsFanin() const; 75 // Return the set of node run to populate the queues (if any). 76 std::vector<const NodeDef*> EnqueueOpsFanin() const; 77 // Return the set nodes used by TensorFlow to initialize the graph. 78 std::vector<const NodeDef*> InitOpsFanin() const; 79 // Return the set of variables accessed during a regular train/inference step. 80 std::vector<const NodeDef*> MainVariables() const; 81 // Return a set of node names that must be preserved. This includes feed and 82 // fetch nodes, keep_ops, init_ops. 83 std::unordered_set<string> NodesToPreserve() const; 84 85 struct OptimizationOptions { 86 // Is it allowed to add nodes to the graph that do not have registered 87 // gradient function. 88 bool allow_non_differentiable_rewrites = true; 89 90 // Tensorflow function execution semantics is slightly different from the 91 // main Tensorflow graph, and we need to make sure that we do not change it 92 // by running Grappler optimizer passes. One main difference is that 93 // functions do not prune ops with side-effects and dataset-output ops (see 94 // PruneFunctionBody in common_runtime/function.cc). 95 bool allow_pruning_stateful_and_dataset_ops = true; 96 97 // If true Grappler will optimize the main graph, and also all functions in 98 // the graph function library (function can't be polymorphic, it can't have 99 // undefined type parameters in the function signature, or placeholder 100 // attributes in the function body). 101 bool optimize_function_library = true; 102 103 // Mark the grapper optimization run in eager mode or not. 104 bool is_eager_mode = false; 105 }; 106 107 const std::unordered_set<string>& devices() const; 108 // Adds a device to a set of available devices, only if it's a valid fully 109 // defined device name. Returns `Status::OK()` if successfully added a device, 110 // and an error otherwise. 111 Status AddDevice(const string& device); 112 // Adds all valid devices from the other Grappler item to the device set. 113 Status AddDevices(const GrapplerItem& other); 114 // Adds all valid devices from the nodes of the graph to the device set. 115 // Returns `Status::OK()` if all device annotations found in a graph are valid 116 // fully defined device names, and an error otherwise. 117 Status InferDevicesFromGraph(); 118 // Clears a set of available devices. 119 void ClearDevices(); 120 121 const OptimizationOptions& optimization_options() const; 122 OptimizationOptions& optimization_options(); 123 124 private: 125 // TODO(ezhulenev) Make GrapplerItem a class and hide all public data members. 126 // TODO(ezhulenev): Migrate all unordered collections to absl. 127 128 // A set of fully defined device names that can be used to place the nodes of 129 // the `graph`. 130 // Example of a fully defined name: "/job:work/replica:1/task:1/device:CPU:0" 131 std::unordered_set<string> devices_; 132 133 OptimizationOptions optimization_options_; 134 }; 135 136 GrapplerItem::OptimizationOptions CreateOptOptionsForEager(); 137 138 } // end namespace grappler 139 } // end namespace tensorflow 140 141 #endif // TENSORFLOW_CORE_GRAPPLER_GRAPPLER_ITEM_H_ 142