• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_set>
22 #include <utility>
23 #include <vector>
24 
25 #include "tensorflow/core/framework/graph.pb.h"
26 #include "tensorflow/core/framework/tensor.h"
27 #include "tensorflow/core/framework/variable.pb.h"
28 #include "tensorflow/core/protobuf/queue_runner.pb.h"
29 
30 namespace tensorflow {
31 namespace grappler {
32 
33 // A TensorFlow model to optimize.
34 // Models are represented by the combination of a graph, one of more fetch
35 // nodes, and potentially a set of nodes to feed.
36 struct GrapplerItem {
37   GrapplerItem() = default;
38   GrapplerItem(const GrapplerItem& other) = default;
39   GrapplerItem(GrapplerItem&& other) = default;
40   GrapplerItem& operator=(const GrapplerItem& other) = default;
41   GrapplerItem& operator=(GrapplerItem&& other) = default;
42   virtual ~GrapplerItem() = default;
43 
44   // Create a copy of this GrapplerItem with graph swapped with the argument.
45   GrapplerItem WithGraph(GraphDef&& graph) const;
46 
47   string id;  // A unique id for this item
48 
49   // Inputs
50   GraphDef graph;
51   std::vector<std::pair<string, Tensor>> feed;
52   std::vector<string> fetch;
53 
54   // Initialization op(s).
55   std::vector<string> init_ops;
56   // Expected initialization time in seconds, or 0 if unknown
57   int64 expected_init_time = 0;
58 
59   // Save/restore ops (if any)
60   string save_op;
61   string restore_op;
62   string save_restore_loc_tensor;
63 
64   // Queue runner(s) required to run the queue(s) of this model.
65   std::vector<QueueRunnerDef> queue_runners;
66 
67   // List of op names to keep in the graph. This includes nodes that are
68   // referenced in various collections, and therefore must be preserved to
69   // ensure that the optimized metagraph can still be loaded.
70   std::vector<string> keep_ops;
71 
72   // Return the set of node evaluated during a regular train/inference step.
73   std::vector<const NodeDef*> MainOpsFanin() const;
74   // Return the set of node run to populate the queues (if any).
75   std::vector<const NodeDef*> EnqueueOpsFanin() const;
76   // Return the set nodes used by TensorFlow to initialize the graph.
77   std::vector<const NodeDef*> InitOpsFanin() const;
78   // Return the set of variables accessed during a regular train/inference step.
79   std::vector<const NodeDef*> MainVariables() const;
80   // Return a set of node names that must be preserved. This includes feed and
81   // fetch nodes, keep_ops, init_ops.
82   std::unordered_set<string> NodesToPreserve() const;
83 
84   struct OptimizationOptions {
85     // Is it allowed to add nodes to the graph that do not have registered
86     // gradient function.
87     bool allow_non_differentiable_rewrites = true;
88 
89     // Tensorflow function execution semantics is slightly different from the
90     // main Tensorflow graph, and we need to make sure that we do not change it
91     // by running Grappler optimizer passes. One main difference is that
92     // functions do not prune ops with side-effects and dataset-output ops (see
93     // PruneFunctionBody in common_runtime/function.cc).
94     bool allow_pruning_stateful_and_dataset_ops = true;
95 
96     // If true Grappler will optimize the main graph, and also all functions in
97     // the graph function library (function can't be polymorphic, it can't have
98     // undefined type parameters in the function signature, or placeholder
99     // attributes in the function body).
100     bool optimize_function_library = true;
101   };
102 
103   const std::unordered_set<string>& devices() const;
104   // Adds a device to a set of available devices, only if it's a valid fully
105   // defined device name. Returns `Status::OK()` if successfully added a device,
106   // and an error otherwise.
107   Status AddDevice(const string& device);
108   // Adds all valid devices from the other Grappler item to the device set.
109   Status AddDevices(const GrapplerItem& other);
110   // Adds all valid devices from the nodes of the graph to the device set.
111   // Returns `Status::OK()` if all device annotations found in a graph are valid
112   // fully defined device names, and an error otherwise.
113   Status InferDevicesFromGraph();
114   // Clears a set of available devices.
115   void ClearDevices();
116 
117   const OptimizationOptions& optimization_options() const;
118   OptimizationOptions& optimization_options();
119 
120  private:
121   // TODO(ezhulenev) Make GrapplerItem a class and hide all public data members.
122   // TODO(ezhulenev): Migrate all unordered collections to absl.
123 
124   // A set of fully defined device names that can be used to place the nodes of
125   // the `graph`.
126   // Example of a fully defined name: "/job:work/replica:1/task:1/device:CPU:0"
127   std::unordered_set<string> devices_;
128 
129   OptimizationOptions optimization_options_;
130 };
131 
132 // Return the transitive fanin of a set of terminal nodes.
133 std::vector<const NodeDef*> ComputeTransitiveFanin(
134     const GraphDef& graph, const std::vector<string>& terminal_nodes);
135 
136 // Return the transitive fanin of a set of terminal nodes. Sets 'ill_formed' to
137 // true if one of the node is missing in the graph, or some node inputs don't
138 // exist.
139 std::vector<const NodeDef*> ComputeTransitiveFanin(
140     const GraphDef& graph, const std::vector<string>& terminal_nodes,
141     bool* ill_formed);
142 
143 }  // end namespace grappler
144 }  // end namespace tensorflow
145 
146 #endif  // TENSORFLOW_CORE_GRAPPLER_GRAPPLER_ITEM_H_
147