1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
|
/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
#include "tensorflow/core/grappler/optimizers/data/map_and_batch_fusion.h"
#include "tensorflow/core/framework/attr_value.pb.h"
#include "tensorflow/core/framework/node_def.pb.h"
#include "tensorflow/core/grappler/clusters/cluster.h"
#include "tensorflow/core/grappler/grappler_item.h"
#include "tensorflow/core/grappler/mutable_graph_view.h"
#include "tensorflow/core/grappler/op_types.h"
#include "tensorflow/core/grappler/optimizers/custom_graph_optimizer_registry.h"
#include "tensorflow/core/grappler/optimizers/data/graph_utils.h"
#include "tensorflow/core/grappler/utils.h"
#include "tensorflow/core/platform/protobuf.h"
namespace tensorflow {
namespace grappler {
namespace {
constexpr char kFusedOpName[] = "MapAndBatchDatasetV2";
NodeDef MakeMapAndBatchNode(const NodeDef& map_node, const NodeDef& batch_node,
MutableGraphView* graph) {
NodeDef new_node;
new_node.set_op(kFusedOpName);
graph_utils::SetUniqueGraphNodeName(kFusedOpName, graph->GetGraph(),
&new_node);
// Set the `input` input argument.
new_node.add_input(map_node.input(0));
// Set the `other_arguments` input arguments.
int num_other_args;
if (map_node.op() == "ParallelMapDataset") {
num_other_args = map_node.input_size() - 2;
} else {
num_other_args = map_node.input_size() - 1;
}
for (int i = 0; i < num_other_args; i++) {
new_node.add_input(map_node.input(i + 1));
}
// Set the `batch_size` input argument.
new_node.add_input(batch_node.input(1));
// Set the `num_parallel_calls` input argument.
if (map_node.op() == "ParallelMapDataset") {
// The type of the `num_parallel_calls` argument in ParallelMapDataset
// and MapAndBatchDataset is different (int32 and int64 respectively)
// so we cannot reuse the same Const node and thus create a new one.
NodeDef* v = graph->GetNode(map_node.input(map_node.input_size() - 1));
NodeDef* tmp = graph_utils::AddScalarConstNode<int64>(
v->attr().at("value").tensor().int_val(0), graph);
new_node.add_input(tmp->name());
} else {
NodeDef* tmp = graph_utils::AddScalarConstNode<int64>(1, graph);
new_node.add_input(tmp->name());
}
// Set the `drop_remainder` input argument.
if (batch_node.op() == "BatchDatasetV2") {
new_node.add_input(batch_node.input(2));
} else {
NodeDef* tmp = graph_utils::AddScalarConstNode<bool>(false, graph);
new_node.add_input(tmp->name());
}
// Set `f` and `Targuments` attributes.
for (auto key : {"f", "Targuments"}) {
graph_utils::CopyAttribute(key, map_node, &new_node);
}
// Set `output_types` and `output_shapes` attributes.
for (auto key : {"output_shapes", "output_types"}) {
graph_utils::CopyAttribute(key, batch_node, &new_node);
}
return new_node;
}
} // namespace
Status MapAndBatchFusion::Optimize(Cluster* cluster, const GrapplerItem& item,
GraphDef* output) {
*output = item.graph;
MutableGraphView graph(output);
std::set<string> nodes_to_delete;
for (const NodeDef& node : item.graph.node()) {
if (node.op() != "BatchDataset" && node.op() != "BatchDatasetV2") {
continue;
}
// Use a more descriptive variable name now that we know the node type.
const NodeDef& batch_node = node;
NodeDef* node2 = graph_utils::GetInputNode(batch_node, graph);
if (node2->op() != "MapDataset" && node2->op() != "ParallelMapDataset") {
continue;
}
// Use a more descriptive variable name now that we know the node type.
NodeDef* map_node = node2;
auto* new_node =
graph.AddNode(MakeMapAndBatchNode(*map_node, batch_node, &graph));
graph.ReplaceInput(batch_node, *new_node);
// Mark the `Map` and `Batch` nodes for removal.
nodes_to_delete.insert(map_node->name());
nodes_to_delete.insert(batch_node.name());
}
graph.DeleteNodes(nodes_to_delete);
return Status::OK();
}
void MapAndBatchFusion::Feedback(Cluster* cluster, const GrapplerItem& item,
const GraphDef& optimize_output,
double result) {
// no-op
}
REGISTER_GRAPH_OPTIMIZER_AS(MapAndBatchFusion, "map_and_batch_fusion");
} // end namespace grappler
} // end namespace tensorflow
|
