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 #include "tensorflow/python/grappler/model_analyzer.h"
17
18 #include <iomanip>
19 #include "tensorflow/core/framework/op.h"
20 #include "tensorflow/core/framework/tensor_shape.pb.h"
21 #include "tensorflow/core/grappler/costs/graph_properties.h"
22 #include "tensorflow/core/grappler/grappler_item.h"
23
24 namespace tensorflow {
25 namespace grappler {
26
ModelAnalyzer(const GrapplerItem & item)27 ModelAnalyzer::ModelAnalyzer(const GrapplerItem& item) : item_(item) {}
28
GenerateReport(bool debug,bool assume_valid_feeds,std::ostream & os)29 Status ModelAnalyzer::GenerateReport(bool debug, bool assume_valid_feeds,
30 std::ostream& os) {
31 GraphProperties properties(item_);
32 TF_RETURN_IF_ERROR(properties.InferStatically(assume_valid_feeds));
33
34 for (const auto& node : item_.MainOpsFanin()) {
35 PrintNodeInfo(node, properties, debug, os);
36 }
37 for (const auto& node : item_.EnqueueOpsFanin()) {
38 PrintNodeInfo(node, properties, debug, os);
39 }
40
41 return Status::OK();
42 }
43
PrintNodeInfo(const NodeDef * node,const GraphProperties & properties,bool debug,std::ostream & os) const44 void ModelAnalyzer::PrintNodeInfo(const NodeDef* node,
45 const GraphProperties& properties, bool debug,
46 std::ostream& os) const {
47 os << node->name() << " [" << node->op() << "]" << std::endl;
48 if (properties.HasOutputProperties(node->name())) {
49 const std::vector<OpInfo::TensorProperties>& props =
50 properties.GetOutputProperties(node->name());
51 for (int i = 0; i < props.size(); ++i) {
52 const OpInfo::TensorProperties& prop = props[i];
53 os << "\t"
54 << "output " << i << " (" << DataTypeString(prop.dtype())
55 << ") has shape ";
56 if (prop.shape().unknown_rank()) {
57 os << "?";
58 } else {
59 os << "[";
60 for (int i = 0; i < prop.shape().dim_size(); ++i) {
61 if (i > 0) {
62 os << ", ";
63 }
64 if (prop.shape().dim(i).size() >= 0) {
65 // Print the actual dimension.
66 os << prop.shape().dim(i).size();
67 } else if (prop.shape().dim(i).size() == -1) {
68 // We don't know anything about the dimension.
69 os << "?";
70 } else {
71 // Symbolic dimension.
72 os << "x" << -prop.shape().dim(i).size();
73 }
74 }
75 os << "]";
76 }
77 os << std::endl;
78 }
79 }
80
81 if (debug) {
82 const OpRegistrationData* op_reg_data;
83 Status status = OpRegistry::Global()->LookUp(node->op(), &op_reg_data);
84 if (!status.ok()) {
85 os << "\tCouldn't find op registration for " << node->op() << std::endl;
86 } else if (!op_reg_data->shape_inference_fn) {
87 os << "\tCouldn't find shape function for op " << node->op() << std::endl;
88 } else if (properties.HasInputProperties(node->name())) {
89 const std::vector<OpInfo::TensorProperties>& props =
90 properties.GetInputProperties(node->name());
91 for (int i = 0; i < props.size(); ++i) {
92 const OpInfo::TensorProperties& prop = props[i];
93 if (prop.has_value()) {
94 os << "\t"
95 << "input " << i << " (" << DataTypeString(prop.dtype())
96 << ") has known value" << std::endl;
97 }
98 }
99 }
100 }
101 }
102
103 } // end namespace grappler
104 } // end namespace tensorflow
105