-
Notifications
You must be signed in to change notification settings - Fork 341
/
graph_constructor.cc
2019 lines (1827 loc) · 74.8 KB
/
graph_constructor.cc
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
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
/* Copyright 2015 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/graph/graph_constructor.h"
#include <algorithm>
#include <set>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include <queue>
#include "absl/algorithm/container.h"
#include "absl/container/flat_hash_set.h"
#include "tensorflow/core/common_runtime/shape_refiner.h"
#include "tensorflow/core/framework/function.h"
#include "tensorflow/core/framework/function.pb.h"
#include "tensorflow/core/framework/graph.pb.h"
#include "tensorflow/core/framework/node_def.pb.h"
#include "tensorflow/core/framework/node_def_builder.h"
#include "tensorflow/core/framework/node_def_util.h"
#include "tensorflow/core/framework/tensor_shape.pb.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/framework/versions.h"
#include "tensorflow/core/framework/versions.pb.h"
#include "tensorflow/core/graph/algorithm.h"
#include "tensorflow/core/graph/graph.h"
#include "tensorflow/core/graph/node_builder.h"
#include "tensorflow/core/graph/tensor_id.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/gtl/flatmap.h"
#include "tensorflow/core/lib/gtl/flatset.h"
#include "tensorflow/core/lib/gtl/inlined_vector.h"
#include "tensorflow/core/lib/strings/scanner.h"
#include "tensorflow/core/lib/strings/str_util.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/macros.h"
#include "tensorflow/core/public/version.h"
namespace tensorflow {
namespace {
// We remove duplicate control inputs before adding edges to the Graph, so we
// can skip expensive duplicates check in 'AddControlEdge'.
static constexpr const bool kDoNotCheckDuplicates = true;
inline bool IsMerge(const NodeDef& node_def) {
return node_def.op() == "Merge" || node_def.op() == "RefMerge" ||
node_def.op() == "_XlaMerge";
}
inline bool IsNextIteration(const NodeDef& node_def) {
return node_def.op() == "NextIteration" ||
node_def.op() == "RefNextIteration";
}
bool IsValidNodeName(StringPiece s, bool allow_internal_ops) {
using ::tensorflow::strings::Scanner;
return Scanner(s)
.One(allow_internal_ops ? Scanner::LETTER_DIGIT_DOT_UNDERSCORE
: Scanner::LETTER_DIGIT_DOT)
.Any(Scanner::LETTER_DIGIT_DASH_DOT_SLASH_UNDERSCORE)
.Eos()
.GetResult();
}
class GraphConstructor {
public:
struct Options {
Options(const GraphConstructorOptions& in) // NOLINT(runtime/explicit)
: allow_internal_ops(in.allow_internal_ops),
expect_device_spec(in.expect_device_spec),
importing(false),
validate_colocation_constraints(false) {}
Options(const ImportGraphDefOptions& in) // NOLINT(runtime/explicit)
: allow_internal_ops(false),
expect_device_spec(false),
prefix(in.prefix.empty() || str_util::EndsWith(in.prefix, "/")
? in.prefix
: in.prefix + "/"),
uniquify_names(in.uniquify_names),
uniquify_prefix(in.uniquify_prefix),
input_map(in.input_map.begin(), in.input_map.end()),
skip_mapped_nodes(in.skip_mapped_nodes),
control_dependencies(in.control_dependencies),
return_tensors(in.return_tensors.begin(), in.return_tensors.end()),
return_nodes(in.return_nodes),
importing(true),
validate_colocation_constraints(in.validate_colocation_constraints),
validate_shape(in.validate_shape),
default_device(in.default_device) {}
bool allow_internal_ops;
bool expect_device_spec;
string prefix;
bool uniquify_names;
bool uniquify_prefix;
std::map<TensorId, TensorId> input_map;
bool skip_mapped_nodes;
std::vector<string> control_dependencies;
std::vector<TensorId> return_tensors;
std::vector<string> return_nodes;
// TODO(ashankar): This bool exists to separate out functionality required
// to make ImportGraphDef a close equivalent of Python's import_graph_def
// without affecting the behavior of ConvertGraphDefToGraph at the time
// ImportGraphDef was added.
//
// That said, the functionality here (shape and op validation) seems
// applicable to ConvertGraphDefToGraph as well, so make an attempt to
// remove this.
bool importing;
bool validate_colocation_constraints;
bool validate_shape = true;
string default_device;
};
typedef gtl::ArraySlice<const NodeDef*> NodeDefSlice;
// versions and library may be nullptr
static Status Construct(
const Options& opts, NodeDefSlice node_defs, const VersionDef* versions,
const FunctionDefLibrary* library, Graph* g, ShapeRefiner* refiner,
std::vector<std::pair<Node*, int>>* return_tensors,
std::vector<Node*>* return_nodes,
std::vector<SafeTensorId>* missing_unused_input_map_keys);
static Status Construct(
const Options& opts, GraphDef&& graph_def, Graph* g,
ShapeRefiner* refiner, std::vector<std::pair<Node*, int>>* return_tensors,
std::vector<Node*>* return_nodes,
std::vector<SafeTensorId>* missing_unused_input_map_keys);
protected:
GraphConstructor(const Options& opts, Graph* g, ShapeRefiner* refiner,
std::vector<std::pair<Node*, int>>* return_tensors,
std::vector<Node*>* return_nodes,
std::vector<SafeTensorId>* missing_unused_input_map_keys)
: opts_(opts),
g_(g),
original_versions_(g->versions()),
prefix_(opts.prefix),
refiner_(refiner),
return_tensors_(return_tensors),
return_nodes_(return_nodes),
missing_unused_input_map_keys_(missing_unused_input_map_keys) {}
virtual ~GraphConstructor() {}
Status TryImport() {
TF_RETURN_IF_ERROR(EnsureNoNameCollisions());
TF_RETURN_IF_ERROR(ValidateInputMapAndControlDependencies());
TF_RETURN_IF_ERROR(BuildNodeIndex());
TF_RETURN_IF_ERROR(InitFromEdges());
// NOTE: Convert() invokes `consume_node_def()` on each node in the input
// graph, so `get_node_def()` is no longer usable once it is called.
TF_RETURN_IF_ERROR(Convert());
TF_RETURN_IF_ERROR(AddBackEdges());
TF_RETURN_IF_ERROR(UpdateVersionDef());
TF_RETURN_IF_ERROR(PopulateReturnTensors());
TF_RETURN_IF_ERROR(PopulateReturnNodes());
TF_RETURN_IF_ERROR(PopulateMissingUnusedInputMapKeys());
UpdateUniquifiedColocationNames();
FixupSourceAndSinkEdges(g_);
return Status::OK();
}
private:
Status EnsureNoNameCollisions();
Status ValidateInputMapAndControlDependencies();
Status BuildNodeIndex();
Status InitFromEdges();
Status Convert();
Status AddBackEdges();
Status UpdateVersionDef();
Status PopulateReturnTensors();
Status PopulateReturnNodes();
Status PopulateMissingUnusedInputMapKeys();
void Undo();
// Prints cycles in the graph.
void PrintCycles();
// Performs DFS starting at `cur_node` and prints any cycles found.
void DFS(int cur_node, std::vector<int>* cur_branch,
std::vector<bool>* is_on_cur_branch,
absl::flat_hash_set<int>* unvisited);
Status IsNodeFullyMapped(const NodeDef& node_def, bool* is_node_mapped);
Status ValidateColocationConstraints(const NodeDef& node_def);
Status MakeNode(NodeDef&& node_def, Node** node);
Status MakeEdge(Node* src, int output_index, Node* dst, int input_index);
Status ValidateShape(Node* node);
Status ModifyNodeDefForImport(NodeDef* node_def);
// Modifies node_def's inputs according to opts_.input_map.
// input_already_exists is a pre-initialized vector of length
// node_def->input_size(). This function will mark inputs that are remapped to
// true.
void RemapNodeDefInputs(NodeDef* node_def,
std::vector<bool>* input_already_exists);
// input_already_exists is a pre-initialized vector of length
// node_def->input_size(). This function will add and mark control inputs as
// true.
void AddControlDependencies(NodeDef* node_def,
std::vector<bool>* input_already_exists);
void AddPrefixToNodeDef(const std::vector<bool>& input_already_exists,
NodeDef* node_def);
// Modifies `node_def` if its name isn't unique, or if any of its inputs'
// names have been uniquified. This must be called in topological order on all
// nodes.
void UniquifyNames(const std::vector<bool>& input_already_exists,
NodeDef* node_def);
// Updates any constructed nodes' colocation group names if the name has been
// updated by UniquifyNames. This is called after all the nodes have been
// constructed so all the names have been uniquified if necessary.
void UpdateUniquifiedColocationNames();
// Returns true if `name` already exists in `g_` (either as a node name or
// prefix).
bool NameExistsInGraph(StringPiece name);
// Returns true if `name` already exists in the GraphDef being imported
// (either as a node name or prefix).
bool NameExistsInGraphDef(StringPiece name);
// Returns a unique version of `original_name`, or `original_name` if it's
// already unique in the graph.
string FindUniqueName(StringPiece original_name);
// Decrement pending count for users of `processed` and add the ones that now
// have all of their pending inputs satisfied to `ready_`.
void UpdatePendingCountAndReady(int processed, bool is_next_iteration);
// Subclasses override the following virtual methods to provide efficient
// access to the original protocol buffer-based graph.
// Returns the number of nodes in the graph.
virtual size_t node_def_count() const = 0;
// Returns the i^th node in the graph. Must not be called after
// consume_node_def(i).
virtual const NodeDef& get_node_def(int i) const = 0;
// Destructively reads the i^th node in the graph, avoiding a copy if
// possible. After calling this method, the result of get_node_def(i) is
// undefined.
virtual NodeDef consume_node_def(int i) = 0;
// Returns the version information for the graph, or nullptr if none is
// available.
virtual const VersionDef* versions() const = 0;
// Returns the function information for the graph, or nullptr if none is
// available.
virtual const FunctionDefLibrary* library() const = 0;
// From constructor
const Options opts_;
Graph* g_;
const VersionDef original_versions_;
// A copy of opts_.prefix, possibly uniquified.
string prefix_;
ShapeRefiner* refiner_;
// May be null. Not owned.
std::vector<std::pair<Node*, int>>* return_tensors_;
// May be null. Not owned.
std::vector<Node*>* return_nodes_;
// May be null. Not owned.
std::vector<SafeTensorId>* missing_unused_input_map_keys_;
// Intermediate datastructure used to populate
// `missing_unused_input_map_keys_`.
std::set<TensorId> used_input_map_keys_;
// Intermediate datastructure used to track the destinations of back edges.
absl::flat_hash_set<int> merge_node_indices_;
// Mapping from node name to the index within node_defs_.
struct NodeInfo {
explicit NodeInfo(int i) : gdef_index(i), node(nullptr) {}
// Containers require that we have a default constructor.
NodeInfo() : NodeInfo(-1) {}
int gdef_index;
Node* node; // nullptr until the NodeDef is converted to a Node.
};
gtl::FlatMap<StringPiece, NodeInfo, StringPieceHasher> gdef_nodes_;
// Storage for StringPiece keys in gdef_nodes_. Typically, the StringPiece key
// will refer to the string stored in `NodeDef::name()`. This intern table is
// only used when the original NodeDef's name is changed.
std::vector<string> string_intern_table_;
// Prefixes already used in the GraphDef being imported.
gtl::FlatSet<StringPiece, StringPieceHasher> gdef_prefixes_;
// Mapping from node name to the existing node in g_.
gtl::FlatMap<StringPiece, Node*, StringPieceHasher> existing_nodes_;
// Prefixes already used in the graph.
gtl::FlatSet<StringPiece, StringPieceHasher> existing_prefixes_;
// Imported node names that have been uniquified. The key is the original
// name, the value is the new unique name.
gtl::FlatMap<string, string> uniquified_names_;
// Index of NodeDefs in node_defs_ with all inputs already converted. We use a
// (sorted) set so nodes are created in the order defined in the GraphDef.
std::set<int> ready_;
// Mapping between index within node_defs_ and the number of inputs that
// still need to be converted.
std::vector<int> pending_count_;
// Mapping between index within node_defs_ and the index within node_defs_ of
// all nodes it outputs to.
std::vector<gtl::InlinedVector<int, 4>> outputs_;
// Used in the conversion from node_defs_ to g_ to represent the ith input
// of a node.
struct InputInfo {
explicit InputInfo(const string& node_name, Node* n, int i)
: name(node_name), node(n), index(i) {}
// Use string instead of StringPiece so we don't have to manage lifetime
string name;
Node* node;
int index;
static bool IsControlInput(const InputInfo& input) {
return input.index == Graph::kControlSlot;
}
static int CompareName(const InputInfo& lhs, const InputInfo& rhs) {
return lhs.name < rhs.name;
}
static bool IsSameName(const InputInfo& lhs, const InputInfo& rhs) {
return lhs.name == rhs.name;
}
};
// Used in the conversion from node_defs_ to g_ to represent an edge from
// the node named 'name' to node 'n'.
struct EdgeInfo {
explicit EdgeInfo(const string& name, int i1, Node* n, int i2)
: src_name(name), src_index(i1), dst_node(n), dst_index(i2) {}
// Use string instead of StringPiece so we don't have to manage lifetime
string src_name;
int src_index;
Node* dst_node;
int dst_index;
};
std::vector<EdgeInfo> back_edges_;
TF_DISALLOW_COPY_AND_ASSIGN(GraphConstructor);
};
// Implementation of GraphConstructor that does not take ownership of the
// input NodeDef messages and thus copies the nodes into the constructed Graph*.
//
// NOTE(mrry): Whenever possible, use NodeDefMovingGraphConstructor, which
// avoids copying each NodeDef into the constructed Graph*.
class NodeDefCopyingGraphConstructor : public GraphConstructor {
public:
NodeDefCopyingGraphConstructor(
const Options& opts, NodeDefSlice node_defs, const VersionDef* versions,
const FunctionDefLibrary* library, Graph* g, ShapeRefiner* refiner,
std::vector<std::pair<Node*, int>>* return_tensors,
std::vector<Node*>* return_nodes,
std::vector<SafeTensorId>* missing_unused_input_map_keys)
: GraphConstructor(opts, g, refiner, return_tensors, return_nodes,
missing_unused_input_map_keys),
node_defs_(node_defs),
versions_(versions),
library_(library) {}
private:
size_t node_def_count() const override { return node_defs_.size(); }
const NodeDef& get_node_def(int i) const override { return *node_defs_[i]; }
NodeDef consume_node_def(int i) override { return *node_defs_[i]; }
const VersionDef* versions() const override { return versions_; }
const FunctionDefLibrary* library() const override { return library_; }
const NodeDefSlice node_defs_;
const VersionDef* const versions_;
const FunctionDefLibrary* const library_;
};
// Implementation of GraphConstructor that takes ownership of the input
// GraphDef, and can perform destructive reads.
class NodeDefMovingGraphConstructor : public GraphConstructor {
public:
NodeDefMovingGraphConstructor(
const Options& opts, GraphDef&& graph_def, Graph* g,
ShapeRefiner* refiner, std::vector<std::pair<Node*, int>>* return_tensors,
std::vector<Node*>* return_nodes,
std::vector<SafeTensorId>* missing_unused_input_map_keys)
: GraphConstructor(opts, g, refiner, return_tensors, return_nodes,
missing_unused_input_map_keys),
graph_def_(std::move(graph_def)),
is_consumed_(graph_def_.node_size(), false) {}
private:
size_t node_def_count() const override { return graph_def_.node().size(); }
const NodeDef& get_node_def(int i) const override {
CHECK(!is_consumed_[i])
<< "NodeDef " << i << " accessed after it was consumed.";
return graph_def_.node(i);
}
NodeDef consume_node_def(int i) override {
CHECK(!is_consumed_[i]) << "NodeDef " << i << " consumed twice.";
is_consumed_[i] = true;
return std::move(*graph_def_.mutable_node(i));
}
const VersionDef* versions() const override { return &graph_def_.versions(); }
const FunctionDefLibrary* library() const override {
return &graph_def_.library();
}
GraphDef graph_def_;
std::vector<bool> is_consumed_;
};
/* static */ Status GraphConstructor::Construct(
const Options& opts, NodeDefSlice node_defs, const VersionDef* versions,
const FunctionDefLibrary* library, Graph* g, ShapeRefiner* refiner,
std::vector<std::pair<Node*, int>>* return_tensors,
std::vector<Node*>* return_nodes,
std::vector<SafeTensorId>* missing_unused_input_map_keys) {
if (versions) {
TF_RETURN_IF_ERROR(CheckVersions(*versions, TF_GRAPH_DEF_VERSION,
TF_GRAPH_DEF_VERSION_MIN_PRODUCER,
"GraphDef", "graph"));
}
NodeDefCopyingGraphConstructor c(opts, node_defs, versions, library, g,
refiner, return_tensors, return_nodes,
missing_unused_input_map_keys);
const Status s = c.TryImport();
if (!s.ok()) c.Undo();
return s;
}
/* static */ Status GraphConstructor::Construct(
const Options& opts, GraphDef&& graph_def, Graph* g, ShapeRefiner* refiner,
std::vector<std::pair<Node*, int>>* return_tensors,
std::vector<Node*>* return_nodes,
std::vector<SafeTensorId>* missing_unused_input_map_keys) {
TF_RETURN_IF_ERROR(CheckVersions(graph_def.versions(), TF_GRAPH_DEF_VERSION,
TF_GRAPH_DEF_VERSION_MIN_PRODUCER,
"GraphDef", "graph"));
NodeDefMovingGraphConstructor c(opts, std::move(graph_def), g, refiner,
return_tensors, return_nodes,
missing_unused_input_map_keys);
const Status s = c.TryImport();
if (!s.ok()) c.Undo();
return s;
}
void GraphConstructor::UpdatePendingCountAndReady(int processed,
bool is_next_iteration) {
for (size_t i = 0; i < outputs_[processed].size(); ++i) {
const int output = outputs_[processed][i];
// We didn't consider NextIteration->Merge edges when computing
// pending_counts_ so we should not have to consider it here either.
bool is_next_iteration_to_merge_edge =
is_next_iteration && merge_node_indices_.count(output) == 1;
if (!is_next_iteration_to_merge_edge) {
int* current_pending_count = &pending_count_[output];
CHECK_GT(*current_pending_count, 0);
(*current_pending_count)--;
if (*current_pending_count == 0) {
ready_.insert(output);
}
}
}
}
// This could be expensive but we don't expect to call it often, if at all (only
// if there are multiple nodes in g_ with the same name)
bool NodeNameInValues(const std::map<TensorId, TensorId>& input_map,
const StringPiece& node_name) {
for (auto iter = input_map.begin(); iter != input_map.end(); ++iter) {
if (iter->second.first == node_name) return true;
}
return false;
}
bool NodeNameInValues(const std::vector<string>& control_dependencies,
const StringPiece& node_name) {
return std::find(control_dependencies.begin(), control_dependencies.end(),
node_name) != control_dependencies.end();
}
// Adds any prefixes of `node_name` (not including the full name itself) to
// `prefixes`.
void AddPrefixes(StringPiece node_name,
gtl::FlatSet<StringPiece, StringPieceHasher>* prefixes) {
size_t idx = -1;
while ((idx = node_name.find('/', idx + 1)) != StringPiece::npos) {
prefixes->insert(node_name.substr(0, idx));
}
}
Status GraphConstructor::EnsureNoNameCollisions() {
existing_nodes_.reserve(g_->num_nodes());
// Populate existing_nodes_ and existing_prefixes_.
for (Node* n : g_->nodes()) {
bool already_exists = !existing_nodes_.insert({n->name(), n}).second;
if (already_exists) {
if (NodeNameInValues(opts_.input_map, n->name())) {
return errors::InvalidArgument(
"cannot resolve input_map because multiple nodes exist with name '",
n->name(), "'");
}
if (NodeNameInValues(opts_.control_dependencies, n->name())) {
return errors::InvalidArgument(
"cannot resolve control_dependencies because multiple nodes exist "
"with name '",
n->name(), "'");
}
}
AddPrefixes(n->name(), &existing_prefixes_);
}
if (prefix_.empty() && opts_.importing && !opts_.uniquify_names) {
for (size_t i = 0; i < node_def_count(); ++i) {
const string& name = get_node_def(i).name();
if (NameExistsInGraph(name)) {
return errors::InvalidArgument("Node name '", name,
"' already exists in the Graph");
}
}
} else if (!prefix_.empty()) {
StringPiece prefix_no_slash(prefix_);
prefix_no_slash.remove_suffix(1);
if (!IsValidNodeName(prefix_no_slash, false)) {
return errors::InvalidArgument("Imported node name prefix '", prefix_,
"' would lead to invalid node names");
}
if (NameExistsInGraph(prefix_no_slash) && opts_.uniquify_prefix) {
prefix_ = strings::StrCat(FindUniqueName(prefix_no_slash), "/");
}
}
return Status::OK();
}
Status GraphConstructor::ValidateInputMapAndControlDependencies() {
for (const auto& mapping : opts_.input_map) {
TensorId src = mapping.first;
TensorId dst = mapping.second;
if (existing_nodes_.count(dst.first) == 0) {
return errors::InvalidArgument(
"node '", dst.first, "' in input_map does not exist in graph ",
"(input_map entry: ", src.ToString(), "->", dst.ToString(), ")");
}
if ((src.second == Graph::kControlSlot) !=
(dst.second == Graph::kControlSlot)) {
return errors::InvalidArgument("input_map entry ", src.ToString(), "->",
dst.ToString(), " between ",
"control edge and non-control edge");
}
}
for (const string& node : opts_.control_dependencies) {
if (existing_nodes_.count(node) == 0) {
return errors::InvalidArgument(
"node '", node,
"' in control_dependencies does not exist in "
"graph");
}
}
return Status::OK();
}
Status GraphConstructor::BuildNodeIndex() {
// Validate the node names and add them to gdef_nodes_ and gdef_prefixes_.
for (int n = 0; n < node_def_count(); ++n) {
const NodeDef& node_def = get_node_def(n);
if (!IsValidNodeName(node_def.name(), opts_.allow_internal_ops)) {
return errors::InvalidArgument(
"Node '", node_def.name(),
"': Node name contains invalid characters");
}
if (!gdef_nodes_
.insert(std::make_pair(StringPiece(node_def.name()), NodeInfo(n)))
.second) {
return errors::InvalidArgument("Node '", node_def.name(),
"' is not unique");
}
// Validate the operation's type.
if (node_def.op().empty()) {
return errors::InvalidArgument("Node '", node_def.name(),
"' does not specify an operation");
}
if (opts_.expect_device_spec && node_def.device().empty()) {
return errors::InvalidArgument("Node '", node_def.name(),
"' is missing a device specification");
}
if (IsMerge(node_def)) {
merge_node_indices_.insert(n);
}
// Validate control edges at end
bool in_control_dependence = false;
for (int i = 0; i < node_def.input_size(); ++i) {
StringPiece input_name = node_def.input(i);
if (!input_name.empty() && absl::StartsWith(input_name, "^")) {
in_control_dependence = true;
} else if (in_control_dependence) {
return errors::InvalidArgument(
"Node '", node_def.name(),
"': Control dependencies must come after regular dependencies");
}
}
// Update gdef_prefixes_.
AddPrefixes(node_def.name(), &gdef_prefixes_);
}
return Status::OK();
}
Status GraphConstructor::InitFromEdges() {
const int num_nodes = node_def_count();
pending_count_.reserve(num_nodes);
outputs_.resize(num_nodes);
gtl::FlatSet<string> next_iteration_nodes;
for (int n = 0; n < node_def_count(); ++n) {
const NodeDef& node_def = get_node_def(n);
if (IsNextIteration(node_def)) {
next_iteration_nodes.insert(node_def.name());
}
}
// Parse the inputs for each node.
for (int n = 0; n < num_nodes; ++n) {
const NodeDef& node_def = get_node_def(n);
int pending_count = node_def.input_size();
if (IsMerge(node_def)) {
// Cycles in the graph are only allowed for while loops. A while loop is
// identified by an edge from a NextIteration node to a Merge node. For
// such Merge nodes, only wait for one non-control input before
// considering the node ready to process in Convert().
int32 num_control_edges = 0;
bool has_loop_back_edge = false;
for (int i = 0; i < node_def.input_size(); ++i) {
StringPiece input_name(node_def.input(i));
if (absl::StartsWith(input_name, "^")) {
num_control_edges++;
} else {
TensorId id(ParseTensorName(input_name));
if (next_iteration_nodes.find(string(id.first)) !=
next_iteration_nodes.end()) {
has_loop_back_edge = true;
}
}
}
if (has_loop_back_edge) {
pending_count = num_control_edges + 1;
}
}
for (int i = 0; i < node_def.input_size(); ++i) {
StringPiece input_name = node_def.input(i);
TensorId id(ParseTensorName(input_name));
if (opts_.input_map.count(id) == 0) {
// If an input is not mapped, then the input should appear in the graph
// being imported.
auto iter = gdef_nodes_.find(id.first);
if (iter == gdef_nodes_.end()) {
return errors::InvalidArgument("Node '", node_def.name(),
"': Unknown input node '",
node_def.input(i), "'");
}
outputs_[iter->second.gdef_index].push_back(n);
} else {
// This input is mapped to an existing edge. Therefore this input is
// as good as being already processed.
--pending_count;
DCHECK_GE(pending_count, 0);
}
}
if (pending_count == 0) {
ready_.insert(n);
}
pending_count_.push_back(pending_count);
}
return Status::OK();
}
Status GraphConstructor::ValidateColocationConstraints(
const NodeDef& node_def) {
if (!opts_.validate_colocation_constraints || !opts_.importing)
return Status::OK();
const auto iter = node_def.attr().find(kColocationAttrName);
if (iter == node_def.attr().end()) return Status::OK();
for (const string& c : iter->second.list().s()) {
StringPiece s(c);
if (absl::ConsumePrefix(&s, kColocationGroupPrefix) &&
gdef_nodes_.find(s) == gdef_nodes_.end()) {
return errors::InvalidArgument(
"Node '", node_def.name(),
"' expects to be colocated with unknown node '", s, "'");
}
}
return Status::OK();
}
Status GraphConstructor::MakeNode(NodeDef&& node_def, Node** node) {
// Add the node to the graph.
Status status;
*node = g_->AddNode(std::move(node_def), &status);
if (!status.ok()) return status;
if (opts_.expect_device_spec) {
(*node)->set_assigned_device_name((*node)->def().device());
}
return Status::OK();
}
Status GraphConstructor::ValidateShape(Node* node) {
if (!opts_.importing || !opts_.validate_shape) return Status::OK();
TF_RETURN_IF_ERROR(refiner_->AddNode(node));
// For nodes with the _output_shapes attribute, override the shape.
std::vector<const TensorShapeProto*> shape_attrs;
const char* kAttrName = "_output_shapes";
if (!TryGetNodeAttr(node->attrs(), kAttrName, &shape_attrs)) {
// No _output_shapes attribute, the AddNode call above was sufficient.
return Status::OK();
}
auto* ic = refiner_->GetContext(node);
DCHECK(ic != nullptr)
<< "ShapeRefiner::AddNode() should have created the InferenceContext";
if (shape_attrs.size() < node->num_outputs()) {
return errors::InvalidArgument(
"Node '", node->name(), "' has ", node->num_outputs(),
" outputs but the ", kAttrName, " attribute specifies shapes for ",
shape_attrs.size(), " outputs");
}
// NOTE(skyewm): we don't raise an error here because some users depend on
// this behavior, even though it's unsafe.
// TODO(b/74619486): raise an error.
if (shape_attrs.size() > node->num_outputs()) {
LOG(WARNING) << "Node '" << node->name() << "' has " << node->num_outputs()
<< " outputs but the " << kAttrName
<< " attribute specifies shapes for " << shape_attrs.size()
<< " outputs. Output shapes may be inaccurate.";
}
for (int i = 0; i < node->num_outputs(); ++i) {
const TensorShapeProto& p = *shape_attrs[i];
shape_inference::ShapeHandle h;
Status s = ic->MakeShapeFromShapeProto(p, &h);
if (!s.ok()) {
return errors::InvalidArgument("Node '", node->name(), " has an invalid ",
kAttrName, " attribute (shape #", i,
" error:'", s.error_message(), "'");
}
s = refiner_->SetShape(node, i, h);
if (!s.ok()) {
// If the output shape is incompatible with what is inferred
// by the graph for a very specific whitelist of ops, then we
// ignore this output shape. This can happen if there is a
// bug in the shape function for some operation, and the
// serialized graph def has the incorrect shape set when
// running on a newer binary with the fixed shape function.
// This is an escape hatch that allows us to correct shape
// functions that are not critical to correct execution but
// would cause graphs to fail if imported after correcting.
const string& op = node->type_string();
const std::vector<string> whitelist = {
// To be removed after 2017/03/08.
"RandomShuffleQueue",
"PaddingFIFOQueue",
"FIFOQueue",
"PriorityQueue",
"QueueSize",
"Stack",
"Barrier",
"BarrierReadySize",
"BarrierIncompleteSize",
"HashTable",
"MutableHashTable",
"MutableHashTableOfTensors",
"Mutex",
"CuckooTable",
"IndexTable",
"WholeFileReader",
"TextLineReader",
"FixedLengthRecordReader",
"TFRecordReader",
"IdentityReader",
"RefSwitch",
"RefEnter",
"RefNextIteration",
"RefMerge",
"RefIdentity",
"LMDBReader",
// To be removed after 2017/04/24.
"ConditionalAccumulator",
"SparseConditionalAccumulator",
"Table",
};
if (std::find(whitelist.begin(), whitelist.end(), op) ==
whitelist.end()) {
return errors::InvalidArgument(
"Node '", node->name(), "' has an ", kAttrName,
" attribute inconsistent with the GraphDef for output #", i, ": ",
s.error_message());
}
}
}
node->ClearAttr(kAttrName);
return Status::OK();
}
Status GraphConstructor::ModifyNodeDefForImport(NodeDef* node_def) {
const OpDef* op_def;
TF_RETURN_IF_ERROR(g_->op_registry()->LookUpOpDef(node_def->op(), &op_def));
AddDefaultsToNodeDef(*op_def, node_def);
TF_RETURN_IF_ERROR(ValidateNodeDef(*node_def, *op_def));
if (versions()) {
TF_RETURN_IF_ERROR(CheckOpDeprecation(*op_def, versions()->producer()));
}
return Status::OK();
}
void RemoveInputs(const std::vector<int>& inputs_to_remove, NodeDef* node_def,
std::vector<bool>* input_already_exists) {
// Remove 'inputs_to_remove' from 'node_def'
NodeDef copy;
copy.mutable_input()->Reserve(node_def->input_size() -
inputs_to_remove.size());
for (int i = 0, j = 0; i < node_def->input_size(); ++i) {
if (j < inputs_to_remove.size() && i == inputs_to_remove[j]) {
++j;
} else {
copy.add_input()->swap(*node_def->mutable_input(i));
}
}
node_def->mutable_input()->Swap(copy.mutable_input());
// Remove 'inputs_to_remove' from 'input_already_exists'
for (int idx : inputs_to_remove) {
input_already_exists->erase(input_already_exists->begin() + idx);
}
DCHECK_EQ(input_already_exists->size(), node_def->input_size());
}
void GraphConstructor::RemapNodeDefInputs(
NodeDef* node_def, std::vector<bool>* input_already_exists) {
DCHECK_EQ(input_already_exists->size(), node_def->input_size());
std::set<TensorId> control_inputs;
std::vector<int> inputs_to_remove;
for (int i = 0; i < node_def->input_size(); ++i) {
auto iter = opts_.input_map.find(ParseTensorName(node_def->input(i)));
if (iter == opts_.input_map.end()) continue;
used_input_map_keys_.insert(iter->first);
TensorId new_input = iter->second;
if (new_input.second == Graph::kControlSlot) {
// Check if we've already remapped a different input to new_input, and if
// so remove this input.
if (control_inputs.count(new_input) > 0) {
inputs_to_remove.push_back(i);
continue;
}
control_inputs.insert(new_input);
}
node_def->set_input(i, new_input.ToString());
(*input_already_exists)[i] = true;
}
if (!inputs_to_remove.empty()) {
RemoveInputs(inputs_to_remove, node_def, input_already_exists);
}
}
void GraphConstructor::AddControlDependencies(
NodeDef* node_def, std::vector<bool>* input_already_exists) {
// To avoid adding redundant control dependencies to every imported node, skip
// nodes that will inherit the dependencies from another imported node.
bool inherits_deps = false;
for (int i = 0; i < node_def->input_size(); ++i) {
// Assume we won't inherit dependencies from remapped inputs that already
// exist in the graph. Even if we're wrong, we'll only add redundant
// dependencies.
if ((*input_already_exists)[i]) continue;
// If this input is a backedge, assume we won't inherit the dependencies.
// TODO(skyewm): we have many redundant ParseTensorName calls. It could be
// worth optimizing these.
TensorId id(ParseTensorName(node_def->input(i)));
auto iter = gdef_nodes_.find(id.first);
DCHECK(iter != gdef_nodes_.end()) << id.first;
if (iter->second.node == nullptr) {
// Input hasn't been created yet, indicating it's a backedge.
continue;
}
inherits_deps = true;
}
if (inherits_deps) return;
// node_def either has no inputs or all remapped inputs, add the control
// dependencies
for (const string& control_dep : opts_.control_dependencies) {
string input = TensorId(control_dep, Graph::kControlSlot).ToString();
bool found = false;
for (int i = node_def->input_size() - 1; i >= 0; --i) {
const string& node_input = node_def->input(i);
if (node_input[0] != '^') {
// Control inputs are at the end. Break when we reach the non-control
// inputs.
break;
}
if (node_input == input) {
// Control dependency already exists
found = true;
break;
}
}
if (found) {
continue;
}
node_def->add_input(input);
input_already_exists->push_back(true);
}
}
void GraphConstructor::AddPrefixToNodeDef(
const std::vector<bool>& input_already_exists, NodeDef* node_def) {
if (prefix_.empty()) return;
node_def->set_name(strings::StrCat(prefix_, node_def->name()));
// Update names of input nodes
for (int i = 0; i < node_def->input_size(); ++i) {
// Skip remapped inputs (which already exist in g_ and are not being
// imported).
if (input_already_exists[i]) continue;
StringPiece input(node_def->input(i));
if (absl::ConsumePrefix(&input, "^")) {
node_def->set_input(i, strings::StrCat("^", prefix_, input));
} else {
node_def->set_input(i, strings::StrCat(prefix_, input));
}
}
// Update names of colocation groups
if (node_def->attr().find(kColocationAttrName) != node_def->attr().end()) {
auto* list =
node_def->mutable_attr()->at(kColocationAttrName).mutable_list();
for (int i = 0; i < list->s_size(); ++i) {
StringPiece v(list->s(i));
if (absl::ConsumePrefix(&v, kColocationGroupPrefix)) {
list->set_s(i, strings::StrCat(kColocationGroupPrefix, prefix_, v));
}
}
}
}
void GraphConstructor::UniquifyNames(
const std::vector<bool>& input_already_exists, NodeDef* node_def) {
if (NameExistsInGraph(node_def->name())) {
string old_name = node_def->name();
node_def->set_name(FindUniqueName(node_def->name()));
uniquified_names_[old_name] = node_def->name();
// Note that we don't have to update gdef_nodes_ or gdef_prefixes_ with
// `name` because we guarantee the original NodeDef names are unique,
// meaning we won't generate this name again.
}
for (int i = 0; i < node_def->input_size(); ++i) {
// Skip remapped inputs (which already exist in g_ and are not being
// imported).
if (input_already_exists[i]) continue;
TensorId id = ParseTensorName(node_def->input(i));
// We require that UniquifyNames() is called on all NodeDefs in topological
// order. This guarantees that node_def's inputs will already be uniquified
// if necessary.
auto iter = uniquified_names_.find(string(id.first));
if (iter == uniquified_names_.end()) continue;
id.first = iter->second;
node_def->set_input(i, id.ToString());
}
}
void GraphConstructor::UpdateUniquifiedColocationNames() {
for (const auto& pair : gdef_nodes_) {
Node* node = pair.second.node;
if (node == nullptr) continue;
std::vector<string> coloc_values;
if (!TryGetNodeAttr(node->attrs(), kColocationAttrName, &coloc_values))
continue;
bool updated = false;
for (size_t i = 0; i < coloc_values.size(); ++i) {