• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2016 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_UTIL_EXAMPLE_PROTO_HELPER_H_
17 #define TENSORFLOW_CORE_UTIL_EXAMPLE_PROTO_HELPER_H_
18 
19 #include <string>
20 #include <unordered_set>
21 #include <vector>
22 
23 #include "tensorflow/core/example/example.pb.h"
24 #include "tensorflow/core/example/feature.pb.h"
25 #include "tensorflow/core/framework/allocator.h"
26 #include "tensorflow/core/framework/graph.pb.h"
27 #include "tensorflow/core/framework/partial_tensor_shape.h"
28 #include "tensorflow/core/framework/tensor.h"
29 #include "tensorflow/core/framework/types.h"
30 #include "tensorflow/core/lib/core/errors.h"
31 #include "tensorflow/core/platform/types.h"
32 #include "tensorflow/core/util/sparse/sparse_tensor.h"
33 
34 // This is a set of helper methods that will make it possible to share
35 // tensorflow::Example proto Tensor conversion code inside the ExampleParserOp
36 // OpKernel as well as in external code.
37 namespace tensorflow {
38 
39 // "Dense" feature configuration.
40 struct FixedLenFeature {
41   string key;
42   DataType dtype;
43   TensorShape shape;
44   Tensor default_value;
45   string values_output_tensor_name;
46 };
47 
48 // "Sparse" feature configuration.
49 struct VarLenFeature {
50   string key;
51   DataType dtype;
52   string values_output_tensor_name;
53   string indices_output_tensor_name;
54   string shapes_output_tensor_name;
55 };
56 
57 // Given a single tensorflow::Example, with an optional example name
58 // at a particular index within a batch, and dense and sparse feature
59 // configurations from fixed_len_features, var_len_features, this method
60 // updates the dense value tensor and the sparse values temporary vector
61 // of tensors. The indexing of the output vectors correspond 1:1 to the
62 // indexing of the feature configuration vectors.
63 //
64 // The fixed_len_features and var_len_features maps are assume to be
65 // have disjoint key fields from the Feature map in the tensorflow.Example
66 // proto.
67 //
68 // For each sparse feature, the sparse values temporary vector holds a
69 // tensor for each Example. Each tensor is either empty or filled, depending
70 // on if the sparse feature value is set for the Example. This
71 // temporary structure is needed because we need to know the total number
72 // of filled elements in the batch to get the proper final sparse tensor
73 // shapes allocated.  After the entire batch is processed,
74 // GetSparseTensorShape can be used to calculate the final shapes and
75 // CopyIntoSparseTensor can be used to copy from the temporary vector
76 // into the final allocated tensors.
77 Status SingleExampleProtoToTensors(
78     const Example& example, const string& name, const int batch_index,
79     const std::vector<FixedLenFeature>& fixed_len_features,
80     const std::vector<VarLenFeature>& var_len_features,
81     std::vector<Tensor*>* dense_values,
82     std::vector<std::vector<Tensor>>* sparse_values_temporary_vector);
83 
84 // The shape of the indices and values tensors associated with a SparseTensor
85 // are dependent on the contents of the batch.
86 struct VarLenFeatureBatchShapes {
87   TensorShape indices_shape;
88   TensorShape values_shape;
89   int max_num_features;
90 };
91 
92 // Get the shape of the sparse values and indices tensors for the batch,
93 // given how many of the tensors in the temporary sparse values vector
94 // are actually filled.
95 Status GetSparseTensorShapes(const VarLenFeature& var_len_feature,
96                              const std::vector<Tensor>& sparse_values_tmp,
97                              const int batch_size,
98                              VarLenFeatureBatchShapes* output_shapes);
99 
100 // A method to convert a batch of tensorflow::Example protos into output
101 // tensors. This method is useful if there already is a batch of deserialized
102 // Example protos in memory (such as a serving use-case) and we do not wish
103 // to incur an extraneous serialize/deserialize.  It is intended
104 // as an outside of OpKernel compatible replacement for the functionality of
105 // ExampleParserOp. In a serving setting, this method could be used to produce
106 // a feed_dict of Tensors that could bypass the ExampleParserOp.
107 //
108 // Note that unlike SingleExampleProtoToTensors, output tensors are
109 // allocated using a provided Allocator within this method.
110 Status BatchExampleProtoToTensors(
111     const std::vector<const Example*>& examples,
112     const std::vector<string>& names,
113     const std::vector<FixedLenFeature>& fixed_len_features,
114     const std::vector<VarLenFeature>& var_len_features, Allocator* allocator,
115     std::vector<Tensor>* output_dense_values_tensor,
116     std::vector<Tensor>* output_sparse_indices_tensor,
117     std::vector<Tensor>* output_sparse_values_tensor,
118     std::vector<Tensor>* output_sparse_shapes_tensor);
119 
120 // Check that the given dtype is one that is compatible with
121 // tensorflow::Example protocol buffer feature values.
122 Status CheckValidType(const DataType& dtype);
123 
124 // Check that the provided Feature proto message's oneof value
125 // matches that of the provided dtype.
126 Status CheckTypesMatch(const Feature& feature, const DataType& dtype,
127                        bool* match);
128 
129 // For a single Example, copy a dense feature value into an output
130 // dense value tensor Out at the provided out_index offset.
131 Status FeatureDenseCopy(const std::size_t out_index, const string& name,
132                         const string& key, const DataType& dtype,
133                         const TensorShape& shape, const Feature& feature,
134                         Tensor* out);
135 
136 // Copy the value a provided Tensor into an output dense_value tensor Out
137 // at the provided out_index offset.
138 void RowDenseCopy(const std::size_t& out_index, const DataType& dtype,
139                   const Tensor& in, Tensor* out);
140 
141 // For a single Example, and given sparse feature return a temporary output
142 // Tensor suitable for being collected in the temporary sparse value vector.
143 Tensor FeatureSparseCopy(const std::size_t batch, const string& key,
144                          const DataType& dtype, const Feature& feature);
145 
146 // Copy a temporary Tensor into the final sparse indices and values
147 // tensor at a given batch index and element offset. This method
148 // assumes that the indices/values Tensors have been properly allocated
149 // for the batch.
150 int64 CopyIntoSparseTensor(const Tensor& in, const int batch,
151                            const int64 offset, Tensor* indices, Tensor* values);
152 
153 // Check that each dense_shape has known rank and inner dimensions; and
154 // update variable_length (whether the outer dimension is None) and
155 // elements_per_stride for each denes_shape.
156 Status GetDenseShapes(const std::vector<PartialTensorShape>& dense_shapes,
157                       std::vector<bool>* variable_length,
158                       std::vector<std::size_t>* elements_per_stride);
159 
160 // Parses the attributes passed to ParseExample.
161 // REQUIRES: Init must be called after construction.
162 struct ParseExampleAttrs {
163  public:
164   template <typename ContextType>
165   Status Init(ContextType* ctx, int op_version = 1) {
166     TF_RETURN_IF_ERROR(ctx->GetAttr("sparse_types", &sparse_types));
167     TF_RETURN_IF_ERROR(ctx->GetAttr("Tdense", &dense_types));
168     TF_RETURN_IF_ERROR(ctx->GetAttr("dense_shapes", &dense_shapes));
169     TF_RETURN_IF_ERROR(
170         GetDenseShapes(dense_shapes, &variable_length, &elements_per_stride));
171     switch (op_version) {
172       case 1:
173         TF_RETURN_IF_ERROR(ctx->GetAttr("Nsparse", &num_sparse));
174         TF_RETURN_IF_ERROR(ctx->GetAttr("Ndense", &num_dense));
175         break;
176       case 2:
177         TF_RETURN_IF_ERROR(
178             ctx->GetAttr("ragged_value_types", &ragged_value_types));
179         TF_RETURN_IF_ERROR(ctx->GetAttr("num_sparse", &num_sparse));
180         TF_RETURN_IF_ERROR(
181             ctx->GetAttr("ragged_split_types", &ragged_split_types));
182         break;
183       default:
184         return errors::InvalidArgument("Unexpected op_version", op_version);
185     }
186     return FinishInit(op_version);
187   }
188 
189   int64 num_sparse;
190   int64 num_dense;
191   int64 num_ragged;
192   std::vector<DataType> sparse_types;
193   std::vector<DataType> dense_types;
194   std::vector<DataType> ragged_value_types;
195   std::vector<DataType> ragged_split_types;
196   std::vector<PartialTensorShape> dense_shapes;
197   std::vector<bool> variable_length;
198   std::vector<std::size_t> elements_per_stride;
199 
200  private:
201   Status FinishInit(int op_version);  // for context-independent parts of Init.
202 };
203 
204 // Parses the attributes passed to ParseSingleExample.
205 // REQUIRES: Init must be called after construction.
206 struct ParseSingleExampleAttrs {
207  public:
208   template <typename ContextType>
InitParseSingleExampleAttrs209   Status Init(ContextType* ctx) {
210     TF_RETURN_IF_ERROR(ctx->GetAttr("sparse_keys", &sparse_keys));
211     TF_RETURN_IF_ERROR(ctx->GetAttr("sparse_types", &sparse_types));
212     TF_RETURN_IF_ERROR(ctx->GetAttr("dense_keys", &dense_keys));
213     TF_RETURN_IF_ERROR(ctx->GetAttr("Tdense", &dense_types));
214     TF_RETURN_IF_ERROR(ctx->GetAttr("dense_shapes", &dense_shapes));
215 
216     int num_sparse;
217     TF_RETURN_IF_ERROR(ctx->GetAttr("num_sparse", &num_sparse));
218     if (num_sparse != sparse_keys.size() || num_sparse != sparse_types.size()) {
219       return errors::InvalidArgument(
220           "num_sparse (", num_sparse, ") must match the size of sparse_keys (",
221           sparse_keys.size(), ") and sparse_types (", sparse_types.size(), ")");
222     }
223 
224     TF_RETURN_IF_ERROR(
225         GetDenseShapes(dense_shapes, &variable_length, &elements_per_stride));
226     return FinishInit();
227   }
228 
229   std::vector<tstring> sparse_keys;
230   std::vector<DataType> sparse_types;
231   std::vector<tstring> dense_keys;
232   std::vector<DataType> dense_types;
233   std::vector<PartialTensorShape> dense_shapes;
234   std::vector<bool> variable_length;
235   std::vector<std::size_t> elements_per_stride;
236 
237  private:
238   Status FinishInit();  // for context-independent parts of Init.
239 };
240 
241 // Parses the attributes passed to ParseSequenceExample.
242 // REQUIRES: Init must be called after construction.
243 struct ParseSequenceExampleAttrs {
244  public:
245   template <typename ContextType>
246   Status Init(ContextType* ctx, int op_version = 1) {
247     switch (op_version) {
248       case 1: {
249         std::vector<string> missing_empty_vector;
250         TF_RETURN_IF_ERROR(ctx->GetAttr(
251             "feature_list_dense_missing_assumed_empty", &missing_empty_vector));
252         for (const string& feature : missing_empty_vector) {
253           feature_list_dense_missing_assumed_empty.insert(feature);
254         }
255       }
256         TF_RETURN_IF_ERROR(
257             ctx->GetAttr("context_sparse_keys", &context_sparse_keys));
258         TF_RETURN_IF_ERROR(
259             ctx->GetAttr("context_dense_keys", &context_dense_keys));
260         TF_RETURN_IF_ERROR(ctx->GetAttr("feature_list_sparse_keys",
261                                         &feature_list_sparse_keys));
262         TF_RETURN_IF_ERROR(
263             ctx->GetAttr("feature_list_dense_keys", &feature_list_dense_keys));
264         TF_RETURN_IF_ERROR(ctx->GetAttr("Ncontext_dense", &num_context_dense));
265         break;
266       case 2:
267         TF_RETURN_IF_ERROR(ctx->GetAttr("context_ragged_value_types",
268                                         &context_ragged_value_types));
269         TF_RETURN_IF_ERROR(ctx->GetAttr("context_ragged_split_types",
270                                         &context_ragged_split_types));
271         TF_RETURN_IF_ERROR(ctx->GetAttr("feature_list_ragged_value_types",
272                                         &feature_list_ragged_value_types));
273         TF_RETURN_IF_ERROR(ctx->GetAttr("feature_list_ragged_split_types",
274                                         &feature_list_ragged_split_types));
275         break;
276       default:
277         return errors::InvalidArgument("Unexpected op_version", op_version);
278     }
279     TF_RETURN_IF_ERROR(
280         ctx->GetAttr("context_sparse_types", &context_sparse_types));
281     TF_RETURN_IF_ERROR(
282         ctx->GetAttr("Nfeature_list_dense", &num_feature_list_dense));
283     TF_RETURN_IF_ERROR(ctx->GetAttr("Ncontext_sparse", &num_context_sparse));
284     TF_RETURN_IF_ERROR(ctx->GetAttr("Tcontext_dense", &context_dense_types));
285     TF_RETURN_IF_ERROR(
286         ctx->GetAttr("feature_list_sparse_types", &feature_list_sparse_types));
287     TF_RETURN_IF_ERROR(
288         ctx->GetAttr("feature_list_dense_types", &feature_list_dense_types));
289     TF_RETURN_IF_ERROR(
290         ctx->GetAttr("Nfeature_list_sparse", &num_feature_list_sparse));
291     TF_RETURN_IF_ERROR(
292         ctx->GetAttr("context_dense_shapes", &context_dense_shapes));
293     TF_RETURN_IF_ERROR(
294         ctx->GetAttr("feature_list_dense_shapes", &feature_list_dense_shapes));
295     return FinishInit(op_version);
296   }
297 
298   std::unordered_set<string> feature_list_dense_missing_assumed_empty;
299   int64 num_context_sparse;
300   int64 num_context_dense;
301   int64 num_context_ragged;
302   int64 num_feature_list_sparse;
303   int64 num_feature_list_dense;
304   int64 num_feature_list_ragged;
305   std::vector<tstring> context_sparse_keys;
306   std::vector<tstring> context_dense_keys;
307   std::vector<tstring> feature_list_sparse_keys;
308   std::vector<tstring> feature_list_dense_keys;
309   std::vector<DataType> context_sparse_types;
310   std::vector<DataType> context_dense_types;
311   std::vector<TensorShape> context_dense_shapes;
312   std::vector<DataType> feature_list_sparse_types;
313   std::vector<DataType> feature_list_dense_types;
314   std::vector<TensorShape> feature_list_dense_shapes;
315   std::vector<DataType> context_ragged_value_types;
316   std::vector<DataType> context_ragged_split_types;
317   std::vector<DataType> feature_list_ragged_value_types;
318   std::vector<DataType> feature_list_ragged_split_types;
319 
320  private:
321   Status FinishInit(int op_version);  // for context-independent parts of Init.
322 };
323 
324 // Parses the attributes passed to ParseSingleSequenceExample.
325 // REQUIRES: Init must be called after construction.
326 struct ParseSingleSequenceExampleAttrs {
327  public:
328   template <typename ContextType>
InitParseSingleSequenceExampleAttrs329   Status Init(ContextType* ctx) {
330     TF_RETURN_IF_ERROR(
331         ctx->GetAttr("context_sparse_types", &context_sparse_types));
332     TF_RETURN_IF_ERROR(ctx->GetAttr("Ncontext_dense", &num_context_dense));
333     TF_RETURN_IF_ERROR(
334         ctx->GetAttr("Nfeature_list_dense", &num_feature_list_dense));
335     TF_RETURN_IF_ERROR(ctx->GetAttr("Ncontext_sparse", &num_context_sparse));
336     TF_RETURN_IF_ERROR(ctx->GetAttr("Tcontext_dense", &context_dense_types));
337     TF_RETURN_IF_ERROR(
338         ctx->GetAttr("feature_list_sparse_types", &feature_list_sparse_types));
339     TF_RETURN_IF_ERROR(
340         ctx->GetAttr("feature_list_dense_types", &feature_list_dense_types));
341     TF_RETURN_IF_ERROR(
342         ctx->GetAttr("Nfeature_list_sparse", &num_feature_list_sparse));
343     TF_RETURN_IF_ERROR(
344         ctx->GetAttr("context_dense_shapes", &context_dense_shapes));
345     TF_RETURN_IF_ERROR(
346         ctx->GetAttr("feature_list_dense_shapes", &feature_list_dense_shapes));
347     return FinishInit();
348   }
349 
350   int64 num_context_sparse;
351   int64 num_context_dense;
352   int64 num_feature_list_sparse;
353   int64 num_feature_list_dense;
354   std::vector<DataType> context_sparse_types;
355   std::vector<DataType> context_dense_types;
356   std::vector<TensorShape> context_dense_shapes;
357   std::vector<DataType> feature_list_sparse_types;
358   std::vector<DataType> feature_list_dense_types;
359   std::vector<TensorShape> feature_list_dense_shapes;
360 
361  private:
362   Status FinishInit();  // for context-independent parts of Init.
363 };
364 
365 }  // namespace tensorflow
366 
367 #endif  // TENSORFLOW_CORE_UTIL_EXAMPLE_PROTO_HELPER_H_
368