Home | History | Annotate | Download | only in data
      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 #ifndef TENSORFLOW_CORE_KERNELS_DATA_DATASET_UTILS_H_
     16 #define TENSORFLOW_CORE_KERNELS_DATA_DATASET_UTILS_H_
     17 
     18 #include "tensorflow/core/framework/dataset.h"
     19 #include "tensorflow/core/framework/tensor.h"
     20 #include "tensorflow/core/kernels/data/captured_function.h"
     21 
     22 namespace tensorflow {
     23 namespace data {
     24 
     25 // This method is used to determine whether we can short-circuit the evaluation
     26 // of the user-defined function `func`. Short-circuting is possible if every
     27 // function output corresponds to one of its inputs (e.g. `f(x) = x`, `f(x,y) =
     28 // (y,x)`, or `f(x) = (x,x)`).
     29 //
     30 // If short-circuiting is possible, the method stores the mapping from output
     31 // indices to input indices in `indices`. Otherwise, `indices` will be empty.
     32 //
     33 // Returns non-ok status if analysis of the function fails.
     34 //
     35 // TODO(jsimsa): Extend this to support constants as well.
     36 Status ComputeShortCircuitIndices(OpKernelConstruction* ctx,
     37                                   const NameAttrList& func,
     38                                   std::vector<int>* indices);
     39 
     40 // Given a vector that maps output indices to input indices, return a vector
     41 // that identifies for which output indices can we move the input (assuming
     42 // output indices are processed left to right).
     43 std::vector<bool> ComputeMoveVector(const std::vector<int>& indices);
     44 
     45 Status MakeIteratorFromInputElement(
     46     IteratorContext* ctx, const std::vector<Tensor>& input_element,
     47     int64 thread_index, const InstantiatedCapturedFunction& inst_captured_func,
     48     StringPiece prefix, std::unique_ptr<IteratorBase>* out_iterator);
     49 
     50 // Returns Status::OK() if `expected` and `received` types match,
     51 // errors::InvalidArgument otherwise.
     52 Status VerifyTypesMatch(const DataTypeVector& expected,
     53                         const DataTypeVector& received);
     54 
     55 // Returns Status::OK() if `expected` and `received` shapes are compatible,
     56 // errors::InvalidArgument otherwise.
     57 Status VerifyShapesCompatible(const std::vector<PartialTensorShape>& expected,
     58                               const std::vector<PartialTensorShape>& received);
     59 
     60 // Helper class for reading data from a VariantTensorData object.
     61 class VariantTensorDataReader : public IteratorStateReader {
     62  public:
     63   explicit VariantTensorDataReader(const VariantTensorData* data);
     64 
     65   // Returns OK iff the initialization was successful.
     66   Status ReadScalar(StringPiece key, int64* val) override;
     67   Status ReadScalar(StringPiece key, string* val) override;
     68   Status ReadTensor(StringPiece key, Tensor* val) override;
     69   bool Contains(StringPiece key) override;
     70 
     71  private:
     72   template <typename T>
     73   Status ReadScalarInternal(StringPiece key, T* val);
     74   Status ReadTensorInternal(StringPiece key, Tensor* val);
     75 
     76   std::map<string, size_t> map_;
     77   const VariantTensorData* data_;  // Not owned.
     78 };
     79 
     80 // Helper class for writing data to a VariantTensorData object.
     81 class VariantTensorDataWriter : public IteratorStateWriter {
     82  public:
     83   // Does not take ownership of data.
     84   explicit VariantTensorDataWriter(VariantTensorData* data) : data_(data) {}
     85   Status WriteScalar(StringPiece key, const int64 val) override;
     86   Status WriteScalar(StringPiece key, const string& val) override;
     87   Status WriteTensor(StringPiece key, const Tensor& val) override;
     88 
     89   // Writes the metadata to `data_`.
     90   Status Flush();
     91 
     92  private:
     93   template <typename T>
     94   Status WriteScalarInternal(StringPiece key, const T& val);
     95   Status WriteTensorInternal(StringPiece key, const Tensor& val);
     96 
     97   VariantTensorData* data_;
     98   std::vector<string> keys_;
     99 };
    100 
    101 // Adds the functions in `to_add` to `base`. If a function with a matching
    102 // signature already exists in `base`, replaces it with the function from
    103 // `to_add`.
    104 Status AddToFunctionLibrary(FunctionLibraryDefinition* base,
    105                             const FunctionLibraryDefinition& to_add);
    106 Status AddToFunctionLibrary(FunctionLibraryDefinition* base,
    107                             const FunctionDefLibrary& to_add);
    108 
    109 // Creates a runner that runs functions with limited parallelism.
    110 std::function<void(std::function<void()>)> RunnerWithMaxParallelism(
    111     std::function<void(std::function<void()>)> runner, int max_parallelism);
    112 
    113 }  // namespace data
    114 }  // namespace tensorflow
    115 
    116 #endif  // TENSORFLOW_CORE_KERNELS_DATA_DATASET_UTILS_H_
    117