repeat_dataset_op.cc

/* Copyright 2017 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/kernels/data/repeat_dataset_op.h"

#include <cstdint>
#include <cstdlib>
#include <functional>
#include <memory>
#include <optional>
#include <string>
#include <utility>
#include <vector>

#include "absl/status/status.h"
#include "absl/status/statusor.h"
#include "absl/strings/str_cat.h"
#include "tensorflow/core/data/global_shuffle_utils.h"
#include "tensorflow/core/data/name_utils.h"
#include "tensorflow/core/framework/dataset.h"
#include "tensorflow/core/framework/partial_tensor_shape.h"
#include "tensorflow/core/framework/tensor.h"
#include "tsl/platform/errors.h"
#include "tsl/platform/statusor.h"
#include "tsl/platform/thread_annotations.h"

namespace tensorflow {
namespace data {

// See documentation in ../../ops/dataset_ops.cc for a high-level
// description of the following op.

/* static */ constexpr const char* const RepeatDatasetOp::kDatasetType;
/* static */ constexpr const char* const RepeatDatasetOp::kInputDataset;
/* static */ constexpr const char* const RepeatDatasetOp::kCount;
/* static */ constexpr const char* const RepeatDatasetOp::kOutputTypes;
/* static */ constexpr const char* const RepeatDatasetOp::kOutputShapes;

namespace {

constexpr char kForeverRepeat[] = "ForeverRepeat";
constexpr char kEmptyRepeat[] = "EmptyRepeat";
constexpr char kFiniteRepeat[] = "FiniteRepeat";
constexpr char kCurIteration[] = "i";
constexpr char kInputImplEmpty[] = "input_impl_empty";
constexpr char kUninitialized[] = "uninitialized";
constexpr int64_t kKnownRatio = 1;

std::string nested_prefix(const std::string& prefix, int64_t epoch) {
  return strings::StrCat(prefix, "[", epoch, "]");
}

// Returns whether `dataset` has an input dataset of the given type. This check
// includes transitive inputs. Returns true if any upstream dataset is a data
// service dataset. Returns false if no upstream dataset is a data service
// dataset, or it's unknown because `dataset` doesn't implement `InputDatasets`.
// TODO(b/269673112): Rewrite the dataset to add an `IsDynamic` attribute to
// signal if the repeated dataset is dynamic or not.
bool HasDataServiceInput(const DatasetBase* dataset) {
  DCHECK(dataset != nullptr);
  if (absl::StartsWith(dataset->type_string(), "DataServiceDataset")) {
    return true;
  }
  std::vector<const DatasetBase*> inputs;
  Status s = dataset->InputDatasets(&inputs);
  if (!s.ok()) {
    return false;
  }
  for (const DatasetBase* input : inputs) {
    if (HasDataServiceInput(input)) {
      return true;
    }
  }
  return false;
}

// Updates an input split provider with the appropriate cardinality count based
// on how many times it is repeated.
class RepeatedSplitProvider : public SplitProvider {
 public:
  explicit RepeatedSplitProvider(std::unique_ptr<SplitProvider> split_provider,
                                 int64_t count)
      : split_provider_(std::move(split_provider)), count_(count) {}

  // Updates the cardinality based on the times the input dataset is repeated.
  int64_t Cardinality() const override {
    if (split_provider_->Cardinality() == 0 || count_ == 0) {
      return 0;
    }
    // From tensorflow/python/data/ops/repeat_op.py, the repeat op uses -1 for
    // infinite repetitions.
    if (count_ < 0) {
      return kInfiniteCardinality;
    }
    if (split_provider_->Cardinality() < 0) {
      return split_provider_->Cardinality();
    }
    return split_provider_->Cardinality() * count_;
  }

  // The following are the same as the input split provider.
  absl::Status GetNext(Tensor* split, bool* end_of_splits) override {
    return split_provider_->GetNext(split, end_of_splits);
  }
  absl::Status Reset() override { return split_provider_->Reset(); }
  absl::Status Save(std::function<std::string(std::string)> full_name,
                    IteratorStateWriter* writer) override {
    return split_provider_->Save(full_name, writer);
  }
  absl::Status Restore(std::function<std::string(std::string)> full_name,
                       IteratorStateReader* reader) override {
    return split_provider_->Restore(full_name, reader);
  }
  void Cancel() override { split_provider_->Cancel(); }

 private:
  const std::unique_ptr<SplitProvider> split_provider_;
  const int64_t count_;
};
}  // namespace

class RepeatDatasetOp::Dataset : public DatasetBase {
 public:
  Dataset(OpKernelContext* ctx, int64_t count, const DatasetBase* input)
      : DatasetBase(DatasetContext(ctx)), count_(count), input_(input) {
    input_->Ref();
    if (input_ != nullptr && !input_->RandomIndexingCompatible().ok()) {
      random_indexing_compatible_ = input_->RandomIndexingCompatible();
    } else if (count <= 0) {
      random_indexing_compatible_ = absl::FailedPreconditionError(
          absl::StrCat("`repeat(", count,
                       ")` does not support random access of tf.data "
                       "datasets."));
    } else {
      random_indexing_compatible_ = absl::OkStatus();
    }
  }

  ~Dataset() override { input_->Unref(); }

  std::unique_ptr<IteratorBase> MakeIteratorInternal(
      const string& prefix) const override {
    if (count_ < 0) {
      return std::make_unique<ForeverIterator>(ForeverIterator::Params{
          this, name_utils::IteratorPrefix(kForeverRepeat, prefix)});
    } else if (count_ == 0) {
      return std::make_unique<EmptyIterator>(EmptyIterator::Params{
          this, name_utils::IteratorPrefix(kEmptyRepeat, prefix)});
    } else {
      return std::make_unique<FiniteIterator>(FiniteIterator::Params{
          this, name_utils::IteratorPrefix(kFiniteRepeat, prefix)});
    }
  }

  absl::Status MakeSplitProviders(std::vector<std::unique_ptr<SplitProvider>>*
                                      split_providers) const override {
    std::vector<std::unique_ptr<SplitProvider>> input_split_providers;
    TF_RETURN_IF_ERROR(input_->MakeSplitProviders(&input_split_providers));

    split_providers->clear();
    for (auto& split_provider : input_split_providers) {
      split_providers->push_back(std::make_unique<RepeatedSplitProvider>(
          std::move(split_provider), count_));
    }
    return absl::OkStatus();
  }

  const DataTypeVector& output_dtypes() const override {
    return input_->output_dtypes();
  }
  const std::vector<PartialTensorShape>& output_shapes() const override {
    return input_->output_shapes();
  }

  string DebugString() const override {
    return name_utils::DatasetDebugString(RepeatDatasetOp::kDatasetType);
  }

  int64_t CardinalityInternal(CardinalityOptions options) const override {
    int64_t n = input_->Cardinality(options);
    if (count_ < 0) {
      if (n == 0) {
        return 0;
      }
      return kInfiniteCardinality;
    }
    if (count_ == 0) {
      return 0;
    }
    if (n == kInfiniteCardinality || n == kUnknownCardinality) {
      return n;
    }
    return count_ * n;
  }

  Status InputDatasets(std::vector<const DatasetBase*>* inputs) const override {
    inputs->push_back(input_);
    return absl::OkStatus();
  }

  Status CheckExternalState() const override {
    return input_->CheckExternalState();
  }

  Status Get(OpKernelContext* ctx, int64 index,
             std::vector<Tensor>* out_tensors) const override {
    TF_RETURN_IF_ERROR(CheckRandomAccessCompatible(index));
    return input_->Get(ctx, index % input_->Cardinality(), out_tensors);
  }

  absl::Status RandomIndexingCompatible() const override {
    return random_indexing_compatible_;
  }

 protected:
  Status AsGraphDefInternal(SerializationContext* ctx,
                            DatasetGraphDefBuilder* b,
                            Node** output) const override {
    Node* input_graph_node = nullptr;
    TF_RETURN_IF_ERROR(b->AddInputDataset(ctx, input_, &input_graph_node));
    Node* count = nullptr;
    TF_RETURN_IF_ERROR(b->AddScalar(count_, &count));
    TF_RETURN_IF_ERROR(b->AddDataset(this, {input_graph_node, count}, output));
    return absl::OkStatus();
  }

 private:
  class EmptyIterator : public DatasetIterator<Dataset> {
   public:
    explicit EmptyIterator(const Params& params)
        : DatasetIterator<Dataset>(params) {}

    bool SymbolicCheckpointCompatible() const override { return true; }

    Status GetNextInternal(IteratorContext* ctx,
                           std::vector<Tensor>* out_tensors,
                           bool* end_of_sequence) override {
      *end_of_sequence = true;
      return absl::OkStatus();
    }

   protected:
    std::shared_ptr<model::Node> CreateNode(
        IteratorContext* ctx, model::Node::Args args) const override {
      return model::MakeKnownRatioNode(std::move(args),
                                       /*ratio=*/kKnownRatio);
    }

    Status SaveInternal(SerializationContext* ctx,
                        IteratorStateWriter* writer) override {
      return absl::OkStatus();
    }
    Status RestoreInternal(IteratorContext* ctx,
                           IteratorStateReader* reader) override {
      return absl::OkStatus();
    }
  };

  class FiniteIterator : public DatasetIterator<Dataset> {
   public:
    explicit FiniteIterator(const Params& params)
        : DatasetIterator<Dataset>(params), i_(0) {}

    bool SymbolicCheckpointCompatible() const override { return true; }

    Status Initialize(IteratorContext* ctx) override {
      mutex_lock l(mu_);
      return dataset()->input_->MakeIterator(
          ctx, this, nested_prefix(prefix(), i_), &input_impl_);
    }

    Status GetNextInternal(IteratorContext* ctx,
                           std::vector<Tensor>* out_tensors,
                           bool* end_of_sequence) override {
      mutex_lock l(mu_);  // TODO(mrry): Make locking less conservative.
      if (!input_impl_) {
        *end_of_sequence = true;
        return absl::OkStatus();
      }
      while (i_ < dataset()->count_) {
        IteratorContextWithIndexMapper ctx_with_index_mapper(ctx, this);
        TF_RETURN_IF_ERROR(input_impl_->GetNext(ctx_with_index_mapper.Get(),
                                                out_tensors, end_of_sequence));
        ctx_with_index_mapper.MergeCheckpoint();
        if (!*end_of_sequence) {
          return absl::OkStatus();
        }
        ctx->PurgeCheckpoint(nested_prefix(prefix(), i_));
        ++i_;
        input_impl_.reset();
        for (const auto& provider : ctx->split_providers()) {
          TF_RETURN_IF_ERROR(provider->Reset());
        }
        TF_RETURN_IF_ERROR(dataset()->input_->MakeIterator(
            ctx, this, nested_prefix(prefix(), i_), &input_impl_));
      }
      *end_of_sequence = true;
      input_impl_.reset();
      return absl::OkStatus();
    }

    IndexMapperFn GetIndexMapper(IndexMapperFn parent_index_mapper)
        const override TF_EXCLUSIVE_LOCKS_REQUIRED(mu_) {
      int64_t input_cardinality = dataset()->input_->Cardinality();
      int64_t repeat_count = i_;
      return [parent_index_mapper, input_cardinality,
              repeat_count](size_t element_position) -> absl::StatusOr<size_t> {
        if (element_position >= input_cardinality) {
          // The input element position is out-of-range. The caller is
          // responsible for handle this case (e.g.: returning end_of_sequence).
          return element_position;
        }

        // First, maps the input indices from
        // [0, input_range] to [0, input_range * repetitions].
        // Then, reduces the shuffled indices to [0, input_range] by taking the
        // mod. This way, the shuffling happens across repetitions.
        size_t repeated_element_position =
            repeat_count * input_cardinality + element_position;
        TF_ASSIGN_OR_RETURN(size_t shuffled_element_position,
                            parent_index_mapper(repeated_element_position));
        return shuffled_element_position % input_cardinality;
      };
    }

   protected:
    std::shared_ptr<model::Node> CreateNode(
        IteratorContext* ctx, model::Node::Args args) const override {
      return model::MakeKnownRatioNode(std::move(args),
                                       /*ratio=*/kKnownRatio);
    }

    Status SaveInternal(SerializationContext* ctx,
                        IteratorStateWriter* writer) override {
      mutex_lock l(mu_);
      TF_RETURN_IF_ERROR(writer->WriteScalar(prefix(), kCurIteration, i_));
      TF_RETURN_IF_ERROR(writer->WriteScalar(
          prefix(), kInputImplEmpty, static_cast<int64_t>(!input_impl_)));
      if (input_impl_) {
        TF_RETURN_IF_ERROR(SaveInput(ctx, writer, input_impl_));
      }
      return absl::OkStatus();
    }

    Status RestoreInternal(IteratorContext* ctx,
                           IteratorStateReader* reader) override {
      mutex_lock l(mu_);
      if (ctx->restored_element_count().has_value()) {
        CardinalityOptions options;
        options.set_compute_level(
            CardinalityOptions::CARDINALITY_COMPUTE_MODERATE);
        const int64_t input_cardinality =
            dataset()->input_->Cardinality(std::move(options));
        i_ = *ctx->restored_element_count() / input_cardinality;
        // For upstream iterators, the restored element count should be the
        // element count within the current repetition.
        IteratorContext::Params params(ctx);
        params.restored_element_count =
            *ctx->restored_element_count() % input_cardinality;
        params.index_mapper = GetIndexMapper(ctx->index_mapper());
        IteratorContext ctx_with_restored_element_count(params);
        return RestoreInput(&ctx_with_restored_element_count, reader,
                            input_impl_);
      }

      TF_RETURN_IF_ERROR(reader->ReadScalar(prefix(), kCurIteration, &i_));
      int64_t input_empty;
      TF_RETURN_IF_ERROR(
          reader->ReadScalar(prefix(), kInputImplEmpty, &input_empty));
      if (static_cast<bool>(!input_empty)) {
        TF_RETURN_IF_ERROR(dataset()->input_->MakeIterator(
            ctx, this, nested_prefix(prefix(), i_), &input_impl_));
        TF_RETURN_IF_ERROR(RestoreInput(ctx, reader, input_impl_));
      } else {
        input_impl_.reset();
      }
      return absl::OkStatus();
    }

   private:
    mutex mu_;
    int64_t i_ TF_GUARDED_BY(mu_);
    std::unique_ptr<IteratorBase> input_impl_ TF_GUARDED_BY(mu_);
  };

  class ForeverIterator : public DatasetIterator<Dataset> {
   public:
    explicit ForeverIterator(const Params& params)
        : DatasetIterator<Dataset>(params),
          has_data_service_input_(HasDataServiceInput(dataset())),
          input_impl_(nullptr),
          i_(0),
          first_call_(true) {}

    bool SymbolicCheckpointCompatible() const override { return true; }

    Status Initialize(IteratorContext* ctx) override {
      mutex_lock l(mu_);
      return dataset()->input_->MakeIterator(
          ctx, this, nested_prefix(prefix(), i_), &input_impl_);
    }

    Status GetNextInternal(IteratorContext* ctx,
                           std::vector<Tensor>* out_tensors,
                           bool* end_of_sequence) override {
      mutex_lock l(mu_);  // TODO(mrry): Make locking less conservative.
      do {
        if (!input_impl_) {
          TF_RETURN_IF_ERROR(dataset()->input_->MakeIterator(
              ctx, this, nested_prefix(prefix(), i_), &input_impl_));
        }
        TF_RETURN_IF_ERROR(
            input_impl_->GetNext(ctx, out_tensors, end_of_sequence));
        DCHECK(!*end_of_sequence || out_tensors->empty());
        if (first_call_ && *end_of_sequence && ctx->split_providers().empty()) {
          // If the first call to GetNext() fails because the end of sequence
          // has been reached, we return EOF unless it repeats a tf.data service
          // dataset, where the repeated elements are non-deterministic.
          // Otherwise, this iterator could loop infinitely.
          if (!has_data_service_input_) {
            input_impl_.reset();
            return absl::OkStatus();
          }
        }
        first_call_ = false;
        if (!*end_of_sequence) {
          return absl::OkStatus();
        }
        ctx->PurgeCheckpoint(nested_prefix(prefix(), i_));
        ++i_;
        for (const auto& provider : ctx->split_providers()) {
          TF_RETURN_IF_ERROR(provider->Reset());
        }
        input_impl_.reset();
        first_call_ = true;
      } while (true);
    }

   protected:
    std::shared_ptr<model::Node> CreateNode(
        IteratorContext* ctx, model::Node::Args args) const override {
      return model::MakeKnownRatioNode(std::move(args),
                                       /*ratio=*/kKnownRatio);
    }

    Status SaveInternal(SerializationContext* ctx,
                        IteratorStateWriter* writer) override {
      mutex_lock l(mu_);
      TF_RETURN_IF_ERROR(writer->WriteScalar(prefix(), kCurIteration, i_));
      TF_RETURN_IF_ERROR(writer->WriteScalar(
          prefix(), kInputImplEmpty, static_cast<int64_t>(!input_impl_)));
      if (input_impl_) {
        TF_RETURN_IF_ERROR(SaveInput(ctx, writer, input_impl_));
      }
      return absl::OkStatus();
    }

    Status RestoreInternal(IteratorContext* ctx,
                           IteratorStateReader* reader) override {
      mutex_lock l(mu_);
      TF_RETURN_IF_ERROR(reader->ReadScalar(prefix(), kCurIteration, &i_));
      int64_t input_empty;
      TF_RETURN_IF_ERROR(
          reader->ReadScalar(prefix(), kInputImplEmpty, &input_empty));
      if (static_cast<bool>(input_empty)) {
        input_impl_.reset();
        first_call_ = true;
      } else {
        TF_RETURN_IF_ERROR(dataset()->input_->MakeIterator(
            ctx, this, nested_prefix(prefix(), i_), &input_impl_));
        TF_RETURN_IF_ERROR(RestoreInput(ctx, reader, input_impl_));
        first_call_ = false;
      }
      return absl::OkStatus();
    }

   private:
    const bool has_data_service_input_;

    mutex mu_;
    std::unique_ptr<IteratorBase> input_impl_ TF_GUARDED_BY(mu_);
    int64_t i_ TF_GUARDED_BY(mu_);
    bool first_call_ TF_GUARDED_BY(mu_);
  };

  const int64_t count_;
  const DatasetBase* const input_;
  absl::Status random_indexing_compatible_;
};

RepeatDatasetOp::RepeatDatasetOp(OpKernelConstruction* ctx)
    : UnaryDatasetOpKernel(ctx) {}

void RepeatDatasetOp::MakeDataset(OpKernelContext* ctx, DatasetBase* input,
                                  DatasetBase** output) {
  // Create a new RepeatDatasetOp::Dataset, insert it in the step-local
  // container, and return it as the output.
  int64_t count;
  OP_REQUIRES_OK(ctx, ParseScalarArgument<int64_t>(ctx, kCount, &count));
  *output = new Dataset(ctx, count, input);
}

namespace {
REGISTER_KERNEL_BUILDER(Name("RepeatDataset").Device(DEVICE_CPU),
                        RepeatDatasetOp);
}  // namespace
}  // namespace data
}  // namespace tensorflow