count_ops.cc

/* Copyright 2020 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 <algorithm>
#include <limits>

#define EIGEN_USE_THREADS

#include "absl/container/flat_hash_map.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/op_requires.h"
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/framework/tensor_types.h"
#include "tensorflow/core/platform/errors.h"
#include "tensorflow/core/platform/types.h"

namespace tensorflow {

// Don't allocate too large `BatchedMap<T>` objects
static int kMaxBatches = std::numeric_limits<int>::max();

template <class T>
using BatchedMap = std::vector<absl::flat_hash_map<int64_t, T>>;

namespace {
// TODO(momernick): Extend this function to work with outputs of rank > 2.
template <class T>
Status OutputSparse(const BatchedMap<T>& per_batch_counts, int64_t num_values,
                    bool is_1d, OpKernelContext* context) {
  int total_values = 0;
  int num_batches = per_batch_counts.size();
  for (const auto& per_batch_count : per_batch_counts) {
    total_values += per_batch_count.size();
  }

  Tensor* indices;
  int inner_dim = is_1d ? 1 : 2;
  TF_RETURN_IF_ERROR(context->allocate_output(
      0, TensorShape({total_values, inner_dim}), &indices));

  Tensor* values;
  TF_RETURN_IF_ERROR(
      context->allocate_output(1, TensorShape({total_values}), &values));

  auto output_indices = indices->matrix<int64_t>();
  auto output_values = values->flat<T>();
  int64_t value_loc = 0;
  for (int b = 0; b < num_batches; ++b) {
    const auto& per_batch_count = per_batch_counts[b];
    std::vector<std::pair<int64_t, T>> pairs(per_batch_count.begin(),
                                             per_batch_count.end());
    std::sort(pairs.begin(), pairs.end());
    for (const auto& x : pairs) {
      if (is_1d) {
        output_indices(value_loc, 0) = x.first;
      } else {
        output_indices(value_loc, 0) = b;
        output_indices(value_loc, 1) = x.first;
      }
      output_values(value_loc) = x.second;
      ++value_loc;
    }
  }
  Tensor* dense_shape;
  if (is_1d) {
    TF_RETURN_IF_ERROR(
        context->allocate_output(2, TensorShape({1}), &dense_shape));
    dense_shape->flat<int64_t>().data()[0] = num_values;
  } else {
    TF_RETURN_IF_ERROR(
        context->allocate_output(2, TensorShape({2}), &dense_shape));
    dense_shape->flat<int64_t>().data()[0] = num_batches;
    dense_shape->flat<int64_t>().data()[1] = num_values;
  }

  return OkStatus();
}

int64_t GetOutputSize(int64_t max_seen, int64_t max_length,
                      int64_t min_length) {
  return max_length >= 0 ? max_length : std::max((max_seen + 1), min_length);
}

}  // namespace

template <class T, class W>
class DenseCount : public OpKernel {
 public:
  explicit DenseCount(OpKernelConstruction* context) : OpKernel(context) {
    OP_REQUIRES_OK(context, context->GetAttr("minlength", &minlength_));
    OP_REQUIRES_OK(context, context->GetAttr("maxlength", &maxlength_));
    OP_REQUIRES_OK(context, context->GetAttr("binary_output", &binary_output_));
  }

  void Compute(OpKernelContext* context) override {
    const Tensor& data = context->input(0);
    const Tensor& weights = context->input(1);
    bool use_weights = weights.NumElements() > 0;

    OP_REQUIRES(context,
                TensorShapeUtils::IsVector(data.shape()) ||
                    TensorShapeUtils::IsMatrix(data.shape()),
                errors::InvalidArgument(
                    "Input must be a 1 or 2-dimensional tensor. Got: ",
                    data.shape().DebugString()));

    // Ensure all values are non-negative.
    const auto data_values = data.flat<T>();
    Eigen::TensorFixedSize<bool, Eigen::Sizes<>, Eigen::RowMajor> nonnegative;
    nonnegative.device(context->eigen_cpu_device()) =
        (data_values >= static_cast<T>(0)).all();
    OP_REQUIRES(
        context, nonnegative(),
        errors::InvalidArgument("Input values must all be non-negative"));

    if (use_weights) {
      OP_REQUIRES(
          context, weights.shape() == data.shape(),
          errors::InvalidArgument(
              "Weights and data must have the same shape. Weight shape: ",
              weights.shape().DebugString(),
              "; data shape: ", data.shape().DebugString()));
    }

    bool is_1d = TensorShapeUtils::IsVector(data.shape());
    int negative_valued_axis = -1;
    int num_batch_dimensions = (data.shape().dims() + negative_valued_axis);

    int num_batch_elements = 1;
    for (int i = 0; i < num_batch_dimensions; ++i) {
      OP_REQUIRES(context, data.shape().dim_size(i) != 0,
                  errors::InvalidArgument(
                      "Invalid input: Shapes dimension cannot be 0."));
      num_batch_elements *= data.shape().dim_size(i);
    }
    int num_value_elements = data.shape().num_elements() / num_batch_elements;
    auto per_batch_counts = BatchedMap<W>(num_batch_elements);

    T max_value = 0;

    const auto weight_values = weights.flat<W>();
    int i = 0;
    for (int b = 0; b < num_batch_elements; ++b) {
      for (int v = 0; v < num_value_elements; ++v) {
        const auto& value = data_values(i);
        if (maxlength_ < 0 || value < maxlength_) {
          if (binary_output_) {
            per_batch_counts[b][value] = 1;
          } else if (use_weights) {
            per_batch_counts[b][value] += weight_values(i);
          } else {
            per_batch_counts[b][value]++;
          }
          if (value > max_value) {
            max_value = value;
          }
        }
        ++i;
      }
    }

    int64_t num_output_values =
        GetOutputSize(max_value, maxlength_, minlength_);
    OP_REQUIRES_OK(context, OutputSparse<W>(per_batch_counts, num_output_values,
                                            is_1d, context));
  }

 private:
  int64_t maxlength_;
  int64_t minlength_;
  bool binary_output_;
};

template <class T, class W>
class SparseCount : public OpKernel {
 public:
  explicit SparseCount(OpKernelConstruction* context) : OpKernel(context) {
    OP_REQUIRES_OK(context, context->GetAttr("minlength", &minlength_));
    OP_REQUIRES_OK(context, context->GetAttr("maxlength", &maxlength_));
    OP_REQUIRES_OK(context, context->GetAttr("binary_output", &binary_output_));
  }

  void Compute(OpKernelContext* context) override {
    const Tensor& indices = context->input(0);
    const Tensor& values = context->input(1);
    const Tensor& shape = context->input(2);
    const Tensor& weights = context->input(3);
    bool use_weights = weights.NumElements() > 0;

    OP_REQUIRES(context, TensorShapeUtils::IsMatrix(indices.shape()),
                errors::InvalidArgument(
                    "Input indices must be a 2-dimensional tensor. Got: ",
                    indices.shape().DebugString()));
    OP_REQUIRES(context, TensorShapeUtils::IsVector(values.shape()),
                errors::InvalidArgument("Input values must be a vector. Got: ",
                                        values.shape().DebugString()));
    OP_REQUIRES(context, TensorShapeUtils::IsVector(shape.shape()),
                errors::InvalidArgument("Input shape must be a vector. Got: ",
                                        shape.shape().DebugString()));
    OP_REQUIRES(context,
                values.shape().dim_size(0) == indices.shape().dim_size(0),
                errors::InvalidArgument(
                    "Number of values must match first dimension of indices.",
                    "Got ", values.shape().dim_size(0),
                    " values, indices shape: ", indices.shape().DebugString()));
    OP_REQUIRES(
        context, shape.shape().dim_size(0) == indices.shape().dim_size(1),
        errors::InvalidArgument(
            "Number of dimensions must match second dimension of indices.",
            "Got ", shape.shape().dim_size(0),
            " dimensions, indices shape: ", indices.shape().DebugString()));
    OP_REQUIRES(context, shape.NumElements() > 0,
                errors::InvalidArgument(
                    "The shape argument requires at least one element."));
    // Validate indices: each index must be valid for the corresponding
    // dimension. This could be possibly done better.
    const auto indices_values = indices.matrix<int64_t>();
    const auto shape_vector = shape.vec<int64_t>();
    int num_values = values.NumElements();  // same as first dim of indices
    int rank = indices.shape().dim_size(1);
    for (int i = 0; i < num_values; ++i) {
      for (int j = 0; j < rank; ++j) {
        OP_REQUIRES(
            context,
            indices_values(i, j) >= 0 && indices_values(i, j) < shape_vector(j),
            errors::InvalidArgument(
                "Invalid index value at ", i, ": dimension ", j, " has value ",
                indices_values(i, j), " which is not in [0, ", shape_vector(j),
                ") (as given by dense shape ", shape.DebugString()));
      }
    }

    // Ensure all values are non-negative.
    const auto values_values = values.flat<T>();
    Eigen::TensorFixedSize<bool, Eigen::Sizes<>, Eigen::RowMajor> nonnegative;
    nonnegative.device(context->eigen_cpu_device()) =
        (values_values >= static_cast<T>(0)).all();
    OP_REQUIRES(
        context, nonnegative(),
        errors::InvalidArgument("Input values must all be non-negative"));

    if (use_weights) {
      OP_REQUIRES(
          context, weights.shape() == values.shape(),
          errors::InvalidArgument(
              "Weights and values must have the same shape. Weight shape: ",
              weights.shape().DebugString(),
              "; values shape: ", values.shape().DebugString()));
    }

    bool is_1d = shape.NumElements() == 1;
    int num_batches = is_1d ? 1 : shape_vector(0);
    OP_REQUIRES(
        context, 0 < num_batches && num_batches < kMaxBatches,
        errors::InvalidArgument("Cannot allocate ", num_batches,
                                " batches, is the dense shape too wide?"));

    const auto weight_values = weights.flat<W>();

    auto per_batch_counts = BatchedMap<W>(num_batches);

    T max_value = 0;

    for (int idx = 0; idx < num_values; ++idx) {
      int batch = is_1d ? 0 : indices_values(idx, 0);
      if (batch >= num_batches) {
        OP_REQUIRES(context, batch < num_batches,
                    errors::InvalidArgument(
                        "Indices value along the first dimension must be ",
                        "lower than the first index of the shape.", "Got ",
                        batch, " as batch and ", num_batches,
                        " as the first dimension of the shape."));
      }
      const auto& value = values_values(idx);
      if (maxlength_ < 0 || value < maxlength_) {
        if (binary_output_) {
          per_batch_counts[batch][value] = 1;
        } else if (use_weights) {
          per_batch_counts[batch][value] += weight_values(idx);
        } else {
          per_batch_counts[batch][value]++;
        }
        if (value > max_value) {
          max_value = value;
        }
      }
    }

    int64_t num_output_values =
        GetOutputSize(max_value, maxlength_, minlength_);
    OP_REQUIRES_OK(context, OutputSparse<W>(per_batch_counts, num_output_values,
                                            is_1d, context));
  }

 private:
  int64_t maxlength_;
  int64_t minlength_;
  bool binary_output_;
  bool validate_;
};

template <class T, class W>
class RaggedCount : public OpKernel {
 public:
  explicit RaggedCount(OpKernelConstruction* context) : OpKernel(context) {
    OP_REQUIRES_OK(context, context->GetAttr("minlength", &minlength_));
    OP_REQUIRES_OK(context, context->GetAttr("maxlength", &maxlength_));
    OP_REQUIRES_OK(context, context->GetAttr("binary_output", &binary_output_));
  }

  void Compute(OpKernelContext* context) override {
    const Tensor& splits = context->input(0);
    const Tensor& values = context->input(1);
    const Tensor& weights = context->input(2);
    bool use_weights = weights.NumElements() > 0;
    bool is_1d = false;

    if (use_weights) {
      OP_REQUIRES(
          context, weights.shape() == values.shape(),
          errors::InvalidArgument(
              "Weights and values must have the same shape. Weight shape: ",
              weights.shape().DebugString(),
              "; values shape: ", values.shape().DebugString()));
    }

    const auto splits_values = splits.flat<int64_t>();
    const auto values_values = values.flat<T>();
    const auto weight_values = weights.flat<W>();
    int num_batches = splits.NumElements() - 1;
    int num_values = values.NumElements();

    OP_REQUIRES(
        context, num_batches > 0,
        errors::InvalidArgument(
            "Must provide at least 2 elements for the splits argument"));
    OP_REQUIRES(context, splits_values(0) == 0,
                errors::InvalidArgument("Splits must start with 0, not with ",
                                        splits_values(0)));
    OP_REQUIRES(context, splits_values(num_batches) == num_values,
                errors::InvalidArgument(
                    "Splits must end with the number of values, got ",
                    splits_values(num_batches), " instead of ", num_values));

    // Ensure all values are non-negative.
    Eigen::TensorFixedSize<bool, Eigen::Sizes<>, Eigen::RowMajor> nonnegative;
    nonnegative.device(context->eigen_cpu_device()) =
        (values_values >= static_cast<T>(0)).all();
    OP_REQUIRES(
        context, nonnegative(),
        errors::InvalidArgument("Input values must all be non-negative"));

    auto per_batch_counts = BatchedMap<W>(num_batches);
    T max_value = 0;
    int batch_idx = 0;

    for (int idx = 0; idx < num_values; ++idx) {
      while (idx >= splits_values(batch_idx)) {
        batch_idx++;
      }
      const auto& value = values_values(idx);
      if (maxlength_ < 0 || value < maxlength_) {
        if (binary_output_) {
          per_batch_counts[batch_idx - 1][value] = 1;
        } else if (use_weights) {
          per_batch_counts[batch_idx - 1][value] += weight_values(idx);
        } else {
          per_batch_counts[batch_idx - 1][value]++;
        }
        if (value > max_value) {
          max_value = value;
        }
      }
    }

    int64_t num_output_values =
        GetOutputSize(max_value, maxlength_, minlength_);
    OP_REQUIRES_OK(context, OutputSparse<W>(per_batch_counts, num_output_values,
                                            is_1d, context));
  }

 private:
  int64_t maxlength_;
  int64_t minlength_;
  bool binary_output_;
  bool validate_;
};

#define REGISTER_W(W_TYPE) \
  REGISTER(int32, W_TYPE)  \
  REGISTER(int64_t, W_TYPE)

#define REGISTER(I_TYPE, W_TYPE)                                     \
                                                                     \
  REGISTER_KERNEL_BUILDER(Name("DenseCountSparseOutput")             \
                              .TypeConstraint<I_TYPE>("T")           \
                              .TypeConstraint<W_TYPE>("output_type") \
                              .Device(DEVICE_CPU),                   \
                          DenseCount<I_TYPE, W_TYPE>)                \
                                                                     \
  REGISTER_KERNEL_BUILDER(Name("SparseCountSparseOutput")            \
                              .TypeConstraint<I_TYPE>("T")           \
                              .TypeConstraint<W_TYPE>("output_type") \
                              .Device(DEVICE_CPU),                   \
                          SparseCount<I_TYPE, W_TYPE>)               \
                                                                     \
  REGISTER_KERNEL_BUILDER(Name("RaggedCountSparseOutput")            \
                              .TypeConstraint<I_TYPE>("T")           \
                              .TypeConstraint<W_TYPE>("output_type") \
                              .Device(DEVICE_CPU),                   \
                          RaggedCount<I_TYPE, W_TYPE>)

TF_CALL_INTEGRAL_TYPES(REGISTER_W);
TF_CALL_float(REGISTER_W);
TF_CALL_double(REGISTER_W);

#undef REGISTER_W
#undef REGISTER

}  // namespace tensorflow