tensorflow--tensorflow/tensorflow/lite/kernels/topk_v2.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 <stdint.h>

#include <algorithm>
#include <iterator>
#include <vector>

#include "tensorflow/lite/core/c/c_api_types.h"
#include "tensorflow/lite/core/c/common.h"
#include "tensorflow/lite/kernels/internal/compatibility.h"
#include "tensorflow/lite/kernels/internal/tensor.h"
#include "tensorflow/lite/kernels/internal/tensor_ctypes.h"
#include "tensorflow/lite/kernels/kernel_util.h"

namespace tflite {
namespace ops {
namespace builtin {
namespace topk_v2 {
constexpr int kInputTensor = 0;
constexpr int kInputTopK = 1;
constexpr int kOutputValues = 0;
constexpr int kOutputIndexes = 1;

namespace {
TfLiteStatus ResizeOutput(TfLiteContext* context, TfLiteNode* node) {
  const TfLiteTensor* top_k;
  TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, kInputTopK, &top_k));
  // INT32 and INT16 number of top result index is supported.
  TF_LITE_ENSURE(context,
                 top_k->type == kTfLiteInt32 || top_k->type == kTfLiteInt16);
  // Check that the tensor contains only one value.
  TF_LITE_ENSURE_EQ(context, NumElements(top_k), 1);

  int32 k;
  if (top_k->type == kTfLiteInt16) {
    k = *GetTensorData<int16_t>(top_k);
  } else {
    // top_k->type == kTfLiteInt32
    k = *GetTensorData<int32_t>(top_k);
  }

  const TfLiteTensor* input;
  TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, kInputTensor, &input));
  const int num_dimensions = NumDimensions(input);
  // Check that input has one or more dimensions.
  TF_LITE_ENSURE_MSG(context, input->dims->size >= 1,
                     "TopK k input must have 1 or more dimensions.");

  // Calculate the size of the last dimension for the output tensors.
  // It should be 0 if k is not positive.
  const int32 output_last_dim = std::max(0, k);

  // Check that k is less or equal the internal dimension *if k is positive*.
  if (k > 0) {
    TF_LITE_ENSURE_MSG(context, k <= input->dims->data[num_dimensions - 1],
                       "TopK k is higher than the internal dimension.");
  }

  TfLiteIntArray* output_indexes_shape = TfLiteIntArrayCreate(num_dimensions);
  TfLiteIntArray* output_values_shape = TfLiteIntArrayCreate(num_dimensions);
  for (int i = 0; i < num_dimensions - 1; ++i) {
    output_indexes_shape->data[i] = input->dims->data[i];
    output_values_shape->data[i] = input->dims->data[i];
  }
  // Use the sanitized dimension size.
  output_indexes_shape->data[num_dimensions - 1] = output_last_dim;
  output_values_shape->data[num_dimensions - 1] = output_last_dim;

  TfLiteTensor* output_indexes;
  TF_LITE_ENSURE_OK(
      context, GetOutputSafe(context, node, kOutputIndexes, &output_indexes));
  TfLiteTensor* output_values;
  TF_LITE_ENSURE_OK(
      context, GetOutputSafe(context, node, kOutputValues, &output_values));
  // Force output types.
  output_values->type = input->type;
  auto resize_tensor = [context](TfLiteTensor* tensor, TfLiteIntArray* new_size,
                                 TfLiteIntArray* delete_on_error) {
    TfLiteStatus status = context->ResizeTensor(context, tensor, new_size);
    if (status != kTfLiteOk) {
      if (delete_on_error != nullptr) {
        TfLiteIntArrayFree(delete_on_error);
      }
    }
    return status;
  };
  TF_LITE_ENSURE_OK(context, resize_tensor(output_indexes, output_indexes_shape,
                                           output_values_shape));
  TF_LITE_ENSURE_OK(context,
                    resize_tensor(output_values, output_values_shape, nullptr));
  return kTfLiteOk;
}

// Class that collects indices of top k values.  Based on template
// tensorflow::gtl::TopN<> but, for optimization, it re-uses the same container.
template <typename T, typename Tidx>
class TopContainer {
 public:
  TopContainer() = delete;
  TopContainer(int32 k, int32 row_size) : k_(k) {
    container_.reserve(std::min(k, row_size) + 1);
  }

  void start_collecting(const T* values) {
    values_ = values;
    container_.clear();
    is_heap_ = false;
  }

  void push(Tidx a) {
    auto comparator = [this](Tidx a, Tidx b) { return compare_fun(a, b); };
    if (!is_heap_) {
      container_.push_back(a);
      if (container_.size() == k_ + 1) {
        std::make_heap(container_.begin(), container_.end(), comparator);
        std::pop_heap(container_.begin(), container_.end(), comparator);
        container_.pop_back();
        is_heap_ = true;
      }
    } else if (comparator(a, container_.front())) {
      // Due to how we defined comparator / compare_fun, container_.front()
      // contains the index of the smallest of the top-k elements seen so far.
      //
      // If control reaches this point, we know that the current index a
      // corresponds to an element which is bigger than the smallest of the
      // top-k elements seen so far.  Hence, we have to update the indices of
      // the top-k elements, by removing the index of the smallest top-k
      // element, adding a, and making sure container_[0:k] is still a heap.
      std::pop_heap(container_.begin(), container_.end(), comparator);
      container_.back() = a;
      std::push_heap(container_.begin(), container_.end(), comparator);
    }
  }

  const std::vector<Tidx>& sorted_result() {
    auto comparator = [this](Tidx a, Tidx b) { return compare_fun(a, b); };
    if (!is_heap_) {
      // Note: due to the way we defined compare_fun (see comments for that
      // function) std::sort puts the indices from container_ in decreasing
      // order of the corresponding elements.
      std::sort(container_.begin(), container_.end(), comparator);
    } else {
      std::sort_heap(container_.begin(), container_.end(), comparator);
    }
    return container_;
  }

 private:
  const int32 k_;

  // container_[0,k) holds the indices of the largest k elements from values_
  // seen so far.  If more than k elements are pushed, then elements are
  // maintained in a min-heap order: container_.front() is
  // the index of the smallest of the top-k elements see so far.
  std::vector<Tidx> container_;

  // Once more than k elements are pushed, the container becomes a min heap,
  // and is_heap_ becomes true.
  bool is_heap_ = false;

  const T* values_ = nullptr;

  // Compares indices a and b based on the corresponding elements from values_.
  //
  // Intuitively, compare_fun(a, b) returns true iff values_[b] < values_[a]
  // (notice the inversion of direction, not a typo); ties (==) are broken in
  // favor of earlier elements (i.e., a < b).
  bool compare_fun(Tidx a, Tidx b) const {
    if (values_[b] < values_[a]) {
      return true;
    } else if (values_[b] > values_[a]) {
      return false;
    } else {
      return a < b;
    }
  }
};

// Mostly modeled on tensorflow/core/kernels/topk_op.cc for CPU.
template <typename T, typename Tidx = int32>
void TopK(int32 row_size, int32 num_rows, const T* data, int32 k,
          Tidx* output_indexes, T* output_values) {
  if (k <= 0) {
    // If k is 0 or negative, there are no top elements to find.
    // The output tensors should already be sized with the last dimension as 0
    // by the Prepare/ResizeOutput functions.
    return;
  }
  TopContainer<T, Tidx> topc(k, row_size);
  for (int row = 0; row < num_rows; ++row) {
    const T* values_row = data + row * row_size;
    topc.start_collecting(values_row);
    for (int32 c = 0; c < row_size; ++c) {
      topc.push(c);
    }

    // Prepare output buffers.
    Tidx* indexes_row = output_indexes + row * k;
    T* output_row = output_values + row * k;
    // We always assume that the output is sorted.
    const auto& top_k = topc.sorted_result();
    std::copy(top_k.begin(), top_k.end(), indexes_row);
    std::transform(top_k.begin(), top_k.end(), output_row,
                   [values_row](const int32 loc) { return values_row[loc]; });
  }
}

}  // namespace

TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
  // Check that the inputs and outputs have the right sizes and types.
  TF_LITE_ENSURE_EQ(context, NumInputs(node), 2);
  TF_LITE_ENSURE_EQ(context, NumOutputs(node), 2);

  const TfLiteTensor* input;
  TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, kInputTensor, &input));
  TfLiteTensor* output_values;
  TF_LITE_ENSURE_OK(
      context, GetOutputSafe(context, node, kOutputValues, &output_values));
  TF_LITE_ENSURE_TYPES_EQ(context, input->type, output_values->type);

  const TfLiteTensor* top_k;
  TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, kInputTopK, &top_k));

  TF_LITE_ENSURE(context,
                 top_k->type != kTfLiteInt32 || top_k->type != kTfLiteInt16);

  // Set output dynamic if the `top_k` tensor is not constant, or the input has
  // dynamic dimensions (indicated by dims signature).
  if (IsConstantOrPersistentTensor(top_k) && !HasUnspecifiedDimension(input)) {
    TF_LITE_ENSURE_OK(context, ResizeOutput(context, node));
  } else {
    TfLiteTensor* output_indexes;
    TF_LITE_ENSURE_OK(
        context, GetOutputSafe(context, node, kOutputIndexes, &output_indexes));
    TfLiteTensor* output_values;
    TF_LITE_ENSURE_OK(
        context, GetOutputSafe(context, node, kOutputValues, &output_values));
    SetTensorToDynamic(output_indexes);
    SetTensorToDynamic(output_values);
  }
  return kTfLiteOk;
}

template <typename idx_type>
TfLiteStatus TopKImpl(TfLiteContext* context, TfLiteNode* node, int32_t k,
                      idx_type* output_indexes) {
  // The tensor can have more than 2 dimensions or even be a vector, the code
  // anyway calls the internal dimension as row;
  const TfLiteTensor* input;
  TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, kInputTensor, &input));
  TfLiteTensor* output_values;
  TF_LITE_ENSURE_OK(
      context, GetOutputSafe(context, node, kOutputValues, &output_values));

  const int32 row_size = input->dims->data[input->dims->size - 1];
  int32 num_rows = 1;
  for (int i = 0; i < input->dims->size - 1; ++i) {
    num_rows *= input->dims->data[i];
  }
  switch (output_values->type) {
    case kTfLiteFloat32:
      TopK(row_size, num_rows, GetTensorData<float>(input), k, output_indexes,
           GetTensorData<float>(output_values));
      break;
    case kTfLiteUInt8:
      TopK(row_size, num_rows, GetTensorData<uint8_t>(input), k, output_indexes,
           output_values->data.uint8);
      break;
    case kTfLiteInt8:
      TopK(row_size, num_rows, GetTensorData<int8_t>(input), k, output_indexes,
           output_values->data.int8);
      break;
    case kTfLiteInt16:
      TopK(row_size, num_rows, GetTensorData<int16_t>(input), k, output_indexes,
           output_values->data.i16);
      break;
    case kTfLiteInt32:
      TopK(row_size, num_rows, GetTensorData<int32_t>(input), k, output_indexes,
           output_values->data.i32);
      break;
    case kTfLiteInt64:
      TopK(row_size, num_rows, GetTensorData<int64_t>(input), k, output_indexes,
           output_values->data.i64);
      break;
    default:
      TF_LITE_KERNEL_LOG(context, "Type %s is currently not supported by TopK.",
                         TfLiteTypeGetName(output_values->type));
      return kTfLiteError;
  }
  return kTfLiteOk;
}

TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
  TfLiteTensor* output_values;
  TF_LITE_ENSURE_OK(
      context, GetOutputSafe(context, node, kOutputValues, &output_values));
  TfLiteTensor* output_indexes;
  TF_LITE_ENSURE_OK(
      context, GetOutputSafe(context, node, kOutputIndexes, &output_indexes));
  if (IsDynamicTensor(output_values)) {
    TF_LITE_ENSURE_OK(context, ResizeOutput(context, node));
  }
  const TfLiteTensor* top_k;
  TF_LITE_ENSURE_OK(context, GetInputSafe(context, node, kInputTopK, &top_k));
  int32 k;

  switch (top_k->type) {
    case kTfLiteInt32:
      k = top_k->data.i32[0];
      break;
    case kTfLiteInt16:
      k = top_k->data.i16[0];
      break;
    default:
      TF_LITE_KERNEL_LOG(context,
                         "Type %s is currently not supported k Type by TopK.",
                         TfLiteTypeGetName(output_values->type));
      return kTfLiteError;
  }

  switch (output_indexes->type) {
    case kTfLiteInt32: {
      return TopKImpl(context, node, k, GetTensorData<int32_t>(output_indexes));
    }
    case kTfLiteInt16: {
      return TopKImpl(context, node, k, GetTensorData<int16_t>(output_indexes));
    }
    default:
      TF_LITE_KERNEL_LOG(
          context, "Output index type %s is currently not supported by TopK.",
          TfLiteTypeGetName(output_values->type));
      return kTfLiteError;
  }
}

}  // namespace topk_v2
TfLiteRegistration* Register_TOPK_V2() {
  static TfLiteRegistration r = {nullptr, nullptr, topk_v2::Prepare,
                                 topk_v2::Eval};
  return &r;
}
}  // namespace builtin
}  // namespace ops
}  // namespace tflite