pytorch/caffe2/operators/self_binning_histogram_op.h

#pragma once

#include "caffe2/core/operator.h"
#include "c10/util/irange.h"

#include <algorithm>
#include <cmath>
#include <limits>

namespace caffe2 {

template <class Context>
class SelfBinningHistogramOp final : public Operator<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  template <class... Args>
  explicit SelfBinningHistogramOp(Args&&... args)
      : Operator<Context>(std::forward<Args>(args)...),
        num_bins_(this->template GetSingleArgument<int>("num_bins", 0)),
        num_edges_(num_bins_ + 1),
        bin_spacing_(this->template GetSingleArgument<std::string>(
            "bin_spacing",
            "linear")),
        logspace_start_(this->template GetSingleArgument<float>("logspace_start", 1e-24)),
        abs_(this->template GetSingleArgument<bool>("abs", false))
         {
    CAFFE_ENFORCE_GE(
        num_bins_, 1, "Number of bins must be greater than or equal to 1.");
    CAFFE_ENFORCE(
        bin_spacing_ == "linear" || bin_spacing_ == "logarithmic",
        "Bin spacing can be one of 'linear' or 'logarithmic'."
    );
    CAFFE_ENFORCE_GT(
      logspace_start_, 0,
      "Logarithmic spacing base is a multiplier and is expected to be >1.");
  }

  bool RunOnDevice() override {
    return DispatchHelper<TensorTypes<float, double>>::call(this, Input(0));
  }

  template <typename T>
  bool DoRunWithType() {
    CheckInputs();

    // Scale the range so that the last count is always 0.
    const T RANGE_SCALING = 1.0001;

    const auto* histogram_values = Output(HISTOGRAM_VALUES);
    histogram_values->Resize(num_edges_);
    auto* histogram_values_data = histogram_values->template mutable_data<T>();
    const auto* histogram_counts = Output(HISTOGRAM_COUNTS);
    histogram_counts->Resize(num_edges_);
    auto* histogram_counts_data =
        histogram_counts->template mutable_data<int64_t>();

    // Calculate the max and min.
    bool first_seen = false;
    // 0 initialization is arbitrary here to suppress linter warnings.
    // The actual initialization check happens through the first_seen variable.
    T max = 0;
    T min = 0;
    int64_t total_count = 0;
    for (const auto input_idx : c10::irange(InputSize())) {
      const auto& x = Input(input_idx);
      const int64_t N = x.numel();
      total_count += N;
      const auto* x_data = x.template data<T>();
      for (const auto data_idx : c10::irange(N)) {
        const T val = this->abs_ ? std::abs(x_data[data_idx]) :  x_data[data_idx];
        if (!first_seen) {
          max = val;
          min = val;
          first_seen = true;
        } else {
          max = std::max(val, max);
          min = std::min(val, min);
        }
      }
    }

    if (!first_seen) {
      math::Set<T, Context>(num_edges_, 0, histogram_values_data, &context_);
      math::Set<int64_t, Context>(
          num_edges_, 0, histogram_counts_data, &context_);
      return true;
    }

    CAFFE_ENFORCE(min <= max, "Incorrect min-max computation min=", min, " max=", max);
    T scaled_max = 0;  // this is set in both branches
    if (bin_spacing_ == "linear") {
      // Let's scale the range so that the last count is 0.
      scaled_max = min + (max - min) * RANGE_SCALING;
      T scaled_range = (scaled_max - min);
      // Avoid underflow by calculating advancement through multiplication.
      for (const auto i : c10::irange(num_edges_)) {
        T advancement_ratio = T(i) / num_bins_;
        histogram_values_data[i] = min + advancement_ratio * scaled_range;
      }
    } else if (bin_spacing_ == "logarithmic") {
      // First, we need to sanitize the range.
      if (min < logspace_start_) {
        min = logspace_start_;
      }
      if (max < logspace_start_) {
        max = logspace_start_;
      }
      T linear_range = max - min;
      linear_range = linear_range * RANGE_SCALING;
      scaled_max = min + linear_range;
      // Calculate base interval using geometric sum.
      // Simply: multiplier = exp((log(max) - log(min))/N)
      // Avoid underflow by delaying division and exp.
      T log_multiplier_numerator =log(scaled_max) - log(min);
      // Avoid underflow by:
      // - Calculating each advancement separately for each i.
      for (const auto i : c10::irange(num_edges_)) {
        T advancement_ratio = T(i)/num_bins_;
        histogram_values_data[i] = min * exp(log_multiplier_numerator * advancement_ratio);
      }
    }

    math::Set<int64_t, Context>(
      num_edges_, 0, histogram_counts_data, &context_);
    if (histogram_values_data[num_edges_-1] <= max) {
      // In cases of min&max being equal (or any unexpected numerical underflow) we
      // may not have a final edge larger than the max.
      histogram_values_data[num_edges_-1] = scaled_max;
      histogram_counts_data[0] = total_count;
    }
    else {
      for (const auto input_idx : c10::irange(InputSize())) {
        const auto& x = Input(input_idx);
        const int64_t N = x.numel();
        const auto* x_data = x.template data<T>();
        for (const auto data_idx : c10::irange(N)) {
          const T val = this->abs_ ? std::abs(x_data[data_idx]) :  x_data[data_idx];
          const auto bisection_it = std::upper_bound(
              histogram_values_data,
              histogram_values_data + num_edges_,
              val);
          const int bisection_idx = bisection_it - histogram_values_data;
          if (bisection_idx > 0 && bisection_idx < num_edges_) {
            histogram_counts_data[bisection_idx - 1]++;
          }
          if (bisection_idx == 0) {
            histogram_counts_data[0]++;
          }
        }
      }
    }

    return true;
  }

 protected:
  OUTPUT_TAGS(HISTOGRAM_VALUES, HISTOGRAM_COUNTS);

 private:
  int num_bins_;
  int num_edges_;
  std::string bin_spacing_;
  float logspace_start_;
  bool abs_; // automatically apply abs() on the input values

  void CheckInputs() {
    const auto& input_zero = Input(0);
    for (const auto i : c10::irange(1, InputSize())) {
      CAFFE_ENFORCE_EQ(
          Input(i).dtype(),
          input_zero.dtype(),
          "All inputs must have the same type; expected ",
          input_zero.dtype().name(),
          " but got ",
          Input(i).dtype().name(),
          " for input ",
          i);
    }
  }
};

} // namespace caffe2