xgboost/src/metric/survival_metric.cu

/*!
 * Copyright 2019-2020 by Contributors
 * \file survival_metric.cu
 * \brief Metrics for survival analysis
 * \author Avinash Barnwal, Hyunsu Cho and Toby Hocking
 */

#include <rabit/rabit.h>
#include <dmlc/registry.h>

#include <memory>
#include <vector>

#include "xgboost/json.h"
#include "xgboost/metric.h"
#include "xgboost/host_device_vector.h"

#include "metric_common.h"
#include "../common/math.h"
#include "../common/survival_util.h"

#if defined(XGBOOST_USE_CUDA)
#include <thrust/execution_policy.h>  // thrust::cuda::par
#include "../common/device_helpers.cuh"
#endif  // XGBOOST_USE_CUDA

using AFTParam = xgboost::common::AFTParam;
using ProbabilityDistributionType = xgboost::common::ProbabilityDistributionType;
template <typename Distribution>
using AFTLoss = xgboost::common::AFTLoss<Distribution>;

namespace xgboost {
namespace metric {
// tag the this file, used by force static link later.
DMLC_REGISTRY_FILE_TAG(survival_metric);

template <typename EvalRow>
class ElementWiseSurvivalMetricsReduction {
 public:
  ElementWiseSurvivalMetricsReduction() = default;
  void Configure(EvalRow policy) {
    policy_ = policy;
  }

  PackedReduceResult CpuReduceMetrics(
      const HostDeviceVector<bst_float>& weights,
      const HostDeviceVector<bst_float>& labels_lower_bound,
      const HostDeviceVector<bst_float>& labels_upper_bound,
      const HostDeviceVector<bst_float>& preds) const {
    size_t ndata = labels_lower_bound.Size();
    CHECK_EQ(ndata, labels_upper_bound.Size());

    const auto& h_labels_lower_bound = labels_lower_bound.HostVector();
    const auto& h_labels_upper_bound = labels_upper_bound.HostVector();
    const auto& h_weights = weights.HostVector();
    const auto& h_preds = preds.HostVector();

    double residue_sum = 0;
    double weights_sum = 0;

    dmlc::OMPException exc;
#pragma omp parallel for reduction(+: residue_sum, weights_sum) schedule(static)
    for (omp_ulong i = 0; i < ndata; ++i) {
      exc.Run([&]() {
        const double wt = h_weights.empty() ? 1.0 : static_cast<double>(h_weights[i]);
        residue_sum += policy_.EvalRow(
          static_cast<double>(h_labels_lower_bound[i]),
          static_cast<double>(h_labels_upper_bound[i]),
          static_cast<double>(h_preds[i])) * wt;
        weights_sum += wt;
      });
    }
    exc.Rethrow();
    PackedReduceResult res{residue_sum, weights_sum};
    return res;
  }

#if defined(XGBOOST_USE_CUDA)

  PackedReduceResult DeviceReduceMetrics(
      const HostDeviceVector<bst_float>& weights,
      const HostDeviceVector<bst_float>& labels_lower_bound,
      const HostDeviceVector<bst_float>& labels_upper_bound,
      const HostDeviceVector<bst_float>& preds) {
    size_t ndata = labels_lower_bound.Size();
    CHECK_EQ(ndata, labels_upper_bound.Size());

    thrust::counting_iterator<size_t> begin(0);
    thrust::counting_iterator<size_t> end = begin + ndata;

    auto s_label_lower_bound = labels_lower_bound.DeviceSpan();
    auto s_label_upper_bound = labels_upper_bound.DeviceSpan();
    auto s_preds = preds.DeviceSpan();
    auto s_weights = weights.DeviceSpan();

    const bool is_null_weight = (weights.Size() == 0);

    auto d_policy = policy_;

    dh::XGBCachingDeviceAllocator<char> alloc;
    PackedReduceResult result = thrust::transform_reduce(
      thrust::cuda::par(alloc),
      begin, end,
      [=] XGBOOST_DEVICE(size_t idx) {
        double weight = is_null_weight ? 1.0 : static_cast<double>(s_weights[idx]);
        double residue = d_policy.EvalRow(
            static_cast<double>(s_label_lower_bound[idx]),
            static_cast<double>(s_label_upper_bound[idx]),
            static_cast<double>(s_preds[idx]));
        residue *= weight;
        return PackedReduceResult{residue, weight};
      },
      PackedReduceResult(),
      thrust::plus<PackedReduceResult>());

    return result;
  }

#endif  // XGBOOST_USE_CUDA

  PackedReduceResult Reduce(
      int device,
      const HostDeviceVector<bst_float>& weights,
      const HostDeviceVector<bst_float>& labels_lower_bound,
      const HostDeviceVector<bst_float>& labels_upper_bound,
      const HostDeviceVector<bst_float>& preds) {
    PackedReduceResult result;

    if (device < 0) {
      result = CpuReduceMetrics(weights, labels_lower_bound, labels_upper_bound, preds);
    }
#if defined(XGBOOST_USE_CUDA)
    else {  // NOLINT
      device_ = device;
      preds.SetDevice(device_);
      labels_lower_bound.SetDevice(device_);
      labels_upper_bound.SetDevice(device_);
      weights.SetDevice(device_);

      dh::safe_cuda(cudaSetDevice(device_));
      result = DeviceReduceMetrics(weights, labels_lower_bound, labels_upper_bound, preds);
    }
#endif  // defined(XGBOOST_USE_CUDA)
    return result;
  }

 private:
  EvalRow policy_;
#if defined(XGBOOST_USE_CUDA)
  int device_{-1};
#endif  // defined(XGBOOST_USE_CUDA)
};

struct EvalIntervalRegressionAccuracy {
  void Configure(const Args& args) {}

  const char* Name() const {
    return "interval-regression-accuracy";
  }

  XGBOOST_DEVICE double EvalRow(
      double label_lower_bound, double label_upper_bound, double log_pred) const {
    const double pred = exp(log_pred);
    return (pred >= label_lower_bound && pred <= label_upper_bound) ? 1.0 : 0.0;
  }

  static double GetFinal(double esum, double wsum) {
    return wsum == 0 ? esum : esum / wsum;
  }
};

/*! \brief Negative log likelihood of Accelerated Failure Time model */
template <typename Distribution>
struct EvalAFTNLogLik {
  void Configure(const Args& args) {
    param_.UpdateAllowUnknown(args);
  }

  const char* Name() const {
    return "aft-nloglik";
  }

  XGBOOST_DEVICE double EvalRow(
      double label_lower_bound, double label_upper_bound, double pred) const {
    return AFTLoss<Distribution>::Loss(
        label_lower_bound, label_upper_bound, pred, param_.aft_loss_distribution_scale);
  }

  static double GetFinal(double esum, double wsum) {
    return wsum == 0 ? esum : esum / wsum;
  }
 private:
  AFTParam param_;
};

template<typename Policy>
struct EvalEWiseSurvivalBase : public Metric {
  EvalEWiseSurvivalBase() = default;

  void Configure(const Args& args) override {
    policy_.Configure(args);
    for (const auto& e : args) {
      if (e.first == "gpu_id") {
        device_ = dmlc::ParseSignedInt<int>(e.second.c_str(), nullptr, 10);
      }
    }
    reducer_.Configure(policy_);
  }

  bst_float Eval(const HostDeviceVector<bst_float>& preds,
                 const MetaInfo& info,
                 bool distributed) override {
    CHECK_EQ(preds.Size(), info.labels_lower_bound_.Size());
    CHECK_EQ(preds.Size(), info.labels_upper_bound_.Size());

    auto result = reducer_.Reduce(
        device_, info.weights_, info.labels_lower_bound_, info.labels_upper_bound_, preds);

    double dat[2] {result.Residue(), result.Weights()};

    if (distributed) {
      rabit::Allreduce<rabit::op::Sum>(dat, 2);
    }
    return static_cast<bst_float>(Policy::GetFinal(dat[0], dat[1]));
  }

  const char* Name() const override {
    return policy_.Name();
  }

 private:
  Policy policy_;
  ElementWiseSurvivalMetricsReduction<Policy> reducer_;
  int device_{-1};  // used only for GPU metric
};

// This class exists because we want to perform dispatch according to the distribution type at
// configuration time, not at prediction time.
struct AFTNLogLikDispatcher : public Metric {
  const char* Name() const override {
    return "aft-nloglik";
  }

  bst_float Eval(const HostDeviceVector<bst_float>& preds,
                 const MetaInfo& info,
                 bool distributed) override {
    CHECK(metric_) << "AFT metric must be configured first, with distribution type and scale";
    return metric_->Eval(preds, info, distributed);
  }

  void Configure(const Args& args) override {
    param_.UpdateAllowUnknown(args);
    switch (param_.aft_loss_distribution) {
    case common::ProbabilityDistributionType::kNormal:
      metric_.reset(new EvalEWiseSurvivalBase<EvalAFTNLogLik<common::NormalDistribution>>());
      break;
    case common::ProbabilityDistributionType::kLogistic:
      metric_.reset(new EvalEWiseSurvivalBase<EvalAFTNLogLik<common::LogisticDistribution>>());
      break;
    case common::ProbabilityDistributionType::kExtreme:
      metric_.reset(new EvalEWiseSurvivalBase<EvalAFTNLogLik<common::ExtremeDistribution>>());
      break;
    default:
      LOG(FATAL) << "Unknown probability distribution";
    }
    Args new_args{args};
    // tparam_ doesn't get propagated to the inner metric object because we didn't use
    // Metric::Create(). I don't think it's a good idea to pollute the metric registry with
    // specialized versions of the AFT metric, so as a work-around, manually pass the GPU ID
    // into the inner metric via configuration.
    new_args.emplace_back("gpu_id", std::to_string(tparam_->gpu_id));
    metric_->Configure(new_args);
  }

  void SaveConfig(Json* p_out) const override {
    auto& out = *p_out;
    out["name"] = String(this->Name());
    out["aft_loss_param"] = ToJson(param_);
  }

  void LoadConfig(const Json& in) override {
    FromJson(in["aft_loss_param"], &param_);
  }

 private:
  AFTParam param_;
  std::unique_ptr<Metric> metric_;
};


XGBOOST_REGISTER_METRIC(AFTNLogLik, "aft-nloglik")
.describe("Negative log likelihood of Accelerated Failure Time model.")
.set_body([](const char* param) {
  return new AFTNLogLikDispatcher();
});

XGBOOST_REGISTER_METRIC(IntervalRegressionAccuracy, "interval-regression-accuracy")
.describe("")
.set_body([](const char* param) {
  return new EvalEWiseSurvivalBase<EvalIntervalRegressionAccuracy>();
});

}  // namespace metric
}  // namespace xgboost