Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

implementation of map ranking algorithm on gpu (#5129)

Browse Source 
* - implementation of map ranking algorithm
  - also effected necessary suggestions mentioned in the earlier ranking pr's
  - made some performance improvements to the ndcg algo as well
This commit is contained in:
sriramch
2019-12-26 15:05:38 -08:00
committed by Rory Mitchell
parent 9b0af6e882
commit ee81ba8e1f
6 changed files with 714 additions and 266 deletions
Download Patch File Download Diff File Expand all files Collapse all files

78
tests/cpp/common/test_device_helpers.cu
View File

@@ -84,3 +84,81 @@ void TestAllocator() {
TEST(bulkAllocator, Test) {
TestAllocator();
}
template <typename T, typename Comp = thrust::less<T>>
void TestUpperBoundImpl(const std::vector<T> &vec, T val_to_find,
const Comp &comp = Comp()) {
EXPECT_EQ(dh::UpperBound(vec.data(), vec.size(), val_to_find, comp),
std::upper_bound(vec.begin(), vec.end(), val_to_find, comp) - vec.begin());
}
template <typename T, typename Comp = thrust::less<T>>
void TestLowerBoundImpl(const std::vector<T> &vec, T val_to_find,
const Comp &comp = Comp()) {
EXPECT_EQ(dh::LowerBound(vec.data(), vec.size(), val_to_find, comp),
std::lower_bound(vec.begin(), vec.end(), val_to_find, comp) - vec.begin());
}
TEST(UpperBound, DataAscending) {
std::vector<int> hvec{0, 3, 5, 5, 7, 8, 9, 10, 10};
// Test boundary conditions
TestUpperBoundImpl(hvec, hvec.front()); // Result 1
TestUpperBoundImpl(hvec, hvec.front() - 1); // Result 0
TestUpperBoundImpl(hvec, hvec.back() + 1); // Result hvec.size()
TestUpperBoundImpl(hvec, hvec.back()); // Result hvec.size()
// Test other values - both missing and present
TestUpperBoundImpl(hvec, 3); // Result 2
TestUpperBoundImpl(hvec, 4); // Result 2
TestUpperBoundImpl(hvec, 5); // Result 4
}
TEST(UpperBound, DataDescending) {
std::vector<int> hvec{10, 10, 9, 8, 7, 5, 5, 3, 0, 0};
const auto &comparator = thrust::greater<int>();
// Test boundary conditions
TestUpperBoundImpl(hvec, hvec.front(), comparator); // Result 2
TestUpperBoundImpl(hvec, hvec.front() + 1, comparator); // Result 0
TestUpperBoundImpl(hvec, hvec.back(), comparator); // Result hvec.size()
TestUpperBoundImpl(hvec, hvec.back() - 1, comparator); // Result hvec.size()
// Test other values - both missing and present
TestUpperBoundImpl(hvec, 9, comparator); // Result 3
TestUpperBoundImpl(hvec, 7, comparator); // Result 5
TestUpperBoundImpl(hvec, 4, comparator); // Result 7
TestUpperBoundImpl(hvec, 8, comparator); // Result 4
}
TEST(LowerBound, DataAscending) {
std::vector<int> hvec{0, 3, 5, 5, 7, 8, 9, 10, 10};
// Test boundary conditions
TestLowerBoundImpl(hvec, hvec.front()); // Result 0
TestLowerBoundImpl(hvec, hvec.front() - 1); // Result 0
TestLowerBoundImpl(hvec, hvec.back()); // Result 7
TestLowerBoundImpl(hvec, hvec.back() + 1); // Result hvec.size()
// Test other values - both missing and present
TestLowerBoundImpl(hvec, 3); // Result 1
TestLowerBoundImpl(hvec, 4); // Result 2
TestLowerBoundImpl(hvec, 5); // Result 2
}
TEST(LowerBound, DataDescending) {
std::vector<int> hvec{10, 10, 9, 8, 7, 5, 5, 3, 0, 0};
const auto &comparator = thrust::greater<int>();
// Test boundary conditions
TestLowerBoundImpl(hvec, hvec.front(), comparator); // Result 0
TestLowerBoundImpl(hvec, hvec.front() + 1, comparator); // Result 0
TestLowerBoundImpl(hvec, hvec.back(), comparator); // Result 8
TestLowerBoundImpl(hvec, hvec.back() - 1, comparator); // Result hvec.size()
// Test other values - both missing and present
TestLowerBoundImpl(hvec, 9, comparator); // Result 2
TestLowerBoundImpl(hvec, 7, comparator); // Result 4
TestLowerBoundImpl(hvec, 4, comparator); // Result 7
TestLowerBoundImpl(hvec, 8, comparator); // Result 3
}

29
tests/cpp/objective/test_ranking_obj.cc
View File

@@ -105,4 +105,33 @@ TEST(Objective, DeclareUnifiedTest(NDCGRankingGPair)) {
ASSERT_NO_THROW(obj->DefaultEvalMetric());
}
TEST(Objective, DeclareUnifiedTest(MAPRankingGPair)) {
std::vector<std::pair<std::string, std::string>> args;
xgboost::GenericParameter lparam = xgboost::CreateEmptyGenericParam(GPUIDX);
std::unique_ptr<xgboost::ObjFunction> obj {
xgboost::ObjFunction::Create("rank:map", &lparam)
};
obj->Configure(args);
CheckConfigReload(obj, "rank:map");
// Test with setting sample weight to second query group
CheckRankingObjFunction(obj,
{0, 0.1f, 0, 0.1f},
{0, 1, 0, 1},
{2.0f, 0.0f},
{0, 2, 4},
{0.95f, -0.95f, 0.0f, 0.0f},
{0.9975f, 0.9975f, 0.0f, 0.0f});
CheckRankingObjFunction(obj,
{0, 0.1f, 0, 0.1f},
{0, 1, 0, 1},
{1.0f, 1.0f},
{0, 2, 4},
{0.475f, -0.475f, 0.475f, -0.475f},
{0.4988f, 0.4988f, 0.4988f, 0.4988f});
ASSERT_NO_THROW(obj->DefaultEvalMetric());
}
} // namespace xgboost

129
tests/cpp/objective/test_ranking_obj_gpu.cu
View File

@@ -19,22 +19,22 @@ RankSegmentSorterTestImpl(const std::vector<uint32_t> &group_indices,
dh::device_vector<T> dlabels(hlabels);
seg_sorter.SortItems(dlabels.data().get(), dlabels.size(), group_indices, Comparator());
EXPECT_EQ(seg_sorter.NumItems(), group_indices.back());
EXPECT_EQ(seg_sorter.NumGroups(), group_indices.size() - 1);
EXPECT_EQ(seg_sorter.GetNumItems(), group_indices.back());
EXPECT_EQ(seg_sorter.GetNumGroups(), group_indices.size() - 1);
// Check the labels
dh::device_vector<T> sorted_dlabels(seg_sorter.NumItems());
sorted_dlabels.assign(thrust::device_ptr<const T>(seg_sorter.Items()),
thrust::device_ptr<const T>(seg_sorter.Items())
+ seg_sorter.NumItems());
dh::device_vector<T> sorted_dlabels(seg_sorter.GetNumItems());
sorted_dlabels.assign(thrust::device_ptr<const T>(seg_sorter.GetItemsPtr()),
thrust::device_ptr<const T>(seg_sorter.GetItemsPtr())
+ seg_sorter.GetNumItems());
thrust::host_vector<T> sorted_hlabels(sorted_dlabels);
EXPECT_EQ(expected_sorted_hlabels, sorted_hlabels);
// Check the indices
dh::device_vector<uint32_t> dorig_pos(seg_sorter.NumItems());
dorig_pos.assign(thrust::device_ptr<const uint32_t>(seg_sorter.OriginalPositions()),
thrust::device_ptr<const uint32_t>(seg_sorter.OriginalPositions())
+ seg_sorter.NumItems());
dh::device_vector<uint32_t> dorig_pos(seg_sorter.GetNumItems());
dorig_pos.assign(thrust::device_ptr<const uint32_t>(seg_sorter.GetOriginalPositionsPtr()),
thrust::device_ptr<const uint32_t>(seg_sorter.GetOriginalPositionsPtr())
+ seg_sorter.GetNumItems());
dh::device_vector<uint32_t> horig_pos(dorig_pos);
EXPECT_EQ(expected_orig_pos, horig_pos);
@@ -125,11 +125,14 @@ TEST(Objective, RankItemCountOnRight) {
}
TEST(Objective, NDCGLambdaWeightComputerTest) {
std::vector<float> hlabels = {3.1f, 1.2f, 2.3f, 4.4f, // Labels
7.8f, 5.01f, 6.96f,
10.3f, 8.7f, 11.4f, 9.45f, 11.4f};
dh::device_vector<bst_float> dlabels(hlabels);
auto segment_label_sorter = RankSegmentSorterTestImpl<float>(
{0, 4, 7, 12}, // Groups
{3.1f, 1.2f, 2.3f, 4.4f, // Labels
7.8f, 5.01f, 6.96f,
10.3f, 8.7f, 11.4f, 9.45f, 11.4f},
hlabels,
{4.4f, 3.1f, 2.3f, 1.2f, // Expected sorted labels
7.8f, 6.96f, 5.01f,
11.4f, 11.4f, 10.3f, 9.45f, 8.7f},
@@ -142,18 +145,114 @@ TEST(Objective, NDCGLambdaWeightComputerTest) {
-1.03f, -2.79f, -3.1f,
104.22f, 103.1f, -101.7f, 100.5f, 45.1f};
dh::device_vector<bst_float> dpreds(hpreds);
xgboost::obj::NDCGLambdaWeightComputer ndcg_lw_computer(dpreds.data().get(),
dpreds.size(),
dlabels.data().get(),
*segment_label_sorter);
// Where will the predictions move from its current position, if they were sorted
// descendingly?
auto dsorted_pred_pos = ndcg_lw_computer.GetSortedPredPos();
auto dsorted_pred_pos = ndcg_lw_computer.GetPredictionSorter().GetIndexableSortedPositions();
thrust::host_vector<uint32_t> hsorted_pred_pos(dsorted_pred_pos);
std::vector<uint32_t> expected_sorted_pred_pos{2, 0, 1, 3,
4, 5, 6,
7, 8, 11, 9, 10};
EXPECT_EQ(expected_sorted_pred_pos, hsorted_pred_pos);
// Check group DCG values
thrust::host_vector<float> hgroup_dcgs(ndcg_lw_computer.GetGroupDcgs());
thrust::host_vector<uint32_t> hgroups(segment_label_sorter->GetGroups());
thrust::host_vector<float> hsorted_labels(segment_label_sorter->GetItems());
EXPECT_EQ(hgroup_dcgs.size(), segment_label_sorter->GetNumGroups());
for (auto i = 0; i < hgroup_dcgs.size(); ++i) {
// Compute group DCG value on CPU and compare
auto gbegin = hgroups[i];
auto gend = hgroups[i + 1];
EXPECT_NEAR(
hgroup_dcgs[i],
xgboost::obj::NDCGLambdaWeightComputer::ComputeGroupDCGWeight(&hsorted_labels[gbegin],
gend - gbegin),
0.01f);
}
}
TEST(Objective, IndexableSortedItemsTest) {
std::vector<float> hlabels = {3.1f, 1.2f, 2.3f, 4.4f, // Labels
7.8f, 5.01f, 6.96f,
10.3f, 8.7f, 11.4f, 9.45f, 11.4f};
dh::device_vector<bst_float> dlabels(hlabels);
auto segment_label_sorter = RankSegmentSorterTestImpl<float>(
{0, 4, 7, 12}, // Groups
hlabels,
{4.4f, 3.1f, 2.3f, 1.2f, // Expected sorted labels
7.8f, 6.96f, 5.01f,
11.4f, 11.4f, 10.3f, 9.45f, 8.7f},
{3, 0, 2, 1, // Expected original positions
4, 6, 5,
9, 11, 7, 10, 8});
segment_label_sorter->CreateIndexableSortedPositions();
thrust::host_vector<uint32_t> sorted_indices(segment_label_sorter->GetIndexableSortedPositions());
std::vector<uint32_t> expected_sorted_indices = {
1, 3, 2, 0,
4, 6, 5,
9, 11, 7, 10, 8};
EXPECT_EQ(expected_sorted_indices, sorted_indices);
}
TEST(Objective, ComputeAndCompareMAPStatsTest) {
std::vector<float> hlabels = {3.1f, 0.0f, 2.3f, 4.4f, // Labels
0.0f, 5.01f, 0.0f,
10.3f, 0.0f, 11.4f, 9.45f, 11.4f};
dh::device_vector<bst_float> dlabels(hlabels);
auto segment_label_sorter = RankSegmentSorterTestImpl<float>(
{0, 4, 7, 12}, // Groups
hlabels,
{4.4f, 3.1f, 2.3f, 0.0f, // Expected sorted labels
5.01f, 0.0f, 0.0f,
11.4f, 11.4f, 10.3f, 9.45f, 0.0f},
{3, 0, 2, 1, // Expected original positions
5, 4, 6,
9, 11, 7, 10, 8});
// Create MAP stats on the device first using the objective
std::vector<bst_float> hpreds{-9.78f, 24.367f, 0.908f, -11.47f,
-1.03f, -2.79f, -3.1f,
104.22f, 103.1f, -101.7f, 100.5f, 45.1f};
dh::device_vector<bst_float> dpreds(hpreds);
xgboost::obj::MAPLambdaWeightComputer map_lw_computer(dpreds.data().get(),
dlabels.data().get(),
*segment_label_sorter);
// Get the device MAP stats on host
thrust::host_vector<xgboost::obj::MAPLambdaWeightComputer::MAPStats> dmap_stats(
map_lw_computer.GetMapStats());
// Compute the MAP stats on host next to compare
thrust::host_vector<uint32_t> hgroups(segment_label_sorter->GetGroups());
for (auto i = 0; i < hgroups.size() - 1; ++i) {
auto gbegin = hgroups[i];
auto gend = hgroups[i + 1];
std::vector<xgboost::obj::ListEntry> lst_entry;
for (auto j = gbegin; j < gend; ++j) {
lst_entry.emplace_back(hpreds[j], hlabels[j], j);
}
std::stable_sort(lst_entry.begin(), lst_entry.end(), xgboost::obj::ListEntry::CmpPred);
// Compute the MAP stats with this list and compare with the ones computed on the device
std::vector<xgboost::obj::MAPLambdaWeightComputer::MAPStats> hmap_stats;
xgboost::obj::MAPLambdaWeightComputer::GetMAPStats(lst_entry, &hmap_stats);
for (auto j = gbegin; j < gend; ++j) {
EXPECT_EQ(dmap_stats[j].hits, hmap_stats[j - gbegin].hits);
EXPECT_NEAR(dmap_stats[j].ap_acc, hmap_stats[j - gbegin].ap_acc, 0.01f);
EXPECT_NEAR(dmap_stats[j].ap_acc_miss, hmap_stats[j - gbegin].ap_acc_miss, 0.01f);
EXPECT_NEAR(dmap_stats[j].ap_acc_add, hmap_stats[j - gbegin].ap_acc_add, 0.01f);
}
}
}
} // namespace xgboost