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

Accept iterator in device dmatrix. (#5783)

Browse Source 
* Remove Device DMatrix.
This commit is contained in:
Jiaming Yuan
2020-07-07 21:44:48 +08:00
committed by GitHub
parent 048d969be4
commit a3ec964346
12 changed files with 495 additions and 343 deletions
Download Patch File Download Diff File Expand all files Collapse all files

149
tests/cpp/data/test_device_dmatrix.cu
View File

@@ -1,149 +0,0 @@
// Copyright (c) 2019 by Contributors
#include <gtest/gtest.h>
#include <xgboost/data.h>
#include "../../../src/data/adapter.h"
#include "../../../src/data/ellpack_page.cuh"
#include "../../../src/data/device_dmatrix.h"
#include "../helpers.h"
#include <thrust/device_vector.h>
#include "../../../src/data/device_adapter.cuh"
#include "../../../src/gbm/gbtree_model.h"
#include "../common/test_hist_util.h"
#include "../../../src/common/compressed_iterator.h"
#include "../../../src/common/math.h"
#include "test_array_interface.h"
using namespace xgboost; // NOLINT
TEST(DeviceDMatrix, RowMajor) {
int num_rows = 1000;
int num_columns = 50;
auto x = common::GenerateRandom(num_rows, num_columns);
auto x_device = thrust::device_vector<float>(x);
auto adapter = common::AdapterFromData(x_device, num_rows, num_columns);
data::DeviceDMatrix dmat(&adapter,
std::numeric_limits<float>::quiet_NaN(), 1, 256);
auto &batch = *dmat.GetBatches<EllpackPage>({0, 256, 0}).begin();
auto impl = batch.Impl();
common::CompressedIterator<uint32_t> iterator(
impl->gidx_buffer.HostVector().data(), impl->NumSymbols());
for(auto i = 0ull; i < x.size(); i++)
{
int column_idx = i % num_columns;
EXPECT_EQ(impl->Cuts().SearchBin(x[i], column_idx), iterator[i]);
}
EXPECT_EQ(dmat.Info().num_col_, num_columns);
EXPECT_EQ(dmat.Info().num_row_, num_rows);
EXPECT_EQ(dmat.Info().num_nonzero_, num_rows * num_columns);
}
TEST(DeviceDMatrix, RowMajorMissing) {
const float kMissing = std::numeric_limits<float>::quiet_NaN();
int num_rows = 10;
int num_columns = 2;
auto x = common::GenerateRandom(num_rows, num_columns);
x[1] = kMissing;
x[5] = kMissing;
x[6] = kMissing;
auto x_device = thrust::device_vector<float>(x);
auto adapter = common::AdapterFromData(x_device, num_rows, num_columns);
data::DeviceDMatrix dmat(&adapter, kMissing, 1, 256);
auto &batch = *dmat.GetBatches<EllpackPage>({0, 256, 0}).begin();
auto impl = batch.Impl();
common::CompressedIterator<uint32_t> iterator(
impl->gidx_buffer.HostVector().data(), impl->NumSymbols());
EXPECT_EQ(iterator[1], impl->GetDeviceAccessor(0).NullValue());
EXPECT_EQ(iterator[5], impl->GetDeviceAccessor(0).NullValue());
// null values get placed after valid values in a row
EXPECT_EQ(iterator[7], impl->GetDeviceAccessor(0).NullValue());
EXPECT_EQ(dmat.Info().num_col_, num_columns);
EXPECT_EQ(dmat.Info().num_row_, num_rows);
EXPECT_EQ(dmat.Info().num_nonzero_, num_rows*num_columns-3);
}
TEST(DeviceDMatrix, ColumnMajor) {
constexpr size_t kRows{100};
std::vector<Json> columns;
thrust::device_vector<double> d_data_0(kRows);
thrust::device_vector<uint32_t> d_data_1(kRows);
columns.emplace_back(GenerateDenseColumn<double>("<f8", kRows, &d_data_0));
columns.emplace_back(GenerateDenseColumn<uint32_t>("<u4", kRows, &d_data_1));
Json column_arr{columns};
std::string str;
Json::Dump(column_arr, &str);
data::CudfAdapter adapter(str);
data::DeviceDMatrix dmat(&adapter, std::numeric_limits<float>::quiet_NaN(),
-1, 256);
auto &batch = *dmat.GetBatches<EllpackPage>({0, 256, 0}).begin();
auto impl = batch.Impl();
common::CompressedIterator<uint32_t> iterator(
impl->gidx_buffer.HostVector().data(), impl->NumSymbols());
for (auto i = 0ull; i < kRows; i++) {
for (auto j = 0ull; j < columns.size(); j++) {
if (j == 0) {
EXPECT_EQ(iterator[i * 2 + j], impl->Cuts().SearchBin(d_data_0[i], j));
} else {
EXPECT_EQ(iterator[i * 2 + j], impl->Cuts().SearchBin(d_data_1[i], j));
}
}
}
EXPECT_EQ(dmat.Info().num_col_, 2);
EXPECT_EQ(dmat.Info().num_row_, kRows);
EXPECT_EQ(dmat.Info().num_nonzero_, kRows*2);
}
// Test equivalence with simple DMatrix
TEST(DeviceDMatrix, Equivalent) {
int bin_sizes[] = {2, 16, 256, 512};
int sizes[] = {100, 1000, 1500};
int num_columns = 5;
for (auto num_rows : sizes) {
auto x = common::GenerateRandom(num_rows, num_columns);
for (auto num_bins : bin_sizes) {
auto dmat = common::GetDMatrixFromData(x, num_rows, num_columns);
auto x_device = thrust::device_vector<float>(x);
auto adapter = common::AdapterFromData(x_device, num_rows, num_columns);
data::DeviceDMatrix device_dmat(
&adapter, std::numeric_limits<float>::quiet_NaN(), 1, num_bins);
const auto &batch = *dmat->GetBatches<EllpackPage>({0, num_bins}).begin();
const auto &device_dmat_batch =
*device_dmat.GetBatches<EllpackPage>({0, num_bins}).begin();
ASSERT_EQ(batch.Impl()->Cuts().Values(), device_dmat_batch.Impl()->Cuts().Values());
ASSERT_EQ(batch.Impl()->gidx_buffer.HostVector(),
device_dmat_batch.Impl()->gidx_buffer.HostVector());
}
}
}
TEST(DeviceDMatrix, IsDense) {
int num_bins = 16;
auto test = [num_bins] (float sparsity) {
HostDeviceVector<float> data;
std::string interface_str = RandomDataGenerator{10, 10, sparsity}
.Device(0).GenerateArrayInterface(&data);
data::CupyAdapter x{interface_str};
std::unique_ptr<data::DeviceDMatrix> device_dmat{ new data::DeviceDMatrix(
&x, std::numeric_limits<float>::quiet_NaN(), 1, num_bins) };
if (sparsity == 0.0) {
ASSERT_TRUE(device_dmat->IsDense()) << sparsity;
} else {
ASSERT_FALSE(device_dmat->IsDense());
}
};
test(0.0);
test(0.1);
}

1
tests/cpp/predictor/test_predictor.cc
View File

@@ -54,6 +54,7 @@ void TestTrainingPrediction(size_t rows, size_t bins,
learner->SetParam("objective", "multi:softprob");
learner->SetParam("num_feature", std::to_string(kCols));
learner->SetParam("num_class", std::to_string(kClasses));
learner->SetParam("max_bin", std::to_string(bins));
learner->Configure();
for (size_t i = 0; i < kIters; ++i) {

80
tests/python-gpu/test_from_cudf.py
View File

@@ -170,3 +170,83 @@ Arrow specification.'''
@pytest.mark.skipif(**tm.no_cudf())
def test_cudf_metainfo_device_dmatrix(self):
_test_cudf_metainfo(xgb.DeviceQuantileDMatrix)
class IterForDMatrixTest(xgb.core.DataIter):
'''A data iterator for XGBoost DMatrix.
`reset` and `next` are required for any data iterator, other functions here
are utilites for demonstration's purpose.
'''
ROWS_PER_BATCH = 100 # data is splited by rows
BATCHES = 16
def __init__(self):
'''Generate some random data for demostration.
Actual data can be anything that is currently supported by XGBoost.
'''
import cudf
self.rows = self.ROWS_PER_BATCH
rng = np.random.RandomState(1994)
self._data = [
cudf.DataFrame(
{'a': rng.randn(self.ROWS_PER_BATCH),
'b': rng.randn(self.ROWS_PER_BATCH)})] * self.BATCHES
self._labels = [rng.randn(self.rows)] * self.BATCHES
self.it = 0 # set iterator to 0
super().__init__()
def as_array(self):
import cudf
return cudf.concat(self._data)
def as_array_labels(self):
return np.concatenate(self._labels)
def data(self):
'''Utility function for obtaining current batch of data.'''
return self._data[self.it]
def labels(self):
'''Utility function for obtaining current batch of label.'''
return self._labels[self.it]
def reset(self):
'''Reset the iterator'''
self.it = 0
def next(self, input_data):
'''Yield next batch of data'''
if self.it == len(self._data):
# Return 0 when there's no more batch.
return 0
input_data(data=self.data(), label=self.labels())
self.it += 1
return 1
@pytest.mark.skipif(**tm.no_cudf())
def test_from_cudf_iter():
rounds = 100
it = IterForDMatrixTest()
# Use iterator
m_it = xgb.DeviceQuantileDMatrix(it)
reg_with_it = xgb.train({'tree_method': 'gpu_hist'}, m_it,
num_boost_round=rounds)
predict_with_it = reg_with_it.predict(m_it)
# Without using iterator
m = xgb.DMatrix(it.as_array(), it.as_array_labels())
assert m_it.num_col() == m.num_col()
assert m_it.num_row() == m.num_row()
reg = xgb.train({'tree_method': 'gpu_hist'}, m,
num_boost_round=rounds)
predict = reg.predict(m)
np.testing.assert_allclose(predict_with_it, predict)

11
tests/python-gpu/test_gpu_demos.py Normal file
View File

@@ -0,0 +1,11 @@
import os
import subprocess
import sys
sys.path.append("tests/python")
import test_demos as td # noqa
def test_data_iterator():
script = os.path.join(td.PYTHON_DEMO_DIR, 'data_iterator.py')
cmd = ['python', script]
subprocess.check_call(cmd)