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

Use matrix for gradient. (#9508)

Browse Source 
- Use the `linalg::Matrix` for storing gradients.
- New API for the custom objective.
- Custom objective for multi-class/multi-target is now required to return the correct shape.
- Custom objective for Python can accept arrays with any strides. (row-major, column-major)
This commit is contained in:
Jiaming Yuan
2023-08-24 05:29:52 +08:00
committed by GitHub
parent 6103dca0bb
commit 972730cde0
77 changed files with 1052 additions and 651 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
R-package/src/init.c
View File

@@ -16,7 +16,7 @@ Check these declarations against the C/Fortran source code.
*/
/* .Call calls */
extern SEXP XGBoosterBoostOneIter_R(SEXP, SEXP, SEXP, SEXP);
extern SEXP XGBoosterTrainOneIter_R(SEXP, SEXP, SEXP, SEXP, SEXP);
extern SEXP XGBoosterCreate_R(SEXP);
extern SEXP XGBoosterCreateInEmptyObj_R(SEXP, SEXP);
extern SEXP XGBoosterDumpModel_R(SEXP, SEXP, SEXP, SEXP);
@@ -53,7 +53,7 @@ extern SEXP XGBGetGlobalConfig_R(void);
extern SEXP XGBoosterFeatureScore_R(SEXP, SEXP);
static const R_CallMethodDef CallEntries[] = {
{"XGBoosterBoostOneIter_R", (DL_FUNC) &XGBoosterBoostOneIter_R, 4},
{"XGBoosterBoostOneIter_R", (DL_FUNC) &XGBoosterTrainOneIter_R, 5},
{"XGBoosterCreate_R", (DL_FUNC) &XGBoosterCreate_R, 1},
{"XGBoosterCreateInEmptyObj_R", (DL_FUNC) &XGBoosterCreateInEmptyObj_R, 2},
{"XGBoosterDumpModel_R", (DL_FUNC) &XGBoosterDumpModel_R, 4},

43
R-package/src/xgboost_R.cc
View File

@@ -48,13 +48,6 @@
using dmlc::BeginPtr;
xgboost::Context const *BoosterCtx(BoosterHandle handle) {
CHECK_HANDLE();
auto *learner = static_cast<xgboost::Learner *>(handle);
CHECK(learner);
return learner->Ctx();
}
xgboost::Context const *DMatrixCtx(DMatrixHandle handle) {
CHECK_HANDLE();
auto p_m = static_cast<std::shared_ptr<xgboost::DMatrix> *>(handle);
@@ -394,21 +387,25 @@ XGB_DLL SEXP XGBoosterUpdateOneIter_R(SEXP handle, SEXP iter, SEXP dtrain) {
return R_NilValue;
}
XGB_DLL SEXP XGBoosterBoostOneIter_R(SEXP handle, SEXP dtrain, SEXP grad, SEXP hess) {
XGB_DLL SEXP XGBoosterTrainOneIter_R(SEXP handle, SEXP dtrain, SEXP iter, SEXP grad, SEXP hess) {
R_API_BEGIN();
CHECK_EQ(length(grad), length(hess))
<< "gradient and hess must have same length";
int len = length(grad);
std::vector<float> tgrad(len), thess(len);
auto ctx = BoosterCtx(R_ExternalPtrAddr(handle));
xgboost::common::ParallelFor(len, ctx->Threads(), [&](xgboost::omp_ulong j) {
tgrad[j] = REAL(grad)[j];
thess[j] = REAL(hess)[j];
});
CHECK_CALL(XGBoosterBoostOneIter(R_ExternalPtrAddr(handle),
R_ExternalPtrAddr(dtrain),
BeginPtr(tgrad), BeginPtr(thess),
len));
CHECK_EQ(length(grad), length(hess)) << "gradient and hess must have same length";
SEXP gdim = getAttrib(grad, R_DimSymbol);
auto n_samples = static_cast<std::size_t>(INTEGER(gdim)[0]);
auto n_targets = static_cast<std::size_t>(INTEGER(gdim)[1]);
SEXP hdim = getAttrib(hess, R_DimSymbol);
CHECK_EQ(INTEGER(hdim)[0], n_samples) << "mismatched size between gradient and hessian";
CHECK_EQ(INTEGER(hdim)[1], n_targets) << "mismatched size between gradient and hessian";
double const *d_grad = REAL(grad);
double const *d_hess = REAL(hess);
auto ctx = xgboost::detail::BoosterCtx(R_ExternalPtrAddr(handle));
auto [s_grad, s_hess] =
xgboost::detail::MakeGradientInterface(ctx, d_grad, d_hess, n_samples, n_targets);
CHECK_CALL(XGBoosterTrainOneIter(R_ExternalPtrAddr(handle), R_ExternalPtrAddr(dtrain),
asInteger(iter), s_grad.c_str(), s_hess.c_str()));
R_API_END();
return R_NilValue;
}
@@ -460,7 +457,7 @@ XGB_DLL SEXP XGBoosterPredictFromDMatrix_R(SEXP handle, SEXP dmat, SEXP json_con
len *= out_shape[i];
}
r_out_result = PROTECT(allocVector(REALSXP, len));
auto ctx = BoosterCtx(R_ExternalPtrAddr(handle));
auto ctx = xgboost::detail::BoosterCtx(R_ExternalPtrAddr(handle));
xgboost::common::ParallelFor(len, ctx->Threads(), [&](xgboost::omp_ulong i) {
REAL(r_out_result)[i] = out_result[i];
});
@@ -669,7 +666,7 @@ XGB_DLL SEXP XGBoosterFeatureScore_R(SEXP handle, SEXP json_config) {
}
out_scores_sexp = PROTECT(allocVector(REALSXP, len));
auto ctx = BoosterCtx(R_ExternalPtrAddr(handle));
auto ctx = xgboost::detail::BoosterCtx(R_ExternalPtrAddr(handle));
xgboost::common::ParallelFor(len, ctx->Threads(), [&](xgboost::omp_ulong i) {
REAL(out_scores_sexp)[i] = out_scores[i];
});

3
R-package/src/xgboost_R.h
View File

@@ -161,12 +161,13 @@ XGB_DLL SEXP XGBoosterUpdateOneIter_R(SEXP ext, SEXP iter, SEXP dtrain);
* \brief update the model, by directly specify gradient and second order gradient,
* this can be used to replace UpdateOneIter, to support customized loss function
* \param handle handle
* \param iter The current training iteration.
* \param dtrain training data
* \param grad gradient statistics
* \param hess second order gradient statistics
* \return R_NilValue
*/
XGB_DLL SEXP XGBoosterBoostOneIter_R(SEXP handle, SEXP dtrain, SEXP grad, SEXP hess);
XGB_DLL SEXP XGBoosterTrainOneIter_R(SEXP handle, SEXP dtrain, SEXP iter, SEXP grad, SEXP hess);
/*!
* \brief get evaluation statistics for xgboost