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

Pass pointer to model parameters. (#5101)

Browse Source 
* Pass pointer to model parameters.

This PR de-duplicates most of the model parameters except the one in
`tree_model.h`.  One difficulty is `base_score` is a model property but can be
changed at runtime by objective function.  Hence when performing model IO, we
need to save the one provided by users, instead of the one transformed by
objective.  Here we created an immutable version of `LearnerModelParam` that
represents the value of model parameter after configuration.
This commit is contained in:
Jiaming Yuan
2019-12-10 12:11:22 +08:00
committed by GitHub
parent 979f74d51a
commit e089e16e3d
33 changed files with 623 additions and 404 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
src/linear/coordinate_common.h
View File

@@ -252,7 +252,7 @@ class CyclicFeatureSelector : public FeatureSelector {
int NextFeature(int iteration, const gbm::GBLinearModel &model,
int group_idx, const std::vector<GradientPair> &gpair,
DMatrix *p_fmat, float alpha, float lambda) override {
return iteration % model.param.num_feature;
return iteration % model.learner_model_param_->num_feature;
}
};
@@ -266,7 +266,7 @@ class ShuffleFeatureSelector : public FeatureSelector {
const std::vector<GradientPair> &gpair,
DMatrix *p_fmat, float alpha, float lambda, int param) override {
if (feat_index_.size() == 0) {
feat_index_.resize(model.param.num_feature);
feat_index_.resize(model.learner_model_param_->num_feature);
std::iota(feat_index_.begin(), feat_index_.end(), 0);
}
std::shuffle(feat_index_.begin(), feat_index_.end(), common::GlobalRandom());
@@ -275,7 +275,7 @@ class ShuffleFeatureSelector : public FeatureSelector {
int NextFeature(int iteration, const gbm::GBLinearModel &model,
int group_idx, const std::vector<GradientPair> &gpair,
DMatrix *p_fmat, float alpha, float lambda) override {
return feat_index_[iteration % model.param.num_feature];
return feat_index_[iteration % model.learner_model_param_->num_feature];
}
protected:
@@ -291,7 +291,7 @@ class RandomFeatureSelector : public FeatureSelector {
int NextFeature(int iteration, const gbm::GBLinearModel &model,
int group_idx, const std::vector<GradientPair> &gpair,
DMatrix *p_fmat, float alpha, float lambda) override {
return common::GlobalRandom()() % model.param.num_feature;
return common::GlobalRandom()() % model.learner_model_param_->num_feature;
}
};
@@ -310,11 +310,11 @@ class GreedyFeatureSelector : public FeatureSelector {
const std::vector<GradientPair> &gpair,
DMatrix *p_fmat, float alpha, float lambda, int param) override {
top_k_ = static_cast<bst_uint>(param);
const bst_uint ngroup = model.param.num_output_group;
const bst_uint ngroup = model.learner_model_param_->num_output_group;
if (param <= 0) top_k_ = std::numeric_limits<bst_uint>::max();
if (counter_.size() == 0) {
counter_.resize(ngroup);
gpair_sums_.resize(model.param.num_feature * ngroup);
gpair_sums_.resize(model.learner_model_param_->num_feature * ngroup);
}
for (bst_uint gid = 0u; gid < ngroup; ++gid) {
counter_[gid] = 0u;
@@ -327,10 +327,10 @@ class GreedyFeatureSelector : public FeatureSelector {
// k-th selected feature for a group
auto k = counter_[group_idx]++;
// stop after either reaching top-K or going through all the features in a group
if (k >= top_k_ || counter_[group_idx] == model.param.num_feature) return -1;
if (k >= top_k_ || counter_[group_idx] == model.learner_model_param_->num_feature) return -1;
const int ngroup = model.param.num_output_group;
const bst_omp_uint nfeat = model.param.num_feature;
const int ngroup = model.learner_model_param_->num_output_group;
const bst_omp_uint nfeat = model.learner_model_param_->num_feature;
// Calculate univariate gradient sums
std::fill(gpair_sums_.begin(), gpair_sums_.end(), std::make_pair(0., 0.));
for (const auto &batch : p_fmat->GetBatches<CSCPage>()) {
@@ -387,8 +387,8 @@ class ThriftyFeatureSelector : public FeatureSelector {
DMatrix *p_fmat, float alpha, float lambda, int param) override {
top_k_ = static_cast<bst_uint>(param);
if (param <= 0) top_k_ = std::numeric_limits<bst_uint>::max();
const bst_uint ngroup = model.param.num_output_group;
const bst_omp_uint nfeat = model.param.num_feature;
const bst_uint ngroup = model.learner_model_param_->num_output_group;
const bst_omp_uint nfeat = model.learner_model_param_->num_feature;
if (deltaw_.size() == 0) {
deltaw_.resize(nfeat * ngroup);
@@ -444,9 +444,9 @@ class ThriftyFeatureSelector : public FeatureSelector {
// k-th selected feature for a group
auto k = counter_[group_idx]++;
// stop after either reaching top-N or going through all the features in a group
if (k >= top_k_ || counter_[group_idx] == model.param.num_feature) return -1;
if (k >= top_k_ || counter_[group_idx] == model.learner_model_param_->num_feature) return -1;
// note that sorted_idx stores the "long" indices
const size_t grp_offset = group_idx * model.param.num_feature;
const size_t grp_offset = group_idx * model.learner_model_param_->num_feature;
return static_cast<int>(sorted_idx_[grp_offset + k] - grp_offset);
}

6
src/linear/updater_coordinate.cc
View File

@@ -35,7 +35,7 @@ class CoordinateUpdater : public LinearUpdater {
void Update(HostDeviceVector<GradientPair> *in_gpair, DMatrix *p_fmat,
gbm::GBLinearModel *model, double sum_instance_weight) override {
tparam_.DenormalizePenalties(sum_instance_weight);
const int ngroup = model->param.num_output_group;
const int ngroup = model->learner_model_param_->num_output_group;
// update bias
for (int group_idx = 0; group_idx < ngroup; ++group_idx) {
auto grad = GetBiasGradientParallel(group_idx, ngroup,
@@ -52,7 +52,7 @@ class CoordinateUpdater : public LinearUpdater {
tparam_.reg_lambda_denorm, cparam_.top_k);
// update weights
for (int group_idx = 0; group_idx < ngroup; ++group_idx) {
for (unsigned i = 0U; i < model->param.num_feature; i++) {
for (unsigned i = 0U; i < model->learner_model_param_->num_feature; i++) {
int fidx = selector_->NextFeature
(i, *model, group_idx, in_gpair->ConstHostVector(), p_fmat,
tparam_.reg_alpha_denorm, tparam_.reg_lambda_denorm);
@@ -65,7 +65,7 @@ class CoordinateUpdater : public LinearUpdater {
inline void UpdateFeature(int fidx, int group_idx, std::vector<GradientPair> *in_gpair,
DMatrix *p_fmat, gbm::GBLinearModel *model) {
const int ngroup = model->param.num_output_group;
const int ngroup = model->learner_model_param_->num_output_group;
bst_float &w = (*model)[fidx][group_idx];
auto gradient =
GetGradientParallel(group_idx, ngroup, fidx, *in_gpair, p_fmat);

25
src/linear/updater_gpu_coordinate.cu
View File

@@ -41,7 +41,7 @@ class GPUCoordinateUpdater : public LinearUpdater { // NOLINT
monitor_.Init("GPUCoordinateUpdater");
}
void LazyInitDevice(DMatrix *p_fmat, const gbm::GBLinearModelParam &model_param) {
void LazyInitDevice(DMatrix *p_fmat, const LearnerModelParam &model_param) {
if (learner_param_->gpu_id < 0) return;
num_row_ = static_cast<size_t>(p_fmat->Info().num_row_);
@@ -88,14 +88,14 @@ class GPUCoordinateUpdater : public LinearUpdater { // NOLINT
gbm::GBLinearModel *model, double sum_instance_weight) override {
tparam_.DenormalizePenalties(sum_instance_weight);
monitor_.Start("LazyInitDevice");
this->LazyInitDevice(p_fmat, model->param);
this->LazyInitDevice(p_fmat, *(model->learner_model_param_));
monitor_.Stop("LazyInitDevice");
monitor_.Start("UpdateGpair");
auto &in_gpair_host = in_gpair->ConstHostVector();
// Update gpair
if (learner_param_->gpu_id >= 0) {
this->UpdateGpair(in_gpair_host, model->param);
this->UpdateGpair(in_gpair_host);
}
monitor_.Stop("UpdateGpair");
@@ -107,8 +107,9 @@ class GPUCoordinateUpdater : public LinearUpdater { // NOLINT
tparam_.reg_alpha_denorm, tparam_.reg_lambda_denorm,
coord_param_.top_k);
monitor_.Start("UpdateFeature");
for (auto group_idx = 0; group_idx < model->param.num_output_group; ++group_idx) {
for (auto i = 0U; i < model->param.num_feature; i++) {
for (auto group_idx = 0; group_idx < model->learner_model_param_->num_output_group;
++group_idx) {
for (auto i = 0U; i < model->learner_model_param_->num_feature; i++) {
auto fidx = selector_->NextFeature(
i, *model, group_idx, in_gpair->ConstHostVector(), p_fmat,
tparam_.reg_alpha_denorm, tparam_.reg_lambda_denorm);
@@ -120,11 +121,12 @@ class GPUCoordinateUpdater : public LinearUpdater { // NOLINT
}
void UpdateBias(DMatrix *p_fmat, gbm::GBLinearModel *model) {
for (int group_idx = 0; group_idx < model->param.num_output_group; ++group_idx) {
for (int group_idx = 0; group_idx < model->learner_model_param_->num_output_group;
++group_idx) {
// Get gradient
auto grad = GradientPair(0, 0);
if (learner_param_->gpu_id >= 0) {
grad = GetBiasGradient(group_idx, model->param.num_output_group);
grad = GetBiasGradient(group_idx, model->learner_model_param_->num_output_group);
}
auto dbias = static_cast<float>(
tparam_.learning_rate *
@@ -133,7 +135,7 @@ class GPUCoordinateUpdater : public LinearUpdater { // NOLINT
// Update residual
if (learner_param_->gpu_id >= 0) {
UpdateBiasResidual(dbias, group_idx, model->param.num_output_group);
UpdateBiasResidual(dbias, group_idx, model->learner_model_param_->num_output_group);
}
}
}
@@ -145,7 +147,7 @@ class GPUCoordinateUpdater : public LinearUpdater { // NOLINT
// Get gradient
auto grad = GradientPair(0, 0);
if (learner_param_->gpu_id >= 0) {
grad = GetGradient(group_idx, model->param.num_output_group, fidx);
grad = GetGradient(group_idx, model->learner_model_param_->num_output_group, fidx);
}
auto dw = static_cast<float>(tparam_.learning_rate *
CoordinateDelta(grad.GetGrad(), grad.GetHess(),
@@ -154,7 +156,7 @@ class GPUCoordinateUpdater : public LinearUpdater { // NOLINT
w += dw;
if (learner_param_->gpu_id >= 0) {
UpdateResidual(dw, group_idx, model->param.num_output_group, fidx);
UpdateResidual(dw, group_idx, model->learner_model_param_->num_output_group, fidx);
}
}
@@ -217,8 +219,7 @@ class GPUCoordinateUpdater : public LinearUpdater { // NOLINT
return num_row_ == 0;
}
void UpdateGpair(const std::vector<GradientPair> &host_gpair,
const gbm::GBLinearModelParam &model_param) {
void UpdateGpair(const std::vector<GradientPair> &host_gpair) {
dh::safe_cuda(cudaMemcpyAsync(
gpair_.data(),
host_gpair.data(),

2
src/linear/updater_shotgun.cc
View File

@@ -27,7 +27,7 @@ class ShotgunUpdater : public LinearUpdater {
gbm::GBLinearModel *model, double sum_instance_weight) override {
auto &gpair = in_gpair->HostVector();
param_.DenormalizePenalties(sum_instance_weight);
const int ngroup = model->param.num_output_group;
const int ngroup = model->learner_model_param_->num_output_group;
// update bias
for (int gid = 0; gid < ngroup; ++gid) {