Skip to content

[CPU-PSLIB] Add config for scale_sparse_grad in config_fleet.py,test=develop #34893

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Aug 16, 2021
Merged

[CPU-PSLIB] Add config for scale_sparse_grad in config_fleet.py,test=develop #34893

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
ef94264
[CPU-PSLIB] Add config for scale_sparse_grad in config_fleet.py,test=…
WorgenZhang Aug 13, 2021
3518522
[CPU-PSLIB] Modify scale_sparse_grad to scale_sparse_gradient_with_ba…
WorgenZhang Aug 16, 2021
9dfa28e
[CPU-PSLIB] Modify scale_sparse_grad to scale_sparse_gradient_with_ba…
WorgenZhang Aug 16, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions paddle/fluid/framework/device_worker.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -212,6 +212,7 @@ class DeviceWorker {
FetchConfig fetch_config_;
bool use_cvm_;
bool no_cvm_;
bool scale_sparse_grad_;
Copy link
Contributor

@fuyinno4 fuyinno4 Aug 16, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

scale_sparse_gradient_with_batch_size

Copy link
Contributor Author

@WorgenZhang WorgenZhang Aug 16, 2021

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

have changed scale_sparse_grad to scale_sparse_gradient_with_batch_size

TrainerDesc trainer_desc_;

// dump params or grads for debug
Expand Down
5 changes: 3 additions & 2 deletions paddle/fluid/framework/downpour_worker.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -89,6 +89,7 @@ void DownpourWorker::Initialize(const TrainerDesc& desc) {
use_cvm_ = desc.use_cvm();
// for sparse value accessor, embedding only
no_cvm_ = desc.no_cvm();
scale_sparse_grad_ = desc.scale_sparse_grad();
scale_datanorm_ = desc.scale_datanorm();
dump_slot_ = desc.dump_slot();
adjust_ins_weight_config_ = desc.adjust_ins_weight_config();
Expand Down Expand Up @@ -591,7 +592,7 @@ void DownpourWorker::TrainFilesWithProfiler() {
*thread_scope_, tid, features_[tid], feature_labels_[tid],
sparse_key_names_[tid], sparse_grad_names_[tid], table.emb_dim(),
&feature_grads_[tid], &push_sparse_status_, cur_batch, use_cvm_,
dump_slot_, &sparse_push_keys_[tid], no_cvm_);
dump_slot_, &sparse_push_keys_[tid], no_cvm_, scale_sparse_grad_);
timeline.Pause();
push_sparse_time += timeline.ElapsedSec();
total_time += timeline.ElapsedSec();
Expand Down Expand Up @@ -866,7 +867,7 @@ void DownpourWorker::TrainFiles() {
*thread_scope_, tid, features_[tid], feature_labels_[tid],
sparse_key_names_[tid], sparse_grad_names_[tid], table.emb_dim(),
&feature_grads_[tid], &push_sparse_status_, cur_batch, use_cvm_,
dump_slot_, &sparse_push_keys_[tid], no_cvm_);
dump_slot_, &sparse_push_keys_[tid], no_cvm_, scale_sparse_grad_);
}
}

Expand Down
3 changes: 2 additions & 1 deletion paddle/fluid/framework/downpour_worker_opt.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -450,11 +450,12 @@ void DownpourWorkerOpt::TrainFiles() {
break;
}
}
bool scale_sparse_grad_ = true;
fleet_ptr_->PushSparseVarsWithLabelAsync(
*thread_scope_, tid, features_[tid], feature_labels_[tid],
sparse_key_names_[tid], sparse_grad_names_[tid], table.emb_dim(),
&feature_grads_[tid], &push_sparse_status_, cur_batch, use_cvm_,
dump_slot_, &sparse_push_keys_[tid], no_cvm_);
dump_slot_, &sparse_push_keys_[tid], no_cvm_, scale_sparse_grad_);
}
}

Expand Down
5 changes: 3 additions & 2 deletions paddle/fluid/framework/fleet/fleet_wrapper.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -870,7 +870,8 @@ void FleetWrapper::PushSparseVarsWithLabelAsync(
std::vector<std::vector<float>>* push_values,
std::vector<::std::future<int32_t>>* push_sparse_status,
const int batch_size, const bool use_cvm, const bool dump_slot,
std::vector<uint64_t>* sparse_push_keys, const bool no_cvm) {
std::vector<uint64_t>* sparse_push_keys, const bool no_cvm,
const bool scale_sparse_grad) {
#ifdef PADDLE_WITH_PSLIB
int offset = 2;
int slot_offset = 0;
Expand Down Expand Up @@ -939,7 +940,7 @@ void FleetWrapper::PushSparseVarsWithLabelAsync(
}
float* g = g_tensor->data<float>();

if (scale_sparse_gradient_with_batch_size_ && grad_dim > 0) {
if (scale_sparse_grad && grad_dim > 0) {
int dim = emb_dim;
Eigen::Map<
Eigen::Matrix<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>>
Expand Down
3 changes: 2 additions & 1 deletion paddle/fluid/framework/fleet/fleet_wrapper.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -209,7 +209,8 @@ class FleetWrapper {
std::vector<std::vector<float>>* push_values,
std::vector<::std::future<int32_t>>* push_sparse_status,
const int batch_size, const bool use_cvm, const bool dump_slot,
std::vector<uint64_t>* sparse_push_keys, const bool no_cvm);
std::vector<uint64_t>* sparse_push_keys, const bool no_cvm,
const bool scale_sparse_grad);

// Push sparse variables to server in async mode
void PushSparseFromTensorWithLabelAsync(
Expand Down
1 change: 1 addition & 0 deletions paddle/fluid/framework/trainer_desc.proto
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,6 +61,7 @@ message TrainerDesc {

optional bool use_ps_gpu = 32 [ default = false ];
optional string user_define_dump_filename = 33;
optional bool scale_sparse_grad = 34 [ default = true ];

// device worker parameters
optional HogwildWorkerParameter hogwild_param = 101;
Expand Down
1 change: 1 addition & 0 deletions python/paddle/fluid/incubate/fleet/parameter_server/pslib/optimizer_factory.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -825,6 +825,7 @@ def _minimize(self,
opt_info["worker_skipped_ops"] = worker_skipped_ops
opt_info["use_cvm"] = strategy.get("use_cvm", False)
opt_info["no_cvm"] = strategy.get("no_cvm", False)
opt_info["scale_sparse_grad"] = strategy.get("scale_sparse_grad", True)
opt_info["worker_class"] = strategy.get("worker_class",
"DownpourWorker")
opt_info["stat_var_names"] = strategy.get("stat_var_names", [])
Expand Down
3 changes: 3 additions & 0 deletions python/paddle/fluid/trainer_desc.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -124,6 +124,9 @@ def _set_use_cvm(self, use_cvm=False):
def _set_no_cvm(self, no_cvm=False):
self.proto_desc.no_cvm = no_cvm

def _set_scale_sparse_grad(self, scale_sparse_grad=True):
self.proto_desc.scale_sparse_grad = scale_sparse_grad

def _set_scale_datanorm(self, scale_datanorm=-1):
self.proto_desc.scale_datanorm = scale_datanorm

Expand Down
3 changes: 3 additions & 0 deletions python/paddle/fluid/trainer_factory.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -95,6 +95,9 @@ def _create_trainer(self, opt_info=None):
trainer._set_use_cvm(opt_info["use_cvm"])
if opt_info.get("no_cvm") is not None:
trainer._set_no_cvm(opt_info["no_cvm"])
if opt_info.get("scale_sparse_grad") is not None:
trainer._set_scale_sparse_grad(opt_info[
"scale_sparse_grad"])
if opt_info.get("scale_datanorm") is not None:
trainer._set_scale_datanorm(opt_info["scale_datanorm"])
if opt_info.get("adjust_ins_weight") is not None:
Expand Down