From 5f98d8003966ebda1dd144ce7aee8a996f608f62 Mon Sep 17 00:00:00 2001 From: wangguibao Date: Thu, 6 Dec 2018 20:49:43 +0800 Subject: [PATCH] AsyncExecutor bugfix: Tensor change to LoDTensor --- paddle/fluid/framework/data_feed.cc | 43 +++++++----------------- paddle/fluid/framework/data_feed.h | 31 +---------------- paddle/fluid/framework/data_feed_test.cc | 15 +++------ 3 files changed, 18 insertions(+), 71 deletions(-) diff --git a/paddle/fluid/framework/data_feed.cc b/paddle/fluid/framework/data_feed.cc index 291d8ffc3..9a4b115ae 100644 --- a/paddle/fluid/framework/data_feed.cc +++ b/paddle/fluid/framework/data_feed.cc @@ -33,11 +33,7 @@ void DataFeed::AddFeedVar(Variable* var, const std::string& name) { CheckInit(); for (size_t i = 0; i < use_slots_.size(); ++i) { if (name == use_slots_[i]) { - if (use_slots_is_dense_[i]) { - feed_vec_[i] = MixTensor(var->GetMutable()); - } else { - feed_vec_[i] = MixTensor(var->GetMutable()); - } + feed_vec_[i] = var->GetMutable(); } } } @@ -350,34 +346,21 @@ void MultiSlotDataFeed::PutToFeedVec( int total_instance = static_cast(offset.back()); if (type[0] == 'f') { // float const auto& feasign = ins_vec[i].GetFloatData(); - if (feed_vec_[i].IsDense()) { - int size_in_each_batch = total_instance / batch_size_; - float* tensor_ptr = feed_vec_[i].GetTensor()->mutable_data( - {batch_size_, size_in_each_batch}, platform::CPUPlace()); - memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float)); - } else { - float* tensor_ptr = feed_vec_[i].GetLoDTensor()->mutable_data( - {total_instance, 1}, platform::CPUPlace()); - memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float)); - LoD data_lod{offset}; - feed_vec_[i].GetLoDTensor()->set_lod(data_lod); - } + float* tensor_ptr = feed_vec_[i]->mutable_data( + {total_instance, 1}, platform::CPUPlace()); + memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float)); } else if (type[0] == 'u') { // uint64 // no uint64_t type in paddlepaddle const auto& feasign = ins_vec[i].GetUint64Data(); - if (feed_vec_[i].IsDense()) { - int size_in_each_batch = total_instance / batch_size_; - int64_t* tensor_ptr = feed_vec_[i].GetTensor()->mutable_data( - {batch_size_, size_in_each_batch}, platform::CPUPlace()); - memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t)); - } else { - int64_t* tensor_ptr = - feed_vec_[i].GetLoDTensor()->mutable_data( - {total_instance, 1}, platform::CPUPlace()); - memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t)); - LoD data_lod{offset}; - feed_vec_[i].GetLoDTensor()->set_lod(data_lod); - } + int64_t* tensor_ptr = feed_vec_[i]->mutable_data( + {total_instance, 1}, platform::CPUPlace()); + memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t)); + } + LoD data_lod{offset}; + feed_vec_[i]->set_lod(data_lod); + if (use_slots_is_dense_[i]) { + int dim = total_instance / batch_size_; + feed_vec_[i]->Resize({batch_size_, dim}); } } } diff --git a/paddle/fluid/framework/data_feed.h b/paddle/fluid/framework/data_feed.h index a7f8d1d31..7cc691970 100644 --- a/paddle/fluid/framework/data_feed.h +++ b/paddle/fluid/framework/data_feed.h @@ -30,35 +30,6 @@ limitations under the License. */ namespace paddle { namespace framework { -// Pack Tensor type and LoDTensor type into MixTensor type, in order -// to record either Tensor or LoDTensor information at the same time. -class MixTensor { - public: - MixTensor() {} - explicit MixTensor(LoDTensor* lodtensor) { - is_dense_ = false; - lodtensor_ = lodtensor; - } - explicit MixTensor(Tensor* tensor) { - is_dense_ = true; - tensor_ = tensor; - } - bool IsDense() { return is_dense_; } - LoDTensor* GetLoDTensor() { - PADDLE_ENFORCE(!is_dense_, "Let a dense var return a LoDTensor ptr."); - return lodtensor_; - } - Tensor* GetTensor() { - PADDLE_ENFORCE(is_dense_, "Let a sparse var return a Tensor ptr."); - return tensor_; - } - - private: - bool is_dense_; - LoDTensor* lodtensor_; - Tensor* tensor_; -}; - // DataFeed is the base virtual class for all ohther DataFeeds. // It is used to read files and parse the data for subsequent trainer. // Example: @@ -133,7 +104,7 @@ class DataFeed { use_slots_index_; // -1: not used; >=0: the index of use_slots_ // The data read by DataFeed will be stored here - std::vector feed_vec_; + std::vector feed_vec_; // the batch size defined by user int default_batch_size_; diff --git a/paddle/fluid/framework/data_feed_test.cc b/paddle/fluid/framework/data_feed_test.cc index 3974f8dba..b3e969871 100644 --- a/paddle/fluid/framework/data_feed_test.cc +++ b/paddle/fluid/framework/data_feed_test.cc @@ -152,19 +152,13 @@ void GetElemSetFromReader(std::vector* reader_elem_set, const auto& multi_slot_desc = data_feed_desc.multi_slot_desc(); std::map lodtensor_targets; - std::map tensor_targets; for (int i = 0; i < multi_slot_desc.slots_size(); ++i) { const auto& slot = multi_slot_desc.slots(i); if (slot.is_used()) { const auto& name = slot.name(); readers[idx]->AddFeedVar(scope->Var(name), name); - if (slot.is_dense()) { - tensor_targets[name] = - &scope->FindVar(name)->Get(); - } else { - lodtensor_targets[name] = - &scope->FindVar(name)->Get(); - } + lodtensor_targets[name] = + &scope->FindVar(name)->Get(); } } readers[idx]->Start(); @@ -175,8 +169,9 @@ void GetElemSetFromReader(std::vector* reader_elem_set, if (!slot.is_used()) { continue; } + const paddle::framework::LoDTensor* tens = + lodtensor_targets[slot.name()]; if (slot.is_dense()) { // dense branch - const paddle::framework::Tensor* tens = tensor_targets[slot.name()]; if (slot.type() == "uint64") { const int64_t* data = tens->data(); int batch_size = tens->dims()[0]; @@ -202,8 +197,6 @@ void GetElemSetFromReader(std::vector* reader_elem_set, PADDLE_THROW("Error type in proto file."); } } else { // sparse branch - const paddle::framework::LoDTensor* tens = - lodtensor_targets[slot.name()]; if (slot.type() == "uint64") { const int64_t* data = tens->data(); for (size_t i = 0; i < tens->NumElements(); ++i) { -- GitLab