From 94dfb7d770ece68cd2c07689bcb3f9fbd6864e04 Mon Sep 17 00:00:00 2001 From: ShenLiang <2282912238@qq.com> Date: Sun, 26 Apr 2020 15:21:21 +0800 Subject: [PATCH] opt the postprocess, test=develop (#24155) --- paddle/fluid/framework/data_set.cc | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/paddle/fluid/framework/data_set.cc b/paddle/fluid/framework/data_set.cc index b0948c8df53..48cb4131d58 100644 --- a/paddle/fluid/framework/data_set.cc +++ b/paddle/fluid/framework/data_set.cc @@ -779,6 +779,9 @@ template class DatasetImpl; void MultiSlotDataset::PostprocessInstance() { // divide pv instance, and merge to input_channel_ if (enable_pv_merge_) { + auto fleet_ptr = FleetWrapper::GetInstance(); + std::shuffle(input_records_.begin(), input_records_.end(), + fleet_ptr->LocalRandomEngine()); input_channel_->Open(); input_channel_->Write(std::move(input_records_)); for (size_t i = 0; i < multi_pv_consume_.size(); ++i) { @@ -799,8 +802,8 @@ void MultiSlotDataset::PostprocessInstance() { multi_consume_channel_[i]->Clear(); } input_channel_->Close(); + this->LocalShuffle(); } - this->LocalShuffle(); } void MultiSlotDataset::SetCurrentPhase(int current_phase) { -- GitLab