Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fpn_ops

Author: jerrywgz

paddle/fluid/API.spec

@@ -128,6 +128,7 @@ paddle.fluid.layers.row_conv (ArgSpec(args=['input', 'future_context_size', 'par
 paddle.fluid.layers.multiplex (ArgSpec(args=['inputs', 'index'], varargs=None, keywords=None, defaults=None), ('document', '013795af319e2e86d3506741941078ee'))
 paddle.fluid.layers.layer_norm (ArgSpec(args=['input', 'scale', 'shift', 'begin_norm_axis', 'epsilon', 'param_attr', 'bias_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(True, True, 1, 1e-05, None, None, None, None)), ('document', 'de6a906950bae9f3c245cb744d22b94e'))
 paddle.fluid.layers.group_norm (ArgSpec(args=['input', 'groups', 'epsilon', 'param_attr', 'bias_attr', 'act', 'data_layout', 'name'], varargs=None, keywords=None, defaults=(1e-05, None, None, None, 'NCHW', None)), ('document', '419c3a24a83cc89219a029cf4092788b'))
+paddle.fluid.layers.spectral_norm (ArgSpec(args=['weight', 'dim', 'power_iters', 'eps', 'name'], varargs=None, keywords=None, defaults=(0, 1, 1e-12, None)), ('document', '3f536aafba30d793287b52d231baff1b'))
 paddle.fluid.layers.softmax_with_cross_entropy (ArgSpec(args=['logits', 'label', 'soft_label', 'ignore_index', 'numeric_stable_mode', 'return_softmax'], varargs=None, keywords=None, defaults=(False, -100, True, False)), ('document', 'bce1b75e3d95b75cacd1099655cbb3c3'))
 paddle.fluid.layers.smooth_l1 (ArgSpec(args=['x', 'y', 'inside_weight', 'outside_weight', 'sigma'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', 'c6b175d253c55baf4b9c0eca9b1dda88'))
 paddle.fluid.layers.one_hot (ArgSpec(args=['input', 'depth'], varargs=None, keywords=None, defaults=None), ('document', '6148b6a555cbfb62fdcd030d8982c18c'))
@@ -261,7 +262,7 @@ paddle.fluid.layers.Switch.default (ArgSpec(args=['self'], varargs=None, keyword
 paddle.fluid.layers.increment (ArgSpec(args=['x', 'value', 'in_place'], varargs=None, keywords=None, defaults=(1.0, True)), ('document', '73bb96ec4783ec1a11e760e8851b0e77'))
 paddle.fluid.layers.array_write (ArgSpec(args=['x', 'i', 'array'], varargs=None, keywords=None, defaults=(None,)), ('document', '40b6d15f4c86b2b09df340d7778ad713'))
 paddle.fluid.layers.create_array (ArgSpec(args=['dtype'], varargs=None, keywords=None, defaults=None), ('document', '2d4f20087080ba5105b55205ad5c5b6a'))
-paddle.fluid.layers.less_than (ArgSpec(args=['x', 'y', 'force_cpu', 'cond'], varargs=None, keywords='ignored', defaults=(None, None)), ('document', '067bbc799c66289ca8b8924c26b6673f'))
+paddle.fluid.layers.less_than (ArgSpec(args=['x', 'y', 'force_cpu', 'cond'], varargs=None, keywords=None, defaults=(None, None)), ('document', '067bbc799c66289ca8b8924c26b6673f'))
 paddle.fluid.layers.equal (ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords=None, defaults=(None,)), ('document', '80c29b1dc64718f0116de90d1ac88a77'))
 paddle.fluid.layers.array_read (ArgSpec(args=['array', 'i'], varargs=None, keywords=None, defaults=None), ('document', '0275133f1dde2aed528b4d3230edf823'))
 paddle.fluid.layers.array_length (ArgSpec(args=['array'], varargs=None, keywords=None, defaults=None), ('document', 'ffb8b9578ec66db565b223d313aa82a2'))
@@ -286,7 +287,7 @@ paddle.fluid.layers.StaticRNN.step_output (ArgSpec(args=['self', 'o'], varargs=N
 paddle.fluid.layers.StaticRNN.update_memory (ArgSpec(args=['self', 'mem', 'var'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.layers.reorder_lod_tensor_by_rank (ArgSpec(args=['x', 'rank_table'], varargs=None, keywords=None, defaults=None), ('document', '3545f529ef04e8f6ecb76b47fa3df01a'))
 paddle.fluid.layers.Print (ArgSpec(args=['input', 'first_n', 'message', 'summarize', 'print_tensor_name', 'print_tensor_type', 'print_tensor_shape', 'print_tensor_lod', 'print_phase'], varargs=None, keywords=None, defaults=(-1, None, -1, True, True, True, True, 'both')), ('document', '5fef91b0e21c93610785f2b1f7161732'))
-paddle.fluid.layers.is_empty (ArgSpec(args=['x', 'cond'], varargs=None, keywords='ignored', defaults=(None,)), ('document', 'bbe578dbb49ad13e15b014e98c22b519'))
+paddle.fluid.layers.is_empty (ArgSpec(args=['x', 'cond'], varargs=None, keywords=None, defaults=(None,)), ('document', 'bbe578dbb49ad13e15b014e98c22b519'))
 paddle.fluid.layers.sigmoid (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '29a25ba78de79152076cacfc5443137d'))
 paddle.fluid.layers.logsigmoid (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '81ccb7acafd06c7728e11581f5d342e3'))
 paddle.fluid.layers.exp (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'e6b3e769413d96aab4176f96db25984b'))
@@ -328,7 +329,8 @@ paddle.fluid.layers.polygon_box_transform (ArgSpec(args=['input', 'name'], varar
 paddle.fluid.layers.yolov3_loss (ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample_ratio', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '991e934c3e09abf0edec7c9c978b4691'))
 paddle.fluid.layers.box_clip (ArgSpec(args=['input', 'im_info', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '397e9e02b451d99c56e20f268fa03f2e'))
 paddle.fluid.layers.multiclass_nms (ArgSpec(args=['bboxes', 'scores', 'score_threshold', 'nms_top_k', 'keep_top_k', 'nms_threshold', 'normalized', 'nms_eta', 'background_label', 'name'], varargs=None, keywords=None, defaults=(0.3, True, 1.0, 0, None)), ('document', 'ca7d1107b6c5d2d6d8221039a220fde0'))
-paddle.fluid.layers.distribute_fpn_proposals (ArgSpec(args=['fpn_rois', 'min_level', 'max_level', 'refer_level', 'refer_scale', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'fdffe52577f7e74c090b030867fefc11'))
+paddle.fluid.layers.distribute_fpn_proposals (ArgSpec(args=['fpn_rois', 'min_level', 'max_level', 'refer_level', 'refer_scale', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '7bb011ec26bace2bc23235aa4a17647d'))
+paddle.fluid.layers.box_decoder_and_assign (ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'box_score', 'box_clip', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '005a5ae47d6c8fff721931d69d072b9f'))
 paddle.fluid.layers.accuracy (ArgSpec(args=['input', 'label', 'k', 'correct', 'total'], varargs=None, keywords=None, defaults=(1, None, None)), ('document', '9808534c12c5e739a10f73ebb0b4eafd'))
 paddle.fluid.layers.auc (ArgSpec(args=['input', 'label', 'curve', 'num_thresholds', 'topk', 'slide_steps'], varargs=None, keywords=None, defaults=('ROC', 4095, 1, 1)), ('document', 'e0e95334fce92d16c2d9db6e7caffc47'))
 paddle.fluid.layers.exponential_decay (ArgSpec(args=['learning_rate', 'decay_steps', 'decay_rate', 'staircase'], varargs=None, keywords=None, defaults=(False,)), ('document', '98a5050bee8522fcea81aa795adaba51'))

paddle/fluid/framework/details/build_strategy.h

@@ -14,6 +14,7 @@
 
 #pragma once
 
+#include <memory>
 #include <string>
 #include <vector>
 
@@ -76,11 +77,11 @@ struct BuildStrategy {
 
   bool fuse_relu_depthwise_conv_{false};
 
-  bool memory_optimize_{false};
+  bool memory_optimize_{true};
   // TODO(dzhwinter):
   // make enable_inplace, memory_optimize_
   // memory_early_delete_ true by default
-  bool enable_inplace_{false};
+  bool enable_inplace_{true};
 
   bool enable_sequential_execution_{false};
 

paddle/fluid/framework/details/memory_optimize_helper.cc

@@ -20,6 +20,9 @@
 #include <numeric>
 #include <sstream>
 #include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/var_desc.h"
 #include "paddle/fluid/platform/cpu_info.h"
 
@@ -302,7 +305,10 @@ std::string OrderedSet::ToString() const {
 
 bool NodeCanReused(ir::Node* node) {
   // valid the node is a var node
-  if (node == nullptr || !node->IsVar() || node->IsCtrlVar()) return false;
+  // vars can be @EMPTY@, @LR_DECAY_REUSE_ID@. For example, while_grad
+  if (node == nullptr || !node->IsVar() || node->IsCtrlVar() ||
+      node->Name() == kEmptyVarName)
+    return false;
 
   bool flag = true;
   // op output force generated in cpu, can not be reused.
@@ -348,10 +354,6 @@ bool NodeCanReused(const VarDesc& node) {
   if (shape.empty() || size < MinChunkSize()) {
     return false;
   }
-  // vars can be @EMPTY@, @LR_DECAY_REUSE_ID@. For example, while_grad
-  std::string name = node.Name();
-  if (!name.empty() && name[0] == '@' && name[name.size() - 1] == '@')
-    return false;
   return true;
 }
 

paddle/fluid/framework/operator.cc

@@ -467,12 +467,6 @@ const Variable* ExecutionContext::InputVar(const std::string& name) const {
   return it->second.empty() ? nullptr : it->second[0];
 }
 
-const Variable* ExecutionContext::LegacyInputVar(
-    const std::string& name) const {
-  auto ipt = op_.Input(name);
-  return ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
-}
-
 Variable* ExecutionContext::OutputVar(const std::string& name) const {
   auto it = ctx_.outputs.find(name);
   if (it == ctx_.outputs.end()) return nullptr;
@@ -483,22 +477,11 @@ Variable* ExecutionContext::OutputVar(const std::string& name) const {
   return it->second.empty() ? nullptr : it->second[0];
 }
 
-Variable* ExecutionContext::LegacyOutputVar(const std::string& name) const {
-  auto opt = op_.Output(name);
-  return opt == kEmptyVarName ? nullptr : scope_.FindVar(opt);
-}
-
 template <>
 const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const {
   return Input<LoDTensor>(name);
 }
 
-template <>
-const Tensor* ExecutionContext::LegacyInput<Tensor>(
-    const std::string& name) const {
-  return LegacyInput<LoDTensor>(name);
-}
-
 template <>
 const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
     const std::string& name) const {
@@ -521,35 +504,11 @@ const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
   return res;
 }
 
-template <>
-const std::vector<const Tensor*> ExecutionContext::LegacyMultiInput<Tensor>(
-    const std::string& name) const {
-  auto names = op().Inputs(name);
-  std::vector<const Tensor*> res;
-  res.reserve(names.size());
-  std::transform(names.begin(), names.end(), std::back_inserter(res),
-                 [&](const std::string& sub_name) -> const Tensor* {
-                   auto var = scope_.FindVar(sub_name);
-                   if (var == nullptr) return nullptr;
-                   PADDLE_ENFORCE(
-                       var->IsType<LoDTensor>(),
-                       "%s should be LoDTensor, but the received type is %s",
-                       sub_name, ToTypeName(var->Type()));
-                   return &(var->Get<LoDTensor>());
-                 });
-  return res;
-}
-
 template <>
 Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const {
   return Output<LoDTensor>(name);
 }
 
-template <>
-Tensor* ExecutionContext::LegacyOutput<Tensor>(const std::string& name) const {
-  return LegacyOutput<LoDTensor>(name);
-}
-
 template <>
 std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
     const std::string& name) const {

paddle/fluid/framework/operator.h

@@ -16,9 +16,11 @@ limitations under the License. */
 
 #include <algorithm>
 #include <atomic>
+#include <memory>
 #include <string>
 #include <tuple>
 #include <unordered_map>
+#include <utility>
 #include <vector>
 
 #include "glog/logging.h"  // For VLOG
@@ -253,31 +255,6 @@ class ExecutionContext {
     return it->second;
   }
 
-  const std::vector<Variable*> LegacyMultiInputVar(
-      const std::string& name) const {
-    auto names = op_.Inputs(name);
-    std::vector<Variable*> res;
-    res.reserve(names.size());
-    std::transform(names.begin(), names.end(), std::back_inserter(res),
-                   [this](const std::string& name) {
-                     return name == kEmptyVarName ? nullptr
-                                                  : scope_.FindVar(name);
-                   });
-    return res;
-  }
-
-  std::vector<Variable*> LegacyMultiOutputVar(const std::string& name) const {
-    auto names = op_.Outputs(name);
-    std::vector<Variable*> res;
-    res.reserve(names.size());
-    std::transform(names.begin(), names.end(), std::back_inserter(res),
-                   [this](const std::string& name) {
-                     return name == kEmptyVarName ? nullptr
-                                                  : scope_.FindVar(name);
-                   });
-    return res;
-  }
-
   template <typename T>
   const T* Input(const std::string& name) const {
     auto* var = InputVar(name);
@@ -290,22 +267,6 @@ class ExecutionContext {
     return var == nullptr ? nullptr : var->GetMutable<T>();
   }
 
-  template <typename T>
-  const T* LegacyInput(const std::string& name) const {
-    auto* var = LegacyInputVar(name);
-    return var == nullptr ? nullptr : &var->Get<T>();
-  }
-
-  template <typename T>
-  T* LegacyOutput(const std::string& name) const {
-    auto var = LegacyOutputVar(name);
-    return var == nullptr ? nullptr : var->GetMutable<T>();
-  }
-
-  const Variable* LegacyInputVar(const std::string& name) const;
-
-  Variable* LegacyOutputVar(const std::string& name) const;
-
   template <typename T>
   const std::vector<const T*> MultiInput(const std::string& name) const {
     auto it = ctx_.inputs.find(name);
@@ -338,32 +299,6 @@ class ExecutionContext {
     return res;
   }
 
-  template <typename T>
-  const std::vector<const T*> LegacyMultiInput(const std::string& name) const {
-    auto names = op_.Inputs(name);
-    std::vector<const T*> res;
-    res.reserve(names.size());
-    std::transform(names.begin(), names.end(), std::back_inserter(res),
-                   [&](const std::string& sub_name) -> const T* {
-                     auto var = scope_.FindVar(sub_name);
-                     return var == nullptr ? nullptr : &var->Get<T>();
-                   });
-    return res;
-  }
-
-  template <typename T>
-  std::vector<T*> LegacyMultiOutput(const std::string& name) const {
-    auto names = op_.Outputs(name);
-    std::vector<T*> res;
-    res.reserve(names.size());
-    std::transform(names.begin(), names.end(), std::back_inserter(res),
-                   [&](const std::string& sub_name) -> T* {
-                     auto var = scope_.FindVar(sub_name);
-                     return var == nullptr ? nullptr : var->GetMutable<T>();
-                   });
-    return res;
-  }
-
   platform::Place GetPlace() const { return device_context_.GetPlace(); }
 
   template <typename DeviceContextType>
@@ -436,24 +371,13 @@ class ExecutionContext {
 template <>
 const Tensor* ExecutionContext::Input<Tensor>(const std::string& name) const;
 
-template <>
-const Tensor* ExecutionContext::LegacyInput<Tensor>(
-    const std::string& name) const;
-
 template <>
 const std::vector<const Tensor*> ExecutionContext::MultiInput<Tensor>(
     const std::string& name) const;
 
-template <>
-const std::vector<const Tensor*> ExecutionContext::LegacyMultiInput<Tensor>(
-    const std::string& name) const;
-
 template <>
 Tensor* ExecutionContext::Output<Tensor>(const std::string& name) const;
 
-template <>
-Tensor* ExecutionContext::LegacyOutput<Tensor>(const std::string& name) const;
-
 template <>
 std::vector<Tensor*> ExecutionContext::MultiOutput<Tensor>(
     const std::string& name) const;

paddle/fluid/inference/api/analysis_predictor.cc

@@ -183,6 +183,9 @@ void AnalysisPredictor::SetMkldnnThreadID(int tid) {
 bool AnalysisPredictor::Run(const std::vector<PaddleTensor> &inputs,
                             std::vector<PaddleTensor> *output_data,
                             int batch_size) {
+  if (UNLIKELY(config_.cpu_math_library_num_threads() > 1)) {
+    paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads());
+  }
   VLOG(3) << "Predictor::predict";
   inference::Timer timer;
   timer.tic();

paddle/fluid/inference/api/api_impl.cc

@@ -131,6 +131,9 @@ NativePaddlePredictor::~NativePaddlePredictor() {
 bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
                                 std::vector<PaddleTensor> *output_data,
                                 int batch_size) {
+  if (UNLIKELY(config_.cpu_math_library_num_threads() > 1)) {
+    paddle::platform::SetNumThreads(config_.cpu_math_library_num_threads());
+  }
   VLOG(3) << "Predictor::predict";
   Timer timer;
   timer.tic();

paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc

@@ -366,15 +366,17 @@ TEST(Analyzer_rnn1, ZeroCopyMultiThread) {
 #define NEW_TENSOR(name__) \
   auto name__##_tensor = predictor->GetInputTensor(#name__);
 
-  auto base_predictor = CreatePaddlePredictor<AnalysisConfig>(config);
+  std::vector<std::unique_ptr<PaddlePredictor>> predictors;
+  predictors.emplace_back(CreatePaddlePredictor<AnalysisConfig>(config));
+  for (int tid = 1; tid < FLAGS_num_threads; tid++) {
+    predictors.emplace_back(predictors.front()->Clone());
+  }
   double total_time_of_threads{0};
   std::vector<std::thread> threads;
 
   for (int tid = 0; tid < FLAGS_num_threads; tid++) {
     threads.emplace_back([&, tid] {
-      // To ensure the thread binding correctly,
-      // please clone inside the threadpool.
-      auto predictor = base_predictor->Clone();
+      auto &predictor = predictors[tid];
       NEW_TENSOR(data_lod_attention);
       NEW_TENSOR(cell_init);
       NEW_TENSOR(data);

paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc

@@ -266,15 +266,17 @@ TEST(Analyzer_seq_pool1, zerocopy_profile_threads) {
   SetConfig(&config);
   config.SwitchUseFeedFetchOps(false);
 
-  auto base_predictor = CreatePaddlePredictor<AnalysisConfig>(config);
+  std::vector<std::unique_ptr<PaddlePredictor>> predictors;
+  predictors.emplace_back(CreatePaddlePredictor<AnalysisConfig>(config));
+  for (int tid = 1; tid < FLAGS_num_threads; tid++) {
+    predictors.emplace_back(predictors.front()->Clone());
+  }
   double total_time_of_threads{0};
   std::vector<std::thread> threads;
 
   for (int tid = 0; tid < FLAGS_num_threads; tid++) {
     threads.emplace_back([&, tid] {
-      // To ensure the thread binding correctly,
-      // please clone inside the threadpool.
-      auto predictor = base_predictor->Clone();
+      auto &predictor = predictors[tid];
       std::vector<std::unique_ptr<ZeroCopyTensor>> inputs;
       PrepareZeroCopyInputs(predictor, &inputs);
       auto output_tensor = predictor->GetOutputTensor(out_var_name);

paddle/fluid/inference/tests/api/tester_helper.h

@@ -17,8 +17,10 @@
 #include <gtest/gtest.h>
 
 #include <algorithm>
+#include <memory>
 #include <string>
 #include <thread>  // NOLINT
+#include <unordered_map>
 #include <vector>
 #ifdef WITH_GPERFTOOLS
 #include <gperftools/profiler.h>
@@ -252,17 +254,19 @@ void TestMultiThreadPrediction(
   int batch_size = FLAGS_batch_size;
   int num_times = FLAGS_repeat;
   std::vector<std::thread> threads;
-  auto main_predictor = CreateTestPredictor(config, use_analysis);
+  std::vector<std::unique_ptr<PaddlePredictor>> predictors;
+  predictors.emplace_back(CreateTestPredictor(config, use_analysis));
+  for (int tid = 1; tid < num_threads; tid++) {
+    predictors.emplace_back(predictors.front()->Clone());
+  }
 
   size_t total_time{0};
   for (int tid = 0; tid < num_threads; ++tid) {
     threads.emplace_back([&, tid]() {
       // Each thread should have local inputs and outputs.
       // The inputs of each thread are all the same.
       std::vector<PaddleTensor> outputs_tid;
-      // To ensure the thread binding correctly,
-      // please clone inside the threadpool.
-      auto predictor = main_predictor->Clone();
+      auto &predictor = predictors[tid];
 #ifdef PADDLE_WITH_MKLDNN
       if (use_analysis) {
         static_cast<AnalysisPredictor *>(predictor.get())

paddle/fluid/operators/detection/CMakeLists.txt

@@ -33,6 +33,7 @@ detection_library(rpn_target_assign_op SRCS rpn_target_assign_op.cc)
 detection_library(generate_proposal_labels_op SRCS generate_proposal_labels_op.cc)
 detection_library(box_clip_op SRCS box_clip_op.cc box_clip_op.cu)
 detection_library(yolov3_loss_op SRCS yolov3_loss_op.cc)
+detection_library(box_decoder_and_assign_op SRCS box_decoder_and_assign_op.cc box_decoder_and_assign_op.cu)
 
 if(WITH_GPU)
   detection_library(generate_proposals_op SRCS generate_proposals_op.cc generate_proposals_op.cu DEPS memory cub)