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

test=develop

Author: NHZlX

cmake/inference_lib.cmake

@@ -18,7 +18,7 @@ function(copy TARGET)
     set(oneValueArgs "")
     set(multiValueArgs SRCS DSTS DEPS)
     cmake_parse_arguments(copy_lib "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-    set(inference_lib_dist_dep ${TARGET} ${inference_lib_dist_dep} PARENT_SCOPE)
+    set(fluid_lib_dist_dep ${TARGET} ${fluid_lib_dist_dep} PARENT_SCOPE)
 
     list(LENGTH copy_lib_SRCS copy_lib_SRCS_len)
     list(LENGTH copy_lib_DSTS copy_lib_DSTS_len)
@@ -185,7 +185,8 @@ copy(cmake_cache
   SRCS ${CMAKE_CURRENT_BINARY_DIR}/CMakeCache.txt
   DSTS ${FLUID_INSTALL_DIR})
 
-add_custom_target(inference_lib_dist DEPENDS ${inference_lib_dist_dep}) 
+# This command generates a complete fluid library for both train and inference
+add_custom_target(fluid_lib_dist DEPENDS ${fluid_lib_dist_dep}) 
 
 # paddle fluid version
 execute_process(

paddle/fluid/API.spec

@@ -75,7 +75,8 @@ paddle.fluid.layers.conv2d_transpose ArgSpec(args=['input', 'num_filters', 'outp
 paddle.fluid.layers.conv3d_transpose ArgSpec(args=['input', 'num_filters', 'output_size', 'filter_size', 'padding', 'stride', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(None, None, 0, 1, 1, None, None, None, True, None, None))
 paddle.fluid.layers.sequence_expand ArgSpec(args=['x', 'y', 'ref_level', 'name'], varargs=None, keywords=None, defaults=(-1, None))
 paddle.fluid.layers.sequence_expand_as ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,))
-paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.sequence_pad ArgSpec(args=['x', 'pad_value', 'maxlen', 'name'], varargs=None, keywords=None, defaults=(None, None))
+paddle.fluid.layers.sequence_unpad ArgSpec(args=['x', 'length', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.lstm_unit ArgSpec(args=['x_t', 'hidden_t_prev', 'cell_t_prev', 'forget_bias', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(0.0, None, None, None))
 paddle.fluid.layers.reduce_sum ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
 paddle.fluid.layers.reduce_mean ArgSpec(args=['input', 'dim', 'keep_dim', 'name'], varargs=None, keywords=None, defaults=(None, False, None))
@@ -127,6 +128,7 @@ paddle.fluid.layers.relu ArgSpec(args=['x', 'name'], varargs=None, keywords=None
 paddle.fluid.layers.log ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.crop ArgSpec(args=['x', 'shape', 'offsets', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
 paddle.fluid.layers.rank_loss ArgSpec(args=['label', 'left', 'right', 'name'], varargs=None, keywords=None, defaults=(None,))
+paddle.fluid.layers.margin_rank_loss ArgSpec(args=['label', 'left', 'right', 'margin', 'name'], varargs=None, keywords=None, defaults=(0.1, None))
 paddle.fluid.layers.elu ArgSpec(args=['x', 'alpha', 'name'], varargs=None, keywords=None, defaults=(1.0, None))
 paddle.fluid.layers.relu6 ArgSpec(args=['x', 'threshold', 'name'], varargs=None, keywords=None, defaults=(6.0, None))
 paddle.fluid.layers.pow ArgSpec(args=['x', 'factor', 'name'], varargs=None, keywords=None, defaults=(1.0, None))

paddle/fluid/CMakeLists.txt

@@ -12,6 +12,5 @@ endif(NOT WIN32)
 if(WITH_INFERENCE)
   # NOTE: please add subdirectory inference at last.
   add_subdirectory(inference)
+  add_subdirectory(train)
 endif()
-
-add_subdirectory(train)

paddle/fluid/framework/details/op_handle_base.h

@@ -64,7 +64,8 @@ class OpHandleBase {
   virtual bool IsMultiDeviceTransfer() { return false; }
 
   const platform::DeviceContext *DeviceContext(platform::Place place) {
-    return dev_ctxes_[place];
+    auto it = dev_ctxes_.find(place);
+    return it != dev_ctxes_.end() ? it->second : nullptr;
   }
 
   void SetDeviceContext(platform::Place place, platform::DeviceContext *ctx_) {

paddle/fluid/framework/executor.cc

@@ -46,6 +46,41 @@ ExecutorPrepareContext::~ExecutorPrepareContext() {
   VLOG(5) << "destroy ExecutorPrepareContext";
 }
 
+template <typename RefCntMap>
+static void DeleteUnusedTensors(const Scope& scope, const OperatorBase* op,
+                                GarbageCollector<Tensor>* gc,
+                                RefCntMap* ref_cnts) {
+  std::unordered_set<Tensor*> erase_tensors;
+
+  auto handler = [&](const VariableNameMap& name_map) {
+    for (auto& name_pair : name_map) {
+      for (auto& name : name_pair.second) {
+        auto it = ref_cnts->find(name);
+        if (it == ref_cnts->end()) continue;
+        if ((it->second)-- == 1) {
+          auto* var = scope.FindVar(name);
+          if (var != nullptr) {
+            VLOG(10) << "Erase tensor \'" << name << "\'";
+            if (var->IsType<LoDTensor>()) {
+              erase_tensors.insert(var->GetMutable<LoDTensor>());
+            } else if (var->IsType<SelectedRows>()) {
+              erase_tensors.insert(
+                  var->GetMutable<SelectedRows>()->mutable_value());
+            }
+          }
+        }
+      }
+    }
+  };
+
+  handler(op->Inputs());
+  handler(op->Outputs());
+
+  if (!erase_tensors.empty()) {
+    gc->Add(erase_tensors);
+  }
+}
+
 Executor::Executor(const platform::Place& place) : place_(place) {}
 
 void Executor::Close() {
@@ -331,9 +366,13 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
   }
 
   int64_t max_memory_size = GetEagerDeletionThreshold();
-
   std::unique_ptr<GarbageCollector<Tensor>> gc;
-  if (max_memory_size >= 0) {
+  // WhileOp would set keep_kids to false
+  // WhileGradOp would need the scopes created in WhileOp
+  // Perhaps, we should not perform eager deletion in WhileOp
+  // The scopes and variables created by WhileOp would be deleted
+  // in WhileGradOp.
+  if (max_memory_size >= 0 && !keep_kids) {
     ctx->ResetReferenceCount();
 #ifdef PADDLE_WITH_CUDA
     if (platform::is_gpu_place(place_)) {
@@ -352,45 +391,8 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
     op->Run(*local_scope, place_);
 
     if (gc != nullptr) {
-      std::vector<std::string> erase_vars;
-      for (auto& input : op->Inputs()) {
-        for (auto& input_name : input.second) {
-          auto it = ctx->cur_ref_cnts_.find(input_name);
-          if (it == ctx->cur_ref_cnts_.end()) continue;
-          if (it->second == 1) {  // should delete it
-            erase_vars.emplace_back(input_name);
-            ctx->cur_ref_cnts_.erase(input_name);
-          } else {
-            --(it->second);
-          }
-        }
-      }
-
-      for (auto& output : op->Outputs()) {
-        for (auto& output_name : output.second) {
-          auto it = ctx->cur_ref_cnts_.find(output_name);
-          if (it == ctx->cur_ref_cnts_.end()) continue;
-          if (it->second == 1) {
-            erase_vars.emplace_back(output_name);
-            ctx->cur_ref_cnts_.erase(output_name);
-          } else {
-            --(it->second);
-          }
-        }
-      }
-
-      if (!erase_vars.empty()) {
-        std::vector<framework::LoDTensor*> erase_tensors;
-        for (auto& name : erase_vars) {
-          auto* var = local_scope->FindVar(name);
-          if (var == nullptr) continue;
-          if (var->IsType<framework::LoDTensor>()) {
-            auto* tensor = var->GetMutable<framework::LoDTensor>();
-            erase_tensors.push_back(tensor);
-          }
-        }
-        if (!erase_tensors.empty()) gc->Add(erase_tensors);
-      }
+      DeleteUnusedTensors(*local_scope, op.get(), gc.get(),
+                          &(ctx->cur_ref_cnts_));
     }
 
     if (FLAGS_benchmark) {

paddle/fluid/framework/executor.h

@@ -32,38 +32,32 @@ template <typename T>
 std::unordered_map<std::string, T> GetNonPersistableReferenceCount(
     const ProgramDesc& prog, size_t block_id) {
   auto& block = prog.Block(block_id);
-  std::unordered_set<std::string> ignored_vars;
   std::unordered_map<std::string, T> ref_cnts;
 
-  for (auto var_desc : block.AllVars()) {
-    auto type = var_desc->Proto()->type().type();
-    if (type != proto::VarType::LOD_TENSOR || var_desc->Persistable()) {
-      ignored_vars.insert(var_desc->Name());  // ignore persistable vars
-    }
-  }
-
-  for (auto op_desc : block.AllOps()) {
-    for (auto& input : op_desc->Inputs()) {
-      for (auto& input_name : input.second) {
-        if (!ignored_vars.count(input_name)) {
-          if (ref_cnts.count(input_name))
-            ++ref_cnts[input_name];
-          else
-            ref_cnts[input_name] = 1;
+  auto update_ref_cnts = [&](OpDesc* op_desc, const VariableNameMap& name_map) {
+    for (auto& name_pair : name_map) {
+      for (auto& name : name_pair.second) {
+        auto* var_desc = block.FindVar(name);
+        if (var_desc == nullptr || var_desc->Persistable()) continue;
+        auto type = var_desc->Proto()->type().type();
+        if (type != proto::VarType::LOD_TENSOR &&
+            type != proto::VarType::SELECTED_ROWS) {
+          continue;
         }
-      }
-    }
 
-    for (auto& output : op_desc->Outputs()) {
-      for (auto output_name : output.second) {
-        if (!ignored_vars.count(output_name)) {
-          if (ref_cnts.count(output_name))
-            ++ref_cnts[output_name];
-          else
-            ref_cnts[output_name] = 1;
+        auto it = ref_cnts.find(name);
+        if (it != ref_cnts.end()) {
+          ++it->second;
+        } else {
+          ref_cnts[name] = 1;
         }
       }
     }
+  };
+
+  for (auto op_desc : block.AllOps()) {
+    update_ref_cnts(op_desc, op_desc->Inputs());
+    update_ref_cnts(op_desc, op_desc->Outputs());
   }
   return ref_cnts;
 }

paddle/fluid/framework/ir/conv_bn_fuse_pass.cc

@@ -44,89 +44,6 @@ namespace ir {
   GET_IR_NODE_FROM_SUBGRAPH(bn_saved_mean, bn_saved_mean, pattern_name);     \
   GET_IR_NODE_FROM_SUBGRAPH(bn_saved_variance, bn_saved_variance, pattern_name)
 
-template <typename UnaryOperation>
-LoDTensor tensor_apply(const LoDTensor& vec, UnaryOperation f) {
-  LoDTensor vec_y;
-  vec_y.Resize(vec.dims());
-  const float* x = vec.data<float>();
-  float* y = vec_y.mutable_data<float>(platform::CPUPlace());
-  for (int64_t i = 0; i < vec.numel(); i++) {
-    y[i] = f(x[i]);
-  }
-  return vec_y;
-}
-
-void tensor_apply_inplace(LoDTensor* vec, float (*f)(float)) {
-  float* data = vec->mutable_data<float>(platform::CPUPlace());
-  for (int64_t i = 0; i < vec->numel(); i++) {
-    data[i] = f(data[i]);
-  }
-}
-
-template <typename BinaryOperation>
-LoDTensor tensor_apply_eltwise(const LoDTensor& vec_a, const LoDTensor& vec_b,
-                               BinaryOperation f) {
-  PADDLE_ENFORCE_EQ(vec_a.dims(), vec_b.dims());
-  LoDTensor vec_y;
-  vec_y.Resize(vec_a.dims());
-  const float* a = vec_a.data<float>();
-  const float* b = vec_b.data<float>();
-  float* y = vec_y.mutable_data<float>(platform::CPUPlace());
-  for (int64_t i = 0; i < vec_a.numel(); i++) {
-    y[i] = f(a[i], b[i]);
-  }
-  return vec_y;
-}
-
-template <typename BinaryOperation>
-LoDTensor tensor_apply_eltwise_broadcast(const LoDTensor& vec_a,
-                                         const LoDTensor& vec_b,
-                                         BinaryOperation f) {
-  PADDLE_ENFORCE_EQ(vec_a.dims().size(), 2);
-  PADDLE_ENFORCE_EQ(vec_b.dims().size(), 2);
-  PADDLE_ENFORCE_EQ(vec_a.dims()[0], vec_b.dims()[0]);
-  PADDLE_ENFORCE_EQ(vec_b.dims()[1], 1);
-  LoDTensor vec_y;
-  vec_y.Resize(vec_a.dims());
-  const float* a = vec_a.data<float>();
-  const float* b = vec_b.data<float>();
-  float* y = vec_y.mutable_data<float>(platform::CPUPlace());
-  size_t a_height = vec_a.dims()[0];
-  size_t a_width = vec_a.dims()[1];
-  for (size_t h = 0; h < a_height; h++) {
-    for (size_t w = 0; w < a_width; ++w) {
-      *(y++) = f(*(a++), b[h]);
-    }
-  }
-  return vec_y;
-}
-
-// reshape to two dimensions {A, B * C * ...}
-void make_tensor_2d(LoDTensor* tensor_to_reshape) {
-  auto dims_count = tensor_to_reshape->dims().size();
-  PADDLE_ENFORCE_GT(dims_count, 0);
-
-  int size2 = 1;
-  for (int i = 1; i < dims_count; i++) {
-    size2 *= tensor_to_reshape->dims()[i];
-  }
-  tensor_to_reshape->Resize(make_ddim({tensor_to_reshape->dims()[0], size2}));
-}
-
-void recompute_conv_weights(LoDTensor* weights, LoDTensor* tmp) {
-  // remember the weights tensor shape {A, B, C, ...}
-  auto weights_shape = weights->dims();
-  // reduce the weights to 2d {A, B * C * ...}
-  make_tensor_2d(weights);
-  // make tmp tensor 2d by adding 1 as second dim {A, 1}
-  make_tensor_2d(tmp);
-
-  *weights =
-      tensor_apply_eltwise_broadcast(*weights, *tmp, std::multiplies<float>());
-  // reshape weights to the original dims {A, B, C, ...}
-  weights->Resize(weights_shape);
-}
-
 void recompute_bias_and_weights(const Scope* scope,
                                 ir::Node* conv_weight,            //
                                 const ir::Node& bn_scale,         //
@@ -135,6 +52,13 @@ void recompute_bias_and_weights(const Scope* scope,
                                 const ir::Node& bn_variance,      //
                                 LoDTensor* eltwise_y_in_tensor,   //
                                 float epsilon) {
+  using EigenVectorArrayMap =
+      Eigen::Map<Eigen::Array<float, Eigen::Dynamic, 1>>;
+  using ConstEigenVectorArrayMap =
+      Eigen::Map<const Eigen::Array<float, Eigen::Dynamic, 1>>;
+  using EigenMatrixArrayMap = Eigen::Map<
+      Eigen::Array<float, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>>;
+
   // Re-compute bias of conv2d from BN
   PADDLE_ENFORCE_EQ(eltwise_y_in_tensor->dims(), bn_bias_tensor.dims());
 
@@ -143,31 +67,38 @@ void recompute_bias_and_weights(const Scope* scope,
       scope->FindVar(bn_variance.Name())->GetMutable<LoDTensor>();
   auto* mean_tensor = scope->FindVar(bn_mean.Name())->GetMutable<LoDTensor>();
 
-  auto std_tensor = LoDTensor();
-  std_tensor.Resize(bn_bias_tensor.dims());
-  std_tensor =
-      tensor_apply(*variance_tensor, [&](float x) { return x + epsilon; });
+  ConstEigenVectorArrayMap scale_array(scale_tensor->data<float>(),
+                                       scale_tensor->numel(), 1);
+  EigenVectorArrayMap variance_array(
+      variance_tensor->mutable_data<float>(platform::CPUPlace()),
+      variance_tensor->numel(), 1);
+  ConstEigenVectorArrayMap mean_array(mean_tensor->data<float>(),
+                                      mean_tensor->numel(), 1);
+  ConstEigenVectorArrayMap bn_bias_array(bn_bias_tensor.data<float>(),
+                                         bn_bias_tensor.numel(), 1);
 
-  using EigenVectorArrayMap =
-      Eigen::Map<Eigen::Array<float, Eigen::Dynamic, 1>>;
+  // variance will not be used anymore, so make it std_array and then tmp_array
+  variance_array += epsilon;
+  variance_array = variance_array.sqrt();
+  variance_array = scale_array / variance_array;
+
+  EigenVectorArrayMap eltwise_y_in_array(
+      eltwise_y_in_tensor->mutable_data<float>(platform::CPUPlace()),
+      eltwise_y_in_tensor->numel(), 1);
 
-  EigenVectorArrayMap std_vec(
-      std_tensor.mutable_data<float>(platform::CPUPlace()), std_tensor.numel(),
-      1);
-  std_vec = std_vec.sqrt();
-  auto tmp_tensor =
-      tensor_apply_eltwise(*scale_tensor, std_tensor, std::divides<float>());
-  auto tensor_minus = tensor_apply_eltwise(*eltwise_y_in_tensor, *mean_tensor,
-                                           std::minus<float>());
-  auto tensor_mul =
-      tensor_apply_eltwise(tensor_minus, tmp_tensor, std::multiplies<float>());
-  *eltwise_y_in_tensor =
-      tensor_apply_eltwise(tensor_mul, bn_bias_tensor, std::plus<float>());
+  eltwise_y_in_array =
+      ((eltwise_y_in_array - mean_array) * variance_array) + bn_bias_array;
 
   // Re-compute weight of conv2d from BN
-  auto* current_param =
-      scope->FindVar(conv_weight->Name())->GetMutable<LoDTensor>();
-  recompute_conv_weights(current_param, &tmp_tensor);
+  auto* weights = scope->FindVar(conv_weight->Name())->GetMutable<LoDTensor>();
+  auto weights_shape = weights->dims();
+  auto weights_shape_2d = flatten_to_2d(weights_shape, 1);
+
+  EigenMatrixArrayMap weights_array_2d(
+      weights->mutable_data<float>(platform::CPUPlace()), weights_shape_2d[0],
+      weights_shape_2d[1]);
+
+  weights_array_2d.colwise() *= variance_array;
 }
 
 std::unique_ptr<ir::Graph> ConvBNFusePass::ApplyImpl(

paddle/fluid/framework/parallel_executor.cc

@@ -307,6 +307,10 @@ ParallelExecutor::~ParallelExecutor() {
       }
     }
   }
+
+  // member_ must be destructed before gcs_ since the destructor of
+  // ReferenceCountOpHandle use raw pointers of gcs_ inside.
+  member_.reset();
 }
 
 }  // namespace framework