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

test=develop

paddle/fluid/framework/details/CMakeLists.txt

@@ -50,7 +50,12 @@ cc_library(data_balance_op_handle SRCS data_balance_op_handle.cc DEPS op_handle_
 cc_library(gather_op_handle SRCS gather_op_handle.cc DEPS op_handle_base scope ddim memory variable_visitor)
 cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base scope)
 
-cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper)
+if(WITH_GPU)
+cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper gpu_info)
+else()
+cc_library(memory_optimize_helper SRCS memory_optimize_helper.cc DEPS graph graph_helper cpu_info)
+endif()
+
 cc_library(memory_optimize_pass SRCS memory_optimize_pass.cc DEPS memory_optimize_helper pass)
 cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_info)
 cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper)

paddle/fluid/framework/details/build_strategy.cc

@@ -240,7 +240,9 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
         continue;
       }
     }
+    VLOG(3) << "Start Apply Pass " << pass->Type();
     graph = pass->Apply(std::move(graph));
+    VLOG(3) << "Finish Apply Pass " << pass->Type();
   }
   return graph;
 }

paddle/fluid/framework/details/inplace_op_pass.cc

@@ -49,7 +49,7 @@ DEFINE_bool(
     "If this option turns on, only these op in whitelist can be inplaced."
     "If it turns off, all of the running op can be candidate of inplaced op."
     "Such as scale, elementwise_add"
-    "By default, it's turned on");
+    "By default, it's turned off");
 
 DECLARE_string(memory_optimize_debug);
 

paddle/fluid/framework/details/memory_optimize_helper.cc

@@ -13,13 +13,19 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/details/memory_optimize_helper.h"
+#include <algorithm>
 #include <deque>
 #include <functional>
-#include <iostream>
+#include <iterator>
 #include <numeric>
 #include <sstream>
 #include <string>
 #include "paddle/fluid/framework/var_desc.h"
+#include "paddle/fluid/platform/cpu_info.h"
+
+#ifdef PADDLE_WITH_CUDA
+#include "paddle/fluid/platform/gpu_info.h"
+#endif  // PADDLE_WITH_CUDA
 
 namespace paddle {
 namespace framework {
@@ -166,6 +172,11 @@ struct NodeComparator {
   bool operator()(ir::Node* lhs, ir::Node* rhs) const {
     auto* lhs_desc = FindVarDescInBlock(lhs);
     auto* rhs_desc = FindVarDescInBlock(rhs);
+    // match data type
+    if (lhs_desc->GetDataType() != rhs_desc->GetDataType()) {
+      return false;
+    }
+    // match shape
     auto lhs_shape = lhs_desc->GetShape();
     auto rhs_shape = rhs_desc->GetShape();
     if ((lhs_shape[0] == -1 && rhs_shape[0] == -1) ||
@@ -230,6 +241,27 @@ ir::Node* OrderedSet::FindBestFitNode(ir::Node* var) const {
   return found_node;
 }
 
+ir::Node* OrderedSet::FindNextBestFitNode(ir::Node* var, ir::Node* prev) const {
+  ir::Node* found_node = nullptr;
+  NodeComparator functor;
+  auto it =
+      std::find_if(nodes_.begin(), nodes_.end(), [&](const NodeVector& v) {
+        if (v.front() == prev)
+          return true;
+        else
+          return false;
+      });
+  PADDLE_ENFORCE(it != nodes_.end(), "Not found previous in node list!");
+  for (it = std::next(it); it != nodes_.end(); ++it) {
+    auto& candidate = it->front();
+    if (functor(var, candidate)) {
+      found_node = candidate;
+      break;
+    }
+  }
+  return found_node;
+}
+
 bool OrderedSet::Has(ir::Node* var) const {
   if (mark_table_.count(var->Name())) {
     auto& node_in_samename = mark_table_.at(var->Name());
@@ -241,10 +273,15 @@ bool OrderedSet::Has(ir::Node* var) const {
   return false;
 }
 
+void OrderedSet::Erase(const std::string& var) {
+  PADDLE_ENFORCE(mark_table_.count(var));
+  nodes_.erase(mark_table_[var]);
+  mark_table_.erase(var);
+}
+
 void OrderedSet::Erase(ir::Node* var) {
-  PADDLE_ENFORCE(mark_table_.count(var->Name()));
-  nodes_.erase(mark_table_[var->Name()]);
-  mark_table_.erase(var->Name());
+  PADDLE_ENFORCE(var != nullptr);
+  Erase(var->Name());
 }
 
 std::string OrderedSet::ToString() const {
@@ -274,14 +311,35 @@ bool NodeCanReused(ir::Node* node) {
   return flag;
 }
 
+int MinChunkSize() {
+  int size{0};
+#ifdef PADDLE_WITH_CUDA
+  size = platform::GpuMinChunkSize();
+#else
+  size = platform::CpuMinChunkSize();
+#endif  // PADDLE_WITH_CUDA
+  return size;
+}
+
 bool NodeCanReused(const VarDesc& node) {
   auto type = node.GetType();
+  // only these types holds bulk of gpu memory
   if (!(type == proto::VarType::LOD_TENSOR ||
         type == proto::VarType::SELECTED_ROWS ||
         type == proto::VarType::LOD_TENSOR_ARRAY)) {
     return false;
   }
-  if (node.Persistable() || node.GetShape().empty()) {
+  // persistable variable is parameter
+  if (node.Persistable()) {
+    return false;
+  }
+  // shape < min_chunk_size is meaningless.
+  // further more, fetched loss always has size = 1
+  // which should not be reused.
+  auto shape = node.GetShape();
+  int size = std::abs(
+      std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>()));
+  if (shape.empty() || size < MinChunkSize()) {
     return false;
   }
   // vars can be @EMPTY@, @LR_DECAY_REUSE_ID@. For example, while_grad
@@ -461,7 +519,9 @@ ir::Node* ControlFlowGraph::GetNodeByName(const std::string& name,
   for (auto* node : ops_) {
     if (node == op) break;
     for (auto& output : node->outputs) {
-      if (output->Name() == name) {
+      PADDLE_ENFORCE((output != nullptr && output->IsVar()),
+                     "Output is empty!");
+      if (output->Var() && output->Name() == name) {
         found_node = output;
       }
     }

paddle/fluid/framework/details/memory_optimize_helper.h

@@ -55,6 +55,7 @@ class OrderedSet {
 
   void Insert(ir::Node* var);
   void Erase(ir::Node* var);
+  void Erase(const std::string& var);
   bool Has(ir::Node* var) const;
   void Clear() {
     mark_table_.clear();
@@ -62,6 +63,7 @@ class OrderedSet {
   }
   // find the bestfit shape node block with var.
   ir::Node* FindBestFitNode(ir::Node* var) const;
+  ir::Node* FindNextBestFitNode(ir::Node* var, ir::Node* prev) const;
   // map store non-const iterator, can not promise const
   int GetNodeIndexInPool(ir::Node* var);
   // pool all node to string

paddle/fluid/framework/details/memory_optimize_helper_test.cc

@@ -107,6 +107,52 @@ TEST(OrderedSet, Normal) {
     ASSERT_EQ(pool.GetNodeIndexInPool(cache), 5);  // match  4:[5,2]
   }
 }
+
+TEST(OrderedSet, FindBestFitNode) {
+  OrderedSet pool;
+  std::vector<std::unique_ptr<ir::Node>> nodes;
+  ProgramDesc prog;
+  BlockDesc* block_desc = prog.MutableBlock(0);
+  auto* op_desc = block_desc->AppendOp();
+  op_desc->SetType("dummy");
+  std::unique_ptr<ir::Node> op = ir::CreateNodeForTest(op_desc);
+
+  {
+    auto desc = block_desc->Var("a");
+    desc->SetShape({128, 128});
+    std::unique_ptr<ir::Node> node = ir::CreateNodeForTest(desc);
+    node->inputs.emplace_back(op.get());
+    nodes.emplace_back(std::move(node));
+  }
+  {
+    auto desc = block_desc->Var("b");
+    desc->SetShape({128, 129});
+    std::unique_ptr<ir::Node> node = ir::CreateNodeForTest(desc);
+    node->inputs.emplace_back(op.get());
+    nodes.emplace_back(std::move(node));
+  }
+  {
+    auto desc = block_desc->Var("c");
+    desc->SetShape({128, 128});
+    std::unique_ptr<ir::Node> node = ir::CreateNodeForTest(desc);
+    node->inputs.emplace_back(op.get());
+    nodes.emplace_back(std::move(node));
+  }
+
+  for (auto& node : nodes) {
+    pool.Insert(node.get());
+  }
+
+  // FindNextBestFitNode
+  auto* n = nodes[0].get();
+  auto* cache = pool.FindBestFitNode(n);
+  PADDLE_ENFORCE(cache->Name() == "a");
+  cache = pool.FindNextBestFitNode(n, cache);
+  PADDLE_ENFORCE(cache->Name() == "c");
+  cache = pool.FindNextBestFitNode(n, cache);
+  PADDLE_ENFORCE(cache->Name() == "b");
+}
+
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/memory_optimize_pass.cc

@@ -69,11 +69,20 @@ std::unique_ptr<ir::Graph> MemoryOptimizePass::ApplyImpl(
     }
 
     for (auto& var : op->outputs) {
-      if (!NodeCanReused(var) || cfg_->Use(op).count(var->Name()) == 0 ||
-          skip_set_.count(var->Name()))
+      if (var->IsVar() && !var->IsCtrlVar() && skip_set_.count(var->Name())) {
+        VLOG(3) << "Skip set contains variable of " << var->Name()
+                << "disable reuse on it. skipped";
         continue;
+      }
+      if (NodeCanReused(var) && cfg_->Use(op).count(var->Name()) == 0) {
         ir::Node* cache = pool_.FindBestFitNode(var);
-
+        while (cache != nullptr && var->Name() == cache->Name()) {
+          VLOG(3) << "The same cache variable is cascade reused. "
+                  << cache->Name() << " is re-filled to the pool after "
+                  << "the reused op is finished. Current op can not "
+                  << "replace it again. Skip this candidate.";
+          cache = pool_.FindNextBestFitNode(var, cache);
+        }
         if (var->Name() == FLAGS_memory_optimize_debug) {
           VLOG(3) << "start match var " << DebugString(var) << " of op "
                   << op->Name();
@@ -82,42 +91,37 @@ std::unique_ptr<ir::Graph> MemoryOptimizePass::ApplyImpl(
                   << ((cache == nullptr) ? "False" : "True");
         }
 
-      if (cache == nullptr) continue;
-      if (var->Name() == cache->Name()) {
-        VLOG(3) << "The same cache variable is cascade reused." << var->Name()
-                << " is re-filled to the pool after"
-                << "the reused op is finished. Current op can not "
-                << "replace it again. Skip this candidate.";
-        continue;
-
+        if (cache != nullptr) {
           int node_idx_in_pool = pool_.GetNodeIndexInPool(cache);
           VLOG(3) << string::Sprintf(
               "!!! %s,  %s => %s, cache idx %d, pool size %d",
               std::to_string(reuse_id++), DebugString(var), DebugString(cache),
               node_idx_in_pool, static_cast<int>(pool_.size()));
+          // NOTE(dzhwinter): update the ProgramDesc/IR Graph
+          // and the CFG Graph on the fly.
+          //
+          // IR Graph define the dependence relationship between nodes.
+          //
+          // ProgramDesc defines the input/output vars. Its used in
+          // CreateOp, CreateVar when running happens.
+          //
+          // CFG Graph store the liveness information, when reuse happens
+          // we also need to update the variable liveness.
+          const std::string var_name = var->Name();
+          const std::string cache_name = cache->Name();
 
-        // update CFG Graph on the fly.
-        // reused var maybe re-fill into the pool
-        cfg_->RenameVarInCFGGraph(var->Name(), cache->Name(), idx);
-        // NOTE(dzhwinter): we need to both update the ProgramDesc
-        // and IR Graph. because op_desc/var_desc is used in CreateOp,
-        // CreateVar when running happens. But IR Graph
-        // define the dependence relationship between nodes.
-        RenameVarInGraphDesc(var->Name(), cache->Name(), idx);
-        RenameVarInGraphNode(var->Name(), cache->Name(), idx, graph.get());
-
-        pool_.Erase(cache);
+          cfg_->RenameVarInCFGGraph(var_name, cache_name, idx);
+          RenameVarInGraphDesc(var_name, cache_name, idx);
+          RenameVarInGraphNode(var_name, cache_name, idx, graph.get());
+          pool_.Erase(cache_name);
+        }
+      }
     }
-
     // fill the pool
-      std::unordered_set<std::string> unlived_vars;
     for (auto var : cfg_->LiveIn(op)) {
       if (cfg_->LiveOut(op).count(var) == 0) {
-          unlived_vars.emplace(var);
-        }
-      }
-      for (auto var : unlived_vars) {
         ir::Node* var_node = cfg_->GetNodeByName(var, op);
+        if (var_node == nullptr || var_node->IsCtrlVar()) continue;
         if (NodeCanReused(var_node) && !pool_.Has(var_node)) {
           pool_.Insert(var_node);
         }
@@ -273,8 +277,7 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var,
     // redirect the input to the latest version of cache_var
     for (auto* node : op->inputs) {
       if (node->Name() == var) {
-        ir::Node* cache_node = graph->CreateVarNode(var_desc.get());
-        var_nodes_[cache_var].emplace_back(cache_node);
+        ir::Node* cache_node = var_nodes_[cache_var].back();
 
         // swap node to cache_node
         cache_node->outputs.insert(cache_node->outputs.end(),
@@ -283,11 +286,15 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var,
         auto* prev_op = node->inputs[0];
         std::replace(prev_op->outputs.begin(), prev_op->outputs.end(), node,
                      cache_node);
-        cache_node->inputs.emplace_back(prev_op);
         for (auto* next_op : node->outputs) {
           std::replace(next_op->inputs.begin(), next_op->inputs.end(), node,
                        cache_node);
         }
+
+        // erase unused node
+        auto& nodes = var_nodes_.at(var);
+        nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end());
+        graph->RemoveNode(node);
       }
     }
 
@@ -307,15 +314,14 @@ void MemoryOptimizePass::RenameVarInGraphNode(const std::string& var,
           std::replace(next_op->inputs.begin(), next_op->inputs.end(), node,
                        cache_node);
         }
-      }
-    }
-  }
 
-  // release node of unused var in graph
-  for (auto* node : var_nodes_[var]) {
+        // erase unused node
+        auto& nodes = var_nodes_.at(var);
+        nodes.erase(std::remove(nodes.begin(), nodes.end(), node), nodes.end());
         graph->RemoveNode(node);
       }
-  var_nodes_.at(var).clear();
+    }
+  }
 }
 
 }  // namespace details

paddle/fluid/framework/inplace_op_inference_test.cc

@@ -179,11 +179,11 @@ TEST(InferInplace, SingleOpInplaceInToOut) {
   op->SetOutput("Out", {"test2_out"});
 
   prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR);
-  prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 64});
+  prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 64, 128, 128});
   prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR);
   prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR);
   prog.MutableBlock(0)->Var("test2_out");
-  prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 128, 128});
 
   auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
   auto in_to_outs = infer_inplace(*op, op->Block());
@@ -201,11 +201,11 @@ TEST(InferInplace, SingleGradOpInplaceInToOut) {
   op->SetOutput(GradVarName("X"), {"test2_a", "test2_b", "test2_c"});
 
   prog.MutableBlock(0)->Var("test2_a")->SetType(proto::VarType::LOD_TENSOR);
-  prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("test2_a")->SetShape({32, 16, 1024, 1024});
   prog.MutableBlock(0)->Var("test2_b")->SetType(proto::VarType::LOD_TENSOR);
   prog.MutableBlock(0)->Var("test2_c")->SetType(proto::VarType::LOD_TENSOR);
   prog.MutableBlock(0)->Var("test2_out");
-  prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("test2_out")->SetShape({32, 16, 1024, 1024});
 
   auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
   auto in_to_outs = infer_inplace(*op, op->Block());
@@ -233,12 +233,12 @@ TEST(InferInplace, MultiOutInplaceInToOut) {
   prog.MutableBlock(0)->Var("o0");
   prog.MutableBlock(0)->Var("y0");
   prog.MutableBlock(0)->Var("z0");
-  prog.MutableBlock(0)->Var("a0")->SetShape({32, 16});
-  prog.MutableBlock(0)->Var("b0")->SetShape({32, 16});
-  prog.MutableBlock(0)->Var("c0")->SetShape({32, 16});
-  prog.MutableBlock(0)->Var("o0")->SetShape({32, 16});
-  prog.MutableBlock(0)->Var("y0")->SetShape({32, 16});
-  prog.MutableBlock(0)->Var("z0")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024});
+  prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024});
+  prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024});
+  prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024});
+  prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024});
+  prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024});
 
   auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
   auto in_to_outs = infer_inplace(*op, op->Block());
@@ -267,12 +267,12 @@ TEST(InferInplace, MultiGradInplaceInToOut) {
   prog.MutableBlock(0)->Var("o0");
   prog.MutableBlock(0)->Var("y0");
   prog.MutableBlock(0)->Var("z0");
-  prog.MutableBlock(0)->Var("a0")->SetShape({32, 16});
-  prog.MutableBlock(0)->Var("b0")->SetShape({32, 16});
-  prog.MutableBlock(0)->Var("c0")->SetShape({32, 16});
-  prog.MutableBlock(0)->Var("o0")->SetShape({32, 16});
-  prog.MutableBlock(0)->Var("y0")->SetShape({32, 16});
-  prog.MutableBlock(0)->Var("z0")->SetShape({32, 16});
+  prog.MutableBlock(0)->Var("a0")->SetShape({32, 16, 1024, 1024});
+  prog.MutableBlock(0)->Var("b0")->SetShape({32, 16, 1024, 1024});
+  prog.MutableBlock(0)->Var("c0")->SetShape({32, 16, 1024, 1024});
+  prog.MutableBlock(0)->Var("o0")->SetShape({32, 16, 1024, 1024});
+  prog.MutableBlock(0)->Var("y0")->SetShape({32, 16, 1024, 1024});
+  prog.MutableBlock(0)->Var("z0")->SetShape({32, 16, 1024, 1024});
 
   auto& infer_inplace = OpInfoMap::Instance().Get(op->Type()).infer_inplace_;
   auto in_to_outs = infer_inplace(*op, op->Block());

paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc

@@ -38,9 +38,13 @@ std::unique_ptr<ir::Graph> IdentityScaleOpCleanPass::ApplyImpl(
                       ->assert_is_op("scale")
                       ->assert_op_attr<float>("scale", 1.)
                       ->assert_op_attr<float>("bias", 0.);
-  auto scale_out = detector.mutable_pattern()
+  auto scale_out =
+      detector.mutable_pattern()
           ->NewNode("scale_out")
-                       ->assert_is_op_output("scale");
+          ->assert_is_op_output("scale")
+          // scale's output var should has only one consumer, or it can't be
+          // removed.
+          ->assert_more([](Node* x) { return x->outputs.size() == 1UL; });
 
   pre_op->LinksTo({scale_in});
   scale_op->LinksFrom({scale_in}).LinksTo({scale_out});

paddle/fluid/imperative/layer.cc

@@ -207,7 +207,7 @@ framework::LoDTensor& VarBase::GradValue() {
 
 std::map<std::string, std::vector<VarBase*>> OpBase::ApplyGrad() {
   if (grad_op_descs_.empty() && backward_id_ <= 0) {
-    LOG(WARNING) << "op with no grad: " << op_desc_->Type();
+    VLOG(3) << "op with no grad: " << op_desc_->Type();
     return {};
   }
 

paddle/fluid/inference/analysis/ir_passes/subgraph_detector.cc

@@ -460,77 +460,6 @@ inline bool CheckNodeIndegreeEquals(const Node &node, size_t n) {
   return node.inputs.size() == n;
 }
 
-NodesTSIterator::NodesTSIterator(const std::vector<Node *> &source) {
-  PADDLE_ENFORCE(!source.empty(),
-                 "Start points of topological sorting should not be empty!");
-  // CHECK all the inputs' in-degree is 0
-  for (auto *node : source) {
-    PADDLE_ENFORCE(CheckNodeIndegreeEquals(*node, 0));
-  }
-
-  std::unordered_set<Node *> visited;
-  std::unordered_set<Node *> to_visit{source.begin(), source.end()};
-
-  std::vector<Node *> inlink_visited;
-  while (!to_visit.empty()) {
-    std::vector<Node *> queue(to_visit.begin(), to_visit.end());
-    for (auto *p : queue) {
-      if (Agent(p).deleted()) {
-        visited.insert(p);
-        to_visit.erase(p);
-      }
-
-      inlink_visited.clear();
-
-      std::copy_if(p->inputs.begin(), p->inputs.end(),
-                   std::back_inserter(inlink_visited),
-                   [&](Node *x) -> bool { return visited.count(x) != 0; });
-
-      if (inlink_visited.size() == p->inputs.size()) {
-        sorted_.push_back(p);
-        for (auto *_ : p->outputs) {
-          if (!visited.count(_)) {
-            to_visit.insert(_);
-          }
-        }
-
-        to_visit.erase(p);
-        visited.insert(p);
-      }
-    }
-  }
-}
-
-NodesTSIterator::NodesTSIterator(const NodesTSIterator &other)
-    : sorted_(other.sorted_), cursor_(other.cursor_) {}
-
-Node &NodesTSIterator::operator*() {
-  PADDLE_ENFORCE_LT(cursor_, sorted_.size());
-  return *sorted_[cursor_];
-}
-
-NodesTSIterator &NodesTSIterator::operator++() {
-  if (++cursor_ >= sorted_.size()) {
-    sorted_.clear();
-    cursor_ = 0;
-  }
-  return *this;
-}
-NodesTSIterator &NodesTSIterator::operator=(const NodesTSIterator &other) {
-  cursor_ = other.cursor_;
-  sorted_ = other.sorted_;
-  return *this;
-}
-
-bool NodesTSIterator::operator==(const NodesTSIterator &other) {
-  return sorted_ == other.sorted_ && cursor_ == other.cursor_;
-}
-
-Node *NodesTSIterator::operator->() {
-  PADDLE_ENFORCE_LT(cursor_, sorted_.size());
-  return sorted_[cursor_];
-}
-
 }  // namespace analysis
 }  // namespace inference
 }  // namespace paddle

paddle/fluid/inference/analysis/ir_passes/subgraph_detector.h

@@ -30,6 +30,7 @@ namespace inference {
 namespace analysis {
 
 using framework::ir::Graph;
+using framework::ir::NodesTSIterator;
 
 const char kIsFunctionNode[] = "__is_function_node__";
 const char kFunctionNodeSubGraph[] = "__function_node_sub_graph__";
@@ -132,32 +133,6 @@ struct Agent {
   framework::ir::Node *x_;
 };
 
-// Topological sorting iterator on nodes.
-struct NodesTSIterator
-    : public std::iterator<std::forward_iterator_tag, framework::ir::Node *> {
-  NodesTSIterator() = default;
-  explicit NodesTSIterator(const std::vector<framework::ir::Node *> &source);
-  NodesTSIterator(NodesTSIterator &&other)
-      : sorted_(std::move(other.sorted_)), cursor_(other.cursor_) {
-    other.cursor_ = 0;
-  }
-  NodesTSIterator(const NodesTSIterator &other);
-
-  framework::ir::Node &operator*();
-  NodesTSIterator &operator++();
-  // TODO(Superjomn) current implementation just compare the first
-  // element, need to compare the graph and all the elements in the queue and
-  // set.
-  NodesTSIterator &operator=(const NodesTSIterator &other);
-  bool operator==(const NodesTSIterator &other);
-  bool operator!=(const NodesTSIterator &other) { return !(*this == other); }
-  framework::ir::Node *operator->();
-
- private:
-  std::vector<framework::ir::Node *> sorted_;
-  size_t cursor_{0};
-};
-
 // The nodes those have no input will be treated as start points.
 static std::vector<framework::ir::Node *> ExtractStartPoints(const Graph &g) {
   std::vector<framework::ir::Node *> result;

paddle/fluid/operators/detection/density_prior_box_op.h

@@ -72,7 +72,7 @@ class DensityPriorBoxOpKernel : public framework::OpKernel<T> {
 #ifdef PADDLE_WITH_MKLML
 #pragma omp parallel for
 #endif
-    for (int i = 0; i < fixed_ratios.size(); i++) {
+    for (size_t i = 0; i < fixed_ratios.size(); i++) {
       sqrt_fixed_ratios.push_back(sqrt(fixed_ratios[i]));
     }
 
@@ -115,11 +115,10 @@ class DensityPriorBoxOpKernel : public framework::OpKernel<T> {
       }
     }
     if (clip) {
-      platform::Transform<platform::CPUDeviceContext> trans;
-      ClipFunctor<T> clip_func;
-      trans(ctx.template device_context<platform::CPUDeviceContext>(),
-            boxes->data<T>(), boxes->data<T>() + boxes->numel(),
-            boxes->data<T>(), clip_func);
+      T* dt = boxes->data<T>();
+      std::transform(dt, dt + boxes->numel(), dt, [](T v) -> T {
+        return std::min<T>(std::max<T>(v, 0.), 1.);
+      });
     }
     framework::Tensor var_t;
     var_t.mutable_data<T>(
@@ -141,7 +140,7 @@ class DensityPriorBoxOpKernel : public framework::OpKernel<T> {
 #pragma omp parallel for collapse(2)
 #endif
     for (int i = 0; i < box_num; ++i) {
-      for (int j = 0; j < variances.size(); ++j) {
+      for (size_t j = 0; j < variances.size(); ++j) {
         e_vars(i, j) = variances[j];
       }
     }

paddle/fluid/operators/detection/prior_box_op.h

@@ -46,13 +46,6 @@ inline void ExpandAspectRatios(const std::vector<float>& input_aspect_ratior,
   }
 }
 
-template <typename T>
-struct ClipFunctor {
-  HOSTDEVICE inline T operator()(T in) const {
-    return std::min<T>(std::max<T>(in, 0.), 1.);
-  }
-};
-
 template <typename T>
 class PriorBoxOpKernel : public framework::OpKernel<T> {
  public:
@@ -101,31 +94,30 @@ class PriorBoxOpKernel : public framework::OpKernel<T> {
     boxes->mutable_data<T>(ctx.GetPlace());
     vars->mutable_data<T>(ctx.GetPlace());
 
-    auto e_boxes = framework::EigenTensor<T, 4>::From(*boxes);
+    T* b_t = boxes->data<T>();
     for (int h = 0; h < feature_height; ++h) {
       for (int w = 0; w < feature_width; ++w) {
         T center_x = (w + offset) * step_width;
         T center_y = (h + offset) * step_height;
         T box_width, box_height;
-        int idx = 0;
         for (size_t s = 0; s < min_sizes.size(); ++s) {
           auto min_size = min_sizes[s];
           if (min_max_aspect_ratios_order) {
             box_width = box_height = min_size / 2.;
-            e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width;
-            e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height;
-            e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width;
-            e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height;
-            idx++;
+            b_t[0] = (center_x - box_width) / img_width;
+            b_t[1] = (center_y - box_height) / img_height;
+            b_t[2] = (center_x + box_width) / img_width;
+            b_t[3] = (center_y + box_height) / img_height;
+            b_t += 4;
             if (max_sizes.size() > 0) {
               auto max_size = max_sizes[s];
               // square prior with size sqrt(minSize * maxSize)
               box_width = box_height = sqrt(min_size * max_size) / 2.;
-              e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width;
-              e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height;
-              e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width;
-              e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height;
-              idx++;
+              b_t[0] = (center_x - box_width) / img_width;
+              b_t[1] = (center_y - box_height) / img_height;
+              b_t[2] = (center_x + box_width) / img_width;
+              b_t[3] = (center_y + box_height) / img_height;
+              b_t += 4;
             }
             // priors with different aspect ratios
             for (size_t r = 0; r < aspect_ratios.size(); ++r) {
@@ -135,11 +127,11 @@ class PriorBoxOpKernel : public framework::OpKernel<T> {
               }
               box_width = min_size * sqrt(ar) / 2.;
               box_height = min_size / sqrt(ar) / 2.;
-              e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width;
-              e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height;
-              e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width;
-              e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height;
-              idx++;
+              b_t[0] = (center_x - box_width) / img_width;
+              b_t[1] = (center_y - box_height) / img_height;
+              b_t[2] = (center_x + box_width) / img_width;
+              b_t[3] = (center_y + box_height) / img_height;
+              b_t += 4;
             }
           } else {
             // priors with different aspect ratios
@@ -147,21 +139,21 @@ class PriorBoxOpKernel : public framework::OpKernel<T> {
               float ar = aspect_ratios[r];
               box_width = min_size * sqrt(ar) / 2.;
               box_height = min_size / sqrt(ar) / 2.;
-              e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width;
-              e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height;
-              e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width;
-              e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height;
-              idx++;
+              b_t[0] = (center_x - box_width) / img_width;
+              b_t[1] = (center_y - box_height) / img_height;
+              b_t[2] = (center_x + box_width) / img_width;
+              b_t[3] = (center_y + box_height) / img_height;
+              b_t += 4;
             }
             if (max_sizes.size() > 0) {
               auto max_size = max_sizes[s];
               // square prior with size sqrt(minSize * maxSize)
               box_width = box_height = sqrt(min_size * max_size) / 2.;
-              e_boxes(h, w, idx, 0) = (center_x - box_width) / img_width;
-              e_boxes(h, w, idx, 1) = (center_y - box_height) / img_height;
-              e_boxes(h, w, idx, 2) = (center_x + box_width) / img_width;
-              e_boxes(h, w, idx, 3) = (center_y + box_height) / img_height;
-              idx++;
+              b_t[0] = (center_x - box_width) / img_width;
+              b_t[1] = (center_y - box_height) / img_height;
+              b_t[2] = (center_x + box_width) / img_width;
+              b_t[3] = (center_y + box_height) / img_height;
+              b_t += 4;
             }
           }
         }
@@ -169,11 +161,10 @@ class PriorBoxOpKernel : public framework::OpKernel<T> {
     }
 
     if (clip) {
-      platform::Transform<platform::CPUDeviceContext> trans;
-      ClipFunctor<T> clip_func;
-      trans(ctx.template device_context<platform::CPUDeviceContext>(),
-            boxes->data<T>(), boxes->data<T>() + boxes->numel(),
-            boxes->data<T>(), clip_func);
+      T* dt = boxes->data<T>();
+      std::transform(dt, dt + boxes->numel(), dt, [](T v) -> T {
+        return std::min<T>(std::max<T>(v, 0.), 1.);
+      });
     }
 
     framework::Tensor var_t;

paddle/fluid/operators/group_norm_op.cc

@@ -170,13 +170,48 @@ class GroupNormGradMaker : public framework::SingleGradOpDescMaker {
   }
 };
 
+class GroupNormInplaceInToOut : public framework::InplaceInToOut {
+ public:
+  using InplaceInToOut::InplaceInToOut;
+
+ protected:
+  std::unordered_map<std::string, std::string> Apply(
+      const framework::OpDesc &op_desc,
+      framework::BlockDesc *block) const override {
+    return {{"X", "Y"}};
+  }
+};
+
+class GroupNormGradInplaceInToOut : public framework::InplaceInToOut {
+ public:
+  using InplaceInToOut::InplaceInToOut;
+
+ protected:
+  std::unordered_map<std::string, std::string> Apply(
+      const framework::OpDesc &op_desc,
+      framework::BlockDesc *block) const override {
+    return {{framework::GradVarName("Y"), framework::GradVarName("X")}};
+  }
+};
+
+class GroupNormOpInferVarType
+    : public framework::PassInDtypeAndVarTypeToOutput {
+ protected:
+  std::unordered_map<std::string, std::string> GetInputOutputWithSameType()
+      const override {
+    return {{"X", /*->*/ "Y"}};
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(group_norm, ops::GroupNormOp, ops::GroupNormOpMaker,
-                  ops::GroupNormGradMaker);
-REGISTER_OPERATOR(group_norm_grad, ops::GroupNormGradOp);
+                  ops::GroupNormOpInferVarType, ops::GroupNormGradMaker,
+                  ops::GroupNormInplaceInToOut);
+REGISTER_OPERATOR(group_norm_grad, ops::GroupNormGradOp,
+                  ops::GroupNormGradInplaceInToOut);
 REGISTER_OP_CPU_KERNEL(
     group_norm, ops::GroupNormKernel<paddle::platform::CPUDeviceContext, float>,
     ops::GroupNormKernel<paddle::platform::CPUDeviceContext, double>);

paddle/fluid/operators/jit/benchmark.cc

@@ -339,6 +339,71 @@ void BenchSoftmaxKernel() {
   }
 }
 
+template <jit::KernelType KT, typename T, typename PlaceType>
+void BenchLayerNormKernel() {
+  const T epsilon = 9.99999975e-06;
+  for (int n : {1, 2, 10}) {
+    for (int x_dim_0 : {1, 9, 17, 50}) {
+      int left = n * x_dim_0;
+      for (int x_dim_1 : TestSizes()) {
+        int right = x_dim_1;
+        int sz = left * right;
+        Tensor x, mean, var, scale, bias, out;
+        x.Resize({n, x_dim_0, x_dim_1});
+        out.Resize({n, x_dim_0, x_dim_1});
+        mean.Resize({n, x_dim_0});
+        var.Resize({n, x_dim_0});
+        scale.Resize({x_dim_1});
+        bias.Resize({x_dim_1});
+
+        RandomVec<T>(sz, x.mutable_data<T>(PlaceType()), -2.f, 2.f);
+        RandomVec<T>(left, mean.mutable_data<T>(PlaceType()), -2.f, 2.f);
+        RandomVec<T>(left, var.mutable_data<T>(PlaceType()), -2.f, 2.f);
+        RandomVec<T>(right, scale.mutable_data<T>(PlaceType()), -2.f, 2.f);
+        RandomVec<T>(right, bias.mutable_data<T>(PlaceType()), -2.f, 2.f);
+
+        const T* scale_data = scale.data<T>();
+        const T* bias_data = bias.data<T>();
+        T* x_data = x.data<T>();
+        T* mean_data = mean.data<T>();
+        T* var_data = var.data<T>();
+        T* out_data = out.mutable_data<T>(PlaceType());
+
+        BenchAllImpls<KT, jit::LayerNormTuples<T>, PlaceType>(
+            right, x_data, out_data, mean_data, var_data, scale_data, bias_data,
+            left, epsilon, right);
+      }
+    }
+  }
+}
+
+template <jit::KernelType KT, typename T, typename PlaceType>
+void BenchCRFDecodingKernel() {
+  constexpr int state_trans_base_idx = 2;
+  for (int seq_len : {1, 11, 17, 50}) {
+    for (int tag_num : TestSizes()) {
+      int x_sz = seq_len * tag_num;
+      int w_sz = (tag_num + state_trans_base_idx) * tag_num;
+      Tensor x, w, alpha, track;
+      x.Resize({seq_len, tag_num});
+      w.Resize({tag_num + state_trans_base_idx, tag_num});
+      alpha.Resize({seq_len, tag_num});
+      track.Resize({seq_len, tag_num});
+
+      RandomVec<T>(x_sz, x.mutable_data<T>(PlaceType()), -2.f, 2.f);
+      RandomVec<T>(w_sz, w.mutable_data<T>(PlaceType()), -2.f, 2.f);
+
+      const T* x_data = x.data<T>();
+      const T* w_data = w.data<T>();
+      T* alpha_data = alpha.mutable_data<T>(PlaceType());
+      int* track_data = track.mutable_data<int>(PlaceType());
+
+      BenchAllImpls<KT, jit::CRFDecodingTuples<T>, PlaceType>(
+          tag_num, seq_len, x_data, w_data, alpha_data, track_data, tag_num);
+    }
+  }
+}
+
 using T = float;
 using CPUPlace = paddle::platform::CPUPlace;
 
@@ -382,6 +447,16 @@ BENCH_FP32_CPU(kMatMul) { BenchMatMulKernel<jit::kMatMul, T, CPUPlace>(); }
 // softmax
 BENCH_FP32_CPU(kSoftmax) { BenchSoftmaxKernel<jit::kSoftmax, T, CPUPlace>(); }
 
+// layernorm
+BENCH_FP32_CPU(kLayerNorm) {
+  BenchLayerNormKernel<jit::kLayerNorm, T, CPUPlace>();
+}
+
+// crfdecoding
+BENCH_FP32_CPU(kCRFDecoding) {
+  BenchCRFDecodingKernel<jit::kCRFDecoding, T, CPUPlace>();
+}
+
 // Benchmark all jit kernels including jitcode, mkl and refer.
 // To use this tool, run command: ./benchmark [options...]
 // Options:

paddle/fluid/operators/jit/test.cc

@@ -292,6 +292,63 @@ struct TestFuncWithRefer<jit::MatMulTuples<T>, std::vector<T>, std::vector<T>,
   }
 };
 
+template <typename T>
+struct TestFuncWithRefer<jit::LayerNormTuples<T>, std::vector<T>,
+                         std::vector<T>, std::vector<T>, std::vector<T>,
+                         std::vector<T>, std::vector<T>, int, float, int> {
+  void operator()(const typename jit::LayerNormTuples<T>::func_type tgt,
+                  std::vector<T>& x, std::vector<T>& outref,  // NOLINT
+                  std::vector<T>& mean, std::vector<T>& var,  // NOLINT
+                  const std::vector<T>& scale, const std::vector<T>& bias,
+                  int left, const float epsilon, int right) {
+    EXPECT_TRUE(tgt != nullptr);
+    EXPECT_EQ(x.size(), static_cast<size_t>(left * right));
+    EXPECT_EQ(outref.size(), static_cast<size_t>(left * right));
+    EXPECT_EQ(mean.size(), static_cast<size_t>(left));
+    EXPECT_EQ(var.size(), static_cast<size_t>(left));
+    EXPECT_EQ(scale.size(), static_cast<size_t>(right));
+    EXPECT_EQ(bias.size(), static_cast<size_t>(right));
+    std::vector<T> outtgt(outref.size());
+    const T* scale_data = scale.data();
+    const T* bias_data = bias.data();
+    T* x_data = x.data();
+    T* mean_data = mean.data();
+    T* var_data = var.data();
+    T* outref_data = outref.data();
+    T* outtgt_data = outtgt.data();
+
+    tgt(x_data, outtgt_data, mean_data, var_data, scale_data, bias_data, left,
+        epsilon, right);
+    ExpectEQ<T>(outtgt_data, outref_data, left * right);
+  }
+};
+
+template <typename T>
+struct TestFuncWithRefer<jit::CRFDecodingTuples<T>, int, std::vector<T>,
+                         std::vector<T>, std::vector<T>, std::vector<int>,
+                         int> {
+  void operator()(const typename jit::CRFDecodingTuples<T>::func_type tgt,
+                  const int seq_len, const std::vector<T>& x,
+                  const std::vector<T>& w, std::vector<T>& alpharef,  // NOLINT
+                  std::vector<int>& trackref, int tag_num) {          // NOLINT
+    constexpr int state_trans_base_idx = 2;
+    EXPECT_TRUE(tgt != nullptr);
+    EXPECT_EQ(x.size(), static_cast<size_t>(seq_len * tag_num));
+    EXPECT_EQ(w.size(),
+              static_cast<size_t>((tag_num + state_trans_base_idx) * tag_num));
+    EXPECT_EQ(alpharef.size(), static_cast<size_t>(seq_len * tag_num));
+    EXPECT_EQ(trackref.size(), static_cast<size_t>(seq_len * tag_num));
+    std::vector<T> alphatgt(alpharef.size());
+    std::vector<int> tracktgt(trackref.size());
+
+    memcpy(trackref.data(), tracktgt.data(), tag_num * sizeof(int));
+    tgt(seq_len, (const T*)x.data(), (const T*)w.data(), alphatgt.data(),
+        tracktgt.data(), tag_num);
+    ExpectEQ<T>(alpharef.data(), alphatgt.data(), seq_len * tag_num);
+    ExpectEQ<int>(trackref.data(), tracktgt.data(), seq_len * tag_num);
+  }
+};
+
 template <jit::KernelType KT, typename KernelTuples, typename PlaceType,
           typename... Args>
 void TestAllImpls(const typename KernelTuples::attr_type& attr, Args... args) {
@@ -640,6 +697,71 @@ void TestNCHW16CMulNCKernel() {
   }
 }
 
+template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+void TestLayerNormKernel() {
+  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
+  const T epsilon = 9.99999975e-06;
+  for (int n : {1, 2, 10}) {
+    for (int x_dim_0 : {1, 9, 17, 50}) {
+      int left = n * x_dim_0;
+      for (int x_dim_1 : TestSizes()) {
+        int right = x_dim_1;
+        auto ref = jit::GetRefer<KT, jit::LayerNormTuples<T>>();
+        EXPECT_TRUE(ref != nullptr);
+        int sz = left * right;
+        std::vector<T> x(sz), mean(left), var(left), scale(right), bias(right),
+            outref(sz);
+        RandomVec<T>(sz, x.data(), -2.f, 2.f);
+        RandomVec<T>(left, mean.data(), -2.f, 2.f);
+        RandomVec<T>(left, var.data(), -2.f, 2.f);
+        RandomVec<T>(right, scale.data(), -2.f, 2.f);
+        RandomVec<T>(right, bias.data(), -2.f, 2.f);
+
+        const T* scale_data = scale.data();
+        const T* bias_data = bias.data();
+        T* x_data = x.data();
+        T* mean_data = mean.data();
+        T* var_data = var.data();
+        T* outref_data = outref.data();
+
+        ref(x_data, outref_data, mean_data, var_data, scale_data, bias_data,
+            left, epsilon, right);
+
+        TestAllImpls<KT, jit::LayerNormTuples<T>, PlaceType, std::vector<T>,
+                     std::vector<T>, std::vector<T>, std::vector<T>,
+                     std::vector<T>, std::vector<T>, int, float>(
+            right, x, outref, mean, var, scale, bias, left, epsilon, right);
+      }
+    }
+  }
+}
+
+template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
+void TestCRFDecodingKernel() {
+  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
+  constexpr int state_trans_base_idx = 2;
+  for (int seq_len : {1, 11, 17, 50}) {
+    for (int tag_num : TestSizes()) {
+      auto ref = jit::GetRefer<KT, jit::CRFDecodingTuples<T>>();
+      EXPECT_TRUE(ref != nullptr);
+      int x_sz = seq_len * tag_num;
+      int w_sz = (tag_num + state_trans_base_idx) * tag_num;
+      std::vector<T> x(x_sz), w(w_sz), alpharef(x_sz);
+      std::vector<int> trackref(x_sz);
+      RandomVec<T>(x_sz, x.data(), -2.f, 2.f);
+      RandomVec<T>(w_sz, w.data(), -2.f, 2.f);
+
+      ref(seq_len, (const T*)x.data(), (const T*)w.data(), alpharef.data(),
+          trackref.data(), tag_num);
+
+      TestAllImpls<KT, jit::CRFDecodingTuples<T>, PlaceType, int,
+                   std::vector<T>, std::vector<T>, std::vector<T>,
+                   std::vector<int>, int>(tag_num, seq_len, x, w, alpharef,
+                                          trackref, tag_num);
+    }
+  }
+}
+
 // XYZNTuple
 TEST(JITKernel, kVMul) {
   TestXYZNKernel<jit::kVMul, float, CPUPlace>();
@@ -761,7 +883,16 @@ TEST(JITKernel, kNCHW16CMulNC) {
   TestNCHW16CMulNCKernel<jit::kNCHW16CMulNC, double, CPUPlace>();
 }
 
-// TODO(yihua/TJ): add crf decoding and layer norm unit tests
+TEST(JITKernel, kLayerNorm) {
+  TestLayerNormKernel<jit::kLayerNorm, float, paddle::platform::CPUPlace>();
+  TestLayerNormKernel<jit::kLayerNorm, double, paddle::platform::CPUPlace>();
+}
+
+TEST(JITKernel, kCRFDecoding) {
+  TestCRFDecodingKernel<jit::kCRFDecoding, float, paddle::platform::CPUPlace>();
+  TestCRFDecodingKernel<jit::kCRFDecoding, double,
+                        paddle::platform::CPUPlace>();
+}
 
 TEST(JITKernel, pool) {
   // TODO(TJ): add some test

paddle/fluid/operators/load_combine_op.cc

@@ -64,7 +64,7 @@ class LoadCombineOp : public framework::OperatorBase {
       auto *tensor = out_var->GetMutable<framework::LoDTensor>();
 
       // Error checking
-      PADDLE_ENFORCE(static_cast<bool>(buffer), "Cannot read more");
+      PADDLE_ENFORCE(static_cast<bool>(*buffer), "Cannot read more");
 
       // Get data from fin to tensor
       DeserializeFromStream(*buffer, tensor, dev_ctx);
@@ -90,6 +90,10 @@ class LoadCombineOp : public framework::OperatorBase {
         tensor->ShareDataWith(fp16_tensor);
       }
     }
+    buffer->peek();
+    PADDLE_ENFORCE(buffer->eof(),
+                   "You are not allowed to load partial data via "
+                   "load_combine_op, use load_op instead.");
   }
 };
 

paddle/fluid/operators/lstm_op.h

@@ -311,6 +311,10 @@ class LSTMGradKernel : public framework::OpKernel<T> {
         lstm_grad.prev_state_grad = c0_g ? ordered_c0_g.data<T>() : nullptr;
       }
 
+      // lstm_value.output_value not used in bp, set to nullptr
+      // lstm_grad.state_active_grad not used in bp, set to nullptr
+      lstm_value.output_value = nullptr;
+      lstm_grad.state_active_grad = nullptr;
       int cur_batch_size = bend - bstart;
       math::LstmUnitGradFunctor<DeviceContext, T>::compute(
           device_ctx, lstm_value, lstm_grad, frame_size, cur_batch_size,

paddle/fluid/operators/lstmp_op.h

@@ -405,6 +405,11 @@ class LSTMPGradKernel : public framework::OpKernel<T> {
       }
 
       int cur_batch_size = bend - bstart;
+      // lstmp_value.output_value not used in bp, set to null
+      // lstmp_grad.state_active_grad not used in bp, set to null
+      lstmp_value.output_value = nullptr;
+      lstmp_grad.state_active_grad = nullptr;
+
       math::LstmUnitGradFunctor<DeviceContext, T>::compute(
           device_ctx, lstmp_value, lstmp_grad, frame_size, cur_batch_size,
           gate_act, cell_act, cand_act);

paddle/fluid/operators/row_conv_op.cc

@@ -109,23 +109,23 @@ from future subsequences in a computationally efficient manner to improve
 unidirectional recurrent neural networks. The row convolution operator is 
 different from the 1D sequence convolution, and is computed as follows:
 
-Given an input sequence $in$ of length $t$ and input dimension $d$, 
-and a filter ($W$) of size $context \times d$, 
+Given an input sequence $X$ of length $t$ and input dimension $D$, 
+and a filter ($W$) of size $context \times D$,
 the output sequence is convolved as:
 
 $$
-out_{i, :} = \\sum_{j=i}^{i + context} in_{j,:} \\cdot W_{i-j, :}
+out_{i} = \\sum_{j=i}^{i + context - 1} X_{j} \\cdot W_{j-i}
 $$
 
 In the above equation:
 
 * $Out_{i}$: The i-th row of output variable with shape [1, D].
 
-* $\\tau$: Future context size.
+* $context$: Future context size.
 
 * $X_{j}$: The j-th row of input variable with shape [1, D].
 
-* $W_{i-j}$: The (i-j)-th row of parameters with shape [1, D].
+* $W_{j-i}$: The (j-i)-th row of parameters with shape [1, D].
 
 More details about row_conv please refer to
 the design document

paddle/fluid/platform/enforce.h

@@ -234,8 +234,10 @@ inline void throw_on_error(ncclResult_t stat, const std::string& msg) {
 #endif  // PADDLE_WITH_CUDA
 
 #define PADDLE_THROW(...)                                            \
+  do {                                                               \
     throw ::paddle::platform::EnforceNotMet(                         \
-      ::paddle::string::Sprintf(__VA_ARGS__), __FILE__, __LINE__)
+        ::paddle::string::Sprintf(__VA_ARGS__), __FILE__, __LINE__); \
+  } while (0)
 
 #define PADDLE_ENFORCE(COND, ...)                                         \
   do {                                                                    \
@@ -274,19 +276,21 @@ inline void throw_on_error(ncclResult_t stat, const std::string& msg) {
   do {                                                      \
     if (UNLIKELY(nullptr == (__VAL))) {                     \
       PADDLE_THROW(#__VAL " should not be null\n%s",        \
-                   paddle::string::Sprintf("" __VA_ARGS__)); \
+                   ::paddle::string::Sprintf(__VA_ARGS__)); \
     }                                                       \
   } while (0)
 
 #define __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, __CMP, __INV_CMP, ...)  \
   do {                                                                  \
-    if (UNLIKELY(!((__VAL0)__CMP(__VAL1)))) {                           \
+    auto __cond1__ = (__VAL0);                                          \
+    auto __cond2__ = (__VAL1);                                          \
+    if (UNLIKELY(!((__cond1__)__CMP(__cond2__)))) {                     \
       PADDLE_THROW("Enforce failed. Expected %s " #__CMP                \
                    " %s, but received %s:%s " #__INV_CMP " %s:%s.\n%s", \
                    #__VAL0, #__VAL1, #__VAL0,                           \
-                   paddle::string::to_string(__VAL0), #__VAL1,          \
-                   paddle::string::to_string(__VAL1),                   \
-                   paddle::string::Sprintf("" __VA_ARGS__));            \
+                   ::paddle::string::to_string(__cond1__), #__VAL1,     \
+                   ::paddle::string::to_string(__cond2__),              \
+                   ::paddle::string::Sprintf(__VA_ARGS__));             \
     }                                                                   \
   } while (0)
 

paddle/fluid/pybind/ir.cc

@@ -13,10 +13,12 @@
 // limitations under the License.
 
 #include "paddle/fluid/pybind/ir.h"
+#include <algorithm>
 #include <string>
 #include <unordered_map>
 #include <unordered_set>
 #include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
 #include "paddle/fluid/framework/ir/graph_pattern_detector.h"
 #include "paddle/fluid/framework/ir/node.h"
 #include "paddle/fluid/framework/op_desc.h"
@@ -27,6 +29,10 @@ namespace py = pybind11;
 using paddle::framework::ir::Graph;
 using paddle::framework::ir::Node;
 using paddle::framework::ir::GraphSafeRemoveNodes;
+using paddle::framework::ir::HasCircle;
+using paddle::framework::ir::GraphNum;
+using paddle::framework::ir::TopologySortOperations;
+using paddle::framework::ir::BuildOperationAdjList;
 using paddle::framework::OpDesc;
 using paddle::framework::ProgramDesc;
 using paddle::framework::VarDesc;
@@ -36,6 +42,12 @@ namespace paddle {
 namespace pybind {
 void BindGraph(py::module *m) {
   m->def("graph_safe_remove_nodes", GraphSafeRemoveNodes);
+  m->def("has_circle", HasCircle);
+  m->def("graph_num", GraphNum);
+  m->def("topology_sort", TopologySortOperations,
+         return_value_policy::reference);
+  m->def("build_adjacency_list", BuildOperationAdjList,
+         return_value_policy::reference);
   py::class_<Graph, std::shared_ptr<Graph>>(
       *m, "Graph",
       "The graph is a Directed Acyclic Single Static Assignment Graph, see "
@@ -46,7 +58,6 @@ void BindGraph(py::module *m) {
       .def("get_float", &Graph::Get<float>)
       .def("get_double", &Graph::Get<double>)
       .def("get_string", &Graph::Get<std::string>)
-      .def("get_program", &Graph::Get<ProgramDesc>)
       .def("get_marked_nodes", &Graph::Get<std::unordered_set<const Node *>>)
       .def("set", [](Graph &self, const std::string &attr_name,
                      int attr) { return self.Set(attr_name, new int(attr)); })
@@ -63,11 +74,6 @@ void BindGraph(py::module *m) {
            [](Graph &self, const std::string &attr_name, double attr) {
              return self.Set(attr_name, new double(attr));
            })
-      .def("set",
-           [](Graph &self, const std::string &attr_name,
-              const ProgramDesc &attr) {
-             return self.Set(attr_name, new ProgramDesc(attr));
-           })
       .def("set",
            [](Graph &self, const std::string &attr_name,
               const std::unordered_set<const Node *> &attr) {
@@ -108,42 +114,42 @@ void BindNode(py::module *m) {
       .def("is_op", &Node::IsOp)
       .def("is_var", &Node::IsVar)
       .def("is_ctrl_var", &Node::IsCtrlVar)
+      .def("clear_inputs", [](Node &self) { self.inputs.clear(); })
       .def("inputs_remove",
            [](Node &self, int node_id) {
-             for (auto it = self.inputs.begin(); it != self.inputs.end();
-                  it++) {
-               if ((*it)->id() == node_id) {
-                 self.inputs.erase(it);
-               }
+             auto pos = std::find_if(
+                 self.inputs.begin(), self.inputs.end(),
+                 [&node_id](const Node *n) { return n->id() == node_id; });
+             if (pos != self.inputs.end()) {
+               self.inputs.erase(pos);
              }
            })
       .def("inputs_remove",
            [](Node &self, Node &node) {
-             for (auto it = self.inputs.begin(); it != self.inputs.end();
-                  it++) {
-               if (*it == &node) {
-                 self.inputs.erase(it);
-               }
+             auto pos =
+                 std::find(self.inputs.begin(), self.inputs.end(), &node);
+             if (pos != self.inputs.end()) {
+               self.inputs.erase(pos);
              }
            })
       .def("inputs_append",
            [](Node &self, Node &node) { self.inputs.push_back(&node); })
+      .def("clear_outputs", [](Node &self) { self.outputs.clear(); })
       .def("outputs_remove",
            [](Node &self, int node_id) {
-             for (auto it = self.outputs.begin(); it != self.outputs.end();
-                  it++) {
-               if ((*it)->id() == node_id) {
-                 self.outputs.erase(it);
-               }
+             auto pos = std::find_if(
+                 self.outputs.begin(), self.outputs.end(),
+                 [&node_id](const Node *n) { return n->id() == node_id; });
+             if (pos != self.outputs.end()) {
+               self.outputs.erase(pos);
              }
            })
       .def("outputs_remove",
            [](Node &self, Node &node) {
-             for (auto it = self.outputs.begin(); it != self.outputs.end();
-                  it++) {
-               if (*it == &node) {
-                 self.outputs.erase(it);
-               }
+             auto pos =
+                 std::find(self.outputs.begin(), self.outputs.end(), &node);
+             if (pos != self.outputs.end()) {
+               self.outputs.erase(pos);
              }
            })
       .def("outputs_append",

paddle/fluid/pybind/pybind.cc

@@ -829,8 +829,7 @@ All parameter, weight, gradient are variables in Paddle.
   m.def("disable_profiler", platform::DisableProfiler);
   m.def("is_profiler_enabled", platform::IsProfileEnabled);
   m.def("reset_profiler", platform::ResetProfiler);
-  m.def("get_pass", [](const py::bytes &binary_str) {
-    std::string pass_type(binary_str);
+  m.def("get_pass", [](const std::string &pass_type) {
     auto pass = framework::ir::PassRegistry::Instance().Get(pass_type);
     return std::shared_ptr<framework::ir::Pass>(std::move(pass));
   });
@@ -838,10 +837,9 @@ All parameter, weight, gradient are variables in Paddle.
   py::class_<ir::Pass, std::shared_ptr<ir::Pass>> pass(m, "Pass");
   pass.def(py::init())
       .def("has", &ir::Pass::Has)
-      .def("set",
-           [](ir::Pass &self, const std::string &attr_name,
-              const ProgramDesc &attr) {
-             return self.Set(attr_name, new ProgramDesc(attr));
+      .def("set_not_owned",
+           [](ir::Pass &self, const std::string &attr_name, ProgramDesc &attr) {
+             self.SetNotOwned<ProgramDesc>(attr_name, &attr);
            })
       .def(
           "set",
@@ -850,7 +848,6 @@ All parameter, weight, gradient are variables in Paddle.
           })
       .def("set", [](ir::Pass &self, const std::string &name,
                      int val) { self.Set<const int>(name, new int(val)); })
-      .def("get_program", &ir::Pass::Get<ProgramDesc>)
       .def("type", &ir::Pass::Type)
       .def("apply", [](ir::Pass &self, std::shared_ptr<ir::Graph> graph) {
         std::unique_ptr<ir::Graph> origin_graph(graph.get());

python/CMakeLists.txt

@@ -64,6 +64,7 @@ if (WITH_TESTING)
   add_subdirectory(paddle/dataset/tests)
   add_subdirectory(paddle/fluid/tests)
   add_subdirectory(paddle/fluid/contrib/tests)
+  add_subdirectory(paddle/fluid/contrib/slim/tests)
 endif()
 install(DIRECTORY ${PADDLE_PYTHON_PACKAGE_DIR}
     DESTINATION opt/paddle/share/wheels

python/paddle/fluid/compiler.py

@@ -177,7 +177,10 @@ class CompiledProgram(object):
 
         # FIXME(dzhwinter): enable_inplace should be after memory_optimize
         # if turn on python memory optimize, turn off the inplace_pass.
-        self._build_strategy.enable_inplace = False if self._program._is_mem_optimized else True
+        if self._build_strategy.memory_optimize is None:
+            self._build_strategy.memory_optimize = False if main._is_mem_optimized else True
+        if self._build_strategy.enable_inplace is None:
+            self._build_strategy.enable_inplace = False if main._is_mem_optimized else True
 
         if self._build_strategy.num_trainers > 1 and trainers_endpoints:
             assert self._build_strategy.num_trainers == len(

python/paddle/fluid/contrib/int8_inference/README.md

@@ -63,10 +63,10 @@ Notes:
 ## 4. How to reproduce the results
 * Small dataset
 ```bash
-python python/paddle/fluid/contrib/tests/test_calibration.py
+FLAGS_use_mkldnn=true python python/paddle/fluid/contrib/tests/test_calibration.py
 ```
 
 * Full dataset
 ```bash
-DATASET=full python python/paddle/fluid/contrib/tests/test_calibration.py
+FLAGS_use_mkldnn=true DATASET=full python python/paddle/fluid/contrib/tests/test_calibration.py
 ```

python/paddle/fluid/contrib/slim/tests/CMakeLists.txt

+file(GLOB TEST_OPS RELATIVE "${CMAKE_CURRENT_SOURCE_DIR}" "test_*.py")
+string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}")
+
+foreach(src ${TEST_OPS})
+    py_test(${src} SRCS ${src}.py)
+endforeach()

python/paddle/fluid/contrib/slim/unitest/configs/config.yaml → python/paddle/fluid/contrib/slim/tests/configs/config.yaml

 version: 1.0
-include: ["./unitest/configs/pruners.yaml", "./unitest/configs/pruners_0.yaml"]
+include: ["./configs/pruners.yaml", "./configs/pruners_0.yaml"]
 pruners:
     pruner_1:
         class: 'RatioPruner'

python/paddle/fluid/contrib/slim/unitest/test_factory.py → python/paddle/fluid/contrib/slim/tests/test_factory.py

@@ -18,7 +18,7 @@ import unittest
 
 class TestFactory(unittest.TestCase):
     def test_parse(self):
-        factory = ConfigFactory('./unitest/configs/config.yaml')
+        factory = ConfigFactory('./configs/config.yaml')
 
         pruner = factory.instance('pruner_1')
         self.assertEquals(pruner.ratios['conv1_1.w'], 0.3)

python/paddle/fluid/contrib/slim/tests/test_graph.py

+#   copyright (c) 2018 paddlepaddle authors. all rights reserved.
+#
+# licensed under the apache license, version 2.0 (the "license");
+# you may not use this file except in compliance with the license.
+# you may obtain a copy of the license at
+#
+#     http://www.apache.org/licenses/license-2.0
+#
+# unless required by applicable law or agreed to in writing, software
+# distributed under the license is distributed on an "as is" basis,
+# without warranties or conditions of any kind, either express or implied.
+# see the license for the specific language governing permissions and
+# limitations under the license.
+
+from __future__ import print_function
+import unittest
+import paddle.fluid as fluid
+import six
+from paddle.fluid.framework import IrGraph
+from paddle.fluid import core
+
+
+def residual_block(num):
+    def conv_bn_layer(input,
+                      ch_out,
+                      filter_size,
+                      stride,
+                      padding,
+                      act='relu',
+                      bias_attr=False):
+        tmp = fluid.layers.conv2d(
+            input=input,
+            filter_size=filter_size,
+            num_filters=ch_out,
+            stride=stride,
+            padding=padding,
+            act=None,
+            bias_attr=bias_attr)
+        return fluid.layers.batch_norm(input=tmp, act=act)
+
+    data = fluid.layers.data(name='image', shape=[1, 32, 32], dtype='float32')
+    label = fluid.layers.data(name='label', shape=[1], dtype='int64')
+    hidden = data
+    for _ in six.moves.xrange(num):
+        conv = conv_bn_layer(hidden, 16, 3, 1, 1, act=None, bias_attr=True)
+        short = conv_bn_layer(hidden, 16, 1, 1, 0, act=None)
+        hidden = fluid.layers.elementwise_add(x=conv, y=short, act='relu')
+    fc = fluid.layers.fc(input=hidden, size=10)
+    loss = fluid.layers.cross_entropy(input=fc, label=label)
+    loss = fluid.layers.mean(loss)
+    return loss
+
+
+class TestGraph(unittest.TestCase):
+    def test_graph_functions(self):
+        main = fluid.Program()
+        startup = fluid.Program()
+        with fluid.program_guard(main, startup):
+            loss = residual_block(2)
+            opt = fluid.optimizer.Adam(learning_rate=0.001)
+            opt.minimize(loss)
+        graph = IrGraph(core.Graph(main.desc), for_test=False)
+        marked_nodes = set()
+        for op in graph.all_ops():
+            if op.name().find('conv2d') > -1:
+                marked_nodes.add(op)
+        graph.draw('.', 'residual', marked_nodes)
+        self.assertFalse(graph.has_circle())
+        self.assertEqual(graph.graph_num(), 1)
+        nodes = graph.topology_sort()
+        self.assertEqual(len(nodes), len(graph.all_ops()))
+        nodes_map = graph.build_adjacency_list()
+        self.assertEqual(len(nodes_map), len(graph.all_ops()))
+        nodes_num = len(graph.all_nodes())
+        graph.safe_remove_nodes(marked_nodes)
+        self.assertEqual(len(graph.all_nodes()), nodes_num - len(marked_nodes))
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/contrib/slim/unitest/test_quantization_pass.py → python/paddle/fluid/contrib/slim/tests/test_quantization_pass.py

@@ -17,9 +17,12 @@ import random
 import numpy as np
 import paddle.fluid as fluid
 import six
-from paddle.fluid.framework import Program
+import paddle
 from paddle.fluid.framework import IrGraph
 from paddle.fluid.contrib.slim.quantization import QuantizationTransformPass
+from paddle.fluid.contrib.slim.quantization import QuantizationFreezePass
+from paddle.fluid.contrib.slim.quantization import ConvertToInt8Pass
+from paddle.fluid.contrib.slim.quantization import TransformForMobilePass
 from paddle.fluid import core
 
 
@@ -65,6 +68,28 @@ def residual_block(num):
     return loss
 
 
+def conv_net(img, label):
+    conv_pool_1 = fluid.nets.simple_img_conv_pool(
+        input=img,
+        filter_size=5,
+        num_filters=20,
+        pool_size=2,
+        pool_stride=2,
+        act="relu")
+    conv_pool_1 = fluid.layers.batch_norm(conv_pool_1)
+    conv_pool_2 = fluid.nets.simple_img_conv_pool(
+        input=conv_pool_1,
+        filter_size=5,
+        num_filters=50,
+        pool_size=2,
+        pool_stride=2,
+        act="relu")
+    prediction = fluid.layers.fc(input=conv_pool_2, size=10, act='softmax')
+    loss = fluid.layers.cross_entropy(input=prediction, label=label)
+    avg_loss = fluid.layers.mean(loss)
+    return avg_loss
+
+
 class TestQuantizationTransformPass(unittest.TestCase):
     def setUp(self):
         self.quantizable_op_and_inputs = {
@@ -171,5 +196,177 @@ class TestQuantizationTransformPass(unittest.TestCase):
         self.residual_block_quant('range_abs_max')
 
 
+class TestQuantizationFreezePass(unittest.TestCase):
+    def freeze_graph(self, use_cuda, seed, quant_type):
+        def build_program(main, startup, is_test):
+            main.random_seed = seed
+            startup.random_seed = seed
+            with fluid.unique_name.guard():
+                with fluid.program_guard(main, startup):
+                    img = fluid.layers.data(
+                        name='image', shape=[1, 28, 28], dtype='float32')
+                    label = fluid.layers.data(
+                        name='label', shape=[1], dtype='int64')
+                    loss = conv_net(img, label)
+                    if not is_test:
+                        opt = fluid.optimizer.Adam(learning_rate=0.001)
+                        opt.minimize(loss)
+            return [img, label], loss
+
+        random.seed(0)
+        np.random.seed(0)
+
+        main = fluid.Program()
+        startup = fluid.Program()
+        test_program = fluid.Program()
+        feeds, loss = build_program(main, startup, False)
+        build_program(test_program, startup, True)
+        test_program = test_program.clone(for_test=True)
+        main_graph = IrGraph(core.Graph(main.desc), for_test=False)
+        test_graph = IrGraph(core.Graph(test_program.desc), for_test=True)
+
+        place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
+        exe = fluid.Executor(place)
+        scope = fluid.Scope()
+        with fluid.scope_guard(scope):
+            exe.run(startup)
+        transform_pass = QuantizationTransformPass(
+            scope=scope, program_exe=exe, activation_quantize_type=quant_type)
+        transform_pass.apply(main_graph)
+        transform_pass.apply(test_graph)
+        dev_name = '_gpu_' if use_cuda else '_cpu_'
+        marked_nodes = set()
+        for op in main_graph.all_ops():
+            if op.name().find('quantize') > -1:
+                marked_nodes.add(op)
+        main_graph.draw('.', 'main' + dev_name + quant_type, marked_nodes)
+        marked_nodes = set()
+        for op in test_graph.all_ops():
+            if op.name().find('quantize') > -1:
+                marked_nodes.add(op)
+        test_graph.draw('.', 'test' + dev_name + quant_type, marked_nodes)
+
+        quantized_main_program = main_graph.to_program()
+        quantized_test_program = test_graph.to_program()
+        iters = 5
+        batch_size = 8
+
+        #train_exe = fluid.ParallelExecutor(
+        #    main_program=quantized_main_program,
+        #    use_cuda=bool(use_cuda),
+        #    loss_name=loss.name,
+        #    scope=scope)
+        train_reader = paddle.batch(
+            paddle.reader.shuffle(
+                paddle.dataset.mnist.train(), buf_size=500),
+            batch_size=batch_size)
+        test_reader = paddle.batch(
+            paddle.dataset.mnist.test(), batch_size=batch_size)
+        feeder = fluid.DataFeeder(feed_list=feeds, place=place)
+        with fluid.scope_guard(scope):
+            for _ in range(iters):
+                data = next(train_reader())
+                loss_v = exe.run(program=quantized_main_program,
+                                 feed=feeder.feed(data),
+                                 fetch_list=[loss])
+                #loss_v = train_exe.run(feed=feeder.feed(data),
+                #                       fetch_list=[loss.name])
+                #print('{}: {}'.format('loss' + dev_name + quant_type, loss_v))
+
+        test_data = next(test_reader())
+        with fluid.program_guard(quantized_test_program):
+            w_var = fluid.framework._get_var('conv2d_1.w_0.quantized',
+                                             quantized_test_program)
+        # Testing
+        with fluid.scope_guard(scope):
+            test_loss1, w_quant = exe.run(program=quantized_test_program,
+                                          feed=feeder.feed(test_data),
+                                          fetch_list=[loss, w_var])
+
+        # Freeze graph for inference, but the weight of fc/conv is still float type.
+        freeze_pass = QuantizationFreezePass(scope=scope, place=place)
+        freeze_pass.apply(test_graph)
+        marked_nodes = set()
+        for op in test_graph.all_ops():
+            if op.name().find('quantize') > -1:
+                marked_nodes.add(op)
+        test_graph.draw('.', 'test_freeze' + dev_name + quant_type,
+                        marked_nodes)
+
+        server_program = test_graph.to_program()
+        with fluid.scope_guard(scope):
+            test_loss2, = exe.run(program=server_program,
+                                  feed=feeder.feed(test_data),
+                                  fetch_list=[loss])
+        self.assertAlmostEqual(test_loss1, test_loss2, delta=5e-3)
+        #print('{}: {}'.format('test_loss1' + dev_name + quant_type, test_loss1))
+        #print('{}: {}'.format('test_loss2' + dev_name + quant_type, test_loss2))
+        w_freeze = np.array(scope.find_var('conv2d_1.w_0').get_tensor())
+        # Maybe failed, this is due to the calculation precision
+        # self.assertAlmostEqual(np.sum(w_freeze), np.sum(w_quant))
+        #print('{}: {}'.format('w_freeze' + dev_name + quant_type,
+        #                      np.sum(w_freeze)))
+        #print('{}: {}'.format('w_quant' + dev_name + quant_type,
+        #                      np.sum(w_quant)))
+
+        # Convert parameter to 8-bit.
+        convert_int8_pass = ConvertToInt8Pass(scope=scope, place=place)
+        convert_int8_pass.apply(test_graph)
+        marked_nodes = set()
+        for op in test_graph.all_ops():
+            if op.name().find('quantize') > -1:
+                marked_nodes.add(op)
+        test_graph.draw('.', 'test_int8' + dev_name + quant_type, marked_nodes)
+        server_program_int8 = test_graph.to_program()
+        # Save the 8-bit parameter and model file.
+        with fluid.scope_guard(scope):
+            fluid.io.save_inference_model('server_int8' + dev_name + quant_type,
+                                          ['image', 'label'], [loss], exe,
+                                          server_program_int8)
+            # Test whether the 8-bit parameter and model file can be loaded successfully.
+            [infer, feed, fetch] = fluid.io.load_inference_model(
+                'server_int8' + dev_name + quant_type, exe)
+        # Check the loaded 8-bit weight.
+        w_8bit = np.array(scope.find_var('conv2d_1.w_0.int8').get_tensor())
+        self.assertEqual(w_8bit.dtype, np.int8)
+        self.assertEqual(np.sum(w_8bit), np.sum(w_freeze))
+        #print('{}: {}'.format('w_8bit' + dev_name + quant_type, np.sum(w_8bit)))
+        #print('{}: {}'.format('w_freeze' + dev_name + quant_type,
+        #                      np.sum(w_freeze)))
+
+        mobile_pass = TransformForMobilePass()
+        mobile_pass.apply(test_graph)
+        marked_nodes = set()
+        for op in test_graph.all_ops():
+            if op.name().find('quantize') > -1:
+                marked_nodes.add(op)
+        test_graph.draw('.', 'test_mobile' + dev_name + quant_type,
+                        marked_nodes)
+
+        mobile_program = test_graph.to_program()
+        with fluid.scope_guard(scope):
+            fluid.io.save_inference_model('mobile_int8' + dev_name + quant_type,
+                                          ['image', 'label'], [loss], exe,
+                                          mobile_program)
+
+    def test_freeze_graph_cuda_dynamic(self):
+        if fluid.core.is_compiled_with_cuda():
+            with fluid.unique_name.guard():
+                self.freeze_graph(True, seed=1, quant_type='abs_max')
+
+    def test_freeze_graph_cpu_dynamic(self):
+        with fluid.unique_name.guard():
+            self.freeze_graph(False, seed=2, quant_type='abs_max')
+
+    def test_freeze_graph_cuda_static(self):
+        if fluid.core.is_compiled_with_cuda():
+            with fluid.unique_name.guard():
+                self.freeze_graph(True, seed=1, quant_type='range_abs_max')
+
+    def test_freeze_graph_cpu_static(self):
+        with fluid.unique_name.guard():
+            self.freeze_graph(False, seed=2, quant_type='range_abs_max')
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/contrib/tests/CMakeLists.txt

@@ -6,5 +6,9 @@ if(APPLE OR WIN32 OR NOT WITH_MKL)
 endif()
 
 foreach(src ${TEST_OPS})
+    if(src MATCHES "test_calibration")
+        py_test(${src} SRCS ${src}.py ENVS FLAGS_use_mkldnn=true)
+    else()
         py_test(${src} SRCS ${src}.py)
+    endif()
 endforeach()

python/paddle/fluid/contrib/tests/test_calibration.py

@@ -199,7 +199,6 @@ class TestCalibrationForResnet50(unittest.TestCase):
 
     def run_program(self, model_path, generate_int8=False, algo='direct'):
         image_shape = [3, 224, 224]
-        os.environ['FLAGS_use_mkldnn'] = 'True'
 
         fluid.memory_optimize(fluid.default_main_program())
 
@@ -241,9 +240,6 @@ class TestCalibrationForResnet50(unittest.TestCase):
             label = label.reshape([-1, 1])
             running_program = calibrator.sampling_program.clone(
             ) if generate_int8 else infer_program.clone()
-            for op in running_program.current_block().ops:
-                if op.has_attr("use_mkldnn"):
-                    op._set_attr("use_mkldnn", True)
 
             t1 = time.time()
             _, acc1, _ = exe.run(

python/paddle/fluid/contrib/tests/test_quantize_transpiler.py

@@ -204,9 +204,11 @@ class TestQuantizeTranspiler(unittest.TestCase):
         build_program(test_program, startup, True)
         test_program = test_program.clone(for_test=True)
 
-        quant_transpiler = QuantizeTranspiler()
-        quant_transpiler.training_transpile(main)
-        quant_transpiler.training_transpile(test_program)
+        quant_type = 'range_abs_max'  # 'range_abs_max' or 'abs_max'
+        quant_transpiler = QuantizeTranspiler(
+            activation_quantize_type=quant_type)
+        quant_transpiler.training_transpile(main, startup)
+        quant_transpiler.training_transpile(test_program, startup)
 
         place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
         exe = fluid.Executor(place)

python/paddle/fluid/framework.py

@@ -16,6 +16,8 @@ from __future__ import print_function
 
 import collections
 from collections import defaultdict
+from collections import Iterable
+import contextlib
 from .wrapped_decorator import signature_safe_contextmanager
 import os
 import re
@@ -1529,12 +1531,16 @@ class Block(object):
 
 class IrGraph(object):
     """
-    IrGraph uses core.Graph as the delegation to accomplish the manipulation.
+    Python IrGraph. Beneath it is a core.Graph, which is used for
+    create a c++ Ir Pass Graph. An IrGraph is just a graph view of
+    a Program. In an IrGraph, both Variables and Operators are graph
+    nodes.
     """
 
     def __init__(self, graph, for_test=False):
         """
-        Construct the IrGraph using core.Graph.
+        Construct an IrGraph using core.Graph.
+
         Args:
             graph(core.Graph): C++ Graph.
             for_test(bool): True for the test graph and false for the train graph.
@@ -1545,23 +1551,81 @@ class IrGraph(object):
         self._for_test = for_test
 
     def is_test(self):
+        """
+        If the graph is used for testing, the function returns true. Otherwise, returns false.
+        """
         return self._for_test
 
-    def all_parameters(self):
-        param_nodes = set()
-        for node in self.graph.nodes():
-            if node.is_var() and node.var() is not None and node.var(
-            ).persistable():
-                param_nodes.add(node)
-        return param_nodes
+    def all_nodes(self):
+        """
+        Return all nodes included in the graph as a set.
+        """
+        return {node for node in self.graph.nodes()}
 
     def all_vars(self):
+        """
+        Return all variable nodes included in the graph as a set.
+        """
         return {node for node in self.graph.nodes() if node.is_var()}
 
+    def all_persistable_vars(self):
+        """
+        Return all persistable variable nodes included in the graph as a set.
+        """
+        persistable_nodes = set()
+        for node in self.graph.nodes():
+            if node.is_var() and node.var() is not None and node.var(
+            ).persistable():
+                persistable_nodes.add(node)
+        return persistable_nodes
+
     def all_ops(self):
+        """
+        Return all operator nodes included in the graph as a set.
+        """
         return {node for node in self.graph.nodes() if node.is_op()}
 
+    def var_node(self, name):
+        """
+        Get a variable node by name from the graph.
+
+        Args:
+            name(str): the name of the variable node.
+
+        Raises:
+            ValueError: The If input's type is not str, or this graph
+            doesn't have a variable with the giving name.
+
+        Returns:
+            core.Node: the variable node with the giving name.
+        """
+        if not isinstance(name, six.string_types):
+            raise TypeError(
+                "var require string as parameter, but get %s instead." %
+                (type(name)))
+        target_var_node = None
+        var_nodes = self.all_vars()
+        for var_node in var_nodes:
+            if var_node.name() == name:
+                target_var_node = var_node
+        if target_var_node is None:
+            raise ValueError("var_node %s not in this graph" % name)
+        return target_var_node
+
     def create_param_node(self, name, var_type, shape, var_dtype):
+        """
+        Create a persistable variable node in the graph. In IrGraph,
+        it can not distinguish between persistable variables and parameters.
+
+        Args:
+            name(str): the name of the persistable variable node.
+            vart_type(core.VarDesc.VarType): the type of the persistable variable node.
+            shape(list): the shape of the persistable variable node.
+            var_dtype(core.VarDesc.VarType): the data type of the persistable variable node.
+
+        Returns:
+            core.Node: the created persistable variable node.
+        """
         var_desc = core.VarDesc(name)
         var_desc.set_type(var_type)
         var_desc.set_shape(shape)
@@ -1570,6 +1634,20 @@ class IrGraph(object):
         return self.graph.create_var_node(var_desc)
 
     def create_var_node(self, name, var_type, shape, var_dtype):
+        """
+        Create a variable node in the graph. The created variable node is
+        not persistable.
+
+        Args:
+            name(str): the name of the variable node.
+            vart_type(core.VarDesc.VarType): the type of the variable node.
+            shape(list): the shape of the variable node.
+            var_dtype(core.VarDesc.VarType): the data type of the variable node.
+
+        Returns:
+            core.Node: the created variable node.
+        """
+
         var_desc = core.VarDesc(name)
         var_desc.set_type(var_type)
         var_desc.set_shape(shape)
@@ -1577,19 +1655,41 @@ class IrGraph(object):
         return self.graph.create_var_node(var_desc)
 
     def create_var_node_from_desc(self, var_desc):
+        """
+        Create a variable node by using an existing VarDesc in the graph.
+        Depend on the giving VarDesc, the created variable node may be persistable.
+
+        Args:
+            var_desc(core.VarDesc): the giving variable description.
+
+        Returns:
+            core.Node: the created variable node.
+        """
         return self.graph.create_var_node(var_desc)
 
     def create_op_node(self, op_type, attrs, inputs, outputs):
+        """
+        Create a operator node in the graph.
+
+        Args:
+            op_type(str): the type of the operator node.
+            attrs(dict): the attributes of the operator node.
+            inputs(dict): the inputs of the operator node.
+            outputs(dict): the outpus of the operator node.
+
+        Returns:
+            core.Node: the created operator node.
+        """
         op_desc = core.OpDesc()
         op_desc.set_type(op_type)
-        for attr, value in attrs.iteritems():
+        for attr, value in six.iteritems(attrs):
             self._update_desc_attr(op_desc, attr, value)
-        for input_name, var_nodes in inputs.iteritems():
+        for input_name, var_nodes in six.iteritems(inputs):
             if not isinstance(var_nodes, list):
                 var_nodes = [var_nodes]
             op_desc.set_input(input_name,
                               [var_node.name() for var_node in var_nodes])
-        for output_name, var_nodes in outputs.iteritems():
+        for output_name, var_nodes in six.iteritems(outputs):
             if not isinstance(var_nodes, list):
                 var_nodes = [var_nodes]
             op_desc.set_output(output_name,
@@ -1597,11 +1697,29 @@ class IrGraph(object):
         return self.graph.create_op_node(op_desc)
 
     def create_op_node_from_desc(self, op_desc):
+        """
+        Create a operator node by using an existing OpDesc in the graph.
+
+        Args:
+            op_desc(core.VarDesc): the giving operator description.
+
+        Returns:
+            core.Node: the created operator node.
+        """
         return self.graph.create_op_node(op_desc)
 
     def update_input_link(self, old_input_node, new_input_node, op_node):
-        assert old_input_node in self.graph.nodes() and new_input_node in self.graph.nodes() and \
-            op_node in self.graph.nodes(), 'Th three arguments must be in the graph nodes.'
+        """
+        Update the input's link of a operator node.
+
+        Args:
+            old_input_node(core.Node): the old input node of the giving op_node.
+            new_input_node(core.Node): the new input node of the giving op_node.
+            op_node(core.Node): the operator node that is needed to update input's link.
+        """
+        assert old_input_node in self.graph.nodes() and new_input_node in \
+        self.graph.nodes() and op_node in self.graph.nodes(), \
+        'The three arguments(old_input_node&new_input_node&op_node) must be in the graph nodes.'
         old_input_node.outputs_remove(op_node)
         op_node.inputs_remove(old_input_node)
         new_input_node.outputs_append(op_node)
@@ -1609,17 +1727,85 @@ class IrGraph(object):
         op_node.op()._rename_input(old_input_node.name(), new_input_node.name())
 
     def link_to(self, node_in, node_out):
+        """
+        Connect two nodes.
+
+        Args:
+            node_in(core.Node): the input node.
+            node_out(core.Node): the output node.
+        """
         assert node_in in self.graph.nodes() and node_out in self.graph.nodes(), \
-            'Th two arguments must be in the graph nodes.'
+            'The two arguments(node_in&node_out) must be in the graph nodes.'
         node_in.outputs_append(node_out)
         node_out.inputs_append(node_in)
 
     def safe_remove_nodes(self, remove_nodes):
+        """
+        Remove nodes safely since links connected to these removed nodes are
+        also removed.
+
+        Args:
+            remove_nodes(set): the nodes prepared to be removed.
+        """
         if not isinstance(remove_nodes, set):
+            if isinstance(remove_nodes, Iterable):
                 remove_nodes = set(remove_nodes)
+            else:
+                remove_nodes = {remove_nodes}
         core.graph_safe_remove_nodes(self.graph, remove_nodes)
 
-    def draw(self, save_path, name, marked_nodes=None):
+    def has_circle(self):
+        """
+        Check if the graph has a circle.
+
+        Returns:
+            bool: True if the graph has a circle else False.
+        """
+        return core.has_circle(self.graph)
+
+    def graph_num(self):
+        """
+        Count the number of unconnected graphs in this graph.
+
+        Returns:
+            int: the number of unconnected graphs.
+        """
+        return core.graph_num(self.graph)
+
+    def topology_sort(self):
+        """
+        Perform the topology sort operation on the graph.
+
+        Notes: the `graph` cannot contain a circle.
+
+        Returns:
+            set(core.Node): nodes in topology order.
+        """
+        return core.topology_sort(self.graph)
+
+    def build_adjacency_list(self):
+        """
+        Build an adjacency list of operations for the `graph`.
+
+        Returns:
+            dict{core.Node: set(core.Node)}: the adjacency list.
+        """
+        return core.build_adjacency_list(self.graph)
+
+    def draw(self, save_path, name, marked_nodes=None, remove_ctr_var=True):
+        """
+        Draw the graph. If `dot` command is installed, the drawn graph
+        will be saved as pdf file type, otherwise dot file type is used.
+
+        Args:
+            save_path(str): the save path of drawn graph.
+            name(str): the name of drawn graph.
+            marked_nodes(set(core.Node)): nodes that are needed to be marked.
+            Default value is None.
+            remove_ctr_var(bool): If it is set True, all control variable nodes
+            in the graph will be removed. Default value is True.
+        """
+
         def _convert_to_pdf(dot_file_path):
             pdf_save_path = os.path.splitext(dot_file_path)[0] + '.pdf'
             exited_code = subprocess.call('dot -Tpdf ' + dot_file_path \
@@ -1629,15 +1815,17 @@ class IrGraph(object):
                 print('The {} is saved as the dot filetype.'.format(
                     dot_file_path))
 
+        if remove_ctr_var:
             remove_ctr_vars = set()
-        ops_num = 0
             for node in self.graph.nodes():
                 if node.is_ctrl_var():
                     remove_ctr_vars.add(node)
-            elif node.is_op():
+            self.safe_remove_nodes(remove_ctr_vars)
+        ops_num = 0
+        for node in self.graph.nodes():
+            if node.is_op():
                 ops_num += 1
         print('Total ops num = {}.'.format(ops_num))
-        self.safe_remove_nodes(remove_ctr_vars)
         if marked_nodes is not None:
             if not isinstance(marked_nodes, set):
                 marked_nodes = set(marked_nodes)
@@ -1652,10 +1840,20 @@ class IrGraph(object):
         _convert_to_pdf(viz_dot_path)
 
     def to_program(self):
+        """
+        Convert the graph into a Program.
+
+        Notes: When the graph includes backward operator nodes, the
+        conversion process may be failed. Usually, this function is
+        only used to convert a test graph.
+
+        Returns:
+            Program: a program converted from the graph.
+        """
         convert_pass = core.get_pass('graph_to_program_pass')
-        convert_pass.set('program', Program().desc)
+        desc = core.ProgramDesc()
+        convert_pass.set_not_owned('program', desc)
         convert_pass.apply(self.graph)
-        desc = convert_pass.get_program('program')
         program = Program._construct_from_desc(desc)
         return program
 

python/paddle/fluid/imperative/layers.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import collections
 import contextlib
 import sys
 import numpy as np
@@ -30,31 +31,45 @@ class Layer(core.Layer):
     def __init__(self, dtype=core.VarDesc.VarType.FP32, name=None):
         self._built = False
         self._dtype = dtype
+        self._parameters = collections.OrderedDict()
+        self._sub_layers = collections.OrderedDict()
+
+    def parameters(self, include_sublayers=True):
+        """Returns a list of Parameters from current and sub-layers.
+
+        Args:
+            include_sublayers: If true, also include the parameters from
+            sublayers.
+
+        Returns a list of Parameters.
+        """
+        ret = [p for p in self._parameters.values()]
+        if include_sublayers:
+            for l in self._sub_layers.values():
+                for p in l.parameters(include_sublayers):
+                    ret.append(p)
+        return ret
 
-    def parameters(self):
-        params = []
-        for key in self.__dict__.keys():
-            value = self.__dict__[key]
-            if isinstance(value, framework.Parameter):
-                params.append(value)
-            elif isinstance(value, core.Layer):
-                params.extend(value.parameters())
-            elif isinstance(value, collections.Container):
-                if len(value) == 0:
-                    continue
-                if isinstance(value[0], framework.Parameter):
-                    params.extend(value)
-                elif isinstance(value[0], core.Layer):
-                    for v in value:
-                        params.extend(v.parameters())
-
-        return params
+    def sublayers(self, include_sublayers=True):
+        """Returns a list of sub layers.
+
+        Args:
+            include_sublayers: If true, also include the layers from sublayers.
+
+        Returns a list of sub layers.
+        """
+        ret = [l for l in self._sub_layers.values()]
+        if include_sublayers:
+            for l in self._sub_layers.values():
+                for sub_l in l.sublayers(include_sublayers):
+                    ret.append(sub_l)
+        return ret
 
     def clear_gradients(self):
         for p in self.parameters():
             p._clear_gradient()
 
-    def _build_once(self, inputs):
+    def _build_once(self, *args):
         pass
 
     def __call__(self, *inputs):
@@ -71,6 +86,66 @@ class Layer(core.Layer):
     def backward(self, *inputs):
         raise ValueError("Layer shouldn't implement backward")
 
+    def add_sublayer(self, name, sublayer):
+        """Adds a sub Layer instance.
+
+          Added sublayer can be access like self.name.
+
+        Args:
+            name: name of this sublayer.
+            sublayer: an instance of Layer.
+        Returns:
+            the sublayer passed in.
+        """
+        assert isinstance(sublayer, core.Layer)
+        self._sub_layers[name] = sublayer
+        return sublayer
+
+    def add_parameter(self, name, parameter):
+        """Adds a Parameter instance.
+
+          Added parameter can be access like self.name.
+
+        Args:
+            name: name of this sublayer.
+            parameter: an instance of Parameter.
+        Returns:
+            the parameter passed in.
+        """
+        assert isinstance(parameter, framework.Parameter)
+        self._parameters[name] = parameter
+        return parameter
+
+    def __getattr__(self, name):
+        if name in self._parameters:
+            return self._parameters[name]
+        elif name in self._sub_layers:
+            return self._sub_layers[name]
+
+    def __setattr__(self, name, value):
+        if isinstance(value, framework.Parameter):
+            params = self.__dict__.get('_parameters', None)
+            if params is None:
+                raise ValueError(
+                    "super(YourLayer, self).__init__() should be called first")
+            params[name] = value
+        elif isinstance(value, core.Layer):
+            layers = self.__dict__.get('_sub_layers', None)
+            if layers is None:
+                raise ValueError(
+                    "super(YourLayer, self).__init__() should be called first")
+            layers[name] = value
+        else:
+            object.__setattr__(self, name, value)
+
+    def __delattr__(self, name):
+        if name in self._parameters:
+            del self._parameters[name]
+        elif name in self._sub_layers:
+            del self._sub_layers[name]
+        else:
+            object.__delattr__(self, name)
+
 
 class PyLayer(core.PyLayer):
     """Layers composed of user-defined python codes."""

python/paddle/fluid/imperative/nn.py

@@ -225,9 +225,6 @@ class FC(layers.Layer):
             act=act,
             name=name)
 
-    def parameters(self):
-        return [self._w, self._b]
-
     def _build_once(self, input):
         input_shape = input.shape
         param_shape = [
@@ -478,9 +475,6 @@ class Embedding(layers.Layer):
             dtype=self._dtype,
             is_bias=False)
 
-    def parameters(self):
-        return [self._w]
-
     def forward(self, input):
         out = self._helper.create_variable_for_type_inference(self._dtype)
         self._helper.append_op(

python/paddle/fluid/layers/nn.py

@@ -3236,7 +3236,7 @@ def group_norm(input,
     # create output
     mean_out = helper.create_variable(dtype=dtype, stop_gradient=True)
     variance_out = helper.create_variable(dtype=dtype, stop_gradient=True)
-    group_norm_out = helper.create_variable(dtype)
+    group_norm_out = helper.create_variable(dtype=dtype)
 
     helper.append_op(
         type="group_norm",
@@ -5936,13 +5936,10 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=False, name=None):
                                 than :attr:`shape`.
         act (str): The non-linear activation to be applied to the reshaped tensor
                    variable.
-        inplace(bool): Must use :attr:`False` if :attr:`x` is used in multiple
-                       operators. If this flag is set :attr:`True`, reuse input
-                       :attr:`x` to reshape, which will change the shape of
-                       tensor variable :attr:`x` and might cause errors when
-                       :attr:`x` is used in multiple operators. If :attr:`False`,
-                       preserve the shape :attr:`x` and create a new output tensor
-                       variable whose data is copied from input x but reshaped.
+        inplace(bool): If ``inplace`` is `True`, the input and output of ``layers.reshape``
+                       are the same variable, otherwise, the input and output of
+                       ``layers.reshape`` are different variables. Note that if :attr:`x`
+                       is more than one layer's input, ``inplace`` must be :attr:`False`.
         name (str): The name of this layer. It is optional.
 
     Returns:
@@ -8335,6 +8332,46 @@ def stack(x, axis=0):
     If :code:`axis` < 0, it would be replaced with :code:`axis+rank(x[0])+1`.
     If :code:`axis` is None, it would be replaced with 0.
 
+    For Example:
+
+    .. code-block:: text
+
+        Case 1:
+          Input:
+            x[0].data = [ [1.0 , 2.0 ] ]
+            x[0].dims = [1, 2]
+            x[1].data = [ [3.0 , 4.0 ] ]
+            x[1].dims = [1, 2]
+            x[2].data = [ [5.0 , 6.0 ] ]
+            x[2].dims = [1, 2]
+
+          Attrs:
+            axis = 0
+
+          Output:
+            Out.data =[ [ [1.0, 2.0] ],
+                        [ [3.0, 4.0] ],
+                        [ [5.0, 6.0] ] ]
+            Out.dims = [3, 1, 2]
+
+        Case 2:
+          Given
+            x[0].data = [ [1.0 , 2.0 ] ]
+            x[0].dims = [1, 2]
+            x[1].data = [ [3.0 , 4.0 ] ]
+            x[1].dims = [1, 2]
+            x[2].data = [ [5.0 , 6.0 ] ]
+            x[2].dims = [1, 2]
+
+          Attrs:
+            axis = 1 or axis = -2
+
+          Output:
+            Out.data =[ [ [1.0, 2.0]
+                          [3.0, 4.0]
+                          [5.0, 6.0] ] ]
+            Out.dims = [1, 3, 2]
+
     Args:
         x (Variable|list(Variable)|tuple(Variable)): Input variables.
         axis (int|None): The axis along which all inputs are stacked.

python/paddle/fluid/layers/tensor.py

@@ -567,7 +567,7 @@ def ones(shape, dtype, force_cpu=False):
     It also sets *stop_gradient* to True.
 
     Args:
-        shape(tuple|list|None): Shape of output tensor
+        shape(tuple|list): Shape of output tensor
         dtype(np.dtype|core.VarDesc.VarType|str): Data type of output tensor
 
     Returns:
@@ -578,6 +578,10 @@ def ones(shape, dtype, force_cpu=False):
 
           data = fluid.layers.ones(shape=[1], dtype='int64')
     """
+    assert isinstance(shape, list) or isinstance(
+        shape, tuple), "The shape's type should be list or tuple."
+    assert reduce(lambda x, y: x * y,
+                  shape) > 0, "The shape is invalid: %s." % (str(shape))
     return fill_constant(value=1.0, **locals())
 
 

python/paddle/fluid/parallel_executor.py

@@ -148,6 +148,8 @@ class ParallelExecutor(object):
             else framework.default_main_program()
         # FIXME(dzhwinter): enable_inplace should be after memory_optimize
         # if turn on python memory optimize, turn off the inplace_pass.
+        if build_strategy.memory_optimize is None:
+            build_strategy.memory_optimize = False if main._is_mem_optimized else True
         if build_strategy.enable_inplace is None:
             build_strategy.enable_inplace = False if main._is_mem_optimized else True
         scope = scope if scope is not None else executor.global_scope()

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -77,6 +77,7 @@ list(REMOVE_ITEM TEST_OPS test_bilinear_interp_op)
 list(REMOVE_ITEM TEST_OPS test_nearest_interp_op)
 list(REMOVE_ITEM TEST_OPS test_imperative_resnet)
 list(REMOVE_ITEM TEST_OPS test_imperative_optimizer)
+list(REMOVE_ITEM TEST_OPS test_ir_memory_optimize_transformer)
 foreach(TEST_OP ${TEST_OPS})
     py_test_modules(${TEST_OP} MODULES ${TEST_OP})
 endforeach(TEST_OP)
@@ -107,6 +108,9 @@ py_test_modules(test_parallel_executor_crf MODULES test_parallel_executor_crf SE
 py_test_modules(test_parallel_executor_fetch_feed MODULES test_parallel_executor_fetch_feed SERIAL)
 set_tests_properties(test_parallel_executor_fetch_feed PROPERTIES TIMEOUT 450)
 py_test_modules(test_parallel_executor_transformer MODULES test_parallel_executor_transformer SERIAL)
+if(NOT WIN32)
+py_test_modules(test_ir_memory_optimize_transformer MODULES test_ir_memory_optimize_transformer SERIAL)
+endif()
 if(NOT APPLE)
     py_test_modules(test_image_classification_resnet MODULES test_image_classification_resnet SERIAL)
     if(CMAKE_BUILD_TYPE STREQUAL "Debug")

python/paddle/fluid/tests/unittests/parallel_executor_test_base.py

@@ -79,7 +79,7 @@ class TestParallelExecutorBase(unittest.TestCase):
             if use_reduce else fluid.BuildStrategy.ReduceStrategy.AllReduce
         build_strategy.fuse_elewise_add_act_ops = fuse_elewise_add_act_ops
         build_strategy.fuse_relu_depthwise_conv = fuse_relu_depthwise_conv
-        build_strategy.memory_optimize = use_ir_memory_optimize
+        build_strategy.memory_optimize = False if memory_opt else use_ir_memory_optimize
         # python memory optimization is conflict with inplace pass.
         # Use ir graph memory optimization after inplace pass is the correct way.
         build_strategy.enable_inplace = False if memory_opt else enable_inplace

python/paddle/fluid/tests/unittests/test_base_layer.py

+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+import numpy as np
+
+import paddle.fluid as fluid
+from paddle.fluid.layer_helper import LayerHelper
+
+
+class L1(fluid.imperative.Layer):
+    def __init__(self):
+        super(L1, self).__init__()
+        self._helper = LayerHelper(
+            'MyLayer',
+            param_attr=fluid.ParamAttr(
+                initializer=fluid.initializer.Constant(value=0.1)))
+
+        self.w1 = self._helper.create_parameter(
+            attr=self._helper.param_attr,
+            shape=[2, 2],
+            dtype='float32',
+            is_bias=False)
+        self.w2 = self._helper.create_parameter(
+            attr=self._helper.param_attr,
+            shape=[2, 2],
+            dtype='float32',
+            is_bias=False)
+
+    def forward(self):
+        return self.w1 + self.w2
+
+
+class L2(fluid.imperative.Layer):
+    def __init__(self):
+        super(L2, self).__init__()
+        self.layer1 = L1()
+        self.layer2 = L1()
+
+    def forward(self):
+        return self.layer1() + self.layer2()
+
+
+class L3(fluid.imperative.Layer):
+    def __init__(self):
+        super(L3, self).__init__()
+        self.layer1 = L2()
+        self.layer2 = L2()
+
+    def forward(self):
+        return self.layer1() + self.layer2()
+
+
+class TestBaseLayer(unittest.TestCase):
+    def test_one_level(self):
+        with fluid.imperative.guard():
+            l = L1()
+            ret = l()
+            self.assertEqual(l.w1.name, "MyLayer_0.w_0")
+            self.assertEqual(l.w2.name, "MyLayer_0.w_1")
+            self.assertTrue(np.allclose(ret._numpy(), 0.2 * np.ones([2, 2])))
+
+    def test_three_level(self):
+        with fluid.imperative.guard():
+            l = L3()
+            ret = l()
+            self.assertTrue(np.allclose(ret._numpy(), 0.8 * np.ones([2, 2])))
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_dist_transpiler.py

@@ -22,6 +22,9 @@ import six
 import unittest
 import numpy as np
 
+import gc
+gc.set_debug(gc.DEBUG_COLLECTABLE)
+
 import paddle.fluid as fluid
 
 
@@ -99,6 +102,12 @@ class TranspilerTest(unittest.TestCase):
         with fluid.unique_name.guard():
             with fluid.program_guard(main, startup):
                 self.transpiler_test_impl()
+        # NOTE: run gc.collect to eliminate pybind side objects to
+        # prevent random double-deallocate when inherited in python.
+        del self.transpiler
+        del main
+        del startup
+        gc.collect()
 
 
 class TestBasicModel(TranspilerTest):
@@ -797,6 +806,7 @@ class TestNCCL2Transpile(TranspilerTest):
             print([op.type for op in startup.global_block().ops])
             self.assertEqual(startup.global_block().ops[-1].type, "gen_nccl_id")
             self.assertIsNotNone(startup.global_block().vars.get("NCCLID"))
+            gc.collect()
         else:
             pass
 

python/paddle/fluid/tests/unittests/test_fuse_elewise_add_act_pass.py

@@ -121,6 +121,8 @@ class TestMNIST(TestParallelExecutorBase):
                 regularization=fluid.regularizer.L2Decay(1e-6))
             return optimizer
 
+        # NOTE(dzh):
+        # need to make it compatible with elewise fuse act
         not_fuse_op_first_loss, not_fuse_op_last_loss = self.check_network_convergence(
             model,
             feed_dict={"image": img,
@@ -128,6 +130,7 @@ class TestMNIST(TestParallelExecutorBase):
             use_cuda=use_cuda,
             fuse_elewise_add_act_ops=False,
             memory_opt=False,
+            use_ir_memory_optimize=False,
             optimizer=_optimizer)
         fuse_op_first_loss, fuse_op_last_loss = self.check_network_convergence(
             model,
@@ -136,6 +139,7 @@ class TestMNIST(TestParallelExecutorBase):
             use_cuda=use_cuda,
             fuse_elewise_add_act_ops=True,
             memory_opt=False,
+            use_ir_memory_optimize=False,
             optimizer=_optimizer)
 
         for loss in zip(not_fuse_op_first_loss, fuse_op_first_loss):

python/paddle/fluid/tests/unittests/test_imperative.py

@@ -333,6 +333,18 @@ class TestImperative(unittest.TestCase):
         self.assertTrue(np.allclose(dy_out, static_out))
         self.assertTrue(np.allclose(dy_grad, static_grad))
 
+        params = mlp.parameters(True)
+        self.assertEqual("FC_0.w_0", params[0].name)
+        self.assertEqual("FC_0.b_0", params[1].name)
+        self.assertEqual("FC_1.w_0", params[2].name)
+        self.assertEqual("FC_1.b_0", params[3].name)
+        self.assertEqual(len(params), 4)
+
+        sublayers = mlp.sublayers(True)
+        self.assertEqual(mlp._fc1, sublayers[0])
+        self.assertEqual(mlp._fc2, sublayers[1])
+        self.assertEqual(len(sublayers), 2)
+
     def test_rnn(self):
         np_inp = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0],
                            [10.0, 11.0, 12.0]])

python/paddle/fluid/tests/unittests/test_imperative_gan.py

@@ -33,9 +33,6 @@ class Discriminator(fluid.imperative.Layer):
         self._fc1 = FC(size=32, act='elu', name="d_fc1")
         self._fc2 = FC(size=1, name="d_fc2")
 
-    def parameters(self):
-        return self._fc1.parameters() + self._fc2.parameters()
-
     def forward(self, inputs):
         x = self._fc1(inputs)
         return self._fc2(x)
@@ -48,10 +45,6 @@ class Generator(fluid.imperative.Layer):
         self._fc2 = FC(size=64, act='elu', name="g_fc2")
         self._fc3 = FC(size=1, name="g_fc3")
 
-    def parameters(self):
-        return self._fc1.parameters() + self._fc2.parameters(
-        ) + self._fc3.parameters()
-
     def forward(self, inputs):
         x = self._fc1(inputs)
         x = self._fc2(x)

python/paddle/fluid/tests/unittests/test_imperative_ptb_rnn.py

@@ -75,16 +75,6 @@ class SimpleLSTMRNN(fluid.imperative.Layer):
             self.hidden_array.append(pre_hidden)
             self.cell_array.append(pre_cell)
 
-    def parameters(self):
-        parameters = list()
-        for param in self.weight_1_arr:
-            parameters.append(param)
-        for param in self.weight_2_arr:
-            parameters.append(param)
-        for bias in self.bias_arr:
-            parameters.append(bias)
-        return parameters
-
     def forward(self, input_embedding, init_hidden=None, init_cell=None):
         res = []
         for index in range(self._num_steps):
@@ -177,12 +167,6 @@ class PtbModel(fluid.imperative.Layer):
     def _build_once(self, input, label, init_hidden, init_cell):
         pass
 
-    def parameters(self):
-        parameters = self.simple_lstm_rnn.parameters() + [
-            self.softmax_weight, self.softmax_bias
-        ] + self.embedding.parameters()
-        return parameters
-
     def forward(self, input, label, init_hidden, init_cell):
 
         init_h = fluid.layers.reshape(

python/paddle/fluid/tests/unittests/test_imperative_resnet.py

@@ -21,7 +21,6 @@ import paddle
 import paddle.fluid as fluid
 from paddle.fluid import core
 from paddle.fluid.layer_helper import LayerHelper
-from paddle.fluid.optimizer import SGDOptimizer
 from paddle.fluid.imperative.nn import Conv2D, Pool2D, BatchNorm, FC
 from paddle.fluid.imperative.base import to_variable
 from test_imperative_base import new_program_scope
@@ -173,11 +172,13 @@ class ResNet(fluid.imperative.Layer):
         for block in range(len(depth)):
             shortcut = False
             for i in range(depth[block]):
-                bottleneck_block = BottleneckBlock(
+                bottleneck_block = self.add_sublayer(
+                    'bb_%d_%d' % (block, i),
+                    BottleneckBlock(
                         num_channels=num_channels,
                         num_filters=num_filters[block],
                         stride=2 if i == 0 and block != 0 else 1,
-                    shortcut=shortcut)
+                        shortcut=shortcut))
                 num_channels = bottleneck_block._num_channels_out
                 self.bottleneck_block_list.append(bottleneck_block)
                 shortcut = True
@@ -223,8 +224,7 @@ class TestImperativeResnet(unittest.TestCase):
                 batch_size=batch_size)
 
             dy_param_init_value = {}
-            for param in fluid.default_main_program().global_block(
-            ).all_parameters():
+            for param in resnet.parameters():
                 dy_param_init_value[param.name] = param._numpy()
 
             for batch_id, data in enumerate(train_reader()):
@@ -247,16 +247,14 @@ class TestImperativeResnet(unittest.TestCase):
                 dy_out = avg_loss._numpy()
 
                 if batch_id == 0:
-                    for param in fluid.default_main_program().global_block(
-                    ).all_parameters():
+                    for param in resnet.parameters():
                         if param.name not in dy_param_init_value:
                             dy_param_init_value[param.name] = param._numpy()
 
                 avg_loss._backward()
 
                 dy_grad_value = {}
-                for param in fluid.default_main_program().global_block(
-                ).all_parameters():
+                for param in resnet.parameters():
                     if not param.stop_gradient:
                         np_array = np.array(param._ivar._grad_ivar().value()
                                             .get_tensor())
@@ -267,8 +265,7 @@ class TestImperativeResnet(unittest.TestCase):
                 resnet.clear_gradients()
 
                 dy_param_value = {}
-                for param in fluid.default_main_program().global_block(
-                ).all_parameters():
+                for param in resnet.parameters():
                     dy_param_value[param.name] = param._numpy()
 
         with new_program_scope():
@@ -349,6 +346,7 @@ class TestImperativeResnet(unittest.TestCase):
         self.assertTrue(np.allclose(static_out, dy_out))
 
         self.assertEqual(len(dy_param_init_value), len(static_param_init_value))
+
         for key, value in six.iteritems(static_param_init_value):
             self.assertTrue(np.allclose(value, dy_param_init_value[key]))
             self.assertTrue(np.isfinite(value.all()))

python/paddle/fluid/tests/unittests/test_ir_memory_optimize_transformer.py

+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import unittest
+import paddle.fluid as fluid
+import paddle.fluid.core as core
+
+os.environ['FLAGS_eager_delete_tensor_gb'] = "0.0"
+os.environ[
+    'RECORDIO_FILENAME'] = '/tmp/ir_memory_optimize_transformer.wmt16.recordio'
+
+from test_parallel_executor_transformer import TestTransformer
+from test_parallel_executor_transformer import transformer
+
+
+# NOTE(dzhwinter): test diferent strategy colisions.
+# open the eager delete tensor strategy by default.
+class TestTransformerWithIR(TestTransformer):
+    def test_main(self):
+        if core.is_compiled_with_cuda():
+            # check python transpiler
+            self.check_network_convergence(
+                transformer,
+                use_cuda=True,
+                memory_opt=True,
+                use_ir_memory_optimize=False)
+            # check IR memory optimize
+            self.check_network_convergence(
+                transformer,
+                use_cuda=True,
+                memory_opt=False,
+                use_ir_memory_optimize=True)
+
+
+if __name__ == '__main__':
+    unittest.main()