From 02cc7d90a606875a44d605c18e17855ce8339652 Mon Sep 17 00:00:00 2001
From: xuwei06 <xuwei06@baidu.com>
Date: Mon, 19 Jun 2017 13:23:24 -0700
Subject: [PATCH] Evaluator for recurrent group

Make the evaluators inside a recurrent goup true evaluator, meaning that their evaluation results are incorporated into the whole evaluator result.
---
 .../gradientmachines/NeuralNetwork.cpp        | 35 ++++++++++++++++++-
 .../gserver/gradientmachines/NeuralNetwork.h  |  2 ++
 .../RecurrentGradientMachine.cpp              | 20 ++---------
 .../RecurrentGradientMachine.h                |  2 --
 4 files changed, 38 insertions(+), 21 deletions(-)

diff --git a/paddle/gserver/gradientmachines/NeuralNetwork.cpp b/paddle/gserver/gradientmachines/NeuralNetwork.cpp
index 4512aacc81f..f245c16bfe9 100644
--- a/paddle/gserver/gradientmachines/NeuralNetwork.cpp
+++ b/paddle/gserver/gradientmachines/NeuralNetwork.cpp
@@ -320,7 +320,7 @@ public:
     }
   }
 
-  virtual void eval(const NeuralNetwork& nn) {
+  virtual void eval(const NeuralNetwork& nn) override {
     for (auto& evaluator : evaluators_) {
       evaluator->eval(nn);
     }
@@ -395,6 +395,30 @@ private:
   }
 };
 
+class SubnetEvaluator : public CombinedEvaluator {
+public:
+  SubnetEvaluator(const std::string& layerName,
+                  std::unique_ptr<Evaluator>&& evaluator)
+      : layerName_(layerName) {
+    addEvaluator(std::move(evaluator));
+  }
+  virtual void eval(const NeuralNetwork& nn) override {
+    const LayerPtr& layer = nn.getLayer(layerName_);
+    CHECK(layer) << "Nonexisted layer: " << layerName_ << " in submodel "
+                 << nn.getName();
+    bool accessed = false;
+    layer->accessSubNetwork([this, &accessed](NeuralNetwork& subnet) {
+      subnet.eval(evaluators_[0].get());
+      accessed = true;
+    });
+    CHECK(accessed) << "There is no subnetwork for layer " << layerName_
+                    << " in submodel " << nn.getName();
+  }
+
+protected:
+  std::string layerName_;
+};
+
 Evaluator* NeuralNetwork::makeEvaluator() const {
   CombinedEvaluator* combinedEvaluator = new CombinedEvaluator();
   auto subModelConfig = std::find_if(config_.sub_models().begin(),
@@ -421,6 +445,15 @@ Evaluator* NeuralNetwork::makeEvaluator() const {
         combinedEvaluator->addEvaluator(std::move(evaluator));
       }
     }
+    for (auto& layer : layers_) {
+      layer->accessSubNetwork(
+          [layer, combinedEvaluator](NeuralNetwork& subnet) {
+            std::unique_ptr<Evaluator> subEvaluator(new SubnetEvaluator(
+                layer->getName(),
+                std::unique_ptr<Evaluator>(subnet.makeEvaluator())));
+            combinedEvaluator->addEvaluator(std::move(subEvaluator));
+          });
+    }
   } else {
     for (const EvaluatorConfig& evalConfig : config_.evaluators()) {
       std::unique_ptr<Evaluator> evaluator(Evaluator::create(evalConfig));
diff --git a/paddle/gserver/gradientmachines/NeuralNetwork.h b/paddle/gserver/gradientmachines/NeuralNetwork.h
index e7b6c438407..12810f64251 100644
--- a/paddle/gserver/gradientmachines/NeuralNetwork.h
+++ b/paddle/gserver/gradientmachines/NeuralNetwork.h
@@ -129,6 +129,8 @@ public:
   static NeuralNetwork* newNeuralNetwork(const std::string& name = "",
                                          NeuralNetwork* rootNetwork = nullptr);
 
+  const std::string& getName() const { return subModelName_; }
+
 protected:
   /**
    * The constructor of NeuralNetwork.
diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
index 3e930380226..5d4b67da843 100644
--- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
+++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp
@@ -287,10 +287,6 @@ void RecurrentGradientMachine::init(
       parameterIds_.push_back(para->getID());
     }
   }
-
-  if (subModelConfig->evaluator_names_size() > 0) {
-    evaluator_.reset(frames_[0]->makeEvaluator());
-  }
 }
 
 void RecurrentGradientMachine::resizeOrCreateFrames(int numFrames) {
@@ -561,9 +557,6 @@ void RecurrentGradientMachine::forward(const std::vector<Argument>& inArgs,
     std::vector<Argument> outArgs;
     frames_[i]->forward(inArgs, &outArgs, passType);
   }
-  if (evaluator_ && passType == PASS_TEST) {
-    this->eval(evaluator_.get());
-  }
 
   reorganizeOutput(passType);
 }
@@ -577,11 +570,6 @@ void RecurrentGradientMachine::backward(const UpdateCallback& callback) {
   for (auto& memoryFrameLine : memoryFrameLines_) {
     memoryFrameLine.bootLayer->backward(nullptr);
   }
-
-  // call printers here so the gradient can be printed
-  if (evaluator_) {
-    this->eval(evaluator_.get());
-  }
 }
 
 void RecurrentGradientMachine::forwardBackward(
@@ -595,9 +583,9 @@ void RecurrentGradientMachine::forwardBackward(
 void RecurrentGradientMachine::eval(Evaluator* evaluator) const {
   // call printers frame by frame
   for (int i = 0; i < maxSequenceLength_; ++i) {
-    LOG(INFO) << "Recurrent Layer Group eval frame " << i << " begin";
+    VLOG(2) << "Recurrent Layer Group eval frame " << i << " begin";
     evaluator->eval(*(frames_[i].get()));
-    LOG(INFO) << "Recurrent Layer Group eval frame " << i << " end";
+    VLOG(2) << "Recurrent Layer Group eval frame " << i << " end";
   }
 }
 
@@ -1093,10 +1081,6 @@ void RecurrentGradientMachine::oneWaySearch(size_t batchSize) {
 
     copyDataOutlinkFrame(machineCur);
 
-    // call value printer
-    if (evaluator_) {
-      evaluator_->eval(*(frames_[machineCur].get()));
-    }
     // check eos
     const IVectorPtr& eosVec =
         eosFrameLine_->layers[machineCur]->getOutput().ids;
diff --git a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h b/paddle/gserver/gradientmachines/RecurrentGradientMachine.h
index 8d94d7e2df2..9f957a9401f 100644
--- a/paddle/gserver/gradientmachines/RecurrentGradientMachine.h
+++ b/paddle/gserver/gradientmachines/RecurrentGradientMachine.h
@@ -428,8 +428,6 @@ protected:
   std::vector<int>
       parameterIds_;  // parameters actually used by this Layer Group
 
-  std::unique_ptr<Evaluator> evaluator_;  // frame printers in this layer group
-
   // store final argument of outFrameLines_
   std::vector<Argument> dataArgs_;
   // store each frame's output argument of outFrameLines_
-- 
GitLab