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

.travis.yml

+group: deprecated-2017Q2
 language: cpp
 cache:
   directories:

doc/api/v2/config/evaluators.rst

@@ -99,3 +99,12 @@ value_printer
 ..  automodule:: paddle.v2.evaluator
     :members:  value_printer
     :noindex:
+
+Detection
+=====
+
+detection_map
+-------------
+..  automodule:: paddle.v2.evaluator
+    :members:  detection_map
+    :noindex:

go/cmd/pserver/pserver.go

@@ -5,18 +5,35 @@ import (
 	"net/http"
 	"net/rpc"
 	"strconv"
+	"time"
 
 	"github.com/namsral/flag"
 
 	"github.com/PaddlePaddle/Paddle/go/pserver"
+	log "github.com/sirupsen/logrus"
 )
 
 func main() {
 	port := flag.Int("port", 0, "port of the pserver")
+	etcdEndpoint := flag.String("etcd-endpoint", "http://127.0.0.1:2379",
+		"comma separated endpoint string for pserver to connect to etcd")
+	etcdTimeout := flag.Int("etcd-timeout", 5, "timeout for etcd calls")
+	logLevel := flag.String("log-level", "info",
+		"log level, possible values: debug, info, warning, error, fatal, panic")
 	flag.Parse()
 
-	s := pserver.NewService()
-	err := rpc.Register(s)
+	level, err := log.ParseLevel(*logLevel)
+	if err != nil {
+		panic(err)
+	}
+	log.SetLevel(level)
+
+	timeout := time.Second * time.Duration((*etcdTimeout))
+	s, err := pserver.NewService(*etcdEndpoint, timeout)
+	if err != nil {
+		panic(err)
+	}
+	err = rpc.Register(s)
 	if err != nil {
 		panic(err)
 	}
@@ -27,7 +44,9 @@ func main() {
 		panic(err)
 	}
 
+	log.Infof("start pserver at port %d", *port)
 	err = http.Serve(l, nil)
+
 	if err != nil {
 		panic(err)
 	}

go/pserver/client_test.go

@@ -7,6 +7,7 @@ import (
 	"strconv"
 	"strings"
 	"testing"
+	"time"
 
 	"github.com/PaddlePaddle/Paddle/go/pserver"
 )
@@ -30,9 +31,12 @@ func init() {
 		port[i] = p
 
 		go func(l net.Listener) {
-			s := pserver.NewService()
+			s, err := pserver.NewService("", time.Second*5)
+			if err != nil {
+				panic(err)
+			}
 			server := rpc.NewServer()
-			err := server.Register(s)
+			err = server.Register(s)
 			if err != nil {
 				panic(err)
 			}

go/pserver/service.go

 package pserver
 
 import (
+	"context"
 	"errors"
 	"fmt"
+	"strconv"
+	"strings"
 	"sync"
+	"time"
+
+	"github.com/PaddlePaddle/Paddle/go/utils/networkhelper"
+	"github.com/coreos/etcd/clientv3"
+	"github.com/coreos/etcd/clientv3/concurrency"
+	log "github.com/sirupsen/logrus"
 )
 
 // ElementType is the type of elements of a Parameter.
@@ -24,6 +33,9 @@ const (
 	Float64
 )
 
+// PsDesired is etcd path for store desired pserver count
+const PsDesired = "/ps_desired"
+
 // Parameter is a piece of data to sync with the parameter server.
 type Parameter struct {
 	Name        string
@@ -47,14 +59,121 @@ type Service struct {
 	mu       sync.Mutex
 	opt      *optimizer
 	paramMap map[string]Parameter
+
+	etcdEndpoints string
+	etcdClient    *clientv3.Client
+	// etcdTimeout is also used as retry intervals.
+	etcdTimeout time.Duration
+	// desired number of pservers in the job.
+	// assume desired will not change during one training job.
+	desired int
+	// FIXME: ensure GetExternalIP gets the correct ip for trainers to connect.
+	externalIP string
 }
 
-// NewService creates a new service.
-func NewService() *Service {
+// NewService creates a new service, will bypass etcd registration if no
+// endpoints specified.
+func NewService(endpoints string, timeout time.Duration) (*Service, error) {
 	s := &Service{opt: newOptimizer(sgd, 0.005)}
 	s.paramMap = make(map[string]Parameter)
 	s.initialized = make(chan struct{})
-	return s
+	s.etcdEndpoints = endpoints
+	s.etcdTimeout = timeout
+
+	var err error
+	s.externalIP, err = networkhelper.GetExternalIP()
+	if err != nil {
+		return nil, err
+	}
+
+	if endpoints != "" {
+		// initialize connection to etcd, try
+		ep := strings.Split(s.etcdEndpoints, ",")
+		for {
+			cli, err := clientv3.New(clientv3.Config{
+				Endpoints:   ep,
+				DialTimeout: s.etcdTimeout,
+			})
+			if err != nil {
+				log.Errorf("connect to etcd error: %v", err)
+				time.Sleep(s.etcdTimeout)
+				continue
+			}
+			s.etcdClient = cli
+			log.Debugf("inited client to %s", s.etcdEndpoints)
+			break
+		}
+		// wait and set s.desired init value
+		for {
+			ctx, cancel := context.WithTimeout(context.Background(), time.Second)
+			resp, err := s.etcdClient.Get(ctx, PsDesired)
+			cancel()
+			if err != nil {
+				log.Errorf("getting %s error: %v", PsDesired, err)
+				time.Sleep(s.etcdTimeout)
+				continue
+			}
+			if len(resp.Kvs) != 0 {
+				s.desired, err = strconv.Atoi(string(resp.Kvs[0].Value))
+				if err != nil {
+					log.Errorf("value of %s invalid %v\n", PsDesired, err)
+					time.Sleep(s.etcdTimeout)
+					// NOTE: wait util ps_desired value change
+					continue
+				}
+				break
+			}
+		}
+		// try register pserver node on etcd
+		for {
+			ctx, cancel := context.WithTimeout(context.Background(), time.Second)
+			_, err := s.registerPserverEtcd(ctx)
+			cancel()
+			if err != nil {
+				log.Warn(err)
+				time.Sleep(s.etcdTimeout)
+				continue
+			}
+			break
+		}
+	} // if endpoints != ""
+	// Bypass etcd registration if no endpoints specified
+	return s, nil
+}
+
+// registerPserverEtcd registers pserver node on etcd using transaction.
+func (s *Service) registerPserverEtcd(ctx context.Context) (*clientv3.TxnResponse, error) {
+	return concurrency.NewSTM(s.etcdClient, func(c concurrency.STM) error {
+		registered := false
+		for i := 0; i < s.desired; i++ {
+			psKey := "/ps/" + strconv.Itoa(i)
+			log.Debugf("checking %s", psKey)
+			ps := c.Get(psKey)
+			log.Debugf("got value (%s) for key: %s", ps, psKey)
+
+			if ps == "" {
+				resp, err := s.etcdClient.Grant(context.TODO(), 5)
+				if err != nil {
+					log.Fatal(err)
+				}
+				// find the first id and write info
+				c.Put(psKey, s.externalIP, clientv3.WithLease(resp.ID))
+				log.Debugf("set pserver node %s with value %s", psKey, s.externalIP)
+				_, kaerr := s.etcdClient.KeepAlive(context.TODO(), resp.ID)
+				if kaerr != nil {
+					log.Errorf("keepalive etcd node error: %v", kaerr)
+					return kaerr
+				}
+				log.Debug("register finished")
+				registered = true
+				break
+			}
+		}
+		if registered == true {
+			return nil
+		}
+		return errors.New("not registerd, may due to already have enough pservers")
+	}, concurrency.WithAbortContext(ctx), concurrency.WithIsolation(concurrency.RepeatableReads))
 }
 
 // InitParam initializes a parameter.

go/pserver/service_test.go

@@ -10,12 +10,15 @@ import (
 )
 
 func TestFull(t *testing.T) {
-	s := pserver.NewService()
+	s, err := pserver.NewService("", time.Second*5)
+	if err != nil {
+		t.Error(err)
+	}
 	var p pserver.Parameter
 	p.Name = "param_a"
 	p.Content = []byte{1, 0, 0, 0, 2, 0, 0, 0, 3, 0, 0, 0}
 	p.ElementType = pserver.Int32
-	err := s.InitParam(pserver.ParameterWithConfig{Param: p, Config: nil}, nil)
+	err = s.InitParam(pserver.ParameterWithConfig{Param: p, Config: nil}, nil)
 	if err != nil {
 		t.FailNow()
 	}
@@ -72,8 +75,11 @@ func TestFull(t *testing.T) {
 }
 
 func TestMultipleInit(t *testing.T) {
-	s := pserver.NewService()
-	err := s.FinishInitParams(0, nil)
+	s, err := pserver.NewService("", time.Second*5)
+	if err != nil {
+		t.Error(err)
+	}
+	err = s.FinishInitParams(0, nil)
 	if err != nil {
 		t.FailNow()
 	}
@@ -85,15 +91,18 @@ func TestMultipleInit(t *testing.T) {
 }
 
 func TestUninitialized(t *testing.T) {
-	s := pserver.NewService()
-	err := s.SendGrad(pserver.Gradient{}, nil)
+	s, err := pserver.NewService("", time.Second*5)
+	err = s.SendGrad(pserver.Gradient{}, nil)
 	if err.Error() != pserver.Uninitialized {
 		t.FailNow()
 	}
 }
 
 func TestBlockUntilInitialized(t *testing.T) {
-	s := pserver.NewService()
+	s, err := pserver.NewService("", time.Second*5)
+	if err != nil {
+		t.Error(err)
+	}
 	ch := make(chan struct{}, 2)
 	errCh := make(chan error, 2)
 	var wg sync.WaitGroup
@@ -133,7 +142,7 @@ func TestBlockUntilInitialized(t *testing.T) {
 	p.Name = "param_a"
 	p.Content = []byte{1, 0, 0, 0, 2, 0, 0, 0, 3, 0, 0, 0}
 	p.ElementType = pserver.Int32
-	err := s.InitParam(pserver.ParameterWithConfig{Param: p, Config: nil}, nil)
+	err = s.InitParam(pserver.ParameterWithConfig{Param: p, Config: nil}, nil)
 	if err != nil {
 		t.FailNow()
 	}

go/utils/networkhelper/helper.go

+package networkhelper
+
+import (
+	"errors"
+	"net"
+)
+
+// GetExternalIP returns the ip address of local network interface, not the
+// loopback device.
+func GetExternalIP() (string, error) {
+	ifaces, err := net.Interfaces()
+	if err != nil {
+		return "", err
+	}
+	for _, iface := range ifaces {
+		if iface.Flags&net.FlagUp == 0 {
+			continue // interface down
+		}
+		if iface.Flags&net.FlagLoopback != 0 {
+			continue // loopback interface
+		}
+		addrs, err := iface.Addrs()
+		if err != nil {
+			return "", err
+		}
+		for _, addr := range addrs {
+			var ip net.IP
+			switch v := addr.(type) {
+			case *net.IPNet:
+				ip = v.IP
+			case *net.IPAddr:
+				ip = v.IP
+			}
+			if ip == nil || ip.IsLoopback() {
+				continue
+			}
+			ip = ip.To4()
+			if ip == nil {
+				continue // not an ipv4 address
+			}
+			return ip.String(), nil
+		}
+	}
+	return "", errors.New("are you connected to the network?")
+}

go/utils/networkhelper/helper_test.go

+package networkhelper
+
+import "testing"
+
+func TestGetIP(t *testing.T) {
+	_, err := GetExternalIP()
+	if err != nil {
+		t.Errorf("GetExternalIP returns error : %v\n", err)
+	}
+}

paddle/gserver/evaluators/DetectionMAPEvaluator.cpp

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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. */
+
+#include "Evaluator.h"
+#include "paddle/gserver/layers/DetectionUtil.h"
+
+using std::map;
+using std::vector;
+using std::pair;
+using std::make_pair;
+
+namespace paddle {
+
+/**
+ * @brief detection map Evaluator
+ *
+ * The config file api is detection_map_evaluator.
+ */
+class DetectionMAPEvaluator : public Evaluator {
+public:
+  DetectionMAPEvaluator()
+      : evaluateDifficult_(false), cpuOutput_(nullptr), cpuLabel_(nullptr) {}
+
+  virtual void start() {
+    Evaluator::start();
+    allTruePos_.clear();
+    allFalsePos_.clear();
+    numPos_.clear();
+  }
+
+  virtual real evalImp(std::vector<Argument>& arguments) {
+    overlapThreshold_ = config_.overlap_threshold();
+    backgroundId_ = config_.background_id();
+    evaluateDifficult_ = config_.evaluate_difficult();
+    apType_ = config_.ap_type();
+
+    MatrixPtr detectTmpValue = arguments[0].value;
+    Matrix::resizeOrCreate(cpuOutput_,
+                           detectTmpValue->getHeight(),
+                           detectTmpValue->getWidth(),
+                           false,
+                           false);
+
+    MatrixPtr labelTmpValue = arguments[1].value;
+    Matrix::resizeOrCreate(cpuLabel_,
+                           labelTmpValue->getHeight(),
+                           labelTmpValue->getWidth(),
+                           false,
+                           false);
+
+    cpuOutput_->copyFrom(*detectTmpValue);
+    cpuLabel_->copyFrom(*labelTmpValue);
+
+    Argument label = arguments[1];
+    const int* labelIndex = label.sequenceStartPositions->getData(false);
+    size_t batchSize = label.getNumSequences();
+
+    vector<map<size_t, vector<NormalizedBBox>>> allGTBBoxes;
+    vector<map<size_t, vector<pair<real, NormalizedBBox>>>> allDetectBBoxes;
+
+    for (size_t n = 0; n < batchSize; ++n) {
+      map<size_t, vector<NormalizedBBox>> bboxes;
+      for (int i = labelIndex[n]; i < labelIndex[n + 1]; ++i) {
+        vector<NormalizedBBox> bbox;
+        getBBoxFromLabelData(cpuLabel_->getData() + i * 6, 1, bbox);
+        int c = cpuLabel_->getData()[i * 6];
+        bboxes[c].push_back(bbox[0]);
+      }
+      allGTBBoxes.push_back(bboxes);
+    }
+
+    size_t n = 0;
+    const real* cpuOutputData = cpuOutput_->getData();
+    for (size_t imgId = 0; imgId < batchSize; ++imgId) {
+      map<size_t, vector<pair<real, NormalizedBBox>>> bboxes;
+      size_t curImgId = static_cast<size_t>((cpuOutputData + n * 7)[0]);
+      while (curImgId == imgId && n < cpuOutput_->getHeight()) {
+        vector<real> label;
+        vector<real> score;
+        vector<NormalizedBBox> bbox;
+        getBBoxFromDetectData(cpuOutputData + n * 7, 1, label, score, bbox);
+        bboxes[label[0]].push_back(make_pair(score[0], bbox[0]));
+        ++n;
+        curImgId = static_cast<size_t>((cpuOutputData + n * 7)[0]);
+      }
+      allDetectBBoxes.push_back(bboxes);
+    }
+
+    for (size_t n = 0; n < batchSize; ++n) {
+      for (map<size_t, vector<NormalizedBBox>>::iterator it =
+               allGTBBoxes[n].begin();
+           it != allGTBBoxes[n].end();
+           ++it) {
+        size_t count = 0;
+        if (evaluateDifficult_) {
+          count = it->second.size();
+        } else {
+          for (size_t i = 0; i < it->second.size(); ++i)
+            if (!(it->second[i].isDifficult)) ++count;
+        }
+        if (numPos_.find(it->first) == numPos_.end() && count != 0) {
+          numPos_[it->first] = count;
+        } else {
+          numPos_[it->first] += count;
+        }
+      }
+    }
+
+    // calcTFPos
+    calcTFPos(batchSize, allGTBBoxes, allDetectBBoxes);
+
+    return 0;
+  }
+
+  virtual void printStats(std::ostream& os) const {
+    real mAP = calcMAP();
+    os << "Detection mAP=" << mAP;
+  }
+
+  virtual void distributeEval(ParameterClient2* client) {
+    LOG(FATAL) << "Distribute detection evaluation not implemented.";
+  }
+
+protected:
+  void calcTFPos(const size_t batchSize,
+                 const vector<map<size_t, vector<NormalizedBBox>>>& allGTBBoxes,
+                 const vector<map<size_t, vector<pair<real, NormalizedBBox>>>>&
+                     allDetectBBoxes) {
+    for (size_t n = 0; n < allDetectBBoxes.size(); ++n) {
+      if (allGTBBoxes[n].size() == 0) {
+        for (map<size_t, vector<pair<real, NormalizedBBox>>>::const_iterator
+                 it = allDetectBBoxes[n].begin();
+             it != allDetectBBoxes[n].end();
+             ++it) {
+          size_t label = it->first;
+          for (size_t i = 0; i < it->second.size(); ++i) {
+            allTruePos_[label].push_back(make_pair(it->second[i].first, 0));
+            allFalsePos_[label].push_back(make_pair(it->second[i].first, 1));
+          }
+        }
+      } else {
+        for (map<size_t, vector<pair<real, NormalizedBBox>>>::const_iterator
+                 it = allDetectBBoxes[n].begin();
+             it != allDetectBBoxes[n].end();
+             ++it) {
+          size_t label = it->first;
+          vector<pair<real, NormalizedBBox>> predBBoxes = it->second;
+          if (allGTBBoxes[n].find(label) == allGTBBoxes[n].end()) {
+            for (size_t i = 0; i < predBBoxes.size(); ++i) {
+              allTruePos_[label].push_back(make_pair(predBBoxes[i].first, 0));
+              allFalsePos_[label].push_back(make_pair(predBBoxes[i].first, 1));
+            }
+          } else {
+            vector<NormalizedBBox> gtBBoxes =
+                allGTBBoxes[n].find(label)->second;
+            vector<bool> visited(gtBBoxes.size(), false);
+            // Sort detections in descend order based on scores
+            std::sort(predBBoxes.begin(),
+                      predBBoxes.end(),
+                      sortScorePairDescend<NormalizedBBox>);
+            for (size_t i = 0; i < predBBoxes.size(); ++i) {
+              real maxOverlap = -1.0;
+              size_t maxIdx = 0;
+              for (size_t j = 0; j < gtBBoxes.size(); ++j) {
+                real overlap =
+                    jaccardOverlap(predBBoxes[i].second, gtBBoxes[j]);
+                if (overlap > maxOverlap) {
+                  maxOverlap = overlap;
+                  maxIdx = j;
+                }
+              }
+              if (maxOverlap > overlapThreshold_) {
+                if (evaluateDifficult_ ||
+                    (!evaluateDifficult_ && !gtBBoxes[maxIdx].isDifficult)) {
+                  if (!visited[maxIdx]) {
+                    allTruePos_[label].push_back(
+                        make_pair(predBBoxes[i].first, 1));
+                    allFalsePos_[label].push_back(
+                        make_pair(predBBoxes[i].first, 0));
+                    visited[maxIdx] = true;
+                  } else {
+                    allTruePos_[label].push_back(
+                        make_pair(predBBoxes[i].first, 0));
+                    allFalsePos_[label].push_back(
+                        make_pair(predBBoxes[i].first, 1));
+                  }
+                }
+              } else {
+                allTruePos_[label].push_back(make_pair(predBBoxes[i].first, 0));
+                allFalsePos_[label].push_back(
+                    make_pair(predBBoxes[i].first, 1));
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  real calcMAP() const {
+    real mAP = 0.0;
+    size_t count = 0;
+    for (map<size_t, size_t>::const_iterator it = numPos_.begin();
+         it != numPos_.end();
+         ++it) {
+      size_t label = it->first;
+      size_t labelNumPos = it->second;
+      if (labelNumPos == 0 || allTruePos_.find(label) == allTruePos_.end())
+        continue;
+      vector<pair<real, size_t>> labelTruePos = allTruePos_.find(label)->second;
+      vector<pair<real, size_t>> labelFalsePos =
+          allFalsePos_.find(label)->second;
+      // Compute average precision.
+      vector<size_t> tpCumSum;
+      getAccumulation(labelTruePos, &tpCumSum);
+      vector<size_t> fpCumSum;
+      getAccumulation(labelFalsePos, &fpCumSum);
+      std::vector<real> precision, recall;
+      size_t num = tpCumSum.size();
+      // Compute Precision.
+      for (size_t i = 0; i < num; ++i) {
+        CHECK_LE(tpCumSum[i], labelNumPos);
+        precision.push_back(static_cast<real>(tpCumSum[i]) /
+                            static_cast<real>(tpCumSum[i] + fpCumSum[i]));
+        recall.push_back(static_cast<real>(tpCumSum[i]) / labelNumPos);
+      }
+      // VOC2007 style
+      if (apType_ == "11point") {
+        vector<real> maxPrecisions(11, 0.0);
+        int startIdx = num - 1;
+        for (int j = 10; j >= 0; --j)
+          for (int i = startIdx; i >= 0; --i) {
+            if (recall[i] < j / 10.) {
+              startIdx = i;
+              if (j > 0) maxPrecisions[j - 1] = maxPrecisions[j];
+              break;
+            } else {
+              if (maxPrecisions[j] < precision[i])
+                maxPrecisions[j] = precision[i];
+            }
+          }
+        for (int j = 10; j >= 0; --j) mAP += maxPrecisions[j] / 11;
+        ++count;
+      } else if (apType_ == "Integral") {
+        // Nature integral
+        real averagePrecisions = 0.;
+        real prevRecall = 0.;
+        for (size_t i = 0; i < num; ++i) {
+          if (fabs(recall[i] - prevRecall) > 1e-6)
+            averagePrecisions += precision[i] * fabs(recall[i] - prevRecall);
+          prevRecall = recall[i];
+        }
+        mAP += averagePrecisions;
+        ++count;
+      } else {
+        LOG(FATAL) << "Unkown ap version: " << apType_;
+      }
+    }
+    if (count != 0) mAP /= count;
+    return mAP * 100;
+  }
+
+  void getAccumulation(vector<pair<real, size_t>> inPairs,
+                       vector<size_t>* accuVec) const {
+    std::stable_sort(
+        inPairs.begin(), inPairs.end(), sortScorePairDescend<size_t>);
+    accuVec->clear();
+    size_t sum = 0;
+    for (size_t i = 0; i < inPairs.size(); ++i) {
+      sum += inPairs[i].second;
+      accuVec->push_back(sum);
+    }
+  }
+
+  std::string getTypeImpl() const { return "detection_map"; }
+
+  real getValueImpl() const { return calcMAP(); }
+
+private:
+  real overlapThreshold_;  // overlap threshold when determining whether matched
+  bool evaluateDifficult_;  // whether evaluate difficult ground truth
+  size_t backgroundId_;     // class index of background
+  std::string apType_;      // how to calculate mAP (Integral or 11point)
+
+  MatrixPtr cpuOutput_;
+  MatrixPtr cpuLabel_;
+
+  map<size_t, size_t> numPos_;  // counts of true objects each classification
+  map<size_t, vector<pair<real, size_t>>>
+      allTruePos_;  // true positive prediction
+  map<size_t, vector<pair<real, size_t>>>
+      allFalsePos_;  // false positive prediction
+};
+
+REGISTER_EVALUATOR(detection_map, DetectionMAPEvaluator);
+
+}  // namespace paddle

paddle/gserver/gradientmachines/NeuralNetwork.cpp

@@ -241,11 +241,14 @@ void NeuralNetwork::forward(const std::vector<Argument>& inArgs,
     dataLayers_[i]->setData(inArgs[i]);
   }
 
+  gLayerStackTrace.set_stage(true);
+
   {
     for (auto& layer : layers_) {
       REGISTER_TIMER_INFO("ForwardTimer", layer->getName().c_str());
       gLayerStackTrace.push(layer->getName());
       layer->forward(passType);
+      gLayerStackTrace.pop(layer->getName());
     }
   }
 
@@ -254,9 +257,6 @@ void NeuralNetwork::forward(const std::vector<Argument>& inArgs,
   for (auto& layer : outputLayers_) {
     outArgs->push_back(layer->getOutput());
   }
-  if (passType == PASS_TEST) {
-    gLayerStackTrace.clear();
-  }
 }
 
 void NeuralNetwork::resetState() {
@@ -283,9 +283,10 @@ void NeuralNetwork::getState(MachineState& machineState) {
 }
 
 void NeuralNetwork::backward(const UpdateCallback& callback) {
-  gLayerStackTrace.pop("");  // tell layer trace is during backward.
+  gLayerStackTrace.set_stage(false);
   FOR_EACH_R(layer, layers_) {
     REGISTER_TIMER_INFO("BackwardTimer", (*layer)->getName().c_str());
+    gLayerStackTrace.push((*layer)->getName());
     if ((*layer)->needGradient()) {
       (*layer)->backward(callback);
     }
@@ -320,7 +321,7 @@ public:
     }
   }
 
-  virtual void eval(const NeuralNetwork& nn) {
+  virtual void eval(const NeuralNetwork& nn) override {
     for (auto& evaluator : evaluators_) {
       evaluator->eval(nn);
     }
@@ -395,6 +396,30 @@ 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 +446,15 @@ 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));

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.

paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp

@@ -208,6 +208,7 @@ void RecurrentGradientMachine::init(
                    });
   CHECK(subModelConfig != config.sub_models().end());
   reversed_ = subModelConfig->reversed();
+  generating_ = subModelConfig->has_generator();
 
   inFrameLines_.resize(subModelConfig->in_links_size());
   for (size_t i = 0; i < inFrameLines_.size(); ++i) {
@@ -287,10 +288,6 @@ void RecurrentGradientMachine::init(
       parameterIds_.push_back(para->getID());
     }
   }
-
-  if (subModelConfig->evaluator_names_size() > 0) {
-    evaluator_.reset(frames_[0]->makeEvaluator());
-  }
 }
 
 void RecurrentGradientMachine::resizeOrCreateFrames(int numFrames) {
@@ -538,7 +535,7 @@ void RecurrentGradientMachine::forward(const std::vector<Argument>& inArgs,
      The outputs are outFramesLines_[i].agentLayer
    */
 
-  if (inFrameLines_.empty() && passType == PASS_TEST) {
+  if (generating_) {
     generateSequence();
     return;
   }  // else forward..
@@ -561,14 +558,14 @@ 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);
 }
 
 void RecurrentGradientMachine::backward(const UpdateCallback& callback) {
+  if (generating_) {
+    return;
+  }
   REGISTER_TIMER_INFO("RecurrentBwTime", "RecurrentBwTime");
   AsyncGpuBlock asyncGpuBlock;
   for (int i = maxSequenceLength_ - 1; i >= 0; --i) {
@@ -577,11 +574,6 @@ 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 +587,9 @@ 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 +1085,6 @@ 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;
@@ -1321,11 +1309,10 @@ void RecurrentGradientMachine::fillGenOutputs() {
 
   batchMachineIdVec_.clear();
   generator_.ids.clear();
+  int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false);
+  starts[0] = 0;
   if (numResults > 1) {
     real* probs = generator_.outArg.in->getData();
-    int* starts =
-        generator_.outArg.sequenceStartPositions->getMutableData(false);
-    starts[0] = 0;
     for (size_t i = 0; i < finalPaths_.size(); ++i) {
       for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
         Path& path = finalPaths_[i][j];
@@ -1348,7 +1335,10 @@ void RecurrentGradientMachine::fillGenOutputs() {
   } else {
     for (size_t i = 0; i < finalPaths_.size(); ++i) {
       CHECK(!finalPaths_[i].empty());
-      generator_.ids = finalPaths_[i][0].ids;
+      generator_.ids.insert(generator_.ids.begin(),
+                            finalPaths_[i][0].ids.begin(),
+                            finalPaths_[i][0].ids.end());
+      starts[i + 1] = starts[i] + finalPaths_[i][0].ids.size();
     }
   }
 }

paddle/gserver/gradientmachines/RecurrentGradientMachine.h

@@ -414,6 +414,7 @@ protected:
     std::vector<int> ids;  // store generated sequences
     Argument outArg;       // final output argument
   };
+  bool generating_;
   Generator generator_;
 
   std::vector<std::unique_ptr<NeuralNetwork>> frames_;
@@ -428,8 +429,6 @@ 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_

paddle/gserver/layers/AgentLayer.cpp

@@ -109,6 +109,40 @@ void GatherAgentLayer::forwardValue(PassType passType) {
   }
 }
 
+namespace {
+
+// dest[index[i]] <- src[i] for each i
+void copyElements(const IVector& srcVec,
+                  const IVector& indexVec,
+                  IVector& destVec) {
+  const int* src = srcVec.getData();
+  const int* index = indexVec.getData();
+  int* dest = destVec.getData();
+  int len = indexVec.getSize();
+  CHECK_EQ(srcVec.getSize(), indexVec.getSize());
+  for (int i = 0; i < len; ++i) {
+    dest[index[i]] = src[i];
+  }
+}
+}
+
+void GatherAgentLayer::forwardIds(PassType passType) {
+  IVectorPtr realId = realLayers_[0]->getOutputLabel();
+  if (!realId) return;
+
+  IVector::resizeOrCreate(output_.ids, allIds_->getSize(), useGpu_);
+  IVectorPtr outId = output_.ids;
+  idsVec_.resize(idIndex_.size());
+
+  for (size_t i = 0; i < realLayers_.size(); ++i) {
+    const IVectorPtr& realId = realLayers_[i]->getOutputLabel();
+    idsVec_[i] = IVector::create(allIds_->getData() + idIndex_[i],
+                                 /* size */ realId->getSize(),
+                                 useGpu_);
+    execViaCpu(&copyElements, *realId, *idsVec_[i], *outId);
+  }
+}
+
 void GatherAgentLayer::backward(const UpdateCallback& callback) {
   (void)callback;
   const MatrixPtr& outputGrad = getOutputGrad();
@@ -136,23 +170,22 @@ void ScatterAgentLayer::forward(PassType passType) {
   CHECK_EQ(realLayer_->getDeviceId(), this->getDeviceId());
 
   int width = this->getSize();
+  if (selectionMode_) {
+    forwardWithSelection(passType);
+  } else {
     if (realOutArg_.hasSeq()) {
-    forwardSequence(passType);
-  } else if (realOutArg_.value || realOutArg_.ids) {
+      output_.subArgFrom(realOutArg_,
+                         /* offset */ idIndex_,
+                         idSize_,
+                         width,
+                         useGpu_,
+                         /* trans */ false,
+                         /* seqFlag */ true,
+                         /* seqStart */ seqStartPosIndex_,
+                         /* seqSize */ numSequences_);
+    } else {
       output_.subArgFrom(
           realOutArg_, /* offset */ idIndex_, idSize_, width, useGpu_);
-  } else {  // used in generation
-    if (realLayer_->getOutput().ids) {
-      IVector::resizeOrCreate(output_.ids, ids_->getSize(), useGpu_);
-      output_.ids->selectFrom(*realLayer_->getOutput().ids, *ids_);
-    }
-    if (realLayer_->getOutput().value) {
-      int height = ids_->getSize();
-      resetOutput(height, width);
-
-      const MatrixPtr& outV = getOutputValue();
-      const MatrixPtr& realV = realLayer_->getOutputValue();
-      outV->selectRows(*realV, *ids_);
     }
   }
 }
@@ -160,6 +193,8 @@ void ScatterAgentLayer::forward(PassType passType) {
 void ScatterAgentLayer::backward(const UpdateCallback& callback) {
   (void)callback;
 
+  CHECK(!selectionMode_);
+
   const MatrixPtr& outputGrad = realOutArg_.grad;
   const MatrixPtr& realGrad = realLayer_->getOutputGrad();
   if (realGrad) {
@@ -174,42 +209,7 @@ void ScatterAgentLayer::backward(const UpdateCallback& callback) {
 REGISTER_LAYER(gather_agent, GatherAgentLayer);
 REGISTER_LAYER(scatter_agent, ScatterAgentLayer);
 
-void GatherAgentLayer::forwardIds(PassType passType) {
-  int height = 0;
-  IVectorPtr idReal = realLayers_[0]->getOutputLabel();
-
-  if (!idReal) return;
-
-  if (output_.subSequenceStartPositions) {
-    int* starts = output_.subSequenceStartPositions->getMutableData(false);
-    // Gather generator.idsVec
-    // if is beam search generation result. Get first result.
-    if (idReal->getData()[idReal->getSize() - 1] == -1) {
-      for (size_t i = 0; i < realLayers_.size(); ++i) {
-        // The first element stores first result size
-        idReal = realLayers_[i]->getOutputLabel();
-        idReal->subVecFrom(*idReal, 1, idReal->getData()[0]);
-      }
-    }
-    for (size_t i = 0; i < realLayers_.size(); ++i) {
-      CHECK(realLayers_[i]->getOutputLabel());
-      starts[i] = height;
-      height += realLayers_[i]->getOutputLabel()->getSize();
-    }
-    starts[realLayers_.size()] = height;
-    output_.sequenceStartPositions->getMutableData(false)[1] = height;
-
-    IVector::resizeOrCreate(output_.ids, height, false);
-    for (size_t i = 0; i < realLayers_.size(); ++i) {
-      output_.ids->subVec(starts[i], starts[i + 1] - starts[i])
-          ->copyFrom(*realLayers_[i]->getOutputLabel());
-    }
-  } else {
-    LOG(FATAL) << "Not implemented";
-  }
-}
-
-void ScatterAgentLayer::forwardSequence(PassType passType) {
+void ScatterAgentLayer::forwardWithSelection(PassType passType) {
   Layer::forward(passType);
   CHECK_EQ(realLayer_->getDeviceId(), this->getDeviceId());
 
@@ -220,17 +220,19 @@ void ScatterAgentLayer::forwardSequence(PassType passType) {
   AsyncGpuBlock asyncGpuBlock;
   REGISTER_TIMER_INFO("SequenceAgentLayerForward", getName().c_str());
 
-  if (realOutArg_.value || realOutArg_.ids) {
-    CHECK(realOutArg_.sequenceStartPositions);
-    output_.subArgFrom(realOutArg_,
-                       /* offset */ idIndex_,
-                       idSize_,
-                       width,
-                       useGpu_,
-                       /* trans */ false,
-                       /* seqFlag */ true,
-                       /* seqStart */ seqStartPosIndex_,
-                       /* seqSize */ numSequences_);
+  if (!input.hasSeq()) {
+    if (realLayer_->getOutput().ids) {
+      IVector::resizeOrCreate(output_.ids, ids_->getSize(), useGpu_);
+      output_.ids->selectFrom(*realLayer_->getOutput().ids, *ids_);
+    }
+    if (realLayer_->getOutput().value) {
+      int height = ids_->getSize();
+      resetOutput(height, width);
+
+      const MatrixPtr& outV = getOutputValue();
+      const MatrixPtr& realV = realLayer_->getOutputValue();
+      outV->selectRows(*realV, *ids_);
+    }
   } else {
     // Putting the generation logic here is really an ugly hack!
     // used in generation

paddle/gserver/layers/AgentLayer.h

@@ -110,6 +110,9 @@ protected:
   // of real layer.
   ICpuGpuVectorPtr inputStartPos_;
 
+  // true for setRealLayer, false for setRealLayerAndOutput
+  bool selectionMode_;
+
 public:
   explicit ScatterAgentLayer(const LayerConfig& config) : Layer(config) {}
 
@@ -137,6 +140,7 @@ public:
     } else {
       cpuIds_ = ids_;
     }
+    selectionMode_ = true;
   }
 
   // set real layer and output, [idIndex, idIndex + idSize) of *ids*
@@ -153,6 +157,7 @@ public:
     idIndex_ = idIndex;
     idSize_ = idSize;
     handleBackward_ = handleBackward;
+    selectionMode_ = false;
   }
 
   void setSequenceStartPositions(const ICpuGpuVectorPtr& sequenceStartPositions,
@@ -166,7 +171,7 @@ public:
   void forward(PassType passType) override;
   void backward(const UpdateCallback& callback) override;
 
-  void forwardSequence(PassType passType);
+  void forwardWithSelection(PassType passType);
 };
 
 }  // namespace paddle

paddle/gserver/tests/test_Evaluator.cpp

@@ -138,6 +138,23 @@ void testEvaluatorAll(TestConfig testConf,
   testEvaluator(testConf, testEvaluatorName, batchSize, false);
 }
 
+TEST(Evaluator, detection_map) {
+  TestConfig config;
+  config.evaluatorConfig.set_type("detection_map");
+  config.evaluatorConfig.set_overlap_threshold(0.5);
+  config.evaluatorConfig.set_background_id(0);
+  config.evaluatorConfig.set_ap_type("Integral");
+  config.evaluatorConfig.set_evaluate_difficult(0);
+
+  config.inputDefs.push_back({INPUT_DATA, "output", 7});
+  config.inputDefs.push_back({INPUT_SEQUENCE_DATA, "label", 6});
+  config.evaluatorConfig.set_evaluate_difficult(false);
+  testEvaluatorAll(config, "detection_map", 100);
+
+  config.evaluatorConfig.set_evaluate_difficult(true);
+  testEvaluatorAll(config, "detection_map", 100);
+}
+
 TEST(Evaluator, classification_error) {
   TestConfig config;
   config.evaluatorConfig.set_type("classification_error");

paddle/trainer/tests/sample_trainer_nest_rnn_gen.conf

@@ -53,7 +53,7 @@ def outer_step(dummy_data):
                            bos_id=0,
                            eos_id=num_words-1,
                            beam_size=2 if beam_flag else 1,
-                           num_results_per_sample=2 if beam_flag else 1,
+                           num_results_per_sample=1,
                            max_length=10)
     return beam_gen
 

paddle/utils/CustomStackTrace.h

@@ -55,13 +55,17 @@ public:
    *        Else, just set status to popping.
    */
   void pop(const T& item) {
-    pushing() = false;
     auto& s = this->stack();
     if (item == s.top()) {
       s.pop();
     }
   }
 
+  /**
+   * @brief Indicate whether we are at forward or backward stage of computation
+   */
+  void set_stage(bool isForward) { pushing() = isForward; }
+
   /**
    * @brief clear current thread stack.
    */

paddle/utils/tests/test_CustomStackTrace.cpp

@@ -72,7 +72,6 @@ TEST(CustomStackTrace, normalTrain) {
       for (size_t i = 0; i < layerSize; ++i) {
         tracer.push("layer_" + paddle::str::to_string(i));
       }
-      tracer.pop("");
       for (size_t i = 0; i < layerSize; ++i) {
         tracer.pop("layer_" + paddle::str::to_string(layerSize - 1 - i));
       }

proto/ModelConfig.proto

@@ -489,6 +489,15 @@ message EvaluatorConfig {
   // Used by ClassificationErrorEvaluator
   // top # classification error
   optional int32 top_k = 13 [default = 1];
+
+  // Used by DetectionMAPEvaluator
+  optional double overlap_threshold = 14 [default = 0.5];
+
+  optional int32 background_id = 15 [default = 0];
+
+  optional bool evaluate_difficult = 16 [default = false];
+
+  optional string ap_type = 17 [default = "11point"];
 }
 
 message LinkConfig {

python/paddle/trainer/config_parser.py

@@ -1280,8 +1280,7 @@ def parse_maxout(maxout, input_layer_name, maxout_conf):
 
 # Define an evaluator
 @config_func
-def Evaluator(
-        name,
+def Evaluator(name,
               type,
               inputs,
               chunk_scheme=None,
@@ -1293,7 +1292,11 @@ def Evaluator(
               num_results=None,
               top_k=None,
               delimited=None,
-        excluded_chunk_types=None, ):
+              excluded_chunk_types=None,
+              overlap_threshold=None,
+              background_id=None,
+              evaluate_difficult=None,
+              ap_type=None):
     evaluator = g_config.model_config.evaluators.add()
     evaluator.type = type
     evaluator.name = MakeLayerNameInSubmodel(name)
@@ -1327,6 +1330,18 @@ def Evaluator(
     if excluded_chunk_types:
         evaluator.excluded_chunk_types.extend(excluded_chunk_types)
 
+    if overlap_threshold is not None:
+        evaluator.overlap_threshold = overlap_threshold
+
+    if background_id is not None:
+        evaluator.background_id = background_id
+
+    if evaluate_difficult is not None:
+        evaluator.evaluate_difficult = evaluate_difficult
+
+    if ap_type is not None:
+        evaluator.ap_type = ap_type
+
 
 class LayerBase(object):
     def __init__(

python/paddle/trainer_config_helpers/evaluators.py

@@ -21,7 +21,8 @@ __all__ = [
     "chunk_evaluator", "sum_evaluator", "column_sum_evaluator",
     "value_printer_evaluator", "gradient_printer_evaluator",
     "maxid_printer_evaluator", "maxframe_printer_evaluator",
-    "seqtext_printer_evaluator", "classification_error_printer_evaluator"
+    "seqtext_printer_evaluator", "classification_error_printer_evaluator",
+    "detection_map_evaluator"
 ]
 
 
@@ -31,10 +32,11 @@ class EvaluatorAttribute(object):
     FOR_RANK = 1 << 2
     FOR_PRINT = 1 << 3
     FOR_UTILS = 1 << 4
+    FOR_DETECTION = 1 << 5
 
     KEYS = [
         "for_classification", "for_regression", "for_rank", "for_print",
-        "for_utils"
+        "for_utils", "for_detection"
     ]
 
     @staticmethod
@@ -57,8 +59,7 @@ def evaluator(*attrs):
     return impl
 
 
-def evaluator_base(
-        input,
+def evaluator_base(input,
                    type,
                    label=None,
                    weight=None,
@@ -72,7 +73,11 @@ def evaluator_base(
                    num_results=None,
                    delimited=None,
                    top_k=None,
-        excluded_chunk_types=None, ):
+                   excluded_chunk_types=None,
+                   overlap_threshold=None,
+                   background_id=None,
+                   evaluate_difficult=None,
+                   ap_type=None):
     """
     Evaluator will evaluate the network status while training/testing.
 
@@ -107,6 +112,14 @@ def evaluator_base(
     :type weight: LayerOutput.
     :param top_k: number k in top-k error rate
     :type top_k: int
+    :param overlap_threshold: In detection tasks to filter detection results
+    :type overlap_threshold: float
+    :param background_id: Identifier of background class
+    :type background_id: int
+    :param evaluate_difficult: Whether to evaluate difficult objects
+    :type evaluate_difficult: bool
+    :param ap_type: How to calculate average persicion
+    :type ap_type: str
     """
     # inputs type assertions.
     assert classification_threshold is None or isinstance(
@@ -136,7 +149,61 @@ def evaluator_base(
         delimited=delimited,
         num_results=num_results,
         top_k=top_k,
-        excluded_chunk_types=excluded_chunk_types, )
+        excluded_chunk_types=excluded_chunk_types,
+        overlap_threshold=overlap_threshold,
+        background_id=background_id,
+        evaluate_difficult=evaluate_difficult,
+        ap_type=ap_type)
+
+
+@evaluator(EvaluatorAttribute.FOR_DETECTION)
+@wrap_name_default()
+def detection_map_evaluator(input,
+                            label,
+                            overlap_threshold=0.5,
+                            background_id=0,
+                            evaluate_difficult=False,
+                            ap_type="11point",
+                            name=None):
+    """
+    Detection mAP Evaluator. It will print mean Average Precision (mAP) for detection.
+
+    The detection mAP Evaluator based on the output of detection_output layer counts
+    the true positive and the false positive bbox and integral them to get the
+    mAP.
+
+    The simple usage is:
+
+    .. code-block:: python
+
+       eval =  detection_map_evaluator(input=det_output,label=lbl)
+
+    :param input: Input layer.
+    :type input: LayerOutput
+    :param label: Label layer.
+    :type label: LayerOutput
+    :param overlap_threshold: The bbox overlap threshold of a true positive.
+    :type overlap_threshold: float
+    :param background_id: The background class index.
+    :type background_id: int
+    :param evaluate_difficult: Whether evaluate a difficult ground truth.
+    :type evaluate_difficult: bool
+    """
+    if not isinstance(input, list):
+        input = [input]
+
+    if label:
+        input.append(label)
+
+    evaluator_base(
+        name=name,
+        type="detection_map",
+        input=input,
+        label=label,
+        overlap_threshold=overlap_threshold,
+        background_id=background_id,
+        evaluate_difficult=evaluate_difficult,
+        ap_type=ap_type)
 
 
 @evaluator(EvaluatorAttribute.FOR_CLASSIFICATION)

python/paddle/trainer_config_helpers/layers.py

@@ -3839,7 +3839,8 @@ def classification_cost(input,
                         weight=None,
                         name=None,
                         evaluator=classification_error_evaluator,
-                        layer_attr=None):
+                        layer_attr=None,
+                        coeff=1.):
     """
     classification cost Layer.
 
@@ -3855,6 +3856,8 @@ def classification_cost(input,
     :param evaluator: Evaluator method.
     :param layer_attr: layer's extra attribute.
     :type layer_attr: ExtraLayerAttribute
+    :param coeff: The coefficient affects the gradient in the backward.
+    :type coeff: float
     :return: LayerOutput object.
     :rtype: LayerOutput
     """
@@ -3868,6 +3871,7 @@ def classification_cost(input,
         name=name,
         type="multi-class-cross-entropy",
         inputs=ipts,
+        coeff=coeff,
         **ExtraLayerAttribute.to_kwargs(layer_attr))
 
     def __add_evaluator__(e):

python/paddle/v2/layer.py

@@ -45,12 +45,12 @@ __all__ = ['data', 'parse_network']
 def __need_to_keep__(name):
     return name in [
         'StaticInput', 'SubsequenceInput', 'GeneratedInput', 'LayerType',
-        'layer_support'
+        'layer_support', 'BaseGeneratedInput'
     ]
 
 
 def __need_to_wrap__(name):
-    return name not in ['AggregateLevel', 'ExpandLevel']
+    return name not in ['AggregateLevel', 'ExpandLevel', 'BaseGeneratedInput']
 
 
 def __convert_name__(inname):
@@ -199,6 +199,15 @@ def __get_used_submodels__(layer_names):
     return submodel_names
 
 
+def __get_submodel_data_out_links__():
+    data_links = set()
+    for submodel in cp.g_config.model_config.sub_models:
+        for link in submodel.out_links:
+            if cp.g_layer_map[link.link_name].type == 'data':
+                data_links.add(link.link_name)
+    return data_links
+
+
 def __get_used_evaluators__(layer_names):
     evaluator_names = set()
     for e in cp.g_config.model_config.evaluators:
@@ -264,6 +273,7 @@ def parse_network(output_layers, extra_layers=None):
     submodel_names = __get_used_submodels__(layer_names)
     submodel_names.add('root')
     evaluator_names = __get_used_evaluators__(layer_names)
+    data_out_links = __get_submodel_data_out_links__()
     input_layer_names = set()
     output_layer_names = set()
 
@@ -279,7 +289,7 @@ def parse_network(output_layers, extra_layers=None):
             continue
         model_config.layers.extend([l])
         if l.type == 'data':
-            if l.name in model_config.output_layer_names:
+            if l.name in data_out_links:
                 """
                 In text generation, the outlink to save the generated word
                 indices is a data_layer defined in recurrent_group. This