diff --git a/go/pserver/optimizer.go b/go/pserver/optimizer.go
index ae7359073494bd9cb6b70b12af4daca064179556..51ffba5c74d82b7f24d5ab6218e47479c4d18658 100644
--- a/go/pserver/optimizer.go
+++ b/go/pserver/optimizer.go
@@ -72,21 +72,34 @@ func newOptimizer(paramWithConfigs ParameterWithConfig, State []byte) *optimizer
 	}
 
 	o.config = c
-	o.opt = C.paddle_create_optimizer((*C.uchar)(&c[0]), C.int(len(c)),
-		C.paddle_element_type(p.ElementType), cbuffer, C.int(paramBufferSize), (*C.char)(cstate), C.int(len(s)))
+	o.opt = C.paddle_create_optimizer(
+		(*C.uchar)(&c[0]),
+		C.int(len(c)),
+		C.paddle_element_type(p.ElementType),
+		cbuffer,
+		C.int(paramBufferSize),
+		(*C.char)(cstate),
+		C.int(len(s)),
+	)
 	return o
 }
 
 func (o *optimizer) GetWeights() []byte {
 	var buffer unsafe.Pointer
+	// we do not own the buffer, no need to free later.
 	bufferLen := C.paddle_optimizer_get_weights(o.opt, &buffer)
 	return cArrayToSlice(buffer, int(bufferLen)*C.sizeof_float)
 }
 
 func (o *optimizer) GetStates() []byte {
 	var cbuffer *C.char
+	// we owns the state buffer, need to free later.
 	cbufferLen := C.paddle_optimizer_get_state(o.opt, &cbuffer)
-	return cArrayToSlice(unsafe.Pointer(cbuffer), int(cbufferLen))
+	buf := cArrayToSlice(unsafe.Pointer(cbuffer), int(cbufferLen))
+	cpy := make([]byte, len(buf))
+	copy(cpy, buf)
+	C.free(unsafe.Pointer(cbuffer))
+	return cpy
 }
 
 func (o *optimizer) UpdateParameter(g Gradient) error {
diff --git a/go/pserver/optimizer_test.go b/go/pserver/optimizer_test.go
index d001e6993e6aed2f5829c1b86928af30f4900c8a..565f56dc286d214e7e9a3ddce389d92d21569cd5 100644
--- a/go/pserver/optimizer_test.go
+++ b/go/pserver/optimizer_test.go
@@ -15,8 +15,12 @@
 package pserver
 
 import (
+	"encoding/binary"
 	"io/ioutil"
+	"math"
 	"testing"
+
+	"github.com/stretchr/testify/assert"
 )
 
 func TestOptimizerCreateRelease(t *testing.T) {
@@ -36,3 +40,39 @@ func TestOptimizerCreateRelease(t *testing.T) {
 	o := newOptimizer(param, nil)
 	o.Cleanup()
 }
+
+func float32Bytes(float float32) []byte {
+	bits := math.Float32bits(float)
+	bytes := make([]byte, 4)
+	binary.LittleEndian.PutUint32(bytes, bits)
+	return bytes
+}
+
+func TestOptimizerState(t *testing.T) {
+	p := Parameter{
+		Name:        "a",
+		ElementType: Int32,
+	}
+	weights := float32Bytes(100)
+	p.Content = weights
+	config, err := ioutil.ReadFile("./client/c/test/testdata/optimizer.pb")
+	if err != nil {
+		t.Fatalf("read optimizer proto failed")
+	}
+	param := ParameterWithConfig{
+		Param:  p,
+		Config: config,
+	}
+	o := newOptimizer(param, nil)
+	s := o.GetStates()
+
+	// clear param content and check if the state is restored.
+	param.Param.Content = float32Bytes(300)
+	o1 := newOptimizer(param, s)
+	s1 := o1.GetStates()
+	assert.Equal(t, s, s1)
+	assert.Equal(t, weights, o.GetWeights())
+	assert.Equal(t, weights, o1.GetWeights())
+	o.Cleanup()
+	o1.Cleanup()
+}
diff --git a/go/pserver/service.go b/go/pserver/service.go
index 25751540a9a2dff043c14e0912bfab1aaa938ab4..29e953acdd8ae16d13cf2307e212f8a18f0f2190 100644
--- a/go/pserver/service.go
+++ b/go/pserver/service.go
@@ -297,6 +297,13 @@ func (s *Service) checkpoint() (err error) {
 		return
 	}
 
+	if _, err = os.Stat(s.checkpointPath); os.IsNotExist(err) {
+		err = os.MkdirAll(s.checkpointPath, os.ModePerm)
+		if err != nil {
+			return
+		}
+	}
+
 	id := uuid.NewV4().String()
 	p := path.Join(s.checkpointPath, id)
 	f, err := os.Create(p)
diff --git a/paddle/optimizer/adadelta_optimizer.cc b/paddle/optimizer/adadelta_optimizer.cc
index 6eec5d846fa5ef6b25e7646200dad1d452dda806..34913c405075ed72af30ed056f74e8b4d7482488 100644
--- a/paddle/optimizer/adadelta_optimizer.cc
+++ b/paddle/optimizer/adadelta_optimizer.cc
@@ -25,19 +25,17 @@ void AdadeltaOptimizer::Update(const Tensor* gradient) {
   }
 }
 
-const char* AdadeltaOptimizer::SerializeState(int* state_len) {
+std::string AdadeltaOptimizer::SerializeState() {
   AdadeltaOptimizerState state;
   state.set_num_sample_passed(num_sample_passed_);
-  std::string lr_str = this->lr_policy_->SerializeState(state_len);
+  std::string lr_str = this->lr_policy_->SerializeState();
   state.mutable_lr_state()->ParseFromString(lr_str);
 
   TensorToProto(*parameter_, state.mutable_parameter());
   TensorToProto(*accum_gradient_, state.mutable_accum_gradient());
   TensorToProto(*accum_delta_, state.mutable_accum_delta());
   TensorToProto(*update_delta_, state.mutable_update_delta());
-  auto str = state.SerializeAsString();
-  *state_len += str.size();
-  return str.c_str();
+  return state.SerializeAsString();
 }
 
 void AdadeltaOptimizer::DeserializeState(const std::string& str) {
diff --git a/paddle/optimizer/adadelta_optimizer.h b/paddle/optimizer/adadelta_optimizer.h
index 1d5eab097f57d049855dd171a1aa6f74c48ae0e7..bc634ee46d60abc9ffc4a31abac5c2f8edaf7aba 100644
--- a/paddle/optimizer/adadelta_optimizer.h
+++ b/paddle/optimizer/adadelta_optimizer.h
@@ -23,7 +23,7 @@ public:
     if (update_delta_) delete update_delta_;
   }
   void Update(const Tensor *gradient);
-  const char *SerializeState(int *state_len);
+  std::string SerializeState();
   void DeserializeState(const std::string &state);
 
 private:
diff --git a/paddle/optimizer/adagrad_optimizer.cc b/paddle/optimizer/adagrad_optimizer.cc
index 5b92610ac547ee11cedf2e49e4d7f1db4b2da646..d915ffb8705eaa96bc96b8071a2c534d4d472273 100644
--- a/paddle/optimizer/adagrad_optimizer.cc
+++ b/paddle/optimizer/adagrad_optimizer.cc
@@ -17,17 +17,15 @@ void AdagradOptimizer::Update(const Tensor* gradient) {
                 learning_rate * decay_ * param[i];
   }
 }
-const char* AdagradOptimizer::SerializeState(int* state_len) {
+std::string AdagradOptimizer::SerializeState() {
   AdagradOptimizerState state;
   state.set_num_sample_passed(num_sample_passed_);
-  std::string lr_str = this->lr_policy_->SerializeState(state_len);
+  std::string lr_str = this->lr_policy_->SerializeState();
   state.mutable_lr_state()->ParseFromString(lr_str);
 
   TensorToProto(*parameter_, state.mutable_parameter());
   TensorToProto(*accum_gradient_, state.mutable_accum_gradient());
-  auto str = state.SerializeAsString();
-  *state_len += str.size();
-  return str.c_str();
+  return state.SerializeAsString();
 }
 
 void AdagradOptimizer::DeserializeState(const std::string& str) {
diff --git a/paddle/optimizer/adagrad_optimizer.h b/paddle/optimizer/adagrad_optimizer.h
index 15d0a965ad0c6967e73b14b465168fa66eb8fba3..b2935f8aff87f710f508c5c5757dd36526ca63f9 100644
--- a/paddle/optimizer/adagrad_optimizer.h
+++ b/paddle/optimizer/adagrad_optimizer.h
@@ -19,7 +19,7 @@ public:
     if (accum_gradient_) delete accum_gradient_;
   }
   void Update(const Tensor *gradient);
-  const char *SerializeState(int *state_len);
+  std::string SerializeState();
   void DeserializeState(const std::string &state);
 
 private:
diff --git a/paddle/optimizer/adam_optimizer.cc b/paddle/optimizer/adam_optimizer.cc
index 1ebb6b1e0f7b4edcbac1b28319fd4de576f85f6a..18e5896a22dc8a3c6292293fffc36ca9e3737b4c 100644
--- a/paddle/optimizer/adam_optimizer.cc
+++ b/paddle/optimizer/adam_optimizer.cc
@@ -22,18 +22,16 @@ void AdamOptimizer::Update(const Tensor *gradient) {
   }
 }
 
-const char *AdamOptimizer::SerializeState(int *state_len) {
+std::string AdamOptimizer::SerializeState() {
   AdamOptimizerState state;
-  std::string lr_str = this->lr_policy_->SerializeState(state_len);
+  std::string lr_str = this->lr_policy_->SerializeState();
   state.mutable_lr_state()->ParseFromString(lr_str);
   state.set_num_sample_passed(num_sample_passed_);
 
   TensorToProto(*parameter_, state.mutable_parameter());
   TensorToProto(*momentums_, state.mutable_momentums());
   TensorToProto(*velocitys_, state.mutable_velocitys());
-  auto str = state.SerializeAsString();
-  *state_len += str.size();
-  return str.c_str();
+  return state.SerializeAsString();
 }
 
 void AdamOptimizer::DeserializeState(const std::string &str) {
diff --git a/paddle/optimizer/adam_optimizer.h b/paddle/optimizer/adam_optimizer.h
index 0ea4c8bb8470504282b4d6c12039791ce896e401..d25cdc0731f65e9875d2fbf67783cce62d88af60 100644
--- a/paddle/optimizer/adam_optimizer.h
+++ b/paddle/optimizer/adam_optimizer.h
@@ -25,7 +25,7 @@ public:
     if (velocitys_) delete velocitys_;
   }
   void Update(const Tensor *gradient);
-  const char *SerializeState(int *state_len);
+  std::string SerializeState();
   void DeserializeState(const std::string &state);
 
 private:
diff --git a/paddle/optimizer/lr_policy.h b/paddle/optimizer/lr_policy.h
index 036c376e10f465c2866a230caf9224f4af5478bc..bbb1ee48214cecdc6b6cd2a400cc9d12d5e8b64a 100644
--- a/paddle/optimizer/lr_policy.h
+++ b/paddle/optimizer/lr_policy.h
@@ -10,7 +10,7 @@ class LrPolicy {
 public:
   virtual ~LrPolicy() {}
   virtual double LearningRate(const uint64_t num_sample_passed) = 0;
-  virtual const char *SerializeState(int *state_len) = 0;
+  virtual std::string SerializeState() = 0;
   virtual void DeserializeState(const std::string &state) = 0;
 };
 
@@ -21,12 +21,10 @@ public:
   double LearningRate(const uint64_t num_sample_passed) {
     return learning_rate_;
   }
-  const char *SerializeState(int *state_len) {
+  std::string SerializeState() {
     LrPolicyState state;
     state.set_learning_rate(learning_rate_);
-    auto str = state.SerializeAsString();
-    *state_len = str.size();
-    return str.c_str();
+    return state.SerializeAsString();
   }
   void DeserializeState(const std::string &str) {
     LrPolicyState state;
@@ -46,14 +44,12 @@ public:
     return std::max(learning_rate_ - lr_decay_a_ * num_sample_passed,
                     lr_decay_b_);
   }
-  const char *SerializeState(int *state_len) {
+  std::string SerializeState() {
     LrPolicyState state;
     state.set_learning_rate(learning_rate_);
     state.set_lr_decay_a(lr_decay_a_);
     state.set_lr_decay_b(lr_decay_b_);
-    auto str = state.SerializeAsString();
-    *state_len = str.size();
-    return str.c_str();
+    return state.SerializeAsString();
   }
   void DeserializeState(const std::string &str) {
     LrPolicyState state;
diff --git a/paddle/optimizer/optimizer.cc b/paddle/optimizer/optimizer.cc
index eb7125adee769c97e16986cabf06ea389bf4c143..a2af139d012433214b825bd68289708098b76da8 100644
--- a/paddle/optimizer/optimizer.cc
+++ b/paddle/optimizer/optimizer.cc
@@ -1,4 +1,7 @@
 #include "optimizer.h"
+#include <glog/logging.h>
+#include <cstdlib>
+#include <cstring>
 #include <string>
 
 #include "parameter_optimizer.h"
@@ -78,7 +81,13 @@ int paddle_optimizer_get_weights(paddle_optimizer* o, void** param_buffer) {
 }
 
 int paddle_optimizer_get_state(paddle_optimizer* o, const char** state) {
-  int state_len = 0;
-  *state = o->impl->SerializeState(&state_len);
+  std::string s = o->impl->SerializeState();
+  int state_len = s.size();
+
+  if (state_len > 0) {
+    *state = (char*)std::malloc(state_len);
+    std::memcpy((void*)*state, (const void*)s.c_str(), state_len);
+  }
+
   return state_len;
 }
diff --git a/paddle/optimizer/parameter_optimizer.cc b/paddle/optimizer/parameter_optimizer.cc
index f6218037925649e741d17f49af972ce2d50f8d3d..db0714635f9366b0404019688daf4708b4a0052f 100644
--- a/paddle/optimizer/parameter_optimizer.cc
+++ b/paddle/optimizer/parameter_optimizer.cc
@@ -32,6 +32,7 @@ ParameterOptimizer *ParameterOptimizer::Create(const std::string &config_proto,
       Tensor *parameter,
       const OptimizerConfig &config) -> ParameterOptimizer * {
     if (config.optimizer() == OptimizerConfig::SGD) {
+      LOG(INFO) << "creating SGD optimizer";
       return new SGDOptimizer(parameter,
                               lr,
                               config.sgd().momentum(),
@@ -39,6 +40,7 @@ ParameterOptimizer *ParameterOptimizer::Create(const std::string &config_proto,
                               config.sgd().nesterov());
     }
     if (config.optimizer() == OptimizerConfig::Adadelta) {
+      LOG(INFO) << "creating Adadelta optimizer";
       return new AdadeltaOptimizer(parameter,
                                    lr,
                                    config.adadelta().rho(),
@@ -46,10 +48,12 @@ ParameterOptimizer *ParameterOptimizer::Create(const std::string &config_proto,
                                    config.adadelta().decay());
     }
     if (config.optimizer() == OptimizerConfig::Adagrad) {
+      LOG(INFO) << "creating Adagrad optimizer";
       return new AdagradOptimizer(
           parameter, lr, config.adagrad().epsilon(), config.adagrad().decay());
     }
     if (config.optimizer() == OptimizerConfig::Adam) {
+      LOG(INFO) << "creating Adam optimizer";
       return new AdamOptimizer(parameter,
                                lr,
                                config.adam().beta_1(),
diff --git a/paddle/optimizer/parameter_optimizer.h b/paddle/optimizer/parameter_optimizer.h
index d89c9abb791f947172078d4dce5b1c366852591b..8319f84e1b820adf5cc0006045f2e13dffa91797 100644
--- a/paddle/optimizer/parameter_optimizer.h
+++ b/paddle/optimizer/parameter_optimizer.h
@@ -28,7 +28,7 @@ public:
                                     Tensor *parameter);
   virtual void Update(const Tensor *gradient) = 0;
   virtual float *get_weight(int *param_size) const;
-  virtual const char *SerializeState(int *state_len) = 0;
+  virtual std::string SerializeState() = 0;
   virtual void DeserializeState(const std::string &state) = 0;
 
 protected:
diff --git a/paddle/optimizer/parameter_optimizer_test.cpp b/paddle/optimizer/parameter_optimizer_test.cpp
index edf4ae37a9beee2911d23dd1ab23e67a18065b1b..c88fa11748716693355042d1784b33d7cfb616f1 100644
--- a/paddle/optimizer/parameter_optimizer_test.cpp
+++ b/paddle/optimizer/parameter_optimizer_test.cpp
@@ -85,6 +85,7 @@ public:
     for (size_t i = 0; i < opts_.size(); ++i) {
       int s = 0;
       float* newp = (float*)opts_[i]->get_weight(&s);
+      EXPECT_EQ(s, kSize);
       for (size_t j = 0; j < kSize; ++j) {
         EXPECT_EQ(newp[j], (*p)[j]);
       }
@@ -99,10 +100,20 @@ public:
   }
 
   void TestCheckPoint() {
+    paddle::optimizer::Tensor* p = FixedTensor(kSize);
     for (size_t i = 0; i < opts_.size(); ++i) {
-      int state_len = 0;
-      std::string state = opts_[i]->SerializeState(&state_len);
+      auto state = opts_[i]->SerializeState();
+      opts_[i]->DeserializeState(state);
+      auto state1 = opts_[i]->SerializeState();
       opts_[i]->DeserializeState(state);
+      EXPECT_EQ(state, state1);
+
+      int s = 0;
+      float* newp = (float*)opts_[i]->get_weight(&s);
+      EXPECT_EQ(s, kSize);
+      for (size_t j = 0; j < kSize; ++j) {
+        EXPECT_EQ(newp[j], (*p)[j]);
+      }
     }
   }
 
diff --git a/paddle/optimizer/serialization_test.cpp b/paddle/optimizer/serialization_test.cpp
index e4d97cbdba545c4ba5adf5b30efd3fc9f3f744ee..4c416f55ee0bd70f9ec6e288b08a5399d8b2bf39 100644
--- a/paddle/optimizer/serialization_test.cpp
+++ b/paddle/optimizer/serialization_test.cpp
@@ -21,7 +21,22 @@ TEST(TensorToProto, Case1) {
   paddle::optimizer::Tensor t(3), t1(3);
   for (size_t i = 0; i < t.size(); ++i) {
     t[i] = i;
-    t1[i] = 0;
+    t1[i] = 10;
+  }
+
+  paddle::TensorProto proto;
+  paddle::optimizer::TensorToProto(t, &proto);
+  paddle::optimizer::ProtoToTensor(proto, &t1);
+  for (size_t i = 0; i < t1.size(); ++i) {
+    EXPECT_EQ(t1[i], t[i]);
+  }
+}
+
+TEST(TensorToProto, Case2) {
+  paddle::optimizer::Tensor t(1), t1(1);
+  for (size_t i = 0; i < t.size(); ++i) {
+    t[i] = i;
+    t1[i] = 10;
   }
 
   paddle::TensorProto proto;
diff --git a/paddle/optimizer/sgd_optimizer.cc b/paddle/optimizer/sgd_optimizer.cc
index 15418faa840c19e776f293700ee886991754fb04..bf2540ecb092437e57a5970264559dc3c6ab4167 100644
--- a/paddle/optimizer/sgd_optimizer.cc
+++ b/paddle/optimizer/sgd_optimizer.cc
@@ -27,16 +27,14 @@ void SGDOptimizer::Update(const Tensor *gradient) {
   }
 }
 
-const char *SGDOptimizer::SerializeState(int *state_len) {
+std::string SGDOptimizer::SerializeState() {
   SGDOptimizerState state;
   state.set_num_sample_passed(num_sample_passed_);
-  std::string lr_str = this->lr_policy_->SerializeState(state_len);
+  std::string lr_str = this->lr_policy_->SerializeState();
   state.mutable_lr_state()->ParseFromString(lr_str);
   TensorToProto(*parameter_, state.mutable_parameter());
   if (momentum_ != 0.0) TensorToProto(*momentums_, state.mutable_momentums());
-  auto str = state.SerializeAsString();
-  *state_len += str.size();
-  return str.c_str();
+  return state.SerializeAsString();
 }
 
 void SGDOptimizer::DeserializeState(const std::string &str) {
diff --git a/paddle/optimizer/sgd_optimizer.h b/paddle/optimizer/sgd_optimizer.h
index b74a902e1aa40a7831b36ab826d72372a3588bcf..6e1a0f0d3f9ecfeb51ccb355d65985a2e6388fb0 100644
--- a/paddle/optimizer/sgd_optimizer.h
+++ b/paddle/optimizer/sgd_optimizer.h
@@ -23,7 +23,7 @@ public:
     if (momentums_) delete momentums_;
   }
   void Update(const Tensor* gradient);
-  const char* SerializeState(int* state_len);
+  std::string SerializeState();
   void DeserializeState(const std::string& state);
 
 private: