diff --git a/paddle/optimizer/Tensor.h b/paddle/optimizer/Tensor.h
index d5ba4b3159f7d8dcdbb8ea9b32d70288caef3982..d779bb50709153dee19dc2709608674875000006 100644
--- a/paddle/optimizer/Tensor.h
+++ b/paddle/optimizer/Tensor.h
@@ -5,7 +5,6 @@
  */
 
 #include <string.h>
-#include "optimizer.h"
 #include "paddle/math/BaseMatrix.h"
 
 namespace paddle {
@@ -15,18 +14,27 @@ template <class T>
 using TensorBase = BaseMatrixT<T>;
 
 template <class T>
-class Tensor : public TensorBase<T> {
+class TensorT : public TensorBase<T> {
 public:
-  Tensor(T* data, int size) : TensorBase<T>(1, size, 0, data, false, false) {}
+  TensorT(T* data, int size) : TensorBase<T>(1, size, 0, data, false, false) {}
+  TensorT(const TensorT& t)
+      : TensorBase<T>(1, t.size(), 0, t.get_buffer(), false, false) {}
+  TensorT& operator=(const TensorT& t) {
+    this->size_ = t.size();
+    this->data_ = t.get_buffer();
+  }
   T* get_buffer() { return this->data_; }
   T& operator[](const int idx) {
-    CHECK(idx >= 0 && idx < this->width_) << " out of index range";
+    CHECK(idx >= 0 && idx < this->width_) << "out of index range";
     return this->data_[idx];
   }
   // TODO: replace with tensorshape
   size_t size() const { return this->width_; }
 };
 
+// TODO(zhihong): design problem of dynamic datatype, need to fix
+typedef TensorT<real> Tensor;
+
 }  // namespace optimizer
 }  // namespace paddle
 
diff --git a/paddle/optimizer/adadelta_optimizer.cc b/paddle/optimizer/adadelta_optimizer.cc
index b76c123ec8038c2f5e6131b5a54d501ede92819c..c5537bde8534c9ce84e4c89bd6637bac9c265047 100644
--- a/paddle/optimizer/adadelta_optimizer.cc
+++ b/paddle/optimizer/adadelta_optimizer.cc
@@ -4,19 +4,17 @@
 namespace paddle {
 namespace optimizer {
 
-template <class T>
-void AdadeltaOptimizer<T>::set_weight(const Tensor<T>* p) {
+void AdadeltaOptimizer::set_weight(Tensor* p) {
   size_t size = p->size();
-  T* gptr = new T[size];
-  accum_gradient = Tensor<T>(gptr, size);
-  T* dptr = new T[size];
-  accum_delta = Tensor<T>(dptr, size);
-  T* dptr_current = new T[size];
-  update_delta = Tensor<T>(dptr_current, size);
+  real* gptr = new real[size];
+  accum_gradient = Tensor(gptr, size);
+  real* dptr = new real[size];
+  accum_delta = Tensor(dptr, size);
+  real* dptr_current = new real[size];
+  update_delta = Tensor(dptr_current, size);
 }
 
-template <class T>
-void AdadeltaOptimizer<T>::update(const Tensor<T>& gradient) {
+void AdadeltaOptimizer::update(const Tensor& gradient) {
   num_sample_passed += 1;
   double learning_rate = lr_policy->get_learning_rate(num_sample_passed);
   for (size_t i = 0; i < parameter_->size(); ++i) {
@@ -33,9 +31,5 @@ void AdadeltaOptimizer<T>::update(const Tensor<T>& gradient) {
         learning_rate * update_delta[i] + learning_rate * decay * parameter_[i];
   }
 }
-
-template class AdadeltaOptimizer<float>;
-template class AdadeltaOptimizer<double>;
-
 }  // namespace optimizer
 }  // namespace paddle
diff --git a/paddle/optimizer/adadelta_optimizer.h b/paddle/optimizer/adadelta_optimizer.h
index 35f3ff86fcfdf9463be15607a4266d4713f3e31e..d9db5d09c2269ef4d830a4fa4cf65c99c4095251 100644
--- a/paddle/optimizer/adadelta_optimizer.h
+++ b/paddle/optimizer/adadelta_optimizer.h
@@ -6,28 +6,27 @@
 namespace paddle {
 namespace optimizer {
 
-template <class T>
-class AdadeltaOptimizer : public ParameterOptimizer<T> {
+class AdadeltaOptimizer : public ParameterOptimizer {
 public:
-  using ParameterOptimizer<T>::parameter_;
-  using ParameterOptimizer<T>::num_sample_passed;
-  using ParameterOptimizer<T>::lr_policy;
+  using ParameterOptimizer::parameter_;
+  using ParameterOptimizer::num_sample_passed;
+  using ParameterOptimizer::lr_policy;
 
   AdadeltaOptimizer(double rho, double epsilon, double decay, BaseLr *lr)
-      : ParameterOptimizer<T>(lr), rho(rho), epsilon(epsilon), decay(decay) {}
+      : ParameterOptimizer(lr), rho(rho), epsilon(epsilon), decay(decay) {}
   ~AdadeltaOptimizer() {
     if (accum_gradient) delete accum_gradient;
     if (accum_delta) delete accum_delta;
     if (update_delta) delete update_delta;
   }
-  void update(const Tensor<T> &gradient);
-  void set_weight(const Tensor<T> *p);
-  T *get_weight() const;
+  void update(const Tensor &gradient);
+  void set_weight(Tensor *p);
+  real *get_weight() const;
 
 private:
-  Tensor<T> *accum_gradient;
-  Tensor<T> *accum_delta;
-  Tensor<T> *update_delta;
+  Tensor *accum_gradient;
+  Tensor *accum_delta;
+  Tensor *update_delta;
 
   double rho;
   double epsilon;
diff --git a/paddle/optimizer/adagrad_optimizer.cc b/paddle/optimizer/adagrad_optimizer.cc
index 7b451cb4074bcf7257ff1a0654dccf2e106e1716..3d47e35896ce84e0e0033a4af00c3fdba089ed24 100644
--- a/paddle/optimizer/adagrad_optimizer.cc
+++ b/paddle/optimizer/adagrad_optimizer.cc
@@ -2,21 +2,18 @@
 
 namespace paddle {
 namespace optimizer {
-template <class T>
 
-template <class T>
-void AdagradOptimizer<T>::set_weight(const Tensor<T>* p) {
+void AdagradOptimizer::set_weight(Tensor* p) {
   size_t size = p->width();
-  T* gptr = new T[size];
-  accum_gradient = Tensor<T>(gptr, size);
-  T* dptr = new T[size];
-  accum_delta = Tensor<T>(dtpr, size);
-  T* dptr_current = new T[size];
-  update_delta = Tensor<T>(dptr_current, size);
+  real* gptr = new real[size];
+  accum_gradient = Tensor(gptr, size);
+  real* dptr = new real[size];
+  accum_delta = Tensor(dtpr, size);
+  real* dptr_current = new real[size];
+  update_delta = Tensor(dptr_current, size);
 }
 
-template <class T>
-void AdagradOptimizer<T>::update(const Tensor<T>& gradient) {
+void AdagradOptimizer::update(const Tensor& gradient) {
   num_sample_passed += 1;
   double learning_rate = lr_policy->get_learning_rate();
   for (size_t i = 0; i < parameter_.size(); ++i) {
@@ -27,7 +24,5 @@ void AdagradOptimizer<T>::update(const Tensor<T>& gradient) {
   }
 }
 
-template class AdagradOptimizer<float>;
-template class AdagradOptimizer<double>;
 }  // namespace optimizer
 }  // namespace paddle
diff --git a/paddle/optimizer/adagrad_optimizer.h b/paddle/optimizer/adagrad_optimizer.h
index a01040f30f99ff1ef2569ae264b753fa8d7bd5c9..0f6ce06f35a076147691785327a6c0ff4a26362b 100644
--- a/paddle/optimizer/adagrad_optimizer.h
+++ b/paddle/optimizer/adagrad_optimizer.h
@@ -6,23 +6,19 @@
 namespace paddle {
 namespace optimizer {
 
-template <class T>
-class AdagradOptimizer : public ParameterOptimizer<T> {
+class AdagradOptimizer : public ParameterOptimizer {
 public:
-  using ParameterOptimizer<T>::parameter_;
-  using ParameterOptimizer<T>::num_sample_passed;
-  using ParameterOptimizer<T>::lr_policy;
   AdagradOptimizer(double epsilon, double decay, BaseLr *lr)
-      : ParameterOptimizer<T>(lr), epsilon(epsilon), decay(decay) {}
+      : ParameterOptimizer(lr), epsilon(epsilon), decay(decay) {}
   ~AdagradOptimizer() {
     if (accum_gradient) delete accum_gradient;
   }
-  void update(const Tensor<T> &gradient);
-  void set_weight(const Tensor<T> *p);
-  T *get_weight() const;
+  void update(const Tensor &gradient);
+  void set_weight(Tensor *p);
+  real *get_weight() const;
 
 private:
-  Tensor<T> *accum_gradient;
+  Tensor *accum_gradient;
   double epsilon;
   double decay;
 };
diff --git a/paddle/optimizer/adam_optimizer.cc b/paddle/optimizer/adam_optimizer.cc
index b2d2ddc596ada8275288fc63fb76d344d523a44c..d9cc3344d5932553c0b3d62be7fca297264646b6 100644
--- a/paddle/optimizer/adam_optimizer.cc
+++ b/paddle/optimizer/adam_optimizer.cc
@@ -3,17 +3,15 @@
 namespace paddle {
 namespace optimizer {
 
-template <class T>
-void AdamOptimizer<T>::set_weight(const Tensor<T> *p) {
+void AdamOptimizer::set_weight(Tensor *p) {
   size_t size = p->width();
-  T *mptr = new T[size];
-  momentums_ = Tensor<T>(mptr, size);
-  T *vptr = new T[size];
-  velocitys_ = Tensor<T>(vtpr, size);
+  real *mptr = new real[size];
+  momentums_ = Tensor(mptr, size);
+  real *vptr = new real[size];
+  velocitys_ = Tensor(vtpr, size);
 }
 
-template <class T>
-void AdamOptimizer<T>::update(const Tensor<T> &gradient) {
+void AdamOptimizer::update(const Tensor &gradient) {
   num_sample_passed += 1;
   double learning_rate = lr_policy->get_learning_rate(num_sample_passed);
   double coef1 = 1.0 - std::pow(beta_1, num_sample_passed);
@@ -28,8 +26,5 @@ void AdamOptimizer<T>::update(const Tensor<T> &gradient) {
                          decay * parameter_[i]);
   }
 }
-
-template class AdamOptimizer<float>;
-template class AdamOptimizer<double>;
 }  // namespace optimizer
 }  // namespace paddle
diff --git a/paddle/optimizer/adam_optimizer.h b/paddle/optimizer/adam_optimizer.h
index cf81bd70a64db123ca6e4214b061781a8c510963..68e2aa0223e02988917d1bde01252724ad8ec4f4 100644
--- a/paddle/optimizer/adam_optimizer.h
+++ b/paddle/optimizer/adam_optimizer.h
@@ -6,15 +6,11 @@
 namespace paddle {
 namespace optimizer {
 
-template <class T>
-class AdamOptimizer : public ParameterOptimizer<T> {
+class AdamOptimizer : public ParameterOptimizer {
 public:
-  using ParameterOptimizer<T>::parameter_;
-  using ParameterOptimizer<T>::num_sample_passed;
-  using ParameterOptimizer<T>::lr_policy;
   AdamOptimizer(
       double beta_1, double beta_2, double epsilon, double decay, BaseLr *lr)
-      : ParameterOptimizer<T>(lr),
+      : ParameterOptimizer(lr),
         beta_1(beta_1),
         beta_2(beta_2),
         epsilon(epsilon),
@@ -23,13 +19,13 @@ public:
     if (momentums_) delete momentums_;
     if (velocitys_) delete velocitys_;
   }
-  void update(const Tensor<T> &gradient);
-  void set_weight(const Tensor<T> *p);
-  T *get_weight() const;
+  void update(const Tensor &gradient);
+  void set_weight(Tensor *p);
+  real *get_weight() const;
 
 private:
-  Tensor<T> *momentums_;
-  Tensor<T> *velocitys_;
+  Tensor *momentums_;
+  Tensor *velocitys_;
   double beta_1;
   double beta_2;
   double epsilon;
diff --git a/paddle/optimizer/lr_policy.h b/paddle/optimizer/lr_policy.h
index 5aad87f45fb0619e514de00e65a5ceb9dac0ef01..e1017cf32dcfc8b5bcd4591e971cff2a5ccd5b58 100644
--- a/paddle/optimizer/lr_policy.h
+++ b/paddle/optimizer/lr_policy.h
@@ -1,6 +1,7 @@
 #ifndef PADDLE_OPTIMIZER_LR_POLICY_H_
 #define PADDLE_OPTIMIZER_LR_POLICY_H_
 
+#include <algorithm>
 #include "OptimizerConfig.pb.h"
 
 namespace paddle {
@@ -19,11 +20,25 @@ protected:
 // constant learning rate policy
 class ConstLr final : public BaseLr {
 public:
+  ConstLr(double lr) : BaseLr(lr){};
   double get_learning_rate(const uint64_t num_sample_passed) {
     return learning_rate;
   }
 };
 
+class LinearLr final : public BaseLr {
+public:
+  LinearLr(double lr, double lr_decay_a, double lr_decay_b)
+      : BaseLr(lr), lr_decay_a(lr_decay_a), lr_decay_b(lr_decay_b) {}
+  double get_learning_rate(const uint64_t num_sample_passed) {
+    return std::max(learning_rate - lr_decay_a * num_sample_passed, lr_decay_b);
+  }
+
+private:
+  double lr_decay_a;
+  double lr_decay_b;
+};
+
 }  // namespace optimizer
 }  // namespace paddle
 
diff --git a/paddle/optimizer/optimizer.cc b/paddle/optimizer/optimizer.cc
index ff6558147eeb866351f724e559e0e1841ac15718..fb2e543bf32570c9744296c50256663aee7da4ff 100644
--- a/paddle/optimizer/optimizer.cc
+++ b/paddle/optimizer/optimizer.cc
@@ -2,8 +2,9 @@
 #include <string>
 
 #include "parameter_optimizer.h"
+using namespace paddle::optimizer;
 
-template <paddle_element_type T>
+template <paddle_element_type VALUE>
 struct EnumToType {};
 
 template <class T>
@@ -26,17 +27,16 @@ MATCH_ENUM_TYPE(int64_t, PADDLE_ELEMENT_TYPE_INT64);
 MATCH_ENUM_TYPE(uint64_t, PADDLE_ELEMENT_TYPE_UINT64);
 MATCH_ENUM_TYPE(float, PADDLE_ELEMENT_TYPE_FLOAT32);
 MATCH_ENUM_TYPE(double, PADDLE_ELEMENT_TYPE_FLOAT64);
- struct paddle_optimizer {
-  /*! \brief optmizer in C++ side */
 
-  paddle::optimizer::ParameterOptimizerBase* impl;
+struct paddle_optimizer {
+  paddle::optimizer::ParameterOptimizer* impl;
 };
 
 paddle_optimizer* paddle_create_optimizer(const unsigned char* config_proto,
                                           int config_proto_len) {
-  paddle_optimizer* optimizer;
+  paddle_optimizer* optimizer = new paddle_optimizer;
   std::string config(config_proto, config_proto + config_proto_len);
-  optimizer->impl->create(config_proto);
+  optimizer->impl = ParameterOptimizer::create(config);
   return optimizer;
 }
 
@@ -49,9 +49,9 @@ int paddle_update_parameter(paddle_optimizer* o,
                             const paddle_element_type data_type,
                             const void* grad_buffer,
                             int num_bytes) {
-  auto type = EnumToType<data_type>::Type;
-  paddle::Tensor<type> gradient(reinterpret_cast<type*>(grad_buffer),
-                                num_bytes);
+  // TOOD(zhihong): datatype not work. need to add the runtime datatype
+  auto grad = reinterpret_cast<const real*>(grad_buffer);
+  Tensor gradient(const_cast<real*>(grad), num_bytes);
   o->impl->update(gradient);
   return PADDLE_SUCCESS;
 }
@@ -60,9 +60,8 @@ int paddle_optimizer_set_weights(paddle_optimizer* o,
                                  const paddle_element_type data_type,
                                  void* param_buffer,
                                  int num_bytes) {
-  auto type = EnumToType<data_type>::Type;
-  paddle::Tensor<type>* param = new paddle::Tensor<type>(
-      reinterpret_cast<type*>(param_buffer), num_bytes);
+  // TOOD(zhihong): datatype not work. need to add the runtime datatype
+  Tensor* param = new Tensor(reinterpret_cast<real*>(param_buffer), num_bytes);
   o->impl->set_weight(param);
   return PADDLE_SUCCESS;
 }
diff --git a/paddle/optimizer/parameter_optimizer.cc b/paddle/optimizer/parameter_optimizer.cc
index 4bd384d77da1deb7f86c72cca6f66947eddb726f..6d9fa5c8024013250e6e0b76707c9f0b467f5646 100644
--- a/paddle/optimizer/parameter_optimizer.cc
+++ b/paddle/optimizer/parameter_optimizer.cc
@@ -10,78 +10,60 @@
 namespace paddle {
 namespace optimizer {
 
-template <class T>
-ParameterOptimizer<T> *ParameterOptimizer<T>::create(
+ParameterOptimizer *ParameterOptimizer::create(
     const ::std::string &config_proto) {
   paddle::OptimizerConfig config;
   CHECK(config.ParseFromString(config_proto) == 0)
       << "error : optimizer config";
-  CHECK(config_valid(config) == 0) << "error : invalid optimizer config ";
 
-  BaseLr *lr = nullptr;
-  switch (config.lr_policy()) {
-    case "ConstLr":
-      lr = new ConstLr(config.lr_config().learning_rate());
-      break;
-  }
-  ParameterOptimizer<T> *opt = nullptr;
-  switch (config.optimizer_name()) {
-    case "SGD":
-      opt = new SGDOptimizer<T>(config.sgd().momentum(),
-                                config.sgd().decay(),
-                                config.sgd().nesterov(),
-                                lr);
-      break;
-    case "Adagrad":
-      opt = new AdagradOptimizer<T>(
+  auto select_lr_policy = [=](const OptimizerConfig &config) -> BaseLr * {
+    std::string s(config.lr_policy());
+    if (s == "ConstLr") return new ConstLr(config.lr_config().learning_rate());
+    if (s == "LinearLr")
+      return new LinearLr(config.lr_config().learning_rate(),
+                          config.lr_config().lr_decay_a(),
+                          config.lr_config().lr_decay_b());
+    // default
+    return new ConstLr(config.lr_config().learning_rate());
+  };
+  BaseLr *lr = select_lr_policy(config);
+  auto select_optimizer =
+      [=](const OptimizerConfig &config) -> ParameterOptimizer * {
+    std::string s(config.optimizer_name());
+    if (s == "SGD") {
+      return new SGDOptimizer(config.sgd().momentum(),
+                              config.sgd().decay(),
+                              config.sgd().nesterov(),
+                              lr);
+    }
+    if (s == "Adadelta") {
+      return new AdagradOptimizer(
           config.adagrad().epsilon(), config.adagrad().decay(), lr);
-      break;
-    case "Adadelta":
-      opt = new AdadeltaOptimizer<T>(config.adadelta().rho(),
-                                     config.adadelta().epsilon(),
-                                     config.adadelta().decay(),
-                                     lr);
-      break;
-    case "Adam":
-      opt = new AdamOptimizer<T>(config.adam().beta_1(),
-                                 config.adam().beta_2(),
-                                 config.adam().epsilon(),
-                                 config.adam().decay(),
-                                 lr);
-      break;
-  }
-
-  return opt;
-}
-
-template <class T>
-T *ParameterOptimizer<T>::get_weight() const {
-  return parameter.get().get_buffer();
-}
-
-template <class T>
-char *ParameterOptimizer<T>::get_config_proto() const {
-  // set config dynamic value for save checkpoint
-  config_.lr_policy().set_learning_rate(
-      lr_policy->get_learning_rate(num_sample_passed));
-  config_.set_num_sample_passed(num_sample_passed);
-  config_.set_iterations(iterations);
-  return config_.SerializeAsString().c_str();
-}
-
-template <class T>
-void ParameterOptimizer<T>::set_weight(const Tensor<T> *p) {
-  parameter_ = p;
+    }
+    if (s == "Adagrad") {
+      return new AdagradOptimizer(
+          config.adagrad().epsilon(), config.adagrad().decay(), lr);
+    }
+    if (s == "Adam") {
+      return new AdadeltaOptimizer(config.adadelta().rho(),
+                                   config.adadelta().epsilon(),
+                                   config.adadelta().decay(),
+                                   lr);
+    }
+    // default
+    return new SGDOptimizer(config.sgd().momentum(),
+                            config.sgd().decay(),
+                            config.sgd().nesterov(),
+                            lr);
+  };
+  return select_optimizer(config);
 }
 
-template <class T>
-bool ParameterOptimizer<T>::config_valid(const ::std::string &config) const {
-  // TODO(zhihong) : add more value checker, failed ASAP
-  return true;
+real *ParameterOptimizer::get_weight() const {
+  return parameter_->get_buffer();
 }
 
-template class ParameterOptimzier<float>;
-template class ParameterOptimzier<double>;
+void ParameterOptimizer::set_weight(Tensor *p) { parameter_ = p; }
 
 }  // namespace optimizer
 }  // namespace paddle
diff --git a/paddle/optimizer/parameter_optimizer.h b/paddle/optimizer/parameter_optimizer.h
index 40994aa86703b0437f0795f1c0323c26596c564a..a4f39836bafc2e7c8c0c1666d58c432573dfe638 100644
--- a/paddle/optimizer/parameter_optimizer.h
+++ b/paddle/optimizer/parameter_optimizer.h
@@ -11,13 +11,6 @@
 namespace paddle {
 namespace optimizer {
 
-class ParameterOptimizerBase {
-private:
-  ParameterOptimizerBase(const ParameterOptimizerBase &) = delete;
-  ParameterOptimizerBase &operator=(const ParameterOptimizerBase &) = delete;
-};
-
-template <class T>
 class ParameterOptimizer {
 public:
   /**
@@ -31,14 +24,13 @@ public:
   virtual ~ParameterOptimizer() { delete parameter_; };
 
   static ParameterOptimizer *create(const ::std::string &config_proto);
-  virtual void update(const Tensor<T> &gradient) = 0;
-  virtual T *get_weight() const;
-  virtual void set_weight(const Tensor<T> *parameter);
+  virtual void update(const Tensor &gradient) = 0;
+  virtual real *get_weight() const;
+  virtual void set_weight(Tensor *parameter);
 
 public:
-  bool config_valid(::std::string &config) const;
   OptimizerConfig config_;
-  Tensor<T> *parameter_;
+  Tensor *parameter_;
 
   // learning rate policy
   BaseLr *lr_policy;
diff --git a/paddle/optimizer/sgd_optimizer.h b/paddle/optimizer/sgd_optimizer.h
index 5115825590d82baded7e2384ed54f5bd75402058..375c99b30b87bb84fb1459d435174a8be4448f07 100644
--- a/paddle/optimizer/sgd_optimizer.h
+++ b/paddle/optimizer/sgd_optimizer.h
@@ -6,31 +6,22 @@
 namespace paddle {
 namespace optimizer {
 
-template <class T>
-class SGDOptimizer : public ParameterOptimizer<T> {
+class SGDOptimizer : public ParameterOptimizer {
 public:
-  using ParameterOptimizer<T>::parameter_;
-  using ParameterOptimizer<T>::num_sample_passed;
-  using ParameterOptimizer<T>::lr_policy;
-
-  SGDOptimizer(double m,
-               double d,
-               bool n,
-               double learning_rate,
-               uint64_t num_sample_passed,
-               BaseLr* lr)
-      : ParameterOptimizer<T>(lr), momentum(m), decay(d), nesterov(n) {}
-  virtual ~SGDOptimizer() {
-    // clear memory by Tensor library
-    delete momentums_;
-  }
-  void update(const Tensor<T>& gradient);
-
-  void set_weight(const Tensor<T>* p);
-  T* get_weight() const;
+  using ParameterOptimizer::parameter_;
+  using ParameterOptimizer::num_sample_passed;
+  using ParameterOptimizer::lr_policy;
+
+  SGDOptimizer(double m, double d, bool n, BaseLr* lr)
+      : ParameterOptimizer(lr), momentum(m), decay(d), nesterov(n) {}
+  virtual ~SGDOptimizer() { delete momentums_; }
+  void update(const Tensor& gradient);
+
+  void set_weight(Tensor* p);
+  real* get_weight() const;
 
 private:
-  Tensor<T>* momentums_;
+  Tensor* momentums_;
   double momentum;
   double decay;
   bool nesterov;
diff --git a/paddle/optimizer/sgd_optmizer.cc b/paddle/optimizer/sgd_optmizer.cc
index cd1635fecdca0cf9803849ff3a6a4b6eb265c32f..03ddc81451715b458b33afdeb10d9bfd24f1b354 100644
--- a/paddle/optimizer/sgd_optmizer.cc
+++ b/paddle/optimizer/sgd_optmizer.cc
@@ -3,23 +3,21 @@
 namespace paddle {
 namespace optimizer {
 
-template <class T>
-void SGDOptimizer<T>::set_weight(const Tensor<T> *p) {
+void SGDOptimizer::set_weight(Tensor *p) {
   //  ParameterOptimizer::set_weight(p);
   size_t size = p->size();
   // TODO: fix it with align aware allocator bind to Tensor
   if (momentum != 0.0) {
-    T *ptr = new T[size];
-    momentums_ = Tensor<T>(ptr, size);
+    real *ptr = new real[size];
+    momentums_ = new Tensor(ptr, size);
   }
 }
 
-template <class T>
-void SGDOptimizer<T>::update(const Tensor<T> &gradient) {
+void SGDOptimizer::update(const Tensor &gradient) {
   num_sample_passed += 1;
   double learning_rate = lr_policy->get_learning_rate(num_sample_passed);
   double velocity = 0.0;
-  Tensor<T> &for (size_t i = 0; i < parameter_->size(); ++i) {
+  for (size_t i = 0; i < parameter_->size(); ++i) {
     if (momentum == 0.0) {
       velocity =
           -learning_rate * gradient[i] - learning_rate * decay * parameter_[i];
@@ -36,8 +34,5 @@ void SGDOptimizer<T>::update(const Tensor<T> &gradient) {
   }
 }
 
-template class SGDOptimizer<float>;
-template class SGDOptimizer<double>;
-
 }  // namespace optimizer
 }  // namespace paddle