diff --git a/paddle/fluid/imperative/CMakeLists.txt b/paddle/fluid/imperative/CMakeLists.txt
index fb57eca658aebbf34dde6eb2f93f45b10aee91c7..373d292b443b7651b785a52a6986b0a0be58ad61 100644
--- a/paddle/fluid/imperative/CMakeLists.txt
+++ b/paddle/fluid/imperative/CMakeLists.txt
@@ -1,3 +1,3 @@
-cc_library(layer SRCS layer.cc DEPS proto_desc)
+cc_library(layer SRCS layer.cc DEPS proto_desc operator)
 cc_library(tracer SRCS tracer.cc DEPS proto_desc)
 cc_library(engine SRCS engine.cc)
diff --git a/paddle/fluid/imperative/layer.cc b/paddle/fluid/imperative/layer.cc
index e0b6ad91e7b98be68b77fb57cb28125b5e06cb23..08379a7ed5473a12265e276c8c5962fdabd93607 100644
--- a/paddle/fluid/imperative/layer.cc
+++ b/paddle/fluid/imperative/layer.cc
@@ -13,7 +13,257 @@
 // limitations under the License.
 
 #include "paddle/fluid/imperative/layer.h"
+#include <deque>
+#include <limits>
+#include <map>
+#include <random>
+#include <utility>
+
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/string/printf.h"
 
 namespace paddle {
-namespace imperative {}  // namespace imperative
+namespace imperative {
+
+using framework::Variable;
+
+void AddTo(Variable* src, Variable* dst) {
+  framework::LoDTensor* dst_tensor = dst->GetMutable<framework::LoDTensor>();
+  framework::LoDTensor* src_tensor = src->GetMutable<framework::LoDTensor>();
+  PADDLE_ENFORCE(dst_tensor->numel() == src_tensor->numel(), "%lld vs %lld",
+                 dst_tensor->numel(), src_tensor->numel());
+  float* dst_data = dst_tensor->mutable_data<float>(platform::CPUPlace());
+  const float* src_data = src_tensor->data<float>();
+  for (size_t i = 0; i < src_tensor->numel(); ++i) {
+    dst_data[i] += src_data[i];
+  }
+}
+
+class Autograd {
+ public:
+  explicit Autograd(framework::Scope* scope) : scope_(scope) {}
+
+  void RunBackward(VarBase* var, framework::Variable* grad) {
+    if (!var->pre_op_) {
+      var->ApplyGrad(scope_, grad);
+      return;
+    }
+    PADDLE_ENFORCE(var->pre_op_->op_desc_);
+    // TODO(panyx0718): Only create vars that "require_grad"
+    std::vector<Variable*> op_grads =
+        CreateOpGrads(var->pre_op_->output_vars_->size());
+    op_grads[var->pre_op_out_idx_] = grad;
+
+    std::deque<std::pair<OpBase*, std::vector<Variable*>>> ready;
+    ready.push_back(std::make_pair(var->pre_op_, op_grads));
+
+    std::map<OpBase*, int> dep_counts = ComputeDepCounts(var->pre_op_);
+    std::map<OpBase*, std::vector<Variable*>> visited;
+
+    while (!ready.empty()) {
+      OpBase* ready_op = ready.front().first;
+      std::vector<Variable*> ready_op_grads = ready.front().second;
+      ready.pop_front();
+      std::vector<Variable*> input_grads = ready_op->ApplyGrad(scope_);
+
+      for (size_t i = 0; i < input_grads.size(); ++i) {
+        if (!input_grads[i]) continue;
+        OpBase* pre_op = ready_op->pre_ops_->at(i);
+        if (!pre_op) continue;
+        int pre_op_out_idx = ready_op->pre_ops_out_idx_->at(i);
+
+        dep_counts[pre_op] -= 1;
+        PADDLE_ENFORCE(dep_counts[pre_op] >= 0);
+        bool pre_op_ready = dep_counts[pre_op] == 0;
+
+        if (pre_op_ready) {
+          if (visited.find(pre_op) == visited.end()) {
+            PADDLE_ENFORCE(pre_op->output_vars_->size() == 1);
+            visited[pre_op] = {input_grads[i]};
+          } else {
+            std::vector<Variable*>& pre_op_grads = visited[pre_op];
+            AccumGrads(pre_op_out_idx, input_grads[i], &pre_op_grads);
+          }
+          ready.push_back(std::make_pair(pre_op, visited[pre_op]));
+        } else {
+          if (visited.find(pre_op) == visited.end()) {
+            // TODO(panyx0718): Only create vars that "require_grad"
+            visited[pre_op] = CreateOpGrads(var->pre_op_->output_vars_->size());
+          } else {
+          }
+          std::vector<Variable*>& pre_op_grads = visited[pre_op];
+          AccumGrads(pre_op_out_idx, input_grads[i], &pre_op_grads);
+        }
+      }
+    }
+  }
+
+ private:
+  void AccumGrads(int grad_idx, Variable* grad,
+                  std::vector<Variable*>* op_grads) {
+    if (!(*op_grads)[grad_idx]) {
+      // FIXME(panyx0718): This should be a deep copy.
+      (*op_grads)[grad_idx] = grad;
+      return;
+    }
+    AddTo(grad, (*op_grads)[grad_idx]);
+  }
+
+  std::map<OpBase*, int> ComputeDepCounts(OpBase* op) {
+    std::map<OpBase*, int> ret;
+
+    std::deque<OpBase*> queue;
+    queue.push_back(op);
+    std::unordered_set<OpBase*> visited;
+    visited.insert(op);
+    while (!queue.empty()) {
+      OpBase* candidate = queue.front();
+      queue.pop_front();
+      for (OpBase* pre_op : *(candidate->pre_ops_)) {
+        if (!pre_op) continue;
+        if (visited.find(pre_op) == visited.end()) {
+          visited.insert(pre_op);
+          queue.push_back(pre_op);
+        }
+        ret[pre_op] += 1;
+      }
+    }
+
+    return ret;
+  }
+
+  std::vector<Variable*> CreateOpGrads(size_t count) {
+    std::vector<Variable*> op_grads;
+    for (size_t i = 0; i < count; ++i) {
+      op_grads.push_back(nullptr);
+    }
+    return op_grads;
+  }
+
+  framework::Scope* scope_;
+};
+
+framework::Variable* CreateVariable(const std::string& name,
+                                    const framework::DDim& dim, float val,
+                                    framework::Scope* scope,
+                                    bool random_name = true) {
+  std::string varname = name;
+  if (random_name) {
+    std::mt19937 rng;
+    rng.seed(std::random_device()());
+    std::uniform_int_distribution<std::mt19937::result_type> dist6(
+        1, std::numeric_limits<int>::max());
+    int id = dist6(rng);
+    varname = string::Sprintf("%s@%d", varname, id);
+  }
+
+  LOG(ERROR) << "creating var " << varname;
+  framework::Variable* var = scope->Var(varname);
+  framework::LoDTensor* tensor = var->GetMutable<framework::LoDTensor>();
+
+  float* data = tensor->mutable_data<float>(dim, platform::CPUPlace());
+  std::fill(data, data + tensor->numel(), val);
+  return var;
+}
+
+framework::LoDTensor& VarBase::Grad() {
+  VLOG(3) << "get var grad " << var_desc_->Name();
+  return *grads_->GetMutable<framework::LoDTensor>();
+}
+
+void VarBase::ApplyGrad(framework::Scope* scope, Variable* grad) {
+  VLOG(3) << "apply var grad " << var_desc_->Name() << " "
+          << grad->Get<framework::LoDTensor>().data<float>()[0];
+  if (!grads_) {
+    grads_ =
+        CreateVariable(string::Sprintf("%s@IGrad", var_desc_->Name()),
+                       var_->Get<framework::LoDTensor>().dims(), 0.0, scope);
+  }
+  AddTo(grad, grads_);
+  VLOG(3) << "grad_ after apply var grad " << var_desc_->Name() << " "
+          << grads_->Get<framework::LoDTensor>().data<float>()[0];
+}
+
+std::vector<Variable*> OpBase::ApplyGrad(framework::Scope* scope) {
+  VLOG(3) << "op grad " << grad_op_desc_->Type();
+
+  for (const std::string& invar : grad_op_desc_->InputArgumentNames()) {
+    block_->FindRecursiveOrCreateVar(invar);
+    framework::Variable* var = scope->Var(invar);
+    LOG(ERROR) << "op grad in var " << invar;
+    if (!var->IsInitialized()) {
+      framework::VarDesc* var_desc = block_->FindVar(invar);
+      if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
+        LOG(ERROR) << "grad op invar init " << invar;
+        var->GetMutable<framework::LoDTensor>();
+      } else {
+        LOG(ERROR) << "tracer doesn't support yet";
+      }
+    } else {
+      var->GetMutable<framework::LoDTensor>()->type();
+    }
+  }
+
+  std::vector<Variable*> ret;
+  for (size_t i = 0; i < input_vars_->size(); ++i) {
+    ret.push_back(nullptr);
+  }
+  for (const std::string& outvar : grad_op_desc_->OutputArgumentNames()) {
+    LOG(ERROR) << "grad outvar " << outvar;
+    block_->FindRecursiveOrCreateVar(outvar);
+    framework::Variable* var = scope->Var(outvar);
+    if (!var->IsInitialized()) {
+      framework::VarDesc* var_desc = block_->FindVar(outvar);
+      if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
+        var->GetMutable<framework::LoDTensor>();
+      } else {
+        LOG(ERROR) << "tracer doesn't support yet";
+      }
+    }
+  }
+  grad_op_desc_->InferShape(*block_);
+  grad_op_desc_->InferVarType(block_);
+  std::unique_ptr<framework::OperatorBase> opbase =
+      framework::OpRegistry::CreateOp(*grad_op_desc_);
+
+  opbase->Run(*scope, platform::CPUPlace());
+
+  for (const std::string& outvar : grad_op_desc_->OutputArgumentNames()) {
+    if (grad_to_var_->find(outvar) != grad_to_var_->end()) {
+      std::string origin_var = (*grad_to_var_)[outvar];
+      for (size_t i = 0; i < input_vars_->size(); ++i) {
+        VarBase* origin_in_var = (*input_vars_)[i];
+        if (origin_in_var->var_desc_->Name() == origin_var) {
+          framework::Variable* var = scope->FindVar(outvar);
+          LOG(ERROR) << "apply grad " << outvar << " with origin "
+                     << origin_var;
+          // TODO(panyx0718): Accumulate.
+          // origin_in_var->grads_ = var;
+          origin_in_var->ApplyGrad(scope, var);
+          ret[i] = var;
+          // TODO(panyx0718): There might be 2 var with the same name. We
+          // currently assume the are the same Variable*. So it doesn't matter
+          // which one is used.
+          break;
+        }
+      }
+    }
+  }
+  return ret;
+}
+
+void VarBase::RunBackward(framework::Scope* scope) {
+  // TODO(panyx0718): Might not be 0th, need to detect.
+  grads_ = CreateVariable(pre_op_->grad_op_desc_->InputArgumentNames()[0],
+                          var_->Get<framework::LoDTensor>().dims(), 1.0, scope,
+                          false);
+  framework::Variable* grad =
+      CreateVariable("init@imperative_grad",
+                     var_->Get<framework::LoDTensor>().dims(), 1.0, scope);
+
+  Autograd(scope).RunBackward(this, grad);
+}
+
+}  // namespace imperative
 }  // namespace paddle
diff --git a/paddle/fluid/imperative/layer.h b/paddle/fluid/imperative/layer.h
index c400b19700094e40a3d234c2c063e27a0596e1bd..0cf0c27a6a9b9f9fc78dfc331977f9ac20e87cb9 100644
--- a/paddle/fluid/imperative/layer.h
+++ b/paddle/fluid/imperative/layer.h
@@ -14,8 +14,10 @@
 
 #pragma once
 
+#include <string>
 #include <vector>
 #include "paddle/fluid/framework/op_desc.h"
+#include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/var_desc.h"
 #include "paddle/fluid/platform/enforce.h"
@@ -27,26 +29,64 @@ class OpBase;
 
 class VarBase {
  public:
-  VarBase() {}
-  virtual ~VarBase() {}
+  VarBase()
+      : pre_op_(nullptr),
+        pre_op_out_idx_(-1),
+        var_desc_(nullptr),
+        var_(nullptr),
+        grads_(nullptr) {}
+
+  virtual ~VarBase() {
+    LOG(ERROR) << "deleting var";
+    LOG(ERROR) << "done deleting var";
+  }
+
+  void ApplyGrad(framework::Scope* scope, framework::Variable* grad);
+
+  void RunBackward(framework::Scope* scope);
+
+  framework::LoDTensor& Grad();
 
   OpBase* pre_op_;
+  int pre_op_out_idx_;
+
   framework::VarDesc* var_desc_;
+  framework::Variable* var_;
+  framework::Variable* grads_;
 };
 
 class OpBase {
  public:
   OpBase()
       : input_vars_(new std::vector<VarBase*>()),
-        output_vars_(new std::vector<VarBase*>()) {}
+        output_vars_(new std::vector<VarBase*>()),
+        pre_ops_(new std::vector<OpBase*>()),
+        pre_ops_out_idx_(new std::vector<int>()),
+        op_desc_(nullptr),
+        grad_op_desc_(nullptr) {}
+
   virtual ~OpBase() {
     delete input_vars_;
     delete output_vars_;
+
+    delete pre_ops_;
+    delete pre_ops_out_idx_;
+
+    if (grad_op_desc_) delete grad_op_desc_;
+    if (grad_to_var_) delete grad_to_var_;
   }
 
+  std::vector<framework::Variable*> ApplyGrad(framework::Scope* scope);
+
   std::vector<VarBase*>* input_vars_;
   std::vector<VarBase*>* output_vars_;
+  std::vector<OpBase*>* pre_ops_;
+  std::vector<int>* pre_ops_out_idx_;
   framework::OpDesc* op_desc_;
+
+  framework::OpDesc* grad_op_desc_;
+  std::unordered_map<std::string, std::string>* grad_to_var_;
+  framework::BlockDesc* block_;
 };
 
 class Layer {
@@ -58,7 +98,7 @@ class Layer {
     return vars;
   }
 
-  virtual void Backward() { LOG(ERROR) << "backward at cpp."; }
+  virtual void Backward() { LOG(ERROR) << "To support customize"; }
 };
 
 }  // namespace imperative
diff --git a/paddle/fluid/imperative/tracer.h b/paddle/fluid/imperative/tracer.h
index 9d7bdda8ccc83aedb25b4aba5868a57d49506370..c3cc0cb33e57012dab904c2176b7647853607d6a 100644
--- a/paddle/fluid/imperative/tracer.h
+++ b/paddle/fluid/imperative/tracer.h
@@ -27,6 +27,20 @@
 namespace paddle {
 namespace imperative {
 
+void CreateGradOp(const framework::OpDesc& op_desc,
+                  const std::unordered_set<std::string>& no_grad_set,
+                  const std::vector<framework::BlockDesc*>& grad_sub_block,
+                  framework::OpDesc** grad_op_desc,
+                  std::unordered_map<std::string, std::string>* grad_to_var) {
+  std::vector<std::unique_ptr<framework::OpDesc>> grad_op_descs =
+      framework::OpInfoMap::Instance()
+          .Get(op_desc.Type())
+          .GradOpMaker()(op_desc, no_grad_set, grad_to_var, grad_sub_block);
+  PADDLE_ENFORCE(grad_op_descs.size() == 1, "Only support 1 grad op now.");
+  // TODO(panyx0718): Leak?
+  *grad_op_desc = grad_op_descs[0].release();
+}
+
 class Tracer {
  public:
   Tracer() {}
@@ -44,6 +58,7 @@ class Tracer {
     for (VarBase* input : inputs) {
       const std::string vname = input->var_desc_->Name();
       framework::Variable* var = scope_->Var(vname);
+      input->var_ = var;
       if (!var->IsInitialized()) {
         framework::VarDesc* var_desc = block_->FindVar(vname);
         if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
@@ -52,11 +67,17 @@ class Tracer {
           LOG(ERROR) << "tracer doesn't support yet";
         }
       }
+      if (input->pre_op_) {
+        op->pre_ops_->push_back(input->pre_op_);
+        op->pre_ops_out_idx_->push_back(input->pre_op_out_idx_);
+      } else {
+        op->pre_ops_->push_back(nullptr);
+      }
     }
 
     *op->output_vars_ = outputs;
-    for (auto output : outputs) {
-      const std::string vname = output->var_desc_->Name();
+    for (size_t i = 0; i < outputs.size(); ++i) {
+      const std::string vname = outputs[i]->var_desc_->Name();
       framework::Variable* var = scope_->Var(vname);
       if (!var->IsInitialized()) {
         framework::VarDesc* var_desc = block_->FindVar(vname);
@@ -66,9 +87,18 @@ class Tracer {
           LOG(ERROR) << "tracer doesn't support yet";
         }
       }
-      output->pre_op_ = op;
+      outputs[i]->var_ = var;
+      outputs[i]->pre_op_ = op;
+      outputs[i]->pre_op_out_idx_ = i;
     }
     op_base->Run(*scope_, platform::CPUPlace());
+
+    framework::OpDesc* grad_op_desc;
+    auto grad_to_var = new std::unordered_map<std::string, std::string>();
+    CreateGradOp(*op_desc, {}, {block_}, &grad_op_desc, grad_to_var);
+    op->grad_op_desc_ = grad_op_desc;
+    op->grad_to_var_ = grad_to_var;
+    op->block_ = block_;
   }
 
   void SetScope(framework::Scope* scope) { scope_ = scope; }
diff --git a/paddle/fluid/pybind/CMakeLists.txt b/paddle/fluid/pybind/CMakeLists.txt
index 7079f9fe04089f1d60c90b2d1758212e1adefd34..b8954cb12628d1f4f333956e0213ddf9c01e592c 100644
--- a/paddle/fluid/pybind/CMakeLists.txt
+++ b/paddle/fluid/pybind/CMakeLists.txt
@@ -1,5 +1,5 @@
 
-set(PYBIND_DEPS pybind python proto_desc memory executor async_executor prune feed_fetch_method pass_builder parallel_executor profiler)
+set(PYBIND_DEPS pybind python proto_desc memory executor async_executor prune feed_fetch_method pass_builder parallel_executor profiler layer)
 set(PYBIND_SRCS pybind.cc exception.cc protobuf.cc const_value.cc recordio.cc async_executor_py.cc imperative.cc)
 
 if(WITH_PYTHON)
diff --git a/paddle/fluid/pybind/imperative.h b/paddle/fluid/pybind/imperative.h
index 5834b83df904bda8389782c5dd65e6f4a23ee4dd..7a9d3a01ea81f11ac85000c3d0153f20e108789a 100644
--- a/paddle/fluid/pybind/imperative.h
+++ b/paddle/fluid/pybind/imperative.h
@@ -42,6 +42,11 @@ class PyOpBase : public imperative::OpBase {
   using imperative::OpBase::OpBase;  // Inherit constructors
 };
 
+class PyVarBase : public imperative::VarBase {
+ public:
+  using imperative::VarBase::VarBase;  // Inherit constructors
+};
+
 void BindTracer(pybind11::module* m);
 
 }  // namespace pybind
diff --git a/paddle/fluid/pybind/pybind.cc b/paddle/fluid/pybind/pybind.cc
index 656d28eb2a1b30eff3793a9fc42ebe06fff8d419..ea07372a289956a31dcf40c4d66e182c93adbf58 100644
--- a/paddle/fluid/pybind/pybind.cc
+++ b/paddle/fluid/pybind/pybind.cc
@@ -34,6 +34,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/reader.h"
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/framework/version.h"
+#include "paddle/fluid/imperative/layer.h"
 #include "paddle/fluid/memory/allocation/allocator_strategy.h"
 #include "paddle/fluid/operators/activation_op.h"
 #include "paddle/fluid/operators/reader/lod_tensor_blocking_queue.h"
@@ -101,8 +102,13 @@ PYBIND11_MODULE(core, m) {
 
   BindException(&m);
 
-  py::class_<imperative::VarBase>(m, "VarBase",
-                                  R"DOC()DOC")
+  py::class_<imperative::VarBase, PyVarBase>(m, "VarBase", R"DOC()DOC")
+      .def(py::init<>())
+      .def("_run_backward",
+           [](imperative::VarBase &self, framework::Scope *scope) {
+             self.RunBackward(scope);
+           })
+      .def("_grad", &imperative::VarBase::Grad)
       .def_property(
           "desc",
           [](const imperative::VarBase &self) { return self.var_desc_; },
@@ -111,13 +117,14 @@ PYBIND11_MODULE(core, m) {
           },
           py::return_value_policy::reference);
 
-  py::class_<imperative::OpBase, PyOpBase>(m, "OpBase",
-                                           R"DOC()DOC")
+  py::class_<imperative::OpBase, PyOpBase>(m, "OpBase", R"DOC()DOC")
       .def(py::init<>())
       .def_property(
           "desc", [](const imperative::OpBase &self) { return self.op_desc_; },
           [](imperative::OpBase &self, framework::OpDesc *op_desc) {
-            self.op_desc_ = op_desc;
+            if (op_desc) {
+              self.op_desc_ = op_desc;
+            }
           },
           py::return_value_policy::reference);
 
diff --git a/python/paddle/fluid/framework.py b/python/paddle/fluid/framework.py
index d4ca6901d5a51b5947fc7a4fcfaaddfaa94c4a40..ba3ffafc85f0492df2d81f545cf25496d3c0793c 100644
--- a/python/paddle/fluid/framework.py
+++ b/python/paddle/fluid/framework.py
@@ -276,6 +276,7 @@ class Variable(core.VarBase):
                  stop_gradient=False,
                  is_data=False,
                  **kwargs):
+        core.VarBase.__init__(self)
         self.block = block
         self.error_clip = error_clip
 
@@ -361,6 +362,12 @@ class Variable(core.VarBase):
         tensor = core.get_variable_tensor(scope, self.desc.name())
         return np.array(tensor)
 
+    def backward(self, scope):
+        self._run_backward(scope)
+
+    def grad(self):
+        return np.array(self._grad())
+
     def __str__(self):
         return self.to_string(True)
 
@@ -983,6 +990,7 @@ class Block(object):
         self.desc = program.desc.block(idx)
         self.vars = collections.OrderedDict()  # var_name --> var
         self.ops = list()  # operator list
+        self._op_descs = list()
         self.program = program
         self.removed_vars = collections.OrderedDict()
 
@@ -1238,13 +1246,12 @@ class Block(object):
         if _in_imperative_mode():
             op_desc = core.OpDesc()
             op = Operator(block=self, desc=op_desc, *args, **kwargs)
-            sys.stderr.write('%s %s!!!\n' % (type(op.inputs), type(op.outputs)))
             _imperative_tracer().trace(op, op.inputs, op.outputs)
-            return
-
-        op_desc = self.desc.append_op()
-        op = Operator(block=self, desc=op_desc, *args, **kwargs)
+        else:
+            op_desc = self.desc.append_op()
+            op = Operator(block=self, desc=op_desc, *args, **kwargs)
         self.ops.append(op)
+        self._op_descs.append(op_desc)
         return op
 
     def _insert_op(self, index, *args, **kwargs):
diff --git a/python/paddle/fluid/tests/unittests/test_imperative.py b/python/paddle/fluid/tests/unittests/test_imperative.py
index af6a2167ce72142e3fb2740e7f2301fbc27f7ee9..18fe8f7c09ae34c5c7e28f07c3fd99539500cbf5 100644
--- a/python/paddle/fluid/tests/unittests/test_imperative.py
+++ b/python/paddle/fluid/tests/unittests/test_imperative.py
@@ -26,6 +26,7 @@ class MyLayer(fluid.imperative.PyLayer):
 
     def forward(self, inputs):
         x = fluid.layers.relu(inputs[0])
+        self._x_for_debug = x
         return [fluid.layers.elementwise_mul(x, x)]
 
 
@@ -43,6 +44,8 @@ class TestImperative(unittest.TestCase):
             x = l(np.array([1.0, 2.0, -1.0], dtype=np.float32))[0]
             self.assertIsNotNone(x)
             sys.stderr.write("%s output: %s\n" % (x, x.numpy(scope=l._scope)))
+            x.backward(l._scope)
+            sys.stderr.write("grad %s\n" % l._x_for_debug.grad())
 
 
 if __name__ == '__main__':