Impl optimizer

paddle/framework/backward.cc

@@ -281,12 +281,16 @@ static void CreateGradVarInBlock(
   auto ops = block_desc->AllOps();
   for (size_t op_index = grad_op_start_index; op_index < ops.size();
        ++op_index) {
+    bool need_infer_shape = false;
     ForEachVarName(ops[op_index]->Outputs(),
                    [&](const std::string& grad_var_name) {
                      if (block_desc->HasVar(grad_var_name)) {
                        return false;
                      }
-                     block_desc->Var(grad_var_name);
+                     need_infer_shape = true;
+                     auto var = block_desc->Var(grad_var_name);
+                     // FIXME(qiao) infer the datatype
+                     var->SetDataType(framework::DataType::FP32);
                      auto it = param_name_map.find(grad_var_name);
                      if (it == param_name_map.end()) {
                        return false;
@@ -298,6 +302,9 @@ static void CreateGradVarInBlock(
                      grad_record.op_idx_ = static_cast<int>(op_index);
                      return false; /* not break */
                    });
+    if (need_infer_shape) {
+      ops[op_index]->InferShape(*block_desc);
+    }
   }
 }
 
@@ -428,10 +435,16 @@ ParamGradInfoMap AppendBackward(
   auto& all_ops = root_block->ops_;
 
   // insert fill one op for target
+  // TODO(qiao) add some check to the target.
   std::string fill_one_op_out = GradVarName(target.Name());
+  std::vector<int64_t> target_shape_desc = target.Shape();
+  std::vector<int> target_shape;
+  std::transform(target_shape_desc.begin(), target_shape_desc.end(),
+                 std::back_inserter(target_shape),
+                 [](int64_t dim) { return static_cast<int>(dim); });
   std::unique_ptr<OpDescBind> fill_one_op(
       new OpDescBind("fill_constant", {}, {{"Out", {fill_one_op_out}}},
-                     {{"shape", std::vector<int>{1}},
+                     {{"shape", target_shape},
                       {"value", static_cast<float>(1.0)},
                       {"data_type", framework::DataType::FP32}}));
   all_ops.push_back(std::move(fill_one_op));
@@ -443,13 +456,22 @@ ParamGradInfoMap AppendBackward(
   auto backward_op_descs = MakeBlockBackward(program_desc, root_block_idx,
                                              &no_grad_var_names, &grad_to_var);
 
-  std::unordered_map<std::string, GradVarInfo> retv;
-
-  // Create Variable
   for (auto& ptr : backward_op_descs) {
     all_ops.push_back(std::move(ptr));
   }
-  root_block->Var(fill_one_op_out);
+  // Create Variable
+
+  // Create target gradient variable
+  std::unordered_map<std::string, GradVarInfo> retv;
+
+  auto var = root_block->Var(fill_one_op_out);
+  // FIXME(qiao) infer the data type
+  var->SetDataType(framework::DataType::FP32);
+  var->SetShape(target.Shape());
+  auto& target_grad = retv[target.Name()];
+  target_grad.name_ = fill_one_op_out;
+  target_grad.block_idx_ = root_block_idx;
+  target_grad.op_idx_ = static_cast<int>(forward_op_num);
 
   // create grad_var for all blocks in this program
   CreateGradVarInBlock(forward_op_num, grad_to_var, root_block, &retv);

paddle/framework/backward_test.cc

@@ -26,6 +26,20 @@ namespace framework {
 
 using DeviceContext = platform::DeviceContext;
 
+class NoneOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext *ctx) const override {}
+};
+
+template <typename Place, typename T>
+class NoneKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &context) const override {}
+};
+
 class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
  public:
   RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
@@ -215,19 +229,51 @@ class MinusOpMaker : public OpProtoAndCheckerMaker {
 namespace f = paddle::framework;
 namespace ops = paddle::operators;
 using EnforceNotMet = paddle::platform::EnforceNotMet;
-REGISTER_OPERATOR(rowwise_add, f::NOP, f::RowWiseAddOpMaker,
+// rowwise_add
+REGISTER_OPERATOR(rowwise_add, f::NoneOp, f::RowWiseAddOpMaker,
                   f::RowWiseAddGradMaker);
-REGISTER_OPERATOR(rowwise_add_grad, f::NOP);
-REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP);
-REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP);
-REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker);
-REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NOP, f::FillZeroOpMaker);
-REGISTER_OP(sum, f::NOP, f::SumOpMaker, sum_grad, f::NOP);
+REGISTER_OP_CPU_KERNEL(rowwise_add,
+                       f::NoneKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OPERATOR(rowwise_add_grad, f::NoneOp);
+REGISTER_OP_CPU_KERNEL(rowwise_add_grad,
+                       f::NoneKernel<paddle::platform::CPUPlace, float>);
+// mul
+REGISTER_OP(mul, f::NoneOp, f::MulOpMaker, mul_grad, f::NoneOp);
+REGISTER_OP_CPU_KERNEL(mul, f::NoneKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(mul_grad,
+                       f::NoneKernel<paddle::platform::CPUPlace, float>);
+// sigmoid
+REGISTER_OP(sigmoid, f::NoneOp, f::SigmoidOpMaker, sigmoid_grad, f::NoneOp);
+REGISTER_OP_CPU_KERNEL(sigmoid,
+                       f::NoneKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NoneOp, f::NoGradOpMaker);
+// fill_zeros_like
+REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NoneOp, f::FillZeroOpMaker);
+REGISTER_OP_CPU_KERNEL(fill_zeros_like,
+                       f::NoneKernel<paddle::platform::CPUPlace, float>);
+// sum
+REGISTER_OP(sum, f::NoneOp, f::SumOpMaker, sum_grad, f::NoneOp);
+REGISTER_OP_CPU_KERNEL(sum, f::NoneKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(sum_grad,
+                       f::NoneKernel<paddle::platform::CPUPlace, float>);
+// fc
 REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker);
-REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad,
-            f::NOP);
-REGISTER_OP(mult_in_out, f::NOP, f::MultInOutOpMaker, mult_in_out_grad, f::NOP);
-REGISTER_OPERATOR(minus, f::NOP, f::MinusOpMaker, f::MinusGradOpDescMaker);
+// many_output_op
+REGISTER_OP(many_output_op, f::NoneOp, f::ManyOutputOpMaker,
+            many_output_op_grad, f::NoneOp);
+// mult_in_out
+REGISTER_OP(mult_in_out, f::NoneOp, f::MultInOutOpMaker, mult_in_out_grad,
+            f::NoneOp);
+REGISTER_OP_CPU_KERNEL(mult_in_out,
+                       f::NoneKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(mult_in_out_grad,
+                       f::NoneKernel<paddle::platform::CPUPlace, float>);
+// minus
+REGISTER_OPERATOR(minus, f::NoneOp, f::MinusOpMaker, f::MinusGradOpDescMaker);
+REGISTER_OP_CPU_KERNEL(minus, f::NoneKernel<paddle::platform::CPUPlace, float>);
+// scale
+REGISTER_OPERATOR(scale, f::NoneOp);
+REGISTER_OP_CPU_KERNEL(scale, f::NoneKernel<paddle::platform::CPUPlace, float>);
 
 TEST(Backward, simple_op_not_need_grad) {
   auto fwd = f::OpRegistry::CreateOp(
@@ -463,6 +509,7 @@ TEST(Backward, simple_single_op) {
   f::ProgramDesc *program_desc = GetNewProgramDesc();
   f::ProgramDescBind &program = f::ProgramDescBind::Instance(program_desc);
   f::BlockDescBind *block = program.Block(0);
+
   f::OpDescBind *op = block->AppendOp();
   op->SetType("rowwise_add");
   op->SetInput("X", {"x"});
@@ -487,7 +534,7 @@ TEST(Backward, simple_single_op) {
   EXPECT_EQ(grad_op->Output(f::GradVarName("b")),
             std::vector<std::string>({f::GradVarName("b")}));
 
-  EXPECT_EQ(var_to_grad.size(), 2UL);
+  EXPECT_EQ(var_to_grad.size(), 3UL);
   EXPECT_EQ(var_to_grad.at("b"), f::GradVarInfo(f::GradVarName("b"), 0, 2));
   EXPECT_EQ(var_to_grad.at("x"), f::GradVarInfo(f::GradVarName("x"), 0, 2));
 
@@ -588,7 +635,7 @@ TEST(Backward, simple_mult_op) {
   EXPECT_EQ(grad_op3->Output(f::GradVarName("b")),
             std::vector<std::string>({f::GradVarName("b3")}));
 
-  EXPECT_EQ(var_to_grad.size(), 6UL);
+  EXPECT_EQ(var_to_grad.size(), 7UL);
   EXPECT_EQ(var_to_grad.at("x1"), f::GradVarInfo(f::GradVarName("x1"), 0, 6));
   EXPECT_EQ(var_to_grad.at("b1"), f::GradVarInfo(f::GradVarName("b1"), 0, 6));
   EXPECT_EQ(var_to_grad.at("out1"),
@@ -666,7 +713,7 @@ TEST(Backward, intermedia_var_no_grad) {
             std::vector<std::string>({f::GradVarName("out1")}));
   EXPECT_EQ(grad_op4->Output(f::GradVarName("Y")), std::vector<std::string>());
 
-  EXPECT_EQ(var_to_grad.size(), 3UL);
+  EXPECT_EQ(var_to_grad.size(), 4UL);
   EXPECT_EQ(var_to_grad.at("x1"), f::GradVarInfo(f::GradVarName("x1"), 0, 6));
   EXPECT_EQ(var_to_grad.at("b1"), f::GradVarInfo(f::GradVarName("b1"), 0, 6));
   EXPECT_EQ(var_to_grad.at("out1"),
@@ -744,7 +791,7 @@ TEST(Backward, var_no_grad) {
   EXPECT_EQ(grad_op1->Output(f::GradVarName("H")),
             std::vector<std::string>({f::GradVarName("h1")}));
 
-  EXPECT_EQ(var_to_grad.size(), 3UL);
+  EXPECT_EQ(var_to_grad.size(), 4UL);
   EXPECT_EQ(var_to_grad.at("y1"), f::GradVarInfo(f::GradVarName("y1"), 0, 3));
   EXPECT_EQ(var_to_grad.at("x1"), f::GradVarInfo(f::GradVarName("x1"), 0, 5));
   EXPECT_EQ(var_to_grad.at("h1"), f::GradVarInfo(f::GradVarName("h1"), 0, 5));
@@ -830,7 +877,7 @@ TEST(Backward, shared_var) {
   EXPECT_EQ(grad_op1->Output(f::GradVarName("b")),
             std::vector<std::string>({f::GradVarName("b1")}));
 
-  EXPECT_EQ(var_to_grad.size(), 5UL);
+  EXPECT_EQ(var_to_grad.size(), 6UL);
   EXPECT_EQ(var_to_grad.at("b3"), f::GradVarInfo(f::GradVarName("b3"), 0, 4));
   EXPECT_EQ(var_to_grad.at("y2"), f::GradVarInfo(f::GradVarName("y2"), 0, 5));
   EXPECT_EQ(var_to_grad.at("out1"),
@@ -863,7 +910,7 @@ TEST(Backward, half_backward) {
   auto ops = block->AllOps();
   ASSERT_EQ(3UL, ops.size());
 
-  EXPECT_EQ(var_to_grad.size(), 1UL);
+  EXPECT_EQ(var_to_grad.size(), 2UL);
   EXPECT_EQ(var_to_grad.at("a"),
             f::GradVarInfo(f::GradVarName("a"), 0, forward_len + 1));
 }

paddle/operators/mul_op.cc

@@ -106,10 +106,10 @@ class MulOpGrad : public framework::OperatorWithKernel {
     auto y_dims = ctx->GetInputDim("Y");
     auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
 
-    auto x_mat_dims =
-        framework::flatten_to_2d(x_dims, Attr<int>("x_num_col_dims"));
-    auto y_mat_dims =
-        framework::flatten_to_2d(y_dims, Attr<int>("y_num_col_dims"));
+    auto x_mat_dims = framework::flatten_to_2d(
+        x_dims, ctx->Attrs().Get<int>("x_num_col_dims"));
+    auto y_mat_dims = framework::flatten_to_2d(
+        y_dims, ctx->Attrs().Get<int>("y_num_col_dims"));
 
     PADDLE_ENFORCE_EQ(
         x_mat_dims[0], out_dims[0],

paddle/pybind/protobuf.cc

@@ -163,6 +163,11 @@ void BindBlockDesc(py::module &m) {
              return self.Var(name);
            },
            py::return_value_policy::reference)
+      .def("has_var",
+           [](BlockDescBind &self, py::bytes byte_name) {
+             std::string name = byte_name;
+             return self.HasVar(name);
+           })
       .def("find_var",
            [](BlockDescBind &self, py::bytes byte_name) {
              std::string name = byte_name;

python/paddle/v2/framework/framework.py

@@ -306,6 +306,14 @@ def parent_idx(self):
     def idx(self):
         return self.desc.id
 
+    def var(self, name):
+        if name not in self.vars:
+            raise ValueError("var %s not in this block" % name)
+        return self.vars[name]
+
+    def all_parameters(self):
+        return {v for k, v in self.vars.iteritems() if isinstance(v, Parameter)}
+
     def create_var(self, *args, **kwargs):
         return Variable(self, *args, **kwargs)
 
@@ -314,7 +322,8 @@ def has_var(self, name):
 
     def create_parameter(self, *args, **kwargs):
         global_block = self.program.global_block()
-        return Parameter(global_block, *args, **kwargs)
+        param = Parameter(global_block, *args, **kwargs)
+        return param
 
     def append_op(self, *args, **kwargs):
         op_desc = self.desc.append_op()
@@ -392,10 +401,16 @@ def __str__(self):
     def global_block(self):
         return self.blocks[0]
 
+    def block(self, index):
+        return self.blocks[index]
+
     def current_block(self):
         return self.blocks[self.current_block_idx]
 
     def append_backward(self, target, no_grad_set):
+        """
+        return map(param_name -> (grad_name, block_index, op_index))
+        """
         assert isinstance(target, Variable)
         param_to_grad_info = self.desc.append_backward(target.desc, no_grad_set)
         self.sync_with_cpp()