Merge branch 'develop' of github.com:baidu/Paddle into feature/change_grad_reg_mechanism

paddle/framework/CMakeLists.txt

@@ -29,7 +29,7 @@ cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
 cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS operator proto_desc)
 cc_library(op_registry SRCS op_registry.cc DEPS grad_op_builder op_proto_maker op_info)
 cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
-cc_test(grad_op_builder_test SRCS grad_op_builder_test.cc DEPS grad_op_builder op_registry add_op)
+cc_test(grad_op_builder_test SRCS grad_op_builder_test.cc DEPS grad_op_builder op_registry sum_op)
 
 py_proto_compile(framework_py_proto SRCS framework.proto)
 # Generate an empty __init__.py to make framework_py_proto as a valid python module.

paddle/framework/grad_op_builder_test.cc

@@ -3,7 +3,7 @@
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/operator.h"
 
-USE_OP(add);
+USE_OP(sum);
 
 namespace paddle {
 namespace framework {
@@ -41,17 +41,24 @@ namespace f = paddle::framework;
 
 TEST(GradOpBuilder, AddTwo) {
   std::shared_ptr<f::OperatorBase> add_op(f::OpRegistry::CreateOp(
-      "add", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {}));
+      "sum", {{"X", {"x", "y"}}}, {{"Out", {"out"}}}, {}));
   std::shared_ptr<f::OperatorBase> grad_add_op =
       f::OpRegistry::CreateGradOp(*add_op);
-  EXPECT_EQ(grad_add_op->Inputs().size(), 4UL);
-  EXPECT_EQ(grad_add_op->Outputs().size(), 2UL);
-  EXPECT_EQ(grad_add_op->Input("X"), "x");
-  EXPECT_EQ(grad_add_op->Input("Y"), "y");
-  EXPECT_EQ(grad_add_op->Input("Out"), "out");
+
+  EXPECT_EQ(grad_add_op->Inputs().size(), 1UL);
+  EXPECT_EQ(grad_add_op->Outputs().size(), 1UL);
   EXPECT_EQ(grad_add_op->Input(f::GradVarName("Out")), f::GradVarName("out"));
-  EXPECT_EQ(grad_add_op->Output(f::GradVarName("X")), f::GradVarName("x"));
-  EXPECT_EQ(grad_add_op->Output(f::GradVarName("Y")), f::GradVarName("y"));
+  auto &outputs = grad_add_op->Outputs(f::GradVarName("X"));
+  EXPECT_EQ(2UL, outputs.size());
+  auto in_output = [&outputs](const std::string &name) {
+    for (auto &output_name : outputs) {
+      if (output_name == name) return true;
+    }
+    return false;
+  };
+
+  EXPECT_TRUE(in_output(f::GradVarName("x")));
+  EXPECT_TRUE(in_output(f::GradVarName("y")));
 }
 
 REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker, mult_io_grad, f::NOP);

paddle/operators/add_op.cc

-/* 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 "paddle/operators/add_op.h"
-
-namespace paddle {
-namespace operators {
-
-class AddOp : public framework::OperatorWithKernel {
- public:
-  using framework::OperatorWithKernel::OperatorWithKernel;
-
- protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {
-    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of AddOp should not be null.");
-    PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) of AddOp should not be null.");
-    PADDLE_ENFORCE(ctx->HasOutput("Out"),
-                   "Output(Out) of AddOp should not be null.");
-
-    auto x_dims = ctx->GetInputDim("X");
-    auto y_dims = ctx->GetInputDim("Y");
-    PADDLE_ENFORCE_EQ(x_dims, y_dims,
-                      "Two input of Add Op's dimension must be same.");
-    ctx->SetOutputDim("Out", x_dims);
-  }
-};
-
-class AddOpMaker : public framework::OpProtoAndCheckerMaker {
- public:
-  AddOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
-      : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "The first input of add op");
-    AddInput("Y", "The second input of add op");
-    AddOutput("Out", "The output of add op");
-    AddComment(R"DOC(
-Two Element Add Operator.
-
-The equation is: Out = X + Y
-)DOC");
-  }
-};
-
-class AddOpGrad : public framework::OperatorWithKernel {
- public:
-  using framework::OperatorWithKernel::OperatorWithKernel;
-
- protected:
-  void InferShape(framework::InferShapeContextBase* ctx) const override {}
-};
-
-}  // namespace operators
-}  // namespace paddle
-
-namespace ops = paddle::operators;
-REGISTER_OP(add, ops::AddOp, ops::AddOpMaker, add_grad, ops::AddOpGrad);
-
-REGISTER_OP_CPU_KERNEL(add, ops::AddKernel<paddle::platform::CPUPlace, float>);

paddle/operators/add_op.cu

-/* 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 "paddle/operators/add_op.h"
-
-namespace ops = paddle::operators;
-REGISTER_OP_GPU_KERNEL(add, ops::AddKernel<paddle::platform::GPUPlace, float>);

paddle/operators/add_op.h

-/* 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. */
-
-#pragma once
-#include "paddle/framework/eigen.h"
-#include "paddle/framework/op_registry.h"
-
-namespace paddle {
-namespace operators {
-
-using Tensor = framework::Tensor;
-template <typename T, int MajorType = Eigen::RowMajor,
-          typename IndexType = Eigen::DenseIndex>
-using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
-
-template <typename Place, typename T>
-class AddKernel : public framework::OpKernel<T> {
- public:
-  void Compute(const framework::ExecutionContext& context) const override {
-    auto* input0 = context.Input<Tensor>("X");
-    auto* input1 = context.Input<Tensor>("Y");
-    auto* output = context.Output<Tensor>("Out");
-
-    output->mutable_data<T>(context.GetPlace());
-
-    auto X = EigenVector<T>::Flatten(*input0);
-    auto Y = EigenVector<T>::Flatten(*input1);
-    auto Z = EigenVector<T>::Flatten(*output);
-
-    auto place = context.GetEigenDevice<Place>();
-
-    Z.device(place) = X + Y;
-  }
-};
-
-}  // namespace operators
-}  // namespace paddle

paddle/operators/sum_op.cc

@@ -43,8 +43,10 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
   SumOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("X", "the input tensors of sum operator.").AsDuplicable();
-    AddOutput("Out", "the output tensor of sum operator.");
+    AddInput("X", "the input tensors of sum operator.")
+        .AsDuplicable()
+        .NotInGradient();
+    AddOutput("Out", "the output tensor of sum operator.").NotInGradient();
     AddComment(R"DOC(
 Sum the input tensors.
 

python/paddle/v2/framework/tests/test_add_op.py

-import unittest
-import numpy as np
-from op_test import OpTest
-
-
-class TestAddOp(OpTest):
-    def setUp(self):
-        self.op_type = "add"
-        self.inputs = {
-            'X': np.random.random((102, 105)).astype("float32"),
-            'Y': np.random.random((102, 105)).astype("float32")
-        }
-        self.outputs = {'Out': self.inputs['X'] + self.inputs['Y']}
-
-    def test_check_output(self):
-        self.check_output()
-
-
-if __name__ == "__main__":
-    unittest.main()

python/paddle/v2/framework/tests/test_cond_op.py

@@ -15,7 +15,7 @@ class PySimpleCond(object):
         for i in range(1, 10, 2):
             array[i] = 0
         self.cond = np.array(array)
-        self.x = np.ones(shape=(10, 1))
+        self.x = np.ones(shape=(10, 1)).astype("float32")
 
     def forward(self):
         self.index_t = np.where(self.cond == 1)

python/paddle/v2/framework/tests/test_gradient_checker.py

-import unittest
-import numpy as np
-import paddle.v2.framework.core as core
-from op_test import get_numeric_gradient
-from op_test import create_op
-
-
-class GetNumericGradientTest(unittest.TestCase):
-    def test_add_op(self):
-        x = np.random.random((10, 1)).astype("float32")
-        y = np.random.random((10, 1)).astype("float32")
-        z = x + y
-        scope = core.Scope()
-        add_op = create_op(scope, "add", {'X': x, 'Y': y}, {'Out': z}, dict())
-        arr = get_numeric_gradient(scope, add_op, {'X': x,
-                                                   'Y': y}, 'X', ['Out'])
-        self.assertAlmostEqual(arr.mean(), 1.0, delta=1e-4)
-
-    def test_softmax_op(self):
-        def stable_softmax(x):
-            """Compute the softmax of vector x in a numerically stable way."""
-            shiftx = x - np.max(x)
-            exps = np.exp(shiftx)
-            return exps / np.sum(exps)
-
-        def label_softmax_grad(Y, dY):
-            dX = Y * 0.0
-            for i in range(Y.shape[0]):
-                d = np.dot(Y[i, :], dY[i, :])
-                dX[i, :] = Y[i, :] * (dY[i, :] - d)
-            return dX
-
-        X = np.random.random((2, 2)).astype("float32")
-        Y = np.apply_along_axis(stable_softmax, 1, X)
-        dY = np.ones(Y.shape)
-        dX = label_softmax_grad(Y, dY)
-
-        scope = core.Scope()
-        softmax_op = create_op(scope, "softmax", {"X": X}, {"Y": Y}, dict())
-
-        arr = get_numeric_gradient(scope, softmax_op, {"X": X}, "X", "Y")
-        np.testing.assert_almost_equal(arr, dX, decimal=1e-2)
-
-
-if __name__ == "__main__":
-    unittest.main()

python/paddle/v2/framework/tests/test_net.py

@@ -15,7 +15,7 @@ def fc(X, W, Y):
 class TestNet(unittest.TestCase):
     def test_net_all(self):
         net = core.Net.create()
-        op1 = Operator("add", X="X", Y="Y", Out="Out")
+        op1 = Operator("sum", X=["X", "Y"], Out="Out")
         net.append_op(op1)
 
         net2 = core.Net.create()
@@ -26,7 +26,7 @@ class TestNet(unittest.TestCase):
 
         expected = '''
 Op(plain_net), inputs:{all[W, X, Y]}, outputs:{all[Out, fc.out, pre_activation]}.
-    Op(add), inputs:{X[X], Y[Y]}, outputs:{Out[Out]}.
+    Op(sum), inputs:{X[X, Y]}, outputs:{Out[Out]}.
     Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}.
         Op(plain_net), inputs:{all[W, X]}, outputs:{all[fc.out, pre_activation]}.
             Op(mul), inputs:{X[X], Y[W]}, outputs:{Out[pre_activation]}.

python/paddle/v2/framework/tests/test_operator.py

@@ -193,10 +193,10 @@ class TestOpDescCreationMethod(unittest.TestCase):
 
 class TestOpCreations(unittest.TestCase):
     def test_all(self):
-        add_op = op.Operator("add", X="a", Y="b", Out="z")
+        add_op = op.Operator("sum", X=["a", "b"], Out="z")
         self.assertIsNotNone(add_op)
         # Invoke C++ DebugString()
-        self.assertEqual('Op(add), inputs:{X[a], Y[b]}, outputs:{Out[z]}.',
+        self.assertEqual('Op(sum), inputs:{X[a, b]}, outputs:{Out[z]}.',
                          str(add_op))