Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into develop

paddle/gserver/gradientmachines/RecurrentGradientMachine.cpp

@@ -1399,8 +1399,8 @@ void RecurrentGradientMachine::createDataOutlinkCopySizeInfo(
              getBeamSize() > 1 ? finalPaths_.size() : finalPaths_[0].size());
     int* starts = inputSeqStartPos->getMutableData(false);
     int seqId = 0;
-    for (int i = 0; i < finalPaths_.size(); ++i) {
+    for (size_t i = 0; i < finalPaths_.size(); ++i) {
-      for (int j = 0; j < finalPaths_[i].size(); ++j) {
+      for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
         copySize[seqId] = getBeamSize() > 1 ? starts[i + 1] - starts[i]
                                             : starts[j + 1] - starts[j];
         batchMachineStartPos_[seqId + 1] =

paddle/gserver/layers/SequenceSliceLayer.cpp

@@ -200,8 +200,9 @@ void SequenceSliceLayer::forward(PassType passType) {
       startIdsOnCpu_ = getInputValue(1);
       endIdsOnCpu_ = getInputValue(2);
     }
-  } else
+  } else {
     copySliceIdsToCpu();
+  }
 
   // calculate the selected row indices in a batch,
   // and build the output sequence information.

paddle/operators/CMakeLists.txt

@@ -70,3 +70,4 @@ op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
 op_library(uniform_random_op
         SRCS uniform_random_op.cc uniform_random_op.cu)
 op_library(scale_op SRCS scale_op.cc scale_op.cu DEPS net_op)
+op_library(minus_op SRCS minus_op.cc minus_op.cu DEPS scale_op)

paddle/operators/minus_op.cc

@@ -0,0 +1,87 @@
+/* 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/minus_op.h"
+#include "paddle/operators/net_op.h"
+
+namespace paddle {
+namespace operators {
+
+class MinusOp : public framework::OperatorWithKernel {
+ public:
+  MinusOp(const std::string &type, const framework::VariableNameMap &inputs,
+          const framework::VariableNameMap &outputs,
+          const framework::AttributeMap &attrs)
+      : OperatorWithKernel(type, inputs, outputs, attrs) {}
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    auto *left_tensor = ctx.Input<framework::Tensor>("X");
+    auto *right_tensor = ctx.Input<framework::Tensor>("Y");
+
+    PADDLE_ENFORCE_EQ(
+        framework::product(left_tensor->dims()),
+        framework::product(right_tensor->dims()),
+        "Minus operator must take two tensor with same num of elements");
+    ctx.Output<framework::Tensor>("Out")->Resize(left_tensor->dims());
+  }
+};
+
+class MinusOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  MinusOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The left tensor of minus operator.").NotInGradient();
+    AddInput("Y", "The right tensor of minus operator.").NotInGradient();
+    AddOutput("Out", "The output tensor of minus operator.").NotInGradient();
+
+    AddComment(R"DOC(Minus Operator
+
+Equation: Out = X - Y
+)DOC");
+  }
+};
+template <typename AttrType>
+class MinusGradOp : public NetOp {
+ public:
+  MinusGradOp(const std::string &type, const framework::VariableNameMap &inputs,
+              const framework::VariableNameMap &outputs,
+              const framework::AttributeMap &attrs)
+      : NetOp(type, inputs, outputs, attrs) {
+    auto out_grad = Input(framework::GradVarName("Out"));
+    auto x_grad = Output(framework::GradVarName("X"));
+    auto y_grad = Output(framework::GradVarName("Y"));
+
+    // x_grad = out_grad
+    AppendOp(framework::OpRegistry::CreateOp("identity", {{"X", {out_grad}}},
+                                             {{"Out", {x_grad}}}, {}));
+
+    framework::AttributeMap scale_attr;
+    scale_attr["scale"] = static_cast<AttrType>(-1);
+    AppendOp(framework::OpRegistry::CreateOp("scale", {{"X", {out_grad}}},
+                                             {{"Out", {y_grad}}}, scale_attr));
+    CompleteAddOp(false);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+USE_OP(scale);
+USE_OP_ITSELF(identity);
+namespace ops = paddle::operators;
+REGISTER_OP(minus, ops::MinusOp, ops::MinusOpMaker, minus_grad,
+            ops::MinusGradOp<float>);
+REGISTER_OP_CPU_KERNEL(minus,
+                       ops::MinusKernel<paddle::platform::CPUPlace, float>);

paddle/operators/minus_op.cu

@@ -0,0 +1,18 @@
+/* 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/minus_op.h"
+
+REGISTER_OP_GPU_KERNEL(
+    minus, paddle::operators::MinusKernel<paddle::platform::GPUPlace, float>);

paddle/operators/minus_op.h

@@ -0,0 +1,39 @@
+/* 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 {
+
+template <typename Place, typename T>
+class MinusKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* left_tensor = context.Input<framework::Tensor>("X");
+    auto* right_tensor = context.Input<framework::Tensor>("Y");
+    auto* out_tensor = context.Output<framework::Tensor>("Out");
+
+    out_tensor->mutable_data<T>(context.GetPlace());
+    auto& dev = context.GetEigenDevice<Place>();
+    framework::EigenVector<T>::Flatten(*out_tensor).device(dev) =
+        framework::EigenVector<T>::Flatten(*left_tensor) -
+        framework::EigenVector<T>::Flatten(*right_tensor);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/net_op_test.cc

@@ -79,7 +79,7 @@ TEST(NetOp, Clone) {
   ASSERT_NE(new_net_op, nullptr);
   ASSERT_TRUE(new_net_op->IsNetOp());
   auto* new_net = static_cast<NetOp*>(new_net_op.get());
-  ASSERT_EQ(2, new_net->ops_.size());
+  ASSERT_EQ(2UL, new_net->ops_.size());
   ASSERT_EQ(new_net->ops_[0]->Type(), "empty");
   ASSERT_EQ(new_net->ops_[1]->Type(), "empty2");
 }

paddle/pybind/CMakeLists.txt

@@ -15,5 +15,6 @@ cc_library(paddle_pybind SHARED
     uniform_random_op
     gaussian_random_op
     fill_zeros_like_op
-    scale_op)
+    scale_op
+    minus_op)
 endif(WITH_PYTHON)

paddle/pybind/pybind.cc

@@ -44,6 +44,7 @@ USE_OP(gaussian_random);
 USE_OP(uniform_random);
 USE_OP(scale);
 USE_OP_ITSELF(identity);
+USE_OP(minus);
 USE_CPU_ONLY_OP(gather);
 
 namespace paddle {

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

@@ -0,0 +1,30 @@
+import unittest
+import numpy as np
+from gradient_checker import GradientChecker, create_op
+from op_test_util import OpTestMeta
+
+
+class MinusOpTest(unittest.TestCase):
+    __metaclass__ = OpTestMeta
+
+    def setUp(self):
+        self.type = "minus"
+        self.inputs = {
+            'X': np.random.random((32, 84)).astype("float32"),
+            'Y': np.random.random((32, 84)).astype("float32")
+        }
+        self.outputs = {'Out': (self.inputs['X'] - self.inputs['Y'])}
+
+
+class MinusGradTest(GradientChecker):
+    def test_left(self):
+        op = create_op("minus")
+        inputs = {
+            "X": np.random.random((10, 10)).astype("float32"),
+            "Y": np.random.random((10, 10)).astype("float32")
+        }
+        self.check_grad(op, inputs, ["X", 'Y'], "Out")
+
+
+if __name__ == '__main__':
+    unittest.main()