Finish framework of squared_l2_distance_op.

paddle/operators/CMakeLists.txt

@@ -73,3 +73,5 @@ op_library(uniform_random_op SRCS uniform_random_op.cc uniform_random_op.cu)
 op_library(lookup_table_op SRCS lookup_table_op.cc lookup_table_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)
+
+op_library(squared_l2_distance_op SRCS squared_l2_distance_op.cc squared_l2_distance_op.cu)

paddle/operators/squared_l2_distance_op.cc

@@ -0,0 +1,82 @@
+/* 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/squared_l2_distance_op.h"
+
+namespace paddle {
+namespace operators {
+
+class SquaredL2DistanceOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
+                            "Input of SquaredL2DistanceOp "
+                            "must be initialized.");
+    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
+                      ctx.Input<Tensor>("Y")->dims(),
+                      "Dimensions of SquaredL2DistanceOp's two inputs "
+                      "must be same.")
+    framework::DDim dims = ctx.Input<Tensor>("X")->dims();
+    ctx.Output<Tensor>("sub_result")->Resize(dims);
+    ctx.Output<Tensor>("Out")->Resize(framework::make_ddim({dims[0], 1}));
+  }
+};
+
+class SquaredL2DistanceOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  SquaredL2DistanceOpMaker(framework::OpProto *proto,
+                           framework::OpAttrChecker *op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "Input value.");
+    AddInput("Y", "Target value.");
+    AddOutput("sub_result",
+              "Buffering substraction result which "
+              "will be reused in backward.")
+        .AsIntermediate();
+    AddOutput("Out", "Squared l2 distance between input and target.");
+    AddComment(R"DOC(
+    SquaredL2DistanceOp will cacluate the squared L2 distances for
+    input and target. Number of distance value equals to the
+    first dimension of input.
+    )DOC");
+  }
+};
+
+class SquaredL2DistanceGradOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(const framework::InferShapeContext &ctx) const override {
+    ctx.Output<Tensor>(framework::GradVarName("X"))
+        ->Resize(ctx.Input<Tensor>("X")->dims());
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(squared_l2_distance, ops::SquaredL2DistanceOp,
+            ops::SquaredL2DistanceOpMaker, squared_l2_distance_grad,
+            ops::SquaredL2DistanceGradOp);
+REGISTER_OP_CPU_KERNEL(
+    squared_l2_distance,
+    ops::SquaredL2DistanceKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    squared_l2_distance_grad,
+    ops::SquaredL2DistanceGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/squared_l2_distance_op.cu

@@ -0,0 +1,25 @@
+/* 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. */
+
+#define EIGEN_USE_GPU
+
+#include "paddle/operators/squared_l2_distance_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(
+    squared_l2_distance,
+    ops::SquaredL2DistanceKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    squared_l2_distance_grad,
+    ops::SquaredL2DistanceGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/squared_l2_distance_op.h

@@ -0,0 +1,84 @@
+/* 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 EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
+template <typename T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+
+template <typename Place, typename T>
+class SquaredL2DistanceKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* input0 = context.Input<Tensor>("X");
+    auto* input1 = context.Input<Tensor>("Y");
+    auto* output0 = context.Output<Tensor>("sub_result");
+    auto* output1 = context.Output<Tensor>("Out");
+
+    output0->mutable_data<T>(context.GetPlace());
+    output1->mutable_data<T>(context.GetPlace());
+
+    auto X = EigenMatrix<T>::From(*input0);
+    auto Y = EigenMatrix<T>::From(*input1);
+    auto subResult = EigenMatrix<T>::From(*output0);
+    auto Z = EigenMatrix<T>::From(*output1);
+
+    auto place = context.GetEigenDevice<Place>();
+    // buffer the substraction result
+    subResult.device(place) = X - Y;
+    const auto& inDims = X.dimensions();
+    const auto& subResMat = subResult.reshape(Eigen::array<int, 2>(
+        {static_cast<int>(inDims[0]), static_cast<int>(X.size() / inDims[0])}));
+    Z.device(place) = subResMat.pow(2).sum(Eigen::array<int, 1>({1}));
+  }
+};
+
+template <typename Place, typename T>
+class SquaredL2DistanceGradKernel : public framework::OpKernel {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* input0 = context.Input<Tensor>("sub_result");
+    auto* OG = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto* IG = context.Output<Tensor>(framework::GradVarName("X"));
+
+    IG->mutable_data<T>(context.GetPlace());
+
+    auto subResult = EigenMatrix<T>::From(*input0);
+    auto outGrad = EigenMatrix<T>::From(*OG);
+    auto inGrad = EigenMatrix<T>::From(*IG);
+
+    const auto& subResDims = subResult.dimensions();
+    int firstDim = static_cast<int>(subResDims[0]);
+    int cols = subResult.size() / firstDim;
+    const auto subResMat =
+        subResult.reshape(Eigen::array<int, 2>({firstDim, cols}));
+    // create a matrix view for input gradient tensor
+    auto inGradMat = inGrad.reshape(Eigen::array<int, 2>({firstDim, cols}));
+    inGradMat.device(context.GetEigenDevice<Place>()) =
+        2 * (outGrad.broadcast(Eigen::array<int, 2>({1, cols}))) * subResMat;
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle

paddle/pybind/CMakeLists.txt

@@ -18,5 +18,6 @@ cc_library(paddle_pybind SHARED
     fill_zeros_like_op
     lookup_table_op
     scale_op
-    minus_op)
+    minus_op
+    squared_l2_distance_op)
 endif(WITH_PYTHON)

paddle/pybind/pybind.cc

@@ -48,6 +48,7 @@ USE_OP_ITSELF(identity);
 USE_OP(minus);
 USE_CPU_ONLY_OP(gather);
 USE_CPU_ONLY_OP(scatter);
+USE_OP(squared_l2_distance);
 
 namespace paddle {
 namespace framework {

python/paddle/v2/framework/tests/CMakeLists.txt

@@ -32,3 +32,4 @@ py_test(test_gradient_checker SRCS test_gradient_checker.py)
 py_test(test_lookup_table SRCS test_lookup_table.py)
 py_test(test_scale_and_identity_op SRCS test_scale_and_identity_op.py)
 py_test(mnist SRCS mnist.py)
+py_test(test_squared_l2_distance_op SRCS test_squared_l2_distance_op.py)

python/paddle/v2/framework/tests/test_squared_l2_distance_op.py

@@ -0,0 +1,25 @@
+import unittest
+from op_test_util import OpTestMeta
+from gradient_checker import GradientChecker, create_op
+import numpy as np
+
+
+class TestSquaredL2DistanceOp(unittest.TestCase):
+    __metaclass__ = OpTestMeta
+
+    def setUp(self):
+        self.type = 'squared_l2_distance'
+        self.inputs = {
+            'X': np.random.uniform(0.1, 1., (2, 3)).astype('float32'),
+            'Y': np.random.uniform(0.1, 1., (2, 3)).astype('float32')
+        }
+        subRes = self.inputs['X'] - self.inputs['Y']
+        output = subRes * subRes
+        self.outputs = {
+            'sub_result': subRes,
+            'Out': np.expand_dims(output.sum(1), 1)
+        }
+
+
+if __name__ == '__main__':
+    unittest.main()