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

doc/api/v2/config/layer.rst

@@ -345,6 +345,11 @@ clip
 ..  autoclass:: paddle.v2.layer.clip
     :noindex:
 
+resize
+------
+..  autoclass:: paddle.v2.layer.resize
+    :noindex:
+
 slope_intercept
 ---------------
 ..  autoclass:: paddle.v2.layer.slope_intercept

doc/howto/dev/new_op_cn.md

@@ -206,7 +206,7 @@ MulOp(const std::string &type, const framework::VariableNameMap &inputs,
 
     - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，注册`ops::MulOpGrad`，类型名为`mul_grad`。
     - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。
-    - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。
+    - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulGradKernel`类。
 
 
 - 在 `.cu`文件中注册GPU Kernel。

doc/howto/dev/new_op_en.md

@@ -205,7 +205,7 @@ The definition of its corresponding backward operator, if applicable, is similar
 
     - `REGISTER_OP` registers the `ops::MulOp` class, type named `mul`, its type `ProtoMaker` is `ops::MulOpMaker`, registering `ops::MulOpGrad` as `mul_grad`.
     - `REGISTER_OP_WITHOUT_GRADIENT` registers an operator without gradient.
-    - `REGISTER_OP_CPU_KERNEL` registers `ops::MulKernel` class and specialized template types `paddle::platform::CPUPlace` and `float`, which also registers `ops::MulKernel`.
+    - `REGISTER_OP_CPU_KERNEL` registers `ops::MulKernel` class and specialized template types `paddle::platform::CPUPlace` and `float`, which also registers `ops::MulGradKernel`.
 
 
 - Registering GPU Kernel in `.cu` files

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/framework/operator.cc

@@ -245,5 +245,12 @@ std::vector<Tensor*> InferShapeContext::MultiOutput<Tensor>(
   return res;
 }
 
+std::ostream& operator<<(std::ostream& os,
+                         const OperatorWithKernel::OpKernelKey& kernel_key) {
+  os << "place[" << kernel_key.place_ << "]:data_type[" << kernel_key.data_type_
+     << "]";
+  return os;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/framework/operator.h

@@ -478,9 +478,25 @@ class OperatorWithKernel : public OperatorBase {
     this->InferShape(&infer_shape_ctx);
 
     ExecutionContext ctx(*this, scope, dev_ctx);
-    auto& opKernel = AllOpKernels().at(type_).at(
-        OpKernelKey(IndicateDataType(ctx), dev_ctx));
-    opKernel->Compute(ctx);
+
+    // check if op[type] has kernel registered.
+    auto& all_op_kernels = AllOpKernels();
+    auto kernels_iter = all_op_kernels.find(type_);
+    if (kernels_iter == all_op_kernels.end()) {
+      PADDLE_THROW("op[%s] has no kernel", type_);
+    }
+
+    // check if op[type] have kernel for kernel_key
+    OpKernelMap& kernels = kernels_iter->second;
+    auto kernel_key = OpKernelKey(IndicateDataType(ctx), dev_ctx);
+    auto kernel_iter = kernels.find(kernel_key);
+
+    if (kernel_iter == kernels.end()) {
+      PADDLE_THROW("op[%s] has no kernel with kernel_key[%s]", type_,
+                   kernel_key);
+    }
+
+    kernel_iter->second->Compute(ctx);
   }
 
   static std::unordered_map<std::string /* op_type */, OpKernelMap>&
@@ -529,5 +545,8 @@ class OperatorWithKernel : public OperatorBase {
   }
 };
 
+std::ostream& operator<<(std::ostream& os,
+                         const OperatorWithKernel::OpKernelKey& kernel_key);
+
 }  // namespace framework
 }  // namespace paddle

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/trainer_config_helpers/layers.py

@@ -142,6 +142,7 @@ __all__ = [
     'img_pool3d_layer',
     'scale_shift_layer',
     'img_conv3d_layer',
+    'resize_layer',
 ]
 
 
@@ -250,6 +251,8 @@ class LayerType(object):
     KMAX_SEQ_SCORE = 'kmax_seq_score'
     SCALE_SHIFT_LAYER = 'scale_shift'
 
+    RESIZE = 'resize'
+
     @staticmethod
     def is_layer_type(type_name):
         """
@@ -6932,3 +6935,23 @@ def scale_shift_layer(input, name=None, param_attr=None, bias_attr=None):
         bias=ParamAttr.to_bias(bias_attr))
     return LayerOutput(
         name, LayerType.SCALE_SHIFT_LAYER, parents=[input], size=input.size)
+
+
+@wrap_name_default("resize")
+def resize_layer(input, size, name=None):
+    """
+    The resize layer resizes the input matrix with a shape of [Height, Width]
+    into the output matrix with a shape of [Height x Width / size, size],
+    where size is the parameter of this layer indicating the output dimension.
+
+    :param input: The input to this layer.
+    :type input: LayerOutput.
+    :param name: The name of this layer. It is optional.
+    :type name: basestring
+    :param size: The resized output dimesion of this layer.
+    :type size: int
+    :return: A LayerOutput object.
+    :rtype: LayerOutput
+    """
+    Layer(name=name, type=LayerType.RESIZE, inputs=Input(input.name), size=size)
+    return LayerOutput(name, LayerType.RESIZE, parents=[input], size=input.size)

python/paddle/trainer_config_helpers/tests/configs/file_list.sh

@@ -10,6 +10,6 @@ test_prelu_layer test_row_conv test_detection_output_layer test_multibox_loss_la
 test_recursive_topology test_gated_unit_layer test_clip_layer test_row_l2_norm_layer
 test_kmax_seq_socre_layer test_sub_nested_seq_select_layer test_scale_shift_layer
 test_seq_slice_layer test_cross_entropy_over_beam test_pooling3D_layer
-test_conv3d_layer test_deconv3d_layer test_BatchNorm3D)
+test_conv3d_layer test_deconv3d_layer test_BatchNorm3D test_resize_layer)
 
 export whole_configs=(test_split_datasource)

python/paddle/trainer_config_helpers/tests/configs/protostr/test_resize_layer.protostr

+type: "nn"
+layers {
+  name: "input"
+  type: "data"
+  size: 300
+  active_type: ""
+}
+layers {
+  name: "__resize_0__"
+  type: "resize"
+  size: 150
+  active_type: ""
+  inputs {
+    input_layer_name: "input"
+  }
+}
+input_layer_names: "input"
+output_layer_names: "__resize_0__"
+sub_models {
+  name: "root"
+  layer_names: "input"
+  layer_names: "__resize_0__"
+  input_layer_names: "input"
+  output_layer_names: "__resize_0__"
+  is_recurrent_layer_group: false
+}
+

python/paddle/trainer_config_helpers/tests/configs/test_resize_layer.py

+from paddle.trainer_config_helpers import *
+
+data = data_layer(name='input', size=300)
+resized = resize_layer(input=data, size=150)
+
+outputs(resized)

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))