Merge remote-tracking branch 'baidu/develop' into executor_impl

doc/api/v1/index_cn.rst

@@ -21,7 +21,7 @@ Model Config API
     trainer_config_helpers/optimizers.rst
     trainer_config_helpers/data_sources.rst
     trainer_config_helpers/layers.rst
-    trainer_config_helpers/activations.rst 
+    trainer_config_helpers/activations.rst
     trainer_config_helpers/poolings.rst
     trainer_config_helpers/networks.rst
     trainer_config_helpers/evaluators.rst

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

@@ -22,7 +22,7 @@ cc_library(attribute SRCS attribute.cc DEPS framework_proto)
 cc_library(proto_desc SRCS var_desc.cc op_desc.cc block_desc.cc program_desc.cc DEPS attribute)
 cc_library(op_proto_maker SRCS op_proto_maker.cc DEPS framework_proto attribute)
 cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
-cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
+cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto proto_desc)
 cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
 

paddle/framework/details/op_registry.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/op_info.h"
+#include "paddle/framework/op_proto_maker.h"
+#include "paddle/framework/operator.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+enum OpInfoFillType {
+  kOperator = 0,
+  kOpProtoAndCheckerMaker = 1,
+  kGradOpDescMaker = 2
+};
+
+template <typename T>
+struct OpInfoFillTypeID {
+  static constexpr OpInfoFillType ID() {
+    return std::is_base_of<OperatorBase, T>::value
+               ? kOperator
+               : (std::is_base_of<OpProtoAndCheckerMaker, T>::value
+                      ? kOpProtoAndCheckerMaker
+                      : (std::is_base_of<GradOpDescMakerBase, T>::value
+                             ? kGradOpDescMaker
+                             : static_cast<OpInfoFillType>(-1)));
+  }
+};
+
+template <typename T, OpInfoFillType = OpInfoFillTypeID<T>::ID()>
+struct OpInfoFiller;
+
+template <size_t I, bool at_end, typename... ARGS>
+class OperatorRegistrarRecursive;
+
+template <size_t I, typename... ARGS>
+class OperatorRegistrarRecursive<I, false, ARGS...> {
+ public:
+  using T = typename std::tuple_element<I, std::tuple<ARGS...>>::type;
+  OperatorRegistrarRecursive(const char* op_type, OpInfo* info) {
+    OpInfoFiller<T> fill;
+    fill(op_type, info);
+    constexpr auto size = sizeof...(ARGS);
+    OperatorRegistrarRecursive<I + 1, I + 1 == size, ARGS...> reg(op_type,
+                                                                  info);
+    (void)(reg);
+  }
+};
+
+template <size_t I, typename... ARGS>
+class OperatorRegistrarRecursive<I, true, ARGS...> {
+ public:
+  OperatorRegistrarRecursive(const char* op_type, OpInfo* info) {}
+};
+
+template <typename T>
+struct OpInfoFiller<T, kOperator> {
+  void operator()(const char* op_type, OpInfo* info) const {
+    info->creator_ = [](const std::string& type, const VariableNameMap& inputs,
+                        const VariableNameMap& outputs,
+                        const AttributeMap& attrs) {
+      return new T(type, inputs, outputs, attrs);
+    };
+  }
+};
+
+template <typename T>
+struct OpInfoFiller<T, kOpProtoAndCheckerMaker> {
+  void operator()(const char* op_type, OpInfo* info) const {
+    info->proto_ = new OpProto;
+    info->checker_ = new OpAttrChecker();
+    auto maker = T(info->proto_, info->checker_);
+    maker.Validate();
+    info->proto_->set_type(op_type);
+    PADDLE_ENFORCE(
+        info->proto_->IsInitialized(),
+        "Fail to initialize %s's OpProto, because %s is not initialized",
+        op_type, info->proto_->InitializationErrorString());
+  }
+};
+
+template <typename T>
+struct OpInfoFiller<T, kGradOpDescMaker> {
+  void operator()(const char* op_type, OpInfo* info) const {
+    info->grad_op_maker_ = new T();
+  }
+};
+}  // namespace details
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/op_info.h

@@ -17,8 +17,8 @@
 #include <map>
 #include <string>
 #include <unordered_map>
-
 #include "paddle/framework/attribute.h"
+#include "paddle/framework/op_desc.h"
 
 namespace paddle {
 namespace framework {
@@ -29,11 +29,18 @@ using OpCreator = std::function<OperatorBase*(
     const std::string& /*type*/, const VariableNameMap& /*inputs*/,
     const VariableNameMap& /*outputs*/, const AttributeMap& /*attrs*/)>;
 
+class GradOpDescMakerBase {
+ public:
+  virtual ~GradOpDescMakerBase() = default;
+  virtual std::vector<OpDescBind> operator()(const OpDescBind&) const = 0;
+};
+
 struct OpInfo {
   OpCreator creator_;
   std::string grad_op_type_;
-  OpProto* proto_;
-  OpAttrChecker* checker_;
+  GradOpDescMakerBase* grad_op_maker_{nullptr};
+  OpProto* proto_{nullptr};
+  OpAttrChecker* checker_{nullptr};
 
   bool HasOpProtoAndChecker() const {
     return proto_ != nullptr && checker_ != nullptr;

paddle/framework/op_registry.h

@@ -21,49 +21,42 @@ limitations under the License. */
 #include <unordered_map>
 #include <unordered_set>
 #include "paddle/framework/attribute.h"
+#include "paddle/framework/details/op_registry.h"
 #include "paddle/framework/framework.pb.h"
 #include "paddle/framework/grad_op_builder.h"
-#include "paddle/framework/op_info.h"
-#include "paddle/framework/op_proto_maker.h"
 #include "paddle/framework/operator.h"
 #include "paddle/framework/scope.h"
 
 namespace paddle {
 namespace framework {
 
+template <typename... ARGS>
+struct OperatorRegistrar {
+  explicit OperatorRegistrar(const char* op_type) : op_type(op_type) {
+    PADDLE_ENFORCE(!OpInfoMap::Instance().Has(op_type),
+                   "'%s' is registered more than once.", op_type);
+    static_assert(sizeof...(ARGS) != 0,
+                  "OperatorRegistrar should be invoked at least by OpClass");
+    details::OperatorRegistrarRecursive<0, false, ARGS...>(op_type, &info);
+  }
+
+  ~OperatorRegistrar() { OpInfoMap::Instance().Insert(op_type, info); }
+
+  const char* op_type;
+
+  OpInfo info;
+};
+
 class OpRegistry {
  public:
   template <typename OpType, typename ProtoMakerType, typename GradOpType>
   static void RegisterOp(const std::string& op_type,
                          const std::string& grad_op_type) {
-    PADDLE_ENFORCE(!OpInfoMap::Instance().Has(op_type),
-                   "'%s' is registered more than once.", op_type);
-    OpInfo op_info;
-    op_info.creator_ = [](
-        const std::string& type, const VariableNameMap& inputs,
-        const VariableNameMap& outputs, const AttributeMap& attrs) {
-      return new OpType(type, inputs, outputs, attrs);
-    };
-    op_info.grad_op_type_ = grad_op_type;
-    if (std::type_index(typeid(ProtoMakerType)) !=
-        std::type_index(typeid(NOPMaker))) {
-      op_info.proto_ = new OpProto;
-      op_info.checker_ = new OpAttrChecker;
-      auto maker = ProtoMakerType(op_info.proto_, op_info.checker_);
-      maker.Validate();
-      op_info.proto_->set_type(op_type);
-      PADDLE_ENFORCE(
-          op_info.proto_->IsInitialized(),
-          "Fail to initialize %s's OpProto, because %s is not initialized",
-          op_type, op_info.proto_->InitializationErrorString());
-    } else {
-      op_info.proto_ = nullptr;
-      op_info.checker_ = nullptr;
-    }
-    OpInfoMap::Instance().Insert(op_type, op_info);
+    OperatorRegistrar<OpType, ProtoMakerType> reg(op_type.c_str());
+    reg.info.grad_op_type_ = grad_op_type;
     // register gradient op
     if (!grad_op_type.empty()) {
-      RegisterOp<GradOpType, NOPMaker, NOP>(grad_op_type, "");
+      OperatorRegistrar<GradOpType> grad_reg(grad_op_type.c_str());
     }
   }
 

paddle/framework/op_registry_test.cc

@@ -173,3 +173,14 @@ TEST(OpRegistry, CustomChecker) {
   int test_attr = op->Attr<int>("test_attr");
   ASSERT_EQ(test_attr, 4);
 }
+
+class CosineOpComplete : public paddle::framework::CosineOp {
+ public:
+  DEFINE_OP_CONSTRUCTOR(CosineOpComplete, paddle::framework::CosineOp);
+  DEFINE_OP_CLONE_METHOD(CosineOpComplete);
+};
+
+TEST(OperatorRegistrar, Test) {
+  using namespace paddle::framework;
+  OperatorRegistrar<CosineOpComplete, CosineOpProtoAndCheckerMaker> reg("cos");
+}
\ No newline at end of file

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

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):
         """
@@ -6473,7 +6476,7 @@ def switch_order_layer(input,
                        act=None,
                        layer_attr=None):
     """
-    This layer switch dimension order of image input. 
+    This layer switch dimension order of image input.
     From order "batchSize, channels, height, width"
     to order "batchSize, height, width, channels".
 
@@ -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)