sequence_pool_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/sequence_pool_op.h"

namespace paddle {
namespace operators {

class SequencePoolOp : 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(X) of SequencePoolOp should not be null.");
    PADDLE_ENFORCE_NOT_NULL(
        ctx.OutputVar("Out"),
        "Output(Out) of SequencePoolOp should not be null.");

    auto* x = ctx.Input<framework::LoDTensor>("X");
    auto dims = x->dims();
    auto lod = x->lod();
    PADDLE_ENFORCE_EQ(lod.size(), 1UL, "Only support one level sequence now.");
    PADDLE_ENFORCE_GE(
        dims[0],
        /*batch size = */ static_cast<int64_t>(lod[0].size() - 1),
        "The first dimension of Input(X) must be large than batch size.");
    dims[0] = lod[0].size() - 1;
    ctx.Output<framework::LoDTensor>("Out")->Resize({dims});
  }
};

class SequencePoolOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  SequencePoolOpMaker(framework::OpProto* proto,
                      framework::OpAttrChecker* op_checker)
      : OpProtoAndCheckerMaker(proto, op_checker) {
    AddInput("X", "A LoDTensor, the variable-length input of SequencePoolOp");
    AddOutput("Out",
              "A LoDTensor, the variable-length output of SequencePoolOp.");
    AddAttr<int>(
        "strategy",
        "(int, default AVERAGE) the pooling strategy of SequencePoolOp.")
        .SetDefault(AVERAGE)
        .InEnum({AVERAGE, SUM, SQRT, MAX, LAST, FIRST});
    AddComment(R"DOC(
    SequencePoolOp pools features of all time-steps of each instance.

    For a mini-batch of 3 variable lengths sentences, containing 2, 3, and 2 words:

    X = [[1, 3], [2, 4, 6], [5, 1]], 

    and X->lod()[0] = [0, 2, 5, 7]

    then, for different strategy, we get: 

    - AVERAGE: Out = [2, 4, 3], where 2=(1+3)/2, 4=(2+4+6)/3, 3=(5+1)/2
    - SUM: Out = [4, 12, 6], where 4=1+3, 12=2+4+6, 6=5+1
    - SQRT: Out = [2.82, 6.93, 4.24], where 2.82=(1+3)/sqrt(2), 6.93=(2+4+6)/sqrt(3), 
                                            4.24=(5+1)/sqrt(2)
    - MAX: Out = [3, 6, 5], where 3=max(1,3), 6=max(2,4,6), 5=max(5,1)
    - LAST: Out = [3, 6, 1], where 3=last(1,3), 6=last(2,4,6), 1=last(5,1)
    - FIRST: Out = [1, 2, 5], where 1=first(1,3), 2=first(2,4,6), 5=first(5,1)

    and X->lod() is nullptr.
    )DOC");
  }
};

class SequencePoolGradOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

 protected:
  void InferShape(const framework::InferShapeContext& ctx) const override {
    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
                            "Gradient of Out should not be null.");
    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
                            "The input X should not be null.");
    auto og_dims =
        ctx.Input<framework::LoDTensor>(framework::GradVarName("Out"))->dims();
    auto x_dims = ctx.Input<framework::LoDTensor>("X")->dims();
    PADDLE_ENFORCE_EQ(og_dims.size(), x_dims.size(),
                      "The rank of output grad must equal to Input(X).");
    for (int64_t i = 1; i < og_dims.size(); ++i) {
      PADDLE_ENFORCE_EQ(og_dims[i], x_dims[i], "The dimension mismatch.");
    }
    auto* x_grad =
        ctx.Output<framework::LoDTensor>(framework::GradVarName("X"));
    x_grad->Resize(x_dims);
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
REGISTER_OP(sequence_pool, ops::SequencePoolOp, ops::SequencePoolOpMaker,
            sequence_pool_grad, ops::SequencePoolGradOp);
REGISTER_OP_CPU_KERNEL(
    sequence_pool, ops::SequencePoolKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(
    sequence_pool_grad,
    ops::SequencePoolGradKernel<paddle::platform::CPUPlace, float>);