dgc_momentum_op.cc 2.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// 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/fluid/operators/optimizers/dgc_momentum_op.h"

17 18
#include <string>

19 20 21 22 23 24 25 26 27
namespace paddle {
namespace operators {

class DGCMomentumOp : public MomentumOp {
 public:
  using MomentumOp::MomentumOp;

 protected:
  void InferShape(framework::InferShapeContext* ctx) const override {
28 29 30
    OP_INOUT_CHECK(ctx->HasInput("current_step"),
                   "Input",
                   "current_step",
31 32
                   "DGCMomentumOp");
    OP_INOUT_CHECK(ctx->HasInput("nranks"), "Input", "nranks", "DGCMomentumOp");
33 34
    OP_INOUT_CHECK(
        ctx->HasOutput("Grad_out"), "Output", "Grad_out", "DGCMomentumOp");
35 36 37 38
    return MomentumOp::InferShape(ctx);
  }

  framework::OpKernelType GetKernelTypeForVar(
39 40
      const std::string& var_name,
      const framework::Tensor& tensor,
41
      const framework::OpKernelType& expected_kernel_type) const override {
42
    if (var_name == "current_step" || var_name == "nranks") {
43 44 45 46 47 48 49 50 51 52 53 54 55
      VLOG(10) << "var_name:" << var_name << " need not to transform";
      return expected_kernel_type;
    }

    return framework::OperatorWithKernel::GetKernelTypeForVar(
        var_name, tensor, expected_kernel_type);
  }
};

class DGCMomentumOpMaker : public MomentumOpMaker {
 public:
  void Make() override {
    AddInput("current_step", "(Tensor) Current step.");
56 57 58 59
    AddInput("nranks", "(Tensor) The number of trainers.");

    AddOutput("Grad_out", "(Tensor) Output grad gradient");

60 61 62 63 64 65 66 67 68 69 70 71 72
    AddAttr<float>("rampup_begin_step",
                   "(float, -1.0)"
                   "The period when begin DGC.")
        .SetDefault(-1.0);

    return MomentumOpMaker::Make();
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
73 74
REGISTER_OP_WITHOUT_GRADIENT(dgc_momentum,
                             ops::DGCMomentumOp,
75 76
                             ops::DGCMomentumOpMaker);

L
Leo Chen 已提交
77 78
REGISTER_OP_CPU_KERNEL(dgc_momentum,
                       ops::DGCMomentumKernel<phi::CPUContext, float>);