You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
139 lines
4.9 KiB
139 lines
4.9 KiB
/* Copyright (c) 2016 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/dgc_op.h"
|
|
#include <string>
|
|
#include <vector>
|
|
#include "paddle/fluid/framework/op_registry.h"
|
|
|
|
namespace paddle {
|
|
namespace operators {
|
|
|
|
class DGCOp : public framework::OperatorWithKernel {
|
|
public:
|
|
using framework::OperatorWithKernel::OperatorWithKernel;
|
|
|
|
void InferShape(framework::InferShapeContext* ctx) const override {
|
|
PADDLE_ENFORCE(ctx->HasInput("U"), "Input(U) of DGCop should not be null.");
|
|
PADDLE_ENFORCE(ctx->HasInput("V"), "Input(V) of DGCop should not be null.");
|
|
PADDLE_ENFORCE(ctx->HasInput("Grad"),
|
|
"Input(Grad) of DGCop should not be null.");
|
|
PADDLE_ENFORCE(ctx->HasInput("current_step"),
|
|
"Input(current_step) of DGCop should not be null.");
|
|
|
|
PADDLE_ENFORCE(ctx->HasOutput("U_out"),
|
|
"Output(U_out) of DGCop should not be null.");
|
|
PADDLE_ENFORCE(ctx->HasOutput("V_out"),
|
|
"Output(V_out) of DGCop should not be null.");
|
|
PADDLE_ENFORCE(ctx->HasOutput("k"),
|
|
"Output(k) of DGCop should not be null.");
|
|
PADDLE_ENFORCE(ctx->HasOutput("EncodeGrad"),
|
|
"Output(EncodeGrad) of DGCop should not be null.");
|
|
}
|
|
|
|
protected:
|
|
framework::OpKernelType GetKernelTypeForVar(
|
|
const std::string& var_name, const framework::Tensor& tensor,
|
|
const framework::OpKernelType& expected_kernel_type) const override {
|
|
if (var_name == "current_step" || var_name == "rampup_step" ||
|
|
var_name == "k") {
|
|
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 DGCOpMaker : public framework::OpProtoAndCheckerMaker {
|
|
public:
|
|
void Make() override {
|
|
AddInput("U", "(Tensor) Middle tensor of DGC");
|
|
AddInput("V", "(Tensor) Middle tensor of DGC");
|
|
AddInput("Grad", "(Tensor) Input gradient");
|
|
AddInput("current_step", "(Tensor) Current step.");
|
|
|
|
AddOutput("U_out",
|
|
"(Tensor) "
|
|
"Output encoded gradient");
|
|
AddOutput("V_out",
|
|
"(Tensor) "
|
|
"Output encoded gradient");
|
|
AddOutput("EncodeGrad",
|
|
"(Tensor) "
|
|
"Output encoded gradient");
|
|
AddOutput("Grad_out",
|
|
"(Tensor) "
|
|
"Output grad gradient");
|
|
AddOutput("k",
|
|
"(Tensor) "
|
|
"Output top-k value");
|
|
|
|
AddAttr<float>("m",
|
|
"(float, 0.9) "
|
|
"The momentum of learning rate.")
|
|
.SetDefault(0.9);
|
|
|
|
AddAttr<bool>("use_nesterov",
|
|
"(bool, true)"
|
|
"The momentum of learning rate.")
|
|
.SetDefault(true);
|
|
|
|
AddAttr<std::vector<float>>("sparsity",
|
|
"(vecotr, float)"
|
|
"The period sparsity of k_select.");
|
|
|
|
AddAttr<float>("rampup_begin_step",
|
|
"(float, 0.0)"
|
|
"The period when begin k_select.")
|
|
.SetDefault(0.0);
|
|
|
|
AddAttr<float>("rampup_step",
|
|
"(float, 0.0)"
|
|
"The period when begin k_select.");
|
|
|
|
AddComment(R"DOC(
|
|
Original paper is https://arxiv.org/abs/1712.01887
|
|
|
|
DGC reduce the communication bandwidth by sending only the important gradients (sparse update):\
|
|
only gradients larger than a threshold are transmitted.
|
|
|
|
To avoid losing information, DGC accumulate the rest of the gradients locally.
|
|
|
|
Eventually, these gradients become large enough to be transmitted.
|
|
|
|
Thus, DGC send the large gradients immediately but eventually send all of the gradients over time.
|
|
|
|
To ensure no loss of accuracy, DGC employs momentum correc-tionandlocal gradient clipping on top of the gradient sparsification to maintain model performance.
|
|
|
|
DGC also uses momentum factor masking and warmup training to overcome the staleness problem caused by reduced communication.
|
|
|
|
This optimizer will do two things:
|
|
|
|
1. Compress the gradient by get TopK import value from tensor \
|
|
and use it for allreduce to reduce network bandwidth.
|
|
|
|
2. Call momentum to optimize on the cost.
|
|
|
|
)DOC");
|
|
}
|
|
};
|
|
|
|
} // namespace operators
|
|
} // namespace paddle
|
|
|
|
namespace ops = paddle::operators;
|
|
REGISTER_OP_WITHOUT_GRADIENT(dgc, ops::DGCOp, ops::DGCOpMaker);
|