Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into softmax_cross_support_high_rank
phlrain
6 years ago
commit
766c868199
@ -0,0 +1,154 @@
|
||||
/* Copyright (c) 2019 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/fluid/operators/cvm_op.h"
|
||||
#include <memory>
|
||||
#include "paddle/fluid/operators/math/math_function.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
using Tensor = framework::Tensor;
|
||||
|
||||
class CVMOp : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
void InferShape(framework::InferShapeContext* ctx) const override {
|
||||
PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
|
||||
PADDLE_ENFORCE(ctx->HasInput("CVM"), "Input(CVM) should be not null.");
|
||||
PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) should be not null.");
|
||||
|
||||
auto x_dims = ctx->GetInputDim("X");
|
||||
auto cvm_dims = ctx->GetInputDim("CVM");
|
||||
PADDLE_ENFORCE_EQ(x_dims.size(), 2UL, "Input(X)'s rank should be 2.");
|
||||
PADDLE_ENFORCE_EQ(cvm_dims.size(), 2UL, "Input(CVM)'s rank should be 2.");
|
||||
PADDLE_ENFORCE_EQ(cvm_dims[1], 2UL,
|
||||
"The 2nd dimension of "
|
||||
"Input(CVM) should be 2.");
|
||||
|
||||
if (ctx->Attrs().Get<bool>("use_cvm")) {
|
||||
ctx->SetOutputDim("Y", {x_dims[0], x_dims[1]});
|
||||
} else {
|
||||
ctx->SetOutputDim("Y", {x_dims[0], x_dims[1] - 2});
|
||||
}
|
||||
ctx->ShareLoD("X", /*->*/ "Y");
|
||||
}
|
||||
|
||||
protected:
|
||||
// Explicitly set that the data type of computation kernel of
|
||||
// cvm
|
||||
// is determined by its input "X".
|
||||
framework::OpKernelType GetExpectedKernelType(
|
||||
const framework::ExecutionContext& ctx) const override {
|
||||
return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
|
||||
platform::CPUPlace());
|
||||
}
|
||||
};
|
||||
|
||||
class CVMGradientOp : public framework::OperatorWithKernel {
|
||||
public:
|
||||
using framework::OperatorWithKernel::OperatorWithKernel;
|
||||
|
||||
void InferShape(framework::InferShapeContext* ctx) const override {
|
||||
PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
|
||||
PADDLE_ENFORCE(ctx->HasInput("CVM"), "Input(CVM) should be not null.");
|
||||
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")),
|
||||
"Input(Y@GRAD) should be not null.");
|
||||
PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
|
||||
"Output(X@GRAD) should be not null.");
|
||||
|
||||
auto x_dims = ctx->GetInputDim("X");
|
||||
auto cvm_dims = ctx->GetInputDim("CVM");
|
||||
auto dy_dims = ctx->GetInputDim(framework::GradVarName("Y"));
|
||||
PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
|
||||
PADDLE_ENFORCE_EQ(dy_dims.size(), 2, "Input(Y@Grad)'s rank should be 2.");
|
||||
PADDLE_ENFORCE_EQ(cvm_dims.size(), 2, "Input(CVM)'s rank should be 2.");
|
||||
|
||||
PADDLE_ENFORCE_EQ(x_dims[0], dy_dims[0],
|
||||
"The 1st dimension of Input(X) and Input(Y@Grad) should "
|
||||
"be equal.");
|
||||
|
||||
PADDLE_ENFORCE_EQ(cvm_dims[1], 2,
|
||||
"When Attr(soft_label) == false, the 2nd dimension of "
|
||||
"Input(CVM) should be 2.");
|
||||
ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
|
||||
ctx->ShareLoD("X", framework::GradVarName("X"));
|
||||
}
|
||||
|
||||
protected:
|
||||
// Explicitly set that the data type of computation kernel of
|
||||
// cvm
|
||||
// is determined by its input "X".
|
||||
framework::OpKernelType GetExpectedKernelType(
|
||||
const framework::ExecutionContext& ctx) const override {
|
||||
return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
|
||||
platform::CPUPlace());
|
||||
}
|
||||
};
|
||||
|
||||
class CVMOpMaker : public framework::OpProtoAndCheckerMaker {
|
||||
public:
|
||||
void Make() override {
|
||||
AddInput("X",
|
||||
"(LodTensor, default LodTensor<float>), a 2-D tensor with shape "
|
||||
"[N x D],"
|
||||
" where N is the batch size and D is the emebdding dim. ");
|
||||
AddInput("CVM",
|
||||
"(Tensor), a 2-D Tensor with shape [N x 2], where N is the batch "
|
||||
"size, 2 is show and click.");
|
||||
AddOutput("Y",
|
||||
"(LodTensor, default LodTensor<float>), a 2-D tensor with shape "
|
||||
"[N x K].");
|
||||
AddAttr<bool>("use_cvm", "bool, use cvm or not").SetDefault(true);
|
||||
AddComment(R"DOC(
|
||||
CVM Operator.
|
||||
|
||||
We assume that input X is a embedding vector with cvm_feature(show and click), which shape is [N * D] (D is 2(cvm_feature) + embedding dim, N is batch_size)
|
||||
if use_cvm is True, we will log(cvm_feature), and output shape is [N * D].
|
||||
if use_cvm is False, we will remove cvm_feature from input, and output shape is [N * (D - 2)].
|
||||
|
||||
)DOC");
|
||||
}
|
||||
};
|
||||
|
||||
class CVMGradOpDescMaker : public framework::SingleGradOpDescMaker {
|
||||
public:
|
||||
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
|
||||
|
||||
protected:
|
||||
std::unique_ptr<framework::OpDesc> Apply() const override {
|
||||
std::unique_ptr<framework::OpDesc> op(new framework::OpDesc());
|
||||
op->SetType("cvm_grad");
|
||||
op->SetInput("X", Input("X"));
|
||||
op->SetInput("CVM", Input("CVM"));
|
||||
op->SetInput(framework::GradVarName("Y"), OutputGrad("Y"));
|
||||
op->SetOutput(framework::GradVarName("X"), InputGrad("X"));
|
||||
op->SetAttrMap(Attrs());
|
||||
return op;
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
namespace ops = paddle::operators;
|
||||
REGISTER_OPERATOR(cvm, ops::CVMOp, ops::CVMOpMaker, ops::CVMGradOpDescMaker);
|
||||
|
||||
REGISTER_OPERATOR(cvm_grad, ops::CVMGradientOp);
|
||||
|
||||
REGISTER_OP_CPU_KERNEL(cvm, ops::CVMOpKernel<float>, ops::CVMOpKernel<double>);
|
||||
|
||||
REGISTER_OP_CPU_KERNEL(cvm_grad, ops::CVMGradOpKernel<float>,
|
||||
ops::CVMGradOpKernel<double>);
|
||||
@ -0,0 +1,105 @@
|
||||
/* 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. */
|
||||
|
||||
#pragma once
|
||||
#include "paddle/fluid/framework/eigen.h"
|
||||
#include "paddle/fluid/framework/op_registry.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
using Tensor = framework::Tensor;
|
||||
using LoDTensor = framework::LoDTensor;
|
||||
|
||||
template <typename T>
|
||||
class CVMOpKernel : public framework::OpKernel<T> {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& context) const override {
|
||||
const LoDTensor* x = context.Input<LoDTensor>("X");
|
||||
const T* x_data = x->data<T>();
|
||||
auto lod = x->lod()[0];
|
||||
int64_t item_size = x->numel() / x->dims()[0];
|
||||
int offset = 2;
|
||||
if (!context.Attr<bool>("use_cvm")) {
|
||||
item_size -= offset;
|
||||
}
|
||||
LoDTensor* y = context.Output<LoDTensor>("Y");
|
||||
T* y_data = y->mutable_data<T>(context.GetPlace());
|
||||
|
||||
int seq_num = static_cast<int>(lod.size()) - 1;
|
||||
for (int i = 0; i < seq_num; ++i) {
|
||||
int64_t seq_len = static_cast<int64_t>(lod[i + 1] - lod[i]);
|
||||
|
||||
for (int j = 0; j < seq_len; ++j) {
|
||||
if (context.Attr<bool>("use_cvm")) {
|
||||
std::memcpy(y_data, x_data, item_size * sizeof(T));
|
||||
y_data[0] = log(y_data[0] + 1);
|
||||
y_data[1] = log(y_data[1] + 1) - y_data[0];
|
||||
x_data += item_size;
|
||||
y_data += item_size;
|
||||
} else {
|
||||
std::memcpy(y_data, x_data + offset, item_size * sizeof(T));
|
||||
x_data += item_size + offset;
|
||||
y_data += item_size;
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
template <typename T>
|
||||
class CVMGradOpKernel : public framework::OpKernel<T> {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& context) const override {
|
||||
LoDTensor* dx = context.Output<LoDTensor>(framework::GradVarName("X"));
|
||||
T* dx_data = dx->mutable_data<T>(context.GetPlace());
|
||||
|
||||
const Tensor* cvm = context.Input<Tensor>("CVM");
|
||||
const T* cvm_data = cvm->data<T>();
|
||||
int offset = 2;
|
||||
const framework::LoDTensor* dOut =
|
||||
context.Input<framework::LoDTensor>(framework::GradVarName("Y"));
|
||||
const T* dout_data = dOut->data<T>();
|
||||
|
||||
auto lod = dx->lod()[0];
|
||||
int64_t item_size = dx->numel() / dx->dims()[0];
|
||||
if (!context.Attr<bool>("use_cvm")) {
|
||||
item_size -= offset;
|
||||
}
|
||||
|
||||
int seq_num = static_cast<int>(lod.size()) - 1;
|
||||
for (int i = 0; i < seq_num; ++i) {
|
||||
int64_t seq_len = static_cast<int64_t>(lod[i + 1] - lod[i]);
|
||||
|
||||
for (int j = 0; j < seq_len; ++j) {
|
||||
if (context.Attr<bool>("use_cvm")) {
|
||||
std::memcpy(dx_data, dout_data, item_size * sizeof(T));
|
||||
dx_data[0] = cvm_data[0];
|
||||
dx_data[1] = cvm_data[1];
|
||||
dx_data += item_size;
|
||||
dout_data += item_size;
|
||||
} else {
|
||||
std::memcpy(dx_data + offset, dout_data, item_size * sizeof(T));
|
||||
dx_data[0] = cvm_data[0];
|
||||
dx_data[1] = cvm_data[1];
|
||||
dx_data += item_size + offset;
|
||||
dout_data += item_size;
|
||||
}
|
||||
}
|
||||
cvm_data += offset;
|
||||
}
|
||||
}
|
||||
};
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,27 @@
|
||||
// Copyright (c) 2018 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/distributed/async_sparse_param_update_recorder.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
namespace distributed {
|
||||
|
||||
std::once_flag AsyncSparseParamUpdateRecorder::init_flag_;
|
||||
std::unique_ptr<AsyncSparseParamUpdateRecorder>
|
||||
AsyncSparseParamUpdateRecorder::recorder_(nullptr);
|
||||
|
||||
} // namespace distributed
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,183 @@
|
||||
// Copyright (c) 2018 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.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <functional>
|
||||
#include <future> // NOLINT
|
||||
#include <memory>
|
||||
#include <string>
|
||||
#include <unordered_map>
|
||||
#include <unordered_set>
|
||||
#include <utility>
|
||||
#include <vector>
|
||||
|
||||
#include <ThreadPool.h>
|
||||
|
||||
#include "paddle/fluid/platform/enforce.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
namespace distributed {
|
||||
|
||||
class ConcurrentSet {
|
||||
public:
|
||||
ConcurrentSet() : pool_(new ::ThreadPool(1)) {}
|
||||
~ConcurrentSet() {}
|
||||
|
||||
std::future<void> Update(const std::vector<int64_t>& rows) {
|
||||
auto task = [this, rows] {
|
||||
if (VLOG_IS_ON(3)) {
|
||||
std::ostringstream sstream;
|
||||
sstream << "[";
|
||||
for (auto& id : rows) {
|
||||
sstream << id << ", ";
|
||||
}
|
||||
sstream << "]";
|
||||
VLOG(3) << "update ids -> " << sstream.str();
|
||||
}
|
||||
for (auto row : rows) {
|
||||
set_.insert(row);
|
||||
}
|
||||
};
|
||||
return pool_->enqueue(std::move(task));
|
||||
}
|
||||
|
||||
std::future<void> GetAndClear(std::vector<int64_t>* result) {
|
||||
auto task = [this, &result] {
|
||||
result->clear();
|
||||
for (auto& id : set_) {
|
||||
result->push_back(id);
|
||||
}
|
||||
if (VLOG_IS_ON(3)) {
|
||||
std::ostringstream sstream;
|
||||
sstream << "[";
|
||||
for (auto& id : *result) {
|
||||
sstream << id << ", ";
|
||||
}
|
||||
sstream << "]";
|
||||
VLOG(3) << "result ids size: " << result->size() << " "
|
||||
<< sstream.str();
|
||||
}
|
||||
set_.clear();
|
||||
};
|
||||
return pool_->enqueue(std::move(task));
|
||||
}
|
||||
|
||||
private:
|
||||
std::unordered_set<int64_t> set_;
|
||||
std::unique_ptr<::ThreadPool> pool_{nullptr};
|
||||
};
|
||||
|
||||
class AsyncSparseParamUpdateRecorder {
|
||||
using TrainerToRows = std::vector<std::unique_ptr<ConcurrentSet>>;
|
||||
|
||||
public:
|
||||
AsyncSparseParamUpdateRecorder(
|
||||
int trainer_num,
|
||||
const std::unordered_map<std::string, std::string>& grad_to_param)
|
||||
: trainer_num_(trainer_num), grad_to_param_(grad_to_param) {
|
||||
if (VLOG_IS_ON(3)) {
|
||||
std::ostringstream sstream;
|
||||
sstream << "[";
|
||||
for (auto& item : grad_to_param) {
|
||||
sstream << item.first << ":" << item.second << ", ";
|
||||
}
|
||||
sstream << "]";
|
||||
VLOG(3) << "trainer_num: " << trainer_num
|
||||
<< " grad_to_param_: " << sstream.str();
|
||||
}
|
||||
for (auto& iter : grad_to_param) {
|
||||
param_to_grad_[iter.second] = iter.first;
|
||||
auto& param_name = iter.second;
|
||||
param_to_updated_rows_[param_name] = TrainerToRows();
|
||||
auto& trainer_to_rows = param_to_updated_rows_[param_name];
|
||||
for (auto i = 0; i < trainer_num; ++i) {
|
||||
trainer_to_rows.emplace_back(new ConcurrentSet());
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
~AsyncSparseParamUpdateRecorder() = default;
|
||||
|
||||
void Update(const std::string& grad_name,
|
||||
const std::vector<int64_t>& update_rows) {
|
||||
VLOG(3) << "update grad: " << grad_name
|
||||
<< " row size: " << update_rows.size();
|
||||
auto& param_name = grad_to_param_.at(grad_name);
|
||||
auto& trainer_to_rows = param_to_updated_rows_.at(param_name);
|
||||
|
||||
std::vector<std::future<void>> fs;
|
||||
for (auto& set : trainer_to_rows) {
|
||||
fs.push_back(set->Update(update_rows));
|
||||
}
|
||||
for (auto& f : fs) {
|
||||
f.wait();
|
||||
}
|
||||
}
|
||||
|
||||
void GetAndClear(const std::string& param_name, int trainer_id,
|
||||
std::vector<int64_t>* result) {
|
||||
VLOG(3) << "GetAndClear param: " << param_name
|
||||
<< " for trainer: " << trainer_id;
|
||||
PADDLE_ENFORCE_LT(trainer_id, trainer_num_);
|
||||
param_to_updated_rows_.at(param_name)[trainer_id]
|
||||
->GetAndClear(result)
|
||||
.wait();
|
||||
}
|
||||
|
||||
bool HasParam(const std::string& param_name) {
|
||||
return param_to_grad_.find(param_name) != param_to_grad_.end();
|
||||
}
|
||||
|
||||
bool HasGrad(const std::string& grad_name) {
|
||||
return grad_to_param_.find(grad_name) != grad_to_param_.end();
|
||||
}
|
||||
|
||||
private:
|
||||
const int trainer_num_;
|
||||
std::unordered_map<std::string, std::string> grad_to_param_;
|
||||
std::unordered_map<std::string, std::string> param_to_grad_;
|
||||
std::unordered_map<std::string, TrainerToRows> param_to_updated_rows_;
|
||||
|
||||
// init recorder
|
||||
public:
|
||||
static void Init(
|
||||
int trainer_num,
|
||||
const std::unordered_map<std::string, std::string>& grad_to_param) {
|
||||
InitImpl(trainer_num, grad_to_param);
|
||||
}
|
||||
|
||||
static AsyncSparseParamUpdateRecorder* GetInstance() {
|
||||
return recorder_.get();
|
||||
}
|
||||
|
||||
private:
|
||||
// Init is called by GetInstance.
|
||||
static void InitImpl(
|
||||
int trainer_num,
|
||||
const std::unordered_map<std::string, std::string>& grad_to_param) {
|
||||
if (recorder_ == nullptr) {
|
||||
recorder_.reset(
|
||||
new AsyncSparseParamUpdateRecorder(trainer_num, grad_to_param));
|
||||
}
|
||||
}
|
||||
|
||||
static std::once_flag init_flag_;
|
||||
static std::unique_ptr<AsyncSparseParamUpdateRecorder> recorder_;
|
||||
};
|
||||
|
||||
} // namespace distributed
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue