提交 057694aa 编写于 作者: W wanghaoshuang

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into dilate_conv

......@@ -18,8 +18,8 @@ cc_test(scope_test SRCS scope_test.cc DEPS scope)
proto_library(framework_proto SRCS framework.proto)
cc_library(attribute SRCS attribute.cc DEPS framework_proto)
cc_library(operator SRCS operator.cc DEPS framework_proto device_context tensor scope attribute)
cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
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)
cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS operator)
......@@ -56,5 +56,6 @@ cc_library(paddle_pybind SHARED
recurrent_op
uniform_random_op
gaussian_random_op
fill_zeros_like_op)
fill_zeros_like_op
scale_op)
endif(WITH_PYTHON)
......@@ -72,8 +72,8 @@ class NoGradOpMaker : public OpProtoAndCheckerMaker {
class FcOp : public operators::NetOp {
public:
FcOp(const std::string &type, const VarNameMap &inputs,
const VarNameMap &outputs, const AttributeMap &attrs)
FcOp(const std::string &type, const VariableNameMap &inputs,
const VariableNameMap &outputs, const AttributeMap &attrs)
: NetOp(type, inputs, outputs, attrs) {
AppendOp(OpRegistry::CreateOp("mul",
{{"X", {Input("X")}}, {"Y", {Input("W")}}},
......
......@@ -20,13 +20,13 @@ namespace framework {
enum class OpArgType { IN, OUT };
static void TransOpArg(const OperatorBase* src_op, const OpArgType& src_type,
bool is_grad, OperatorBase::VarNameMap* vars) {
bool is_grad, VariableNameMap* vars) {
const auto& src_inout =
src_type == OpArgType::IN ? src_op->Inputs() : src_op->Outputs();
auto& dst_inout = *vars;
const OpProto* proto = OpRegistry::op_info_map().at(src_op->Type()).proto_;
auto& proto = OpInfoMap::Instance().Get(src_op->Type()).Proto();
const auto& src_arg_list =
src_type == OpArgType::IN ? proto->inputs() : proto->outputs();
src_type == OpArgType::IN ? proto.inputs() : proto.outputs();
for (const auto& arg : src_arg_list) {
if (arg.not_in_gradient() && !is_grad) continue;
const std::string src_name = arg.name();
......@@ -40,26 +40,18 @@ static void TransOpArg(const OperatorBase* src_op, const OpArgType& src_type,
}
OperatorBase* BuildGradOp(const OperatorBase* op) {
auto it = OpRegistry::op_info_map().find(op->Type());
PADDLE_ENFORCE(it != OpRegistry::op_info_map().end(),
"'%s' has not been registered.", op->Type());
PADDLE_ENFORCE(it->second.proto_ != nullptr, "'%s' has no OpProto.",
op->Type());
std::string grad_op_type = it->second.grad_op_type_;
PADDLE_ENFORCE(!grad_op_type.empty(), "'%s' has no gradient operator.",
op->Type());
auto& info = OpInfoMap::Instance().Get(op->Type());
PADDLE_ENFORCE(info.HasGradientOp());
OperatorBase::VarNameMap inputs;
OperatorBase::VarNameMap outputs;
VariableNameMap inputs;
VariableNameMap outputs;
TransOpArg(op, OpArgType::IN, false, &inputs); // I
TransOpArg(op, OpArgType::OUT, false, &inputs); // O
TransOpArg(op, OpArgType::OUT, true, &inputs); // OG
TransOpArg(op, OpArgType::IN, true, &outputs); // IG
it = OpRegistry::op_info_map().find(grad_op_type);
PADDLE_ENFORCE(it != OpRegistry::op_info_map().end(),
"'%s' has not been registered.", grad_op_type);
return it->second.creator_(grad_op_type, inputs, outputs, op->Attrs());
auto& grad_info = OpInfoMap::Instance().Get(info.grad_op_type_);
return grad_info.Creator()(info.grad_op_type_, inputs, outputs, op->Attrs());
}
} // namespace framework
......
/* 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/framework/op_info.h"
namespace paddle {
namespace framework {
static OpInfoMap* g_op_info_map = nullptr;
OpInfoMap& OpInfoMap::Instance() {
if (g_op_info_map == nullptr) {
g_op_info_map = new OpInfoMap();
}
return *g_op_info_map;
}
} // namespace framework
} // namespace paddle
/* 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 <functional>
#include <map>
#include <string>
#include <unordered_map>
#include "paddle/framework/attribute.h"
namespace paddle {
namespace framework {
class OperatorBase;
using VariableNameMap = std::map<std::string, std::vector<std::string>>;
using OpCreator = std::function<OperatorBase*(
const std::string& /*type*/, const VariableNameMap& /*inputs*/,
const VariableNameMap& /*outputs*/, const AttributeMap& /*attrs*/)>;
struct OpInfo {
OpCreator creator_;
std::string grad_op_type_;
OpProto* proto_;
OpAttrChecker* checker_;
bool HasOpProtoAndChecker() const {
return proto_ != nullptr && checker_ != nullptr;
}
const OpProto& Proto() const {
PADDLE_ENFORCE_NOT_NULL(proto_, "Operator Proto has not been registered");
PADDLE_ENFORCE(proto_->IsInitialized(),
"Operator Proto must be initialized in op info");
return *proto_;
}
const OpAttrChecker& Checker() const {
PADDLE_ENFORCE_NOT_NULL(checker_,
"Operator Checker has not been registered");
return *checker_;
}
const OpCreator& Creator() const {
PADDLE_ENFORCE_NOT_NULL(creator_,
"Operator Creator has not been registered");
return creator_;
}
bool HasGradientOp() const { return !grad_op_type_.empty(); }
};
class OpInfoMap {
public:
static OpInfoMap& Instance();
OpInfoMap(const OpInfoMap& o) = delete;
OpInfoMap(OpInfoMap&& o) = delete;
OpInfoMap& operator=(const OpInfoMap& o) = delete;
OpInfoMap& operator=(OpInfoMap&& o) = delete;
bool Has(const std::string& op_type) const {
return map_.find(op_type) != map_.end();
}
void Insert(const std::string& type, const OpInfo& info) {
PADDLE_ENFORCE(!Has(type), "Operator %s has been registered", type);
map_.insert({type, info});
}
const OpInfo& Get(const std::string& type) const {
auto it = map_.find(type);
PADDLE_ENFORCE(it != map_.end(), "Operator %s are not found", type);
return it->second;
}
template <typename Callback>
void IterAllInfo(Callback callback) {
for (auto& it : map_) {
callback(it.first, it.second);
}
}
private:
OpInfoMap() = default;
std::unordered_map<std::string, const OpInfo> map_;
};
} // namespace framework
} // namespace paddle
......@@ -19,32 +19,18 @@ limitations under the License. */
namespace paddle {
namespace framework {
std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const std::string& type,
const VarNameMap& inputs,
const VarNameMap& outputs,
AttributeMap attrs) {
auto it = op_info_map().find(type);
PADDLE_ENFORCE(it != op_info_map().end(),
"Operator '%s' has not been registered.", type);
it->second.checker_->Check(attrs);
auto op = it->second.creator_(type, inputs, outputs, attrs);
std::unique_ptr<OperatorBase> OpRegistry::CreateOp(
const std::string& type, const VariableNameMap& inputs,
const VariableNameMap& outputs, AttributeMap attrs) {
auto& info = OpInfoMap::Instance().Get(type);
info.Checker().Check(attrs);
auto op = info.Creator()(type, inputs, outputs, attrs);
return std::unique_ptr<OperatorBase>(op);
}
std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
VarNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
VarNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
AttributeMap attrs;
for (auto& attr : op_desc.attrs()) {
attrs[attr.name()] = GetAttrValue(attr);
}
return CreateOp(op_desc.type(), inputs, outputs, attrs);
}
OperatorBase::VarNameMap OpRegistry::ConvertOpDescVarsToVarNameMap(
static VariableNameMap ConvertOpDescVarsToVarNameMap(
const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars) {
VarNameMap ret_val;
VariableNameMap ret_val;
for (auto& var : op_desc_vars) {
auto& var_names = ret_val[var.parameter()];
auto& var_names_in_proto = var.arguments();
......@@ -55,6 +41,17 @@ OperatorBase::VarNameMap OpRegistry::ConvertOpDescVarsToVarNameMap(
return ret_val;
}
std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
VariableNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
VariableNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
AttributeMap attrs;
for (auto& attr : op_desc.attrs()) {
attrs[attr.name()] = GetAttrValue(attr);
}
return CreateOp(op_desc.type(), inputs, outputs, attrs);
}
std::unique_ptr<OperatorBase> OpRegistry::CreateGradOp(const OperatorBase& op) {
PADDLE_ENFORCE(!op.IsNetOp(), "Use framework::Backward to get backward ops");
return std::unique_ptr<OperatorBase>(BuildGradOp(&op));
......
......@@ -23,6 +23,7 @@ limitations under the License. */
#include "paddle/framework/attribute.h"
#include "paddle/framework/framework.pb.h"
#include "paddle/framework/grad_op_builder.h"
#include "paddle/framework/op_info.h"
#include "paddle/framework/operator.h"
#include "paddle/framework/scope.h"
......@@ -30,28 +31,16 @@ namespace paddle {
namespace framework {
class OpRegistry {
using VarNameMap = OperatorBase::VarNameMap;
using OpCreator = std::function<OperatorBase*(
const std::string& /*type*/, const VarNameMap& /*inputs*/,
const VarNameMap& /*outputs*/, const AttributeMap& /*attrs*/)>;
public:
struct OpInfo {
OpCreator creator_;
std::string grad_op_type_;
OpProto* proto_;
OpAttrChecker* checker_;
};
template <typename OpType, typename ProtoMakerType, typename GradOpType>
static void RegisterOp(const std::string& op_type,
const std::string& grad_op_type) {
PADDLE_ENFORCE(op_info_map().count(op_type) == 0,
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 VarNameMap& inputs,
const VarNameMap& outputs,
const AttributeMap& attrs) {
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;
......@@ -70,7 +59,7 @@ class OpRegistry {
op_info.proto_ = nullptr;
op_info.checker_ = nullptr;
}
op_info_map().insert(std::make_pair(op_type, op_info));
OpInfoMap::Instance().Insert(op_type, op_info);
// register gradient op
if (!grad_op_type.empty()) {
RegisterOp<GradOpType, NOPMaker, NOP>(grad_op_type, "");
......@@ -78,21 +67,13 @@ class OpRegistry {
}
static std::unique_ptr<OperatorBase> CreateOp(const std::string& type,
const VarNameMap& inputs,
const VarNameMap& outputs,
const VariableNameMap& inputs,
const VariableNameMap& outputs,
AttributeMap attrs);
static std::unique_ptr<OperatorBase> CreateOp(const OpDesc& op_desc);
static VarNameMap ConvertOpDescVarsToVarNameMap(
const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars);
static std::unique_ptr<OperatorBase> CreateGradOp(const OperatorBase& op);
static std::unordered_map<std::string, const OpInfo>& op_info_map() {
static std::unordered_map<std::string, const OpInfo> op_info_map_;
return op_info_map_;
}
};
class Registrar {
......
......@@ -115,8 +115,8 @@ void OperatorBase::Rename(const std::string& old_name,
}
OperatorBase::OperatorBase(const std::string& type,
const OperatorBase::VarNameMap& inputs,
const OperatorBase::VarNameMap& outputs,
const VariableNameMap& inputs,
const VariableNameMap& outputs,
const AttributeMap& attrs)
: type_(type), inputs_(inputs), outputs_(outputs), attrs_(attrs) {
static std::atomic<size_t> gUniqId(0UL);
......@@ -141,18 +141,10 @@ std::vector<std::string> OperatorBase::OutputVars(bool has_intermediate) const {
}
return ret_val;
}
auto it = OpRegistry::op_info_map().find(type_);
PADDLE_ENFORCE(
it != OpRegistry::op_info_map().end(),
"Operator %s not registered, cannot figure out intermediate outputs",
type_);
PADDLE_ENFORCE(
it->second.proto_ != nullptr,
"Operator %s has no OpProto, cannot figure out intermediate outputs",
type_);
auto& info = OpInfoMap::Instance().Get(Type());
// get all OpProto::Var for outputs
for (auto& o : it->second.proto_->outputs()) {
for (auto& o : info.Proto().outputs()) {
// ignore all intermediate output
if (o.intermediate()) continue;
auto out = outputs_.find(o.name());
......
......@@ -19,6 +19,7 @@ limitations under the License. */
#include <unordered_map>
#include <vector>
#include "op_info.h"
#include "paddle/framework/attribute.h"
#include "paddle/framework/framework.pb.h"
#include "paddle/framework/scope.h"
......@@ -62,10 +63,8 @@ class ExecutionContext;
*/
class OperatorBase {
public:
using VarNameMap = std::map<std::string, std::vector<std::string>>;
OperatorBase(const std::string& type, const VarNameMap& inputs,
const VarNameMap& outputs, const AttributeMap& attrs);
OperatorBase(const std::string& type, const VariableNameMap& inputs,
const VariableNameMap& outputs, const AttributeMap& attrs);
virtual ~OperatorBase() {}
......@@ -93,8 +92,8 @@ class OperatorBase {
/// rename inputs outputs name
void Rename(const std::string& old_name, const std::string& new_name);
const VarNameMap& Inputs() const { return inputs_; }
const VarNameMap& Outputs() const { return outputs_; }
const VariableNameMap& Inputs() const { return inputs_; }
const VariableNameMap& Outputs() const { return outputs_; }
//! Get a input with argument's name described in `op_proto`
const std::string& Input(const std::string& name) const;
//! Get a input which has multiple variables.
......@@ -122,30 +121,32 @@ class OperatorBase {
// I (Inputs)opear
// O (Outputs)
// OG (Output Gradients)
VarNameMap inputs_;
VariableNameMap inputs_;
// NOTE: in case of OpGrad, outputs_ contains
// IG (Inputs Gradients)
VarNameMap outputs_;
VariableNameMap outputs_;
AttributeMap attrs_;
};
// Macro for define a clone method.
// If you are writing an kernel operator, `Clone` will be defined when you
// register it. i.e. `Clone` method is not needed to define by yourself.
#define DEFINE_OP_CLONE_METHOD(CLS) \
#define DEFINE_OP_CLONE_METHOD(cls) \
std::unique_ptr<OperatorBase> Clone() const final { \
return std::unique_ptr<OperatorBase>(new CLS(*this)); \
return std::unique_ptr<OperatorBase>(new cls(*this)); \
}
// Macro for define a default constructor for Operator.
// You can also use
// using PARENT_CLASS::PARENT_CLASS;
// to use parent's constructor.
#define DEFINE_OP_CONSTRUCTOR(CLS, PARENT_CLS) \
CLS(const std::string& type, const VarNameMap& inputs, \
const VarNameMap& outputs, const paddle::framework::AttributeMap& attrs) \
: PARENT_CLS(type, inputs, outputs, attrs) {}
#define DEFINE_OP_CONSTRUCTOR(cls, parent_cls) \
cls(const std::string& type, \
const ::paddle::framework::VariableNameMap& inputs, \
const ::paddle::framework::VariableNameMap& outputs, \
const paddle::framework::AttributeMap& attrs) \
: parent_cls(type, inputs, outputs, attrs) {}
class NOP : public OperatorBase {
public:
......@@ -389,8 +390,8 @@ class OperatorWithKernel : public OperatorBase {
using OpKernelMap =
std::unordered_map<OpKernelKey, std::unique_ptr<OpKernel>, OpKernelHash>;
OperatorWithKernel(const std::string& type, const VarNameMap& inputs,
const VarNameMap& outputs, const AttributeMap& attrs)
OperatorWithKernel(const std::string& type, const VariableNameMap& inputs,
const VariableNameMap& outputs, const AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {}
void InferShape(const Scope& scope) const override {
......
......@@ -23,8 +23,8 @@ static int op_run_num = 0;
class OpWithoutKernelTest : public OperatorBase {
public:
OpWithoutKernelTest(const std::string& type, const VarNameMap& inputs,
const VarNameMap& outputs, const AttributeMap& attrs)
OpWithoutKernelTest(const std::string& type, const VariableNameMap& inputs,
const VariableNameMap& outputs, const AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs), x(1) {}
void InferShape(const Scope& scope) const override {}
void Run(const Scope& scope,
......@@ -249,8 +249,9 @@ TEST(OpKernel, multi_inputs) {
class OperatorClone : public paddle::framework::OperatorBase {
public:
DEFINE_OP_CLONE_METHOD(OperatorClone);
OperatorClone(const std::string& type, const VarNameMap& inputs,
const VarNameMap& outputs,
OperatorClone(const std::string& type,
const paddle::framework::VariableNameMap& inputs,
const paddle::framework::VariableNameMap& outputs,
const paddle::framework::AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {}
void InferShape(const paddle::framework::Scope& scope) const override {}
......
......@@ -42,6 +42,8 @@ USE_OP(fill_zeros_like);
USE_OP_ITSELF(recurrent_op);
USE_OP(gaussian_random);
USE_OP(uniform_random);
USE_OP(scale);
USE_OP_ITSELF(identity);
USE_CPU_ONLY_OP(gather);
namespace paddle {
......@@ -139,19 +141,16 @@ All parameter, weight, gradient are variables in Paddle.
//! @note: Be careful! PyBind will return std::string as an unicode, not
//! Python str. If you want a str object, you should cast them in Python.
m.def("get_all_op_protos", []() -> std::vector<py::bytes> {
auto &op_info_map = OpRegistry::op_info_map();
std::vector<py::bytes> ret_values;
for (auto it = op_info_map.begin(); it != op_info_map.end(); ++it) {
const OpProto *proto = it->second.proto_;
if (proto == nullptr) {
continue;
}
PADDLE_ENFORCE(proto->IsInitialized(), "OpProto must all be initialized");
OpInfoMap::Instance().IterAllInfo([&ret_values](const std::string &type,
const OpInfo &info) {
if (!info.HasOpProtoAndChecker()) return;
std::string str;
PADDLE_ENFORCE(proto->SerializeToString(&str),
PADDLE_ENFORCE(info.Proto().SerializeToString(&str),
"Serialize OpProto Error. This could be a bug of Paddle.");
ret_values.push_back(py::bytes(str));
}
ret_values.emplace_back(str);
});
return ret_values;
});
m.def_submodule(
......
......@@ -105,7 +105,10 @@ class Tensor {
template <typename T>
inline Tensor Slice(const int& begin_idx, const int& end_idx) const;
platform::Place place() const { return holder_->place(); }
platform::Place place() const {
PADDLE_ENFORCE_NOT_NULL(holder_, "Tensor get place() must contains holder");
return holder_->place();
}
private:
template <typename T>
......
......@@ -1012,11 +1012,6 @@ void RecurrentGradientMachine::generateSequence() {
/* width */ resultNum,
false,
/* useGpu */ false);
Matrix::resizeOrCreate(generator_.outArg.value,
/* height */ maxGenWordCount,
/* width */ 1,
false,
/* useGpu */ false);
}
ICpuGpuVector::resizeOrCreate(generator_.outArg.sequenceStartPositions,
numSequences + 1,
......@@ -1026,7 +1021,7 @@ void RecurrentGradientMachine::generateSequence() {
} else {
oneWaySearch(numSequences);
}
if (dataArgsSize_) createDataOutlink(batchMachineIdVec_);
if (dataArgsSize_) createDataOutlink();
size_t size = generator_.ids.size();
generator_.outArg.ids->resize(size);
......@@ -1106,6 +1101,7 @@ void RecurrentGradientMachine::oneWaySearch(size_t batchSize) {
}
batchMachineIdVec_.clear();
batchMachineStartPos_.clear();
int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false);
starts[0] = 0;
generator_.ids.clear();
......@@ -1312,13 +1308,20 @@ void RecurrentGradientMachine::fillGenOutputs() {
finalPaths_[i].resize(minFinalPathsSize);
}
batchMachineIdVec_.clear();
generator_.ids.clear();
int* starts = generator_.outArg.sequenceStartPositions->getMutableData(false);
starts[0] = 0;
if (numResults > 1) {
real* probs = generator_.outArg.in->getData();
int idsProbSaveSize = 0;
for (auto inSeq : finalPaths_) {
for (auto path : inSeq) idsProbSaveSize += path.ids.size();
idsProbSaveSize += inSeq.size();
}
Matrix::resizeOrCreate(
generator_.outArg.value, idsProbSaveSize, 1, false, false);
real* idsProb = generator_.outArg.value->getData();
real* probs = generator_.outArg.in->getData();
size_t curPos = 0;
for (size_t i = 0; i < finalPaths_.size(); ++i) {
for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
......@@ -1333,24 +1336,16 @@ void RecurrentGradientMachine::fillGenOutputs() {
curPos += genLen;
idsProb[curPos++] = -1.0;
probs[i * numResults + j] = path.logProb;
if (!j && dataArgsSize_) {
// in beam search, here only reserved the top 1 generated result
// for out_links that are not the generated word indices.
batchMachineIdVec_.insert(batchMachineIdVec_.end(),
path.machineIdVec.begin(),
path.machineIdVec.end());
}
}
starts[i + 1] = generator_.ids.size();
}
} else {
for (size_t i = 0; i < finalPaths_.size(); ++i) {
CHECK(!finalPaths_[i].empty());
generator_.ids.insert(generator_.ids.begin(),
finalPaths_[i][0].ids.begin(),
finalPaths_[i][0].ids.end());
starts[i + 1] = starts[i] + finalPaths_[i][0].ids.size();
Path& path = finalPaths_[i][0];
generator_.ids.insert(
generator_.ids.begin(), path.ids.begin(), path.ids.end());
starts[i + 1] = starts[i] + path.ids.size();
}
}
}
......@@ -1364,25 +1359,76 @@ void RecurrentGradientMachine::copyDataOutlinkFrame(size_t machineCur) {
}
}
void RecurrentGradientMachine::createDataOutlink(
std::vector<int>& machineIdVec) {
size_t seqNum =
getBeamSize() > 1UL ? finalPaths_.size() : finalPaths_[0].size();
std::vector<int> starts(seqNum + 1, 0);
for (size_t i = 0; i < seqNum; ++i) {
size_t seqLen = getBeamSize() > 1UL ? finalPaths_[i][0].ids.size()
: finalPaths_[0][i].ids.size();
starts[i + 1] = starts[i] + seqLen;
void RecurrentGradientMachine::createDataOutlinkSelRowsInfo(
bool isSeq, std::vector<Argument>& outArgs) {
batchMachineIdVec_.clear();
size_t seqIdx = 0;
for (size_t i = 0; i < finalPaths_.size(); ++i) {
for (size_t j = 0; j < finalPaths_[i].size(); ++j) {
std::vector<int>& machineIdVec = finalPaths_[i][j].machineIdVec;
if (isSeq) {
for (size_t i = 0; i < machineIdVec.size(); ++i) {
size_t rowId = machineIdVec[i];
int* seqPos =
outArgs[i].sequenceStartPositions->getMutableData(false);
batchMachineIdVec_.push_back(seqPos[rowId]);
}
} else {
batchMachineIdVec_.insert(
batchMachineIdVec_.end(), machineIdVec.begin(), machineIdVec.end());
}
seqIdx++;
}
}
}
void RecurrentGradientMachine::createDataOutlinkCopySizeInfo(
bool isSeq, std::vector<Argument>& outArgs, std::vector<int>& copySize) {
size_t totalSeqNum = std::accumulate(
finalPaths_.begin(),
finalPaths_.end(),
0UL,
[](size_t a, const std::vector<Path>& b) { return a + b.size(); });
copySize.resize(totalSeqNum, 1);
batchMachineStartPos_.resize(totalSeqNum + 1, 0);
if (isSeq) {
ICpuGpuVectorPtr inputSeqStartPos = outArgs[0].sequenceStartPositions;
CHECK_EQ(static_cast<size_t>(inputSeqStartPos->getSize() - 1),
getBeamSize() > 1 ? finalPaths_.size() : finalPaths_[0].size());
int* starts = inputSeqStartPos->getMutableData(false);
int seqId = 0;
for (int i = 0; i < finalPaths_.size(); ++i) {
for (int j = 0; j < finalPaths_[i].size(); ++j) {
copySize[seqId] = getBeamSize() > 1 ? starts[i + 1] - starts[i]
: starts[j + 1] - starts[j];
batchMachineStartPos_[seqId + 1] =
batchMachineStartPos_[seqId] + finalPaths_[i][j].ids.size();
seqId++;
}
}
} else {
for (size_t i = 0; i < finalPaths_[0].size(); ++i)
batchMachineStartPos_[i + 1] =
batchMachineStartPos_[i] + finalPaths_[0][i].ids.size();
}
}
void RecurrentGradientMachine::createDataOutlink() {
for (size_t i = 0; i < dataArgsSize_; i++) {
bool isSeq = dataArgsFrame_[i][0].hasSeq();
std::vector<int> copySize;
createDataOutlinkCopySizeInfo(isSeq, dataArgsFrame_[i], copySize);
createDataOutlinkSelRowsInfo(isSeq, dataArgsFrame_[i]);
dataArgs_[i].concat(dataArgsFrame_[i],
machineIdVec,
starts,
batchMachineIdVec_,
batchMachineStartPos_,
copySize,
useGpu_,
HPPL_STREAM_1,
PASS_TEST);
auto dataAgent =
dynamic_cast<DataLayer*>(outFrameLines_[i + 1].agentLayer.get());
CHECK_NOTNULL(dataAgent);
......
......@@ -190,7 +190,7 @@ public:
std::vector<int> ids;
/**
* @brief idsProb, log probability of each generated words.
* @brief idsProb, log probability of each generated word.
*/
std::vector<real> idsProb;
......@@ -472,15 +472,43 @@ private:
void copyDataOutlinkFrame(size_t machineCur);
/*
* @brief In generation, if the layer group has more than 1 outlink, outlinks
* except the first one are data outlinks. This function creates the data
* outlinks.
* @note In beam search, only one generated sequence with the hightest log
* probabilites are retained.
* @param machineIdVec : select a row of output matrix in each frame
* that the generation process expanded.
* @brief In generation, if the layer group has more than 1 outlink, outlink
* except the first one is a data outlink. In RecurrentLayerGroup, each time
* step is a separate Network, outputs of a layer inside the
* RecurrentLayerGroup are stored in separate Arguments. If one layer is
* specified as an outlink of RecurrentLayerGroup. This function will
* collect outputs in each time step of each generated sequence which are
* dispersed in separate Arguments to form a new single Argument as output of
* RecurrentLayerGroup.
*/
void createDataOutlink(std::vector<int>& machineIdVec);
void createDataOutlink();
/*
* @brief decide to select how many rows from the Matrix stored the forward
* pass results from a start position.
*
* @param isSeq: a flag indicating whetehr the layer to be output of the
* RecurrentGradientMachine is a sequence or not
* @param outArgs: all of the the returned Arguments of the forward pass
* during the generation process.
* @param copySize: the returned result, number of rows to select from the
* Matrix stored the forward pass results from a start position.
*/
void createDataOutlinkCopySizeInfo(bool isSeq,
std::vector<Argument>& outArgs,
std::vector<int>& copySize);
/*
* @brief decide index of the start row for each time step of a generated
* sequence in Matrix stored the entire beam search batch's forward pass
* results.
*
* @param isSeq: a flag indicating whether the layer to be output of the
* RecurrentGradientMachine is a sequence or not
* @param outArgs: all of the returned Arguments of the forward pass
* during the generation process.
*/
void createDataOutlinkSelRowsInfo(bool isSeq, std::vector<Argument>& outArgs);
/*
* @brief used in beam search, connect previous frame to form recurrent link
......@@ -543,6 +571,7 @@ private:
std::vector<int> topIds_;
std::vector<int> seqIds_;
std::vector<int> batchMachineIdVec_;
std::vector<int> batchMachineStartPos_;
std::vector<std::vector<Path>> finalPaths_;
std::vector<real> minFinalPathLogProb_;
BeamSearchControlCallbacks* beamSearchCtrlCallbacks_;
......
......@@ -69,3 +69,4 @@ op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
DEPS framework_proto tensor op_registry operator net_op)
op_library(uniform_random_op
SRCS uniform_random_op.cc uniform_random_op.cu)
op_library(scale_op SRCS scale_op.cc scale_op.cu DEPS net_op)
......@@ -68,10 +68,15 @@ std::string NetOp::DebugString() const {
bool NetOp::IsNetOp() const { return true; }
std::vector<std::string> NetOp::OutputVars(bool has_intermediate) const {
std::vector<std::string> all;
for (auto& pair : this->outputs_) {
for (auto& var_name : pair.second) {
all.push_back(var_name);
}
}
if (has_intermediate) {
return this->outputs_.at(kAll);
return all;
}
auto& all = this->outputs_.at(kAll);
std::vector<std::string> ret_val;
for (auto& each : all) {
if (!Contains(intermediate_outputs_, each)) {
......@@ -81,9 +86,8 @@ std::vector<std::string> NetOp::OutputVars(bool has_intermediate) const {
return ret_val;
}
NetOp::NetOp(const std::string& type,
const framework::OperatorBase::VarNameMap& inputs,
const framework::OperatorBase::VarNameMap& outputs,
NetOp::NetOp(const std::string& type, const framework::VariableNameMap& inputs,
const framework::VariableNameMap& outputs,
const framework::AttributeMap& attrs)
: framework::OperatorBase(type, inputs, outputs, attrs) {}
......
......@@ -38,8 +38,10 @@ class NetOp : public framework::OperatorBase {
public:
static const char kAll[];
NetOp() : framework::OperatorBase("plain_net", {}, {}, {}) {}
NetOp(const std::string& type, const VarNameMap& inputs,
const VarNameMap& outputs, const framework::AttributeMap& attrs);
NetOp(const std::string& type, const framework::VariableNameMap& inputs,
const framework::VariableNameMap& outputs,
const framework::AttributeMap& attrs);
NetOp(const NetOp& o) : framework::OperatorBase(o.type_, {}, {}, o.attrs_) {
this->ops_.reserve(o.ops_.size());
......
......@@ -131,8 +131,8 @@ const rnn::ArgumentName RecurrentGradientOp::kArgName{
"memories", "pre_memories", "boot_memories@grad"};
RecurrentOp::RecurrentOp(const std::string& type,
const framework::OperatorBase::VarNameMap& inputs,
const framework::OperatorBase::VarNameMap& outputs,
const framework::VariableNameMap& inputs,
const framework::VariableNameMap& outputs,
const framework::AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {
rnn::InitArgument(kArgName, &arg_, *this);
......@@ -223,8 +223,8 @@ void RecurrentGradientAlgorithm::InferShape(const Scope& scope) const {
}
RecurrentGradientOp::RecurrentGradientOp(
const std::string& type, const framework::OperatorBase::VarNameMap& inputs,
const framework::OperatorBase::VarNameMap& outputs,
const std::string& type, const framework::VariableNameMap& inputs,
const framework::VariableNameMap& outputs,
const framework::AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {
rnn::InitArgument(kArgName, &arg_, *this);
......
......@@ -114,8 +114,9 @@ class RecurrentGradientAlgorithm {
class RecurrentOp : public framework::OperatorBase {
public:
RecurrentOp(const std::string& type, const VarNameMap& inputs,
const VarNameMap& outputs, const framework::AttributeMap& attrs);
RecurrentOp(const std::string& type, const framework::VariableNameMap& inputs,
const framework::VariableNameMap& outputs,
const framework::AttributeMap& attrs);
RecurrentOp(const RecurrentOp& o)
: framework::OperatorBase(
......@@ -150,8 +151,9 @@ class RecurrentOp : public framework::OperatorBase {
class RecurrentGradientOp : public framework::OperatorBase {
public:
RecurrentGradientOp(const std::string& type, const VarNameMap& inputs,
const VarNameMap& outputs,
RecurrentGradientOp(const std::string& type,
const framework::VariableNameMap& inputs,
const framework::VariableNameMap& outputs,
const framework::AttributeMap& attrs);
RecurrentGradientOp(const RecurrentGradientOp& o)
......
/* 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/scale_op.h"
#include "paddle/operators/net_op.h"
namespace paddle {
namespace operators {
class ScaleOp : public framework::OperatorWithKernel {
public:
ScaleOp(const std::string &type, const framework::VariableNameMap &inputs,
const framework::VariableNameMap &outputs,
const framework::AttributeMap &attrs)
: OperatorWithKernel(type, inputs, outputs, attrs) {}
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
auto *in = ctx.Input<framework::Tensor>("X");
auto *out = ctx.Output<framework::Tensor>("Out");
out->Resize(in->dims());
}
};
template <typename AttrType>
class ScaleOpMaker : public framework::OpProtoAndCheckerMaker {
public:
ScaleOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "The input tensor of scale operator.").NotInGradient();
AddOutput("Out", "The output tensor of scale operator.").NotInGradient();
AddComment(R"DOC(Scale operator
The equation is: Out = scale*X
)DOC");
AddAttr<AttrType>("scale", "scale of scale operator.").SetDefault(1.0);
}
};
// Identity Op's gradient is identity op, too.
// Grad(Out=scale(X)) => Grad(X) = scale(Grad(Out))
template <typename AttrType>
class ScaleGradOp : public NetOp {
public:
ScaleGradOp(const std::string &type, const framework::VariableNameMap &inputs,
const framework::VariableNameMap &outputs,
const framework::AttributeMap &attrs)
: NetOp(type, inputs, outputs, attrs) {
AppendOp(framework::OpRegistry::CreateOp(
"scale", {{"X", {Input(framework::GradVarName("Out"))}}},
{{"Out", {Output(framework::GradVarName("X"))}}},
{{"scale", GetAttr<AttrType>("scale")}}));
CompleteAddOp(false);
}
};
// identity is a alias of scale op. This is also a example for creating a alias
// operator.
template <typename AttrType>
class IdentityOpMaker : public framework::OpProtoAndCheckerMaker {
public:
IdentityOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "input tensor of identity op");
AddOutput("Out", "output tensor of identity op");
AddComment("identity operator. Just a alias of scale op which scale = 1.0");
}
};
template <typename AttrType>
class IdentityOp : public NetOp {
public:
IdentityOp(const std::string &type, const framework::VariableNameMap &inputs,
const framework::VariableNameMap &outputs,
const framework::AttributeMap &attrs)
: NetOp(type, inputs, outputs, attrs) {
AppendOp(framework::OpRegistry::CreateOp(
"scale", {{"X", {Input("X")}}}, {{"Out", {Output("Out")}}},
{{"scale", static_cast<AttrType>(1)}}));
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(scale, ops::ScaleOp, ops::ScaleOpMaker<float>, scale_grad,
ops::ScaleGradOp<float>);
REGISTER_OP_CPU_KERNEL(scale,
ops::ScaleKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_WITHOUT_GRADIENT(identity, ops::IdentityOp<float>,
ops::IdentityOpMaker<float>);
/* 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/scale_op.h"
REGISTER_OP_GPU_KERNEL(
scale, paddle::operators::ScaleKernel<paddle::platform::GPUPlace, float>);
/* 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/eigen.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
template <typename Place, typename T, typename AttrType = T>
class ScaleKernel : public framework::OpKernel {
public:
virtual void Compute(const framework::ExecutionContext& context) const {
auto* tensor = context.Output<framework::Tensor>("Out");
auto* in = context.Input<framework::Tensor>("X");
tensor->mutable_data<T>(in->place());
auto scale = static_cast<T>(context.op_.GetAttr<AttrType>("scale"));
auto eigen_out = framework::EigenVector<T>::Flatten(*tensor);
auto eigen_in = framework::EigenVector<T>::Flatten(*in);
auto& dev = context.GetEigenDevice<Place>();
eigen_out.device(dev) = scale * eigen_in;
}
};
} // namespace operators
} // namespace paddle
......@@ -276,17 +276,21 @@ int32_t Argument::resizeAndCopyFrom(const Argument& src,
void Argument::concat(const std::vector<Argument>& args,
const std::vector<int>& selectRows,
const std::vector<int>& seqStartPos,
const std::vector<int>& copySize,
bool useGpu,
hl_stream_t stream,
PassType passType) {
CHECK(!subSequenceStartPositions)
<< "undefined behavior for subsequence positions";
size_t batchSize = selectRows.size();
size_t batchSize = 0;
for (size_t i = 0; i < copySize.size(); ++i)
batchSize += copySize[i] * (seqStartPos[i + 1] - seqStartPos[i]);
auto copyArg = [batchSize, stream](MatrixPtr& dst,
MatrixPtr src,
int startRow,
int pos,
int desStartRow,
int srcStartRow,
int size,
bool useGpu) {
if (!src) {
......@@ -300,14 +304,14 @@ void Argument::concat(const std::vector<Argument>& args,
dst->resize(batchSize, width);
}
MatrixPtr tmpMatrix = dst->subMatrix(startRow, size);
tmpMatrix->copyFrom(*src->subMatrix(pos, size), stream);
MatrixPtr tmpMatrix = dst->subMatrix(desStartRow, size);
tmpMatrix->copyFrom(*src->subMatrix(srcStartRow, size), stream);
};
auto copyIds = [batchSize, stream](IVectorPtr& dst,
const IVectorPtr& src,
int startRow,
int pos,
int desStartRow,
int srcStartRow,
int size,
bool useGpu) {
if (!src) {
......@@ -315,13 +319,14 @@ void Argument::concat(const std::vector<Argument>& args,
return;
}
IVector::resizeOrCreate(dst, batchSize, useGpu);
dst->subVec(startRow, size)->copyFrom(*src->subVec(pos, size), stream);
dst->subVec(desStartRow, size)
->copyFrom(*src->subVec(srcStartRow, size), stream);
};
auto copyStrs = [batchSize, stream](SVectorPtr& dst,
const SVectorPtr& src,
int startRow,
int pos,
int desStartRow,
int srcStartRow,
int size,
bool useGpu) {
if (!src) {
......@@ -333,30 +338,31 @@ void Argument::concat(const std::vector<Argument>& args,
} else {
dst->resize(batchSize);
}
std::copy(
src->begin() + pos, src->begin() + pos + size, dst->begin() + startRow);
std::copy(src->begin() + srcStartRow,
src->begin() + srcStartRow + size,
dst->begin() + desStartRow);
};
dataId = args[0].dataId;
CHECK_NE(seqStartPos.size(), 0UL);
size_t sampleNum = seqStartPos.size() - 1;
for (size_t i = 0; i < sampleNum; ++i) {
int desStartRow = 0;
for (size_t i = 0; i < copySize.size(); ++i) {
int startPos = seqStartPos[i];
int endPos = seqStartPos[i + 1];
CHECK_GE(args.size(), static_cast<size_t>(endPos - startPos));
for (int j = startPos; j < endPos; ++j) {
const Argument& arg = args[j - startPos];
CHECK_EQ(arg.dataId, dataId) << "Arguments in concat should have"
<< " same dataId";
const int copySize = 1;
const int rowIdx = selectRows[j];
copyArg(in, arg.in, j, rowIdx, copySize, useGpu);
copyArg(value, arg.value, j, rowIdx, copySize, useGpu);
CHECK_EQ(arg.dataId, dataId) << "Arguments to concatenate should have "
<< "the same dataId.";
const int srcStartRow = selectRows[j];
copyArg(in, arg.in, desStartRow, srcStartRow, copySize[i], useGpu);
copyArg(value, arg.value, desStartRow, srcStartRow, copySize[i], useGpu);
if (passType != PASS_TEST) {
copyArg(grad, arg.grad, j, rowIdx, copySize, useGpu);
copyArg(grad, arg.grad, desStartRow, srcStartRow, copySize[i], useGpu);
}
copyIds(ids, arg.ids, j, rowIdx, copySize, useGpu);
copyStrs(strs, arg.strs, j, rowIdx, copySize, useGpu);
copyIds(ids, arg.ids, desStartRow, srcStartRow, copySize[i], useGpu);
copyStrs(strs, arg.strs, desStartRow, srcStartRow, copySize[i], useGpu);
desStartRow += copySize[i];
}
}
ICpuGpuVector::resizeOrCreate(
......
......@@ -240,6 +240,7 @@ struct Argument {
void concat(const std::vector<Argument>& args,
const std::vector<int>& selectRows,
const std::vector<int>& seqStartPos,
const std::vector<int>& copySize,
bool useGpu,
hl_stream_t stream,
PassType passType);
......
......@@ -28,3 +28,4 @@ py_test(test_uniform_random_op SRCS test_uniform_random_op.py)
py_test(test_recurrent_op SRCS test_recurrent_op.py)
py_test(test_sgd_op SRCS test_sgd_op.py)
py_test(test_gradient_checker SRCS test_gradient_checker.py)
py_test(test_scale_and_identity_op SRCS test_scale_and_identity_op.py)
......@@ -160,8 +160,13 @@ class GradientChecker(unittest.TestCase):
grad_tensor.set(data, place)
# run backward op
for name in backward_op.outputs():
backward_outs = backward_op.outputs()
backward_names = [
item for key in backward_outs for item in backward_outs[key]
]
for name in backward_names:
scope.new_var(name)
backward_op.infer_shape(scope)
backward_op.run(scope, ctx)
......
import unittest
from op_test_util import OpTestMeta
from gradient_checker import GradientChecker, create_op
import numpy as np
from paddle.v2.framework.op import Operator
class IdentityTest(unittest.TestCase):
__metaclass__ = OpTestMeta
def setUp(self):
self.type = "identity"
self.inputs = {'X': np.random.random((32, 784)).astype("float32")}
self.outputs = {'Out': self.inputs['X']}
class IdentityGradOpTest(GradientChecker):
def test_normal(self):
op = create_op("identity")
inputs = {"X": np.random.random((10, 10)).astype("float32")}
self.check_grad(op, inputs, set("X"), "Out")
class ScaleTest(unittest.TestCase):
__metaclass__ = OpTestMeta
def setUp(self):
self.type = "scale"
self.inputs = {'X': np.random.random((32, 784)).astype("float32")}
self.attrs = {'scale': -2.3}
self.outputs = {'Out': self.inputs['X'] * self.attrs['scale']}
class ScaleGradTest(GradientChecker):
def test_normal(self):
op = Operator("scale", X="X", Out="Out", scale=3.2)
self.check_grad(op,
{"X": np.random.random((10, 10)).astype("float32")},
set("X"), "Out")
if __name__ == '__main__':
unittest.main()
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册