ce_debug
frankwhzhang 6 years ago
commit f4cc5881b0

@ -1,6 +1,7 @@
add_subdirectory(memory)
add_subdirectory(platform)
add_subdirectory(framework)
add_subdirectory(imperative)
add_subdirectory(operators)
add_subdirectory(string)
add_subdirectory(recordio)

@ -33,11 +33,7 @@ void DataFeed::AddFeedVar(Variable* var, const std::string& name) {
CheckInit();
for (size_t i = 0; i < use_slots_.size(); ++i) {
if (name == use_slots_[i]) {
if (use_slots_is_dense_[i]) {
feed_vec_[i] = MixTensor(var->GetMutable<Tensor>());
} else {
feed_vec_[i] = MixTensor(var->GetMutable<LoDTensor>());
}
feed_vec_[i] = var->GetMutable<LoDTensor>();
}
}
}
@ -301,6 +297,7 @@ bool MultiSlotDataFeed::ParseOneInstance(std::vector<MultiSlotType>* instance) {
"the data, please check if the data contains unresolvable "
"characters.\nplease check this error line: %s",
str);
if (idx != -1) {
(*instance)[idx].Init(all_slots_type_[i]);
if ((*instance)[idx].GetType()[0] == 'f') { // float
@ -337,6 +334,7 @@ void MultiSlotDataFeed::AddInstanceToInsVec(
(*ins_vec)[i].InitOffset();
}
}
for (size_t i = 0; i < instance.size(); ++i) {
(*ins_vec)[i].AddIns(instance[i]);
}
@ -348,36 +346,25 @@ void MultiSlotDataFeed::PutToFeedVec(
const auto& type = ins_vec[i].GetType();
const auto& offset = ins_vec[i].GetOffset();
int total_instance = static_cast<int>(offset.back());
if (type[0] == 'f') { // float
const auto& feasign = ins_vec[i].GetFloatData();
if (feed_vec_[i].IsDense()) {
int size_in_each_batch = total_instance / batch_size_;
float* tensor_ptr = feed_vec_[i].GetTensor()->mutable_data<float>(
{batch_size_, size_in_each_batch}, platform::CPUPlace());
memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float));
} else {
float* tensor_ptr = feed_vec_[i].GetLoDTensor()->mutable_data<float>(
{total_instance, 1}, platform::CPUPlace());
memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float));
LoD data_lod{offset};
feed_vec_[i].GetLoDTensor()->set_lod(data_lod);
}
float* tensor_ptr = feed_vec_[i]->mutable_data<float>(
{total_instance, 1}, platform::CPUPlace());
memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(float));
} else if (type[0] == 'u') { // uint64
// no uint64_t type in paddlepaddle
const auto& feasign = ins_vec[i].GetUint64Data();
if (feed_vec_[i].IsDense()) {
int size_in_each_batch = total_instance / batch_size_;
int64_t* tensor_ptr = feed_vec_[i].GetTensor()->mutable_data<int64_t>(
{batch_size_, size_in_each_batch}, platform::CPUPlace());
memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t));
} else {
int64_t* tensor_ptr =
feed_vec_[i].GetLoDTensor()->mutable_data<int64_t>(
{total_instance, 1}, platform::CPUPlace());
memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t));
LoD data_lod{offset};
feed_vec_[i].GetLoDTensor()->set_lod(data_lod);
}
int64_t* tensor_ptr = feed_vec_[i]->mutable_data<int64_t>(
{total_instance, 1}, platform::CPUPlace());
memcpy(tensor_ptr, &feasign[0], total_instance * sizeof(int64_t));
}
LoD data_lod{offset};
feed_vec_[i]->set_lod(data_lod);
if (use_slots_is_dense_[i]) {
int dim = total_instance / batch_size_;
feed_vec_[i]->Resize({batch_size_, dim});
}
}
}

@ -30,35 +30,6 @@ limitations under the License. */
namespace paddle {
namespace framework {
// Pack Tensor type and LoDTensor type into MixTensor type, in order
// to record either Tensor or LoDTensor information at the same time.
class MixTensor {
public:
MixTensor() {}
explicit MixTensor(LoDTensor* lodtensor) {
is_dense_ = false;
lodtensor_ = lodtensor;
}
explicit MixTensor(Tensor* tensor) {
is_dense_ = true;
tensor_ = tensor;
}
bool IsDense() { return is_dense_; }
LoDTensor* GetLoDTensor() {
PADDLE_ENFORCE(!is_dense_, "Let a dense var return a LoDTensor ptr.");
return lodtensor_;
}
Tensor* GetTensor() {
PADDLE_ENFORCE(is_dense_, "Let a sparse var return a Tensor ptr.");
return tensor_;
}
private:
bool is_dense_;
LoDTensor* lodtensor_;
Tensor* tensor_;
};
// DataFeed is the base virtual class for all ohther DataFeeds.
// It is used to read files and parse the data for subsequent trainer.
// Example:
@ -133,7 +104,7 @@ class DataFeed {
use_slots_index_; // -1: not used; >=0: the index of use_slots_
// The data read by DataFeed will be stored here
std::vector<MixTensor> feed_vec_;
std::vector<LoDTensor*> feed_vec_;
// the batch size defined by user
int default_batch_size_;

@ -152,19 +152,13 @@ void GetElemSetFromReader(std::vector<MultiTypeSet>* reader_elem_set,
const auto& multi_slot_desc = data_feed_desc.multi_slot_desc();
std::map<std::string, const paddle::framework::LoDTensor*>
lodtensor_targets;
std::map<std::string, const paddle::framework::Tensor*> tensor_targets;
for (int i = 0; i < multi_slot_desc.slots_size(); ++i) {
const auto& slot = multi_slot_desc.slots(i);
if (slot.is_used()) {
const auto& name = slot.name();
readers[idx]->AddFeedVar(scope->Var(name), name);
if (slot.is_dense()) {
tensor_targets[name] =
&scope->FindVar(name)->Get<paddle::framework::Tensor>();
} else {
lodtensor_targets[name] =
&scope->FindVar(name)->Get<paddle::framework::LoDTensor>();
}
lodtensor_targets[name] =
&scope->FindVar(name)->Get<paddle::framework::LoDTensor>();
}
}
readers[idx]->Start();
@ -175,8 +169,9 @@ void GetElemSetFromReader(std::vector<MultiTypeSet>* reader_elem_set,
if (!slot.is_used()) {
continue;
}
const paddle::framework::LoDTensor* tens =
lodtensor_targets[slot.name()];
if (slot.is_dense()) { // dense branch
const paddle::framework::Tensor* tens = tensor_targets[slot.name()];
if (slot.type() == "uint64") {
const int64_t* data = tens->data<int64_t>();
int batch_size = tens->dims()[0];
@ -202,8 +197,6 @@ void GetElemSetFromReader(std::vector<MultiTypeSet>* reader_elem_set,
PADDLE_THROW("Error type in proto file.");
}
} else { // sparse branch
const paddle::framework::LoDTensor* tens =
lodtensor_targets[slot.name()];
if (slot.type() == "uint64") {
const int64_t* data = tens->data<int64_t>();
for (size_t i = 0; i < tens->NumElements(); ++i) {

@ -97,7 +97,7 @@ void ExecutorThreadWorker::SetDevice() {
static unsigned concurrency_cap = std::thread::hardware_concurrency();
int thread_id = this->thread_id_;
if (thread_id < concurrency_cap) {
if (static_cast<unsigned>(thread_id) < concurrency_cap) {
unsigned proc = thread_id;
cpu_set_t mask;

@ -16,7 +16,9 @@ limitations under the License. */
#include <string>
#include <vector>
#include "glog/logging.h"
#include "paddle/fluid/framework/var_type.h"
#include "paddle/fluid/framework/variable.h"
#include "paddle/fluid/platform/place.h"
namespace paddle {
namespace framework {
@ -53,5 +55,12 @@ LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name,
return tensor;
}
LoDTensor& GetVariableTensor(const Scope& scope, const std::string& var_name) {
Variable* var = scope.FindVar(var_name);
PADDLE_ENFORCE(var, "%s no in scope", var_name);
PADDLE_ENFORCE(var->IsType<LoDTensor>(), "Only support lod tensor now.");
return *var->GetMutable<LoDTensor>();
}
} // namespace framework
} // namespace paddle

@ -27,5 +27,7 @@ void SetFeedVariable(Scope* scope, const LoDTensor& input,
LoDTensor& GetFetchVariable(const Scope& scope, const std::string& var_name,
size_t index);
LoDTensor& GetVariableTensor(const Scope& scope, const std::string& var_name);
} // namespace framework
} // namespace paddle

@ -38,9 +38,8 @@ void CheckProgram(const ProgramDesc &program) {
switch (role_id) {
case _INT(OpRole::kForward):
if (visit.find(_INT(OpRole::kBackward)) != visit.end()) {
LOG(ERROR)
<< "Cannot add backward operator before forward operator %s."
<< op->Type();
LOG(ERROR) << "Cannot add backward operator before forward operator "
<< op->Type();
}
break;
case _INT(OpRole::kBackward):

@ -0,0 +1,3 @@
cc_library(layer SRCS layer.cc DEPS proto_desc operator)
cc_library(tracer SRCS tracer.cc DEPS proto_desc)
cc_library(engine SRCS engine.cc)

@ -0,0 +1,53 @@
// 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/imperative/engine.h"
#include <mutex> // NOLINT
#include <vector>
#include "glog/logging.h"
namespace paddle {
namespace imperative {
static std::once_flag init_engine;
static Engine* engine;
class DummyEngine : public Engine {
public:
void Enqueue(Runnable* runnable) override {
queued_runnables_.push_back(runnable);
}
size_t Size() const override { return queued_runnables_.size(); }
void Sync() override {
for (Runnable* l : queued_runnables_) {
LOG(INFO) << "running " << reinterpret_cast<void*>(l);
}
queued_runnables_.clear();
}
private:
std::vector<Runnable*> queued_runnables_;
};
Engine* GetEngine() {
std::call_once(init_engine, []() { engine = new DummyEngine(); });
return engine;
}
} // namespace imperative
} // namespace paddle

@ -0,0 +1,39 @@
// 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 <cstddef>
#include <cstdint>
namespace paddle {
namespace imperative {
struct Runnable {};
class Engine {
public:
virtual ~Engine() {}
virtual void Enqueue(Runnable* runnable) = 0;
virtual size_t Size() const = 0;
virtual void Sync() = 0;
};
Engine* GetEngine();
} // namespace imperative
} // namespace paddle

@ -0,0 +1,221 @@
// 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/imperative/layer.h"
#include <deque>
#include <limits>
#include <map>
#include <random>
#include <utility>
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/string/printf.h"
namespace paddle {
namespace imperative {
using framework::Variable;
void AddTo(Variable* src, Variable* dst) {
framework::LoDTensor* dst_tensor = dst->GetMutable<framework::LoDTensor>();
framework::LoDTensor* src_tensor = src->GetMutable<framework::LoDTensor>();
PADDLE_ENFORCE(dst_tensor->numel() == src_tensor->numel(), "%lld vs %lld",
dst_tensor->numel(), src_tensor->numel());
float* dst_data = dst_tensor->mutable_data<float>(platform::CPUPlace());
const float* src_data = src_tensor->data<float>();
for (size_t i = 0; i < src_tensor->numel(); ++i) {
dst_data[i] += src_data[i];
}
}
class Autograd {
public:
explicit Autograd(framework::Scope* scope) : scope_(scope) {}
void RunBackward(VarBase* var) {
PADDLE_ENFORCE(var->pre_op_->op_desc_);
// TODO(panyx0718): Only create for vars that "require_grad"
(*var->pre_op_->output_vars_)[var->pre_op_out_idx_]->grads_ = var->grads_;
std::deque<OpBase*> ready;
ready.push_back(var->pre_op_);
std::map<OpBase*, int> dep_counts = ComputeDepCounts(var->pre_op_);
while (!ready.empty()) {
OpBase* ready_op = ready.front();
ready.pop_front();
std::vector<Variable*> input_grads = ready_op->ApplyGrad(scope_);
for (size_t i = 0; i < input_grads.size(); ++i) {
if (!input_grads[i]) continue;
OpBase* pre_op = ready_op->pre_ops_->at(i);
if (!pre_op) continue;
dep_counts[pre_op] -= 1;
PADDLE_ENFORCE(dep_counts[pre_op] >= 0);
bool pre_op_ready = dep_counts[pre_op] == 0;
if (pre_op_ready) {
ready.push_back(pre_op);
}
}
}
}
private:
std::map<OpBase*, int> ComputeDepCounts(OpBase* op) {
std::map<OpBase*, int> ret;
std::deque<OpBase*> queue;
queue.push_back(op);
std::unordered_set<OpBase*> visited;
visited.insert(op);
while (!queue.empty()) {
OpBase* candidate = queue.front();
queue.pop_front();
for (OpBase* pre_op : *(candidate->pre_ops_)) {
if (!pre_op) continue;
if (visited.find(pre_op) == visited.end()) {
visited.insert(pre_op);
queue.push_back(pre_op);
}
ret[pre_op] += 1;
}
}
return ret;
}
framework::Scope* scope_;
};
framework::Variable* CreateVariable(const std::string& name,
const framework::DDim& dim, float val,
framework::Scope* scope,
bool random_name = true) {
std::string varname = name;
if (random_name) {
std::mt19937 rng;
rng.seed(std::random_device()());
std::uniform_int_distribution<std::mt19937::result_type> dist6(
1, std::numeric_limits<int>::max());
int id = dist6(rng);
varname = string::Sprintf("%s@%d", varname, id);
}
VLOG(3) << "creating var " << varname;
framework::Variable* var = scope->Var(varname);
framework::LoDTensor* tensor = var->GetMutable<framework::LoDTensor>();
float* data = tensor->mutable_data<float>(dim, platform::CPUPlace());
std::fill(data, data + tensor->numel(), val);
return var;
}
framework::LoDTensor& VarBase::Grad() {
VLOG(3) << "get var grad " << var_desc_->Name();
return *grads_->GetMutable<framework::LoDTensor>();
}
void VarBase::ApplyGrad(framework::Scope* scope, Variable* grad) {
VLOG(3) << "apply var grad " << var_desc_->Name() << " "
<< grad->Get<framework::LoDTensor>().data<float>()[0];
if (!grads_) {
grads_ =
CreateVariable(string::Sprintf("%s@IGrad", var_desc_->Name()),
var_->Get<framework::LoDTensor>().dims(), 0.0, scope);
}
AddTo(grad, grads_);
VLOG(3) << "grad_ after apply var grad " << var_desc_->Name() << " "
<< grads_->Get<framework::LoDTensor>().data<float>()[0];
}
std::vector<Variable*> OpBase::ApplyGrad(framework::Scope* scope) {
VLOG(3) << "op grad " << grad_op_desc_->Type();
for (const std::string& grad_invar : grad_op_desc_->InputArgumentNames()) {
if (grad_to_var_->find(grad_invar) == grad_to_var_->end()) {
// grad op inputs can be forward inputs, so not in grad_to_var.
continue;
}
VLOG(3) << "op grad in var " << grad_invar;
block_->FindRecursiveOrCreateVar(grad_invar);
framework::Variable* var = scope->Var(grad_invar);
const std::string& invar = grad_to_var_->at(grad_invar);
for (VarBase* varbase : *output_vars_) {
// Use the accumulated grads_ by sharing the input with grads_.
if (varbase->var_desc_->Name() == invar) {
var->GetMutable<framework::LoDTensor>()->ShareDataWith(
varbase->grads_->Get<framework::LoDTensor>());
break;
}
}
}
for (const std::string& outvar : grad_op_desc_->OutputArgumentNames()) {
VLOG(3) << "grad outvar " << outvar;
block_->FindRecursiveOrCreateVar(outvar);
framework::Variable* var = scope->Var(outvar);
if (!var->IsInitialized()) {
framework::VarDesc* var_desc = block_->FindVar(outvar);
if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
var->GetMutable<framework::LoDTensor>();
} else {
LOG(ERROR) << "tracer doesn't support yet";
}
}
}
grad_op_desc_->InferShape(*block_);
grad_op_desc_->InferVarType(block_);
std::unique_ptr<framework::OperatorBase> opbase =
framework::OpRegistry::CreateOp(*grad_op_desc_);
opbase->Run(*scope, platform::CPUPlace());
// `ret` matches exactly with `input_vars_` of forward op.
std::vector<Variable*> ret;
for (size_t i = 0; i < input_vars_->size(); ++i) {
bool found = false;
for (const std::string& outvar : grad_op_desc_->OutputArgumentNames()) {
Variable* var = scope->FindVar(outvar);
VarBase* origin_var = (*input_vars_)[i];
std::string orig_var = grad_to_var_->at(outvar);
PADDLE_ENFORCE(origin_var->var_desc_->Name() == orig_var);
VLOG(3) << "apply grad " << outvar << " with origin " << orig_var;
origin_var->ApplyGrad(scope, var);
found = true;
ret.push_back(var);
// TODO(panyx0718): There might be another outvar with the same name.
// In that case, it doesn't matter the first one or the second one is
// used.
break;
}
if (!found) {
ret.push_back(nullptr);
}
}
return ret;
}
void VarBase::RunBackward(framework::Scope* scope) {
grads_ = CreateVariable(framework::GradVarName(var_desc_->Name()),
var_->Get<framework::LoDTensor>().dims(), 1.0, scope,
false);
if (!pre_op_) return;
Autograd(scope).RunBackward(this);
}
} // namespace imperative
} // namespace paddle

@ -0,0 +1,102 @@
// 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 <string>
#include <vector>
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/var_desc.h"
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
namespace imperative {
class OpBase;
class VarBase {
public:
VarBase()
: pre_op_(nullptr),
pre_op_out_idx_(-1),
var_desc_(nullptr),
var_(nullptr),
grads_(nullptr) {}
virtual ~VarBase() {}
void ApplyGrad(framework::Scope* scope, framework::Variable* grad);
void RunBackward(framework::Scope* scope);
framework::LoDTensor& Grad();
OpBase* pre_op_;
int pre_op_out_idx_;
framework::VarDesc* var_desc_;
framework::Variable* var_;
framework::Variable* grads_;
};
class OpBase {
public:
OpBase()
: input_vars_(new std::vector<VarBase*>()),
output_vars_(new std::vector<VarBase*>()),
pre_ops_(new std::vector<OpBase*>()),
pre_ops_out_idx_(new std::vector<int>()),
op_desc_(nullptr),
grad_op_desc_(nullptr) {}
virtual ~OpBase() {
delete input_vars_;
delete output_vars_;
delete pre_ops_;
delete pre_ops_out_idx_;
if (grad_op_desc_) delete grad_op_desc_;
if (grad_to_var_) delete grad_to_var_;
}
std::vector<framework::Variable*> ApplyGrad(framework::Scope* scope);
std::vector<VarBase*>* input_vars_;
std::vector<VarBase*>* output_vars_;
std::vector<OpBase*>* pre_ops_;
std::vector<int>* pre_ops_out_idx_;
framework::OpDesc* op_desc_;
framework::OpDesc* grad_op_desc_;
std::unordered_map<std::string, std::string>* grad_to_var_;
framework::BlockDesc* block_;
};
class Layer {
public:
virtual ~Layer() {}
virtual std::vector<VarBase> Forward(const std::vector<VarBase>& inputs) {
std::vector<VarBase> vars;
return vars;
}
virtual void Backward() { LOG(ERROR) << "To support customize"; }
};
} // namespace imperative
} // namespace paddle

@ -0,0 +1,19 @@
// 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/imperative/tracer.h"
namespace paddle {
namespace imperative {} // namespace imperative
} // namespace paddle

@ -0,0 +1,128 @@
// 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 <map>
#include <string>
#include <vector>
#include "paddle/fluid/framework/op_desc.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/imperative/engine.h"
#include "paddle/fluid/imperative/layer.h"
namespace paddle {
namespace imperative {
void CreateGradOp(const framework::OpDesc& op_desc,
const std::unordered_set<std::string>& no_grad_set,
const std::vector<framework::BlockDesc*>& grad_sub_block,
framework::OpDesc** grad_op_desc,
std::unordered_map<std::string, std::string>* grad_to_var) {
std::vector<std::unique_ptr<framework::OpDesc>> grad_op_descs =
framework::OpInfoMap::Instance()
.Get(op_desc.Type())
.GradOpMaker()(op_desc, no_grad_set, grad_to_var, grad_sub_block);
PADDLE_ENFORCE(grad_op_descs.size() == 1, "Only support 1 grad op now.");
// TODO(panyx0718): Leak?
*grad_op_desc = grad_op_descs[0].release();
}
class Tracer {
public:
explicit Tracer(framework::BlockDesc* root_block) : root_block_(root_block) {
root_scope_ = new framework::Scope();
scopes_[root_block_] = root_scope_;
}
virtual ~Tracer() { delete root_scope_; }
void Trace(OpBase* op, const std::vector<VarBase*>& inputs,
const std::vector<VarBase*>& outputs,
framework::BlockDesc* block) {
framework::Scope* scope = GetScope(block);
framework::OpDesc* op_desc = op->op_desc_;
VLOG(3) << "tracer tracing " << op_desc->Type();
op_desc->InferShape(*block);
op_desc->InferVarType(block);
std::unique_ptr<framework::OperatorBase> op_base =
framework::OpRegistry::CreateOp(*op_desc);
*op->input_vars_ = inputs;
for (VarBase* input : inputs) {
const std::string vname = input->var_desc_->Name();
framework::Variable* var = scope->Var(vname);
input->var_ = var;
if (!var->IsInitialized()) {
framework::VarDesc* var_desc = block->FindVar(vname);
if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
var->GetMutable<framework::LoDTensor>();
} else {
LOG(ERROR) << "tracer doesn't support yet";
}
}
if (input->pre_op_) {
op->pre_ops_->push_back(input->pre_op_);
op->pre_ops_out_idx_->push_back(input->pre_op_out_idx_);
} else {
op->pre_ops_->push_back(nullptr);
}
}
*op->output_vars_ = outputs;
for (size_t i = 0; i < outputs.size(); ++i) {
const std::string vname = outputs[i]->var_desc_->Name();
framework::Variable* var = scope->Var(vname);
if (!var->IsInitialized()) {
framework::VarDesc* var_desc = block->FindVar(vname);
if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
var->GetMutable<framework::LoDTensor>();
} else {
LOG(ERROR) << "tracer doesn't support yet";
}
}
outputs[i]->var_ = var;
outputs[i]->pre_op_ = op;
outputs[i]->pre_op_out_idx_ = i;
}
op_base->Run(*scope, platform::CPUPlace());
framework::OpDesc* grad_op_desc;
auto grad_to_var = new std::unordered_map<std::string, std::string>();
CreateGradOp(*op_desc, {}, {block}, &grad_op_desc, grad_to_var);
op->grad_op_desc_ = grad_op_desc;
op->grad_to_var_ = grad_to_var;
op->block_ = block;
}
framework::Scope* GetScope(framework::BlockDesc* block) {
if (scopes_.find(block) != scopes_.end()) {
return scopes_.at(block);
}
framework::BlockDesc* parent_block = block->ParentBlock();
PADDLE_ENFORCE(scopes_.find(parent_block) != scopes_.end());
framework::Scope* scope = &scopes_[parent_block]->NewScope();
scopes_[block] = scope;
return scope;
}
private:
std::map<framework::BlockDesc*, framework::Scope*> scopes_;
framework::BlockDesc* root_block_;
framework::Scope* root_scope_;
};
} // namespace imperative
} // namespace paddle

@ -103,6 +103,7 @@ struct Argument {
// Model specified with program and parameters files.
DECL_ARGUMENT_FIELD(model_program_path, ModelProgramPath, std::string);
DECL_ARGUMENT_FIELD(model_params_path, ModelParamsPath, std::string);
DECL_ARGUMENT_FIELD(model_from_memory, ModelFromMemory, bool);
// The overall graph to work on.
DECL_ARGUMENT_UNIQUE_FIELD(main_graph, MainGraph, framework::ir::Graph);

@ -46,7 +46,7 @@ void IrGraphBuildPass::RunImpl(Argument *argument) {
argument->model_params_path_valid()) {
auto program =
LoadModel(argument->model_program_path(), argument->model_params_path(),
argument->scope_ptr(), place);
argument->scope_ptr(), place, argument->model_from_memory());
argument->SetMainProgram(program.release());
} else {
PADDLE_THROW(
@ -68,9 +68,14 @@ std::unique_ptr<framework::ProgramDesc> IrGraphBuildPass::LoadModel(
std::unique_ptr<framework::ProgramDesc> IrGraphBuildPass::LoadModel(
const std::string &program_path, const std::string &params_path,
framework::Scope *scope, const platform::Place &place) {
framework::Scope *scope, const platform::Place &place,
bool model_from_memory) {
framework::Executor exe(place);
return Load(&exe, scope, program_path, params_path);
if (!model_from_memory) {
return Load(&exe, scope, program_path, params_path);
} else {
return LoadFromMemory(&exe, scope, program_path, params_path);
}
}
std::string IrGraphBuildPass::repr() const { return "ir-graph-build-pass"; }

@ -24,7 +24,7 @@ namespace inference {
namespace analysis {
/*
* Load program and parameter to memory from the disk.
* Load program and parameter to memory from the disk or directly from memory.
*/
class IrGraphBuildPass : public AnalysisPass {
public:
@ -38,7 +38,8 @@ class IrGraphBuildPass : public AnalysisPass {
const platform::Place &place);
std::unique_ptr<framework::ProgramDesc> LoadModel(
const std::string &program_path, const std::string &params_path,
framework::Scope *scope, const platform::Place &place);
framework::Scope *scope, const platform::Place &place,
bool model_from_memory);
std::string model_binary_str_;
};

@ -53,6 +53,7 @@ contrib::AnalysisConfig::AnalysisConfig(const contrib::AnalysisConfig &other) {
use_tensorrt_ = other.use_tensorrt_;
tensorrt_max_batchsize_ = other.tensorrt_max_batchsize_;
tensorrt_workspace_size_ = other.tensorrt_workspace_size_;
model_from_memory_ = other.model_from_memory_;
if (use_gpu) {
pass_builder_.reset(new GpuPassStrategy(
@ -80,6 +81,8 @@ contrib::AnalysisConfig::AnalysisConfig(contrib::AnalysisConfig &&other) {
use_tensorrt_ = other.use_tensorrt_;
tensorrt_max_batchsize_ = other.tensorrt_max_batchsize_;
tensorrt_workspace_size_ = other.tensorrt_workspace_size_;
model_from_memory_ = other.model_from_memory_;
pass_builder_ = std::move(other.pass_builder_);
}
@ -102,4 +105,13 @@ void contrib::AnalysisConfig::EnableTensorRtEngine(int workspace_size,
pass_builder()->InsertPass(1, "tensorrt_subgraph_pass");
}
void contrib::AnalysisConfig::SetModelBuffer(const char *prog_buffer,
size_t prog_buffer_size,
const char *param_buffer,
size_t param_buffer_size) {
prog_file = std::string(prog_buffer, prog_buffer + prog_buffer_size);
param_file = std::string(param_buffer, param_buffer + param_buffer_size);
model_from_memory_ = true;
}
} // namespace paddle

@ -308,6 +308,7 @@ void AnalysisPredictor::OptimizeInferenceProgram() {
argument_.SetUseGPU(config_.use_gpu);
argument_.SetGPUDeviceId(config_.device);
argument_.SetModelFromMemory(config_.model_from_memory_);
// Analyze inference_program
if (!config_.model_dir.empty()) {
argument_.SetModelDir(config_.model_dir);
@ -448,20 +449,24 @@ bool AnalysisPredictor::LoadProgramDesc() {
return false;
}
std::string pb_content;
// Read binary
std::ifstream fin(filename, std::ios::in | std::ios::binary);
PADDLE_ENFORCE(static_cast<bool>(fin), "Cannot open file %s", filename);
fin.seekg(0, std::ios::end);
pb_content.resize(fin.tellg());
fin.seekg(0, std::ios::beg);
fin.read(&(pb_content.at(0)), pb_content.size());
fin.close();
// Create ProgramDesc
framework::proto::ProgramDesc proto;
proto.ParseFromString(pb_content);
if (!config_.model_from_memory()) {
std::string pb_content;
// Read binary
std::ifstream fin(filename, std::ios::in | std::ios::binary);
PADDLE_ENFORCE(static_cast<bool>(fin.is_open()), "Cannot open file %s",
filename);
fin.seekg(0, std::ios::end);
pb_content.resize(fin.tellg());
fin.seekg(0, std::ios::beg);
fin.read(&(pb_content.at(0)), pb_content.size());
fin.close();
proto.ParseFromString(pb_content);
} else {
proto.ParseFromString(config_.prog_file);
}
inference_program_.reset(new framework::ProgramDesc(proto));
return true;
}
@ -469,6 +474,7 @@ bool AnalysisPredictor::LoadProgramDesc() {
bool AnalysisPredictor::LoadParameters() {
PADDLE_ENFORCE_NOT_NULL(inference_program_.get(),
"The inference program should be loaded first.");
const auto &global_block = inference_program_->MutableBlock(0);
// create a temporary program to load parameters.

@ -52,10 +52,13 @@ struct AnalysisConfig : public NativeConfig {
bool use_tensorrt() const { return use_tensorrt_; }
void EnableMKLDNN();
// NOTE this is just for internal development, please not use it.
// NOT stable yet.
bool use_mkldnn() const { return use_mkldnn_; }
// Specify the memory buffer of program and parameter
void SetModelBuffer(const char* prog_buffer, size_t prog_buffer_size,
const char* program_buffer, size_t program_buffer_size);
bool model_from_memory() const { return model_from_memory_; }
friend class ::paddle::AnalysisPredictor;
protected:
@ -64,6 +67,7 @@ struct AnalysisConfig : public NativeConfig {
int tensorrt_workspace_size_;
int tensorrt_max_batchsize_;
std::unique_ptr<PassStrategy> pass_builder_;
bool model_from_memory_{false};
};
// Configurations for Anakin engine.

@ -69,7 +69,8 @@ bool IsPersistable(const framework::VarDesc* var) {
void LoadPersistables(framework::Executor* executor, framework::Scope* scope,
const framework::ProgramDesc& main_program,
const std::string& dirname,
const std::string& param_filename) {
const std::string& param_filename,
bool model_from_memory = false) {
const framework::BlockDesc& global_block = main_program.Block(0);
framework::ProgramDesc* load_program = new framework::ProgramDesc();
@ -108,6 +109,7 @@ void LoadPersistables(framework::Executor* executor, framework::Scope* scope,
op->SetType("load_combine");
op->SetOutput("Out", paramlist);
op->SetAttr("file_path", {param_filename});
op->SetAttr("model_from_memory", {model_from_memory});
op->CheckAttrs();
}
@ -130,16 +132,17 @@ std::unique_ptr<framework::ProgramDesc> Load(framework::Executor* executor,
"model version %ld is not supported.",
main_program->Version());
LoadPersistables(executor, scope, *main_program, dirname, "");
// model_from_memory is false in seperate parameters.
LoadPersistables(executor, scope, *main_program, dirname, "",
false /* model_from_memory */);
return main_program;
}
std::unique_ptr<framework::ProgramDesc> Load(
framework::Executor* executor, framework::Scope* scope,
const std::string& prog_filename, const std::string& param_filename) {
std::string model_filename = prog_filename;
std::string program_desc_str;
ReadBinaryFile(model_filename, &program_desc_str);
ReadBinaryFile(prog_filename, &program_desc_str);
std::unique_ptr<framework::ProgramDesc> main_program(
new framework::ProgramDesc(program_desc_str));
@ -147,7 +150,22 @@ std::unique_ptr<framework::ProgramDesc> Load(
"model version %ld is not supported.",
main_program->Version());
LoadPersistables(executor, scope, *main_program, "", param_filename);
LoadPersistables(executor, scope, *main_program, "", param_filename,
false /* model_from_memory */);
return main_program;
}
std::unique_ptr<framework::ProgramDesc> LoadFromMemory(
framework::Executor* executor, framework::Scope* scope,
const std::string& prog_buffer, const std::string& param_buffer) {
std::unique_ptr<framework::ProgramDesc> main_program(
new framework::ProgramDesc(prog_buffer));
PADDLE_ENFORCE(framework::IsProgramVersionSupported(main_program->Version()),
"model version %ld is not supported.",
main_program->Version());
LoadPersistables(executor, scope, *main_program, "", param_buffer,
true /* model_filename */);
return main_program;
}

@ -30,7 +30,8 @@ void Init(const std::vector<std::string> argv);
void LoadPersistables(framework::Executor* executor, framework::Scope* scope,
const framework::ProgramDesc& main_program,
const std::string& dirname,
const std::string& param_filename);
const std::string& param_filename,
bool model_from_memory);
std::unique_ptr<framework::ProgramDesc> Load(framework::Executor* executor,
framework::Scope* scope,
@ -41,6 +42,10 @@ std::unique_ptr<framework::ProgramDesc> Load(framework::Executor* executor,
const std::string& prog_filename,
const std::string& param_filename);
std::unique_ptr<framework::ProgramDesc> LoadFromMemory(
framework::Executor* executor, framework::Scope* scope,
const std::string& prog_buffer, const std::string& param_buffer);
// Save the variables from a scope to disk.
void SaveVars(const framework::Scope& scope,
const std::vector<std::string>& vars, const std::string& dirname,

@ -93,9 +93,17 @@ void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
}
}
void SetConfig(contrib::AnalysisConfig *cfg) {
cfg->prog_file = FLAGS_infer_model + "/__model__";
cfg->param_file = FLAGS_infer_model + "/param";
void SetConfig(contrib::AnalysisConfig *cfg, bool memory_load = false) {
if (memory_load) {
std::string buffer_prog, buffer_param;
ReadBinaryFile(FLAGS_infer_model + "/__model__", &buffer_prog);
ReadBinaryFile(FLAGS_infer_model + "/param", &buffer_param);
cfg->SetModelBuffer(&buffer_prog[0], buffer_prog.size(), &buffer_param[0],
buffer_param.size());
} else {
cfg->prog_file = FLAGS_infer_model + "/__model__";
cfg->param_file = FLAGS_infer_model + "/param";
}
cfg->use_gpu = false;
cfg->device = 0;
cfg->specify_input_name = true;
@ -114,9 +122,9 @@ void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {
}
// Easy for profiling independently.
TEST(Analyzer_Chinese_ner, profile) {
void profile(bool memory_load = false) {
contrib::AnalysisConfig cfg;
SetConfig(&cfg);
SetConfig(&cfg, memory_load);
std::vector<PaddleTensor> outputs;
std::vector<std::vector<PaddleTensor>> input_slots_all;
@ -138,6 +146,12 @@ TEST(Analyzer_Chinese_ner, profile) {
}
}
TEST(Analyzer_Chinese_ner, profile) { profile(); }
TEST(Analyzer_Chinese_ner, profile_memory_load) {
profile(true /* memory_load */);
}
// Check the fuse status
TEST(Analyzer_Chinese_ner, fuse_statis) {
contrib::AnalysisConfig cfg;

@ -49,8 +49,6 @@ std::ostream &operator<<(std::ostream &os, const NativeConfig &config) {
os << GenSpaces(num_spaces) << "device: " << config.device << "\n";
os << GenSpaces(num_spaces)
<< "fraction_of_gpu_memory: " << config.fraction_of_gpu_memory << "\n";
os << GenSpaces(num_spaces) << "prog_file: " << config.prog_file << "\n";
os << GenSpaces(num_spaces) << "param_file: " << config.param_file << "\n";
os << GenSpaces(num_spaces)
<< "specify_input_name: " << config.specify_input_name << "\n";
os << GenSpaces(num_spaces)
@ -65,6 +63,13 @@ std::ostream &operator<<(std::ostream &os,
os << GenSpaces(num_spaces) << "contrib::AnalysisConfig {\n";
num_spaces++;
os << *reinterpret_cast<const NativeConfig *>(&config);
if (!config.model_from_memory()) {
os << GenSpaces(num_spaces) << "prog_file: " << config.prog_file << "\n";
os << GenSpaces(num_spaces) << "param_file: " << config.param_file << "\n";
} else {
os << GenSpaces(num_spaces)
<< "prog_file and param_file: load from memory \n";
}
os << GenSpaces(num_spaces) << "enable_ir_optim: " << config.enable_ir_optim
<< "\n";
os << GenSpaces(num_spaces)

Some files were not shown because too many files have changed in this diff Show More

Loading…
Cancel
Save