Skip to content

P
Paddle

机器未来 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle

体验新版 GitCode，发现更多精彩内容 >>

提交 f1f8327c 编写于 作者: T tensor-tang
浏览文件
操作

Merge remote-tracking branch 'ups/develop' into refine/mem

上级 a0c5fd83 49f23e63
展开全部 隐藏空白更改
内联 并排
Showing with 2122 addition and 1555 deletion
paddle/contrib/CMakeLists.txt
浏览文件 @ f1f8327c
......@@ -14,4 +14,3 @@
#
add_subdirectory(inference)
add_subdirectory(tape)
paddle/contrib/tape/CMakeLists.txt 已删除 100644 → 0
浏览文件 @ a0c5fd83
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
if(APPLE)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=pessimizing-move")
endif(APPLE)
cc_library(tape_variable SRCS variable.cc DEPS ${FLUID_CORE_MODULES} device_context framework_proto proto_desc operator)
cc_library(tape SRCS tape.cc DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB} tape_variable)
cc_test(test_tape
SRCS test_tape.cc
DEPS tape tape_variable)
paddle/contrib/tape/README.md 已删除 100644 → 0
浏览文件 @ a0c5fd83
# Dynamic Graph on Fluid
PaddlePaddle Fluid is targeting the autodiff without tape, which, however, is very
challenging and we are still way from there. DyNet and PyTorch provide a good design
idea, the *tape*, that significantly eases the challenge. Also, DyNet provides
a C++ API that is as convenient as Python but with higher efficiency and could
conveniently integrate with industrial/production systems. This package, `tape`,
combines the good of
1. tape from PyTorch and DyNet
2. C++ API and core from DyNet
3. rich set of operators from PaddlePaddle
## Overview
We can implement Dynet-like Tape(See this [survey](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/survey/dynamic_graph.md))
by wrapping Paddle Fluid's `Operator` and `Variable`.
The user API is straight forward since
1. it is imperative. And it uses host language's control flow logic.
1. it avoids extra concepts such as `Scope` and `Executor`.
All of these benefits come at the cost of just adding one line `reset_global_tape`
at every iteration.
## Code Structure
In short, the `Tape` contains a vector of `OpHandle`s. And an `OpHandle` contains its
`type`, the pointers to the `Variable`s, and necessary attributes.
```c++
class Variable {
public:
VriableHandle Grad(); // returns its gradient variable
private:
framework::VarDesc desc_; // compile time infershape, necessary for lazy execution
framework::Variable var_; // run time variable, holds data memory
};
using VariableHandle = shared_ptr<Variable>;
struct OpHandle {
string type_;
map<string, vector<VariableHandle>> inputs_;
map<string, vector<VariableHandle>> outputs_;
AttributeMap attrs_;
};
class Tape {
public:
void AddOp(OpHandle); // add op
void Forward(); // execute the tape_
void Backward(); // execute the backward of the tape_
private:
vector<OpHandle> tape_;
};
```
We uses `Function` to indicate layers. It takes care of parameter
initialization and `AddOp` to the Tape when it is called.
```c++
class Linear {
public:
Linear(int in_dim, int out_dim, const std::string &act)
: w_(new Variable("LinearWeight")),
b_(new Variable("LinearBias")),
act_(act) {
Tape init_tape;
std::string initializer = "fill_constant";
framework::AttributeMap attrs;
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
attrs["shape"] = std::vector<int>{in_dim, out_dim};
attrs["value"] = 1.0f;
init_tape.AddOp(initializer, {}, {{"Out", {w_}}}, attrs);
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
attrs["shape"] = std::vector<int>{out_dim};
attrs["value"] = 1.0f;
init_tape.AddOp(initializer, {}, {{"Out", {b_}}}, attrs);
init_tape.Forward();
}
VariableHandle operator()(VariableHandle input) {
VariableHandle pre_bias(new Variable("linear"));
get_global_tape().AddOp("mul",
{{"X", {input}}, {"Y", {w_}}},
{{"Out", {pre_bias}}},
{{"x_num_col_dims", 1}, {"y_num_col_dims", 1}});
VariableHandle pre_act(new Variable("linear"));
get_global_tape().AddOp("elementwise_add",
{{"X", {pre_bias}}, {"Y", {b_}}},
{{"Out", {pre_act}}},
{{"axis", 1}});
VariableHandle post_act(new Variable("linear"));
get_global_tape().AddOp(act_,
{{"X", {pre_act}}},
{{"Out", {post_act}}},
{});
return post_act;
}
std::vector<VariableHandle> Params() { return {w_, b_}; }
private:
VariableHandle w_;
VariableHandle b_;
std::string act_;
};
```
## User API
```c++
// Model function
paddle::tape::Linear linear1(3, 3, "relu"); // init weight and bias
paddle::tape::Linear linear2(3, 3, "relu"); // init weight and bias
paddle::tape::Mean mean;
// Optimizer
paddle::tape::SGD sgd(0.001);
// Data Feeder
paddle::tape::Fill data_feeder(...);
VariableHandle input(new paddle::tape::Variable("input"));
VariableHandle label(new paddle::tape::Variable("label"));
for (int i = 0; i < 2; ++i) {
reset_global_tape();
data_feeder(input, label);
auto loss = softmax(linear2(linear1(input)), label); // compile time InferShape & InferVarType
LOG(INFO) << loss.value(); // Run forward up to loss
// Run backward, store gradient of w at w->Grad()
get_global_tape.Backward(loss);
// Update w
sgd(linear1.Params());
sgd(linear2.Params());
}
```
<details>
<summary></summary>
digraph G {
subgraph cluster_0 {
node [shape=record,style=filled];
style=filled;
color=lightgrey;
linear1 [label="{type: mul | {input | {<before_mul1>X: before_mul1 |<weight1> Y: weight1}} | {output |<before_bias1> Out: before_bias1}}"];
elementwise_add1 [label="{type: elementwise_add | {input | {<before_bias1>X: before_bias1 |<bias1> Y: bias1}} | {output |<before_act1> Out: before_act1}}"];
relu1 [label="{type: relu | {input | {<before_act1>X: before_act1 }} | {output |<after_act1> Out: after_act1}}"];
linear1 -> elementwise_add1->relu1;
label = "forward tape";
}
linear1:before_mul1->before_mul1
linear1:weight1->weight1
linear1:before_bias1->before_bias1
elementwise_add1:bias1->bias1
elementwise_add1:before_bias1->before_bias1
elementwise_add1:before_act1->before_act1
relu1:before_act1->before_act1
relu1:after_act1->after_act1
subgraph cluster_1 {
node [shape=record,style=filled];
style=filled;
color=lightgrey;
linear1_grad [label="{type: mul_grad | {input | {<before_mul1>X: before_mul1 |<weight1> Y: weight1|<before_bias1_grad> Out_grad: before_bias1_grad}} | {output |{<before_mul1_grad>X_grad: before_mul1_grad |<weight1_grad> Y_grad: weight1_grad}}}"];
elementwise_add1_grad [label="{type: elementwise_add_grad | {input | <before_act1_grad> Out_grad: before_act1_grad} | {output |{<before_bias1_grad>X_grad: before_bias1_grad |<bias1_grad> Y_grad: bias1_grad}}}"];
relu1_grad [label="{type: relu_grad | {input |<after_act1_grad> Out_grad: after_act1_grad} | {ouput | {<before_act1_grad>X_grad: before_act1_grad }}}"];
linear1_grad -> elementwise_add1_grad ->relu1_grad [dir=back];
label = "backward tape";
}
relu1_grad:after_act1_grad->after_act1_grad
relu1_grad:before_act1_grad->before_act1_grad
elementwise_add1_grad:before_act1_grad->before_act1_grad
elementwise_add1_grad:before_bias1_grad->before_bias1_grad
elementwise_add1_grad:bias1_grad->bias1_grad
linear1_grad:before_mul1->before_mul1
linear1_grad:weight1->weight1
linear1_grad:before_bias1_grad->before_bias1_grad
linear1_grad:before_mul1_grad->before_mul1_grad
linear1_grad:weight1_grad->weight1_grad
subgraph cluster_2 {
node [shape=record];
label = "Linear1";
weight1
bias1
}
weight1 -> weight1_grad [ label="Grad()", style="dashed" ];
bias1 -> bias1_grad [ label="Grad()", style="dashed"];
}
</details>
![Image](https://github.com/tonyyang-svail/Paddle/blob/cpp_tap/paddle/contrib/tape/computation_graph.png)
## Code Reuse
We want to stay close to Paddle Fluid as much as possible.
### Reuse All Operators
As all Ops are registered at `OpInfoMap`, the effort of adding a new `Function`
is about 10 lines of code, similar to expose an operator to Python.
### Reuse Compile Time InferShape and InferVarType
Note that all the symbolic information is stored at `tape::Varaible::desc_`, instead
of `ProgramDesc.block.vars`, we create a temporary `BlockDesc` to do `InferShape` and
`InferVarType` every time we `AddOp` to the tape.
### Reuse Operator::Run
We use smart pointer, instead of `Scope`, to manage memory. So we create a temporary
`Scope` for every `Operator::Run()`.
## Possible Feature
### Release Memory on Backward
We can release memory aggressively. During backward, we can delete the OpHandle once
we have finished its backward. Since all the variable is managed by smart pointer, the
memory is automatically released when its `ref_count` goes to 0.
### Kernel Fusion
As a symbolic representation of the Tape is constructed first before the actual
execution, it would be possible to perform graph optimization. One use case is kernel
fusion.
paddle/contrib/tape/function.h 已删除 100644 → 0
浏览文件 @ a0c5fd83
// 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 "paddle/contrib/tape/tape.h"
#include "paddle/contrib/tape/variable.h"
#include "paddle/fluid/framework/type_defs.h"
namespace paddle {
namespace tape {
class Function {};
class Fill {
public:
Fill(const std::string &initializer, const framework::AttributeMap &attrs)
: initializer_(initializer), attrs_(attrs) {}
void operator()(VariableHandle var) {
get_global_tape().AddOp(initializer_, {}, {{"Out", {var}}}, attrs_);
}
private:
const std::string initializer_;
const framework::AttributeMap attrs_;
};
class Mean {
public:
VariableHandle operator()(VariableHandle var) {
VariableHandle out(new Variable("mean"));
get_global_tape().AddOp("mean", {{"X", {var}}}, {{"Out", {out}}}, {});
return out;
}
};
class Linear {
public:
Linear(int in_dim, int out_dim, const std::string &act)
: w_(new Variable("LinearWeight")),
b_(new Variable("LinearBias")),
act_(act) {
Tape init_tape;
std::string initializer = "fill_constant";
framework::AttributeMap attrs;
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
attrs["shape"] = std::vector<int>{in_dim, out_dim};
attrs["value"] = 1.0f;
init_tape.AddOp(initializer, {}, {{"Out", {w_}}}, attrs);
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
attrs["shape"] = std::vector<int>{out_dim};
attrs["value"] = 1.0f;
init_tape.AddOp(initializer, {}, {{"Out", {b_}}}, attrs);
init_tape.Forward();
}
VariableHandle operator()(VariableHandle input) {
VariableHandle pre_bias(new Variable("linear"));
get_global_tape().AddOp("mul",
{{"X", {input}}, {"Y", {w_}}},
{{"Out", {pre_bias}}},
{{"x_num_col_dims", 1}, {"y_num_col_dims", 1}});
VariableHandle pre_act(new Variable("linear"));
get_global_tape().AddOp("elementwise_add",
{{"X", {pre_bias}}, {"Y", {b_}}},
{{"Out", {pre_act}}},
{{"axis", 1}});
VariableHandle post_act(new Variable("linear"));
get_global_tape().AddOp(
act_, {{"X", {pre_act}}}, {{"Out", {post_act}}}, {});
return post_act;
}
std::vector<VariableHandle> Params() { return {w_, b_}; }
private:
VariableHandle w_;
VariableHandle b_;
std::string act_;
};
class SGD {
public:
SGD(float learning_rate) : learning_rate_(new Variable("sgd")) {
Tape init_tape;
std::string initializer = "fill_constant";
framework::AttributeMap attrs;
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
attrs["shape"] = std::vector<int>{1};
attrs["value"] = learning_rate;
init_tape.AddOp(initializer, {}, {{"Out", {learning_rate_}}}, attrs);
init_tape.Forward();
}
void operator()(VariableHandle input) {
PADDLE_ENFORCE(get_global_tape().HasBeenBackwarded(),
"optimization must happen after the backward");
Tape temp_tape;
temp_tape.AddOp("sgd",
{{"Param", {input}},
{"LearningRate", {learning_rate_}},
{"Grad", {input->Grad()}}},
{{"ParamOut", {input}}},
{});
temp_tape.Forward();
}
private:
VariableHandle learning_rate_;
};
}
}
paddle/contrib/tape/tape.cc 已删除 100644 → 0
浏览文件 @ a0c5fd83
// 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/contrib/tape/tape.h"
#include <list>
#include <map>
#include <memory>
#include <string>
#include <vector>
#include "paddle/fluid/framework/data_type.h"
#include "paddle/fluid/framework/dim.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/pybind/pybind.h"
namespace paddle {
namespace tape {
// borrowed from
// https://stackoverflow.com/questions/874134/find-if-string-ends-with-another-string-in-c
inline bool ends_with(std::string const &value, std::string const &ending) {
if (ending.size() > value.size()) return false;
return std::equal(ending.rbegin(), ending.rend(), value.rbegin());
}
std::ostream &operator<<(std::ostream &os, const framework::VarDesc &var_desc) {
os << var_desc.Name();
os << "[" << var_desc.GetType() << "]";
os << "[" << var_desc.GetDataType() << "]";
os << "{";
for (auto &i : var_desc.GetShape()) {
os << i << ",";
}
os << "}";
return os;
}
std::string to_string(const std::string &type,
const VariableHandleMap &in_vars,
const VariableHandleMap &out_vars,
const framework::AttributeMap &attrs) {
std::stringstream ss;
ss << type << " ";
for (auto &param_name : in_vars) {
for (auto &var : param_name.second) {
ss << param_name.first << ":(" << var->Desc() << ") ";
}
}
for (auto &param_name : out_vars) {
for (auto &var : param_name.second) {
ss << param_name.first << ":(" << var->Desc() << ") ";
}
}
return ss.str();
}
framework::OpDesc CreateOpDesc(const std::string &type,
const VariableHandleMap &in_vars,
const VariableHandleMap &out_vars,
const framework::AttributeMap &attrs) {
framework::VariableNameMap inputs;
for (auto &param_name : in_vars) {
for (auto &var : param_name.second) {
inputs[param_name.first].emplace_back(var->Name());
}
}
framework::VariableNameMap outputs;
for (auto &param_name : out_vars) {
for (auto &var : param_name.second) {
outputs[param_name.first].emplace_back(var->Name());
}
}
return framework::OpDesc(type, inputs, outputs, attrs);
}
void InferShapeAndVarType(const std::string &type,
const VariableHandleMap &in_vars,
VariableHandleMap *out_vars,
const framework::AttributeMap &attrs) {
framework::OpDesc op_desc = CreateOpDesc(type, in_vars, *out_vars, attrs);
// Create a temporary block for compile-time
framework::ProgramDesc program_desc;
framework::BlockDesc *block_desc = program_desc.MutableBlock(0);
PADDLE_ENFORCE(block_desc);
for (auto &param_name : in_vars) {
for (auto &var : param_name.second) {
*block_desc->Var(var->Name())->Proto() = *var->MutableDesc()->Proto();
}
}
for (auto &param_name : *out_vars) {
for (auto &var : param_name.second) {
*block_desc->Var(var->Name())->Proto() = *var->MutableDesc()->Proto();
}
}
LOG(INFO) << "- " << to_string(type, in_vars, *out_vars, attrs);
op_desc.InferShape(*block_desc);
op_desc.InferVarType(block_desc);
for (auto &param_name : *out_vars) {
for (auto &var : param_name.second) {
*var->MutableDesc()->Proto() = *block_desc->Var(var->Name())->Proto();
}
}
LOG(INFO) << "+ " << to_string(type, in_vars, *out_vars, attrs);
}
void Tape::AddOp(const std::string &type,
const VariableHandleMap &in_vars,
VariableHandleMap out_vars,
const framework::AttributeMap &attrs) {
InferShapeAndVarType(type, in_vars, &out_vars, attrs);
tape_.emplace_back(type, in_vars, out_vars, attrs);
}
// Temporary Scope for Operator::Run()
class ScopeWrapper : public framework::Scope {
public:
ScopeWrapper(const VariableHandleMap &in_vars,
const VariableHandleMap &out_vars) {
for (auto &v : in_vars) {
for (auto &vv : v.second) {
if (!vars_.count(vv->Name())) {
vars_[vv->Name()].reset(vv->Var());
}
}
}
for (auto &v : out_vars) {
for (auto &vv : v.second) {
if (!vars_.count(vv->Name())) {
vars_[vv->Name()].reset(vv->Var());
}
}
}
}
~ScopeWrapper() {
for (auto &pair : vars_) {
pair.second.release();
}
}
};
void Tape::Forward() {
LOG(INFO) << "Starting forward -------------------------";
PADDLE_ENFORCE(!has_been_backwarded_);
while (current_position_ < tape_.size()) {
OpHandle &op = tape_[current_position_];
// Create Output Tensor, this is only necessary for OpWithKernel
for (auto &param2var : op.outputs_) {
for (auto &var : param2var.second) {
var->InitializeVariable();
}
}
framework::OpDesc op_desc =
CreateOpDesc(op.type_, op.inputs_, op.outputs_, op.attrs_);
ScopeWrapper scope(op.inputs_, op.outputs_);
framework::OpRegistry::CreateOp(op_desc)->Run(scope, platform::CPUPlace());
current_position_++;
}
LOG(INFO) << "Finishing forward -------------------------";
}
void Tape::Backward(VariableHandle target) {
PADDLE_ENFORCE(!has_been_backwarded_);
Forward();
// TODO(tonyyang-svail): check output of last op is target
backward_tape_.reset(new Tape());
framework::AttributeMap attrs;
// FIXME(tonyyang-svail): Need to infer_data_type
attrs["dtype"] = framework::proto::VarType::Type::VarType_Type_FP32;
attrs["shape"] = std::vector<int>{1};
attrs["value"] = 1.0f;
backward_tape_->AddOp(
"fill_constant", {}, {{"Out", {target->Grad()}}}, attrs);
for (auto it = tape_.rbegin(); it != tape_.rend(); ++it) {
framework::OpDesc op_desc =
CreateOpDesc(it->type_, it->inputs_, it->outputs_, it->attrs_);
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, {}, &grad_to_var, {});
for (auto &op_desc : grad_op_descs) {
std::unordered_map<std::string, VariableHandle> name2var;
for (auto &param2vars : it->inputs_) {
for (auto &a : param2vars.second) {
name2var[a->Name()] = a;
}
}
for (auto &param2vars : it->outputs_) {
for (auto &a : param2vars.second) {
name2var[a->Name()] = a;
}
}
VariableHandleMap in_vars;
VariableHandleMap out_vars;
std::map<const framework::VariableNameMap *, VariableHandleMap *>
loop_over{{&op_desc->Inputs(), &in_vars},
{&op_desc->Outputs(), &out_vars}};
for (auto &each : loop_over) {
auto &vmp = *each.first;
auto &vhm = *each.second;
for (auto &p2a : vmp) {
for (auto &argu : p2a.second) {
if (name2var.count(argu)) {
vhm[p2a.first].push_back(name2var[argu]);
} else {
PADDLE_ENFORCE(ends_with(argu, framework::kGradVarSuffix),
argu.c_str());
std::string name = argu.substr(
0, argu.size() - std::strlen(framework::kGradVarSuffix));
PADDLE_ENFORCE(name2var.count(name), name.c_str());
vhm[p2a.first].push_back(name2var[name]->Grad());
}
}
}
}
backward_tape_->AddOp(
op_desc->Type(), in_vars, out_vars, op_desc->GetAttrMap());
}
// TODO(tonyyang-svail): how to fill empty grad?
// TODO(tonyyang-svail): Sum var grad is necessary
}
backward_tape_->Forward();
has_been_backwarded_ = true;
}
Tape &get_global_tape() {
static Tape T;
return T;
}
void reset_global_tape() { get_global_tape() = Tape(); }
}
}
paddle/contrib/tape/variable.h 已删除 100644 → 0
浏览文件 @ a0c5fd83
// 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 <memory>
#include "paddle/fluid/framework/operator.h" // framework::kGradVarSuffix
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/variable.h"
namespace paddle {
namespace tape {
class Variable;
using VariableHandle = std::shared_ptr<Variable>;
/*
* Combination of
* framework::VarDesc desc_;
* framework::Variable var_;
*/
class Variable {
public:
Variable(const std::string pre_fix)
: desc_(pre_fix + std::to_string(count())) {}
Variable(const std::string pre_fix, bool is_grad)
: desc_(pre_fix + (is_grad ? framework::kGradVarSuffix
: std::to_string(count()))) {}
~Variable() { LOG(INFO) << "Deleting " << Name(); }
// Instantiate LoDTensor/SelectedRow
void InitializeVariable();
VariableHandle Grad() {
if (grad_.expired()) {
VariableHandle new_grad(new Variable(desc_.Name(), true));
grad_ = new_grad;
return new_grad;
} else {
return VariableHandle(grad_);
}
}
// Stochastic Gradient Descent with Momentum
// VariableHandle Momentum ();
// void init(const std::string& initializer,
// const framework::AttributeMap& attrs);
// void value() {};
const framework::VarDesc& Desc() const { return desc_; }
framework::VarDesc* MutableDesc() { return &desc_; }
// TODO(tonyyang-svail): No need to expose name
std::string Name() const { return desc_.Name(); }
framework::Variable* Var() { return &var_; }
private:
int count() {
static int counter = 0;
return counter++;
}
framework::VarDesc desc_;
framework::Variable var_;
std::weak_ptr<Variable> grad_;
};
}
}
paddle/fluid/inference/analysis/CMakeLists.txt
浏览文件 @ f1f8327c
set(FLUID_CORE_MODULES proto_desc memory lod_tensor executor init)
cc_library(analysis SRCS dot.cc node.cc data_flow_graph.cc graph_traits.cc subgraph_splitter.cc fluid_to_data_flow_graph_pass.cc
DEPS paddle_fluid)
cc_library(analysis SRCS pass_manager.cc dot.cc node.cc data_flow_graph.cc graph_traits.cc subgraph_splitter.cc
fluid_to_data_flow_graph_pass.cc
data_flow_graph_to_fluid_pass.cc
tensorrt_subgraph_pass.cc
dfg_graphviz_draw_pass.cc
DEPS framework_proto)
cc_test(test_node SRCS node_tester.cc DEPS analysis)
cc_test(test_dot SRCS dot_tester.cc DEPS analysis)
set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests)
cc_test(test_data_flow_graph SRCS data_flow_graph_tester.cc DEPS analysis ${FLUID_CORE_MODULES} paddle_fluid
ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model)
set_tests_properties(test_data_flow_graph PROPERTIES DEPENDS test_word2vec)
function (inference_analysis_test TARGET)
set(options "")
set(oneValueArgs "")
set(multiValueArgs SRCS)
cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
cc_test(test_subgraph_splitter
SRCS subgraph_splitter_tester.cc
DEPS analysis paddle_fluid tensor
ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model)
set_tests_properties(test_subgraph_splitter PROPERTIES DEPENDS test_word2vec)
cc_test(${TARGET}
SRCS "${analysis_test_SRCS}"
DEPS analysis
ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model --fraction_of_gpu_memory_to_use=0.5)
set_tests_properties(${TARGET} PROPERTIES DEPENDS test_word2vec)
endfunction(inference_analysis_test)
cc_test(test_dfg_graphviz_draw_pass
SRCS dfg_graphviz_draw_pass_tester.cc
DEPS analysis
ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model)
set_tests_properties(test_dfg_graphviz_draw_pass PROPERTIES DEPENDS test_word2vec)
inference_analysis_test(test_data_flow_graph SRCS data_flow_graph_tester.cc)
inference_analysis_test(test_data_flow_graph_to_fluid_pass SRCS data_flow_graph_to_fluid_pass_tester.cc)
inference_analysis_test(test_fluid_to_data_flow_graph_pass SRCS fluid_to_data_flow_graph_pass_tester.cc)
inference_analysis_test(test_subgraph_splitter SRCS subgraph_splitter_tester.cc)
inference_analysis_test(test_dfg_graphviz_draw_pass SRCS dfg_graphviz_draw_pass_tester.cc)
#inference_analysis_test(test_tensorrt_subgraph_pass SRCS tensorrt_subgraph_pass_tester.cc)
inference_analysis_test(test_pass_manager SRCS pass_manager_tester.cc)
paddle/contrib/tape/variable.cc → paddle/fluid/inference/analysis/argument.cc
浏览文件 @ f1f8327c
......@@ -12,22 +12,4 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/contrib/tape/variable.h"
namespace paddle {
namespace tape {
void Variable::InitializeVariable() {
LOG(INFO) << "Initialzing " << desc_.Name() << " as " << desc_.GetType();
framework::proto::VarType::Type var_type = desc_.GetType();
if (var_type == framework::proto::VarType::LOD_TENSOR) {
var_.GetMutable<framework::LoDTensor>();
} else if (var_type == framework::proto::VarType::SELECTED_ROWS) {
var_.GetMutable<framework::SelectedRows>();
} else {
PADDLE_THROW("Variable type %d is not in [LOD_TENSOR, SELECTED_ROWS]",
var_type);
}
}
}
}
#include "paddle/fluid/inference/analysis/argument.h"
paddle/contrib/tape/tape.h → paddle/fluid/inference/analysis/argument.h
浏览文件 @ f1f8327c
......@@ -11,54 +11,45 @@
// 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 <memory>
#include <string>
#include <vector>
#include "paddle/contrib/tape/variable.h"
namespace paddle {
namespace tape {
using VariableHandleMap = std::map<std::string, std::vector<VariableHandle>>;
struct OpHandle {
OpHandle(const std::string &type,
const VariableHandleMap &in_vars,
const VariableHandleMap &out_vars,
const framework::AttributeMap &attrs)
: type_(type), inputs_(in_vars), outputs_(out_vars), attrs_(attrs) {}
/*
* This file defines the class Argument, which is the input and output of the
* analysis module. All the fields that needed either by Passes or PassManagers
* are contained in Argument.
*
* TODO(Superjomn) Find some way better to contain the fields when it grow too
* big.
*/
std::string type_;
VariableHandleMap inputs_;
VariableHandleMap outputs_;
framework::AttributeMap attrs_;
};
class Tape {
public:
void AddOp(const std::string &type,
const VariableHandleMap &in_vars,
VariableHandleMap out_vars,
const framework::AttributeMap &attrs);
void Forward();
void Backward(VariableHandle target);
bool HasBeenBackwarded() { return has_been_backwarded_; }
#pragma once
private:
bool has_been_backwarded_ = false;
size_t current_position_ = 0;
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/inference/analysis/data_flow_graph.h"
std::vector<OpHandle> tape_;
std::shared_ptr<Tape> backward_tape_;
namespace paddle {
namespace inference {
namespace analysis {
/*
* The argument definition of both Pass and PassManagers.
*
* All the fields should be registered here for clearness.
*/
struct Argument {
// The graph that process by the Passes or PassManagers.
std::unique_ptr<DataFlowGraph> main_dfg;
// The original program desc.
std::unique_ptr<framework::proto::ProgramDesc> origin_program_desc;
};
Tape &get_global_tape();
#define UNLIKELY(condition) __builtin_expect(static_cast<bool>(condition), 0)
#define ANALYSIS_ARGUMENT_CHECK_FIELD(field__) \
if (UNLIKELY(!(field__))) { \
LOG(ERROR) << "field " << #field__ << " should be set."; \
return false; \
}
void reset_global_tape();
}
}
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/inference/analysis/data_flow_graph.cc
浏览文件 @ f1f8327c
......@@ -14,6 +14,7 @@ limitations under the License. */
#include "paddle/fluid/inference/analysis/data_flow_graph.h"
#include "paddle/fluid/inference/analysis/dot.h"
#include "paddle/fluid/inference/analysis/node.h"
namespace paddle {
namespace inference {
......@@ -57,19 +58,7 @@ std::string DataFlowGraph::DotString() const {
// Add nodes
for (size_t i = 0; i < nodes.size(); i++) {
const Node &node = nodes.Get(i);
switch (node.type()) {
case Node::Type::kValue:
dot.AddNode(node.repr(), node.dot_attrs());
break;
case Node::Type::kFunction:
dot.AddNode(node.repr(), node.dot_attrs());
break;
case Node::Type::kFunctionBlock:
dot.AddNode(node.repr(), node.dot_attrs());
break;
default:
PADDLE_THROW("unsupported Node type %d", static_cast<int>(node.type()));
}
dot.AddNode(node.repr(), node.dot_attrs());
}
// Add edges
......
paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc 0 → 100644
浏览文件 @ f1f8327c
// 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/inference/analysis/data_flow_graph_to_fluid_pass.h"
#include "paddle/fluid/framework/proto_desc.h"
namespace paddle {
namespace inference {
namespace analysis {
bool DataFlowGraphToFluidPass::Initialize(Argument* argument) {
ANALYSIS_ARGUMENT_CHECK_FIELD(argument)
ANALYSIS_ARGUMENT_CHECK_FIELD(argument->origin_program_desc)
desc_ = argument->origin_program_desc.get();
// Here some logic from program_desc.cc and will not add new interfaces into
// framework::ProgramDesc class, use some UT to assure the correctness.
auto* block = desc_->mutable_blocks()->Add();
block->set_idx(framework::kRootBlockIndex);
block->set_parent_idx(framework::kNoneBlockIndex);
return true;
}
bool DataFlowGraphToFluidPass::Finalize() { return true; }
void DataFlowGraphToFluidPass::Run(DataFlowGraph* graph) {
auto traits = GraphTraits<DataFlowGraph>(graph);
for (auto it = traits.nodes().begin(); it != traits.nodes().end(); ++it) {
if (it->deleted()) continue;
switch (it->type()) {
case Node::Type::kFunction:
LOG(INFO) << "add function " << it->name();
AddFluidOp(&(*it));
break;
case Node::Type::kFunctionBlock:
AddEngineOp(&(*it));
break;
default:
continue;
}
}
}
void DataFlowGraphToFluidPass::AddFluidOp(Node* node) {
LOG(INFO) << "processing func " << node->name();
auto* ori_op = static_cast<framework::proto::OpDesc*>(node->pb_desc());
// currently only the main block is analyzed.
auto* main_block = desc_->mutable_blocks(framework::kRootBlockIndex);
auto* op = main_block->add_ops();
LOG(INFO) << "to copy the op";
*op = *ori_op; // copy the attributes, by default, these will not be changed
// by analysis phrase.
// The inputs and outputs of the existing ops are not changed by tensorrt
// subgraph pass.
// NOTE It might be changed by other passes in the long run.
}
void DataFlowGraphToFluidPass::AddEngineOp(Node* node) {
// auto* ori_op = static_cast<framework::proto::OpDesc*>(node->extra_info());
// auto* main_block = desc_->mutable_blocks(framework::kRootBlockIndex);
// auto* op = main_block->add_ops();
// TODO(Superjomn) Here need to expose some arguments for default setting.
}
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h 0 → 100644
浏览文件 @ f1f8327c
/* 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. */
/*
* This file implements the transformation from fluid ProgramDesc to data flow
* graph.
*/
#pragma once
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/inference/analysis/data_flow_graph.h"
#include "paddle/fluid/inference/analysis/pass.h"
namespace paddle {
namespace inference {
namespace analysis {
class DataFlowGraphToFluidPass final : public DataFlowGraphPass {
public:
DataFlowGraphToFluidPass() = default;
bool Initialize(Argument *argument) override;
bool Finalize() override;
void Run(DataFlowGraph *graph) override;
std::string repr() const override { return "DFG to fluid"; }
std::string description() const override {
return "Transform a DFG to a Fluid ProgramDesc";
}
Pass *CreatePrinterPass(std::ostream &os,
const std::string &banner) const override {
return nullptr;
}
protected:
// Add a Fluid Op into the ProgramDesc.
void AddFluidOp(Node *node);
// Add a EngineOp into the ProgramDesc.
void AddEngineOp(Node *node);
private:
framework::proto::ProgramDesc *desc_;
};
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc
浏览文件 @ f1f8327c
......@@ -27,13 +27,12 @@ namespace inference {
namespace analysis {
TEST_F(DFG_Tester, Test) {
framework::proto::ProgramDesc new_desc;
DataFlowGraph graph;
FluidToDataFlowGraphPass pass0;
DataFlowGraphToFluidPass pass1;
pass0.Initialize(desc);
pass1.Initialize(&new_desc);
ASSERT_TRUE(pass0.Initialize(&argument));
ASSERT_TRUE(pass1.Initialize(&argument));
pass0.Run(&graph);
pass1.Run(&graph);
......
paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.cc 0 → 100644
浏览文件 @ f1f8327c
/* 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/inference/analysis/dfg_graphviz_draw_pass.h"
namespace paddle {
namespace inference {
namespace analysis {
void DFG_GraphvizDrawPass::Run(DataFlowGraph *graph) {
auto content = Draw(graph);
std::ofstream file(GenDotPath());
file.write(content.c_str(), content.size());
file.close();
LOG(INFO) << "draw dot to " << GenDotPath();
}
std::string DFG_GraphvizDrawPass::Draw(DataFlowGraph *graph) {
Dot dot;
// Add nodes
for (size_t i = 0; i < graph->nodes.size(); i++) {
const Node &node = graph->nodes.Get(i);
if (config_.display_deleted_node || !node.deleted()) {
dot.AddNode(node.repr(), node.dot_attrs());
}
}
// Add edges
for (size_t i = 0; i < graph->nodes.size(); i++) {
const Node &node = graph->nodes.Get(i);
if (!config_.display_deleted_node && node.deleted()) continue;
for (auto &in : node.inlinks) {
if (!config_.display_deleted_node && in->deleted()) continue;
for (auto &in : node.inlinks) {
dot.AddEdge(in->repr(), node.repr(), {});
}
}
}
return dot.Build();
}
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h
浏览文件 @ f1f8327c
......@@ -21,6 +21,7 @@ limitations under the License. */
#include <fstream>
#include <string>
#include "paddle/fluid/inference/analysis/dot.h"
#include "paddle/fluid/inference/analysis/pass.h"
namespace paddle {
......@@ -32,35 +33,39 @@ namespace analysis {
*/
class DFG_GraphvizDrawPass : public DataFlowGraphPass {
public:
DFG_GraphvizDrawPass(const std::string& dir, const std::string& id)
: dir_(dir), id_(id) {}
bool Initialize() override { return Pass::Initialize(); }
void Run(DataFlowGraph* graph) override {
auto content = Draw(graph);
std::ofstream file(GenDotPath());
file.write(content.c_str(), content.size());
file.close();
LOG(INFO) << "draw dot to " << GenDotPath();
}
struct Config {
Config(const std::string &dir, const std::string &id,
bool display_deleted_node = false)
: dir(dir), id(id), display_deleted_node(display_deleted_node) {}
// The directory to store the .dot or .png files.
const std::string dir;
// The identifier for this dot file.
const std::string id;
// Whether to display deleted nodes, default false.
const bool display_deleted_node;
};
DFG_GraphvizDrawPass(const Config &config) : config_(config) {}
bool Initialize(Argument *argument) override { return true; }
void Run(DataFlowGraph *graph) override;
bool Finalize() override { return Pass::Finalize(); }
Pass* CreatePrinterPass(std::ostream& os,
const std::string& banner) const override {
return nullptr;
std::string repr() const override { return "DFG graphviz drawer"; }
std::string description() const override {
return "Debug a DFG by draw with graphviz";
}
private:
// Path of the dot file to output.
std::string GenDotPath() const {
return dir_ + "/" + "graph_" + id_ + ".dot";
return config_.dir + "/" + "graph_" + config_.id + ".dot";
}
std::string Draw(DataFlowGraph* graph) { return graph->DotString(); }
std::string Draw(DataFlowGraph *graph);
std::string dir_;
std::string id_;
Config config_;
};
} // namespace analysis
......
paddle/fluid/inference/analysis/dfg_graphviz_draw_pass_tester.cc
浏览文件 @ f1f8327c
......@@ -24,9 +24,10 @@ namespace inference {
namespace analysis {
TEST_F(DFG_Tester, dfg_graphviz_draw_pass_tester) {
auto dfg = ProgramDescToDFG(desc);
DFG_GraphvizDrawPass pass("./", "test");
pass.Initialize();
auto dfg = ProgramDescToDFG(*argument.origin_program_desc);
DFG_GraphvizDrawPass::Config config("./", "test");
DFG_GraphvizDrawPass pass(config);
pass.Initialize(&argument);
pass.Run(&dfg);
// test content
......@@ -38,7 +39,8 @@ TEST_F(DFG_Tester, dfg_graphviz_draw_pass_tester) {
while (std::getline(file, line)) {
no++;
}
ASSERT_EQ(no, 82);
// DFG is sensitive to ProgramDesc, be careful to change the existing models.
ASSERT_EQ(no, 112);
}
} // namespace analysis
......
paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc
浏览文件 @ f1f8327c
......@@ -21,19 +21,23 @@ namespace paddle {
namespace inference {
namespace analysis {
FluidToDataFlowGraphPass::FluidToDataFlowGraphPass() {}
bool FluidToDataFlowGraphPass::Initialize() { return Pass::Initialize(); }
bool FluidToDataFlowGraphPass::Initialize(
const framework::proto::ProgramDesc &desc) {
desc_ = &desc;
bool FluidToDataFlowGraphPass::Initialize(Argument *argument) {
ANALYSIS_ARGUMENT_CHECK_FIELD(argument);
ANALYSIS_ARGUMENT_CHECK_FIELD(argument->origin_program_desc);
PADDLE_ENFORCE(argument);
if (!argument->main_dfg) {
LOG(INFO) << "Init DFG";
argument->main_dfg.reset(new DataFlowGraph);
}
desc_ = argument->origin_program_desc.get();
return true;
}
bool FluidToDataFlowGraphPass::Finalize() { return Pass::Finalize(); }
void FluidToDataFlowGraphPass::Run(DataFlowGraph *graph) {
PADDLE_ENFORCE(graph);
PADDLE_ENFORCE(desc_);
// insert vars
std::unordered_map<std::string, size_t> var2id;
auto &main_block = desc_->blocks(framework::kRootBlockIndex);
......@@ -41,7 +45,7 @@ void FluidToDataFlowGraphPass::Run(DataFlowGraph *graph) {
const auto &var = main_block.vars(i);
auto *v = graph->nodes.Create(Node::Type::kValue);
v->SetName(var.name());
v->SetExtraInfo(const_cast<void *>(static_cast<const void *>(&var)));
v->SetPbDesc(const_cast<void *>(static_cast<const void *>(&var)));
var2id[var.name()] = v->id();
}
for (int i = 0; i < main_block.ops_size(); i++) {
......@@ -51,7 +55,7 @@ void FluidToDataFlowGraphPass::Run(DataFlowGraph *graph) {
static_cast<Function *>(o)->SetFuncType(op.type());
// Link to the original protobuf message's memory, make it easier to
// generate from a data flow graph to fluid ProgramDesc.
o->SetExtraInfo(const_cast<void *>(static_cast<const void *>(&op)));
o->SetPbDesc(const_cast<void *>(static_cast<const void *>(&op)));
// set inputs and outputs
// TODO(Superjomn) make sure the InputNames is the real variable name.
for (int j = 0; j < op.inputs_size(); j++) {
......
paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h
浏览文件 @ f1f8327c
......@@ -34,13 +34,18 @@ namespace analysis {
*/
class FluidToDataFlowGraphPass final : public DataFlowGraphPass {
public:
FluidToDataFlowGraphPass();
bool Initialize() override;
bool Initialize(const framework::proto::ProgramDesc &desc) override;
FluidToDataFlowGraphPass() = default;
bool Initialize(Argument *argument) override;
bool Finalize() override;
void Run(DataFlowGraph *graph) override;
std::string repr() const override { return "fluid-to-data-flow-graph"; }
std::string description() const override {
return "transform a fluid ProgramDesc to a data flow graph.";
}
Pass *CreatePrinterPass(std::ostream &os,
const std::string &banner) const override;
......
paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc
浏览文件 @ f1f8327c
......@@ -23,11 +23,11 @@ namespace analysis {
TEST_F(DFG_Tester, Init) {
FluidToDataFlowGraphPass pass;
pass.Initialize();
pass.Initialize(desc);
pass.Initialize(&argument);
DataFlowGraph graph;
pass.Run(&graph);
ASSERT_GT(graph.nodes.size(), 0);
// Analysis is sensitive to ProgramDesc, careful to change the original model.
ASSERT_EQ(graph.nodes.size(), 37);
pass.Finalize();
LOG(INFO) << '\n' << graph.DotString();
}
......
paddle/fluid/inference/analysis/helper.h
浏览文件 @ f1f8327c
......@@ -62,6 +62,7 @@ struct DataTypeNamer {
SET_TYPE(int);
SET_TYPE(bool);
SET_TYPE(float);
SET_TYPE(void *);
}
std::unordered_map<decltype(typeid(int).hash_code()), // NOLINT
......
paddle/fluid/inference/analysis/node.cc
浏览文件 @ f1f8327c
......@@ -40,6 +40,9 @@ Node *NodeMap::Create(Node::Type type) {
case Node::Type::kValue:
nodes_.emplace_back(new Value);
break;
case Node::Type::kFunctionBlock:
nodes_.emplace_back(new FunctionBlock);
break;
default:
PADDLE_THROW("Not supported node type.");
}
......
paddle/fluid/inference/analysis/node.h
浏览文件 @ f1f8327c
......@@ -71,12 +71,17 @@ class Node {
// Get an additional attribute and convert it to T data type. NOTE this will
// silently create a new attribute if not exists.
Attr &attr(const std::string &name) { return attrs_[name]; }
Attr &attr(const std::string &name) const { return attrs_[name]; }
int id() const { return id_; }
bool deleted() const { return deleted_; }
// The Protobuf description is set/get with a void* to decouple Node interface
// from a specific kind of Protobuf message.
void SetPbDesc(void *pb) { attr("pb_desc").Pointer() = pb; }
void *pb_desc() const { return attr("pb_desc").Pointer(); }
void SetDeleted() { deleted_ = true; }
bool deleted() const { return deleted_; }
void SetName(const std::string &name) { name_ = name; }
const std::string &name() const { return name_; }
......@@ -84,29 +89,25 @@ class Node {
void SetType(Type type) { type_ = type; }
Type type() const { return type_; }
void *extra_info() const { return extra_info_; }
void SetExtraInfo(void *extra_info) { extra_info_ = extra_info; }
// Input links.
std::vector<Node *> inlinks;
// Output links.
std::vector<Node *> outlinks;
// A helper class to maintain the status from Pass.
// TODO(superjomn) add a checker here to ensure the T is primary.
struct Attr {
// NOTE T should be a primary type or a struct combined by several primary
// types.
// NOTE the STL containers should not use here.
// Some usages
// Attr attr;
// T data;
// attr.data.assign((char*)data, sizeof(data));
// Attr attr;
// attr.Bool() = true;
bool &Bool() { return As<bool>(); }
float &Float() { return As<float>(); }
int32_t &Int32() { return As<int32_t>(); }
int64_t &Int64() { return As<int64_t>(); }
void *&Pointer() { return As<void *>(); }
private:
template <typename T>
......@@ -130,6 +131,7 @@ class Node {
size_t type_hash_{std::numeric_limits<size_t>::max()};
};
// Type checks.
bool IsFunction() const { return type_ == Node::Type::kFunction; }
bool IsValue() const { return type_ == Node::Type::kValue; }
bool IsFunctionBlock() const { return type_ == Node::Type::kFunctionBlock; }
......@@ -148,9 +150,6 @@ class Node {
Type type_{Type::kNone};
// Mark this node is deleted by some pass.
bool deleted_{false};
void *extra_info_;
mutable std::unordered_map<std::string, Attr> attrs_;
};
......
paddle/fluid/inference/analysis/pass.h
浏览文件 @ f1f8327c
......@@ -19,6 +19,7 @@ limitations under the License. */
#include <string>
#include "paddle/fluid/framework/framework.pb.h"
#include "paddle/fluid/inference/analysis/argument.h"
#include "paddle/fluid/inference/analysis/data_flow_graph.h"
#include "paddle/fluid/inference/analysis/helper.h"
#include "paddle/fluid/inference/analysis/node.h"
......@@ -30,19 +31,24 @@ namespace analysis {
class Pass {
public:
Pass() = default;
virtual ~Pass() {}
virtual ~Pass() = default;
// Virtual method overridden by subclasses to do only necessary initialization
// before any pass is run.
virtual bool Initialize() { return false; }
// virtual bool Initialize() { return false; }
// There is some passes such as FlowToDataFlowGraphPass that needs a
// ProgramDesc. Here use the native ProgramDesc ProtoBuf message, so that it
// only couple with the proto file.
virtual bool Initialize(const framework::proto::ProgramDesc &desc) {
return false;
}
// virtual bool Initialize(const framework::proto::ProgramDesc &desc) { return
// false; }
// There are some Passes such as DataFlowGraphToFluidPass that will output a
// ProgramDesc.
virtual bool Initialize(framework::proto::ProgramDesc *desc) { return false; }
// virtual bool Initialize(framework::proto::ProgramDesc *desc) { return
// false; }
// Mutable Pass.
virtual bool Initialize(Argument *argument) { return false; }
// Readonly Pass.
virtual bool Initialize(const Argument &argument) { return false; }
// Virtual method overriden by subclasses to do any necessary clean up after
// all passes have run.
......@@ -50,7 +56,9 @@ class Pass {
// Get a Pass appropriate to print the Node this pass operates on.
virtual Pass *CreatePrinterPass(std::ostream &os,
const std::string &banner) const = 0;
const std::string &banner) const {
return nullptr;
}
// Run on a single Node.
virtual void Run(Node *x) { LOG(FATAL) << "not valid"; }
......@@ -60,6 +68,11 @@ class Pass {
virtual void Run(FunctionBlock *x) { LOG(FATAL) << "not valid"; }
// Run on a single DataFlowGraph.
virtual void Run(DataFlowGraph *x) { LOG(FATAL) << "not valid"; }
// Human-readable short representation.
virtual std::string repr() const = 0;
// Human-readable long description.
virtual std::string description() const = 0;
};
// NodePass process on any Node types.
......
paddle/fluid/inference/analysis/pass_manager.cc 0 → 100644
浏览文件 @ f1f8327c
/* 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/inference/analysis/pass_manager.h"
#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h"
namespace paddle {
namespace inference {
namespace analysis {
void DfgPassManager::RunAll() {
PADDLE_ENFORCE(argument_);
for (auto& pass : data_) {
VLOG(4) << "Running pass [" << pass->repr() << "]";
pass->Run(argument_->main_dfg.get());
}
}
void NodePassManager::RunAll() {
PADDLE_ENFORCE(argument_);
PADDLE_ENFORCE(argument_->main_dfg.get());
auto trait =
GraphTraits<DataFlowGraph>(argument_->main_dfg.get()).nodes_in_DFS();
for (auto& node : trait) {
for (auto& pass : data_) {
pass->Run(&node);
}
}
}
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/inference/analysis/pass_manager.h 0 → 100644
浏览文件 @ f1f8327c
/* 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. */
/*
* This file defines the logic of pass management. The analysis for inference is
* a pipeline of Passes, a PassManager is a agency that helps to manage the
* executation of the Passes.
*
* There are two modes of Passes, the first one is called NodePass and takes
* an Node as input and output; the second one is called DFGPass and takes a
* DFG(Data Flow Graph) as input and output. It is hard to put all the passes in
* the same pipeline, there are two kinds of PassManagers, both takes a DFG as
* input and output a DFG, but the Passes inside are different:
*
* 1. NodePassManager: the passes inside are all NodePasses, it can have
* different graph trivial algorithm, for example, DFS_NodePassManager will
* trigger the passes in depth first order;
* 2. DfgPassManager: the passes inside are all DfgPasses.
*/
#pragma once
#include <string>
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/inference/analysis/pass.h"
namespace paddle {
namespace inference {
namespace analysis {
/*
* PassManager is the base class for all pass managers, a pass manager has
* several Pass-es registered, and execute them in the linear order.
*/
class PassManager : public OrderedRegistry<Pass> {
public:
PassManager() = default;
// Call all the passes' Initialize methods. The desc and data_flow_graph are
// globally shared, so pass them as the arguemnts for all the pass managers.
virtual bool Initialize(const Argument& argument) { return false; }
virtual bool Initialize(Argument* argument) {
argument_ = argument;
for (auto& pass : data_) {
LOG(INFO) << "Initializing pass " << pass->repr();
if (!pass->Initialize(argument)) {
LOG(ERROR) << "Failed to initialize pass [" << pass->repr() << "]";
return false;
}
}
return true;
}
// Call all the passes' Finalize methods.
virtual bool Finalize() {
for (auto& pass : data_) {
if (!pass->Finalize()) {
LOG(ERROR) << "Failed to finalize pass [" << pass->repr() << "]";
return false;
}
}
return true;
}
// Run all the passes.
virtual void RunAll() = 0;
// Short identifier.
virtual std::string repr() const = 0;
// Long description.
virtual std::string description() const = 0;
virtual ~PassManager() = default;
protected:
Argument* argument_{nullptr};
};
/*
* A pass manager that process a DFG.
*/
class DfgPassManager : public PassManager {
public:
DfgPassManager() = default;
void RunAll() override;
virtual ~DfgPassManager() = default;
};
/*
* A pass manager that process a Node each time.
*/
class NodePassManager : public PassManager {
public:
NodePassManager() = default;
void RunAll() override;
virtual ~NodePassManager() = default;
};
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/inference/analysis/pass_manager_tester.cc 0 → 100644
浏览文件 @ f1f8327c
/* 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/inference/analysis/pass_manager.h"
#include "paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h"
#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h"
#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h"
#include "paddle/fluid/inference/analysis/ut_helper.h"
#include <gtest/gtest.h>
namespace paddle {
namespace inference {
namespace analysis {
class TestDfgPassManager final : public DfgPassManager {
public:
TestDfgPassManager() = default;
virtual ~TestDfgPassManager() = default;
// Short identifier.
std::string repr() const override { return "test-pass-manager"; }
// Long description.
std::string description() const override { return "test doc"; }
};
class TestNodePassManager final : public NodePassManager {
public:
virtual ~TestNodePassManager() = default;
std::string repr() const override { return "test-node-pass-manager"; }
std::string description() const override { return "test doc"; }
};
class TestNodePass final : public NodePass {
public:
virtual ~TestNodePass() = default;
bool Initialize(Argument* argument) override { return true; }
void Run(Node* node) override {
LOG(INFO) << "- Processing node " << node->repr();
}
std::string repr() const override { return "test-node"; }
std::string description() const override { return "some doc"; }
};
TEST_F(DFG_Tester, DFG_pass_manager) {
TestDfgPassManager manager;
DFG_GraphvizDrawPass::Config config("./", "dfg.dot");
manager.Register("fluid-to-flow-graph", new FluidToDataFlowGraphPass);
manager.Register("graphviz", new DFG_GraphvizDrawPass(config));
manager.Register("dfg-to-fluid", new DataFlowGraphToFluidPass);
ASSERT_TRUE(manager.Initialize(&argument));
manager.RunAll();
}
TEST_F(DFG_Tester, Node_pass_manager) {
// Pre-process: initialize the DFG with the ProgramDesc first.
FluidToDataFlowGraphPass pass0;
pass0.Initialize(&argument);
pass0.Run(argument.main_dfg.get());
TestNodePassManager manager;
manager.Register("test-node-pass", new TestNodePass);
ASSERT_TRUE(manager.Initialize(&argument));
manager.RunAll();
}
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/inference/analysis/subgraph_splitter_tester.cc
浏览文件 @ f1f8327c
......@@ -19,22 +19,23 @@ namespace paddle {
namespace inference {
namespace analysis {
SubGraphSplitter::NodeInsideSubgraphTeller teller = [](const Node* node) {
if (node->type() != Node::Type::kFunction) return false;
const auto* func = static_cast<const Function*>(node);
if (func->func_type() == "elementwise_add" || func->func_type() == "relu" ||
func->func_type() == "conv2d" || func->func_type() == "mul" ||
func->func_type() == "sigmoid" || func->func_type() == "softmax") {
LOG(INFO) << "sub-graph marked " << node->repr();
return true;
}
return false;
};
TEST_F(DFG_Tester, Split) {
auto desc = LoadProgramDesc();
auto dfg = ProgramDescToDFG(desc);
LOG(INFO) << "spliter\n" << dfg.DotString();
SubGraphSplitter::NodeInsideSubgraphTeller teller = [](const Node* node) {
if (node->type() != Node::Type::kFunction) return false;
const auto* func = static_cast<const Function*>(node);
if (func->func_type() == "elementwise_add" || func->func_type() == "relu" ||
func->func_type() == "conv2d" || func->func_type() == "mul" ||
func->func_type() == "sigmoid" || func->func_type() == "softmax") {
LOG(INFO) << "sub-graph marked " << node->repr();
return true;
}
return false;
};
ASSERT_GT(dfg.nodes.size(), 5UL);
auto subgraphs = SubGraphSplitter(&dfg, teller)();
......@@ -62,6 +63,28 @@ TEST_F(DFG_Tester, Split) {
ASSERT_EQ(subgraphs.back().size(), 6UL);
}
TEST_F(DFG_Tester, Fuse) {
auto desc = LoadProgramDesc();
auto dfg = ProgramDescToDFG(desc);
size_t count0 = dfg.nodes.size();
SubGraphFuse fuse(&dfg, teller);
fuse();
int count1 = 0;
for (auto& node : dfg.nodes.nodes()) {
if (node->deleted()) {
LOG(INFO) << "deleted " << node->repr();
}
count1 += node->deleted();
}
// At least one nodes should be deleted.
ASSERT_EQ(dfg.nodes.size(), count0 + 1); // added a new FunctionBlock
ASSERT_EQ(6UL, count1);
}
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/contrib/tape/test_tape.cc → paddle/fluid/inference/analysis/tensorrt_subgraph_pass.cc
浏览文件 @ f1f8327c
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
// 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.
......@@ -12,50 +12,22 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include "gtest/gtest.h"
#include "paddle/contrib/tape/function.h"
#include "paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h"
#include "paddle/fluid/inference/analysis/subgraph_splitter.h"
using namespace paddle::tape;
namespace paddle {
namespace inference {
namespace analysis {
TEST(Tape, TestMLP) {
LOG(INFO) << "TestMLP";
Linear linear1(3, 3, "relu");
Linear linear2(3, 3, "relu");
Mean mean;
TensorRTSubGraphPass::TensorRTSubGraphPass(
const TensorRTSubGraphPass::NodeInsideSubgraphTeller &teller)
: node_inside_subgraph_teller_(teller) {}
SGD sgd(0.001);
std::string initializer = "fill_constant";
paddle::framework::AttributeMap attrs;
attrs["dtype"] = paddle::framework::proto::VarType::Type::VarType_Type_FP32;
attrs["shape"] = std::vector<int>{3, 3};
attrs["value"] = 1.0f;
Fill filler(initializer, attrs);
for (int i = 0; i < 2; ++i) {
reset_global_tape();
VariableHandle input(new Variable("input"));
filler(input);
auto loss = mean(linear2(linear1(input)));
get_global_tape().Backward(loss);
for (auto w : linear1.Params()) {
sgd(w);
}
for (auto w : linear2.Params()) {
sgd(w);
}
}
void TensorRTSubGraphPass::Run(DataFlowGraph *graph) {
SubGraphFuse(graph, node_inside_subgraph_teller_);
}
int main(int argc, char** argv) {
std::vector<paddle::platform::Place> places;
places.emplace_back(paddle::platform::CPUPlace());
paddle::platform::DeviceContextPool::Init(places);
} // analysis
} // inference
testing::InitGoogleTest(&argc, argv);
return RUN_ALL_TESTS();
}
} // paddle
paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h 0 → 100644
浏览文件 @ f1f8327c
/* 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 "paddle/fluid/inference/analysis/node.h"
#include "paddle/fluid/inference/analysis/pass.h"
#include "paddle/fluid/inference/analysis/subgraph_splitter.h"
namespace paddle {
namespace inference {
namespace analysis {
/*
* Parse the graph and replace TensorRT supported nodes with SubGraphNode
*/
class TensorRTSubGraphPass : public DataFlowGraphPass {
public:
// Tell whether to transform a sub-graph into TensorRT.
using NodeInsideSubgraphTeller = SubGraphFuse::NodeInsideSubgraphTeller;
TensorRTSubGraphPass(const NodeInsideSubgraphTeller& teller);
bool Initialize(Argument* argument) override { return true; }
// This class get a sub-graph as input and determine whether to transform this
// sub-graph into TensorRT.
void Run(DataFlowGraph* graph) override;
private:
NodeInsideSubgraphTeller node_inside_subgraph_teller_;
};
} // namespace analysis
} // namespace inference
} // paddle
paddle/fluid/inference/analysis/tensorrt_subgraph_pass_tester.cc 0 → 100644
浏览文件 @ f1f8327c
/* 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/inference/analysis/tensorrt_subgraph_pass.h"
#include <gflags/gflags.h>
#include <gtest/gtest.h>
#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h"
#include "paddle/fluid/inference/analysis/ut_helper.h"
namespace paddle {
namespace inference {
namespace analysis {
DEFINE_string(model_dir, "", "inference test model dir");
TEST(TensorRTSubGraph, single_pass) {
auto desc = LoadProgramDesc();
auto dfg = ProgramDescToDFG(desc);
SubGraphSplitter::NodeInsideSubgraphTeller teller = [](const Node* node) {
if (node->type() != Node::Type::kFunction) return false;
const auto* func = static_cast<const Function*>(node);
if (func->func_type() == "elementwise_add" || func->func_type() == "relu" ||
func->func_type() == "conv2d" || func->func_type() == "mul" ||
func->func_type() == "sigmoid" || func->func_type() == "softmax") {
LOG(INFO) << "sub-graph marked " << node->repr();
return true;
}
return false;
};
DFG_GraphvizDrawPass::Config config{"./", "test"};
DFG_GraphvizDrawPass dfg_pass(config);
dfg_pass.Initialize();
DFG_GraphvizDrawPass dfg_pass1(config);
dfg_pass1.Initialize();
dfg_pass.Run(&dfg);
TensorRTSubGraphPass trt_pass(std::move(teller));
trt_pass.Initialize();
trt_pass.Run(&dfg);
dfg_pass1.Run(&dfg);
// Check the TRT op's block desc
for (auto node : dfg.nodes.nodes()) {
if (node->IsFunctionBlock()) {
}
}
}
TEST(TensorRTSubGraph, pass_manager) {}
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/inference/analysis/ut_helper.h
浏览文件 @ f1f8327c
......@@ -15,33 +15,46 @@ limitations under the License. */
#pragma once
#include <gflags/gflags.h>
#include <gtest/gtest.h>
#include <fstream>
#include <string>
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/inference/analysis/data_flow_graph.h"
#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h"
#include "paddle/fluid/inference/analysis/ut_helper.h"
#include "paddle/fluid/inference/io.h"
namespace paddle {
namespace inference {
// Read ProgramDesc from a __model__ file, defined in io.cc
extern void ReadBinaryFile(const std::string& filename, std::string* contents);
namespace analysis {
DEFINE_string(inference_model_dir, "", "inference test model dir");
static framework::proto::ProgramDesc LoadProgramDesc(
const std::string& model_dir = FLAGS_inference_model_dir) {
paddle::platform::CPUPlace place;
paddle::framework::Executor executor(place);
paddle::framework::Scope scope;
auto program = Load(&executor, &scope, model_dir);
return *program->Proto();
std::string msg;
std::string net_file = FLAGS_inference_model_dir + "/__model__";
std::ifstream fin(net_file, std::ios::in | std::ios::binary);
PADDLE_ENFORCE(static_cast<bool>(fin), "Cannot open file %s", net_file);
fin.seekg(0, std::ios::end);
msg.resize(fin.tellg());
fin.seekg(0, std::ios::beg);
fin.read(&(msg.at(0)), msg.size());
fin.close();
framework::proto::ProgramDesc program_desc;
program_desc.ParseFromString(msg);
return program_desc;
}
static DataFlowGraph ProgramDescToDFG(
const framework::proto::ProgramDesc& desc) {
DataFlowGraph graph;
FluidToDataFlowGraphPass pass;
pass.Initialize(desc);
Argument argument;
argument.origin_program_desc.reset(new framework::proto::ProgramDesc(desc));
pass.Initialize(&argument);
pass.Run(&graph);
pass.Finalize();
return graph;
......@@ -49,9 +62,12 @@ static DataFlowGraph ProgramDescToDFG(
class DFG_Tester : public ::testing::Test {
protected:
void SetUp() override { desc = LoadProgramDesc(FLAGS_inference_model_dir); }
void SetUp() override {
auto desc = LoadProgramDesc(FLAGS_inference_model_dir);
argument.origin_program_desc.reset(new framework::proto::ProgramDesc(desc));
}
framework::proto::ProgramDesc desc;
Argument argument;
};
} // namespace analysis
......
paddle/fluid/memory/detail/system_allocator.cc
浏览文件 @ f1f8327c
......@@ -43,14 +43,16 @@ void* CPUAllocator::Alloc(size_t* index, size_t size) {
*index = 0; // unlock memory
void* p;
void* p = nullptr;
#ifdef PADDLE_WITH_MKLDNN
// refer to https://github.com/01org/mkl-dnn/blob/master/include/mkldnn.hpp
// memory alignment
PADDLE_ENFORCE_EQ(posix_memalign(&p, 4096ul, size), 0);
PADDLE_ENFORCE_EQ(posix_memalign(&p, 4096ul, size), 0, "Alloc %ld error!",
size);
#else
PADDLE_ENFORCE_EQ(posix_memalign(&p, 32ul, size), 0);
PADDLE_ENFORCE_EQ(posix_memalign(&p, 32ul, size), 0, "Alloc %ld error!",
size);
#endif
PADDLE_ENFORCE(p, "Fail to allocate CPU memory: size = %d .", size);
......
paddle/fluid/operators/activation_mkldnn_op.cc
浏览文件 @ f1f8327c
......@@ -12,16 +12,20 @@
See the License for the specific language governing permissions and
limitations under the License. */
#include "mkldnn.hpp"
#include "paddle/fluid/operators/activation_op.h"
#include "paddle/fluid/operators/mkldnn_activation_op.h"
#include "paddle/fluid/platform/mkldnn_helper.h"
namespace paddle {
namespace operators {
using paddle::framework::Tensor;
using paddle::platform::MKLDNNDeviceContext;
using framework::DataLayout;
using framework::Tensor;
using mkldnn::memory;
using mkldnn::primitive;
using mkldnn::stream;
using platform::GetMKLDNNFormat;
using platform::MKLDNNDeviceContext;
using platform::to_void_cast;
namespace {
std::string gethash(const mkldnn::memory::dims &operand_dims,
......@@ -35,188 +39,260 @@ std::string gethash(const mkldnn::memory::dims &operand_dims,
};
return dim2str(operand_dims) + std::to_string(algorithm);
}
} // namespace
template <typename Functor>
class MKLDNNActivationKernel
: public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
const auto *x = ctx.Input<Tensor>("X");
PADDLE_ENFORCE(x->layout() == DataLayout::kMKLDNN &&
x->format() != memory::format::format_undef,
"Wrong layout/format set for Input x tensor");
Functor functor;
auto attrs = functor.GetAttrs();
for (auto &attr : attrs) {
*attr.second = ctx.Attr<float>(attr.first);
}
functor(ctx);
}
};
template <typename T, typename ExecContext>
void eltwise_forward(const ExecContext &ctx, mkldnn::algorithm algorithm,
const T alpha = 0, const T beta = 0) {
template <typename Functor>
class MKLDNNActivationGradKernel
: public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
public:
void Compute(const framework::ExecutionContext &ctx) const override {
const auto *diff_y = ctx.Input<Tensor>(framework::GradVarName("Out"));
PADDLE_ENFORCE(diff_y->layout() == DataLayout::kMKLDNN &&
diff_y->format() != memory::format::format_undef,
"Wrong layout/format set for Input OutGrad tensor");
Functor functor;
auto attrs = functor.GetAttrs();
for (auto &attr : attrs) {
*attr.second = ctx.Attr<float>(attr.first);
}
functor(ctx);
}
};
template <typename T>
void eltwise_forward(const framework::ExecutionContext &ctx,
mkldnn::algorithm algorithm, const T alpha = 0,
const T beta = 0) {
PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
"It must use CPUPlace.");
auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
const auto &mkldnn_engine = dev_ctx.GetEngine();
// get buffers
const auto *src = ctx.template Input<Tensor>("X");
const auto *src_data = src->template data<T>();
const auto *x = ctx.Input<Tensor>("X");
auto *y = ctx.Output<Tensor>("Out");
auto *dst = ctx.template Output<Tensor>("Out");
T *dst_data = dst->template mutable_data<T>(ctx.GetPlace());
const T *x_data = x->data<T>();
T *y_data = y->mutable_data<T>(ctx.GetPlace());
// get memory dim
PADDLE_ENFORCE(src->dims().size() == 2 || src->dims().size() == 4,
PADDLE_ENFORCE(x->dims().size() == 2 || x->dims().size() == 4,
"Input dim must be with 2 or 4");
std::vector<int> src_tz = framework::vectorize2int(src->dims());
std::vector<int> src_tz = framework::vectorize2int(x->dims());
auto src_format =
src_tz.size() == 2 ? mkldnn::memory::format::nc : x->format();
const std::string key = gethash(src_tz, algorithm);
const std::string key_src_data =
key + ctx.op().Output("Out") + "@eltwise_fwd_src_data";
const std::string key_src_mem = key + "@eltwise_fwd_src_mem";
const std::string key_dst_mem = key + "@eltwise_fwd_dst_mem";
const std::string key_fwd = key + "@eltwise_fwd";
const std::string key_src_layout =
key + ctx.op().Output("Out") + "@eltwise_fwd_src_layout";
const std::string key_with_layout = key + std::to_string(src_format);
const std::string key_src_mem = key_with_layout + "@eltwise_fwd_src_mem";
const std::string key_dst_mem = key_with_layout + "@eltwise_fwd_dst_mem";
const std::string key_fwd = key_with_layout + "@eltwise_fwd";
const std::string key_fwd_pd = key_with_layout + "@eltwise_fwd_pd";
// save input data and layout to be referred in backward path
auto p_src_data = std::make_shared<const T *>(x_data);
dev_ctx.SetBlob(key_src_data, p_src_data);
auto p_src_layout = std::make_shared<memory::format>(src_format);
dev_ctx.SetBlob(key_src_layout, p_src_layout);
auto p_fwd = std::static_pointer_cast<mkldnn::eltwise_forward>(
dev_ctx.GetBlob(key_fwd));
// save input data to be referred in backward path
auto p_src_data = std::make_shared<const T *>(src_data);
dev_ctx.SetBlob(key_src_data, p_src_data);
std::shared_ptr<memory> dst_memory;
if (p_fwd == nullptr) {
// create memory description
auto data_md = src_tz.size() == 2
? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
mkldnn::memory::format::nc)
: platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
mkldnn::memory::format::nchw);
// create memory primitives
auto p_src_mem = std::make_shared<mkldnn::memory>(mkldnn::memory(
{data_md, mkldnn_engine}, platform::to_void_cast(src_data)));
dev_ctx.SetBlob(key_src_mem, p_src_mem);
auto p_dst_mem = std::make_shared<mkldnn::memory>(mkldnn::memory(
{data_md, mkldnn_engine}, platform::to_void_cast(dst_data)));
dev_ctx.SetBlob(key_dst_mem, p_dst_mem);
auto fwd_desc = mkldnn::eltwise_forward::desc(
mkldnn::prop_kind::forward_training, algorithm, data_md, alpha, beta);
auto p_fwd_pd = std::make_shared<mkldnn::eltwise_forward::primitive_desc>(
fwd_desc, mkldnn_engine);
const std::string key_fwd_pd = key + "eltwise_fwd_pd";
dev_ctx.SetBlob(key_fwd_pd, p_fwd_pd);
p_fwd = std::make_shared<mkldnn::eltwise_forward>(
*p_fwd_pd, *(p_src_mem.get()), *(p_dst_mem.get()));
// create mkldnn memory for input X
auto src_md = platform::MKLDNNMemDesc(
src_tz, platform::MKLDNNGetDataType<T>(), src_format);
auto src_memory = std::shared_ptr<memory>(
new memory({src_md, mkldnn_engine}, to_void_cast(x_data)));
// save src_memory to be referred in backward path
dev_ctx.SetBlob(key_src_mem, src_memory);
// create primitive descriptor for activation forward and save it
auto forward_desc = mkldnn::eltwise_forward::desc(
mkldnn::prop_kind::forward_training, algorithm,
src_memory->get_primitive_desc().desc(), alpha, beta);
auto forward_pd = std::make_shared<mkldnn::eltwise_forward::primitive_desc>(
forward_desc, mkldnn_engine);
// save prim desc into global device context to be referred in backward path
dev_ctx.SetBlob(key_fwd_pd, forward_pd);
// create mkldnn memory for output y
dst_memory =
std::make_shared<memory>(forward_pd->dst_primitive_desc(), y_data);
dev_ctx.SetBlob(key_dst_mem, dst_memory);
// create activation primitive
p_fwd = std::make_shared<mkldnn::eltwise_forward>(*forward_pd, *src_memory,
*dst_memory);
dev_ctx.SetBlob(key_fwd, p_fwd);
} else {
// primitives already exist
auto p_src_mem =
auto src_memory =
std::static_pointer_cast<mkldnn::memory>(dev_ctx.GetBlob(key_src_mem));
PADDLE_ENFORCE(p_src_mem != nullptr,
"Fail to find eltwise p_src_mem in device context.");
auto p_dst_mem =
PADDLE_ENFORCE(src_memory != nullptr,
"Fail to find eltwise src_memory in device context.");
dst_memory =
std::static_pointer_cast<mkldnn::memory>(dev_ctx.GetBlob(key_dst_mem));
PADDLE_ENFORCE(p_dst_mem != nullptr,
"Fail to find eltwise p_src_mem in device context.");
PADDLE_ENFORCE(dst_memory != nullptr,
"Fail to find eltwise dst_memory in device context.");
p_src_mem->set_data_handle(platform::to_void_reinterpret_cast(src_data));
p_dst_mem->set_data_handle(dst_data);
src_memory->set_data_handle(platform::to_void_cast(x_data));
dst_memory->set_data_handle(y_data);
}
// push primitive to stream and wait until it's executed
std::vector<mkldnn::primitive> pipeline = {*(p_fwd.get())};
mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
std::vector<primitive> pipeline;
pipeline.push_back(*p_fwd);
stream(stream::kind::eager).submit(pipeline).wait();
y->set_layout(DataLayout::kMKLDNN);
y->set_format(GetMKLDNNFormat(*dst_memory));
}
template <typename T, typename ExecContext>
void eltwise_grad(const ExecContext &ctx, mkldnn::algorithm algorithm,
const T alpha = 0, const T beta = 0) {
template <typename T>
void eltwise_grad(const framework::ExecutionContext &ctx,
mkldnn::algorithm algorithm, const T alpha = 0,
const T beta = 0) {
auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();
const auto &mkldnn_engine = dev_ctx.GetEngine();
// get buffers
const auto *out = ctx.template Input<Tensor>("Out");
auto *dout = ctx.template Input<Tensor>(framework::GradVarName("Out"));
const auto *diff_dst = dout->template data<T>();
const auto *diff_y = ctx.Input<Tensor>(framework::GradVarName("Out"));
auto *diff_x = ctx.Output<Tensor>(framework::GradVarName("X"));
auto *dx =
ctx.template Output<framework::Tensor>(framework::GradVarName("X"));
const T *diff_src = dx->template mutable_data<T>(ctx.GetPlace());
const T *diff_y_data = diff_y->data<T>();
T *diff_x_data = diff_x->mutable_data<T>(ctx.GetPlace());
// get memory dim
std::vector<int> src_tz = framework::vectorize2int(out->dims());
std::vector<int> diff_dst_tz = framework::vectorize2int(diff_y->dims());
const std::string key = gethash(src_tz, algorithm);
const std::string key_diff_src_mem = key + "@eltwise_diff_src_mem";
const std::string key_diff_dst_mem = key + "@eltwise_diff_dst_mem";
const std::string key_grad = key + "@eltwise_grad";
auto diff_y_format =
diff_dst_tz.size() == 2 ? mkldnn::memory::format::nc : diff_y->format();
const std::string key = gethash(diff_dst_tz, algorithm);
const std::string key_src_data =
key + ctx.op().Input("Out") + "@eltwise_fwd_src_data";
const std::string key_src_layout =
key + ctx.op().Input("Out") + "@eltwise_fwd_src_layout";
const auto p_src_layout =
std::static_pointer_cast<memory::format>(dev_ctx.GetBlob(key_src_layout));
const std::string key_src_mem =
key + std::to_string(*p_src_layout) + "@eltwise_fwd_src_mem";
const std::string key_fwd_pd =
key + std::to_string(*p_src_layout) + "@eltwise_fwd_pd";
const std::string key_with_layouts =
key + std::to_string(*p_src_layout) + "-" + std::to_string(diff_y_format);
const std::string key_diff_src_mem =
key_with_layouts + "@eltwise_diff_src_mem";
const std::string key_diff_dst_mem =
key_with_layouts + "@eltwise_diff_dst_mem";
const std::string key_grad = key_with_layouts + "@eltwise_grad";
const auto p_src_data =
std::static_pointer_cast<T *>(dev_ctx.GetBlob(key_src_data));
const std::string key_src_mem = key + "@eltwise_fwd_src_mem";
auto p_src_mem =
auto src_memory =
std::static_pointer_cast<mkldnn::memory>(dev_ctx.GetBlob(key_src_mem));
p_src_mem->set_data_handle(*p_src_data.get());
PADDLE_ENFORCE(src_memory != nullptr,
"Fail to find src_memory in device context");
src_memory->set_data_handle(*p_src_data.get());
std::shared_ptr<memory> diff_src_memory;
auto p_grad = std::static_pointer_cast<mkldnn::eltwise_forward::primitive>(
auto p_grad = std::static_pointer_cast<mkldnn::eltwise_backward>(
dev_ctx.GetBlob(key_grad));
if (p_grad == nullptr) {
// create memory description
auto data_md = src_tz.size() == 2
? platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
mkldnn::memory::format::nc)
: platform::MKLDNNMemDesc(src_tz, mkldnn::memory::f32,
mkldnn::memory::format::nchw);
// create memory primitives
std::shared_ptr<void> p_diff_src_mem =
std::make_shared<mkldnn::memory>(mkldnn::memory(
{data_md, mkldnn_engine}, platform::to_void_cast(diff_src)));
dev_ctx.SetBlob(key_diff_src_mem, p_diff_src_mem);
std::shared_ptr<void> p_diff_dst_mem =
std::make_shared<mkldnn::memory>(mkldnn::memory(
{data_md, mkldnn_engine}, platform::to_void_cast(diff_dst)));
dev_ctx.SetBlob(key_diff_dst_mem, p_diff_dst_mem);
auto bwd_desc = mkldnn::eltwise_backward::desc(algorithm, data_md, data_md,
alpha, beta);
const std::string key_fwd_pd = key + "eltwise_fwd_pd";
auto *p_fwd_pd = static_cast<mkldnn::eltwise_forward::primitive_desc *>(
dev_ctx.GetBlob(key_fwd_pd).get());
auto eltwise_bwd_prim_desc = mkldnn::eltwise_backward::primitive_desc(
bwd_desc, mkldnn_engine, *p_fwd_pd);
// create mkldnn memory for input diff_y
auto diff_dst_md = platform::MKLDNNMemDesc(
diff_dst_tz, platform::MKLDNNGetDataType<T>(), diff_y_format);
auto diff_dst_memory = std::shared_ptr<memory>(
new memory({diff_dst_md, mkldnn_engine}, to_void_cast(diff_y_data)));
dev_ctx.SetBlob(key_diff_dst_mem, diff_dst_memory);
// retrieve eltwise primitive desc from device context
auto forward_pd =
std::static_pointer_cast<mkldnn::eltwise_forward::primitive_desc>(
dev_ctx.GetBlob(key_fwd_pd));
PADDLE_ENFORCE(forward_pd != nullptr,
"Fail to find eltwise_fwd_pd in device context");
// ceate primitive descriptor for activation backward
auto backward_desc = mkldnn::eltwise_backward::desc(
algorithm, diff_dst_memory->get_primitive_desc().desc(),
src_memory->get_primitive_desc().desc(), alpha, beta);
auto backward_pd = mkldnn::eltwise_backward::primitive_desc(
backward_desc, mkldnn_engine, *forward_pd);
// create mkldnn memory for output diff_src
diff_src_memory = std::make_shared<memory>(
backward_pd.diff_src_primitive_desc(), diff_x_data);
dev_ctx.SetBlob(key_diff_src_mem, diff_src_memory);
// create activation backward primitive
p_grad = std::make_shared<mkldnn::eltwise_backward>(
eltwise_bwd_prim_desc, *static_cast<mkldnn::memory *>(p_src_mem.get()),
*(static_cast<mkldnn::memory *>(p_diff_dst_mem.get())),
*(static_cast<mkldnn::memory *>(p_diff_src_mem.get())));
backward_pd, *src_memory, *diff_dst_memory, *diff_src_memory);
dev_ctx.SetBlob(key_grad, p_grad);
} else {
// primitives already exist
auto p_diff_src_mem = std::static_pointer_cast<mkldnn::memory>(
diff_src_memory = std::static_pointer_cast<mkldnn::memory>(
dev_ctx.GetBlob(key_diff_src_mem));
auto p_diff_dst_mem = std::static_pointer_cast<mkldnn::memory>(
auto diff_dst_memory = std::static_pointer_cast<mkldnn::memory>(
dev_ctx.GetBlob(key_diff_dst_mem));
p_diff_src_mem->set_data_handle(
platform::to_void_reinterpret_cast(diff_src));
p_diff_dst_mem->set_data_handle(
platform::to_void_reinterpret_cast(diff_dst));
diff_src_memory->set_data_handle(
platform::to_void_reinterpret_cast(diff_x_data));
diff_dst_memory->set_data_handle(
platform::to_void_reinterpret_cast(diff_y_data));
}
// push primitive to stream and wait until it's executed
std::vector<mkldnn::primitive> pipeline = {*(p_grad.get())};
mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
std::vector<primitive> pipeline;
pipeline.push_back(*p_grad);
stream(stream::kind::eager).submit(pipeline).wait();
diff_x->set_layout(DataLayout::kMKLDNN);
diff_x->set_format(GetMKLDNNFormat(*diff_src_memory));
}
} // anonymous namespace
template <typename T, mkldnn::algorithm algorithm>
struct MKLDNNActivationFunc : public BaseActivationFunctor<T> {
template <typename ExecContext>
void operator()(const ExecContext &ctx) const {
void operator()(const framework::ExecutionContext &ctx) const {
eltwise_forward<T>(ctx, algorithm);
}
};
template <typename T, mkldnn::algorithm algorithm>
struct MKLDNNActivationGradFunc : public BaseActivationFunctor<T> {
template <typename ExecContext>
void operator()(const ExecContext &ctx) const {
void operator()(const framework::ExecutionContext &ctx) const {
eltwise_grad<T>(ctx, algorithm);
}
};
......
paddle/fluid/operators/activation_op.cc
浏览文件 @ f1f8327c
......@@ -19,18 +19,20 @@ limitations under the License. */
namespace paddle {
namespace operators {
#define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT) \
class OP_NAME##OpMaker \
: public ::paddle::framework::OpProtoAndCheckerMaker { \
public: \
void Make() override { \
AddInput("X", "Input of " #OP_NAME " operator"); \
AddOutput("Out", "Output of " #OP_NAME " operator").Reuse("X"); \
AddAttr<bool>("use_mkldnn", \
"(default false) Only used in mkldnn kernel") \
.SetDefault(false); \
AddComment(OP_COMMENT); \
} \
using paddle::framework::Tensor;
#define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT) \
class OP_NAME##OpMaker \
: public ::paddle::framework::OpProtoAndCheckerMaker { \
public: \
void Make() override { \
AddInput("X", "Input of " #OP_NAME " operator"); \
AddOutput("Out", "Output of " #OP_NAME " operator").Reuse("X"); \
AddAttr<bool>("use_mkldnn", \
"(bool, default false) Only used in mkldnn kernel") \
.SetDefault(false); \
AddComment(#OP_COMMENT); \
} \
}
#define REGISTER_ACTIVATION_OP_GRAD_MAKER(OP_NAME, KERNEL_TYPE) \
......@@ -58,7 +60,6 @@ framework::OpKernelType GetKernelType(const framework::ExecutionContext& ctx,
const framework::OperatorWithKernel& oper,
const std::string& name) {
framework::LibraryType library{framework::LibraryType::kPlain};
framework::DataLayout layout = framework::DataLayout::kAnyLayout;
#ifdef PADDLE_WITH_MKLDNN
auto it = oper.Attrs().find("use_mkldnn");
......@@ -82,6 +83,7 @@ class ActivationOp : public framework::OperatorWithKernel {
ctx->ShareLoD("X", /*->*/ "Out");
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return GetKernelType(ctx, *this, "X");
......@@ -96,6 +98,7 @@ class ActivationOpGrad : public framework::OperatorWithKernel {
ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out"));
}
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return GetKernelType(ctx, *this, "Out");
......@@ -133,7 +136,7 @@ $out = \max(x, 0)$
__attribute__((unused)) constexpr char TanhDoc[] = R"DOC(
Tanh Activation Operator.
$$out = \frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$
$$out = \\frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$
)DOC";
......@@ -443,7 +446,7 @@ class SwishOpMaker : public framework::OpProtoAndCheckerMaker {
AddComment(R"DOC(
Swish Activation Operator.
$$out = \frac{x}{1 + e^{- \beta x}}$$
$$out = \\frac{x}{1 + e^{- \beta x}}$$
)DOC");
}
......
paddle/fluid/operators/chunk_eval_op.cc
浏览文件 @ f1f8327c
......@@ -91,32 +91,31 @@ class ChunkEvalOpMaker : public framework::OpProtoAndCheckerMaker {
"(int64_t). The number of chunks both in Inference and Label on the "
"given mini-batch.");
AddAttr<int>("num_chunk_types",
"(int). The number of chunk type. See below for details.");
AddAttr<std::string>(
"chunk_scheme",
"(string, default IOB). The labeling scheme indicating "
"how to encode the chunks. Must be IOB, IOE, IOBES or plain. See below "
"for details.")
"The number of chunk type. See the description for details.");
AddAttr<std::string>("chunk_scheme",
"The labeling scheme indicating "
"how to encode the chunks. Must be IOB, IOE, IOBES or "
"plain. See the description"
"for details.")
.SetDefault("IOB");
AddAttr<std::vector<int>>("excluded_chunk_types",
"(list<int>) A list including chunk type ids "
"A list including chunk type ids "
"indicating chunk types that are not counted. "
"See below for details.")
"See the description for details.")
.SetDefault(std::vector<int>{});
AddComment(R"DOC(
For some basics of chunking, please refer to
‘Chunking with Support Vector Machines <https://aclanthology.info/pdf/N/N01/N01-1025.pdf>’.
'Chunking with Support Vector Machines <https://aclanthology.info/pdf/N/N01/N01-1025.pdf>'.
CheckEvalOp computes the precision, recall, and F1-score of chunk detection,
ChunkEvalOp computes the precision, recall, and F1-score of chunk detection,
and supports IOB, IOE, IOBES and IO (also known as plain) tagging schemes.
Here is a NER example of labeling for these tagging schemes:
Li Ming works at Agricultural Bank of China in Beijing.
IO: I-PER I-PER O O I-ORG I-ORG I-ORG I-ORG O I-LOC
IOB: B-PER I-PER O O B-ORG I-ORG I-ORG I-ORG O B-LOC
IOE: I-PER E-PER O O I-ORG I-ORG I-ORG E-ORG O E-LOC
IOBES: B-PER E-PER O O I-ORG I-ORG I-ORG E-ORG O S-LOC
Li Ming works at Agricultural Bank of China in Beijing.
IO I-PER I-PER O O I-ORG I-ORG I-ORG I-ORG O I-LOC
IOB B-PER I-PER O O B-ORG I-ORG I-ORG I-ORG O B-LOC
IOE I-PER E-PER O O I-ORG I-ORG I-ORG E-ORG O E-LOC
IOBES B-PER E-PER O O I-ORG I-ORG I-ORG E-ORG O S-LOC
There are three chunk types(named entity types) including PER(person), ORG(organization)
and LOC(LOCATION), and we can see that the labels have the form <tag type>-<chunk type>.
......@@ -124,31 +123,31 @@ and LOC(LOCATION), and we can see that the labels have the form <tag type>-<chun
Since the calculations actually use label ids rather than labels, extra attention
should be paid when mapping labels to ids to make CheckEvalOp work. The key point
is that the listed equations are satisfied by ids.
tag_type = label % num_tag_type
chunk_type = label / num_tag_type
tag_type = label % num_tag_type
chunk_type = label / num_tag_type
where `num_tag_type` is the num of tag types in the tagging scheme, `num_chunk_type`
is the num of chunk types, and `tag_type` get its value from the following table.
Scheme Begin Inside End Single
plain 0 - - -
IOB 0 1 - -
IOE - 0 1 -
IOBES 0 1 2 3
Scheme Begin Inside End Single
plain 0 - - -
IOB 0 1 - -
IOE - 0 1 -
IOBES 0 1 2 3
Still use NER as example, assuming the tagging scheme is IOB while chunk types are ORG,
PER and LOC. To satisfy the above equations, the label map can be like this:
B-ORG 0
I-ORG 1
B-PER 2
I-PER 3
B-LOC 4
I-LOC 5
O 6
B-ORG 0
I-ORG 1
B-PER 2
I-PER 3
B-LOC 4
I-LOC 5
O 6
It’s not hard to verify the equations noting that the num of chunk types
It's not hard to verify the equations noting that the num of chunk types
is 3 and the num of tag types in IOB scheme is 2. For example, the label
id of I-LOC is 5, the tag type id of I-LOC is 1, and the chunk type id of
I-LOC is 2, which consistent with the results from the equations.
......
paddle/fluid/operators/clip_by_norm_op.cc
浏览文件 @ f1f8327c
......@@ -54,10 +54,19 @@ be linearly scaled to make the L2 norm of $Out$ equal to $max\_norm$, as
shown in the following formula:
$$
Out = \frac{max\_norm * X}{norm(X)},
Out = \\frac{max\\_norm * X}{norm(X)},
$$
where $norm(X)$ represents the L2 norm of $X$.
Examples:
.. code-block:: python
data = fluid.layer.data(
name='data', shape=[2, 4, 6], dtype='float32')
reshaped = fluid.layers.clip_by_norm(
x=data, max_norm=0.5)
)DOC");
}
};
......
paddle/fluid/operators/conv_transpose_op.cc
浏览文件 @ f1f8327c
......@@ -156,7 +156,7 @@ Parameters(strides, paddings) are two elements. These two elements represent hei
and width, respectively.
The input(X) size and output(Out) size may be different.
Example:
For an example:
Input:
Input shape: $(N, C_{in}, H_{in}, W_{in})$
Filter shape: $(C_{in}, C_{out}, H_f, W_f)$
......
paddle/fluid/operators/cos_sim_op.cc
浏览文件 @ f1f8327c
......@@ -76,9 +76,9 @@ class CosSimOpMaker : public framework::OpProtoAndCheckerMaker {
.AsIntermediate();
AddComment(R"DOC(
Cosine Similarity Operator.
**Cosine Similarity Operator**
$Out = X^T * Y / (\sqrt{X^T * X} * \sqrt{Y^T * Y})$
$Out = \frac{X^T * Y}{(\sqrt{X^T * X} * \sqrt{Y^T * Y})}$
The input X and Y must have the same shape, except that the 1st dimension
of input Y could be just 1 (different from input X), which will be
......
paddle/fluid/operators/crf_decoding_op.cc
浏览文件 @ f1f8327c
......@@ -53,21 +53,18 @@ sequence of observed tags.
The output of this operator changes according to whether Input(Label) is given:
1. Input(Label) is given:
This happens in training. This operator is used to co-work with the chunk_eval
operator.
When Input(Label) is given, the crf_decoding operator returns a row vector
with shape [N x 1] whose values are fixed to be 0, indicating an incorrect
prediction, or 1 indicating a tag is correctly predicted. Such an output is the
input to chunk_eval operator.
This happens in training. This operator is used to co-work with the chunk_eval
operator.
When Input(Label) is given, the crf_decoding operator returns a row vector
with shape [N x 1] whose values are fixed to be 0, indicating an incorrect
prediction, or 1 indicating a tag is correctly predicted. Such an output is the
input to chunk_eval operator.
2. Input(Label) is not given:
This is the standard decoding process.
This is the standard decoding process.
The crf_decoding operator returns a row vector with shape [N x 1] whose values
range from 0 to maximum tag number - 1. Each element indicates an index of a
range from 0 to maximum tag number - 1, Each element indicates an index of a
predicted tag.
)DOC");
}
......
paddle/fluid/operators/detail/grpc_client.cc
浏览文件 @ f1f8327c
......@@ -245,7 +245,7 @@ void GRPCClient::Proceed() {
if (c->status_.ok()) {
c->Process();
} else {
LOG(ERROR) << "var: " << c->var_h_.String()
LOG(FATAL) << "var: " << c->var_h_.String()
<< " grpc error:" << c->status_.error_message();
}
delete c;
......
paddle/fluid/operators/detection/iou_similarity_op.cc
浏览文件 @ f1f8327c
......@@ -68,15 +68,16 @@ class IOUSimilarityOpMaker : public framework::OpProtoAndCheckerMaker {
"representing pairwise iou scores.");
AddComment(R"DOC(
IOU Similarity Operator.
**IOU Similarity Operator**
Computes intersection-over-union (IOU) between two box lists.
Box list 'X' should be a LoDTensor and 'Y' is a common Tensor,
boxes in 'Y' are shared by all instance of the batched inputs of X.
Given two boxes A and B, the calculation of IOU is as follows:
Box list 'X' should be a LoDTensor and 'Y' is a common Tensor,
boxes in 'Y' are shared by all instance of the batched inputs of X.
Given two boxes A and B, the calculation of IOU is as follows:
$$
IOU(A, B) =
\frac{area(A\cap B)}{area(A)+area(B)-area(A\cap B)}
\\frac{area(A\\cap B)}{area(A)+area(B)-area(A\\cap B)}
$$
)DOC");
......
paddle/fluid/operators/detection/polygon_box_transform_op.cc
浏览文件 @ f1f8327c
......@@ -83,11 +83,13 @@ class PolygonBoxTransformOpMaker : public framework::OpProtoAndCheckerMaker {
AddComment(R"DOC(
PolygonBoxTransform Operator.
PolygonBoxTransform Operator is used to transform the coordinate shift to the real coordinate.
The input is the final geometry output in detection network.
We use 2*n numbers to denote the coordinate shift from n corner vertices of
the polygon_box to the pixel location. As each distance offset contains two numbers (xi, yi),
the geometry output contains 2*n channels.
PolygonBoxTransform Operator is used to transform the coordinate shift to the real coordinate.
)DOC");
}
};
......
paddle/fluid/operators/linear_chain_crf_op.cc
浏览文件 @ f1f8327c
......@@ -84,6 +84,7 @@ CRF. Please refer to http://www.cs.columbia.edu/~mcollins/fb.pdf and
http://cseweb.ucsd.edu/~elkan/250Bwinter2012/loglinearCRFs.pdf for details.
Equation:
1. Denote Input(Emission) to this operator as $x$ here.
2. The first D values of Input(Transition) to this operator are for starting
weights, denoted as $a$ here.
......@@ -106,6 +107,7 @@ Finally, the linear chain CRF operator outputs the logarithm of the conditional
likelihood of each training sample in a mini-batch.
NOTE:
1. The feature function for a CRF is made up of the emission features and the
transition features. The emission feature weights are NOT computed in
this operator. They MUST be computed first before this operator is called.
......
paddle/fluid/operators/lstm_op.cc
浏览文件 @ f1f8327c
......@@ -184,34 +184,32 @@ Long-Short Term Memory (LSTM) Operator.
The defalut implementation is diagonal/peephole connection
(https://arxiv.org/pdf/1402.1128.pdf), the formula is as follows:
$$
i_t = \sigma(W_{ix}x_{t} + W_{ih}h_{t-1} + W_{ic}c_{t-1} + b_i) \\
$$ i_t = \\sigma(W_{ix}x_{t} + W_{ih}h_{t-1} + W_{ic}c_{t-1} + b_i) $$
f_t = \sigma(W_{fx}x_{t} + W_{fh}h_{t-1} + W_{fc}c_{t-1} + b_f) \\
$$ f_t = \\sigma(W_{fx}x_{t} + W_{fh}h_{t-1} + W_{fc}c_{t-1} + b_f) $$
\tilde{c_t} = act_g(W_{cx}x_t + W_{ch}h_{t-1} + b_c) \\
$$ \\tilde{c_t} = act_g(W_{cx}x_t + W_{ch}h_{t-1} + b_c) $$
o_t = \sigma(W_{ox}x_{t} + W_{oh}h_{t-1} + W_{oc}c_t + b_o) \\
$$ o_t = \\sigma(W_{ox}x_{t} + W_{oh}h_{t-1} + W_{oc}c_t + b_o) $$
c_t = f_t \odot c_{t-1} + i_t \odot \tilde{c_t} \\
$$ c_t = f_t \\odot c_{t-1} + i_t \\odot \\tilde{c_t} $$
h_t = o_t \odot act_h(c_t)
$$
$$ h_t = o_t \\odot act_h(c_t) $$
where the W terms denote weight matrices (e.g. $W_{xi}$ is the matrix
of weights from the input gate to the input), $W_{ic}, W_{fc}, W_{oc}$
are diagonal weight matrices for peephole connections. In our implementation,
we use vectors to reprenset these diagonal weight matrices. The b terms
denote bias vectors ($b_i$ is the input gate bias vector), $\sigma$
is the non-line activations, such as logistic sigmoid function, and
$i, f, o$ and $c$ are the input gate, forget gate, output gate,
and cell activation vectors, respectively, all of which have the same size as
the cell output activation vector $h$.
The $\odot$ is the element-wise product of the vectors. $act_g$ and $act_h$
are the cell input and cell output activation functions and `tanh` is usually
used for them. $\tilde{c_t}$ is also called candidate hidden state,
which is computed based on the current input and the previous hidden state.
- W terms denote weight matrices (e.g. $W_{xi}$ is the matrix
of weights from the input gate to the input), $W_{ic}, W_{fc}, W_{oc}$
are diagonal weight matrices for peephole connections. In our implementation,
we use vectors to reprenset these diagonal weight matrices.
- The b terms denote bias vectors ($b_i$ is the input gate bias vector).
- $\sigma$ is the non-line activations, such as logistic sigmoid function.
- $i, f, o$ and $c$ are the input gate, forget gate, output gate,
and cell activation vectors, respectively, all of which have the same size as
the cell output activation vector $h$.
- The $\odot$ is the element-wise product of the vectors.
- $act_g$ and $act_h$ are the cell input and cell output activation functions
and `tanh` is usually used for them.
- $\tilde{c_t}$ is also called candidate hidden state,
which is computed based on the current input and the previous hidden state.
Set `use_peepholes` False to disable peephole connection. The formula
is omitted here, please refer to the paper
......
paddle/fluid/operators/pool_op.cc
浏览文件 @ f1f8327c
......@@ -204,8 +204,6 @@ void Pool2dOpMaker::Make() {
// TODO(dzhwinter): need to registered layout transform function
AddComment(R"DOC(
Pool2d Operator.
The pooling2d operation calculates the output based on
the input, pooling_type and ksize, strides, paddings parameters.
Input(X) and output(Out) are in NCHW format, where N is batch size, C is the
......@@ -215,19 +213,28 @@ These two elements represent height and width, respectively.
The input(X) size and output(Out) size may be different.
Example:
Input:
X shape: $(N, C, H_{in}, W_{in})$
Output:
Out shape: $(N, C, H_{out}, W_{out})$
For ceil_mode = false:
$$
H_{out} = \frac{(H_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1 \\
W_{out} = \frac{(W_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1
H_{out} = \\frac{(H_{in} - ksize[0] + 2 * paddings[0])}{strides[0]} + 1
$$
$$
W_{out} = \\frac{(W_{in} - ksize[1] + 2 * paddings[1])}{strides[1]} + 1
$$
For ceil_mode = true:
$$
H_{out} = \frac{(H_{in} - ksize[0] + 2 * paddings[0] + strides[0] - 1)}{strides[0]} + 1 \\
W_{out} = \frac{(W_{in} - ksize[1] + 2 * paddings[1] + strides[1] - 1)}{strides[1]} + 1
H_{out} = \\frac{(H_{in} - ksize[0] + 2 * paddings[0] + strides[0] - 1)}{strides[0]} + 1
$$
$$
W_{out} = \\frac{(W_{in} - ksize[1] + 2 * paddings[1] + strides[1] - 1)}{strides[1]} + 1
$$
)DOC");
......
paddle/fluid/operators/roi_pool_op.cc
浏览文件 @ f1f8327c
......@@ -139,7 +139,20 @@ class ROIPoolOpMaker : public framework::OpProtoAndCheckerMaker {
"The pooled output width.")
.SetDefault(1);
AddComment(R"DOC(
ROIPool operator
**ROIPool Operator**
Region of interest pooling (also known as RoI pooling) is to perform
is to perform max pooling on inputs of nonuniform sizes to obtain
fixed-size feature maps (e.g. 7*7).
The operator has three steps:
1. Dividing each region proposal into equal-sized sections with
the pooled_width and pooled_height
2. Finding the largest value in each section
3. Copying these max values to the output buffer
ROI Pooling for Faster-RCNN. The link below is a further introduction:
https://stackoverflow.com/questions/43430056/what-is-roi-layer-in-fast-rcnn
......
paddle/fluid/operators/scale_op.cc
浏览文件 @ f1f8327c
......@@ -41,13 +41,13 @@ class ScaleOpMaker : public framework::OpProtoAndCheckerMaker {
AddInput("X", "(Tensor) Input tensor of scale operator.");
AddOutput("Out", "(Tensor) Output tensor of scale operator.");
AddComment(R"DOC(
Scale operator
**Scale operator**
Multiply the input tensor with a float scalar to scale the input tensor.
$$Out = scale*X$$
)DOC");
AddAttr<float>("scale",
"(float, default 1.0)"
"The scaling factor of the scale operator.")
AddAttr<float>("scale", "The scaling factor of the scale operator.")
.SetDefault(1.0);
}
};
......
paddle/fluid/operators/shape_op.cc
浏览文件 @ f1f8327c
......@@ -36,10 +36,13 @@ class ShapeOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("Input", "(Tensor), The input tensor.");
AddOutput("Out", "(Tensor), The shape of input tensor.");
AddOutput("Out",
"(Tensor), The shape of input tensor, the data type of the shape"
" is int64_t, will be on the same device with the input Tensor.");
AddComment(R"DOC(
Shape Operator.
Get the shape of input tensor.
Shape Operator
Get the shape of input tensor. Only support CPU input Tensor now.
)DOC");
}
};
......
paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc
浏览文件 @ f1f8327c
......@@ -113,14 +113,14 @@ The logistic loss is given as follows:
$$loss = -Labels * \log(\sigma(X)) - (1 - Labels) * \log(1 - \sigma(X))$$
We know that $$\sigma(X) = (1 / (1 + \exp(-X)))$$. By substituting this we get:
We know that $$\sigma(X) = \\frac{1}{1 + \exp(-X)}$$. By substituting this we get:
$$loss = X - X * Labels + \log(1 + \exp(-X))$$
For stability and to prevent overflow of $$\exp(-X)$$ when X < 0,
we reformulate the loss as follows:
$$loss = \max(X, 0) - X * Labels + \log(1 + \exp(-|X|))$$
$$loss = \max(X, 0) - X * Labels + \log(1 + \exp(-\|X\|))$$
Both the input `X` and `Labels` can carry the LoD (Level of Details) information.
However the output only shares the LoD with input `X`.
......
paddle/fluid/operators/slice_op.cc
浏览文件 @ f1f8327c
......@@ -95,23 +95,26 @@ of that dimension. If the value passed to start or end is larger than
the n (the number of elements in this dimension), it represents n.
For slicing to the end of a dimension with unknown size, it is recommended
to pass in INT_MAX. If axes are omitted, they are set to [0, ..., ndim-1].
Example 1:
Given:
data = [ [1, 2, 3, 4], [5, 6, 7, 8], ]
axes = [0, 1]
starts = [1, 0]
ends = [2, 3]
Then:
result = [ [5, 6, 7], ]
Example 2:
Given:
data = [ [1, 2, 3, 4], [5, 6, 7, 8], ]
starts = [0, 1]
ends = [-1, 1000]
Then:
result = [ [2, 3, 4], ]
Following examples will explain how slice works:
.. code-block:: text
Cast1:
Given:
data = [ [1, 2, 3, 4], [5, 6, 7, 8], ]
axes = [0, 1]
starts = [1, 0]
ends = [2, 3]
Then:
result = [ [5, 6, 7], ]
Cast2:
Given:
data = [ [1, 2, 3, 4], [5, 6, 7, 8], ]
starts = [0, 1]
ends = [-1, 1000]
Then:
result = [ [2, 3, 4], ]
)DOC");
}
};
......
paddle/fluid/operators/tensorrt_engine_op_test.cc
浏览文件 @ f1f8327c
......@@ -240,7 +240,7 @@ void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) {
}
// Test with a larger FC layer.
TEST(TensorRTEngineOp, fc) { Execute(40, 256, 256); }
TEST(TensorRTEngineOp, fc) { Execute(40, 28, 28); }
} // namespace operators
} // namespace paddle
......
paddle/fluid/operators/uniform_random_batch_size_like_op.cc
浏览文件 @ f1f8327c
......@@ -35,10 +35,10 @@ class UniformRandomBatchSizeLikeOpMaker : public BatchSizeLikeOpMaker {
protected:
void Apply() override {
AddComment(R"DOC(
Uniform random operator
UniformRandomBatchSizeLike operator.
This operator initializes a tensor with the same batch_size as the Input tensor
with random values sampled from a uniform distribution.
with random values sampled from a uniform distribution.
)DOC");
AddAttr<float>("min",
......
python/paddle/fluid/framework.py
浏览文件 @ f1f8327c
......@@ -1034,6 +1034,37 @@ class Block(object):
class Program(object):
"""
Python Program. Beneath it is a ProgramDesc, which is used for
create c++ Program. A program is a self-contained programing
language like container. It has at least one Block, when the
control flow op like conditional_block, while_op is included,
it will contains nested block.
Please reference the framework.proto for details.
Notes: we have default_startup_program and default_main_program
by default, a pair of them will shared the parameters.
The default_startup_program only run once to initialize parameters,
default_main_program run in every minibatch and adjust the weights.
Args:
None
Returns:
Python Program
Examples:
.. code-block:: python
main_program = Program()
startup_program = Program()
with fluid.program_guard(main_program=main_program, startup_program=startup_program):
fluid.layers.data(name="x", shape=[-1, 784], dtype='float32')
fluid.layers.data(name="y", shape=[-1, 1], dtype='int32')
fluid.layers.fc(name="fc", shape=[10], dtype='float32', act="relu")
"""
def __init__(self):
self.desc = core.ProgramDesc()
self.blocks = [Block(self, 0)]
......@@ -1099,6 +1130,8 @@ class Program(object):
def clone(self, for_test=False):
"""Clone the Program object
Args:
for_test(bool): indicate whether clone for test.
Set for_test to False when we want to clone the program for training.
Set for_test to True when we want to clone the program for testing.
......@@ -1109,8 +1142,9 @@ class Program(object):
the is_test attributes in these operators will be set to True for
testing purposes, otherwise, they remain unchanged.
Returns(Program):
The cloned Program object.
Returns:
Program: The cloned Program object.
"""
if for_test:
p = self.inference_optimize()
......@@ -1228,6 +1262,7 @@ class Program(object):
def copy_param_info_from(self, other):
"""
Copy the information of parameters from other program.
Args:
other(Program): Other program
......@@ -1246,6 +1281,7 @@ class Program(object):
def copy_data_info_from(self, other):
"""
Copy the information of data variables from other program.
Args:
other(Program): Other program
......@@ -1299,6 +1335,7 @@ class Parameter(Variable):
def to_string(self, throw_on_error, with_details=False):
"""
To debug string.
Args:
throw_on_error(bool): raise exception when self is not initialized
when throw_on_error is True
......
python/paddle/fluid/layers/control_flow.py
浏览文件 @ f1f8327c
......@@ -27,7 +27,6 @@ __all__ = [
'merge_lod_tensor',
'BlockGuard',
'BlockGuardWithCompletion',
'StaticRNNMemoryLink',
'WhileGuard',
'While',
'Switch',
......@@ -56,34 +55,36 @@ __all__ = [
def split_lod_tensor(input, mask, level=0):
"""
**split_lod_tensor**
This function takes in an input that contains the complete lod information,
and takes in a mask which is used to mask certain parts of the input.
The output is the true branch and the false branch with the mask applied to
the input at a certain level in the tensor.
the input at a certain level in the tensor. Mainly used in IfElse to split
data into two parts.
Args:
input(tuple|list|None): The input tensor that contains complete
lod information needed to construct the output.
mask(list): A bool column vector which masks the input.
level(int): The specific lod level to rank.
level(int): The specific lod level to split.
Returns:
Variable: The true branch of tensor as per the mask applied to input.
Variable: The false branch of tensor as per the mask applied to input.
tuple(Variable, Variable):
The true branch of tensor as per the mask applied to input.
The false branch of tensor as per the mask applied to input.
Examples:
.. code-block:: python
x = layers.data(name='x', shape=[1])
x = fluid.layers.data(name='x', shape=[1])
x.persistable = True
y = layers.data(name='y', shape=[1])
y = fluid.layers.data(name='y', shape=[1])
y.persistable = True
out_true, out_false = layers.split_lod_tensor(
out_true, out_false = fluid.layers.split_lod_tensor(
input=x, mask=y, level=level)
"""
helper = LayerHelper('split_lod_tensor', **locals())
out_true = helper.create_tmp_variable(dtype=input.dtype)
......@@ -106,8 +107,9 @@ def merge_lod_tensor(in_true, in_false, x, mask, level=0):
This function takes in an input :math:`x`, the True branch, the False
branch and a binary :math:`mask`. Using this information, this function
merges the True and False branches of the tensor into a single Output
at a certain lod level indiacted by :math:`level`.
merges the True and False branches of the tensor into a single tensor as
output at a certain lod level indicated by :math:`level`. Used in IfElse
to merge the output if True block and False Block.
Args:
in_true(tuple|list|None): The True branch to be merged.
......@@ -115,7 +117,7 @@ def merge_lod_tensor(in_true, in_false, x, mask, level=0):
x(tuple|list|None): The input tensor that contains complete
lod information needed to construct the output.
mask(list): A bool column vector which masks the input.
level(int): The specific lod level to rank.
level(int): The specific lod level to merge.
Returns:
Variable: The merged output tensor.
......@@ -410,16 +412,17 @@ class StaticRNNMemoryLink(object):
"""
StaticRNNMemoryLink class.
Args:
init: the initial variable for Memory
init: Variable
pre_mem: the memory variable in previous time step
pre_mem: Variable
mem: the memory variable in current time step
mem: Variable
StaticRNNMemoryLink class is used to create a link between two
memory cells of a StaticRNN.
NOTE: This is a internal data structure of a very low-level API.
Please use StaticRNN instead.
Args:
init(Variable): the initial variable for Memory.
pre_mem(Variable): the memory variable in previous time step.
mem(Variable): the memory variable in current time step.
"""
def __init__(self, init, pre_mem, mem=None):
......@@ -819,17 +822,25 @@ def max_sequence_len(rank_table):
def lod_tensor_to_array(x, table):
""" Convert a LOD_TENSOR to an LOD_TENSOR_ARRAY.
"""
Convert a LoDTensor to a LoDTensorArray.
This function split a LoDTesnor to a LoDTensorArray according to its LoD
information. LoDTensorArray is an alias of C++ std::vector<LoDTensor> in
PaddlePaddle. The generated LoDTensorArray of this function can be further read
or written by `read_from_array()` and `write_to_array()` operators. However,
this function is generally an internal component of PaddlePaddle `DynamicRNN`.
Users should not use it directly.
Args:
x (Variable|list): The LOD tensor to be converted to a LOD tensor array.
x (Variable|list): The LoDTensor to be converted to a LoDTensorArray.
table (ParamAttr|list): The variable that stores the level of lod
which is ordered by sequence length in
descending order.
descending order. It is generally generated
by `layers.lod_rank_table()` API.
Returns:
Variable: The variable of type array that has been converted from a
tensor.
Variable: The LoDTensorArray that has been converted from the input tensor.
Examples:
.. code-block:: python
......@@ -894,8 +905,7 @@ def increment(x, value=1.0, in_place=True):
in_place (bool): If the increment should be performed in-place.
Returns:
Variable: The tensor variable storing the transformation of
element-wise increment of each value in the input.
Variable: The elementwise-incremented object.
Examples:
.. code-block:: python
......@@ -937,7 +947,7 @@ def array_write(x, i, array=None):
Variable: The output LOD_TENSOR_ARRAY where the input tensor is written.
Examples:
.. code-block::python
.. code-block:: python
tmp = fluid.layers.zeros(shape=[10], dtype='int32')
i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=10)
......@@ -958,14 +968,17 @@ def array_write(x, i, array=None):
def create_array(dtype):
"""This function creates an array of type :math:`LOD_TENSOR_ARRAY` using the
LayerHelper.
"""
**Create LoDTensorArray**
This function creates an array of LOD_TENSOR_ARRAY . It is mainly used to
implement RNN with array_write, array_read and While.
Args:
dtype (int|float): The data type of the elements in the array.
dtype (int|float): The data type of the elements in the lod_tensor_array.
Returns:
Variable: The tensor variable storing the elements of data type.
Variable: The lod_tensor_array variable storing the elements of data type.
Examples:
.. code-block:: python
......@@ -1048,16 +1061,34 @@ def equal(x, y, cond=None, **ignored):
def array_read(array, i):
"""This function performs the operation to read the data in as an
"""
This function performs the operation to read the data in as an
LOD_TENSOR_ARRAY.
.. code-block:: text
Given:
array = [0.6, 0.1, 0.3, 0.1]
And:
i = 2
Then:
output = 0.3
Args:
array (Variable|list): The input tensor that will be written to an array.
i (Variable|list): The subscript index in tensor array, that points the
place where data will be written to.
array (Variable|list): The input tensor that store data to be read.
i (Variable|list): The index of the data to be read from input array.
Returns:
Variable: The tensor type variable that has the data written to it.
Examples:
.. code-block::python
.. code-block:: python
tmp = fluid.layers.zeros(shape=[10], dtype='int32')
i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=10)
arr = layers.array_read(tmp, i=i)
......@@ -1114,9 +1145,14 @@ def shrink_memory(x, i, table):
def array_length(array):
"""This function performs the operation to find the length of the input
"""
**Get the Length of Input LoDTensorArray**
This function performs the operation to find the length of the input
LOD_TENSOR_ARRAY.
Related API: array_read, array_write, While.
Args:
array (LOD_TENSOR_ARRAY): The input array that will be used
to compute the length.
......@@ -1125,12 +1161,13 @@ def array_length(array):
Variable: The length of the input LoDTensorArray.
Examples:
.. code-block::python
.. code-block:: python
tmp = fluid.layers.zeros(shape=[10], dtype='int32')
i = fluid.layers.fill_constant(shape=[1], dtype='int64', value=10)
arr = fluid.layers.array_write(tmp, i=i)
arr_len = fluid.layers.array_length(arr)
"""
helper = LayerHelper('array_length', **locals())
tmp = helper.create_tmp_variable(dtype='int64')
......@@ -1141,6 +1178,13 @@ def array_length(array):
class ConditionalBlockGuard(BlockGuard):
"""
ConditionalBlockGuard is derived from BlockGuard. It is dedicated for
holding a ConditionalBlock, and helping users entering and exiting the
ConditionalBlock via Python's 'with' keyword. However, ConditionalBlockGuard
is generally an internal component of IfElse, users should not use it directly.
"""
def __init__(self, block):
if not isinstance(block, ConditionalBlock):
raise TypeError("block should be conditional block")
......@@ -1214,6 +1258,42 @@ class ConditionalBlock(object):
class Switch(object):
"""
Switch class works just like a `if-elif-else`. Can be used in learning rate scheduler
to modify learning rate
The Semantics:
1. A `switch` control-flow checks cases one-by-one.
2. The condition of each case is a boolean value, which is a scalar Variable.
3. It runs the first matched case, or the default case if there is one.
4. Once it matches a case, it runs the corresponding branch and only that branch.
Examples:
.. code-block:: python
lr = fluid.layers.tensor.create_global_var(
shape=[1],
value=0.0,
dtype='float32',
persistable=True,
name="learning_rate")
one_var = tensor.fill_constant(
shape=[1], dtype='float32', value=1.0)
two_var = tensor.fill_constant(
shape=[1], dtype='float32', value=2.0)
with fluid.layers.control_flow.Switch() as switch:
with switch.case(global_step == zero_var):
fluid.layers.tensor.assign(input=one_var, output=lr)
with switch.default():
fluid.layers.tensor.assign(input=two_var, output=lr)
"""
def __init__(self, name=None):
self.helper = LayerHelper('switch', name=name)
self.inside_scope = False
......@@ -1243,7 +1323,8 @@ class Switch(object):
return ConditionalBlockGuard(cond_block)
def default(self):
"""create a default case for this switch
"""
create a default case for this switch
"""
pre_cond_num = len(self.pre_not_conditions)
if pre_cond_num == 0:
......@@ -1825,26 +1906,26 @@ def reorder_lod_tensor_by_rank(x, rank_table):
def is_empty(x, cond=None, **ignored):
"""
**Is Empty**
This layer returns the truth value of whether the variable is empty.
Test whether a Variable is empty.
Args:
x(Variable): Operand of *is_empty*
cond(Variable|None): Optional output variable to store the result
of *is_empty*
x (Variable): The Variable to be tested.
cond (Variable|None): Output parameter. Returns the test result
of given 'x'. Default: None
Returns:
Variable: The tensor variable storing the output of *is_empty*.
Variable: A bool scalar. True if 'x' is an empty Variable.
Raises:
TypeError: If input cond is not a variable, or cond's dtype is
not bool
not bool.
Examples:
.. code-block:: python
less = fluid.layers.is_empty(x=input)
res = fluid.layers.is_empty(x=input)
# or:
fluid.layers.is_empty(x=input, cond=res)
"""
helper = LayerHelper("is_empty", **locals())
if cond is None:
......
python/paddle/fluid/layers/detection.py
浏览文件 @ f1f8327c
......@@ -620,7 +620,7 @@ def prior_box(input,
offset=0.5,
name=None):
"""
**Prior box operator**
**Prior Box Operator**
Generate prior boxes for SSD(Single Shot MultiBox Detector) algorithm.
Each position of the input produce N prior boxes, N is determined by
......@@ -649,26 +649,30 @@ def prior_box(input,
name(str): Name of the prior box op. Default: None.
Returns:
boxes(Variable): the output prior boxes of PriorBox.
The layout is [H, W, num_priors, 4].
H is the height of input, W is the width of input,
num_priors is the total
box count of each position of input.
Variances(Variable): the expanded variances of PriorBox.
The layout is [H, W, num_priors, 4].
H is the height of input, W is the width of input
num_priors is the total
box count of each position of input
tuple: A tuple with two Variable (boxes, variances)
boxes: the output prior boxes of PriorBox.
The layout is [H, W, num_priors, 4].
H is the height of input, W is the width of input,
num_priors is the total
box count of each position of input.
variances: the expanded variances of PriorBox.
The layout is [H, W, num_priors, 4].
H is the height of input, W is the width of input
num_priors is the total
box count of each position of input
Examples:
.. code-block:: python
box, var = prior_box(
input=conv1,
image=images,
min_sizes=[100.],
flip=True,
clip=True)
box, var = fluid.layers.prior_box(
input=conv1,
image=images,
min_sizes=[100.],
flip=True,
clip=True)
"""
helper = LayerHelper("prior_box", **locals())
dtype = helper.input_dtype()
......@@ -738,11 +742,9 @@ def multi_box_head(inputs,
stride=1,
name=None):
"""
**Prior_boxes**
Generate prior boxes for SSD(Single Shot MultiBox Detector)
algorithm. The details of this algorithm, please refer the
section 2.2 of SSD paper (SSD: Single Shot MultiBox Detector)
section 2.2 of SSD paper `SSD: Single Shot MultiBox Detector
<https://arxiv.org/abs/1512.02325>`_ .
Args:
......@@ -783,24 +785,27 @@ def multi_box_head(inputs,
name(str): Name of the prior box layer. Default: None.
Returns:
mbox_loc(Variable): The predicted boxes' location of the inputs.
The layout is [N, H*W*Priors, 4]. where Priors
is the number of predicted boxes each position of each input.
mbox_conf(Variable): The predicted boxes' confidence of the inputs.
The layout is [N, H*W*Priors, C]. where Priors
is the number of predicted boxes each position of each input
and C is the number of Classes.
boxes(Variable): the output prior boxes of PriorBox.
The layout is [num_priors, 4]. num_priors is the total
box count of each position of inputs.
Variances(Variable): the expanded variances of PriorBox.
The layout is [num_priors, 4]. num_priors is the total
box count of each position of inputs
tuple: A tuple with four Variables. (mbox_loc, mbox_conf, boxes, variances)
mbox_loc: The predicted boxes' location of the inputs. The layout
is [N, H*W*Priors, 4]. where Priors is the number of predicted
boxes each position of each input.
mbox_conf: The predicted boxes' confidence of the inputs. The layout
is [N, H*W*Priors, C]. where Priors is the number of predicted boxes
each position of each input and C is the number of Classes.
boxes: the output prior boxes of PriorBox. The layout is [num_priors, 4].
num_priors is the total box count of each position of inputs.
variances: the expanded variances of PriorBox. The layout is
[num_priors, 4]. num_priors is the total box count of each position of inputs
Examples:
.. code-block:: python
mbox_locs, mbox_confs, box, var = layers.multi_box_head(
mbox_locs, mbox_confs, box, var = fluid.layers.multi_box_head(
inputs=[conv1, conv2, conv3, conv4, conv5, conv5],
image=images,
num_classes=21,
......
python/paddle/fluid/layers/io.py
浏览文件 @ f1f8327c
......@@ -109,10 +109,35 @@ class BlockGuardServ(BlockGuard):
class ListenAndServ(object):
"""
ListenAndServ class.
**ListenAndServ Layer**
ListenAndServ is used to create a rpc server bind and listen
on specific TCP port, this server will run the sub-block when
received variables from clients.
Args:
endpoint(string): IP:port string which the server will listen on.
inputs(list): a list of variables that the server will get from clients.
fan_in(int): how many client are expected to report to this server, default: 1.
optimizer_mode(bool): whether to run the server as a parameter server, default: True.
ListenAndServ class is used to wrap listen_and_serv op to create a server
which can receive variables from clients and run a block.
Examples:
.. code-block:: python
with fluid.program_guard(main):
serv = layers.ListenAndServ(
"127.0.0.1:6170", ["X"], optimizer_mode=False)
with serv.do():
x = layers.data(
shape=[32, 32],
dtype='float32',
name="X",
append_batch_size=False)
fluid.initializer.Constant(value=1.0)(x, main.global_block())
layers.scale(x=x, scale=10.0, out=out_var)
exe = fluid.Executor(place)
exe.run(main)
"""
def __init__(self, endpoint, inputs, fan_in=1, optimizer_mode=True):
......@@ -544,6 +569,41 @@ def shuffle(reader, buffer_size):
def batch(reader, batch_size):
"""
This layer is a reader decorator. It takes a reader and adds
'batching' decoration on it. When reading with the result
decorated reader, output data will be automatically organized
to the form of batches.
Args:
reader(Variable): The reader to be decorated with 'batching'.
batch_size(int): The batch size.
Returns:
Variable: The reader which has been decorated with 'batching'.
Examples:
.. code-block:: python
raw_reader = fluid.layers.io.open_files(filenames=['./data1.recordio',
'./data2.recordio'],
shapes=[(3,224,224), (1)],
lod_levels=[0, 0],
dtypes=['float32', 'int64'],
thread_num=2,
buffer_size=2)
batch_reader = fluid.layers.batch(reader=raw_reader, batch_size=5)
# If we read data with the raw_reader:
# data = fluid.layers.read_file(raw_reader)
# We can only get data instance by instance.
#
# However, if we read data with the batch_reader:
# data = fluid.layers.read_file(batch_reader)
# Each 5 adjacent instances will be automatically combined together
# to become a batch. So what we get('data') is a batch data instead
# of an instance.
"""
return __create_unshared_decorated_reader__(
'create_batch_reader', reader, {'batch_size': int(batch_size)})
......@@ -589,15 +649,41 @@ def parallel(reader):
{})
def read_file(file_obj):
def read_file(reader):
"""
Execute the given reader and get data via it.
A reader is also a Variable. It can be a raw reader generated by
`fluid.layers.open_files()` or a decorated one generated by
`fluid.layers.double_buffer()` and so on.
Args:
reader(Variable): The reader to execute.
Returns:
Tuple[Variable]: Data read via the given reader.
Examples:
.. code-block:: python
data_file = fluid.layers.open_files(
filenames=['mnist.recordio'],
shapes=[(-1, 748), (-1, 1)],
lod_levels=[0, 0],
dtypes=["float32", "int64"])
data_file = fluid.layers.double_buffer(
fluid.layers.batch(data_file, batch_size=64))
input, label = fluid.layers.read_file(data_file)
"""
helper = LayerHelper('read_file')
out = [
helper.create_tmp_variable(
stop_gradient=True, dtype='float32')
for _ in range(len(file_obj.desc.shapes()))
for _ in range(len(reader.desc.shapes()))
]
helper.append_op(
type='read', inputs={'Reader': [file_obj]}, outputs={'Out': out})
type='read', inputs={'Reader': [reader]}, outputs={'Out': out})
if len(out) == 1:
return out[0]
else:
......
python/paddle/fluid/layers/layer_function_generator.py
浏览文件 @ f1f8327c
......@@ -49,6 +49,13 @@ _single_dollar_pattern_ = re.compile(r"\$([^\$]+)\$")
_two_bang_pattern_ = re.compile(r"!!([^!]+)!!")
def escape_math(text):
return _two_bang_pattern_.sub(
r'$$\1$$',
_single_dollar_pattern_.sub(r':math:`\1`',
_two_dollar_pattern_.sub(r"!!\1!!", text)))
def _generate_doc_string_(op_proto):
"""
Generate docstring by OpProto
......@@ -60,12 +67,6 @@ def _generate_doc_string_(op_proto):
str: the document string
"""
def escape_math(text):
return _two_bang_pattern_.sub(
r'$$\1$$',
_single_dollar_pattern_.sub(
r':math:`\1`', _two_dollar_pattern_.sub(r"!!\1!!", text)))
if not isinstance(op_proto, framework_pb2.OpProto):
raise TypeError("OpProto should be `framework_pb2.OpProto`")
......@@ -233,9 +234,6 @@ def autodoc(comment=""):
return __impl__
_inline_math_single_dollar = re.compile(r"\$([^\$]+)\$")
def templatedoc(op_type=None):
"""
Decorator of layer function. It will use the docstring from the layer
......@@ -253,9 +251,6 @@ def templatedoc(op_type=None):
def trim_ending_dot(msg):
return msg.rstrip('.')
def escape_inline_math(msg):
return _inline_math_single_dollar.sub(repl=r':math:`\1`', string=msg)
def __impl__(func):
if op_type is None:
op_type_name = func.__name__
......@@ -269,7 +264,7 @@ def templatedoc(op_type=None):
for line in comment_lines:
line = line.strip()
if len(line) != 0:
comment += escape_inline_math(line)
comment += escape_math(line)
comment += " "
elif len(comment) != 0:
comment += "\n \n "
......
python/paddle/fluid/layers/learning_rate_scheduler.py
浏览文件 @ f1f8327c
......@@ -71,21 +71,40 @@ def noam_decay(d_model, warmup_steps):
def exponential_decay(learning_rate, decay_steps, decay_rate, staircase=False):
"""Applies exponential decay to the learning rate.
"""
Applies exponential decay to the learning rate.
When training a model, it is often recommended to lower the learning rate as the
training progresses. By using this function, the learning rate will be decayed by
'decay_rate' every 'decay_steps' steps.
>>> if staircase == True:
>>> decayed_learning_rate = learning_rate * decay_rate ^ floor(global_step / decay_steps)
>>> else:
>>> decayed_learning_rate = learning_rate * decay_rate ^ (global_step / decay_steps)
```python
decayed_learning_rate = learning_rate *
decay_rate ^ (global_step / decay_steps)
```
Args:
learning_rate: A scalar float32 value or a Variable. This
will be the initial learning rate during training
decay_steps: A Python `int32` number.
decay_rate: A Python `float` number.
staircase: Boolean. If set true, decay the learning rate every decay_steps.
learning_rate(Variable|float): The initial learning rate.
decay_steps(int): See the decay computation above.
decay_rate(float): The decay rate. See the decay computation above.
staircase(Boolean): If True, decay the learning rate at discrete intervals.
Default: False
Returns:
The decayed learning rate
Variable: The decayed learning rate
Examples:
.. code-block:: python
base_lr = 0.1
sgd_optimizer = fluid.optimizer.SGD(
learning_rate=fluid.layers.exponential_decay(
learning_rate=base_lr,
decay_steps=10000,
decay_rate=0.5,
staircase=True))
sgd_optimizer.minimize(avg_cost)
"""
global_step = _decay_step_counter()
......@@ -129,22 +148,39 @@ def natural_exp_decay(learning_rate, decay_steps, decay_rate, staircase=False):
def inverse_time_decay(learning_rate, decay_steps, decay_rate, staircase=False):
"""Applies inverse time decay to the initial learning rate.
"""
Applies inverse time decay to the initial learning rate.
>>> if staircase:
When training a model, it is often recommended to lower the learning rate as the
training progresses. By using this function, an inverse decay function will be
applied to the initial learning rate.
>>> if staircase == True:
>>> decayed_learning_rate = learning_rate / (1 + decay_rate * floor(global_step / decay_step))
>>> else:
>>> decayed_learning_rate = learning_rate / (1 + decay_rate * global_step / decay_step)
Args:
learning_rate: A scalar float32 value or a Variable. This
will be the initial learning rate during training.
decay_steps: A Python `int32` number.
decay_rate: A Python `float` number.
staircase: Boolean. If set true, decay the learning rate every decay_steps.
learning_rate(Variable|float): The initial learning rate.
decay_steps(int): See the decay computation above.
decay_rate(float): The decay rate. See the decay computation above.
staircase(Boolean): If True, decay the learning rate at discrete intervals.
Default: False
Returns:
The decayed learning rate
Variable: The decayed learning rate
Examples:
.. code-block:: python
base_lr = 0.1
sgd_optimizer = fluid.optimizer.SGD(
learning_rate=fluid.layers.inverse_time_decay(
learning_rate=base_lr,
decay_steps=10000,
decay_rate=0.5,
staircase=True))
sgd_optimizer.minimize(avg_cost)
"""
global_step = _decay_step_counter()
......@@ -163,25 +199,28 @@ def polynomial_decay(learning_rate,
end_learning_rate=0.0001,
power=1.0,
cycle=False):
"""Applies polynomial decay to the initial learning rate.
"""
Applies polynomial decay to the initial learning rate.
.. code-block:: python
if cycle:
decay_steps = decay_steps * ceil(global_step / decay_steps)
else:
global_step = min(global_step, decay_steps)
decayed_learning_rate = (learning_rate - end_learning_rate) *
(1 - global_step / decay_steps) ^ power + end_learning_rate
>>> if cycle:
>>> decay_steps = decay_steps * ceil(global_step / decay_steps)
>>> else:
>>> global_step = min(global_step, decay_steps)
>>> decayed_learning_rate = (learning_rate - end_learning_rate) *
>>> (1 - global_step / decay_steps) ^ power +
>>> end_learning_rate
Args:
learning_rate: A scalar float32 value or a Variable. This
will be the initial learning rate during training
decay_steps: A Python `int32` number.
end_learning_rate: A Python `float` number.
power: A Python `float` number
cycle: Boolean. If set true, decay the learning rate every decay_steps.
learning_rate(Variable|float32): A scalar float32 value or a Variable. This
will be the initial learning rate during training.
decay_steps(int32): A Python `int32` number.
end_learning_rate(float): A Python `float` number.
power(float): A Python `float` number.
cycle(bool): If set true, decay the learning rate every decay_steps.
Returns:
The decayed learning rate
Variable: The decayed learning rate
"""
global_step = _decay_step_counter()
......
python/paddle/fluid/layers/metric.py
浏览文件 @ f1f8327c
......@@ -27,8 +27,32 @@ __all__ = ['accuracy', 'auc']
def accuracy(input, label, k=1, correct=None, total=None):
"""
accuracy layer.
Refer to the https://en.wikipedia.org/wiki/Precision_and_recall
This function computes the accuracy using the input and label.
The output is the top k inputs and their indices.
If the correct label occurs in top k predictions, then correct will increment by one.
Note: the dtype of accuracy is determined by input. the input and label dtype can be different.
Args:
input(Variable): The input of accuracy layer, which is the predictions of network.
Carry LoD information is supported.
label(Variable): The label of dataset.
k(int): The top k predictions for each class will be checked.
correct(Variable): The correct predictions count.
total(Variable): The total entries count.
Returns:
Variable: The correct rate.
Examples:
.. code-block:: python
data = fluid.layers.data(name="data", shape=[-1, 32, 32], dtype="float32")
label = fluid.layers.data(name="data", shape=[-1,1], dtype="int32")
predict = fluid.layers.fc(input=data, size=10)
acc = fluid.layers.accuracy(input=predict, label=label, k=5)
"""
helper = LayerHelper("accuracy", **locals())
topk_out, topk_indices = nn.topk(input, k=k)
......
python/paddle/fluid/layers/nn.py
浏览文件 @ f1f8327c
此差异已折叠。
python/paddle/fluid/layers/ops.py
浏览文件 @ f1f8327c
此差异已折叠。
Markdown is supported
0% 或 .
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 或