Skip to content

P
Paddle

PaddlePaddle / Paddle
接近 2 年 前同步成功

通知 2323
Star 20933
Fork 5424
P
Paddle
提交 9942a304 编写于 作者: Y yuyang18
浏览文件
操作

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

上级 1c19f1ab 1d7e60fd
显示空白变更内容
内联 并排
Showing with 3564 addition and 1545 deletion
benchmark/fluid/fluid_benchmark.py
浏览文件 @ 9942a304
......@@ -97,7 +97,7 @@ def dist_transpile(trainer_id, args):
return train_program, fluid.default_startup_program()
else:
raise ValueError(
'TRAINING_ROLE environment variable must be either TRAINER or PSERVER'
'PADDLE_TRAINING_ROLE environment variable must be either TRAINER or PSERVER'
)
......
benchmark/fluid/kube_gen_job.py
浏览文件 @ 9942a304
......@@ -108,10 +108,10 @@ def gen_job():
tn_container["ports"][0]["containerPort"] = spreadport
envs.append({"name": "PADDLE_JOB_NAME", "value": args.jobname})
envs.append({"name": "TRAINERS", "value": str(args.trainers)})
envs.append({"name": "PADDLE_TRAINERS", "value": str(args.trainers)})
envs.append({"name": "PSERVERS", "value": str(args.pservers)})
envs.append({"name": "ENTRY", "value": args.entry})
envs.append({"name": "PADDLE_INIT_PORT", "value": str(args.port)})
envs.append({"name": "PADDLE_PSERVER_PORT", "value": str(args.port)})
envs.append({"name": "PADDLE_PSERVER_PORT", "value": str(args.port)})
# NOTE: these directories below are cluster specific, please modify
# this settings before you run on your own cluster.
......@@ -167,16 +167,22 @@ def gen_job():
tn_container["volumeMounts"] = volumeMounts
ps_container["env"] = envs
ps_container["env"].append({"name": "TRAINING_ROLE", "value": "PSERVER"})
ps_container["env"].append({
"name": "PADDLE_TRAINING_ROLE",
"value": "PSERVER"
})
tn_container["env"] = envs
if args.disttype == "pserver":
tn_container["env"].append({
"name": "TRAINING_ROLE",
"name": "PADDLE_TRAINING_ROLE",
"value": "TRAINER"
})
elif args.disttype == "nccl2" or args.disttype == "local":
# NCCL2 have no training role, set to plain WORKER
tn_container["env"].append({"name": "TRAINING_ROLE", "value": "WORKER"})
tn_container["env"].append({
"name": "PADDLE_TRAINING_ROLE",
"value": "WORKER"
})
os.mkdir(args.jobname)
if args.disttype == "pserver":
......
cmake/external/mkldnn.cmake
浏览文件 @ 9942a304
......@@ -45,7 +45,8 @@ IF(${CBLAS_PROVIDER} STREQUAL "MKLML")
ELSE()
MESSAGE(FATAL_ERROR "Should enable MKLML when build MKLDNN")
ENDIF()
SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result -Wno-unused-result")
SET(MKLDNN_FLAG "-Wno-error=strict-overflow -Wno-error=unused-result")
SET(MKLDNN_FLAG "${MKLDNN_FLAG} -Wno-unused-result -Wno-unused-value")
SET(MKLDNN_CFLAG "${CMAKE_C_FLAGS} ${MKLDNN_FLAG}")
SET(MKLDNN_CXXFLAG "${CMAKE_CXX_FLAGS} ${MKLDNN_FLAG}")
ExternalProject_Add(
......
doc/fluid/howto/cluster/fluid_cluster_train_cn.md
浏览文件 @ 9942a304
......@@ -168,13 +168,13 @@ cd /paddle/python/paddle/fluid/tests/book
第二步，启动Parameter Server：
```bash
PADDLE_INIT_PORT=6174 PADDLE_INIT_PSERVERS=192.168.1.2 TRAINERS=2 POD_IP=192.168.1.2 PADDLE_INIT_TRAINER_ID=1 TRAINING_ROLE=PSERVER python test_fit_a_line.py
PADDLE_PSERVER_PORT=6174 PADDLE_PSERVER_IPS=192.168.1.2 PADDLE_TRAINERS=2 PADDLE_CURRENT_IP=192.168.1.2 PADDLE_TRAINER_ID=1 PADDLE_TRAINING_ROLE=PSERVER python test_fit_a_line.py
```
执行命令后请等待出现提示： ```Server listening on 192.168.1.2:6174 ```, 表示Paramter Server已经正常启动。
第三步，启动Trainer：
```bash
PADDLE_INIT_PORT=6174 PADDLE_INIT_PSERVERS=192.168.1.3 TRAINERS=2 POD_IP=192.168.1.3 PADDLE_INIT_TRAINER_ID=1 TRAINING_ROLE=TRAINER python test_fit_a_line.py
PADDLE_PSERVER_PORT=6174 PADDLE_PSERVER_IPS=192.168.1.3 PADDLE_TRAINERS=2 PADDLE_CURRENT_IPP=192.168.1.3 PADDLE_TRAINER_ID=1 PADDLE_TRAINING_ROLE=TRAINER python test_fit_a_line.py
```
由于我们定义的Trainer的数量是2个，因此需要在另外一个计算节点上再启动一个Trainer。
......
doc/fluid/howto/cluster/fluid_recordio.md
浏览文件 @ 9942a304
......@@ -114,8 +114,8 @@ def gen_train_list(file_pattern, trainers, trainer_id):
ret_list.append(f)
return ret_list
trainers = int(os.getenv("TRAINERS"))
trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
trainers = int(os.getenv("PADDLE_TRAINERS"))
trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
data_file = fluid.layers.io.open_files(
filenames=gen_train_list("./mnist-[0-9]*.recordio", 2, 0),
thread_num=1,
......
paddle/contrib/CMakeLists.txt
浏览文件 @ 9942a304
......@@ -14,4 +14,3 @@
#
add_subdirectory(inference)
add_subdirectory(tape)
paddle/contrib/tape/README.md 已删除 100644 → 0
浏览文件 @ 1c19f1ab
# 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 @ 1c19f1ab
// 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 @ 1c19f1ab
// 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 @ 1c19f1ab
// 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/framework/executor.cc
浏览文件 @ 9942a304
......@@ -295,13 +295,14 @@ void Executor::Run(const ProgramDesc& program, Scope* scope,
std::unique_ptr<ExecutorPrepareContext> Executor::Prepare(
const ProgramDesc& program, int block_id) {
auto* ctx = new ExecutorPrepareContext(program, block_id);
std::unique_ptr<ExecutorPrepareContext> ctx(
new ExecutorPrepareContext(program, block_id));
PADDLE_ENFORCE_LT(static_cast<size_t>(block_id), program.Size());
auto& block = program.Block(block_id);
for (auto& op_desc : block.AllOps()) {
ctx->ops_.push_back(OpRegistry::CreateOp(*op_desc));
}
return std::unique_ptr<ExecutorPrepareContext>(ctx);
return ctx;
}
std::vector<std::shared_ptr<ExecutorPrepareContext>> Executor::Prepare(
......
paddle/fluid/inference/analysis/CMakeLists.txt
浏览文件 @ 9942a304
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(test_dfg_graphviz_draw_pass
SRCS dfg_graphviz_draw_pass_tester.cc
cc_test(${TARGET}
SRCS "${analysis_test_SRCS}"
DEPS analysis
ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model)
set_tests_properties(test_dfg_graphviz_draw_pass PROPERTIES DEPENDS test_word2vec)
ARGS --inference_model_dir=${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)
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 @ 9942a304
......@@ -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 @ 9942a304
......@@ -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 @ 9942a304
......@@ -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()));
}
}
// Add edges
......
paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.cc 0 → 100644
浏览文件 @ 9942a304
// 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 @ 9942a304
/* 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 @ 9942a304
......@@ -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 @ 9942a304
/* 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 @ 9942a304
......@@ -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 @ 9942a304
......@@ -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 @ 9942a304
......@@ -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 @ 9942a304
......@@ -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 @ 9942a304
......@@ -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 @ 9942a304
......@@ -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 @ 9942a304
......@@ -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 @ 9942a304
......@@ -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.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 @ 9942a304
......@@ -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 @ 9942a304
/* 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 @ 9942a304
/* 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 @ 9942a304
/* 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 @ 9942a304
......@@ -19,12 +19,7 @@ namespace paddle {
namespace inference {
namespace analysis {
TEST_F(DFG_Tester, Split) {
auto desc = LoadProgramDesc();
auto dfg = ProgramDescToDFG(desc);
LOG(INFO) << "spliter\n" << dfg.DotString();
SubGraphSplitter::NodeInsideSubgraphTeller teller = [](const Node* node) {
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" ||
......@@ -34,7 +29,13 @@ TEST_F(DFG_Tester, Split) {
return true;
}
return false;
};
};
TEST_F(DFG_Tester, Split) {
auto desc = LoadProgramDesc();
auto dfg = ProgramDescToDFG(desc);
LOG(INFO) << "spliter\n" << dfg.DotString();
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 @ 9942a304
......@@ -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 @ 9942a304
/* 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 @ 9942a304
/* 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 @ 9942a304
......@@ -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/operators/activation_mkldnn_op.cc
浏览文件 @ 9942a304
......@@ -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 @ 9942a304
......@@ -19,6 +19,8 @@ limitations under the License. */
namespace paddle {
namespace operators {
using paddle::framework::Tensor;
#define REGISTER_ACTIVATION_OP_MAKER(OP_NAME, OP_COMMENT) \
class OP_NAME##OpMaker \
: public ::paddle::framework::OpProtoAndCheckerMaker { \
......@@ -27,9 +29,9 @@ namespace operators {
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") \
"(bool, default false) Only used in mkldnn kernel") \
.SetDefault(false); \
AddComment(OP_COMMENT); \
AddComment(#OP_COMMENT); \
} \
}
......@@ -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");
......@@ -140,7 +143,7 @@ $$out = \\frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$
__attribute__((unused)) constexpr char TanhShrinkDoc[] = R"DOC(
TanhShrink Activation Operator.
$$out = x - \frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$
$$out = x - \\frac{e^{x} - e^{-x}}{e^{x} + e^{-x}}$$
)DOC";
......@@ -382,7 +385,7 @@ class STanhOpMaker : public framework::OpProtoAndCheckerMaker {
AddComment(R"DOC(
STanh Activation Operator.
$$out = b * \frac{e^{a * x} - e^{-a * x}}{e^{a * x} + e^{-a * x}}$$
$$out = b * \\frac{e^{a * x} - e^{-a * x}}{e^{a * x} + e^{-a * x}}$$
)DOC");
}
......
paddle/fluid/operators/concat_op.h
浏览文件 @ 9942a304
......@@ -60,34 +60,45 @@ template <typename DeviceContext, typename T>
class ConcatGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const {
auto* in = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
auto* out_grad =
ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
auto ins = ctx.MultiInput<framework::Tensor>("X");
auto out_var_names = ctx.Outputs(framework::GradVarName("X"));
auto outs = ctx.MultiOutput<framework::Tensor>(framework::GradVarName("X"));
int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
// get output tensor that the name is not kEmptyVarName
std::vector<framework::Tensor*> outputs;
for (size_t j = 0; j < outs.size(); ++j) {
if (out_var_names[j] != framework::kEmptyVarName) {
outs[j]->mutable_data<T>(ctx.GetPlace());
outputs.push_back(outs[j]);
} else {
outputs.push_back(nullptr);
}
}
// Sometimes direct copies will be faster, this maybe need deeply analysis.
if (axis == 0 && outs.size() < 10) {
size_t input_offset = 0;
auto in_stride = framework::stride_numel(in->dims());
const auto in_stride = framework::stride_numel(out_grad->dims());
for (auto& out : outs) {
out->mutable_data<T>(ctx.GetPlace());
auto out_stride = framework::stride_numel(out->dims());
StridedNumelCopyWithAxis<T>(ctx.device_context(), axis, out->data<T>(),
out_stride, in->data<T>() + input_offset,
in_stride, out_stride[axis]);
for (size_t i = 0; i < outs.size(); ++i) {
auto out_stride = framework::stride_numel(ins[i]->dims());
auto* out = outputs[i];
if (out != nullptr) {
StridedNumelCopyWithAxis<T>(
ctx.device_context(), axis, out->data<T>(), out_stride,
out_grad->data<T>() + input_offset, in_stride, out_stride[axis]);
}
input_offset += out_stride[axis];
}
} else {
std::vector<framework::Tensor> outputs(outs.size());
for (size_t j = 0; j < outs.size(); ++j) {
outs[j]->mutable_data<T>(ctx.GetPlace());
outputs[j] = *outs[j];
}
auto& dev_ctx = ctx.template device_context<DeviceContext>();
paddle::operators::math::ConcatGradFunctor<DeviceContext, T>
concat_grad_functor;
concat_grad_functor(dev_ctx, *in, static_cast<int>(axis), &outputs);
concat_grad_functor(dev_ctx, *out_grad, ins, static_cast<int>(axis),
&outputs);
}
}
};
......
paddle/fluid/operators/detection_map_op.cc
浏览文件 @ 9942a304
......@@ -175,12 +175,12 @@ class DetectionMAPOpMaker : public framework::OpProtoAndCheckerMaker {
AddComment(R"DOC(
Detection mAP evaluate operator.
The general steps are as follows. First, calculate the true positive and
false positive according to the input of detection and labels, then
calculate the mAP evaluate value.
Supporting '11 point' and 'integral' mAP algorithm. Please get more information
from the following articles:
https://sanchom.wordpress.com/tag/average-precision/
https://arxiv.org/abs/1512.02325
false positive according to the input of detection and labels, then
calculate the mAP evaluate value.
Supporting '11 point' and 'integral' mAP algorithm. Please get more information
from the following articles:
https://sanchom.wordpress.com/tag/average-precision/
https://arxiv.org/abs/1512.02325
)DOC");
}
......
paddle/fluid/operators/gaussian_random_mkldnn_op.cc 0 → 100644
浏览文件 @ 9942a304
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <string>
#include "paddle/fluid/operators/mean_op.h"
namespace paddle {
namespace operators {
using framework::DataLayout;
template <typename T>
class GaussianMKLDNNKernel : public paddle::framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
float mean = context.Attr<float>("mean");
float std = context.Attr<float>("std");
auto* tensor = context.Output<framework::Tensor>("Out");
T* data = tensor->mutable_data<T>(context.GetPlace());
unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
std::minstd_rand engine;
if (seed == 0) {
seed = std::random_device()();
}
engine.seed(seed);
std::normal_distribution<T> dist(mean, std);
int64_t size = tensor->numel();
for (int64_t i = 0; i < size; ++i) {
data[i] = dist(engine);
}
// The format of output is set as the mkldnn's format
// TODO(@mozga-intel) The format of matrix sets inside the another layers.
tensor->set_layout(DataLayout::kMKLDNN);
tensor->set_format(mkldnn::memory::format::oihw);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_KERNEL(gaussian_random, MKLDNN, ::paddle::platform::CPUPlace,
ops::GaussianMKLDNNKernel<float>);
paddle/fluid/operators/gaussian_random_op.cc
浏览文件 @ 9942a304
......@@ -15,6 +15,10 @@ limitations under the License. */
#include <random>
#include "paddle/fluid/framework/op_registry.h"
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
namespace paddle {
namespace operators {
......@@ -62,9 +66,20 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
framework::LibraryType library{framework::LibraryType::kPlain};
framework::DataLayout layout{framework::DataLayout::kAnyLayout};
#ifdef PADDLE_WITH_MKLDNN
if (library == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library = framework::LibraryType::kMKLDNN;
layout = framework::DataLayout::kMKLDNN;
}
#endif
return framework::OpKernelType(
static_cast<framework::proto::VarType::Type>(ctx.Attr<int>("dtype")),
ctx.device_context());
ctx.device_context(), layout, library);
}
};
......@@ -95,7 +110,9 @@ class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker {
"(int, default 5(FP32)) "
"Output data type.")
.SetDefault(framework::proto::VarType::FP32);
AddAttr<bool>("use_mkldnn",
"(bool, default false) Only used in mkldnn kernel")
.SetDefault(false);
AddComment(R"DOC(
GaussianRandom Operator.
......
paddle/fluid/operators/math/concat.cc
浏览文件 @ 9942a304
......@@ -70,35 +70,40 @@ template <typename T>
class ConcatGradFunctor<platform::CPUDeviceContext, T> {
public:
void operator()(const platform::CPUDeviceContext& context,
const framework::Tensor& input, const int axis,
std::vector<framework::Tensor>* outputs) {
const framework::Tensor& input,
const std::vector<const framework::Tensor*>& ref_inputs,
const int axis, std::vector<framework::Tensor*>* outputs) {
// TODO(zcd): Add input data validity checking
int num = outputs->size();
size_t num = outputs->size();
int input_rows = 1;
auto dim_0 = outputs->at(0).dims();
auto dim_0 = ref_inputs[0]->dims();
for (int i = 0; i < axis; ++i) {
input_rows *= dim_0[i];
}
int input_cols = 0;
std::vector<int64_t> output_cols(outputs->size());
for (int i = 0; i < num; ++i) {
int t_cols = outputs->at(i).numel() / input_rows;
for (size_t i = 0; i < num; ++i) {
int t_cols = ref_inputs[i]->numel() / input_rows;
input_cols += t_cols;
output_cols[i] = t_cols;
}
auto cpu_place = boost::get<platform::CPUPlace>(context.GetPlace());
// computation
for (int k = 0; k < input_rows; ++k) {
for (size_t k = 0; k < input_rows; ++k) {
const T* src_ptr = input.data<T>() + k * input_cols;
int col_idx = 0;
for (int j = 0; j < num; ++j) {
int col_len = output_cols[j];
T* dst_ptr = outputs->at(j).data<T>() + k * col_len;
auto* out_tensor = outputs->at(j);
if (out_tensor != nullptr) {
T* dst_ptr = out_tensor->data<T>() + k * col_len;
memory::Copy(cpu_place, dst_ptr, cpu_place, src_ptr + col_idx,
sizeof(T) * col_len);
}
col_idx += col_len;
}
}
......
paddle/fluid/operators/math/concat.cu
浏览文件 @ 9942a304
......@@ -22,43 +22,24 @@ namespace paddle {
namespace operators {
namespace math {
template <typename T>
__device__ T upper_bound(const T* first, T count, T val) {
const T* orig = first;
const T* it = nullptr;
T step = 0;
while (count > 0) {
it = first;
step = count / 2;
it += step;
if (!(val < *it)) {
first = ++it;
count -= step + 1;
} else {
count = step;
}
}
return first - orig;
}
template <typename T>
__global__ void KernelConcat(T** inputs, const int* input_cols, int col_size,
const int output_rows, const int output_cols,
T* output) {
int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
int segment = upper_bound<int>(input_cols, col_size, tid_x) - 1;
int curr_offset = input_cols[segment];
int curr_segment = segment;
int curr_segment = 0;
int curr_offset = input_cols[0];
for (; tid_x < output_cols; tid_x += blockDim.x * gridDim.x) {
T curr_col_offset;
while ((curr_col_offset = input_cols[curr_segment + 1]) <= tid_x) {
int curr_col_offset = input_cols[curr_segment + 1];
while (curr_col_offset <= tid_x) {
curr_offset = curr_col_offset;
++curr_segment;
curr_col_offset = input_cols[curr_segment + 1];
}
int local_col = tid_x - curr_offset;
int segment_width = curr_col_offset - curr_offset;
T* input_ptr = inputs[curr_segment];
int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
for (; tid_y < output_rows; tid_y += blockDim.y * gridDim.y)
......@@ -89,24 +70,26 @@ __global__ void KernelConcatGrad(const T* input_data, const int in_row,
const int in_col, const int* out_cols,
int out_cols_size, T** outputs_data) {
int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
int segment = upper_bound<int>(out_cols, out_cols_size, tid_x) - 1;
int curr_offset = out_cols[segment];
int curr_segment = segment;
int curr_segment = 0;
int curr_offset = out_cols[0];
for (; tid_x < in_col; tid_x += blockDim.x * gridDim.x) {
T curr_col_offset;
while ((curr_col_offset = out_cols[curr_segment + 1]) <= tid_x) {
int curr_col_offset = out_cols[curr_segment + 1];
while (curr_col_offset <= tid_x) {
curr_offset = curr_col_offset;
++curr_segment;
curr_col_offset = out_cols[curr_segment + 1];
}
int local_col = tid_x - curr_offset;
int segment_width = curr_col_offset - curr_offset;
T* output_ptr = outputs_data[curr_segment];
if (output_ptr != nullptr) {
int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y)
output_ptr[tid_y * segment_width + local_col] =
input_data[tid_y * in_col + tid_x];
}
}
}
template <typename T>
......@@ -118,11 +101,13 @@ __global__ void KernelConcatGrad(const T* input_data, const int in_row,
int split = tid_x / fixed_out_col;
int in_offset = tid_x - split * fixed_out_col;
T* output_ptr = outputs_data[split];
if (output_ptr != nullptr) {
int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y)
output_ptr[tid_y * fixed_out_col + in_offset] =
input_data[tid_y * in_col + tid_x];
}
}
}
/*
......@@ -203,17 +188,18 @@ template <typename T>
class ConcatGradFunctor<platform::CUDADeviceContext, T> {
public:
void operator()(const platform::CUDADeviceContext& context,
const framework::Tensor& input, const int axis,
std::vector<framework::Tensor>* outputs) {
const framework::Tensor& input,
const std::vector<const framework::Tensor*>& ref_inputs,
const int axis, std::vector<framework::Tensor*>* outputs) {
// TODO(zcd): Add input data validity checking
int o_num = outputs->size();
int out_row = 1;
auto dim_0 = outputs->at(0).dims();
auto dim_0 = ref_inputs[0]->dims();
for (int i = 0; i < axis; ++i) {
out_row *= dim_0[i];
}
int out_col = outputs->at(0).numel() / out_row;
int out0_col = ref_inputs[0]->numel() / out_row;
int in_col = 0, in_row = out_row;
bool sameShape = true;
......@@ -223,13 +209,17 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
outputs_cols[0] = 0;
for (int i = 0; i < o_num; ++i) {
int t_col = outputs->at(i).numel() / out_row;
int t_col = ref_inputs.at(i)->numel() / out_row;
if (sameShape) {
if (t_col != out_col) sameShape = false;
if (t_col != out0_col) sameShape = false;
}
in_col += t_col;
outputs_cols[i + 1] = in_col;
outputs_ptr[i] = outputs->at(i).data<T>();
if (outputs->at(i) != nullptr) {
outputs_ptr[i] = outputs->at(i)->data<T>();
} else {
outputs_ptr[i] = nullptr;
}
}
T** dev_out_gpu_data =
......@@ -255,7 +245,7 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
if (sameShape) {
KernelConcatGrad<<<grid_size, block_size, 0, context.stream()>>>(
input.data<T>(), in_row, in_col, out_col, dev_out_gpu_data);
input.data<T>(), in_row, in_col, out0_col, dev_out_gpu_data);
} else {
const int* dev_outs_col_data = outputs_cols.CUDAData(context.GetPlace());
KernelConcatGrad<<<grid_size, block_size, 0, context.stream()>>>(
......
paddle/fluid/operators/math/concat.h
浏览文件 @ 9942a304
......@@ -57,7 +57,8 @@ template <typename DeviceContext, typename T>
class ConcatGradFunctor {
public:
void operator()(const DeviceContext& context, const framework::Tensor& input,
const int axis, std::vector<framework::Tensor>* outputs);
const std::vector<const framework::Tensor*>& ref_inputs,
const int axis, std::vector<framework::Tensor*>* outputs);
};
} // namespace math
......
paddle/fluid/operators/parallel_do_op.cc
浏览文件 @ 9942a304
......@@ -295,7 +295,7 @@ class ParallelDoGradOp : public framework::OperatorBase {
auto sum_op = framework::OpRegistry::CreateOp(
"sum", {{"X", {s, tmp_name}}}, {{"Out", {s}}},
framework::AttributeMap{});
framework::AttributeMap{{"use_mkldnn", {false}}});
VLOG(10) << sum_op->DebugStringEx(sub_scopes[0]);
sum_op->Run(*sub_scopes[0], places[0]);
WaitOnPlace(places[0]);
......
paddle/fluid/operators/recurrent_op.cc
浏览文件 @ 9942a304
......@@ -429,7 +429,8 @@ class RecurrentGradOp : public RecurrentBase {
auto sum_op = framework::OpRegistry::CreateOp(
"sum", {{"X", {pg_names[param_id], new_inside_name}}},
{{"Out", {pg_names[param_id]}}}, framework::AttributeMap{});
{{"Out", {pg_names[param_id]}}},
framework::AttributeMap{{"use_mkldnn", {false}}});
sum_op->Run(cur_scope, place);
cur_scope.Rename(new_inside_name, inside_grad_name);
......
paddle/fluid/operators/sum_mkldnn_op.cc 0 → 100644
浏览文件 @ 9942a304
// 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.
/*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 "mkldnn.hpp"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/operators/math/selected_rows_functor.h"
#include "paddle/fluid/operators/sum_op.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/mkldnn_helper.h"
namespace paddle {
namespace operators {
using paddle::framework::Tensor;
using paddle::platform::MKLDNNDeviceContext;
using paddle::platform::CPUDeviceContext;
using framework::DataLayout;
using mkldnn::memory;
using mkldnn::primitive;
using mkldnn::stream;
using mkldnn::sum;
using mkldnn::reorder;
using platform::to_void_cast;
template <typename T>
class SumMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
public:
void Compute(const paddle::framework::ExecutionContext& ctx) const override {
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();
auto in_vars = ctx.MultiInputVar("X");
const int N = in_vars.size();
auto out_var = ctx.OutputVar("Out");
bool in_place = out_var == in_vars[0];
if (out_var->IsType<framework::LoDTensor>()) {
LoDTensor* output = ctx.Output<LoDTensor>("Out");
T* output_data = output->mutable_data<T>(ctx.GetPlace());
std::vector<int> dst_tz = framework::vectorize2int(output->dims());
auto src_tz = dst_tz;
memory::format output_format{memory::format::format_undef};
std::vector<float> scales;
std::vector<memory::primitive_desc> srcs_mpd;
std::vector<mkldnn::memory> srcs_mem;
PADDLE_ENFORCE(in_vars[0]->IsType<LoDTensor>(),
"Input[0] must be LoDTensors");
auto& input0 = in_vars[0]->Get<LoDTensor>();
PADDLE_ENFORCE(input0.layout() == DataLayout::kMKLDNN &&
input0.format() != memory::format::format_undef,
"Wrong layout/format for inputs[0]");
memory::format input_format = input0.format();
if (src_tz.size() == 1 && (input_format == memory::format::nchw ||
input_format == memory::format::nhwc)) {
input_format = memory::format::x;
}
if (src_tz.size() == 2 && (input_format == memory::format::nchw ||
input_format == memory::format::nhwc)) {
input_format = memory::format::nc;
}
for (int i = in_place ? 1 : 0; i < N; i++) {
PADDLE_ENFORCE(in_vars[i]->IsType<LoDTensor>(),
"all inputs must be all LoDTensors");
auto& input = in_vars[i]->Get<LoDTensor>();
PADDLE_ENFORCE(input.layout() == DataLayout::kMKLDNN &&
input.format() != memory::format::format_undef,
"Wrong layout/format for inputs");
if (input.numel() == 0) {
continue;
}
const T* input_data = input.data<T>();
auto src_md =
memory::desc(src_tz, memory::data_type::f32, input_format);
auto src_mpd = memory::primitive_desc(src_md, mkldnn_engine);
auto src_mem = memory(src_mpd, to_void_cast(input_data));
srcs_mpd.push_back(src_mpd);
srcs_mem.push_back(src_mem);
scales.push_back(1.0);
}
auto dst_md =
memory::desc(dst_tz, memory::data_type::f32, memory::format::any);
auto sum_pd = sum::primitive_desc(dst_md, scales, srcs_mpd);
std::shared_ptr<memory> dst_mem;
if (in_place) {
dst_mem.reset(new memory(sum_pd.dst_primitive_desc()));
} else {
dst_mem.reset(new memory(sum_pd.dst_primitive_desc(), output_data));
}
std::vector<mkldnn::primitive::at> inputs;
for (size_t i = 0; i < srcs_mem.size(); ++i) {
inputs.push_back(srcs_mem[i]);
}
auto sum_prim = mkldnn::sum(sum_pd, inputs, *dst_mem);
output_format = (memory::format)platform::GetMKLDNNFormat(sum_pd);
primitive reorder_prim;
std::shared_ptr<memory> target_mem;
if (in_place) {
output_format = input_format;
target_mem.reset(new memory(
{{{src_tz}, memory::data_type::f32, output_format}, mkldnn_engine},
output_data));
reorder_prim = reorder(*dst_mem, *target_mem);
}
std::vector<primitive> pipeline;
pipeline.push_back(sum_prim);
if (in_place) pipeline.push_back(reorder_prim);
stream(stream::kind::eager).submit(pipeline).wait();
output->set_layout(DataLayout::kMKLDNN);
output->set_format(output_format);
} else if (out_var->IsType<framework::SelectedRows>()) {
// TODO(@mozga-intel) Add MKLDNN SelectedRows support
std::unique_ptr<framework::SelectedRows> in0;
if (in_place) {
// If is in_place, we store the input[0] to in0
auto& in_sel0 = in_vars[0]->Get<SelectedRows>();
auto& rows = in_sel0.rows();
in0.reset(new framework::SelectedRows(rows, in_sel0.height()));
in0->mutable_value()->ShareDataWith(in_sel0.value());
}
auto get_selected_row = [&](size_t i) -> const SelectedRows& {
if (i == 0 && in0) {
return *in0.get();
} else {
return in_vars[i]->Get<SelectedRows>();
}
};
auto* out = ctx.Output<SelectedRows>("Out");
out->mutable_rows()->clear();
auto* out_value = out->mutable_value();
// Runtime InferShape
size_t first_dim = 0;
for (int i = 0; i < N; i++) {
auto& sel_row = get_selected_row(i);
first_dim += sel_row.rows().size();
}
auto in_dim =
framework::vectorize(get_selected_row(N - 1).value().dims());
in_dim[0] = static_cast<int64_t>(first_dim);
out_value->Resize(framework::make_ddim(in_dim));
// if all the input sparse vars are empty, no need to
// merge these vars.
if (first_dim == 0UL) {
return;
}
out_value->mutable_data<T>(ctx.GetPlace());
math::SelectedRowsAddTo<CPUDeviceContext, T> functor;
int64_t offset = 0;
for (int i = 0; i < N; i++) {
auto& sel_row = get_selected_row(i);
if (sel_row.rows().size() == 0) {
continue;
}
PADDLE_ENFORCE_EQ(out->height(), sel_row.height());
functor(ctx.template device_context<CPUDeviceContext>(), sel_row,
offset, out);
offset += sel_row.value().numel();
}
} else if (out_var->IsType<framework::LoDTensorArray>()) {
// TODO(@mozga-intel) Add MKLDNN LoDTensorArray support
auto& out_array = *out_var->GetMutable<framework::LoDTensorArray>();
for (size_t i = in_place ? 1 : 0; i < in_vars.size(); ++i) {
PADDLE_ENFORCE(in_vars[i]->IsType<framework::LoDTensorArray>(),
"Only support all inputs are TensorArray");
auto& in_array = in_vars[i]->Get<framework::LoDTensorArray>();
for (size_t i = 0; i < in_array.size(); ++i) {
if (in_array[i].numel() != 0) {
if (i >= out_array.size()) {
out_array.resize(i + 1);
}
if (out_array[i].numel() == 0) {
framework::TensorCopy(in_array[i], in_array[i].place(),
ctx.device_context(), &out_array[i]);
out_array[i].set_lod(in_array[i].lod());
} else {
PADDLE_ENFORCE(out_array[i].lod() == in_array[i].lod());
auto in = EigenVector<T>::Flatten(in_array[i]);
auto result = EigenVector<T>::Flatten(out_array[i]);
result.device(*ctx.template device_context<MKLDNNDeviceContext>()
.eigen_device()) = result + in;
}
}
}
}
} else {
PADDLE_THROW("Unexpected branch, output variable type is %s",
out_var->Type().name());
}
}
};
} // namespace operators
} // namespace paddle
REGISTER_OP_KERNEL(sum, MKLDNN, ::paddle::platform::CPUPlace,
paddle::operators::SumMKLDNNOpKernel<float>);
paddle/fluid/operators/sum_op.cc
浏览文件 @ 9942a304
......@@ -18,6 +18,10 @@ limitations under the License. */
#include "paddle/fluid/framework/var_type_inference.h"
#include "paddle/fluid/operators/detail/safe_ref.h"
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
namespace paddle {
namespace operators {
using framework::Tensor;
......@@ -63,6 +67,18 @@ class SumOp : public framework::OperatorWithKernel {
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
auto x_vars = ctx.MultiInputVar("X");
framework::LibraryType library{framework::LibraryType::kPlain};
framework::DataLayout layout{framework::DataLayout::kAnyLayout};
#ifdef PADDLE_WITH_MKLDNN
if (library == framework::LibraryType::kPlain &&
platform::CanMKLDNNBeUsed(ctx)) {
library = framework::LibraryType::kMKLDNN;
layout = framework::DataLayout::kMKLDNN;
}
#endif
if (x_vars[0]->IsType<framework::LoDTensor>()) {
int dtype = -1;
for (auto& x_var : x_vars) {
......@@ -80,26 +96,27 @@ class SumOp : public framework::OperatorWithKernel {
"Sum operator should have at least one tensor");
return framework::OpKernelType(
static_cast<framework::proto::VarType::Type>(dtype),
ctx.device_context());
static_cast<framework::proto::VarType::Type>(dtype), ctx.GetPlace(),
layout, library);
} else if (x_vars[0]->IsType<framework::SelectedRows>()) {
for (auto& var : x_vars) {
auto& value = var->Get<framework::SelectedRows>().value();
if (value.IsInitialized()) {
return framework::OpKernelType(framework::ToDataType(value.type()),
ctx.device_context());
ctx.device_context(), layout, library);
}
}
// if input sparse vars are not initialized, use an default kernel type.
return framework::OpKernelType(framework::proto::VarType::FP32,
ctx.device_context());
ctx.device_context(), layout, library);
} else if (x_vars[0]->IsType<framework::LoDTensorArray>()) {
for (auto& x_var : x_vars) {
auto& array = x_var->Get<framework::LoDTensorArray>();
for (auto& each : array) {
if (each.numel() != 0) {
return framework::OpKernelType(framework::ToDataType(each.type()),
ctx.device_context());
ctx.device_context(), layout,
library);
}
}
}
......@@ -116,6 +133,9 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker {
AddInput("X", "(vector<Tensor>) The input tensors of sum operator.")
.AsDuplicable();
AddOutput("Out", "(Tensor) The output tensor of sum operator.").Reuse("X");
AddAttr<bool>("use_mkldnn",
"(bool, default false) Only used in mkldnn kernel")
.SetDefault(false);
AddComment(R"DOC(
Sum operator.
......@@ -132,7 +152,6 @@ class SumOpVarTypeInference : public framework::VarTypeInference {
framework::BlockDesc* block) const override {
auto& inputs = op_desc.Input("X");
auto var_type = framework::proto::VarType::SELECTED_ROWS;
for (auto& name : op_desc.Input("X")) {
VLOG(10) << name << " "
<< block->FindRecursiveOrCreateVar(name).GetType();
......@@ -206,6 +225,7 @@ namespace ops = paddle::operators;
REGISTER_OPERATOR(sum, ops::SumOp, ops::SumOpMaker, ops::SumGradMaker,
ops::SumOpVarTypeInference);
REGISTER_OP_CPU_KERNEL(
sum, ops::SumKernel<paddle::platform::CPUDeviceContext, float>,
ops::SumKernel<paddle::platform::CPUDeviceContext, double>,
......
paddle/fluid/operators/tensorrt_engine_op_test.cc
浏览文件 @ 9942a304
......@@ -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/while_op.cc
浏览文件 @ 9942a304
......@@ -203,11 +203,11 @@ class WhileGradOp : public framework::OperatorBase {
->set_lod(inside_tensor.lod());
}
}
auto new_inside_name = cur_scope.Rename(inside_grad_name);
auto sum_op = framework::OpRegistry::CreateOp(
"sum", {{"X", {pg_names[param_id], new_inside_name}}},
{{"Out", {pg_names[param_id]}}}, framework::AttributeMap{});
{{"Out", {pg_names[param_id]}}},
framework::AttributeMap{{"use_mkldnn", {false}}});
sum_op->Run(cur_scope, dev_place);
cur_scope.Rename(new_inside_name, inside_grad_name);
}
......
paddle/fluid/platform/cpu_info.cc
浏览文件 @ 9942a304
......@@ -28,9 +28,15 @@ DEFINE_double(fraction_of_cpu_memory_to_use, 1,
"Default use 100% of CPU memory for PaddlePaddle,"
"reserve the rest for page tables, etc");
DEFINE_uint64(
initial_cpu_memory_in_mb, 500,
"Default initial 500MB of CPU memory for PaddlePaddle, in MD unit.");
DEFINE_uint64(initial_cpu_memory_in_mb,
#ifdef PADDLE_WITH_MKLDNN
/* Aligned with mozga-intel, MKLDNN need at least 5000 MB
* to obtain the best performance*/
5000,
#else
500,
#endif
"Initial CPU memory for PaddlePaddle, in MD unit.");
DEFINE_double(
fraction_of_cuda_pinned_memory_to_use, 0.5,
......@@ -59,10 +65,7 @@ inline size_t CpuTotalPhysicalMemory() {
size_t CpuMaxAllocSize() {
// For distributed systems, it requires configuring and limiting
// the fraction of memory to use.
return std::min(
static_cast<size_t>(FLAGS_fraction_of_cpu_memory_to_use *
CpuTotalPhysicalMemory()),
static_cast<size_t>(FLAGS_initial_cpu_memory_in_mb * 1 << 20));
return FLAGS_fraction_of_cpu_memory_to_use * CpuTotalPhysicalMemory();
}
size_t CpuMinChunkSize() {
......@@ -71,8 +74,11 @@ size_t CpuMinChunkSize() {
}
size_t CpuMaxChunkSize() {
// Allow to allocate the maximum chunk size is roughly 3% of CPU memory.
return CpuMaxAllocSize() / 32;
// Allow to allocate the maximum chunk size is roughly 3% of CPU memory,
// or the initial_cpu_memory_in_mb.
return std::min(
static_cast<size_t>(CpuMaxAllocSize() / 32),
static_cast<size_t>(FLAGS_initial_cpu_memory_in_mb * 1 << 20));
}
size_t CUDAPinnedMaxAllocSize() {
......
paddle/fluid/platform/mkldnn_helper.h
浏览文件 @ 9942a304
......@@ -99,5 +99,11 @@ inline mkldnn::memory::format GetMKLDNNFormat(const mkldnn::memory memory) {
memory.get_primitive_desc().desc().data.format);
}
inline mkldnn::memory::format GetMKLDNNFormat(
const mkldnn::sum::primitive_desc& memory) {
return static_cast<mkldnn::memory::format>(
memory.dst_primitive_desc().desc().data.format);
}
} // namespace platform
} // namespace paddle
paddle/testing/paddle_gtest_main.cc
浏览文件 @ 9942a304
......@@ -30,8 +30,9 @@ int main(int argc, char** argv) {
new_argv.push_back(
strdup("--tryfromenv=fraction_of_gpu_memory_to_use,use_pinned_memory"));
#else
new_argv.push_back(strdup("--tryfromenv=use_pinned_memory,use_mkldnn"));
new_argv.push_back(strdup("--undefok=use_mkldnn"));
new_argv.push_back(strdup(
"--tryfromenv=use_pinned_memory,use_mkldnn,initial_cpu_memory_in_mb"));
new_argv.push_back(strdup("--undefok=use_mkldnn,initial_cpu_memory_in_mb"));
#endif
int new_argc = static_cast<int>(new_argv.size());
char** new_argv_address = new_argv.data();
......
python/paddle/fluid/__init__.py
浏览文件 @ 9942a304
......@@ -118,7 +118,7 @@ def __bootstrap__():
read_env_flags = [
'use_pinned_memory', 'check_nan_inf', 'benchmark', 'warpctc_dir',
'eager_delete_scope', 'use_mkldnn'
'eager_delete_scope', 'use_mkldnn', 'initial_cpu_memory_in_mb'
]
if core.is_compiled_with_cuda():
read_env_flags += [
......
python/paddle/fluid/average.py
浏览文件 @ 9942a304
......@@ -36,6 +36,25 @@ def _is_number_or_matrix_(var):
class WeightedAverage(object):
"""
Calculate weighted average.
The average calculating is accomplished via Python totally.
They do not change Paddle's Program, nor do anything to
modify NN model's configuration. They are completely
wrappers of Python functions.
Examples:
.. code-block:: python
avg = fluid.average.WeightedAverage()
avg.add(value=2.0, weight=1)
avg.add(value=4.0, weight=2)
avg.eval()
# The result is 3.333333333.
# For (2.0 * 1 + 4.0 * 2) / (1 + 2) = 3.333333333
"""
def __init__(self):
warnings.warn(
"The %s is deprecated, please use fluid.metrics.Accuracy instead." %
......
python/paddle/fluid/backward.py
浏览文件 @ 9942a304
......@@ -132,9 +132,9 @@ def _addup_repetitive_outputs_(op_descs):
for idx, op_desc in enumerate(op_descs):
for var_name in op_desc.input_arg_names():
if len(renamed_vars[var_name]) > 1:
pending_sum_ops.append(
(_create_op_desc_("sum", {"X": renamed_vars[var_name]},
{"Out": [var_name]}, {}), idx))
pending_sum_ops.append((_create_op_desc_(
"sum", {"X": renamed_vars[var_name]}, {"Out": [var_name]},
{"use_mkldnn": False}), idx))
renamed_vars[var_name] = [var_name]
for var_name in op_desc.output_arg_names():
if var_name == core.empty_var_name(
......@@ -161,8 +161,9 @@ def _addup_repetitive_outputs_(op_descs):
renamed_vars[var_name].append(new_name)
for var_name, inputs in renamed_vars.iteritems():
if len(inputs) > 1:
pending_sum_ops.append((_create_op_desc_(
"sum", {"X": inputs}, {"Out": [var_name]}, {}), len(op_descs)))
pending_sum_ops.append(
(_create_op_desc_("sum", {"X": inputs}, {"Out": [var_name]},
{"use_mkldnn": False}), len(op_descs)))
# sum_op descs are sorted according to their insert position
for p in reversed(pending_sum_ops):
op_descs.insert(p[1], p[0])
......@@ -434,18 +435,65 @@ def _get_stop_gradients_(program):
def append_backward(loss, parameter_list=None, no_grad_set=None,
callbacks=None):
"""
Append backward part to main_program
Append backward part to main_program.
Args:
loss(Variable): The variable generated by cost function.
parameter_list(list[string]): Parameters that need to be updated by
optimizer. If None, it means all parameters need to be updated.
no_grad_set(set): Variables that have no gradients in Block 0.
All variables with `step_gradient=True` from all blocks will be
automatically added.
A complete neural network training is made up of forward and backward
propagation. However, when we configure a network, we only need to
specify its forwrd part. The backward part is generated automatically
according to the forward part by this function.
Return:
(list[(Variable,Variable)]): list of (parameter, gradient) pair.
In most cases, users do not need to invoke this function manually. It
will be automatically invoked by the optimizer's `minimize` function.
Args:
loss(Variable): The loss variable of the network.
parameter_list(list[string]|None): Names of parameters that need
to be updated by optimizers.
If it is None, all parameters
will be updated.
Default: None
no_grad_set(set|None): Variables in the Block 0 whose gradients
should be ignored. All variables with
`step_gradient=True` from all blocks will
be automatically added into this set.
Default: None
callbacks(list[callable object]|None): The callbacks are used for
doing some custom jobs during
backward part building. All
callable objects in it will
be invoked once each time a
new gradient operator is added
into the program. The callable
object must has two input
parameters: 'block' and 'context'.
The 'block' is the block which
the new gradient operator will
be added to. The 'context' is a
map, whose keys are gradient
variable names and values are
corresponding original variables.
In addition to this, the 'context'
has another special key-value pair:
the key is string '__current_op_desc__'
and the value is the op_desc of the
gradient operator who has just
triggered the callable object.
Returns:
list[(Variable,Variable)]: Pairs of parameter and its
corresponding gradients. The key is the parameter and the
value is gradient variable.
Raises:
AssertionError: If `loss` is not an instance of Variable.
Examples:
.. code-block:: python
# network configuration code
# ...
avg_loss = fluid.layers.mean(loss)
param_grad_list = fluid.backward.append_backward(loss=avg_loss)
"""
assert isinstance(loss, framework.Variable)
......
python/paddle/fluid/clip.py
浏览文件 @ 9942a304
......@@ -24,8 +24,6 @@ __all__ = [
'GradientClipByValue',
'GradientClipByNorm',
'GradientClipByGlobalNorm',
'append_gradient_clip_ops',
'error_clip_callback',
]
......@@ -38,6 +36,25 @@ class BaseErrorClipAttr(object):
class ErrorClipByValue(BaseErrorClipAttr):
"""
Clips tensor values to the range [min, max].
Given a tensor t, this operation clips its value to min and max inplace.
- Any values less than min are set to min.
- Any values greater than max are set to max.
Args:
max (float): The maximum value to clip by.
min (float, optional): The minimum value to clip by. if not set by user, \
will be set to -max by framework.
Examples:
.. code-block:: python
var = fluid.framework.Variable(..., error_clip=ErrorClipByValue(max=5.0), ...)
"""
def __init__(self, max, min=None):
max = float(max)
if min is None:
......@@ -99,6 +116,31 @@ class NullGradientClipAttr(BaseGradientClipAttr):
class GradientClipByValue(BaseGradientClipAttr):
"""
Clips gradient values to the range [min, max].
Given a tensor t, this operation clips its value to min and max inplace.
- Any values less than min are set to min.
- Any values greater than max are set to max.
Args:
max (float): The maximum value to clip by.
min (float, optional): The minimum value to clip by. if not set by user, \
will be set to -max by framework.
Examples:
.. code-block:: python
w_param_attrs = ParamAttr(name=None,
initializer=UniformInitializer(low=-1.0, high=1.0, seed=0),
learning_rate=1.0,
regularizer=L1Decay(1.0),
trainable=True,
clip=GradientClipByValue(-1.0, 1.0))
y_predict = fluid.layers.fc(input=x, size=1, param_attr=w_param_attrs)
"""
def __init__(self, max, min=None):
max = float(max)
if min is None:
......@@ -120,6 +162,37 @@ class GradientClipByValue(BaseGradientClipAttr):
class GradientClipByNorm(BaseGradientClipAttr):
"""
Clips tensor values to a maximum L2-norm.
This operator limits the L2 norm of the input :math:`X` within :math:`max\_norm`.
If the L2 norm of :math:`X` is less than or equal to :math:`max\_norm`, :math:`Out`
will be the same as :math:`X`. If the L2 norm of :math:`X` is greater than
:math:`max\_norm`, :math:`X` will be linearly scaled to make the L2 norm of
:math:`Out` equal to :math:`max\_norm`, as shown in the following formula:
.. math::
Out = \\frac{max\_norm * X}{norm(X)},
where :math:`norm(X)` represents the L2 norm of :math:`X`.
Args:
clip_norm (float): The maximum norm value
Examples:
.. code-block:: python
w_param_attrs = ParamAttr(name=None,
initializer=UniformInitializer(low=-1.0, high=1.0, seed=0),
learning_rate=1.0,
regularizer=L1Decay(1.0),
trainable=True,
clip=GradientClipByNorm(clip_norm=2.0))
y_predict = fluid.layers.fc(input=x, size=1, param_attr=w_param_attrs)
"""
def __init__(self, clip_norm):
self.clip_norm = clip_norm
......@@ -135,6 +208,44 @@ class GradientClipByNorm(BaseGradientClipAttr):
class GradientClipByGlobalNorm(BaseGradientClipAttr):
"""
Clips values of multiple tensors by the ratio of the sum of their norms.
Given a list of tensors t_list, and a clipping ratio clip_norm, this
operation returns a list of clipped tensors list_clipped and the global
norm (global_norm) of all tensors in t_list.
To perform the clipping, the values :math:`t\_list[i]` are set to:
.. math::
t\_list[i] = t\_list[i] * \\frac{clip\_norm}{\max(global\_norm, clip\_norm)}
where:
.. math::
global\_norm = \sqrt{\sum_{i=0}^{N-1}(l2norm(t\_list[i]))^2}
If :math:`clip\_norm > global\_norm` then the entries in t_list remain as they are,
otherwise they're all shrunk by the global ratio.
Args:
clip_norm (float): The maximum norm value
group_name (str, optional): The group name for this clip.
Examples:
.. code-block:: python
p_g_clip = fluid.backward.append_backward(loss=avg_cost_clip)
with fluid.program_guard(main_program=prog_clip):
fluid.clip.set_gradient_clip(
fluid.clip.GradientClipByGlobalNorm(clip_norm=2.0))
p_g_clip = fluid.clip.append_gradient_clip_ops(p_g_clip)
"""
def __init__(self, clip_norm, group_name="default_group"):
if not isinstance(group_name, basestring):
raise TypeError("'group_name' must be a basestring.")
......@@ -184,13 +295,14 @@ class GradientClipByGlobalNorm(BaseGradientClipAttr):
def set_gradient_clip(clip, param_list=None, program=None):
"""
To specify parameters that require gradient clip.
Args:
clip(BaseGradientClipAttr): An instance of some derived class of BaseGradientClipAttr,
which describes the type and detailed attributes of required gradient clip.
param_list(list, None by default): Parameters that require gradient clip.
param_list(list(Variable)): Parameters that require gradient clip.
It can be a list of parameter or a list of parameter's name.
When it's None, all parameters in the program will be included.
program(Program, None by default): The program where parameters are.
program(Program): The program where parameters are.
Will be the default main program when assigned with None.
"""
if not isinstance(clip, BaseGradientClipAttr):
......
python/paddle/fluid/data_feeder.py
浏览文件 @ 9942a304
......@@ -29,6 +29,13 @@ class DataToLoDTensorConverter(object):
self.place = place
self.lod_level = lod_level
self.shape = shape
negtive_count = 0
for s in self.shape:
if s < 0:
negtive_count += 1
if negtive_count > 1:
self.shape = None
break
if dtype == core.VarDesc.VarType.FP32:
self.dtype = 'float32'
elif dtype == core.VarDesc.VarType.INT64:
......@@ -61,7 +68,9 @@ class DataToLoDTensorConverter(object):
self._feed_impl_(each_data, lod[1:], lod_level - 1)
def done(self):
arr = numpy.array(self.data, dtype=self.dtype).reshape(self.shape)
arr = numpy.array(self.data, dtype=self.dtype)
if self.shape:
arr = arr.reshape(self.shape)
t = core.LoDTensor()
t.set(arr, self.place)
if self.lod_level > 0:
......
python/paddle/fluid/framework.py
浏览文件 @ 9942a304
......@@ -642,6 +642,12 @@ class Operator(object):
def set_attr(self, name, val):
self.attrs[name] = val
if isinstance(val, Block):
self.desc.set_block_attr(name, val.desc)
elif isinstance(val, core.BlockDesc) or \
isinstance(val, core.ProgramDesc):
self.desc.set_serialized_attr(name, val.serialize_to_string())
else:
self.desc.set_attr(name, val)
@property
......
python/paddle/fluid/inferencer.py
浏览文件 @ 9942a304
......@@ -27,13 +27,30 @@ __all__ = ['Inferencer', ]
class Inferencer(object):
def __init__(self, infer_func, param_path, place=None, parallel=False):
"""
:param infer_func: a function that will return predict Variable
:param param_path: the path where the inference model is saved by fluid.io.save_params
:param place: place to do the inference
:param parallel: use parallel_executor to run the inference, it will use multi CPU/GPU.
Inferencer High Level API.
Args:
infer_func (Python func): Infer function that will return predict Variable
param_path (str): The path where the inference model is saved by fluid.io.save_params
place (Place): place to do the inference
parallel (bool): use parallel_executor to run the inference, it will use multi CPU/GPU.
Examples:
.. code-block:: python
def inference_program():
x = fluid.layers.data(name='x', shape=[13], dtype='float32')
y_predict = fluid.layers.fc(input=x, size=1, act=None)
return y_predict
place = fluid.CPUPlace()
inferencer = fluid.Inferencer(
infer_func=inference_program, param_path="/tmp/model", place=place)
"""
def __init__(self, infer_func, param_path, place=None, parallel=False):
self.param_path = param_path
self.scope = core.Scope()
self.parallel = parallel
......@@ -60,9 +77,20 @@ class Inferencer(object):
def infer(self, inputs, return_numpy=True):
"""
:param inputs: a map of {"input_name": input_var} that will be feed into the inference program
to get the predict value
:return: the predict value of the inference model
Do Inference for Inputs
Args:
inputs (map): a map of {"input_name": input_var} that will be feed into the inference program
return_numpy (bool): transform return value into numpy or not
Returns:
Tensor or Numpy: the predict value of the inference model for the inputs
Examples:
.. code-block:: python
tensor_x = numpy.random.uniform(0, 10, [batch_size, 13]).astype("float32")
results = inferencer.infer({'x': tensor_x})
"""
if not isinstance(inputs, dict):
raise ValueError(
......
python/paddle/fluid/initializer.py
浏览文件 @ 9942a304
......@@ -19,26 +19,39 @@ from framework import convert_np_dtype_to_dtype_
from core import VarDesc
__all__ = [
'Constant', 'Uniform', 'Normal', 'Xavier', 'Bilinear', 'force_init_on_cpu',
'init_on_cpu', 'ConstantInitializer', 'UniformInitializer',
'NormalInitializer', 'XavierInitializer', 'BilinearInitializer'
'Constant', 'Uniform', 'Normal', 'Xavier', 'Bilinear', 'MSRA',
'force_init_on_cpu', 'init_on_cpu', 'ConstantInitializer',
'UniformInitializer', 'NormalInitializer', 'XavierInitializer',
'BilinearInitializer', 'MSRAInitializer'
]
_force_init_on_cpu_ = False
def force_init_on_cpu():
"""
The flag of whether force to init variables on CPU.
Examples:
.. code-block:: python
if force_init_on_cpu():
pass
"""
return _force_init_on_cpu_
@contextlib.contextmanager
def init_on_cpu():
"""
Switch program with `with` statement
Force the variable to be inited on CPU.
Examples:
>>> with init_on_cpu():
>>> step = layers.create_global_var()
.. code-block:: python
with init_on_cpu():
step = layers.create_global_var()
"""
global _force_init_on_cpu_
......@@ -104,14 +117,18 @@ class Initializer(object):
class ConstantInitializer(Initializer):
"""Implements the constant initializer
"""
def __init__(self, value=0.0, force_cpu=False):
"""Constructor for ConstantInitializer
Args:
value: constant value to initialize the variable
value (float): constant value to initialize the variable
Examples:
.. code-block:: python
fc = fluid.layers.fc(input=x, size=10,
param_attr=fluid.initializer.Constant(value=2.0))
"""
def __init__(self, value=0.0, force_cpu=False):
assert value is not None
super(ConstantInitializer, self).__init__()
self._value = value
......@@ -146,16 +163,20 @@ class ConstantInitializer(Initializer):
class UniformInitializer(Initializer):
"""Implements the random uniform distribution initializer
"""
def __init__(self, low=-1.0, high=1.0, seed=0):
"""Constructor for UniformInitializer
Args:
low: lower boundary of the uniform distribution
high: upper boundary of the uniform distribution
seed: random seed
low (float): lower boundary of the uniform distribution
high (float): upper boundary of the uniform distribution
seed (int): random seed
Examples:
.. code-block:: python
fc = fluid.layers.fc(input=x, size=10,
param_attr=fluid.initializer.Uniform(low=-0.5, high=0.5))
"""
def __init__(self, low=-1.0, high=1.0, seed=0):
assert low is not None
assert high is not None
assert high >= low
......@@ -196,17 +217,21 @@ class UniformInitializer(Initializer):
class NormalInitializer(Initializer):
"""Implements the random Normal(Gaussian) distribution initializer
"""
def __init__(self, loc=0.0, scale=1.0, seed=0):
"""Constructor for NormalInitializer
"""Implements the Random Normal(Gaussian) distribution initializer
Args:
loc: mean of the normal distribution
scale: standard deviation of the normal distribution
seed: random seed
loc (float): mean of the normal distribution
scale (float): standard deviation of the normal distribution
seed (int): random seed
Examples:
.. code-block:: python
fc = fluid.layers.fc(input=x, size=10,
param_attr=fluid.initializer.Normal(loc=0.0, scale=2.0))
"""
def __init__(self, loc=0.0, scale=1.0, seed=0):
assert loc is not None
assert scale is not None
assert seed is not None
......@@ -246,39 +271,49 @@ class NormalInitializer(Initializer):
class XavierInitializer(Initializer):
"""Implements the Xavier initializer
"""
This class implements the Xavier weight initializer from the paper
Understanding the difficulty of training deep feedforward neural
networks[1] by Xavier Glorot and Yoshua Bengio.
`Understanding the difficulty of training deep feedforward neural
networks <http://proceedings.mlr.press/v9/glorot10a/glorot10a.pdf>`_
by Xavier Glorot and Yoshua Bengio.
This initializer is designed to keep the scale of the gradients
approximately same in all the layers. In case of Uniform distribution,
the range is [-x, x], where x = sqrt(6 / (fan_in + fan_out)).
the range is [-x, x], where
.. math::
x = \sqrt{\\frac{6.0}{fan\_in + fan\_out}}
In case of Normal distribution, the mean is 0 and the standard deviation
is sqrt(2/ (fan_in + fan_out)).
is
References:
[1] Understanding the difficulty of training deep feedforward neural
networks. International conference on artificial intelligence and
statistics.
(http://proceedings.mlr.press/v9/glorot10a.html)
"""
.. math::
\sqrt{\\frac{2.0}{fan\_in + fan\_out}}
def __init__(self, uniform=True, fan_in=None, fan_out=None, seed=0):
"""Constructor for XavierInitializer
Args:
uniform: whether to use uniform or normal distribution
fan_in: fan_in for Xavier initialization. If None, it is
uniform (bool): whether to use uniform or normal distribution
fan_in (float): fan_in for Xavier initialization. If None, it is
inferred from the variable.
fan_out: fan_out for Xavier initialization. If None, it is
fan_out (float): fan_out for Xavier initialization. If None, it is
inferred from the variable.
seed: random seed
seed (int): random seed
Note:
It is recommended to set fan_in and fan_out to None for most cases.
Examples:
.. code-block:: python
fc = fluid.layers.fc(
input=queries, size=10,
param_attr=fluid.initializer.Xavier(uniform=False))
Note: It is recommended to set fan_in and fan_out to None for
most cases.
"""
def __init__(self, uniform=True, fan_in=None, fan_out=None, seed=0):
assert uniform is not None
assert seed is not None
super(XavierInitializer, self).__init__()
......@@ -342,30 +377,42 @@ class MSRAInitializer(Initializer):
"""Implements the MSRA initializer a.k.a. Kaiming Initializer
This class implements the weight initialization from the paper
Delving Deep into Rectifiers: Surpassing Human-Level Performance on
ImageNet Classification[1] by Kaiming He, Xiangyu Zhang, Shaoqing Ren
and Jian Sun. This is a robust initialization method that particularly
considers the rectifier nonlinearities. In case of Uniform distribution,
the range is [-x, x], where x = sqrt(6 / fan_in). In case of Normal
distribution, the mean is 0 and the standard deviation
is sqrt(2/ fan_in).
References:
[1] Delving Deep into Rectifiers: Surpassing Human-Level Performance
on ImageNet Classification
(https://arxiv.org/abs/1502.01852)
"""
`Delving Deep into Rectifiers: Surpassing Human-Level Performance on
ImageNet Classification <https://arxiv.org/abs/1502.01852>`_
by Kaiming He, Xiangyu Zhang, Shaoqing Ren and Jian Sun. This is a
robust initialization method that particularly considers the rectifier
nonlinearities. In case of Uniform distribution, the range is [-x, x], where
def __init__(self, uniform=True, fan_in=None, seed=0):
"""Constructor for MSRAInitializer
.. math::
x = \sqrt{\\frac{6.0}{fan\_in}}
In case of Normal distribution, the mean is 0 and the standard deviation
is
.. math::
\sqrt{\\frac{2.0}{fan\_in}}
Args:
uniform: whether to use uniform or normal distribution
fan_in: fan_in for MSRAInitializer. If None, it is
uniform (bool): whether to use uniform or normal distribution
fan_in (float): fan_in for MSRAInitializer. If None, it is\
inferred from the variable.
seed: random seed
seed (int): random seed
Note:
It is recommended to set fan_in to None for most cases.
Examples:
.. code-block:: python
fc = fluid.layers.fc(
input=queries, size=10,
param_attr=fluid.initializer.MSRA(uniform=False))
"""
Note: It is recommended to set fan_in to None for most cases.
def __init__(self, uniform=True, fan_in=None, seed=0):
"""Constructor for MSRAInitializer
"""
assert uniform is not None
assert seed is not None
......@@ -425,34 +472,37 @@ class MSRAInitializer(Initializer):
class BilinearInitializer(Initializer):
"""Implements the bilinear initializer.
"""
This initializer can be used in transposed convolution operator to
act as upsampling. Users can upsample a feature map with shape of
(B, C, H, W) by any integer factor. The usage is:
>>> factor = 2
>>> w_attr = ParamAttr(learning_rate=0., regularizer=L2Decay(0.),
>>> initializer=Bilinear())
>>> conv_up = fluid.layers.conv2d_transpose(
>>> input,
>>> num_filters=C,
>>> output_size=None,
>>> filter_size=2 * factor - factor % 2,
>>> padding=ceil((factor - 1) / 2.),
>>> stride=factor,
>>> groups=C,
>>> param_attr=w_attr,
>>> bias_attr=False)
Where, `num_filters=C` and `groups=C` means this is channel-wise tranposed
Examples:
.. code-block:: python
factor = 2
w_attr = ParamAttr(learning_rate=0., regularizer=L2Decay(0.),
initializer=Bilinear())
conv_up = fluid.layers.conv2d_transpose(
input,
num_filters=C,
output_size=None,
filter_size=2 * factor - factor % 2,
padding=ceil((factor - 1) / 2.),
stride=factor,
groups=C,
param_attr=w_attr,
bias_attr=False)
Where, `num_filters=C` and `groups=C` means this is channel-wise transposed
convolution. The filter shape will be (C, 1, K, K) where K is `filer_size`,
This initializer will set a (K, K) interpolation kernel for every channel
of the filter identically. The resulting shape of the output feature map
will be (B, C, factor * H, factor * W). Note that the learning rate and the
weight decay are set to 0 in order to keep coefficient values of bilinear
interpolation unchanged during training.
"""
def __init__(self):
......@@ -469,7 +519,7 @@ class BilinearInitializer(Initializer):
be added.
Returns:
the initialization op
Operator: the initialization op
Raises:
ValueError: If type of `var` and `block` is not right.
......
python/paddle/fluid/io.py
浏览文件 @ 9942a304
......@@ -30,20 +30,42 @@ __all__ = [
def is_parameter(var):
"""Check whether the variable is a Parameter.
This function checks whether the input variable is a Parameter.
"""
Check whether the given variable is an instance of Parameter.
Args:
var : The input variable.
var(Variable): The variable to be checked.
Returns:
boolean result whether the variable is a Parameter.
bool: True if the given `var` is an instance of Parameter,
False if not.
Examples:
.. code-block:: python
param = fluid.default_main_program().global_block().var('fc.w')
res = fluid.io.is_parameter(param)
"""
return isinstance(var, Parameter)
def is_persistable(var):
"""
Check whether the given variable is persistable.
Args:
var(Variable): The variable to be checked.
Returns:
bool: True if the given `var` is persistable
False if not.
Examples:
.. code-block:: python
param = fluid.default_main_program().global_block().var('fc.w')
res = fluid.io.is_persistable(param)
"""
if var.desc.type() == core.VarDesc.VarType.FEED_MINIBATCH or \
var.desc.type() == core.VarDesc.VarType.FETCH_LIST:
return False
......@@ -68,20 +90,69 @@ def save_vars(executor,
predicate=None,
filename=None):
"""
Save variables to directory by executor.
Save variables to the given directory by executor.
There are two ways to specify variables to be saved: The first way, list
variables in a list and assign it to the `vars`. The second way, assign the
`main_program` with an existing program, then all variables in the program
will be saved. The first way has a higher priority. In other words, if `vars`
are assigned, the `main_program` and the `predicate` will be ignored.
:param executor: executor that save variable
:param dirname: directory path
:param main_program: program. If vars is None, then filter all variables in this
program which fit `predicate`. Default default_main_program.
:param predicate: The Predicate describes a callable that returns a variable
as a bool. If it returns true, the corresponding input variable will be saved.
:param vars: variables need to be saved. If vars is specified, program & predicate
will be ignored
:param filename: The name of a single file that all vars are saved to.
If it is None, save variables to separate files.
The `dirname` are used to specify the folder where to save variables.
If you prefer to save variables in separate files in the folder `dirname`,
set `filename` None; if you prefer to save all variables in a single file,
use `filename` to specify it.
:return: None
Args:
executor(Executor): The executor to run for saving variables.
dirname(str): The directory path.
main_program(Program|None): The program whose variables will be saved.
If it is None, the default main program will
be used automatically.
Default: None
vars(list[Variable]|None): The list that contains all variables to save.
It has a higher priority than the `main_program`.
Default: None
predicate(function|None): If it is not None, only variables in the
`main_program` that makes predicate(variable)==True
will be saved. It only works when we are using the
`main_program` to specify variables (In other words
`vars` is None).
Default: None
filename(str|None): The file which to save all variables. If you prefer to save
variables separately, set it to None.
Default: None
Returns:
None
Raises:
TypeError: If `main_program` is not an instance of Program nor None.
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
param_path = "./my_paddle_model"
# The first usage: using `main_program` to specify variables
def name_has_fc(var):
res = "fc" in var.name
return res
prog = fluid.default_main_program()
fluid.io.save_vars(executor=exe, dirname=path, main_program=prog,
vars=None)
# All variables in `main_program` whose name includes "fc" will be saved.
# And variables are going to be saved separately.
# The second usage: using `vars` to specify variables
var_list = [var_a, var_b, var_c]
fluid.io.save_vars(executor=exe, dirname=path, vars=var_list,
filename="vars_file")
# var_a, var_b and var_c will be saved. And they are going to be
# saved in the same file named 'var_file' in the path "./my_paddle_model".
"""
if vars is None:
if main_program is None:
......@@ -129,7 +200,42 @@ def save_vars(executor,
def save_params(executor, dirname, main_program=None, filename=None):
"""
Save all parameters to directory with executor.
This function filters out all parameters from the give `main_program`
and then save them to the folder `dirname` or the file `filename`.
Use the `dirname` to specify the saving folder. If you would like to
save parameters in separate files, set `filename` None; if you would
like to save all parameters in a single file, use `filename` to specify
the file name.
NOTICE: Some variables are not Parameter while they are necessary for
training. So you can NOT save and continue your training just by
`save_params()` and `load_params()`. Please use `save_persistables()`
and `load_persistables()` instead.
Args:
executor(Executor): The executor to run for saving parameters.
dirname(str): The saving directory path.
main_program(Program|None): The program whose parameters will be
saved. If it is None, the default
main program will be used automatically.
Default: None
filename(str|None): The file to save all parameters. If you prefer
to save parameters in differnet files, set it
to None.
Default: None
Returns:
None
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
param_path = "./my_paddle_model"
prog = fluid.default_main_program()
fluid.io.save_params(executor=exe, dirname=param_path,
main_program=None)
"""
save_vars(
executor,
......@@ -142,7 +248,37 @@ def save_params(executor, dirname, main_program=None, filename=None):
def save_persistables(executor, dirname, main_program=None, filename=None):
"""
Save all persistables to directory with executor.
This function filters out all variables with `persistable==True` from the
give `main_program` and then saves these variables to the folder `dirname`
or file `filename`.
The `dirname` is used to specify the folder where persistable variables
are going to be saved. If you would like to save variables in separate
files, set `filename` None; if you would like to save all variables in a
single file, use `filename` to specify the file name.
Args:
executor(Executor): The executor to run for saving persistable variables.
dirname(str): The directory path.
main_program(Program|None): The program whose persistbale variables will
be saved. If it is None, the default main
program will be used automatically.
Default: None
filename(str|None): The file to saved all variables. If you prefer to
save variables in differnet files, set it to None.
Default: None
Returns:
None
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
param_path = "./my_paddle_model"
prog = fluid.default_main_program()
fluid.io.save_persistables(executor=exe, dirname=param_path,
main_program=None)
"""
save_vars(
executor,
......@@ -160,20 +296,69 @@ def load_vars(executor,
predicate=None,
filename=None):
"""
Load variables from directory by executor.
Load variables from the given directory by executor.
There are two ways to specify variables to be loaded: The first way, list
variables in a list and assign it to the `vars`. The second way, assign the
`main_program` with an existing program, then all variables in the program
will be loaded. The first way has a higher priority. In other words if `vars`
are assigned, the `main_program` and the `predicate` will be ignored.
The `dirname` are used to specify the folder where to load variables.
If variables were saved in separate files in the folder `dirname`,
set `filename` None; if all variables were saved in a single file,
use `filename` to specify it.
:param executor: executor that load variable
:param dirname: directory path
:param main_program: program. If vars is None, then filter all variables in this
program which fit `predicate`. Default default_main_program().
:param predicate: The Predicate describes a callable that returns a variable
as a bool. If it returns true, the corresponding input variable will be loaded.
:param vars: variables need to be loaded. If vars is specified, program &
predicate will be ignored
:param filename: The name of the single file that all vars are loaded from.
If it is None, load variables from separate files.
Args:
executor(Executor): The executor to run for loading variables.
dirname(str): The directory path.
main_program(Program|None): The program whose variables will be loaded.
If it is None, the default main program will
be used automatically.
Default: None
vars(list[Variable]|None): The list that contains all variables to load.
It has a higher priority than the `main_program`.
Default: None
predicate(function|None): If it is not None, only variables in the
`main_program` that makes predicate(variable)==True
will be loaded. It only works when we are using the
`main_program` to specify variables (In other words
`vars` is None).
Default: None
filename(str|None): The file which saved all required variables. If variables
were saved in differnet files, set it to None.
Default: None
Returns:
None
Raises:
TypeError: If `main_program` is not an instance of Program nor None.
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
param_path = "./my_paddle_model"
# The first usage: using `main_program` to specify variables
def name_has_fc(var):
res = "fc" in var.name
return res
:return: None
prog = fluid.default_main_program()
fluid.io.load_vars(executor=exe, dirname=path, main_program=prog,
vars=None)
# All variables in `main_program` whose name includes "fc" will be loaded.
# And all the variables are supposed to have been saved in differnet files.
# The second usage: using `vars` to specify variables
var_list = [var_a, var_b, var_c]
fluid.io.load_vars(executor=exe, dirname=path, vars=var_list,
filename="vars_file")
# var_a, var_b and var_c will be loaded. And they are supposed to haven
# been saved in the same file named 'var_file' in the path "./my_paddle_model".
"""
if vars is None:
if main_program is None:
......@@ -221,7 +406,42 @@ def load_vars(executor,
def load_params(executor, dirname, main_program=None, filename=None):
"""
load all parameters from directory by executor.
This function filters out all parameters from the give `main_program`
and then trys to load these parameters from the folder `dirname` or
the file `filename`.
Use the `dirname` to specify the folder where parameters were saved. If
parameters were saved in separate files in the folder `dirname`, set
`filename` None; if all parameters were saved in a single file, use
`filename` to specify the file name.
NOTICE: Some variables are not Parameter while they are necessary for
training. So you can NOT save and continue your training just by
`save_params()` and `load_params()`. Please use `save_persistables()`
and `load_persistables()` instead.
Args:
executor(Executor): The executor to run for loading parameters.
dirname(str): The directory path.
main_program(Program|None): The program whose parameters will be
loaded. If it is None, the default
main program will be used automatically.
Default: None
filename(str|None): The file which saved all parameters. If parameters
were saved in differnet files, set it to None.
Default: None
Returns:
None
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
param_path = "./my_paddle_model"
prog = fluid.default_main_program()
fluid.io.load_params(executor=exe, dirname=param_path,
main_program=None)
"""
load_vars(
executor,
......@@ -233,7 +453,37 @@ def load_params(executor, dirname, main_program=None, filename=None):
def load_persistables(executor, dirname, main_program=None, filename=None):
"""
load all persistables from directory by executor.
This function filters out all variables with `persistable==True` from the
give `main_program` and then trys to load these variables from the folder
`dirname` or the file `filename`.
Use the `dirname` to specify the folder where persistable variables were
saved. If variables were saved in separate files, set `filename` None;
if all variables were saved in a single file, use `filename` to specify
the file name.
Args:
executor(Executor): The executor to run for loading persistable variables.
dirname(str): The directory path.
main_program(Program|None): The program whose persistbale variables will
be loaded. If it is None, the default main
program will be used automatically.
Default: None
filename(str|None): The file which saved all variables. If variables were
saved in differnet files, set it to None.
Default: None
Returns:
None
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
param_path = "./my_paddle_model"
prog = fluid.default_main_program()
fluid.io.load_persistables(executor=exe, dirname=param_path,
main_program=None)
"""
load_vars(
executor,
......@@ -306,22 +556,48 @@ def save_inference_model(dirname,
model_filename=None,
params_filename=None):
"""
Build a model especially for inference,
and save it to directory by the executor.
Prune the given `main_program` to build a new program especially for inference,
and then save it and all related parameters to given `dirname` by the `executor`.
Args:
dirname(str): The directory path to save the inference model.
feeded_var_names(list[str]): Names of variables that need to be feeded data
during inference.
target_vars(list[Variable]): Variables from which we can get inference
results.
executor(Executor): The executor that saves the inference model.
main_program(Program|None): The original program, which will be pruned to
build the inference model. If is setted None,
the default main program will be used.
Default: None.
model_filename(str|None): The name of file to save the inference program
itself. If is setted None, a default filename
`__model__` will be used.
params_filename(str|None): The name of file to save all related parameters.
If it is setted None, parameters will be saved
in separate files .
:param dirname: directory path
:param feeded_var_names: Names of variables that need to be feeded data during inference
:param target_vars: Variables from which we can get inference results.
:param executor: executor that save inference model
:param main_program: original program, which will be pruned to build the inference model.
Default default_main_program().
:param model_filename: The name of file to save inference program.
If not specified, default filename `__model__` will be used.
:param params_filename: The name of file to save parameters.
It is used for the case that all parameters are saved in a single binary file.
If not specified, parameters are considered saved in separate files.
Returns:
None
Raises:
ValueError: If `feed_var_names` is not a list of basestring.
ValueError: If `target_vars` is not a list of Variable.
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
path = "./infer_model"
fluid.io.save_inference_model(dirname=path, feeded_var_names=['img'],
target_vars=[predict_var], executor=exe)
# In this exsample, the function will prune the default main program
# to make it suitable for infering the `predict_var`. The pruned
# inference program is going to be saved in the "./infer_model/__model__"
# and parameters are going to be saved in separate files under folder
# "./infer_model".
:return: None
"""
if isinstance(feeded_var_names, basestring):
feeded_var_names = [feeded_var_names]
......@@ -382,18 +658,49 @@ def load_inference_model(dirname,
"""
Load inference model from a directory
:param dirname: directory path
:param executor: executor that load inference model
:param model_filename: The name of file to load inference program.
If not specified, default filename `__model__` will be used.
:param params_filename: The name of file to load parameters.
It is used for the case that all parameters are saved in a single binary file.
If not specified, parameters are considered saved in separate files.
Args:
dirname(str): The directory path
executor(Executor): The executor to run for loading inference model.
model_filename(str|None): The name of file to load inference program.
If it is None, the default filename
'__model__' will be used.
Default: None
params_filename(str|None): The name of file to load all parameters.
It is only used for the case that all
parameters were saved in a single binary
file. If parameters were saved in separate
files, set it as 'None'.
Returns:
tuple: The return of this function is a tuple with three elements:
(program, feed_target_names, fetch_targets). The `program` is a
Program, it's the program for inference. The `feed_target_names` is
a list of str, it contains Names of variables that need to feed
data in the inference program. The `fetch_targets` is a list of
Variable. It contains variables from which we can get inference
results.
Raises:
ValueError: If `dirname` is not a existing directory.
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
path = "./infer_model"
[inference_program, feed_target_names, fetch_targets] =
fluid.io.load_inference_model(dirname=path, executor=exe)
results = exe.run(inference_program,
feed={feed_target_names[0]: tensor_img},
fetch_list=fetch_targets)
# In this exsample, the inference program was saved in the
# "./infer_model/__model__" and parameters were saved in
# separate files in ""./infer_model".
# After getting inference program, feed target names and
# fetch targets, we can use an Executor to run the inference
# program to get the inference result.
:return: [program, feed_target_names, fetch_targets]
program: program especially for inference.
feed_target_names: Names of variables that need to feed data
fetch_targets: Variables from which we can get inference results.
"""
if not os.path.isdir(dirname):
raise ValueError("There is no directory named '%s'", dirname)
......@@ -424,12 +731,25 @@ def load_inference_model(dirname,
def get_parameter_value(para, executor):
"""
Get the LoDTensor for the parameter
Get the LoDTensor value of the given parameter.
Args:
para(Parameter): The parameter to get value from.
executor(Executor): The executor to run for retrieving the value.
Returns:
numpy.array: The given parameter's values.
Raises:
AssertionError: If the `para` is not an instance of Parameter.
:param executor: executor for retrieving the value
:param para: the given parameter
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
param = fluid.default_main_program().global_block().var('fc.w')
p = fluid.io.get_parameter_value(param, exe)
:return: the LoDTensor for the parameter
"""
assert is_parameter(para)
......@@ -441,14 +761,30 @@ def get_parameter_value(para, executor):
def get_parameter_value_by_name(name, executor, program=None):
"""
Get the LoDTensor for paramter with the given name
Get the LoDTensor value of a certain parameter by its name.
Args:
name(str): The parameter's name.
executor(Executor): The executor to run for retrieving the value.
program(Program | None): The program where to find the parameter.
If it's set to be None, the function will
try to find the parameter in the default
main program.
:param executor: executor for retrieving the value
:param name: the name of the parameter
:param program: the program where the variable is found
Default default_main_program().
Returns:
numpy.array: The parameter's values.
:return: the LoDTensor for the variable
Raises:
TypeError: If given `name` is not an instance of basestring.
TypeError: If the parameter with the given name doesn't exist.
AssertionError: If there is a varibale named `name` in the
given program but it is not a Parameter.
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
p = fluid.io.get_parameter_value('fc.w', exe)
"""
if program is None:
program = default_main_program()
......@@ -470,16 +806,58 @@ def save_checkpoint(executor,
main_program=None,
max_num_checkpoints=3):
"""
Save Checkpoint will save persistable LodTensor variables from main_program in checkpoint directory,
the directory named by serial number from 0 to (n -1), save_checkpoint use LRU strategy
to keep numbers of checkpoint directory, the numbers of checkpoint directory are max_num_checkpoints at most,
The interval between two saved checkpoints must greater than save_interval_secs.
This function filters out all checkpoint variables from the give
main_program and then saves these variables to the `checkpoint_dir`
directory.
In the training precess, we generally save a checkpoint in each
iteration. So there might be a lot of checkpoints in the
`checkpoint_dir`. To avoid them taking too much disk space, the
`max_num_checkpoints` are introduced to limit the total number of
checkpoints. If the number of existing checkpints is greater than
the `max_num_checkpoints`, oldest ones will be scroll deleted.
A variable is a checkpoint variable and will be saved if it meets
all following conditions:
1. It's persistable.
2. It's type is not FEED_MINIBATCH nor FETCH_LIST nor RAW.
3. It's name contains no "@GRAD" nor ".trainer_" nor ".block".
:param executor executor for save the value
:param checkpoint_dir the checkpoint directory
:param trainer_id currect trainer id, if id is equal to 0, the trainer is chief
:param main_program will save all variables in program
:param max_num_checkpoints will keep numbers of checkpoint serials not bigger than max_num_checkpoints
Args:
executor(Executor): The executor to run for save checkpoint.
checkpoint_dir(str): The folder where to save checkpoints.
trainer_id(int): currect trainer id, if id is equal to 0, the trainer
is chief.
trainer_args(dict|None): Current training arguments. Such as 'epoch_id'
and 'step_id'.
Defaut: None
main_program(Program|None): The program whose checkpoint variables will
be saved. If it is None, the default main program will be used.
max_num_checkpoints(int): The max number of total number of existing
checkpoints.
Default: 3
Returns:
None
Raises:
ValueError: If `checkpoint_dir` is None.
AssertionError: If `trainer_args` is not a dict.
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
path = "./checkpoints"
prog = fluid.default_main_program()
trainer_args = {"epoch_id": 200,
"step_id": 20} # just an example
fluid.io.save_checkpoint(executor=exe,
checkpoint_dir=path,
trainer_id=0,
trainer_args=trainer_args,
main_program=prog,
max_num_checkpoints=3)
"""
if checkpoint_dir is None:
raise ValueError("'checkpoint_dir' should not be None")
......@@ -503,13 +881,50 @@ def save_checkpoint(executor,
def load_checkpoint(executor, checkpoint_dir, serial, main_program):
"""
Load checkpoint from a directory by executor,
it will find the most recent saved checkpoint file and load it auto.
This function filters out all checkpoint variables from the give
main_program and then try to load these variables from the
`checkpoint_dir` directory.
In the training precess, we generally save a checkpoint in each
iteration. So there are more than one checkpoint in the
`checkpoint_dir` (each checkpoint has its own sub folder), use
`serial` to specify which serial of checkpoint you would like to
load.
A variable is a checkpoint variable and will be loaded if it meets
all following conditions:
1. It's persistable.
2. It's type is not FEED_MINIBATCH nor FETCH_LIST nor RAW.
3. It's name contains no "@GRAD" nor ".trainer_" nor ".block".
Args:
executor(Executor): The executor to run for loading checkpoint.
checkpoint_dir(str): The folder where all checkpoints are.
serial(int): The serial of checkpoint you would like to load.
main_program(Program): The program whose checkpoint variables will
be loaded.
:param executor executor for load the value
:param checkpoint_dir the checkpoint directory
:param serial the serial folder in checkpoint directory will be load
:param main_program will load all variables in program
Returns:
None
Raises:
ValueError: If `checkpoint_dir` is None.
ValueError: If `serial` is None or `serial` is less than 0.
ValueError: If `main_program` is None.
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
path = "./checkpoints"
prog = fluid.default_main_program()
fluid.io.load_checkpoint(executor=exe, checkpoint_dir=path,
serial=9, main_program=prog)
# In this example, `load_checkpoint` function
# will first filters out all checkpoint variables in the default
# main program, and then try to load these variables form the
# folder "./checkpoints/checkpoint_9/__model__".
"""
if checkpoint_dir is None:
......@@ -530,8 +945,8 @@ def clean_checkpoint(checkpoint_dir, delete_dir=False):
clean the checkpoint dir, when the train exits normally, the trainer will call clean_checkpoint to delete checkpoint directory saved before.
delete_dir only works when the directory is empty, otherwise, OSError is raised.
:param checkpoint_dir
:param delete_dir
: param checkpoint_dir
: param delete_dir
"""
if checkpoint_dir is None:
......@@ -547,13 +962,40 @@ def load_persist_vars_without_grad(executor,
program,
has_model_dir=False):
"""
load_persist_vars_without_grad will load variables from a directory by an executor,
the variable named end with "@GRAD" will not be loaded.
This function filters out all checkpoint variables from the give
program and then trys to load these variables from the given directory.
A variable is a checkpoint variable if it meets all following
conditions:
1. It's persistable.
2. It's type is not FEED_MINIBATCH nor FETCH_LIST nor RAW.
3. It's name contains no "@GRAD" nor ".trainer_" nor ".block".
:param executor executor for load the value
:param dirname the checkpoint directory
:param program will load all variables in program
:param has_model_dir if has_model_dir is True, will load variables from sub directory named __model__
Args:
executor(Executor): The executor to run for loading variables.
dirname(str): The directory path.
program(Program): The program whose checkpoint variables will
be loaded.
has_model_dir(bool): if True, the function loads variables
from a sub directory named '__model__'.
Default: False
Returns:
None
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
param_path = "./my_paddle_model"
prog = fluid.default_main_program()
fluid.io.load_persist_vars_without_grad(executor=exe,
dirname=param_path, program=prog, has_model_dir=True)
# In this example, `load_persist_vars_without_grad` function
# will first filters out all checkpoint variables in the default
# main program, and then trys to load these variables form the
# folder "./my_paddle_model/__model__".
"""
if has_model_dir:
......@@ -569,12 +1011,38 @@ def load_persist_vars_without_grad(executor,
def save_persist_vars_without_grad(executor, dirname, program):
"""
save_persist_vars_without_grad will save variables to a directory by an executor,
the variable named end with "@GRAD" will not be saved.
This function filters out all checkpoint variables from the give
program and then save these variables to a sub-folder '__model__' of
the given directory.
A variable is a checkpoint variable if it meets all following
conditions:
1. It's persistable.
2. It's type is not FEED_MINIBATCH nor FETCH_LIST nor RAW.
3. It's name contains no "@GRAD" nor ".trainer_" nor ".block".
Args:
executor(Executor): The executor to run for saving variables.
dirname(str): The directory path.
program(Program): The program whose checkpoint variables will
be saved.
Returns:
None
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
param_path = "./my_paddle_model"
prog = fluid.default_main_program()
fluid.io.save_persist_vars_without_grad(executor=exe,
dirname=param_path, program=prog)
:param executor executor for load the value
:param dirname the checkpoint directory
:param program will load all variables in program
# In this example, `save_persist_vars_without_grad` function
# will first filters out all checkpoint variables in the default
# main program, and then saves these variables to the folder
# "./my_paddle_model/__model__".
"""
cur_dir = _get_model_dir(dirname)
save_vars(
......@@ -620,7 +1088,7 @@ def _is_checkpoint_var(var):
the checkpoint will not save or load all the variables.
var type is FEED_MINIBATCH/FETCH_LIST/RAW or var name ends with @GRAD are discarded.
:param var
: param var
"""
if var.desc.type() == core.VarDesc.VarType.FEED_MINIBATCH or \
var.desc.type() == core.VarDesc.VarType.FETCH_LIST or \
......@@ -701,7 +1169,7 @@ def _write_success(dirname):
"""
write an empty file named "_SUCCESS" in checkpoint dir, indicate this checkpoint is correct.
:param dirname
: param dirname
"""
success_file = os.path.join(dirname, SUCCESS_MARK_FILENAME)
with open(success_file, 'a') as f:
......@@ -713,7 +1181,7 @@ def get_latest_checkpoint_serial(checkpoint_dir):
"""
get the latest file in checkpoint directory, the _SUCCESS file must exist in the directory
:param checkpoint_dir
: param checkpoint_dir
"""
if not checkpoint_dir:
return -1
......
python/paddle/fluid/layers/control_flow.py
浏览文件 @ 9942a304
......@@ -185,7 +185,9 @@ def Print(input,
Returns:
Variable: Output tensor, same data with input tensor.
Examples:
.. code-block:: python
value = some_layer(...)
......@@ -1201,6 +1203,31 @@ class ConditionalBlockGuard(BlockGuard):
class ConditionalBlock(object):
'''
**ConditionalBlock**
ConditionalBlock is an operator that bind a block to a specific condition,
if the condition matches, the corresponding block will be executed.
Args:
inputs (Variable): bool conditions.
is_scalar_condition (bool): whether the branch is controled by a scalar.
name(str): name of this ConditionalBlock.
Examples:
.. code-block:: python
cond = layers.less_than(x=label, y=limit)
true_image, false_image = layers.split_lod_tensor(
input=image, mask=cond)
true_cond = layers.ConditionalBlock([true_image])
with true_cond.block():
...
with false_cond.block():
...
'''
def __init__(self, inputs, is_scalar_condition=False, name=None):
for each_input in inputs:
if not isinstance(each_input, Variable):
......
python/paddle/fluid/layers/detection.py
浏览文件 @ 9942a304
......@@ -16,7 +16,7 @@ All layers just related to the detection neural network.
"""
from layer_function_generator import generate_layer_fn
from layer_function_generator import autodoc
from layer_function_generator import autodoc, templatedoc
from ..layer_helper import LayerHelper
import tensor
import nn
......@@ -155,7 +155,7 @@ def detection_output(loc,
return nmsed_outs
@autodoc()
@templatedoc()
def detection_map(detect_res,
label,
class_num,
......@@ -166,6 +166,47 @@ def detection_map(detect_res,
input_states=None,
out_states=None,
ap_version='integral'):
"""
${comment}
Args:
detect_res: ${detect_res_comment}
label: ${label_comment}
class_num: ${class_num_comment}
background_label: ${background_label_comment}
overlap_threshold: ${overlap_threshold_comment}
evaluate_difficult: ${evaluate_difficult_comment}
has_state: ${has_state_comment}
input_states: If not None, It contains 3 elements:
1. pos_count ${pos_count_comment}.
2. true_pos ${true_pos_comment}.
3. false_pos ${false_pos_comment}.
out_states: If not None, it contains 3 elements.
1. accum_pos_count ${accum_pos_count_comment}.
2. accum_true_pos ${accum_true_pos_comment}.
3. accum_false_pos ${accum_false_pos_comment}.
ap_version: ${ap_type_comment}
Returns:
${map_comment}
Examples:
.. code-block:: python
detect_res = fluid.layers.data(
name='detect_res',
shape=[10, 6],
append_batch_size=False,
dtype='float32')
label = fluid.layers.data(
name='label',
shape=[10, 6],
append_batch_size=False,
dtype='float32')
map_out = fluid.layers.detection_map(detect_res, label, 21)
"""
helper = LayerHelper("detection_map", **locals())
def __create_var(type):
......
python/paddle/fluid/layers/nn.py
浏览文件 @ 9942a304
......@@ -93,6 +93,7 @@ __all__ = [
'mean_iou',
'relu',
'log',
'crop',
]
......@@ -197,7 +198,10 @@ def fc(input,
else:
pre_bias = helper.create_tmp_variable(dtype)
helper.append_op(
type="sum", inputs={"X": mul_results}, outputs={"Out": pre_bias})
type="sum",
inputs={"X": mul_results},
outputs={"Out": pre_bias},
attrs={"use_mkldnn": use_mkldnn})
# add bias
pre_activation = helper.append_bias_op(pre_bias, dim_start=num_flatten_dims)
# add activation
......@@ -2674,18 +2678,35 @@ def sequence_expand(x, y, ref_level=-1, name=None):
def beam_search(pre_ids, ids, scores, beam_size, end_id, level=0):
'''
**beam search**
This function implements the beam search algorithm.
Beam search is a classical algorithm for selecting candidate words
in a machine translation task.
Refer to `Beam search <https://en.wikipedia.org/wiki/Beam_search>`_
for more details.
Args:
pre_ids (Variable): ${pre_ids_comment}
ids (Variable): ${ids_comment}
scores (Variable): ${scores_comment}
beam_size (int): ${beam_size_comment}
end_id (int): ${end_id_comment}
level (int): ${level_comment}
pre_ids (Variable): ids in previous step.
ids (Variable): a LoDTensor of shape of [None,k]
scores (Variable): a LoDTensor that has the same shape and LoD with `ids`
beam_size (int): beam size for beam search
end_id (int): the token id which indicates the end of a sequence
level (int): the level of LoDTensor
Returns:
tuple: a tuple of beam_search output variables: selected_ids, selected_scores
tuple: a tuple of beam_search output variables: `selected_ids`, `selected_scores`
Examples:
.. code-block:: python
# current_score is a Tensor of shape (num_batch_size, embed_size), which
# consists score of each candidate word.
topk_scores, topk_indices = pd.topk(current_score, k=50)
selected_ids, selected_scores = pd.beam_search(
pre_ids, topk_indices, topk_scores, beam_size, end_id=10, level=0)
'''
helper = LayerHelper('beam_search', **locals())
score_type = scores.dtype
......@@ -5003,3 +5024,101 @@ def mean_iou(input, label, num_classes):
},
attrs={"num_classes": num_classes})
return out_mean_iou, out_wrong, out_correct
def crop(x, shape=None, offsets=None, name=None):
"""
Crop input into output, as specified by offsets and shape.
.. code-block:: text
* Case 1:
Given
X = [[0, 1, 2, 0, 0]
[0, 3, 4, 0, 0]
[0, 0, 0, 0, 0]],
and
shape = [2, 2],
offsets = [0, 1],
output is:
Out = [[1, 2],
[3, 4]].
* Case 2:
Given
X = [[0, 1, 2, 5, 0]
[0, 3, 4, 6, 0]
[0, 0, 0, 0, 0]],
and shape is tensor
shape = [[0, 0, 0]
[0, 0, 0]]
and
offsets = [0, 1],
output is:
Out = [[1, 2, 5],
[3, 4, 6]].
Args:
x (Variable): The input tensor variable.
shape (Variable|list/tuple of integer): The output shape is specified
by `shape`, which can a Variable or a list/tupe of integer.
If a tensor Variable, it's rank must be the same as `x`. This way
is suitable for the case that the output shape may be changed each
iteration. If a list/tupe of integer, it's length must be the same
as the rank of `x`
offsets (Variable|list/tuple of integer|None): Specifies the copping
offsets at each dimension. It can be a Variable or or a list/tupe
of integer. If a tensor Variable, it's rank must be the same as `x`.
This way is suitable for the case that the offsets may be changed
each iteration. If a list/tupe of integer, it's length must be the
same as the rank of `x`. If None, the offsets are 0 at each
dimension.
name(str|None): A name for this layer(optional). If set None, the layer
will be named automatically.
Returns:
Variable: The cropped tensor variable.
Raises:
ValueError: If shape is not a list, tuple or Variable.
Examples:
.. code-block:: python
x = fluid.layers.data(name="x", shape=[3, 5], dtype="float32")
y = fluid.layers.data(name="y", shape=[2, 3], dtype="float32")
crop = fluid.layers.crop(x, shape=y)
# or
z = fluid.layers.data(name="z", shape=[3, 5], dtype="float32")
crop = fluid.layers.crop(z, shape=[2, 3])
"""
helper = LayerHelper('crop', **locals())
if not (isinstance(shape, list) or isinstance(shape, tuple) or \
isinstance(shape, Variable)):
raise ValueError("The shape should be a list, tuple or Variable.")
if offsets is None:
offsets = [0] * len(x.shape)
out = helper.create_tmp_variable(x.dtype)
ipts = {'X': x}
attrs = {}
if isinstance(shape, Variable):
ipts['Y'] = shape
else:
attrs['shape'] = shape
if isinstance(offsets, Variable):
ipts['Offsets'] = offsets
else:
attrs['offsets'] = offsets
helper.append_op(
type='crop',
inputs=ipts,
outputs={'Out': out},
attrs=None if len(attrs) == 0 else attrs)
return out
python/paddle/fluid/layers/tensor.py
浏览文件 @ 9942a304
......@@ -230,7 +230,11 @@ def sums(input, out=None):
helper = LayerHelper('sum', **locals())
if out is None:
out = helper.create_tmp_variable(dtype=helper.input_dtype())
helper.append_op(type='sum', inputs={'X': input}, outputs={'Out': out})
helper.append_op(
type='sum',
inputs={'X': input},
outputs={'Out': out},
attrs={'use_mkldnn': False})
return out
......
python/paddle/fluid/optimizer.py
浏览文件 @ 9942a304
......@@ -26,10 +26,10 @@ from clip import append_gradient_clip_ops, error_clip_callback
from contextlib import contextmanager
__all__ = [
'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad',
'SGD', 'Momentum', 'Adagrad', 'Adam', 'Adamax', 'DecayedAdagrad', 'Ftrl',
'SGDOptimizer', 'MomentumOptimizer', 'AdagradOptimizer', 'AdamOptimizer',
'AdamaxOptimizer', 'DecayedAdagradOptimizer', 'RMSPropOptimizer',
'Adadelta', 'ModelAverage', 'Optimizer'
'FtrlOptimizer', 'Adadelta', 'ModelAverage', 'Optimizer', 'RMSPropOptimizer'
]
......@@ -192,15 +192,15 @@ class Optimizer(object):
"""Add optimization operators to update gradients to variables.
Args:
loss: the target that this optimization is for.
parameters_and_grads: a list of (variable, gradient) pair to update.
loss(Variable): the target that this optimization is for.
parameters_and_grads(list(tuple(Variable, Variable))):
a list of (variable, gradient) pair to update.
Returns:
return_op_list: a list of operators that will complete one step of
optimization. This will include parameter update ops, global step
update ops and any other custom ops required by subclasses to manage
their internal state.
:param startup_program:
"""
# This is a default implementation of create_optimization_pass that
# can be shared by most optimizers. This implementation assumes that
......@@ -268,7 +268,22 @@ class Optimizer(object):
class SGDOptimizer(Optimizer):
""" Simple SGD optimizer without any state.
"""
Optimizer of the stochastic gradient descent algorithm.
.. math::
param\_out = param - learning\_rate * grad
Args:
learning_rate (float|Variable): the learning rate used to update parameters. \
Can be a float value or a Variable with one float value as data element.
Examples:
.. code-block:: python
sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.2)
sgd_optimizer.minimize(cost)
"""
def __init__(self, learning_rate, **kwargs):
......@@ -294,7 +309,37 @@ class SGDOptimizer(Optimizer):
class MomentumOptimizer(Optimizer):
"""Simple Momentum optimizer with velocity state
"""
Simple Momentum optimizer with velocity state
This optimizer has a flag for Nestrov Momentum.
The update equations are as follows:
.. math::
& velocity = mu * velocity + gradient
& if (use\_nesterov):
&\quad param = param - gradient * learning\_rate + mu * velocity * learning\_rate
& else:
&\quad param = param - learning\_rate * velocity
Args:
learning_rate (float|Variable): the learning rate used to update parameters. \
Can be a float value or a Variable with one float value as data element.
momentum (float): momentum factor
use_nesterov (bool): enables Nesterov momentum
Examples:
.. code-block:: python
optimizer = fluid.optimizer.Momentum(learning_rate=0.2, momentum=0.1)
optimizer.minimize(cost)
"""
_velocity_acc_str = "velocity"
......@@ -338,7 +383,32 @@ class MomentumOptimizer(Optimizer):
class AdagradOptimizer(Optimizer):
"""Simple Adagrad optimizer with moment state
"""
**Adaptive Gradient Algorithm (Adagrad)**
The update is done as follows:
.. math::
moment\_out &= moment + grad * grad
param\_out &= param - \\frac{learning\_rate * grad}{\sqrt{moment\_out} + \epsilon}
The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
does not have the epsilon attribute. It is added here in our implementation
as also proposed here: http://cs231n.github.io/neural-networks-3/#ada
for numerical stability to avoid the division by zero error.
Args:
learning_rate (float|Variable): the learning rate used to update parameters. \
Can be a float value or a Variable with one float value as data element.
epsilon (float): a small float value for numerical stability.
Examples:
.. code-block:: python
optimizer = fluid.optimizer.Adagrad(learning_rate=0.2)
optimizer.minimize(cost)
"""
_moment_acc_str = "moment"
......@@ -379,7 +449,40 @@ class AdagradOptimizer(Optimizer):
class AdamOptimizer(Optimizer):
"""Implements the Adam Optimizer
"""
This implements the Adam optimizer from Section 2 of the Adam
paper : https://arxiv.org/abs/1412.6980.
Adam is a first-order gradient-based optimization method based on
adaptive estimates of lower-order moments.
Adam updates:
.. math::
t & = t + 1
moment\_1\_out & = {\\beta}_1 * moment\_1 + (1 - {\\beta}_1) * grad
moment\_2\_out & = {\\beta}_2 * moment\_2 + (1 - {\\beta}_2) * grad * grad
learning\_rate & = learning\_rate * \\
\\frac{\sqrt{1 - {\\beta}_2^t}}{1 - {\\beta}_1^t}
param\_out & = param - learning\_rate * \\frac{moment\_1}{\sqrt{moment\_2} + \epsilon}
Args:
learning_rate (float|Variable): the learning rate used to update parameters. \
Can be a float value or a Variable with one float value as data element.
beta1 (float): The exponential decay rate for the 1st moment estimates.
beta2 (float): The exponential decay rate for the 2nd moment estimates.
epsilon (float): a small float value for numerical stability.
Examples:
.. code-block:: python
optimizer = fluid.optimizer.Adam(learning_rate=0.2)
optimizer.minimize(cost)
"""
_moment1_acc_str = "moment1"
_moment2_acc_str = "moment2"
......@@ -484,7 +587,42 @@ class AdamOptimizer(Optimizer):
class AdamaxOptimizer(Optimizer):
"""Implements the Adamax Optimizer
"""
We implement the Adamax optimizer from Section 7 of the Adam
paper: https://arxiv.org/abs/1412.6980. Adamax is a variant of the
Adam algorithm based on the infinity norm.
Adamax updates:
.. math::
t & = t + 1
moment\_out & = {\\beta}_1 * moment + (1 - {\\beta}_1) * grad
inf\_norm\_out & = max({\\beta}_2 * inf\_norm + \epsilon, |grad|)
learning\_rate & = \\frac{learning\_rate}{1 - {\\beta}_1^t}
param\_out & = param - learning\_rate * \\frac{moment\_out}{inf\_norm\_out}
The original paper does not have an epsilon attribute.
However, it is added here for numerical stability to prevent the
division by 0 error.
Args:
learning_rate (float|Variable): the learning rate used to update parameters. \
Can be a float value or a Variable with one float value as data element.
beta1 (float): The exponential decay rate for the 1st moment estimates.
beta2 (float): The exponential decay rate for the 2nd moment estimates.
epsilon (float): a small float value for numerical stability.
Examples:
.. code-block:: python
optimizer = fluid.optimizer.Adamax(learning_rate=0.2)
optimizer.minimize(cost)
"""
_moment_acc_str = "moment"
_inf_norm_acc_str = "inf_norm"
......@@ -568,7 +706,34 @@ class AdamaxOptimizer(Optimizer):
class DecayedAdagradOptimizer(Optimizer):
"""Simple Decayed Adagrad optimizer with moment state
"""
**Decayed Adagrad Optimizer**
The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
The update is done as follows:
.. math::
moment\_out & = decay * moment + (1 - decay) * grad * grad
param\_out & = param - \\frac{learning\_rate * grad}{\sqrt{moment\_out} + \epsilon}
The original paper(http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
does not have an epsilon attribute. It is added here for numerical
stability to avoid the division by zero error.
Args:
learning_rate (float|Variable): the learning rate used to update parameters. \
Can be a float value or a Variable with one float value as data element.
decay (float): decay rate.
epsilon (float): a small float value for numerical stability.
Examples:
.. code-block:: python
optimizer = fluid.optimizer.DecayedAdagrad(learning_rate=0.2)
optimizer.minimize(cost)
"""
_moment_acc_str = "moment"
......@@ -614,6 +779,7 @@ class DecayedAdagradOptimizer(Optimizer):
class AdadeltaOptimizer(Optimizer):
"""
**Adadelta Optimizer**
Simple Adadelta optimizer with average squared grad state and
average squared update state.
The details of adadelta please refer to this
......@@ -628,7 +794,7 @@ class AdadeltaOptimizer(Optimizer):
E(dx_t^2) &= \\rho * E(dx_{t-1}^2) + (1-\\rho) * (-g*learning\\_rate)^2
Args:
learning_rate(float): global leraning rate
learning_rate(float): global learning rate
rho(float): rho in equation
epsilon(float): epsilon in equation
......@@ -703,37 +869,37 @@ class RMSPropOptimizer(Optimizer):
.. math::
r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2 \\\\
r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
w & = w - \\frac{\\eta} {\\sqrt{r(w,t) + \\epsilon}} \\nabla Q_{i}(w)
The first equation calculates moving average of the squared gradient for
each weight. Then dividing the gradient by :math: `sqrt{v(w,t)}`.
each weight. Then dividing the gradient by :math:`sqrt{v(w,t)}`.
In some cases, adding a momentum term :math: `\\beta` is beneficial.
In our implementation, Nesterov momentum is used:
.. math::
r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2 \\\\
r(w, t) & = \\rho r(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2
v(w, t) & = \\beta v(w, t-1) + \\frac{\\eta} {\\sqrt{v(w,t) +
\\epsilon}} \\nabla Q_{i}(w)
w & = w - v(w, t)
where, :math: `\\rho` is a hyperparameter and typical values are 0.9, 0.95
where, :math:`\\rho` is a hyperparameter and typical values are 0.9, 0.95
and so on. :math: `beta` is the momentum term. :math: `\\epsilon` is a
smoothing term to avoid division by zero, usually set somewhere in range
from 1e-4 to 1e-8.
Args:
learning_rate(float): global leraning rate.
learning_rate(float): global learning rate.
rho(float): rho is :math: `\\rho` in equation, set 0.95 by default.
epsilon(float): :math: `\\epsilon` in equation is smoothing term to
avoid division by zero, set 1e-6 by default.
momentum(float): :math: `\\beta` in equation is the momentum term,
momentum(float): :math:`\\beta` in equation is the momentum term,
set 0.0 by default.
Raises:
......@@ -810,6 +976,113 @@ class RMSPropOptimizer(Optimizer):
return rmsprop_op
class FtrlOptimizer(Optimizer):
"""
FTRL (Follow The Regularized Leader) Optimizer.
The paper that proposed Follow The Regularized Leader (FTRL):
(https://www.eecs.tufts.edu/~dsculley/papers/ad-click-prediction.pdf)
.. math::
&new\_accum = squared\_accum + grad^2
&if (lr\_power == -0.5):
&\quad linear\_accum += grad - \\frac{\\sqrt{new\_accum} - \\sqrt{squared\_accum}}{learning\_rate * param}
&else:
&\quad linear\_accum += grad - \\frac{new\_accum^{-lr\_power} - accum^{-lr\_power}}{learning\_rate * param}
&x = l1 * sign(linear\_accum) - linear\_accum
&if (lr\_power == -0.5):
&\quad y = \\frac{\\sqrt{new\_accum}}{learning\_rate} + (2 * l2)
&\quad pre\_shrink = \\frac{x}{y}
&\quad param = (abs(linear\_accum) > l1).select(pre\_shrink, 0.0)
&else:
&\quad y = \\frac{new\_accum^{-lr\_power}}{learning\_rate} + (2 * l2)
&\quad pre\_shrink = \\frac{x}{y}
&\quad param = (abs(linear\_accum) > l1).select(pre\_shrink, 0.0)
&squared\_accum += grad^2
Args:
learning_rate (float|Variable): global learning rate.
l1 (float):
l2 (float):
lr_power (float):
Raises:
ValueError: If learning_rate, rho, epsilon, momentum are None.
Examples:
.. code-block:: python
optimizer = fluid.optimizer.Ftrl(0.0001)
_, params_grads = optimizer.minimize(cost)
"""
_squared_acc_str = "squared"
_linear_acc_str = "linear"
def __init__(self, learning_rate, l1=0.0, l2=0.0, lr_power=-0.5, **kwargs):
super(FtrlOptimizer, self).__init__(
learning_rate=learning_rate, **kwargs)
if learning_rate is None:
raise ValueError("learning_rate is not set.")
self.type = "ftrl"
self._l1 = l1
self._l2 = l2
self._lr_power = lr_power
def _create_accumulators(self, block, parameters):
if not isinstance(block, framework.Block):
raise TypeError("block is not instance of framework.Block.")
for p in parameters:
self._add_accumulator(self._squared_acc_str, p)
self._add_accumulator(self._linear_acc_str, p)
def _append_optimize_op(self, block, param_and_grad):
if not isinstance(block, framework.Block):
raise TypeError("block is not instance of framework.Block.")
squared_acc = self._get_accumulator(self._squared_acc_str,
param_and_grad[0])
linear_acc = self._get_accumulator(self._linear_acc_str,
param_and_grad[0])
ftrl_op = block.append_op(
type=self.type,
inputs={
"Param": param_and_grad[0],
"Grad": param_and_grad[1],
"SquaredAccumulator": squared_acc,
"LinearAccumulator": linear_acc,
"LearningRate": self._create_param_lr(param_and_grad),
},
outputs={
"ParamOut": param_and_grad[0],
"SquaredAccumOut": squared_acc,
"LinearAccumOut": linear_acc
},
attrs={"l1": self._l1,
"l2": self._l1,
"lr_power": self._lr_power})
return ftrl_op
# We short the class name, since users will use the optimizer with the package
# name. The sample code:
#
......@@ -826,6 +1099,7 @@ Adamax = AdamaxOptimizer
DecayedAdagrad = DecayedAdagradOptimizer
Adadelta = AdadeltaOptimizer
RMSProp = RMSPropOptimizer
Ftrl = FtrlOptimizer
class ModelAverage(Optimizer):
......@@ -844,7 +1118,9 @@ class ModelAverage(Optimizer):
max_average_window: The maximum size of average window.
Examples:
...
.. code-block:: python
optimizer = fluid.optimizer.Momentum()
_, params_grads = optimizer.minimize(cost)
model_average = fluid.optimizer.ModelAverage(params_grads, 0.15,
......
python/paddle/fluid/profiler.py
浏览文件 @ 9942a304
......@@ -42,6 +42,9 @@ def cuda_profiler(output_file, output_mode=None, config=None):
counters/options for profiling by `config` argument. The default config
is ['gpustarttimestamp', 'gpustarttimestamp', 'gridsize3d',
'threadblocksize', 'streamid', 'enableonstart 0', 'conckerneltrace'].
Then users can use NVIDIA Visual Profiler
(https://developer.nvidia.com/nvidia-visual-profiler) tools to load this
this output file to visualize results.
Args:
output_file (string) : The output file name, the result will be
......@@ -50,6 +53,33 @@ def cuda_profiler(output_file, output_mode=None, config=None):
Comma separated values format. It should be 'kvp' or 'csv'.
config (list of string) : The profiler options and counters can refer
to "Compute Command Line Profiler User Guide".
Raises:
ValueError: If `output_mode` is not in ['kvp', 'csv'].
Examples:
.. code-block:: python
import paddle.fluid as fluid
import paddle.fluid.profiler as profiler
epoc = 8
dshape = [4, 3, 28, 28]
data = fluid.layers.data(name='data', shape=[3, 28, 28], dtype='float32')
conv = fluid.layers.conv2d(data, 20, 3, stride=[1, 1], padding=[1, 1])
place = fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
output_file = 'cuda_profiler.txt'
with profiler.cuda_profiler(output_file, 'csv') as nvprof:
for i in range(epoc):
input = np.random.random(dshape).astype('float32')
exe.run(fluid.default_main_program(), feed={'data': input})
# then use NVIDIA Visual Profiler (nvvp) to load this output file
# to visualize results.
"""
if output_mode is None:
output_mode = 'csv'
......@@ -69,19 +99,52 @@ def cuda_profiler(output_file, output_mode=None, config=None):
def reset_profiler():
"""The profiler clear interface.
reset_profiler will clear the previous time record.
"""
Clear the previous time record. This interface does not work for
`fluid.profiler.cuda_profiler`, it only works for
`fluid.profiler.start_profiler`, `fluid.profiler.stop_profiler`,
and `fluid.profiler.profiler`.
Examples:
.. code-block:: python
import paddle.fluid.profiler as profiler
with profiler.profiler(state, 'total', '/tmp/profile'):
for iter in range(10):
if iter == 2:
profiler.reset_profiler()
# ...
"""
core.reset_profiler()
def start_profiler(state):
"""Enable the profiler.
"""
Enable the profiler. Uers can use `fluid.profiler.start_profiler` and
`fluid.profiler.stop_profiler` to insert the code, except the usage of
`fluid.profiler.profiler` interface.
Args:
state (string) : The profiling state, which should be 'CPU', 'GPU'
or 'All'. 'CPU' means only profile CPU. 'GPU' means profiling
GPU as well. 'All' also generates timeline.
Raises:
ValueError: If `state` is not in ['CPU', 'GPU', 'All'].
Examples:
.. code-block:: python
import paddle.fluid.profiler as profiler
profiler.start_profiler('GPU')
for iter in range(10):
if iter == 2:
profiler.reset_profiler()
# except each iteration
profiler.stop_profiler('total', '/tmp/profile')
"""
if core.is_profiler_enabled():
return
......@@ -97,7 +160,10 @@ def start_profiler(state):
def stop_profiler(sorted_key=None, profile_path='/tmp/profile'):
"""Stop the profiler.
"""
Stop the profiler. Uers can use `fluid.profiler.start_profiler` and
`fluid.profiler.stop_profiler` to insert the code, except the usage of
`fluid.profiler.profiler` interface.
Args:
sorted_key (string) : If None, the profiling results will be printed
......@@ -111,6 +177,23 @@ def stop_profiler(sorted_key=None, profile_path='/tmp/profile'):
The `ave` means sorting by the average execution time.
profile_path (string) : If state == 'All', it will write a profile
proto output file.
Raises:
ValueError: If `sorted_key` is not in
['calls', 'total', 'max', 'min', 'ave'].
Examples:
.. code-block:: python
import paddle.fluid.profiler as profiler
profiler.start_profiler('GPU')
for iter in range(10):
if iter == 2:
profiler.reset_profiler()
# except each iteration
profiler.stop_profiler('total', '/tmp/profile')
"""
if not core.is_profiler_enabled():
return
......@@ -137,7 +220,12 @@ def profiler(state, sorted_key=None, profile_path='/tmp/profile'):
Different from cuda_profiler, this profiler can be used to profile both CPU
and GPU program. By defalut, it records the CPU and GPU operator kernels,
if you want to profile other program, you can refer the profiling tutorial
to add more records.
to add more records in C++ code.
If the state == 'All', a profile proto file will be written to
`profile_path`. This file records timeline information during the execution.
Then users can visualize this file to see the timeline, please refer
https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/howto/optimization/timeline.md
Args:
state (string) : The profiling state, which should be 'CPU' or 'GPU',
......@@ -156,6 +244,25 @@ def profiler(state, sorted_key=None, profile_path='/tmp/profile'):
The `ave` means sorting by the average execution time.
profile_path (string) : If state == 'All', it will write a profile
proto output file.
Raises:
ValueError: If `state` is not in ['CPU', 'GPU', 'All']. If `sorted_key` is
not in ['calls', 'total', 'max', 'min', 'ave'].
Examples:
.. code-block:: python
import paddle.fluid.profiler as profiler
with profiler.profiler('All', 'total', '/tmp/profile') as prof:
for pass_id in range(pass_num):
for batch_id, data in enumerate(train_reader()):
exe.run(fluid.default_main_program(),
feed=feeder.feed(data),
fetch_list=[],
use_program_cache=True)
# ...
"""
start_profiler(state)
yield
......
python/paddle/fluid/regularizer.py
浏览文件 @ 9942a304
......@@ -16,8 +16,8 @@ import framework
from . import core
__all__ = [
'append_regularization_ops', 'WeightDecayRegularizer', 'L1Decay', 'L2Decay',
'L1DecayRegularizer', 'L2DecayRegularizer'
'append_regularization_ops', 'L1Decay', 'L2Decay', 'L1DecayRegularizer',
'L2DecayRegularizer'
]
......@@ -36,7 +36,8 @@ def append_regularization_ops(parameters_and_grads, regularization=None):
set. It will be applied with regularizer.
Returns:
list of (parameters, gradients) pair with the regularized gradient
list[(Variable, Variable)]: list of (parameters, gradients) \
pair with the regularized gradient
Raises:
Exception: Unknown regularization type
......@@ -100,6 +101,24 @@ class WeightDecayRegularizer(object):
class L2DecayRegularizer(WeightDecayRegularizer):
"""Implements the L2 Weight Decay Regularization
Small values of L2 can help prevent over fitting the training data.
.. math::
L2WeightDecay = reg\_coeff * parameter
Args:
regularization_coeff(float): regularization coeff
Examples:
.. code-block:: python
optimizer = fluid.optimizer.Adagrad(
learning_rate=1e-4,
regularization=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.1))
optimizer.minimize(avg_cost)
"""
def __init__(self, regularization_coeff=0.0):
......@@ -154,6 +173,27 @@ class L2DecayRegularizer(WeightDecayRegularizer):
class L1DecayRegularizer(WeightDecayRegularizer):
"""Implements the L1 Weight Decay Regularization
L1 regularization encourages sparsity.
.. math::
L1WeightDecay = reg\_coeff * sign(parameter)
Args:
regularization_coeff(float): regularization coeff
Examples:
.. code-block:: python
program = fluid.framework.Program()
block = program.global_block()
mul_x = block.create_parameter(
dtype="float32",
shape=[5, 10],
lod_level=0,
name="mul.x",
regularizer=fluid.regularizer.L1DecayRegularizer(0.5))
"""
def __init__(self, regularization_coeff=0.0):
......
python/paddle/fluid/tests/book/notest_understand_sentiment.py
浏览文件 @ 9942a304
......@@ -194,16 +194,16 @@ def train(word_dict,
if is_local:
train_loop(fluid.default_main_program())
else:
port = os.getenv("PADDLE_INIT_PORT", "6174")
pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip...
port = os.getenv("PADDLE_PSERVER_PORT", "6174")
pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip...
eplist = []
for ip in pserver_ips.split(","):
eplist.append(':'.join([ip, port]))
pserver_endpoints = ",".join(eplist) # ip:port,ip:port...
trainers = int(os.getenv("TRAINERS"))
trainers = int(os.getenv("PADDLE_TRAINERS"))
current_endpoint = os.getenv("POD_IP") + ":" + port
trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
training_role = os.getenv("TRAINING_ROLE", "TRAINER")
trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
t = fluid.DistributeTranspiler()
t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
if training_role == "PSERVER":
......
python/paddle/fluid/tests/book/test_fit_a_line.py
浏览文件 @ 9942a304
......@@ -69,16 +69,16 @@ def train(use_cuda, save_dirname, is_local):
if is_local:
train_loop(fluid.default_main_program())
else:
port = os.getenv("PADDLE_INIT_PORT", "6174")
pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip...
port = os.getenv("PADDLE_PSERVER_PORT", "6174")
pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip...
eplist = []
for ip in pserver_ips.split(","):
eplist.append(':'.join([ip, port]))
pserver_endpoints = ",".join(eplist) # ip:port,ip:port...
trainers = int(os.getenv("TRAINERS"))
trainers = int(os.getenv("PADDLE_TRAINERS"))
current_endpoint = os.getenv("POD_IP") + ":" + port
trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
training_role = os.getenv("TRAINING_ROLE", "TRAINER")
trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
t = fluid.DistributeTranspiler()
t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
if training_role == "PSERVER":
......
python/paddle/fluid/tests/book/test_image_classification.py
浏览文件 @ 9942a304
......@@ -178,16 +178,16 @@ def train(net_type, use_cuda, save_dirname, is_local):
if is_local:
train_loop(fluid.default_main_program())
else:
port = os.getenv("PADDLE_INIT_PORT", "6174")
pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip...
port = os.getenv("PADDLE_PSERVER_PORT", "6174")
pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip...
eplist = []
for ip in pserver_ips.split(","):
eplist.append(':'.join([ip, port]))
pserver_endpoints = ",".join(eplist) # ip:port,ip:port...
trainers = int(os.getenv("TRAINERS"))
trainers = int(os.getenv("PADDLE_TRAINERS"))
current_endpoint = os.getenv("POD_IP") + ":" + port
trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
training_role = os.getenv("TRAINING_ROLE", "TRAINER")
trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
t = fluid.DistributeTranspiler()
t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
if training_role == "PSERVER":
......
python/paddle/fluid/tests/book/test_label_semantic_roles.py
浏览文件 @ 9942a304
......@@ -209,16 +209,16 @@ def train(use_cuda, save_dirname=None, is_local=True):
if is_local:
train_loop(fluid.default_main_program())
else:
port = os.getenv("PADDLE_INIT_PORT", "6174")
pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip...
port = os.getenv("PADDLE_PSERVER_PORT", "6174")
pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip...
eplist = []
for ip in pserver_ips.split(","):
eplist.append(':'.join([ip, port]))
pserver_endpoints = ",".join(eplist) # ip:port,ip:port...
trainers = int(os.getenv("TRAINERS"))
trainers = int(os.getenv("PADDLE_TRAINERS"))
current_endpoint = os.getenv("POD_IP") + ":" + port
trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
training_role = os.getenv("TRAINING_ROLE", "TRAINER")
trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
t = fluid.DistributeTranspiler()
t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
if training_role == "PSERVER":
......
python/paddle/fluid/tests/book/test_machine_translation.py
浏览文件 @ 9942a304
......@@ -200,16 +200,16 @@ def train_main(use_cuda, is_sparse, is_local=True):
if is_local:
train_loop(framework.default_main_program())
else:
port = os.getenv("PADDLE_INIT_PORT", "6174")
pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip...
port = os.getenv("PADDLE_PSERVER_PORT", "6174")
pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip...
eplist = []
for ip in pserver_ips.split(","):
eplist.append(':'.join([ip, port]))
pserver_endpoints = ",".join(eplist) # ip:port,ip:port...
trainers = int(os.getenv("TRAINERS"))
trainers = int(os.getenv("PADDLE_TRAINERS"))
current_endpoint = os.getenv("POD_IP") + ":" + port
trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
training_role = os.getenv("TRAINING_ROLE", "TRAINER")
trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
t = fluid.DistributeTranspiler()
t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
if training_role == "PSERVER":
......
python/paddle/fluid/tests/book/test_recognize_digits.py
浏览文件 @ 9942a304
......@@ -151,16 +151,16 @@ def train(nn_type,
if is_local:
train_loop(fluid.default_main_program())
else:
port = os.getenv("PADDLE_INIT_PORT", "6174")
pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip...
port = os.getenv("PADDLE_PSERVER_PORT", "6174")
pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip...
eplist = []
for ip in pserver_ips.split(","):
eplist.append(':'.join([ip, port]))
pserver_endpoints = ",".join(eplist) # ip:port,ip:port...
trainers = int(os.getenv("TRAINERS"))
trainers = int(os.getenv("PADDLE_TRAINERS"))
current_endpoint = os.getenv("POD_IP") + ":" + port
trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
training_role = os.getenv("TRAINING_ROLE", "TRAINER")
trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
t = fluid.DistributeTranspiler()
t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
if training_role == "PSERVER":
......
python/paddle/fluid/tests/book/test_recommender_system.py
浏览文件 @ 9942a304
......@@ -220,16 +220,16 @@ def train(use_cuda, save_dirname, is_local=True):
if is_local:
train_loop(fluid.default_main_program())
else:
port = os.getenv("PADDLE_INIT_PORT", "6174")
pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip...
port = os.getenv("PADDLE_PSERVER_PORT", "6174")
pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip...
eplist = []
for ip in pserver_ips.split(","):
eplist.append(':'.join([ip, port]))
pserver_endpoints = ",".join(eplist) # ip:port,ip:port...
trainers = int(os.getenv("TRAINERS"))
trainers = int(os.getenv("PADDLE_TRAINERS"))
current_endpoint = os.getenv("POD_IP") + ":" + port
trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
training_role = os.getenv("TRAINING_ROLE", "TRAINER")
trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
t = fluid.DistributeTranspiler()
t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
if training_role == "PSERVER":
......
python/paddle/fluid/tests/book/test_word2vec.py
浏览文件 @ 9942a304
......@@ -125,16 +125,16 @@ def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True):
if is_local:
train_loop(fluid.default_main_program())
else:
port = os.getenv("PADDLE_INIT_PORT", "6174")
pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip...
port = os.getenv("PADDLE_PSERVER_PORT", "6174")
pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip...
eplist = []
for ip in pserver_ips.split(","):
eplist.append(':'.join([ip, port]))
pserver_endpoints = ",".join(eplist) # ip:port,ip:port...
trainers = int(os.getenv("TRAINERS"))
trainers = int(os.getenv("PADDLE_TRAINERS"))
current_endpoint = os.getenv("POD_IP") + ":" + port
trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
training_role = os.getenv("TRAINING_ROLE", "TRAINER")
trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
t = fluid.DistributeTranspiler()
t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
if training_role == "PSERVER":
......
python/paddle/fluid/tests/unittests/test_concat_op.py
浏览文件 @ 9942a304
......@@ -43,7 +43,7 @@ class TestConcatOp(OpTest):
self.axis = 1
class TestConcatOp2(OpTest):
class TestConcatOp2(TestConcatOp):
def init_test_data(self):
self.x0 = np.random.random((2, 3, 4, 5)).astype('float32')
self.x1 = np.random.random((2, 3, 4, 5)).astype('float32')
......@@ -51,5 +51,16 @@ class TestConcatOp2(OpTest):
self.axis = 1
class TestConcatOp3(TestConcatOp):
def init_test_data(self):
self.x0 = np.random.random((1, 256, 170, 256)).astype('float32')
self.x1 = np.random.random((1, 128, 170, 256)).astype('float32')
self.x2 = np.random.random((1, 128, 170, 256)).astype('float32')
self.axis = 1
def test_check_grad(self):
pass
if __name__ == '__main__':
unittest.main()
paddle/contrib/tape/CMakeLists.txt → python/paddle/fluid/tests/unittests/test_gaussian_random_mkldnn_op.py
浏览文件 @ 9942a304
# Copyright (c) 2016 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.
......@@ -11,15 +11,16 @@
# 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)
import unittest
from test_gaussian_random_op import TestGaussianRandomOp
class TestMKLDNN(TestGaussianRandomOp):
def init_kernel_type(self):
self.use_mkldnn = True
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)
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/tests/unittests/test_gaussian_random_op.py
浏览文件 @ 9942a304
......@@ -25,7 +25,15 @@ class TestGaussianRandomOp(unittest.TestCase):
def setUp(self):
self.op_type = "gaussian_random"
self.inputs = {}
self.attrs = {"shape": [1000, 784], "mean": .0, "std": 1., "seed": 10}
self.use_mkldnn = False
self.init_kernel_type()
self.attrs = {
"shape": [1000, 784],
"mean": .0,
"std": 1.,
"seed": 10,
"use_mkldnn": self.use_mkldnn
}
self.outputs = ["Out"]
......@@ -58,6 +66,9 @@ class TestGaussianRandomOp(unittest.TestCase):
self.assertAlmostEqual(numpy.mean(tensor), .0, delta=0.1)
self.assertAlmostEqual(numpy.std(tensor), 1., delta=0.1)
def init_kernel_type(self):
pass
if __name__ == "__main__":
unittest.main()
python/paddle/fluid/tests/unittests/test_layers.py
浏览文件 @ 9942a304
......@@ -401,6 +401,15 @@ class TestBook(unittest.TestCase):
self.assertIsNotNone(output)
print(str(program))
def test_maxout(self):
program = Program()
with program_guard(program):
x = layers.data(name='x', shape=[3, 5], dtype="float32")
y = layers.data(name='y', shape=[2, 3], dtype="float32")
output = layers.crop(x, shape=y)
self.assertIsNotNone(output)
print(str(program))
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/tests/unittests/test_optimizer.py
浏览文件 @ 9942a304
......@@ -434,5 +434,71 @@ class TestDecayedAdagradOptimizer(unittest.TestCase):
self.assertAlmostEqual(init_ops[1].attr('value'), 0.0)
class TestFtrlOptimizer(unittest.TestCase):
class MockFtrl(optimizer.FtrlOptimizer):
def get_accumulators(self):
return self._accumulators
def get_squared_str(self):
return self._squared_acc_str
def get_linear_str(self):
return self._linear_acc_str
def test_ftrl_optimizer(self):
init_program = framework.Program()
program = framework.Program()
block = program.global_block()
mul_x = block.create_parameter(
dtype="float32",
shape=[5, 10],
lod_level=0,
name="mul.x",
optimize_attr={'learning_rate': 1.1})
mul_y = block.create_var(
dtype="float32", shape=[10, 8], lod_level=0, name="mul.y")
mul_out = block.create_var(
dtype="float32", shape=[5, 8], lod_level=0, name="mul.out")
block.append_op(
type="mul",
inputs={"X": mul_x,
"Y": mul_y},
outputs={"Out": mul_out},
attrs={"x_num_col_dims": 1})
mean_out = block.create_var(
dtype="float32", shape=[1], lod_level=0, name="mean.out")
block.append_op(
type="mean", inputs={"X": mul_out}, outputs={"Out": mean_out})
learning_rate = 0.01
ftrl_optimizer = self.MockFtrl(
learning_rate=learning_rate, l1=0.0, l2=0.0, lr_power=-0.5)
params_grads = append_backward(mean_out)
self.assertEqual(len(params_grads), 1)
self.assertEqual(len(ftrl_optimizer.get_accumulators()), 0)
opts = ftrl_optimizer.create_optimization_pass(params_grads, mul_out,
init_program)
self.assertEqual(len(opts), 3)
self.assertEqual([op.type for op in opts],
["fill_constant", "elementwise_mul", "ftrl"])
# Check accumulators
accumulators = ftrl_optimizer.get_accumulators()
self.assertEqual(len(accumulators), 2)
self.assertTrue(ftrl_optimizer.get_squared_str() in accumulators)
self.assertTrue(ftrl_optimizer.get_linear_str() in accumulators)
squared_acc = accumulators[ftrl_optimizer.get_squared_str()]
linear_acc = accumulators[ftrl_optimizer.get_linear_str()]
self.assertEqual(len(squared_acc), 1)
self.assertEqual(len(linear_acc), 1)
self.assertTrue(mul_x.name in squared_acc)
self.assertTrue(mul_x.name in linear_acc)
# Check init_program
init_ops = init_program.global_block().ops
self.assertEqual(len(init_ops), 3)
self.assertEqual(init_ops[0].type, "fill_constant")
self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate)
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/tests/unittests/test_sum_mkldnn_op.py 0 → 100644
浏览文件 @ 9942a304
# 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.
import unittest
from test_sum_op import TestSumOp
class TestMKLDNN(TestSumOp):
def init_kernel_type(self):
self.use_mkldnn = True
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/tests/unittests/test_sum_op.py
浏览文件 @ 9942a304
......@@ -20,12 +20,15 @@ from op_test import OpTest
class TestSumOp(OpTest):
def setUp(self):
self.op_type = "sum"
self.use_mkldnn = False
self.init_kernel_type()
x0 = np.random.random((3, 4)).astype('float32')
x1 = np.random.random((3, 4)).astype('float32')
x2 = np.random.random((3, 4)).astype('float32')
self.inputs = {"X": [("x0", x0), ("x1", x1), ("x2", x2)]}
y = x0 + x1 + x2
self.outputs = {'Out': y}
self.attrs = {'use_mkldnn': self.use_mkldnn}
def test_check_output(self):
self.check_output()
......@@ -33,6 +36,9 @@ class TestSumOp(OpTest):
def test_check_grad(self):
self.check_grad(['x0'], 'Out')
def init_kernel_type(self):
pass
if __name__ == "__main__":
unittest.main()
python/paddle/fluid/transpiler/distribute_transpiler.py
浏览文件 @ 9942a304
......@@ -44,7 +44,7 @@ import numpy as np
from ps_dispatcher import RoundRobin, HashName, PSDispatcher
from .. import core, framework
from ..framework import Program, default_main_program, \
default_startup_program, \
default_startup_program, Block, \
Variable, Parameter, grad_var_name
from details import *
......@@ -471,7 +471,7 @@ class DistributeTranspiler:
self._append_pserver_ops(block, op, endpoint, grad_to_block_id,
self.origin_program, merged_var)
else:
self._append_pserver_non_opt_ops(block, op, endpoint)
self._append_pserver_non_opt_ops(block, op)
def __op_have_grad_input__(op):
for varname in op.input_arg_names:
......@@ -479,13 +479,39 @@ class DistributeTranspiler:
return varname
return ""
def __clone_lr_op_sub_block__(op, program, new_block):
if not op.has_attr('sub_block'):
return
origin_block_desc = op.attr('sub_block')
origin_block = self.origin_program.block(origin_block_desc.id)
assert isinstance(origin_block, Block)
# we put the new sub block to new block to follow the block
# hierarchy of the original blocks
new_sub_block = program.create_block(new_block.idx)
# clone vars
for var in origin_block.vars:
new_sub_block.clone_variable(var)
# clone ops
for op in origin_block.ops:
self._clone_lr_op(program, new_sub_block, op)
# clone sub_block of op
__clone_lr_op_sub_block__(op, program, new_sub_block)
# reset the block of op
op.set_attr('sub_block', new_sub_block)
# append lr decay ops to the child block if exists
lr_ops = self._get_lr_ops()
if len(lr_ops) > 0:
lr_decay_block = pserver_program.create_block(
pserver_program.num_blocks - 1)
for _, op in enumerate(lr_ops):
self._append_pserver_non_opt_ops(lr_decay_block, op, endpoint)
self._append_pserver_non_opt_ops(lr_decay_block, op)
# append sub blocks to pserver_program in lr_decay_op
__clone_lr_op_sub_block__(op, pserver_program, lr_decay_block)
# append op to the current block
grad_to_block_id = []
......@@ -798,7 +824,8 @@ class DistributeTranspiler:
table_opt_block.append_op(
type="sum",
inputs={"X": pserver_side_table_grad_list},
outputs={"Out": [grad_var]})
outputs={"Out": [grad_var]},
attrs={"use_mkldnn": False})
else:
# in async_mode, for table gradient, it also need to be splited to each parameter server
origin_grad_name = grad_var.name
......@@ -1030,7 +1057,8 @@ class DistributeTranspiler:
optimize_block.append_op(
type="sum",
inputs={"X": vars2merge},
outputs={"Out": merged_var})
outputs={"Out": merged_var},
attrs={"use_mkldnn": False})
# TODO(panyx0718): What if it's SELECTED_ROWS.
if not merged_var.type == core.VarDesc.VarType.SELECTED_ROWS:
optimize_block.append_op(
......@@ -1116,7 +1144,29 @@ class DistributeTranspiler:
break
return grad_block
def _append_pserver_non_opt_ops(self, optimize_block, opt_op, endpoint):
def _clone_lr_op(self, program, block, op):
inputs = self._get_input_map_from_op(
self.origin_program.global_block().vars, op)
for key, varlist in inputs.iteritems():
if not isinstance(varlist, list):
varlist = [varlist]
for var in varlist:
if var not in program.global_block().vars:
block.clone_variable(var)
outputs = self._get_output_map_from_op(
self.origin_program.global_block().vars, op)
for key, varlist in outputs.iteritems():
if not isinstance(varlist, list):
varlist = [varlist]
for var in varlist:
if var not in program.global_block().vars:
block.clone_variable(var)
block.append_op(
type=op.type, inputs=inputs, outputs=outputs, attrs=op.attrs)
def _append_pserver_non_opt_ops(self, optimize_block, opt_op):
program = optimize_block.program
# Append the ops for parameters that do not need to be optimized/updated
inputs = self._get_input_map_from_op(
......
python/paddle/fluid/transpiler/inference_transpiler.py
浏览文件 @ 9942a304
......@@ -19,16 +19,30 @@ from ..executor import global_scope
class InferenceTranspiler:
'''
Convert the fluid program to optimized inference program.
There are several optimizations, only fuse batch normalization is supported now.
Examples:
.. code-block:: python
# As InferenceTranspiler will modify the original program,
# please clone before use it.
inference_transpiler_program = program.clone()
t = fluid.InferenceTranspiler()
t.transpile(inference_transpiler_program, place)
'''
def transpile(self, program, place, scope=None):
'''
Transpile the program. Support only fuse batch normalization now.
:param program: program to transpile
:type program: Program
:param place: inference place
:type place: Place
:param scope: inference scope
:type scope: Scope or None
Run the transpiler.
Args:
program (Program): program to transpile
place (Place): inference place
scope (Scope|None): inference Scope
'''
if not isinstance(program, Program):
raise TypeError("program should be as Program type")
......@@ -49,36 +63,43 @@ class InferenceTranspiler:
can be integrated with them. Doing so will give us a forward acceleration,
especially in environments like mobile or embedded.
For input X:
- Conv process: X = input * W + bias
- Batch norm process: X' = (X - mean) / std
- Scale Process: Y = a * X' + b
For input :math:`X`:
- Conv process: :math:`X = input * W + bias`
- Batch norm process: :math:`X' = (X - mean) / std`
- Scale Process: :math:`Y = a * X' + b`
After fuse into one operation:
Y = (input * W + bias - mean) / std * a + b
= input * a * W / std + ((bias - mean) / std * a + b)
.. math::
Y &= (input * W + bias - mean) / std * a + b \\\\
&= input * a * W / std + ((bias - mean) / std * a + b)
The operator transformation is:
- before:
- conv->batch_norm->any_other_op (bias == 0)
- conv->elementwise_add->batch_norm->any_other_op (bias != 0)
- after:
- conv->elementwise_add->any_other_op
The transpile stages are:
1. insert elementwise_add op when bias == 0.
2. fuse the batch_norm's parameters to conv and elementwise_add operators.
3. remove batch_norm ops which are not used in any other ops.
4. adjust the input of any_other_op to be the output of elementwise_add operator.
5. remove unused variables.
:param program: program to transpile
:type program: Program
:param place: inference place
:type place: Place
:param scope: inference scope
:type scope: Scope
Args:
program (Program): program to transpile
place (Place): inference place
scope (Scope): inference Scope
'''
self.scope = scope
self.place = place
......
tools/print_signatures.py 0 → 100644
浏览文件 @ 9942a304
# 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.
"""
Print all signature of a python module in alphabet order.
Usage:
./print_signature "paddle.fluid" > signature.txt
"""
import importlib
import inspect
import collections
import sys
import pydoc
member_dict = collections.OrderedDict()
def visit_member(parent_name, member):
cur_name = ".".join([parent_name, member.__name__])
if inspect.isclass(member):
for name, value in inspect.getmembers(member):
if hasattr(value, '__name__') and (not name.startswith("_") or
name == "__init__"):
visit_member(cur_name, value)
elif callable(member):
try:
member_dict[cur_name] = inspect.getargspec(member)
except TypeError: # special for PyBind method
member_dict[cur_name] = " ".join([
line.strip() for line in pydoc.render_doc(member).split('\n')
if "->" in line
])
else:
raise RuntimeError("Unsupported generate signature of member, type {0}".
format(str(type(member))))
def visit_all_module(mod):
for member_name in (
name
for name in (mod.__all__ if hasattr(mod, "__all__") else dir(mod))
if not name.startswith("_")):
instance = getattr(mod, member_name, None)
if instance is None:
continue
if inspect.ismodule(instance):
visit_all_module(instance)
else:
visit_member(mod.__name__, instance)
visit_all_module(importlib.import_module(sys.argv[1]))
for name in member_dict:
print name, member_dict[name]
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