Skip to content

P
Paddle

机器未来 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 9386ac0a 编写于 作者: Y Yibing Liu
浏览文件
操作

Enhance cuda code & unittest for argsort_op

上级 9c69fdf5 c318aa5f
展开全部 隐藏空白更改
内联 并排
Showing with 6433 addition and 1607 deletion
README.md
浏览文件 @ 9386ac0a
......@@ -4,7 +4,6 @@
[![Build Status](https://travis-ci.org/PaddlePaddle/Paddle.svg?branch=develop)](https://travis-ci.org/PaddlePaddle/Paddle)
[![Documentation Status](https://img.shields.io/badge/docs-latest-brightgreen.svg?style=flat)](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/index_en.html)
[![Documentation Status](https://img.shields.io/badge/中文文档-最新-brightgreen.svg)](http://www.paddlepaddle.org/docs/develop/documentation/zh/getstarted/index_cn.html)
[![Coverage Status](https://coveralls.io/repos/github/PaddlePaddle/Paddle/badge.svg?branch=develop)](https://coveralls.io/github/PaddlePaddle/Paddle?branch=develop)
[![Release](https://img.shields.io/github/release/PaddlePaddle/Paddle.svg)](https://github.com/PaddlePaddle/Paddle/releases)
[![License](https://img.shields.io/badge/license-Apache%202-blue.svg)](LICENSE)
......
benchmark/fluid/args.py
浏览文件 @ 9386ac0a
......@@ -122,5 +122,9 @@ def parse_args():
type=str,
default="",
help='Directory that contains all the training recordio files.')
parser.add_argument(
'--use_inference_transpiler',
action='store_true',
help='If set, uses inference transpiler to optimize the program.')
args = parser.parse_args()
return args
benchmark/fluid/fluid_benchmark.py 100644 → 100755
浏览文件 @ 9386ac0a
......@@ -131,6 +131,11 @@ def train(avg_loss, infer_prog, optimizer, train_reader, test_reader, batch_acc,
exe = fluid.Executor(place)
exe.run(startup_prog)
# Use inference_transpiler to speedup
if args.use_inference_transpiler:
t = fluid.InferenceTranspiler()
t.transpile(infer_prog, place)
if not args.use_reader_op:
feed_var_list = [
var for var in train_prog.global_block().vars.itervalues()
......
cmake/external/anakin.cmake
浏览文件 @ 9386ac0a
......@@ -26,13 +26,15 @@ function(fetch_include_recursively root_dir)
endforeach()
endfunction()
# download library
message(STATUS "Download Anakin library from ${ANAKIN_LIBRARY_URL}")
execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_INSTALL_DIR}")
execute_process(COMMAND bash -c "rm -rf ${ANAKIN_INSTALL_DIR}/*")
execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; wget -q ${ANAKIN_LIBRARY_URL}")
execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_INSTALL_DIR}")
execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; tar xzf anakin_release_simple.tar.gz")
if (NOT EXISTS "${ANAKIN_INSTALL_DIR}")
# download library
message(STATUS "Download Anakin library from ${ANAKIN_LIBRARY_URL}")
execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_INSTALL_DIR}")
execute_process(COMMAND bash -c "rm -rf ${ANAKIN_INSTALL_DIR}/*")
execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; wget -q ${ANAKIN_LIBRARY_URL}")
execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_INSTALL_DIR}")
execute_process(COMMAND bash -c "cd ${ANAKIN_INSTALL_DIR}; tar xzf anakin_release_simple.tar.gz")
endif()
if (WITH_ANAKIN)
message(STATUS "Anakin for inference is enabled")
......
cmake/external/grpc.cmake
浏览文件 @ 9386ac0a
......@@ -40,12 +40,12 @@ ExternalProject_Add(
# NOTE(wuyi):
# this package is generated by following steps:
# 1. git clone -b v1.8.x https://github.com/grpc/grpc.git
# 2. submodule update --init
# 2. git submodule update --init
# 3. keep only zlib, cares, protobuf, boringssl under "third_party",
# checkout and clean other dirs under third_party
# 4. remove .git, and package the directory.
URL "http://paddlepaddledeps.bj.bcebos.com/grpc-v1.8.x.tar.gz"
URL_MD5 "c9c58ee7d0e8929a63155af6a2ecdbd0"
URL "http://paddlepaddledeps.bj.bcebos.com/grpc-v1.10.x.tar.gz"
URL_MD5 "1f268a2aff6759839dccd256adcc91cf"
PREFIX ${GRPC_SOURCES_DIR}
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
......
cmake/external/openblas.cmake
浏览文件 @ 9386ac0a
......@@ -114,7 +114,12 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c)
FILE(WRITE ${dummyfile} "const char *dummy_cblas = \"${dummyfile}\";")
ADD_LIBRARY(cblas STATIC ${dummyfile})
TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES})
IF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
TARGET_LINK_LIBRARIES(cblas dynload_mklml)
ELSE()
TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES})
ENDIF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
IF(NOT ${CBLAS_FOUND})
ADD_DEPENDENCIES(cblas extern_openblas)
......
cmake/generic.cmake
浏览文件 @ 9386ac0a
......@@ -96,6 +96,20 @@ if(NOT APPLE AND NOT ANDROID)
set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -pthread -ldl -lrt")
endif(NOT APPLE AND NOT ANDROID)
set_property(GLOBAL PROPERTY FLUID_MODULES "")
# find all fluid modules is used for paddle fluid static library
# for building inference libs
function(find_fluid_modules TARGET_NAME)
get_filename_component(__target_path ${TARGET_NAME} ABSOLUTE)
string(REGEX REPLACE "^${PADDLE_SOURCE_DIR}/" "" __target_path ${__target_path})
string(FIND "${__target_path}" "fluid" pos)
if(pos GREATER 1)
get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
set(fluid_modules ${fluid_modules} ${TARGET_NAME})
set_property(GLOBAL PROPERTY FLUID_MODULES "${fluid_modules}")
endif()
endfunction(find_fluid_modules)
function(merge_static_libs TARGET_NAME)
set(libs ${ARGN})
list(REMOVE_DUPLICATES libs)
......@@ -195,6 +209,15 @@ function(cc_library TARGET_NAME)
list(REMOVE_ITEM cc_library_DEPS warpctc)
add_dependencies(${TARGET_NAME} warpctc)
endif()
# Only deps libmklml.so, not link
if("${cc_library_DEPS};" MATCHES "mklml;")
list(REMOVE_ITEM cc_library_DEPS mklml)
if(NOT "${TARGET_NAME}" MATCHES "dynload_mklml")
list(APPEND cc_library_DEPS dynload_mklml)
endif()
add_dependencies(${TARGET_NAME} mklml)
target_link_libraries(${TARGET_NAME} "-L${MKLML_LIB_DIR} -liomp5 -Wl,--as-needed")
endif()
target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
endif()
......@@ -241,6 +264,7 @@ function(cc_test TARGET_NAME)
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
if (${cc_test_SERIAL})
set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
endif()
endif()
endfunction(cc_test)
......@@ -305,6 +329,7 @@ function(nv_test TARGET_NAME)
add_test(${TARGET_NAME} ${TARGET_NAME})
if (nv_test_SERIAL)
set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
endif()
endif()
endfunction(nv_test)
......@@ -552,7 +577,7 @@ function(py_test TARGET_NAME)
set(multiValueArgs SRCS DEPS ARGS ENVS)
cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
add_test(NAME ${TARGET_NAME}
COMMAND env PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
COMMAND env FLAGS_init_allocated_mem=true PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
endif()
......
cmake/inference_lib.cmake
浏览文件 @ 9386ac0a
......@@ -12,19 +12,6 @@
# See the License for the specific language governing permissions and
# limitations under the License.
set_property(GLOBAL PROPERTY FLUID_MODULES "")
# find all fluid modules is used for paddle fluid static library
function(find_fluid_modules TARGET_NAME)
get_filename_component(__target_path ${TARGET_NAME} ABSOLUTE)
string(REGEX REPLACE "^${PADDLE_SOURCE_DIR}/" "" __target_path ${__target_path})
string(FIND "${__target_path}" "fluid" pos)
if(pos GREATER 1)
get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
set(fluid_modules ${fluid_modules} ${TARGET_NAME})
set_property(GLOBAL PROPERTY FLUID_MODULES "${fluid_modules}")
endif()
endfunction(find_fluid_modules)
# make package for paddle fluid shared and static library
function(copy TARGET)
set(options "")
......@@ -149,21 +136,33 @@ copy(memory_lib
DSTS ${dst_dir}/${module} ${dst_dir}/${module}/detail
)
set(module "inference")
copy(inference_lib DEPS paddle_fluid_shared paddle_fluid
SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
DSTS ${dst_dir}/${module} ${dst_dir}/${module}
)
set(inference_deps paddle_fluid_shared paddle_fluid)
if(WITH_CONTRIB)
set(contrib_dst_dir "${FLUID_INSTALL_DIR}/contrib/inference")
copy(contrib_inference_lib DEPS paddle_inference_api
message(STATUS "installing contrib")
set(contrib_dst_dir "${FLUID_INSTALL_DIR}/contrib/inference")
if (WITH_ANAKIN)
copy(contrib_anakin_inference_lib DEPS paddle_inference_api inference_anakin_api
SRCS
${PADDLE_BINARY_DIR}/paddle/contrib/inference/libinference_anakin_api* # compiled anakin api
${PADDLE_BINARY_DIR}/third_party/install/anakin/*.tar.gz # anakin release
DSTS ${contrib_dst_dir}/anakin ${contrib_dst_dir}/anakin)
list(APPEND inference_deps contrib_anakin_inference_lib)
endif()
copy(contrib_inference_lib DEPS paddle_inference_api paddle_inference_api_shared
SRCS ${PADDLE_SOURCE_DIR}/paddle/contrib/inference/paddle_inference_api.h
${PADDLE_BINARY_DIR}/paddle/contrib/inference/libpaddle_inference_api.*
DSTS ${contrib_dst_dir} ${contrib_dst_dir}
)
${PADDLE_BINARY_DIR}/paddle/contrib/inference/libpaddle_inference_api*
DSTS ${contrib_dst_dir} ${contrib_dst_dir})
list(APPEND inference_deps contrib_inference_lib)
endif()
set(module "inference")
copy(inference_lib DEPS ${inference_deps}
SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
DSTS ${dst_dir}/${module} ${dst_dir}/${module}
)
set(module "platform")
copy(platform_lib DEPS profiler_py_proto
SRCS ${src_dir}/${module}/*.h ${src_dir}/${module}/dynload/*.h ${src_dir}/${module}/details/*.h
......
doc/about/about_us.rst 0 → 100644
浏览文件 @ 9386ac0a
=========
关于我们
=========
什么是PaddlePaddle
--------------------
- PaddlePaddle是百度自主研发并开源的深度学习框架，它能够让开发者和企业安全、快速地实现自己的AI想法
- 项目团队汇聚了全球顶级的深度学习科学家，致力于为开发者和企业提供最好的深度学习研发体验
- 框架具有易学、易用、安全、高效四大特性，是最适合中国开发者和企业的深度学习工具
PaddlePaddle的技术特色
-------------------------
- 新一代深度学习框架： PaddlePaddle是基于“深度学习编程语言”的新一代深度学习框架，在保证性能的同时，极大的提升了框架对模型的表达能力，能够描述任意潜在可能出现的模型
- 对大规模计算更加友好：经过百度内多种大规模计算业务的打磨，PaddlePaddle在分布式计算上表现优异，基于EDL技术能够节约大量计算资源，同时也能支持大规模稀疏模型的训练
- 提供可视化的深度学习：通过Visual DL可以帮助开发者方便的观测训练整体趋势、数据样本质量和中间结果、参数分布和变化趋势、以及模型的结构，帮助开发者更便捷的完成编程过程
提供基于PaddlePaddle的教育体系
--------------------------------
- 深度学习课程：百度与中国市场顶级的教育、培训机构共同开发了深度学习精品课程以及学习教材，帮助开发者从零掌握深度学习
- 深度学习实训：对于目的是科研和学习的用户，PaddlePaddle提供了无需安装、线上运行的开发环境，并提供算法、算力、数据支持
- 线下培训：提供丰富、高质量的线下教育活动，如青年教师培训、线下实战营、沙龙等多种形式的培训和交流
提供基于PaddlePaddle的AI服务
------------------------------
- EadyDL：可以帮助零算法基础的企业快速完成一个深度学习任务，只需少量的数据即可得到优质的模型
- AI市场：提供标准化的AI 能力、产品的交易机制，帮助企业快速找到所需，有效开展AI业务
- 深度学习竞赛： PaddlePaddle汇聚顶尖深度学习开发者，企业可以发布自己的商业问题，通过竞赛方式快速找到最优的解决方案
你对PaddlePaddle有任何的问题都可以通过以下方式联系到我们
-----------------------------------------------------------
- 学习/使用问题：可以在 `PaddlePaddle开源社区 <https://github.com/PaddlePaddle/Paddle/issues>`_，以及 `PaddlePaddle中文社区 <http://ai.baidu.com/forum/topic/list/168>`_ 向我们反馈
- 对PaddlePaddle框架发展的建议：可发送邮件至Paddle-better@baidu.com
我们期待与你一起打造世界顶级深度学习框架，共同推动AI技术的进步
PaddlePaddle团队
doc/fluid/api/evaluator.rst → doc/fluid/api/average.rst
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
=========
evaluator
=========
=============
fluid.average
=============
.. _api_fluid_average_WeightedAverage:
WeightedAverage
---------------
.. autoclass:: paddle.fluid.average.WeightedAverage
:members:
:noindex:
doc/fluid/api/backward.rst 0 → 100644
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
==============
fluid.backward
==============
.. _api_fluid_backward_append_backward:
append_backward
---------------
.. autofunction:: paddle.fluid.backward.append_backward
:noindex:
.. _api_fluid_backward_calc_gradient:
calc_gradient
-------------
.. autofunction:: paddle.fluid.backward.calc_gradient
:noindex:
doc/fluid/api/clip.rst
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
====
clip
====
==========
fluid.clip
==========
.. _api_fluid_clip_ErrorClipByValue:
ErrorClipByValue
----------------
......@@ -12,6 +14,8 @@ ErrorClipByValue
:members:
:noindex:
.. _api_fluid_clip_GradientClipByValue:
GradientClipByValue
-------------------
......@@ -19,6 +23,8 @@ GradientClipByValue
:members:
:noindex:
.. _api_fluid_clip_GradientClipByNorm:
GradientClipByNorm
------------------
......@@ -26,6 +32,8 @@ GradientClipByNorm
:members:
:noindex:
.. _api_fluid_clip_GradientClipByGlobalNorm:
GradientClipByGlobalNorm
------------------------
......@@ -33,15 +41,3 @@ GradientClipByGlobalNorm
:members:
:noindex:
append_gradient_clip_ops
------------------------
.. autofunction:: paddle.fluid.clip.append_gradient_clip_ops
:noindex:
error_clip_callback
-------------------
.. autofunction:: paddle.fluid.clip.error_clip_callback
:noindex:
doc/fluid/api/data.rst 已删除 100644 → 0
浏览文件 @ 9c69fdf5
==================================
Data Reader Interface and DataSets
==================================
.. toctree::
:maxdepth: 1
data/data_reader.rst
data/image.rst
data/dataset.rst
doc/fluid/api/data_feeder.rst
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
===========
data_feeder
===========
=================
fluid.data_feeder
=================
.. _api_fluid_data_feeder_DataFeeder:
DataFeeder
----------
......
doc/fluid/api/executor.rst
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
========
executor
========
==============
fluid.executor
==============
.. _api_fluid_executor_Executor:
Executor
--------
......@@ -12,24 +14,32 @@ Executor
:members:
:noindex:
.. _api_fluid_executor_global_scope:
global_scope
------------
.. autofunction:: paddle.fluid.executor.global_scope
:noindex:
.. _api_fluid_executor_scope_guard:
scope_guard
-----------
.. autofunction:: paddle.fluid.executor.scope_guard
:noindex:
switch_scope
------------
.. _api_fluid_executor__switch_scope:
_switch_scope
-------------
.. autofunction:: paddle.fluid.executor.switch_scope
.. autofunction:: paddle.fluid.executor._switch_scope
:noindex:
.. _api_fluid_executor_fetch_var:
fetch_var
---------
......
doc/fluid/api/fluid.rst 0 → 100644
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
=====
fluid
=====
.. _api_fluid_Block:
Block
-----
.. autoclass:: paddle.fluid.Block
:members:
:noindex:
.. _api_fluid_Variable:
Variable
--------
.. autoclass:: paddle.fluid.Variable
:members:
:noindex:
.. _api_fluid_Program:
Program
-------
.. autoclass:: paddle.fluid.Program
:members:
:noindex:
.. _api_fluid_Operator:
Operator
--------
.. autoclass:: paddle.fluid.Operator
:members:
:noindex:
.. _api_fluid_default_startup_program:
default_startup_program
-----------------------
.. autofunction:: paddle.fluid.default_startup_program
:noindex:
.. _api_fluid_default_main_program:
default_main_program
--------------------
.. autofunction:: paddle.fluid.default_main_program
:noindex:
.. _api_fluid_program_guard:
program_guard
-------------
.. autofunction:: paddle.fluid.program_guard
:noindex:
.. _api_fluid_get_var:
get_var
-------
.. autofunction:: paddle.fluid.get_var
:noindex:
.. _api_fluid_Executor:
Executor
--------
.. autoclass:: paddle.fluid.Executor
:members:
:noindex:
.. _api_fluid_global_scope:
global_scope
------------
.. autofunction:: paddle.fluid.global_scope
:noindex:
.. _api_fluid_scope_guard:
scope_guard
-----------
.. autofunction:: paddle.fluid.scope_guard
:noindex:
.. _api_fluid__switch_scope:
_switch_scope
-------------
.. autofunction:: paddle.fluid._switch_scope
:noindex:
.. _api_fluid_fetch_var:
fetch_var
---------
.. autofunction:: paddle.fluid.fetch_var
:noindex:
.. _api_fluid_Go:
Go
--
.. autoclass:: paddle.fluid.Go
:members:
:noindex:
.. _api_fluid_make_channel:
make_channel
------------
.. autofunction:: paddle.fluid.make_channel
:noindex:
.. _api_fluid_channel_send:
channel_send
------------
.. autofunction:: paddle.fluid.channel_send
:noindex:
.. _api_fluid_channel_recv:
channel_recv
------------
.. autofunction:: paddle.fluid.channel_recv
:noindex:
.. _api_fluid_channel_close:
channel_close
-------------
.. autofunction:: paddle.fluid.channel_close
:noindex:
.. _api_fluid_Select:
Select
------
.. autoclass:: paddle.fluid.Select
:members:
:noindex:
.. _api_fluid_Trainer:
Trainer
-------
.. autoclass:: paddle.fluid.Trainer
:members:
:noindex:
.. _api_fluid_BeginEpochEvent:
BeginEpochEvent
---------------
.. autoclass:: paddle.fluid.BeginEpochEvent
:members:
:noindex:
.. _api_fluid_EndEpochEvent:
EndEpochEvent
-------------
.. autoclass:: paddle.fluid.EndEpochEvent
:members:
:noindex:
.. _api_fluid_BeginStepEvent:
BeginStepEvent
--------------
.. autoclass:: paddle.fluid.BeginStepEvent
:members:
:noindex:
.. _api_fluid_EndStepEvent:
EndStepEvent
------------
.. autoclass:: paddle.fluid.EndStepEvent
:members:
:noindex:
.. _api_fluid_CheckpointConfig:
CheckpointConfig
----------------
.. autoclass:: paddle.fluid.CheckpointConfig
:members:
:noindex:
.. _api_fluid_Inferencer:
Inferencer
----------
.. autoclass:: paddle.fluid.Inferencer
:members:
:noindex:
.. _api_fluid_DistributeTranspiler:
DistributeTranspiler
--------------------
.. autoclass:: paddle.fluid.DistributeTranspiler
:members:
:noindex:
.. _api_fluid_memory_optimize:
memory_optimize
---------------
.. autofunction:: paddle.fluid.memory_optimize
:noindex:
.. _api_fluid_release_memory:
release_memory
--------------
.. autofunction:: paddle.fluid.release_memory
:noindex:
.. _api_fluid_ParallelExecutor:
ParallelExecutor
----------------
.. autoclass:: paddle.fluid.ParallelExecutor
:members:
:noindex:
.. _api_fluid_ExecutionStrategy:
ExecutionStrategy
-----------------
.. autoclass:: paddle.fluid.ExecutionStrategy
:members:
:noindex:
.. _api_fluid_BuildStrategy:
BuildStrategy
-------------
.. autoclass:: paddle.fluid.BuildStrategy
:members:
:noindex:
.. _api_fluid_create_lod_tensor:
create_lod_tensor
-----------------
.. autofunction:: paddle.fluid.create_lod_tensor
:noindex:
.. _api_fluid_create_random_int_lodtensor:
create_random_int_lodtensor
---------------------------
.. autofunction:: paddle.fluid.create_random_int_lodtensor
:noindex:
.. _api_fluid_LoDTensor:
LoDTensor
---------
.. autoclass:: paddle.fluid.LoDTensor
:members:
:noindex:
.. _api_fluid_CPUPlace:
CPUPlace
--------
.. autoclass:: paddle.fluid.CPUPlace
:members:
:noindex:
.. _api_fluid_CUDAPlace:
CUDAPlace
---------
.. autoclass:: paddle.fluid.CUDAPlace
:members:
:noindex:
.. _api_fluid_CUDAPinnedPlace:
CUDAPinnedPlace
---------------
.. autoclass:: paddle.fluid.CUDAPinnedPlace
:members:
:noindex:
.. _api_fluid_Tensor:
Tensor
------
.. autoclass:: paddle.fluid.Tensor
:members:
:noindex:
.. _api_fluid_ParamAttr:
ParamAttr
---------
.. autoclass:: paddle.fluid.ParamAttr
:members:
:noindex:
.. _api_fluid_WeightNormParamAttr:
WeightNormParamAttr
-------------------
.. autoclass:: paddle.fluid.WeightNormParamAttr
:members:
:noindex:
.. _api_fluid_DataFeeder:
DataFeeder
----------
.. autoclass:: paddle.fluid.DataFeeder
:members:
:noindex:
.. _api_fluid_Scope:
Scope
-----
.. autoclass:: paddle.fluid.Scope
:members:
:noindex:
doc/fluid/api/gen_doc.py
浏览文件 @ 9386ac0a
......@@ -29,19 +29,27 @@ def parse_arg():
class DocGenerator(object):
def __init__(self, module_name, stream=sys.stdout):
def __init__(self, module_name=None, stream=sys.stdout):
if module_name == "":
module_name = None
self.stream = stream
self.module_name = module_name
if not hasattr(fluid, module_name):
raise ValueError("Cannot find fluid.{0}".format(module_name))
if module_name is None:
self.module_name = "fluid"
else:
self.module = getattr(fluid, module_name)
self.module_name = "fluid." + module_name
if module_name is None:
self.module = fluid
else:
if not hasattr(fluid, module_name):
raise ValueError("Cannot find fluid.{0}".format(module_name))
else:
self.module = getattr(fluid, module_name)
self.stream.write('''.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
''')
self._print_header_(module_name, dot='=', is_title=True)
self._print_header_(self.module_name, dot='=', is_title=True)
def print_submodule(self, submodule_name):
submodule = getattr(self.module, submodule_name)
......@@ -60,25 +68,29 @@ class DocGenerator(object):
self._print_header_(name, dot='=', is_title=False)
def print_item(self, name):
item = getattr(self.module, name)
item = getattr(self.module, name, None)
if item is None:
return
if isinstance(item, types.TypeType):
self.print_class(name)
elif isinstance(item, types.FunctionType):
self.print_method(name)
else:
raise RuntimeError("Unsupported item {0}".format(name))
pass
def print_class(self, name):
self._print_ref_(name)
self._print_header_(name, dot='-', is_title=False)
self.stream.write('''.. autoclass:: paddle.fluid.{0}.{1}
self.stream.write('''.. autoclass:: paddle.{0}.{1}
:members:
:noindex:
'''.format(self.module_name, name))
def print_method(self, name):
self._print_ref_(name)
self._print_header_(name, dot='-', is_title=False)
self.stream.write('''.. autofunction:: paddle.fluid.{0}.{1}
self.stream.write('''.. autofunction:: paddle.{0}.{1}
:noindex:
'''.format(self.module_name, name))
......@@ -94,6 +106,10 @@ class DocGenerator(object):
self.stream.write('\n')
self.stream.write('\n')
def _print_ref_(self, name):
self.stream.write(".. _api_{0}_{1}:\n\n".format("_".join(
self.module_name.split(".")), name))
def main():
args = parse_arg()
......
doc/fluid/api/gen_doc.sh
浏览文件 @ 9386ac0a
#!/bin/bash
python gen_doc.py layers --submodules control_flow device io nn ops tensor detection learning_rate_scheduler metric > layers.rst
python gen_doc.py layers --submodules control_flow device io nn ops tensor learning_rate_scheduler detection metric_op tensor > layers.rst
for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer transpiler
for module in data_feeder clip metrics executor initializer io nets optimizer param_attr profiler regularizer transpiler recordio_writer backward average profiler
do
python gen_doc.py ${module} > ${module}.rst
done
python gen_doc.py "" > fluid.rst
doc/fluid/api/index_en.rst
浏览文件 @ 9386ac0a
======================
Fluid
======================
=============
API Reference
=============
.. toctree::
:maxdepth: 1
fluid.rst
layers.rst
data_feeder.rst
executor.rst
......@@ -18,3 +19,8 @@ Fluid
regularizer.rst
io.rst
data.rst
transpiler.rst
recordio_writer.rst
backward.rst
average.rst
profiler.rst
doc/fluid/api/initializer.rst
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
===========
initializer
===========
=================
fluid.initializer
=================
.. _api_fluid_initializer_Constant:
Constant
--------
......@@ -12,6 +14,8 @@ Constant
:members:
:noindex:
.. _api_fluid_initializer_Uniform:
Uniform
-------
......@@ -19,6 +23,8 @@ Uniform
:members:
:noindex:
.. _api_fluid_initializer_Normal:
Normal
------
......@@ -26,6 +32,8 @@ Normal
:members:
:noindex:
.. _api_fluid_initializer_Xavier:
Xavier
------
......@@ -33,6 +41,8 @@ Xavier
:members:
:noindex:
.. _api_fluid_initializer_Bilinear:
Bilinear
--------
......@@ -40,18 +50,33 @@ Bilinear
:members:
:noindex:
.. _api_fluid_initializer_MSRA:
MSRA
----
.. autoclass:: paddle.fluid.initializer.MSRA
:members:
:noindex:
.. _api_fluid_initializer_force_init_on_cpu:
force_init_on_cpu
-----------------
.. autofunction:: paddle.fluid.initializer.force_init_on_cpu
:noindex:
.. _api_fluid_initializer_init_on_cpu:
init_on_cpu
-----------
.. autofunction:: paddle.fluid.initializer.init_on_cpu
:noindex:
.. _api_fluid_initializer_ConstantInitializer:
ConstantInitializer
-------------------
......@@ -59,6 +84,8 @@ ConstantInitializer
:members:
:noindex:
.. _api_fluid_initializer_UniformInitializer:
UniformInitializer
------------------
......@@ -66,6 +93,8 @@ UniformInitializer
:members:
:noindex:
.. _api_fluid_initializer_NormalInitializer:
NormalInitializer
-----------------
......@@ -73,6 +102,8 @@ NormalInitializer
:members:
:noindex:
.. _api_fluid_initializer_XavierInitializer:
XavierInitializer
-----------------
......@@ -80,6 +111,8 @@ XavierInitializer
:members:
:noindex:
.. _api_fluid_initializer_BilinearInitializer:
BilinearInitializer
-------------------
......@@ -87,3 +120,12 @@ BilinearInitializer
:members:
:noindex:
.. _api_fluid_initializer_MSRAInitializer:
MSRAInitializer
---------------
.. autoclass:: paddle.fluid.initializer.MSRAInitializer
:members:
:noindex:
doc/fluid/api/io.rst
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
==
io
==
========
fluid.io
========
.. _api_fluid_io_save_vars:
save_vars
---------
......@@ -11,84 +13,112 @@ save_vars
.. autofunction:: paddle.fluid.io.save_vars
:noindex:
.. _api_fluid_io_save_params:
save_params
-----------
.. autofunction:: paddle.fluid.io.save_params
:noindex:
.. _api_fluid_io_save_persistables:
save_persistables
-----------------
.. autofunction:: paddle.fluid.io.save_persistables
:noindex:
.. _api_fluid_io_load_vars:
load_vars
---------
.. autofunction:: paddle.fluid.io.load_vars
:noindex:
.. _api_fluid_io_load_params:
load_params
-----------
.. autofunction:: paddle.fluid.io.load_params
:noindex:
.. _api_fluid_io_load_persistables:
load_persistables
-----------------
.. autofunction:: paddle.fluid.io.load_persistables
:noindex:
.. _api_fluid_io_save_inference_model:
save_inference_model
--------------------
.. autofunction:: paddle.fluid.io.save_inference_model
:noindex:
.. _api_fluid_io_load_inference_model:
load_inference_model
--------------------
.. autofunction:: paddle.fluid.io.load_inference_model
:noindex:
.. _api_fluid_io_get_inference_program:
get_inference_program
---------------------
.. autofunction:: paddle.fluid.io.get_inference_program
:noindex:
.. _api_fluid_io_save_checkpoint:
save_checkpoint
---------------
.. autofunction:: paddle.fluid.io.save_checkpoint
:noindex:
.. _api_fluid_io_load_checkpoint:
load_checkpoint
---------------
.. autofunction:: paddle.fluid.io.load_checkpoint
:noindex:
.. _api_fluid_io_clean_checkpoint:
clean_checkpoint
----------------
.. autofunction:: paddle.fluid.io.clean_checkpoint
:noindex:
.. _api_fluid_io_load_persist_vars_without_grad:
load_persist_vars_without_grad
------------------------------
.. autofunction:: paddle.fluid.io.load_persist_vars_without_grad
:noindex:
.. _api_fluid_io_save_persist_vars_without_grad:
save_persist_vars_without_grad
------------------------------
.. autofunction:: paddle.fluid.io.save_persist_vars_without_grad
:noindex:
.. _api_fluid_io_get_latest_checkpoint_serial:
get_latest_checkpoint_serial
----------------------------
......
doc/fluid/api/layers.rst
浏览文件 @ 9386ac0a
此差异已折叠。
doc/fluid/api/metrics.rst
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
=======
metrics
=======
=============
fluid.metrics
=============
.. _api_fluid_metrics_MetricBase:
MetricBase
----------
......@@ -12,6 +14,8 @@ MetricBase
:members:
:noindex:
.. _api_fluid_metrics_CompositeMetric:
CompositeMetric
---------------
......@@ -19,6 +23,26 @@ CompositeMetric
:members:
:noindex:
.. _api_fluid_metrics_Precision:
Precision
---------
.. autoclass:: paddle.fluid.metrics.Precision
:members:
:noindex:
.. _api_fluid_metrics_Recall:
Recall
------
.. autoclass:: paddle.fluid.metrics.Recall
:members:
:noindex:
.. _api_fluid_metrics_Accuracy:
Accuracy
--------
......@@ -26,6 +50,8 @@ Accuracy
:members:
:noindex:
.. _api_fluid_metrics_ChunkEvaluator:
ChunkEvaluator
--------------
......@@ -33,6 +59,8 @@ ChunkEvaluator
:members:
:noindex:
.. _api_fluid_metrics_EditDistance:
EditDistance
------------
......@@ -40,6 +68,8 @@ EditDistance
:members:
:noindex:
.. _api_fluid_metrics_DetectionMAP:
DetectionMAP
------------
......@@ -47,6 +77,8 @@ DetectionMAP
:members:
:noindex:
.. _api_fluid_metrics_Auc:
Auc
---
......
doc/fluid/api/nets.rst
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
====
nets
====
==========
fluid.nets
==========
.. _api_fluid_nets_simple_img_conv_pool:
simple_img_conv_pool
--------------------
......@@ -11,18 +13,24 @@ simple_img_conv_pool
.. autofunction:: paddle.fluid.nets.simple_img_conv_pool
:noindex:
.. _api_fluid_nets_sequence_conv_pool:
sequence_conv_pool
------------------
.. autofunction:: paddle.fluid.nets.sequence_conv_pool
:noindex:
.. _api_fluid_nets_glu:
glu
---
.. autofunction:: paddle.fluid.nets.glu
:noindex:
.. _api_fluid_nets_scaled_dot_product_attention:
scaled_dot_product_attention
----------------------------
......
doc/fluid/api/optimizer.rst
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
=========
optimizer
=========
===============
fluid.optimizer
===============
.. _api_fluid_optimizer_SGD:
SGD
---
......@@ -12,6 +14,8 @@ SGD
:members:
:noindex:
.. _api_fluid_optimizer_Momentum:
Momentum
--------
......@@ -19,6 +23,8 @@ Momentum
:members:
:noindex:
.. _api_fluid_optimizer_Adagrad:
Adagrad
-------
......@@ -26,6 +32,8 @@ Adagrad
:members:
:noindex:
.. _api_fluid_optimizer_Adam:
Adam
----
......@@ -33,6 +41,8 @@ Adam
:members:
:noindex:
.. _api_fluid_optimizer_Adamax:
Adamax
------
......@@ -40,6 +50,8 @@ Adamax
:members:
:noindex:
.. _api_fluid_optimizer_DecayedAdagrad:
DecayedAdagrad
--------------
......@@ -47,6 +59,17 @@ DecayedAdagrad
:members:
:noindex:
.. _api_fluid_optimizer_Ftrl:
Ftrl
----
.. autoclass:: paddle.fluid.optimizer.Ftrl
:members:
:noindex:
.. _api_fluid_optimizer_SGDOptimizer:
SGDOptimizer
------------
......@@ -54,6 +77,8 @@ SGDOptimizer
:members:
:noindex:
.. _api_fluid_optimizer_MomentumOptimizer:
MomentumOptimizer
-----------------
......@@ -61,6 +86,8 @@ MomentumOptimizer
:members:
:noindex:
.. _api_fluid_optimizer_AdagradOptimizer:
AdagradOptimizer
----------------
......@@ -68,6 +95,8 @@ AdagradOptimizer
:members:
:noindex:
.. _api_fluid_optimizer_AdamOptimizer:
AdamOptimizer
-------------
......@@ -75,6 +104,8 @@ AdamOptimizer
:members:
:noindex:
.. _api_fluid_optimizer_AdamaxOptimizer:
AdamaxOptimizer
---------------
......@@ -82,6 +113,8 @@ AdamaxOptimizer
:members:
:noindex:
.. _api_fluid_optimizer_DecayedAdagradOptimizer:
DecayedAdagradOptimizer
-----------------------
......@@ -89,6 +122,8 @@ DecayedAdagradOptimizer
:members:
:noindex:
.. _api_fluid_optimizer_RMSPropOptimizer:
RMSPropOptimizer
----------------
......@@ -96,6 +131,17 @@ RMSPropOptimizer
:members:
:noindex:
.. _api_fluid_optimizer_FtrlOptimizer:
FtrlOptimizer
-------------
.. autoclass:: paddle.fluid.optimizer.FtrlOptimizer
:members:
:noindex:
.. _api_fluid_optimizer_Adadelta:
Adadelta
--------
......@@ -103,6 +149,8 @@ Adadelta
:members:
:noindex:
.. _api_fluid_optimizer_ModelAverage:
ModelAverage
------------
......@@ -110,6 +158,8 @@ ModelAverage
:members:
:noindex:
.. _api_fluid_optimizer_Optimizer:
Optimizer
---------
......@@ -117,3 +167,12 @@ Optimizer
:members:
:noindex:
.. _api_fluid_optimizer_RMSPropOptimizer:
RMSPropOptimizer
----------------
.. autoclass:: paddle.fluid.optimizer.RMSPropOptimizer
:members:
:noindex:
doc/fluid/api/param_attr.rst
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
==========
param_attr
==========
================
fluid.param_attr
================
.. _api_fluid_param_attr_ParamAttr:
ParamAttr
---------
......@@ -12,6 +14,8 @@ ParamAttr
:members:
:noindex:
.. _api_fluid_param_attr_WeightNormParamAttr:
WeightNormParamAttr
-------------------
......
doc/fluid/api/profiler.rst
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
========
profiler
========
==============
fluid.profiler
==============
.. _api_fluid_profiler_cuda_profiler:
cuda_profiler
-------------
......@@ -11,24 +13,32 @@ cuda_profiler
.. autofunction:: paddle.fluid.profiler.cuda_profiler
:noindex:
.. _api_fluid_profiler_reset_profiler:
reset_profiler
--------------
.. autofunction:: paddle.fluid.profiler.reset_profiler
:noindex:
.. _api_fluid_profiler_profiler:
profiler
--------
.. autofunction:: paddle.fluid.profiler.profiler
:noindex:
.. _api_fluid_profiler_start_profiler:
start_profiler
--------------
.. autofunction:: paddle.fluid.profiler.start_profiler
:noindex:
.. _api_fluid_profiler_stop_profiler:
stop_profiler
-------------
......
doc/fluid/api/recordio_writer.rst 0 → 100644
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
=====================
fluid.recordio_writer
=====================
.. _api_fluid_recordio_writer_convert_reader_to_recordio_file:
convert_reader_to_recordio_file
-------------------------------
.. autofunction:: paddle.fluid.recordio_writer.convert_reader_to_recordio_file
:noindex:
.. _api_fluid_recordio_writer_convert_reader_to_recordio_files:
convert_reader_to_recordio_files
--------------------------------
.. autofunction:: paddle.fluid.recordio_writer.convert_reader_to_recordio_files
:noindex:
doc/fluid/api/regularizer.rst
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
===========
regularizer
===========
=================
fluid.regularizer
=================
.. _api_fluid_regularizer_append_regularization_ops:
append_regularization_ops
-------------------------
......@@ -11,12 +13,7 @@ append_regularization_ops
.. autofunction:: paddle.fluid.regularizer.append_regularization_ops
:noindex:
WeightDecayRegularizer
----------------------
.. autoclass:: paddle.fluid.regularizer.WeightDecayRegularizer
:members:
:noindex:
.. _api_fluid_regularizer_L1Decay:
L1Decay
-------
......@@ -25,6 +22,8 @@ L1Decay
:members:
:noindex:
.. _api_fluid_regularizer_L2Decay:
L2Decay
-------
......@@ -32,6 +31,8 @@ L2Decay
:members:
:noindex:
.. _api_fluid_regularizer_L1DecayRegularizer:
L1DecayRegularizer
------------------
......@@ -39,6 +40,8 @@ L1DecayRegularizer
:members:
:noindex:
.. _api_fluid_regularizer_L2DecayRegularizer:
L2DecayRegularizer
------------------
......
doc/fluid/api/transpiler.rst
浏览文件 @ 9386ac0a
.. THIS FILE IS GENERATED BY `gen_doc.{py|sh}`
!DO NOT EDIT THIS FILE MANUALLY!
==========
transpiler
==========
================
fluid.transpiler
================
.. _api_fluid_transpiler_DistributeTranspiler:
DistributeTranspiler
--------------------
......@@ -12,12 +14,7 @@ DistributeTranspiler
:members:
:noindex:
InferenceTranspiler
-------------------
.. autoclass:: paddle.fluid.transpiler.InferenceTranspiler
:members:
:noindex:
.. _api_fluid_transpiler_memory_optimize:
memory_optimize
---------------
......@@ -25,12 +22,16 @@ memory_optimize
.. autofunction:: paddle.fluid.transpiler.memory_optimize
:noindex:
.. _api_fluid_transpiler_release_memory:
release_memory
--------------
.. autofunction:: paddle.fluid.transpiler.release_memory
:noindex:
.. _api_fluid_transpiler_HashName:
HashName
--------
......@@ -38,9 +39,12 @@ HashName
:members:
:noindex:
.. _api_fluid_transpiler_RoundRobin:
RoundRobin
----------
.. autoclass:: paddle.fluid.transpiler.RoundRobin
:members:
:noindex:
doc/fluid/design/concepts/lod_tensor.md
浏览文件 @ 9386ac0a
......@@ -173,6 +173,7 @@ are transformed into offsets of elements/words as follows:
## Slicing of LoD Tensors
When we use the above 2-level LoD Tensor as the input to a nested-RNN, we need to retrieve certain sequences. Here we define the sequence identified by branch <i,j,...> as the **<i,j,...>-slice**.
For example, the <2>-slice of above example is
......@@ -189,3 +190,22 @@ and the <2,0>-slice of above slice is
10 12
||
```
## Length Representation vs Offset Representation
The offset representation is an implementation-oriented decision and it makes understanding the idea behind LoDTensor difficult.
Hence, we encapsulate this implementation detail in C++ and expose the original length representation in our Python API.
Specifically, we call this length representation `recursive_sequence_lengths` and users can use the following code to set or get the `recursive_sequence_lengths` of a LoDTensor in Python:
```Python
# length representation of lod called recursive_sequence_lengths
recursive_seq_lens = [[3, 1, 2], [2, 2, 1, 3, 1, 2]]
# Create a LoDTensor that has the above recursive_sequence_lengths info.
# This recursive_sequence_lengths will be converted to an offset representation of LoD in the C++ implementation under the hood.
tensor = fluid.LoDTensor(lod)
# Set/Change the recursive_sequence_lengths info of LoDTensor
tensor.set_recursive_sequence_lengths([[3, 1, 2]])
# Get the recursive_sequence_lengths info of a LoDTensor (the offset-based LoD representation stored in C++ will be converted
# back to length-based recursive_sequence_lengths), new_recursive_seq_lens = [[3, 1, 2]]
new_recursive_seq_lens = tensor.recursive_sequence_lengths()
```
doc/fluid/design/concepts/python_data_feeding.md 0 → 100644
浏览文件 @ 9386ac0a
# Python Data Feeding
In the former implementation of Paddle Fluid, there are two ways to feed data:
- Use `reader_op` in backend C++ side. This method only supports data feeding from recordio files and random data generators, but supports many kinds of `decorated_readers`. For examples, `double_buffer_reader` uses two threads to achieve better performance: one for time-consuming I/O operations, and the other for `Executor::Run()`. See [C++ Data Feeding](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/fluid/design/concepts/cpp_data_feeding.md) for details.
- Feed data directly using `DataFeeder.feed()` in Python codes. It is more flexible than the first way. Many kinds of preprocessing steps can be performed before feeding using Python or any other languages, instead of adding many uncommon `operators` in C++ side. But this method is less efficient: the program cannot read the next mini-batch data before `Executor::Run()` ends. Moreover, `decorated_readers` such as `double_buffer_reader` cannot be used for better performance.
In this document, we design a Python Data Feeding process combining the efficiency of the first way and the flexibility of the second way. A data queue `LoDTensorBlockingQueue` is designed to be shared by the Python and C++ side, while `LoDTensorArray` is pushed into the queue in Python side and `reader_op` in C++ side reads out the data from the queue.
## Design of LoDTensorBlockingQueue
`LoDTensorBlockingQueue` is a blocking queue with a fixed `capacity` and accepts `std::vector<framework::LoDTensor>` with shapes indicated by `dims`. Since `LoDTensorBlockingQueue` must be constructed using `capacity` and `dims`, it cannot be a `Variable` type. Therefore, a `LoDTensorBlockingQueueHolder` is designed to defer construction of `LoDTensorBlockingQueue`.
```C++
class LoDTensorBlockingQueueHolder;
class LoDTensorBlockingQueue {
friend class LoDTensorBlockingQueueHolder;
private:
// `LoDTensorBlockingQueue` can only be constructed by
// `LoDTensorBlockingQueueHolder::InitOnce()`
LoDTensorBlockingQueue(size_t capacity, const std::vector<framework::DDim>& dims);
public:
size_t Size() const { return queue_.Size(); } // Get the current size of the queue
size_t Cap() const { return queue_.Cap(); }// Get the capacity of the queue
void Close() { return queue_.Close(); }
bool IsClosed() const { return queue_.IsClosed(); }
// Block if Size() == Cap()
// Return false only when queue_.IsClosed() == true
bool Push(const std::vector<framework::LoDTensor> &lod_tensor_vec);
// Block if Size() == 0.
// *Success == false when queue_.IsClosed() == true
std::vector<framework::LoDTensor> Pop(bool *success = nullptr);
private:
// Use reader::BlockingQueue as the inner data structure
BlockingQueue<std::vector<framework::LoDTensor>> queue_;
std::vector<framework::DDim> dims_;
};
class LoDTensorBlockingQueueHolder {
public:
// Call the constructor of `LoDTensorBlockingQueue` to create queue_
// `InitOnce` can only called once, otherwise an exception would raise
void InitOnce(size_t capacity, const std::vector<framework::DDim>& dims) {
PADDLE_ENFORCE(queue_ == nullptr);
queue_.reset(new LoDTensorBlockingQueue(capacity, dims));
}
const std::shared_ptr<LoDTensorBlockingQueue>& GetQueue() const { return queue_; }
private:
std::shared_ptr<LoDTensorBlockingQueue> queue_;
};
```
There are some major things that must be concerned:
- `LoDTensorBlockingQueueHolder` should be a `Variable` in global scope, so that `reader_op` can find it when reading data.
- A `Variable` of `LoDTensorBlockingQueueHolder` but not `VarDesc` must be created in Python code before `Executor::Run()` so that `Executor::Run()` can get the feeding data when it is called.
- `Create_reader_op` should accept the name of the `LoDTensorBlockingQueueHolder` variable as an input.
## Release of the GIL in pybind
`Pybind11::gil_scoped_release` is used to release GIL (Global Interpreter Lock) when `LoDTensorBlockingQueue::Push()` or `Executor::Run()` method are invoked in Python side, making `LoDTensorBlockingQueue::Push()` and `Executor::Run()` run in parallel.
## Design of PyReader
`PyReader` is a reader which holds a `LoDTensorBlockingQueue` object.
```C++
class PyReader : public ReaderBase {
public:
explicit PyReader(const std::shared_ptr<LoDTensorBlockingQueue>& queue);
void ReadNext(std::vector<framework::LoDTensor>* out) override {
bool success;
*out = queue_->Pop(&success);
if (!success) out->clear();
}
void ReInit() override { return; }
private:
std::shared_ptr<LoDTensorBlockingQueue> queue_;
};
```
## Design of CreatePyReaderOp
`CreatePyReaderOp` is used to create the `PyReader` object. It requires an input `blocking_queue` which indicates the name of the `LoDTensorBlockingQueueHolder` variable.
```C++
class CreatePyReaderOp : public framework::OperatorBase {
public:
using framework::OperatorBase::OperatorBase;
private:
void RunImpl(const framework::Scope& scope,
const platform::Place& dev_place) const override {
auto* out = scope.FindVar(Output("Out"))
->template GetMutable<framework::ReaderHolder>();
if (out->Get() != nullptr) return;
const std::string& queue_name = Input("blocking_queue");
auto* queue_holder_var = scope.FindVar(queue_name);
PADDLE_ENFORCE(queue_holder_var != nullptr);
auto* queue_holder = queue_holder_var
->template GetMutable<framework::LoDTensorBlockingQueueHolder>();
out->Reset(new PyReader(queue_holder->GetQueue()));
}
};
```
## Design of Python codes
The design of Python codes are as follows. First, we construct a variable of `LoDTensorBlockingQueueHolder` and init it with given parameters, returning the `LoDTensorBlockingQueue` object after initialization. After that, a layer of `CreatePyReaderOp` is constructed and accepts the name of the `LoDTensorBlockingQueueHolder` variable. The `LoDTensorBlockingQueue` object and result of the layer are both returned.
```Python
def py_reader(capacity, shapes):
queue_name = unique_name.generate("lod_tensor_blocking_queue")
var = global_scope().var(feeder_name) # create LoDTensorBlockingQueueHolder Variable
feed_queue = core.init_lod_tensor_blocking_queue(var, capacity, shapes) # init the queue
out = create_var()
create_py_reader_op_with_queue_name(
inputs={'blocking_queue': queue_name},
outputs={'Out':[out]})
return out, feed_queue
```
doc/fluid/design/multi_devices/kernel_selection.md
浏览文件 @ 9386ac0a
......@@ -74,10 +74,10 @@ void OperatorWithKernel::Run(
auto kernel_type_for_var = this->GetKernelTypeForVar(...);
if (kernel_type_for_var.place_ != expected_kernel_key.place_) {
auto* trans_var = new_scope.Var(var_name);
auto* out = DataTransform(expected_kernel_key,
auto* out = TransformData(expected_kernel_key,
kernel_type_for_var,
*tensor_in);
CopyVariableWithTensor(...);
SetTensorToVariable(...);
}
}
......
doc/fluid/howto/optimization/host_memory_profiling_cn.md
浏览文件 @ 9386ac0a
## 堆内存分析和优化
# 堆内存分析和优化
计算机程序都可能有内存泄漏的风险。**内存泄漏**一般是由于程序在堆(heap)上分配了内存而没有释放，随着程序的运行占用的内存越来越大，一方面会影响程序的稳定性，可能让运行速度越来越慢，或者造成oom，甚至会影响运行程序的机器的稳定性，造成宕机。
......@@ -20,11 +20,11 @@ Paddle也提供了基于gperftool的[CPU性能分析教程](https://github.com/P
对于堆内存的分析，主要用到thread-caching malloc和heap-profiling using tcmalloc。
## 使用流程
#### 环境
## 环境
本教程基于paddle提供的Docker开发环境paddlepaddle/paddle:latest-dev，基于Ubuntu 16.04.4 LTS环境。
#### 使用流程
## 使用流程
- 安装google-perftools
......
doc/fluid/howto/optimization/timeline_cn.md 0 → 100644
浏览文件 @ 9386ac0a
# 如何使用timeline工具做性能分析
1. 在训练的主循环外加上`with profiler.profiler(...)`。运行之后，代码会在`/tmp/profile`目录下生成一个profile的记录文件。
**提示：**
请不要在timeline记录信息时运行太多次迭代，因为timeline中的记录数量和迭代次数是成正比的。
```python
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=[])
...
```
1. 运行`python paddle/tools/timeline.py`来处理`/tmp/profile`，这个程序默认会生成一个`/tmp/timeline`文件，你也可以用命令行参数来修改这个路径，请参考[timeline.py](https://github.com/PaddlePaddle/Paddle/blob/develop/tools/timeline.py)。
1. 打开chrome浏览器，访问<chrome://tracing/>，用`load`按钮来加载生成的`timeline`文件。
![chrome tracing](./tracing.jpeg)
1. 结果如下图所示，可以放到来查看timetime的细节信息。
![chrome timeline](./timeline.jpeg)
paddle/contrib/inference/CMakeLists.txt
浏览文件 @ 9386ac0a
......@@ -19,6 +19,9 @@ endif(APPLE)
set(inference_deps paddle_inference_api paddle_fluid_api)
if(WITH_GPU AND TENSORRT_FOUND)
set(inference_deps ${inference_deps} paddle_inference_tensorrt_subgraph_engine)
endif()
function(inference_api_test TARGET_NAME)
if (WITH_TESTING)
......@@ -43,6 +46,10 @@ cc_library(paddle_inference_api
SRCS paddle_inference_api.cc paddle_inference_api_impl.cc
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
cc_library(paddle_inference_api_shared SHARED
SRCS paddle_inference_api.cc paddle_inference_api_impl.cc
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
cc_test(test_paddle_inference_api
SRCS test_paddle_inference_api.cc
DEPS paddle_inference_api)
......@@ -50,17 +57,30 @@ cc_test(test_paddle_inference_api
inference_api_test(test_paddle_inference_api_impl
ARGS test_word2vec test_image_classification)
if (WITH_ANAKIN AND WITH_TESTING) # only needed in CI
if(WITH_GPU AND TENSORRT_FOUND)
cc_library(paddle_inference_tensorrt_subgraph_engine
SRCS paddle_inference_api_tensorrt_subgraph_engine.cc
DEPS paddle_inference_api analysis tensorrt_engine paddle_inference_api paddle_fluid_api)
inference_api_test(test_paddle_inference_api_tensorrt_subgraph_engine ARGS test_word2vec)
endif()
if (WITH_ANAKIN) # only needed in CI
# Due to Anakin do not have official library releases and the versions of protobuf and cuda do not match Paddle's,
# so anakin library will not be merged to our official inference library. To use anakin prediction API, one need to
# compile the libinference_anakin_api.a and compile with anakin.so.
nv_library(inference_anakin_api SHARED SRCS paddle_inference_api.cc paddle_inference_api_anakin_engine.cc)
nv_library(inference_anakin_api SRCS paddle_inference_api.cc paddle_inference_api_anakin_engine.cc)
nv_library(inference_anakin_api_shared SHARED SRCS paddle_inference_api.cc paddle_inference_api_anakin_engine.cc)
target_compile_options(inference_anakin_api BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
target_compile_options(inference_anakin_api_shared BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
target_link_libraries(inference_anakin_api anakin anakin_saber_common)
cc_test(inference_anakin_test SRCS paddle_inference_api_anakin_engine_tester.cc
target_link_libraries(inference_anakin_api_shared anakin anakin_saber_common)
if (WITH_TESTING)
cc_test(inference_anakin_test SRCS paddle_inference_api_anakin_engine_tester.cc
ARGS --model=${ANAKIN_INSTALL_DIR}/mobilenet_v2.anakin.bin
DEPS inference_anakin_api)
target_compile_options(inference_anakin_test BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
target_compile_options(inference_anakin_test BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
endif(WITH_TESTING)
endif()
if(WITH_TESTING)
......
paddle/contrib/inference/demo/CMakeLists.txt
浏览文件 @ 9386ac0a
......@@ -14,3 +14,48 @@
#
inference_api_test(simple_on_word2vec ARGS test_word2vec)
option(WITH_INFERENCE_DEMO "Compile with Inference demo" OFF)
if(NOT WITH_INFERENCE_DEMO)
return()
endif()
set(DEMO_INSTALL_DIR "${PADDLE_BINARY_DIR}/inference_demo")
set(URL_ROOT http://paddlemodels.bj.bcebos.com/inference-vis-demos%2F)
function(inference_download_test_demo TARGET)
if (NOT WITH_TESTING)
return()
endif()
set(options "")
set(oneValueArgs URL)
set(multiValueArgs SRCS)
cmake_parse_arguments(tests "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
set(test_dir "${DEMO_INSTALL_DIR}/${TARGET}")
message(STATUS "inference demo ${test_dir}")
if(NOT EXISTS "${test_dir}")
message(STATUS "Download ${TARGET} model from ${tests_URL}")
execute_process(COMMAND bash -c "mkdir -p ${test_dir}")
execute_process(COMMAND bash -c "cd ${test_dir}; wget -q ${tests_URL}")
execute_process(COMMAND bash -c "cd ${test_dir}; tar xzf *.tar.gz")
endif()
cc_test(${TARGET} SRCS "${tests_SRCS}"
DEPS paddle_inference_api paddle_fluid
ARGS --data=${test_dir}/data.txt
--modeldir=${test_dir}/model
--refer=${test_dir}/result.txt)
endfunction()
# disable mobilenet test
#inference_download_test_demo(mobilenet_inference_demo
# SRCS vis_demo.cc
# URL ${URL_ROOT}mobilenet.tar.gz)
inference_download_test_demo(se_resnext50_inference_demo
SRCS vis_demo.cc
URL ${URL_ROOT}se_resnext50.tar.gz)
inference_download_test_demo(ocr_inference_demo
SRCS vis_demo.cc
URL ${URL_ROOT}ocr.tar.gz)
paddle/contrib/inference/demo/README.md 0 → 100644
浏览文件 @ 9386ac0a
# Infernce Demos
Input data format:
- Each line contains a single record
- Each record's format is
```
<space splitted floats as data>\t<space splitted ints as shape>
```
Follow the C++ codes in `vis_demo.cc`.
## MobileNet
To execute the demo, simply run
```sh
./mobilenet_inference_demo --modeldir <model> --data <datafile>
```
## SE-ResNeXt-50
To execute the demo, simply run
```sh
./se_resnext50_inference_demo --modeldir <model> --data <datafile>
```
## OCR
To execute the demo, simply run
```sh
./ocr_inference_demo --modeldir <model> --data <datafile>
```
paddle/contrib/inference/demo/simple_on_word2vec.cc
浏览文件 @ 9386ac0a
......@@ -21,6 +21,7 @@ limitations under the License. */
#include <memory>
#include <thread>
#include "paddle/contrib/inference/paddle_inference_api.h"
namespace paddle {
namespace demo {
......
paddle/contrib/inference/demo/utils.h 0 → 100644
浏览文件 @ 9386ac0a
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include <string>
#include <vector>
#include "paddle/contrib/inference/paddle_inference_api.h"
namespace paddle {
namespace demo {
static void split(const std::string& str,
char sep,
std::vector<std::string>* pieces) {
pieces->clear();
if (str.empty()) {
return;
}
size_t pos = 0;
size_t next = str.find(sep, pos);
while (next != std::string::npos) {
pieces->push_back(str.substr(pos, next - pos));
pos = next + 1;
next = str.find(sep, pos);
}
if (!str.substr(pos).empty()) {
pieces->push_back(str.substr(pos));
}
}
/*
* Get a summary of a PaddleTensor content.
*/
static std::string SummaryTensor(const PaddleTensor& tensor) {
std::stringstream ss;
int num_elems = tensor.data.length() / PaddleDtypeSize(tensor.dtype);
ss << "data[:10]\t";
switch (tensor.dtype) {
case PaddleDType::INT64: {
for (int i = 0; i < std::min(num_elems, 10); i++) {
ss << static_cast<int64_t*>(tensor.data.data())[i] << " ";
}
break;
}
case PaddleDType::FLOAT32:
for (int i = 0; i < std::min(num_elems, 10); i++) {
ss << static_cast<float*>(tensor.data.data())[i] << " ";
}
break;
}
return ss.str();
}
} // namespace demo
} // namespace paddle
paddle/contrib/inference/demo/vis_demo.cc 0 → 100644
浏览文件 @ 9386ac0a
/* 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 contains demo for mobilenet, se-resnext50 and ocr.
*/
#include <gflags/gflags.h>
#include <glog/logging.h> // use glog instead of PADDLE_ENFORCE to avoid importing other paddle header files.
#include <gtest/gtest.h>
#include <fstream>
#include <iostream>
#include "paddle/contrib/inference/demo/utils.h"
#include "paddle/contrib/inference/paddle_inference_api.h"
#ifdef PADDLE_WITH_CUDA
DECLARE_double(fraction_of_gpu_memory_to_use);
#endif
namespace paddle {
namespace demo {
DEFINE_string(modeldir, "", "Directory of the inference model.");
DEFINE_string(refer, "", "path to reference result for comparison.");
DEFINE_string(
data,
"",
"path of data; each line is a record, format is "
"'<space splitted floats as data>\t<space splitted ints as shape'");
struct Record {
std::vector<float> data;
std::vector<int32_t> shape;
};
void split(const std::string& str, char sep, std::vector<std::string>* pieces);
Record ProcessALine(const std::string& line) {
LOG(INFO) << "process a line";
std::vector<std::string> columns;
split(line, '\t', &columns);
CHECK_EQ(columns.size(), 2UL)
<< "data format error, should be <data>\t<shape>";
Record record;
std::vector<std::string> data_strs;
split(columns[0], ' ', &data_strs);
for (auto& d : data_strs) {
record.data.push_back(std::stof(d));
}
std::vector<std::string> shape_strs;
split(columns[1], ' ', &shape_strs);
for (auto& s : shape_strs) {
record.shape.push_back(std::stoi(s));
}
LOG(INFO) << "data size " << record.data.size();
LOG(INFO) << "data shape size " << record.shape.size();
return record;
}
void CheckOutput(const std::string& referfile, const PaddleTensor& output) {
std::string line;
std::ifstream file(referfile);
std::getline(file, line);
auto refer = ProcessALine(line);
file.close();
size_t numel = output.data.length() / PaddleDtypeSize(output.dtype);
LOG(INFO) << "predictor output numel " << numel;
LOG(INFO) << "reference output numel " << refer.data.size();
EXPECT_EQ(numel, refer.data.size());
switch (output.dtype) {
case PaddleDType::INT64: {
for (size_t i = 0; i < numel; ++i) {
EXPECT_EQ(static_cast<int64_t*>(output.data.data())[i], refer.data[i]);
}
break;
}
case PaddleDType::FLOAT32:
for (size_t i = 0; i < numel; ++i) {
EXPECT_NEAR(
static_cast<float*>(output.data.data())[i], refer.data[i], 1e-5);
}
break;
}
}
/*
* Use the native fluid engine to inference the demo.
*/
void Main(bool use_gpu) {
NativeConfig config;
config.param_file = FLAGS_modeldir + "/__params__";
config.prog_file = FLAGS_modeldir + "/__model__";
config.use_gpu = use_gpu;
config.device = 0;
#ifdef PADDLE_WITH_CUDA
config.fraction_of_gpu_memory = FLAGS_fraction_of_gpu_memory_to_use;
#endif
LOG(INFO) << "init predictor";
auto predictor =
CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
LOG(INFO) << "begin to process data";
// Just a single batch of data.
std::string line;
std::ifstream file(FLAGS_data);
std::getline(file, line);
auto record = ProcessALine(line);
file.close();
// Inference.
PaddleTensor input{
.name = "xx",
.shape = record.shape,
.data = PaddleBuf(record.data.data(), record.data.size() * sizeof(float)),
.dtype = PaddleDType::FLOAT32};
LOG(INFO) << "run executor";
std::vector<PaddleTensor> output;
predictor->Run({input}, &output);
LOG(INFO) << "output.size " << output.size();
auto& tensor = output.front();
LOG(INFO) << "output: " << SummaryTensor(tensor);
// compare with reference result
CheckOutput(FLAGS_refer, tensor);
}
TEST(demo, vis_demo_cpu) { Main(false /*use_gpu*/); }
#ifdef PADDLE_WITH_CUDA
TEST(demo, vis_demo_gpu) { Main(true /*use_gpu*/); }
#endif
} // namespace demo
} // namespace paddle
paddle/contrib/inference/high_level_api.md
浏览文件 @ 9386ac0a
# Inference High-level APIs
This document describes the high-level inference APIs one can use to easily deploy a Paddle model for an application.
This document describes the high-level inference APIs, one can use them to deploy a Paddle model for an application quickly.
The APIs are described in `paddle_inference_api.h`, just one header file, and two libaries `libpaddle_fluid.so` and `libpaddle_fluid_api.so` are needed.
The APIs are described in `paddle_inference_api.h`, just one header file, and two libaries `libpaddle_fluid.so` and `libpaddle_fluid_api.so` are needed for a deployment.
## PaddleTensor
We provide the `PaddleTensor` data structure is to give a general tensor interface.
We provide the `PaddleTensor` data structure to give a general tensor interface.
The definition is
......@@ -17,18 +17,19 @@ struct PaddleTensor {
};
```
The data is stored in a continuous memory `PaddleBuf`, and tensor's data type is specified by a `PaddleDType`.
The `name` field is used to specify the name of input variable,
that is important when there are multiple inputs and need to distiuish which variable to set.
The data is stored in a continuous memory `PaddleBuf,` and a `PaddleDType` specifies tensor's data type.
The `name` field is used to specify the name of an input variable,
that is important when there are multiple inputs and need to distinguish which variable to set.
## engine
The inference APIs has two different underlying implementation, currently there are two valid engines:
The inference APIs has two different underlying engines
- the native engine, which is consists of the native operators and framework,
- the Anakin engine, which is a Anakin library embeded.
- the Anakin engine, which has an Anakin library embedded.
The native engine takes a native Paddle model as input, and supports any model that trained by Paddle,
but the Anakin engine can only take the Anakin model as input(user need to manully transform the format first) and currently not all Paddle models are supported.
the Anakin engine is faster for some model,
but it can only take the Anakin model as input(user need to transform the format first manually) and currently not all Paddle models are supported.
```c++
enum class PaddleEngineKind {
......@@ -38,10 +39,10 @@ enum class PaddleEngineKind {
```
## PaddlePredictor and how to create one
The main interface is `PaddlePredictor`, there are following methods
The main interface is `PaddlePredictor,` there are following methods
- `bool Run(const std::vector<PaddleTensor>& inputs, std::vector<PaddleTensor>* output_data)`
- take inputs and output `output_data`
- take inputs and output `output_data.`
- `Clone` to clone a predictor from an existing one, with model parameter shared.
There is a factory method to help create a predictor, and the user takes the ownership of this object.
......@@ -51,9 +52,9 @@ template <typename ConfigT, PaddleEngineKind engine = PaddleEngineKind::kNative>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor(const ConfigT& config);
```
By specifying the engine kind and config, one can get an specific implementation.
By specifying the engine kind and config, one can get a specific implementation.
## Reference
- [paddle_inference_api.h](./paddle_inference_api.h)
- [demos](./demo)
- [some demos](./demo)
paddle/contrib/inference/high_level_api_cn.md 0 → 100644
浏览文件 @ 9386ac0a
# Paddle 预测 API
为了更简单方便的预测部署，Fluid 提供了一套高层 API 用来隐藏底层不同的优化实现。
预测库包含:
- 头文件 `paddle_inference_api.h` 定义了所有的接口
- 库文件`libpaddle_fluid.so` 或 `libpaddle_fluid.a`
- 库文件 `libpaddle_inference_api.so` 或 `libpaddle_inference_api.a`
下面是详细的一些 API 概念介绍
## PaddleTensor
PaddleTensor 定义了预测最基本的输入输出的数据格式，其定义是
```c++
struct PaddleTensor {
std::string name; // variable name.
std::vector<int> shape;
PaddleBuf data; // blob of data.
PaddleDType dtype;
};
```
- `name` 用于指定输入数据对应的 模型中variable 的名字 （暂时没有用，但会在后续支持任意 target 时启用）
- `shape` 表示一个 Tensor 的 shape
- `data` 数据以连续内存的方式存储在`PaddleBuf` 中，`PaddleBuf` 可以接收外面的数据或者独立`malloc`内存，详细可以参考头文件中相关定义。
- `dtype` 表示 Tensor 的数据类型
## engine
高层 API 底层有多种优化实现，我们称之为 engine，目前有三种 engine
- 原生 engine，由 paddle 原生的 forward operator 组成，可以天然支持所有paddle 训练出的模型，
- Anakin engine，封装了 [Anakin](https://github.com/PaddlePaddle/Anakin) ，在某些模型上性能不错，但只能接受自带模型格式，无法支持所有 paddle 模型，
- TensorRT mixed engine，用子图的方式支持了 [TensorRT](https://developer.nvidia.com/tensorrt) ，支持所有paddle 模型，并自动切割部分计算子图到 TensorRT 上加速（WIP）
其实现为
```c++
enum class PaddleEngineKind {
kNative = 0, // Use the native Fluid facility.
kAnakin, // Use Anakin for inference.
kAutoMixedTensorRT // Automatically mixing TensorRT with the Fluid ops.
};
```
## 预测部署过程
总体上分为以下步骤
1. 用合适的配置创建 `PaddlePredictor`
2. 创建输入用的 `PaddleTensor`，传入到 `PaddlePredictor` 中
3. 获取输出的 `PaddleTensor` ，将结果取出
下面完整演示一个简单的模型，部分细节代码隐去
```c++
#include "paddle_inference_api.h"
// 创建一个 config，并修改相关设置
paddle::NativeConfig config;
config.model_dir = "xxx";
config.use_gpu = false;
// 创建一个原生的 PaddlePredictor
auto predictor =
paddle::CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
// 创建输入 tensor
int64_t data[4] = {1, 2, 3, 4};
paddle::PaddleTensor tensor{.name = "",
.shape = std::vector<int>({4, 1}),
.data = PaddleBuf(data, sizeof(data)),
.dtype = PaddleDType::INT64};
// 创建输出 tensor，输出 tensor 的内存可以复用
std::vector<paddle::PaddleTensor> outputs;
// 执行预测
CHECK(predictor->Run(slots, &outputs));
// 获取 outputs ...
```
编译时，联编 `libpaddle_fluid.a/.so` 和 `libpaddle_inference_api.a/.so` 便可。
## 详细代码参考
- [inference demos](./demo)
- [复杂单线程/多线程例子](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/contrib/inference/test_paddle_inference_api_impl.cc)
paddle/contrib/inference/paddle_inference_api.cc
浏览文件 @ 9386ac0a
......@@ -16,6 +16,19 @@ limitations under the License. */
namespace paddle {
int PaddleDtypeSize(PaddleDType dtype) {
switch (dtype) {
case PaddleDType::FLOAT32:
return sizeof(float);
case PaddleDType::INT64:
return sizeof(int64_t);
default:
//
assert(false);
return -1;
}
}
PaddleBuf::PaddleBuf(PaddleBuf&& other)
: data_(other.data_),
length_(other.length_),
......@@ -62,4 +75,4 @@ void PaddleBuf::Free() {
}
}
} // namespace paddle
\ No newline at end of file
} // namespace paddle
paddle/contrib/inference/paddle_inference_api.h
浏览文件 @ 9386ac0a
......@@ -15,7 +15,7 @@ limitations under the License. */
/*
* This file contains the definition of a simple Inference API for Paddle.
*
* ATTENTION: It requires some C++ features, for lower version C++ or C, we
* ATTENTION: It requires some C++11 features, for lower version C++ or C, we
* might release another API.
*/
......@@ -73,12 +73,12 @@ struct PaddleTensor {
};
enum class PaddleEngineKind {
kNative = 0, // Use the native Fluid facility.
kAnakin, // Use Anakin for inference.
kNative = 0, // Use the native Fluid facility.
kAnakin, // Use Anakin for inference.
kAutoMixedTensorRT, // Automatically mix Fluid with TensorRT.
// TODO(Superjomn) support following engines latter.
// kTensorRT, // Use TensorRT for inference.
// kAutoMixedAnakin, // Automatically mix Fluid with Anakin.
// kAutoMixedTensorRT, // Automatically mix Fluid with TensorRT.
};
/*
......@@ -130,6 +130,11 @@ struct AnakinConfig : public PaddlePredictor::Config {
int max_batch_size{-1};
};
struct TensorRTConfig : public NativeConfig {
// Determine whether a subgraph will be executed by TRT.
int min_subgraph_size{1};
};
// A factory to help create different predictors.
//
// FOR EXTENSION DEVELOPER:
......@@ -140,4 +145,7 @@ struct AnakinConfig : public PaddlePredictor::Config {
// Similarly, each engine kind should map to a unique predictor implementation.
template <typename ConfigT, PaddleEngineKind engine = PaddleEngineKind::kNative>
std::unique_ptr<PaddlePredictor> CreatePaddlePredictor(const ConfigT& config);
int PaddleDtypeSize(PaddleDType dtype);
} // namespace paddle
paddle/contrib/inference/paddle_inference_api_impl.cc
浏览文件 @ 9386ac0a
......@@ -89,6 +89,7 @@ bool NativePaddlePredictor::Init(
LOG(ERROR) << "fail to load inference model.";
return false;
}
ctx_ = executor_->Prepare(*inference_program_, 0);
executor_->CreateVariables(
*inference_program_, sub_scope_ ? sub_scope_ : scope_.get(), 0);
......@@ -119,6 +120,7 @@ bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
return false;
}
for (size_t i = 0; i < feed_target_names_.size(); ++i) {
VLOG(4) << "setting " << i << "-th target";
feed_targets[feed_target_names_[i]] = &feeds[i];
}
// get fetch variable
......@@ -130,14 +132,16 @@ bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
}
// Run the inference program
// if share variables, we need not create variables
VLOG(4) << "Run prepared context";
executor_->RunPreparedContext(
ctx_.get(),
sub_scope_ != nullptr ? sub_scope_ : scope_.get(),
&feed_targets,
&fetch_targets,
false /* don't create variable eatch time */);
VLOG(4) << "Finish prepared context";
if (!GetFetch(fetchs, output_data)) {
LOG(ERROR) << "fail to get fetchs";
LOG(ERROR) << "fail to get fetches";
return false;
}
VLOG(3) << "predict cost: " << timer.toc() << "ms";
......
paddle/contrib/inference/paddle_inference_api_impl.h
浏览文件 @ 9386ac0a
......@@ -44,7 +44,7 @@ class NativePaddlePredictor : public PaddlePredictor {
~NativePaddlePredictor() override;
private:
protected:
bool SetFeed(const std::vector<PaddleTensor> &input_datas,
std::vector<framework::LoDTensor> *feeds);
bool GetFetch(const std::vector<framework::LoDTensor> &fetchs,
......
paddle/contrib/inference/paddle_inference_api_tensorrt_subgraph_engine.cc 0 → 100644
浏览文件 @ 9386ac0a
// 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/inference/paddle_inference_api.h"
#include "paddle/contrib/inference/paddle_inference_api_impl.h"
#include "paddle/fluid/inference/analysis/analyzer.h"
#include "paddle/fluid/inference/utils/singleton.h"
namespace paddle {
using inference::analysis::Argument;
using inference::Singleton;
using inference::analysis::Analyzer;
using framework::proto::ProgramDesc;
class TensorRTSubgraphPredictor : public NativePaddlePredictor {
public:
explicit TensorRTSubgraphPredictor(const TensorRTConfig& config)
: NativePaddlePredictor(config), config_(config) {}
bool Init(const std::shared_ptr<framework::Scope>& parent_scope) {
VLOG(3) << "Predictor::init()";
if (config_.use_gpu) {
place_ = paddle::platform::CUDAPlace(config_.device);
} else {
place_ = paddle::platform::CPUPlace();
}
if (parent_scope) {
scope_ = parent_scope;
sub_scope_ = &(parent_scope->NewScope());
} else {
paddle::framework::InitDevices(false);
scope_.reset(new paddle::framework::Scope());
}
executor_.reset(new paddle::framework::Executor(place_));
// Initialize the inference program
if (!config_.model_dir.empty()) {
// Parameters are saved in separate files sited in
// the specified `dirname`.
inference_program_ = paddle::inference::Load(
executor_.get(), scope_.get(), config_.model_dir);
} else if (!config_.prog_file.empty() && !config_.param_file.empty()) {
// All parameters are saved in a single file.
// The file names should be consistent with that used
// in Python API `fluid.io.save_inference_model`.
inference_program_ = paddle::inference::Load(
executor_.get(), scope_.get(), config_.prog_file, config_.param_file);
} else {
LOG(ERROR) << "fail to load inference model.";
return false;
}
// Analyze inference_program
Argument argument;
argument.origin_program_desc.reset(
new ProgramDesc(*inference_program_->Proto()));
Singleton<Analyzer>::Global().Run(&argument);
CHECK(argument.transformed_program_desc);
VLOG(5) << "transformed program:\n"
<< argument.transformed_program_desc->SerializeAsString();
VLOG(5) << "to prepare executor";
*inference_program_->Proto() = *argument.transformed_program_desc;
ctx_ = executor_->Prepare(*inference_program_, 0);
VLOG(5) << "to create variables";
executor_->CreateVariables(
*inference_program_, sub_scope_ ? sub_scope_ : scope_.get(), 0);
// Get the feed_target_names and fetch_target_names
feed_target_names_ = inference_program_->GetFeedTargetNames();
fetch_target_names_ = inference_program_->GetFetchTargetNames();
return true;
}
private:
TensorRTConfig config_;
};
template <>
std::unique_ptr<PaddlePredictor>
CreatePaddlePredictor<TensorRTConfig, PaddleEngineKind::kAutoMixedTensorRT>(
const TensorRTConfig& config) {
VLOG(3) << "create TensorRTSubgraphPredictor";
if (config.use_gpu) {
// 1. GPU memeroy
PADDLE_ENFORCE_GT(
config.fraction_of_gpu_memory,
0.f,
"fraction_of_gpu_memory in the config should be set to range (0., 1.]");
PADDLE_ENFORCE_GE(config.device, 0, "Invalid device id %d", config.device);
std::vector<std::string> flags;
if (config.fraction_of_gpu_memory >= 0.0f ||
config.fraction_of_gpu_memory <= 0.95f) {
flags.push_back("dummpy");
std::string flag = "--fraction_of_gpu_memory_to_use=" +
std::to_string(config.fraction_of_gpu_memory);
flags.push_back(flag);
VLOG(3) << "set flag: " << flag;
framework::InitGflags(flags);
}
}
std::unique_ptr<PaddlePredictor> predictor(
new TensorRTSubgraphPredictor(config));
if (!dynamic_cast<TensorRTSubgraphPredictor*>(predictor.get())
->Init(nullptr)) {
return nullptr;
}
return std::move(predictor);
}
} // namespace paddle
paddle/contrib/inference/test_paddle_inference_api_tensorrt_subgraph_engine.cc 0 → 100644
浏览文件 @ 9386ac0a
// 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 <gflags/gflags.h>
#include <glog/logging.h>
#include <gtest/gtest.h>
#include "paddle/contrib/inference/paddle_inference_api.h"
namespace paddle {
DEFINE_string(dirname, "", "Directory of the inference model.");
void Main(bool use_gpu) {
//# 1. Create PaddlePredictor with a config.
TensorRTConfig config;
config.model_dir = FLAGS_dirname + "word2vec.inference.model";
config.use_gpu = use_gpu;
config.fraction_of_gpu_memory = 0.15;
config.device = 0;
auto predictor =
CreatePaddlePredictor<TensorRTConfig,
PaddleEngineKind::kAutoMixedTensorRT>(config);
for (int batch_id = 0; batch_id < 3; batch_id++) {
//# 2. Prepare input.
int64_t data[4] = {1, 2, 3, 4};
PaddleTensor tensor{.name = "",
.shape = std::vector<int>({4, 1}),
.data = PaddleBuf(data, sizeof(data)),
.dtype = PaddleDType::INT64};
// For simplicity, we set all the slots with the same data.
std::vector<PaddleTensor> slots(4, tensor);
//# 3. Run
std::vector<PaddleTensor> outputs;
CHECK(predictor->Run(slots, &outputs));
//# 4. Get output.
ASSERT_EQ(outputs.size(), 1UL);
LOG(INFO) << "output buffer size: " << outputs.front().data.length();
const size_t num_elements = outputs.front().data.length() / sizeof(float);
// The outputs' buffers are in CPU memory.
for (size_t i = 0; i < std::min(5UL, num_elements); i++) {
LOG(INFO) << static_cast<float*>(outputs.front().data.data())[i];
}
}
}
TEST(paddle_inference_api_tensorrt_subgraph_engine, main) { Main(true); }
} // namespace paddle
\ No newline at end of file
paddle/fluid/framework/data_layout_transform.cc
浏览文件 @ 9386ac0a
......@@ -147,10 +147,9 @@ void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
"Input tensor type is not supported: ", in.type().name());
memory::data_type out_type = in_type;
memory::format in_format =
in_tz.size() == 2 ? memory::format::nc : in.format();
memory::format out_format =
out_tz.size() == 2 ? memory::format::nc : ToMKLDNNFormat(out_layout);
auto in_format = platform::MKLDNNFormatForSize(in_tz.size(), in.format());
auto out_format =
platform::MKLDNNFormatForSize(in_tz.size(), ToMKLDNNFormat(out_layout));
void* in_data = GetDataFromTensor(in, in_type);
......
paddle/fluid/framework/data_layout_transform.h
浏览文件 @ 9386ac0a
......@@ -61,6 +61,7 @@ inline MKLDNNDataType ToMKLDNNDataType(const std::type_index type) {
if (iter != dict.end()) return iter->second;
return MKLDNNDataType::data_undef;
}
#endif
void TransDataLayoutFromMKLDNN(const OpKernelType& kernel_type_for_var,
......
paddle/fluid/framework/data_transform.cc
浏览文件 @ 9386ac0a
......@@ -18,17 +18,21 @@ limitations under the License. */
#include "paddle/fluid/framework/data_layout_transform.h"
#include "paddle/fluid/framework/data_type_transform.h"
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
namespace paddle {
namespace framework {
static void PassTensorData(Tensor* from, Tensor* to) {
static void PassTensorData(Tensor *from, Tensor *to) {
to->ShareDataWith(*from);
*from = Tensor();
}
void DataTransform(const OpKernelType& expected_kernel_type,
const OpKernelType& kernel_type_for_var,
const Tensor& input_tensor, Tensor* output_tensor) {
void TransformData(const OpKernelType &expected_kernel_type,
const OpKernelType &kernel_type_for_var,
const Tensor &input_tensor, Tensor *output_tensor) {
bool transformed = false;
Tensor in;
in.ShareDataWith(input_tensor);
......@@ -47,9 +51,13 @@ void DataTransform(const OpKernelType& expected_kernel_type,
#ifdef PADDLE_WITH_MKLDNN
// Case1 - transform from Non-MKLDNN OPKernel to MKLDNN OPKernel
// Just set layout/format. No real transform occur
auto out_format = platform::MKLDNNFormatForSize(in.dims().size(),
ToMKLDNNFormat(lin));
out.ShareDataWith(input_tensor);
out.set_layout(DataLayout::kMKLDNN);
out.set_format(ToMKLDNNFormat(lin));
out.set_format(out_format);
#endif
} else {
// Case2 - transfrom from MKLDNN OPKernel to Non-MKLDNN OPKernel
......@@ -85,17 +93,17 @@ void DataTransform(const OpKernelType& expected_kernel_type,
output_tensor->ShareDataWith(in);
}
void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor,
Variable* out_var) {
void SetTensorToVariable(const Variable &in_var, const Tensor &tensor,
Variable *out_var) {
if (in_var.IsType<LoDTensor>()) {
auto& in_lod_tensor = in_var.Get<LoDTensor>();
auto* tran_lod_tensor = out_var->GetMutable<LoDTensor>();
auto &in_lod_tensor = in_var.Get<LoDTensor>();
auto *tran_lod_tensor = out_var->GetMutable<LoDTensor>();
tran_lod_tensor->set_lod(in_lod_tensor.lod());
tran_lod_tensor->set_layout(in_lod_tensor.layout());
tran_lod_tensor->ShareDataWith(tensor);
} else if (in_var.IsType<SelectedRows>()) {
auto& in_selected_rows = in_var.Get<SelectedRows>();
auto* trans_selected_rows = out_var->GetMutable<SelectedRows>();
auto &in_selected_rows = in_var.Get<SelectedRows>();
auto *trans_selected_rows = out_var->GetMutable<SelectedRows>();
trans_selected_rows->set_height(in_selected_rows.height());
trans_selected_rows->set_rows(in_selected_rows.rows());
trans_selected_rows->mutable_value()->ShareDataWith(tensor);
......
paddle/fluid/framework/data_transform.h
浏览文件 @ 9386ac0a
......@@ -30,12 +30,15 @@ limitations under the License. */
namespace paddle {
namespace framework {
void DataTransform(const OpKernelType& expected_kernel_type,
const OpKernelType& kernel_type_for_var,
const Tensor& input_tensor, Tensor* out);
void CopyVariableWithTensor(const Variable& in_var, const Tensor& tensor,
Variable* out_var);
void TransformData(const OpKernelType &expected_kernel_type,
const OpKernelType &kernel_type_for_var,
const Tensor &input_tensor, Tensor *out);
/**
* Set OutVar from InVar, except the tensor is shared with `tensor`
*/
void SetTensorToVariable(const Variable &in_var, const Tensor &tensor,
Variable *out_var);
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/multi_devices_graph_builder.cc
浏览文件 @ 9386ac0a
......@@ -207,53 +207,56 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
is_forwarding = false;
} else {
int op_dev_id = GetOpDeviceID(*op);
if (op_dev_id == -1) { // var on all device
CreateComputationalOps(&result, *op, places_.size());
} else {
if (op_dev_id != -1) { // This op only runs on one specific device.
CreateComputationalOp(&result, *op, op_dev_id);
for (auto &var_name : op->OutputArgumentNames()) {
var_name_on_devices_.emplace(var_name, op_dev_id);
}
}
if (!is_forwarding && places_.size() > 1) {
// Currently, we assume that once gradient is generated, it can be
// broadcast, and each gradient is only broadcast once.
if (static_cast<bool>(boost::get<int>(op->GetAttr(
OpProtoAndCheckerMaker::OpRoleAttrName())) &
static_cast<int>(OpRole::kBackward))) {
try {
auto backward_vars =
boost::get<std::vector<std::string>>(op->GetNullableAttr(
OpProtoAndCheckerMaker::OpRoleVarAttrName()));
PADDLE_ENFORCE_EQ(backward_vars.size() % 2, 0);
for (size_t i = 0; i < backward_vars.size(); i += 2) {
auto &p_name = backward_vars[i];
auto &g_name = backward_vars[i + 1];
VLOG(10) << "Bcast " << g_name << " for parameter " << p_name;
switch (strategy_.reduce_) {
case BuildStrategy::ReduceStrategy::kReduce:
cur_device_id = GetAppropriateDeviceID({g_name});
CreateReduceOp(&result, g_name, cur_device_id);
var_name_on_devices_.emplace(g_name, cur_device_id);
bcast_var_name_set[cur_device_id].emplace(p_name);
break;
case BuildStrategy::ReduceStrategy::kAllReduce:
if (IsSparseGradient(g_name)) {
CreateReduceOp(&result, g_name, 0);
CreateBroadcastOp(&result, g_name, 0);
} else {
InsertAllReduceOp(&result, g_name);
}
break;
default:
LOG(FATAL) << "Unknown reduce strategy ";
break;
} else {
// This op runs on all devices, and its output may have parameter's
// gradients.
CreateComputationalOps(&result, *op, places_.size());
if (!is_forwarding && places_.size() > 1) {
// Currently, we assume that once gradient is generated, it can be
// broadcast, and each gradient is only broadcast once.
if (static_cast<bool>(boost::get<int>(op->GetAttr(
OpProtoAndCheckerMaker::OpRoleAttrName())) &
static_cast<int>(OpRole::kBackward))) {
try {
auto backward_vars =
boost::get<std::vector<std::string>>(op->GetNullableAttr(
OpProtoAndCheckerMaker::OpRoleVarAttrName()));
PADDLE_ENFORCE_EQ(backward_vars.size() % 2, 0);
for (size_t i = 0; i < backward_vars.size(); i += 2) {
auto &p_name = backward_vars[i];
auto &g_name = backward_vars[i + 1];
VLOG(10) << "Bcast " << g_name << " for parameter " << p_name;
switch (strategy_.reduce_) {
case BuildStrategy::ReduceStrategy::kReduce:
cur_device_id = GetAppropriateDeviceID({g_name});
CreateReduceOp(&result, g_name, cur_device_id);
var_name_on_devices_.emplace(g_name, cur_device_id);
bcast_var_name_set[cur_device_id].emplace(p_name);
break;
case BuildStrategy::ReduceStrategy::kAllReduce:
if (IsSparseGradient(g_name)) {
CreateReduceOp(&result, g_name, 0);
CreateBroadcastOp(&result, g_name, 0);
} else {
InsertAllReduceOp(&result, g_name);
}
break;
default:
LOG(FATAL) << "Unknown reduce strategy ";
break;
}
}
} catch (boost::bad_get e) {
}
} catch (boost::bad_get e) {
}
}
}
......@@ -470,7 +473,7 @@ void MultiDevSSAGraphBuilder::ConnectOp(SSAGraph *result, OpHandleBase *op,
void MultiDevSSAGraphBuilder::CreateDistTrainOp(SSAGraph *result,
const OpDesc &op) const {
int op_dev_id = -1;
if (op.Type() == "split_byref") {
if (op.Type() == "split_byref" || op.Type() == "split_selected_rows") {
op_dev_id = GetVarDeviceID(op.InputArgumentNames()[0]);
if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
op_dev_id = GetAppropriateDeviceID(op.InputArgumentNames());
......@@ -483,6 +486,9 @@ void MultiDevSSAGraphBuilder::CreateDistTrainOp(SSAGraph *result,
}
} else if (op.Type() == "concat") {
op_dev_id = GetVarDeviceID(op.InputArgumentNames()[0]);
for (auto &varname : op.OutputArgumentNames()) {
var_name_on_devices_.emplace(varname, op_dev_id);
}
} else {
PADDLE_ENFORCE(
"the distribute training related op should be in [split_byref, "
......
paddle/fluid/framework/details/ssa_graph_builder.h
浏览文件 @ 9386ac0a
......@@ -30,7 +30,7 @@ class SSAGraphBuilder {
SSAGraphBuilder() {}
virtual ~SSAGraphBuilder() {}
virtual std::unique_ptr<SSAGraph> Build(const ProgramDesc &program) const = 0;
virtual int GetVarDeviceID(const std::string &var_name) const { return -1; }
virtual int GetVarDeviceID(const std::string &var_name) const = 0;
DISABLE_COPY_AND_ASSIGN(SSAGraphBuilder);
......
paddle/fluid/framework/details/ssa_graph_checker.h
浏览文件 @ 9386ac0a
......@@ -16,6 +16,8 @@
#include "paddle/fluid/framework/details/ssa_graph_builder.h"
#include <string>
namespace paddle {
namespace framework {
namespace details {
......@@ -33,6 +35,10 @@ class SSAGraghBuilderWithChecker : public SSAGraphBuilder {
return graph;
}
int GetVarDeviceID(const std::string& var_name) const override {
return builder_->GetVarDeviceID(var_name);
}
bool IsValidGraph(const SSAGraph* graph) const;
private:
......
paddle/fluid/framework/details/ssa_graph_printer.h
浏览文件 @ 9386ac0a
......@@ -15,6 +15,7 @@
#pragma once
#include <iosfwd>
#include <string>
#include "paddle/fluid/framework/details/ssa_graph_builder.h"
namespace paddle {
......@@ -55,6 +56,10 @@ class SSAGraghBuilderWithPrinter : public SSAGraphBuilder {
return graph;
}
int GetVarDeviceID(const std::string& var_name) const override {
return builder_->GetVarDeviceID(var_name);
}
private:
std::unique_ptr<SSAGraphPrinter> printer_;
std::unique_ptr<SSAGraphBuilder> builder_;
......
paddle/fluid/framework/lod_tensor.cc
浏览文件 @ 9386ac0a
......@@ -68,7 +68,7 @@ std::ostream &operator<<(std::ostream &os, const LoDTensor &t) {
// only print first ten elements
int64_t size = t.numel() < 10 ? t.numel() : 10;
for (int64_t i = 0; i < size; ++i) {
if (t.type().hash_code() == typeid(float).hash_code()) {
if (t.type().hash_code() == typeid(float).hash_code()) { // NOLINT
os << t.data<float>()[i] << " ";
} else if (t.type().hash_code() == typeid(int64_t).hash_code()) {
os << t.data<int64_t>()[i] << " ";
......
paddle/fluid/framework/op_kernel_type.h
浏览文件 @ 9386ac0a
......@@ -97,7 +97,7 @@ inline bool NeedTransformLayout(const DataLayout& l, const DataLayout& r) {
return ret;
}
inline bool TransFromNeeded(const OpKernelType& l, const OpKernelType& r) {
inline bool NeedTransform(const OpKernelType& l, const OpKernelType& r) {
return (!platform::places_are_same_class(l.place_, r.place_)) ||
(l.data_type_ != r.data_type_) ||
NeedTransformLayout(l.data_layout_, r.data_layout_);
......
paddle/fluid/framework/operator.cc
浏览文件 @ 9386ac0a
......@@ -620,8 +620,6 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
"There are no kernels which are registered in the %s operator.", type_);
}
ExecutionContext ctx(*this, scope, *dev_ctx);
OpKernelMap& kernels = kernels_iter->second;
// TODO(dzhwinter) : kernel fallback mechanism will be added when all the
......@@ -631,7 +629,8 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
// Do selection
// }
auto expected_kernel_key = this->GetExpectedKernelType(ctx);
auto expected_kernel_key =
this->GetExpectedKernelType(ExecutionContext(*this, scope, *dev_ctx));
VLOG(3) << "expected_kernel_key:" << expected_kernel_key;
auto kernel_iter = kernels.find(expected_kernel_key);
......@@ -640,56 +639,34 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
KernelTypeToString(expected_kernel_key));
}
// do data transform
Scope& new_scope = scope.NewScope();
// do data transformScope &transfer_scope;
std::vector<std::string> transfered_inplace_vars;
auto* transfer_scope =
TryTransferData(scope, expected_kernel_key, &transfered_inplace_vars);
std::vector<std::string> inplace_vars;
for (auto& var_name_item : this->Inputs()) {
for (auto& var_name : var_name_item.second) {
auto* var = scope.FindVar(var_name);
if (var && VarIsTensor(var)) {
auto* tensor_in = GetTensorFromVar(var);
if (tensor_in->IsInitialized()) {
auto kernel_type_for_var = this->GetKernelTypeForVar(
var_name_item.first, *tensor_in, expected_kernel_key);
if (TransFromNeeded(kernel_type_for_var, expected_kernel_key)) {
auto out_var_names = OutputVars(true);
if (std::find(out_var_names.begin(), out_var_names.end(),
var_name) != out_var_names.end()) {
inplace_vars.push_back(var_name);
}
VLOG(3) << "Transform Variable " << var_name << " from "
<< kernel_type_for_var << " to " << expected_kernel_key;
auto* trans_var = new_scope.Var(var_name);
std::shared_ptr<Tensor> out(new Tensor);
DataTransform(expected_kernel_key, kernel_type_for_var, *tensor_in,
out.get());
CopyVariableWithTensor(*var, *(out.get()), trans_var);
}
}
}
}
// exec scope is the scope that kernel actually executed on.
const Scope& exec_scope =
(transfer_scope == nullptr ? scope : *transfer_scope);
if (!(expected_kernel_key.place_ == dev_ctx->GetPlace())) {
dev_ctx = pool.Get(expected_kernel_key.place_);
}
auto* new_dev_ctx = pool.Get(expected_kernel_key.place_);
kernel_iter->second->Compute(
ExecutionContext(*this, new_scope, *new_dev_ctx));
kernel_iter->second->Compute(ExecutionContext(*this, exec_scope, *dev_ctx));
for (auto& var_name : inplace_vars) {
VLOG(3) << "share inplace var " + var_name + " back to it's original scope";
auto* original_tensor = GetMutableTensorFromVar(scope.FindVar(var_name));
auto* transformed_tensor = GetTensorFromVar(new_scope.FindVar(var_name));
original_tensor->ShareDataWith(*transformed_tensor);
if (!transfered_inplace_vars.empty()) {
// there is inplace variable has been transfered.
TransferInplaceVarsBack(scope, transfered_inplace_vars, *transfer_scope);
}
/*For profiling/benchmark only*/
if (FLAGS_benchmark) {
new_dev_ctx->Wait();
dev_ctx->Wait();
}
if (FLAGS_check_nan_inf) {
for (auto& vname : OutputVars(true)) {
auto* var = new_scope.FindVar(vname);
auto* var = exec_scope.FindVar(vname);
if (var == nullptr) continue;
if (var->IsType<framework::LoDTensor>()) {
CheckTensorNANOrInf(vname, var->Get<framework::LoDTensor>());
......@@ -697,6 +674,64 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
}
}
}
void OperatorWithKernel::TransferInplaceVarsBack(
const Scope& scope, const std::vector<std::string>& inplace_vars,
const Scope& transfer_scope) const {
for (auto& var_name : inplace_vars) {
VLOG(3) << "share inplace var " + var_name + " back to it's original scope";
auto* original_tensor = GetMutableTensorFromVar(scope.FindVar(var_name));
auto* transformed_tensor =
GetTensorFromVar(transfer_scope.FindVar(var_name));
original_tensor->ShareDataWith(*transformed_tensor);
}
}
Scope* OperatorWithKernel::TryTransferData(
const Scope& scope, const OpKernelType& expected_kernel_key,
std::vector<std::string>* transfered_inplace_vars) const {
Scope* new_scope = nullptr;
for (auto& var_name_item : Inputs()) {
for (auto& var_name : var_name_item.second) {
auto* var = scope.FindVar(var_name);
// Only tensor can be tranfer to another device.
if (var == nullptr || !VarIsTensor(var)) {
continue;
}
auto* tensor_in = GetTensorFromVar(var);
if (!tensor_in->IsInitialized()) {
continue;
}
auto kernel_type_for_var = GetKernelTypeForVar(
var_name_item.first, *tensor_in, expected_kernel_key);
if (!NeedTransform(kernel_type_for_var, expected_kernel_key)) {
continue;
}
auto out_var_names = OutputVars(true);
if (std::find(out_var_names.begin(), out_var_names.end(), var_name) !=
out_var_names.end()) {
transfered_inplace_vars->emplace_back(var_name);
}
VLOG(3) << "Transform Variable " << var_name << " from "
<< kernel_type_for_var << " to " << expected_kernel_key;
if (new_scope == nullptr) {
new_scope = &scope.NewScope();
}
auto* trans_var = new_scope->Var(var_name);
Tensor out;
TransformData(expected_kernel_key, kernel_type_for_var, *tensor_in, &out);
SetTensorToVariable(*var, out, trans_var);
}
}
return new_scope;
}
proto::VarType::Type OperatorWithKernel::IndicateDataType(
const ExecutionContext& ctx) const {
......
paddle/fluid/framework/operator.h
浏览文件 @ 9386ac0a
......@@ -384,6 +384,20 @@ class OperatorWithKernel : public OperatorBase {
// same.
proto::VarType::Type IndicateDataType(const ExecutionContext& ctx) const;
void RunImpl(const Scope& scope, const platform::Place& place) const final;
/**
* Transfer data from scope to a transfered scope. If there is no data need to
* be tranfered, it returns nullptr.
*
* * transfered_inplace_vars is a output vector.
*/
Scope* TryTransferData(
const Scope& scope, const OpKernelType& expected_kernel_key,
std::vector<std::string>* transfered_inplace_vars) const;
void TransferInplaceVarsBack(const Scope& scope,
const std::vector<std::string>& inplace_vars,
const Scope& exec_scope) const;
};
extern bool OpSupportGPU(const std::string& op_type);
......
paddle/fluid/framework/parallel_executor.cc
浏览文件 @ 9386ac0a
......@@ -133,17 +133,18 @@ ParallelExecutor::ParallelExecutor(
void ParallelExecutor::BCastParamsToGPUs(
const std::unordered_set<std::string> &vars) const {
// the the initialize bcast, all vars would be bcast from device(0), otherwise
// the the initializing bcast, all vars would be bcast from device(0),
// otherwise
// bcast from the specified device.
bool initialize = builder_.get() == nullptr ? true : false;
bool initializing = builder_.get() == nullptr ? true : false;
for (auto &var : vars) {
int var_dev_id =
builder_.get() == nullptr ? -1 : builder_->GetVarDeviceID(var);
if (!initialize && var_dev_id == -1) continue;
if (!initializing && var_dev_id == -1) continue;
framework::Variable *main_var = nullptr;
if (initialize) {
if (initializing) {
main_var = member_->local_scopes_[0]->FindVar(var);
} else {
main_var = member_->local_scopes_[var_dev_id]->FindVar(var);
......@@ -164,7 +165,8 @@ void ParallelExecutor::BCastParamsToGPUs(
auto place = member_->places_[i];
void *buffer;
if ((initialize && i == 0) || (!initialize && i == var_dev_id)) {
if ((initializing && i == 0) ||
(!initializing && static_cast<int>(i) == var_dev_id)) {
buffer = const_cast<void *>(main_tensor.data<void>());
} else {
auto local_scope = member_->local_scopes_[i];
......@@ -181,8 +183,16 @@ void ParallelExecutor::BCastParamsToGPUs(
platform::NCCLGroupGuard guard;
for (size_t i = 0; i < member_->places_.size(); ++i) {
auto &nccl_ctx = member_->nccl_ctxs_->at(member_->places_[i]);
platform::dynload::ncclBcast(buffers[i], numel, data_type, 0,
nccl_ctx.comm_, nccl_ctx.stream());
if (initializing) {
platform::dynload::ncclBcast(buffers[i], numel, data_type, 0,
nccl_ctx.comm_, nccl_ctx.stream());
} else {
if (var_dev_id >= 0) {
platform::dynload::ncclBcast(buffers[i], numel, data_type,
var_dev_id, nccl_ctx.comm_,
nccl_ctx.stream());
}
}
}
member_->nccl_ctxs_->WaitAll();
}
......
paddle/fluid/framework/tensor_util.cc
浏览文件 @ 9386ac0a
......@@ -69,7 +69,22 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
PADDLE_ENFORCE(platform::is_gpu_place(ctx_place));
auto stream =
reinterpret_cast<const platform::CUDADeviceContext&>(ctx).stream();
memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size, stream);
if (platform::is_same_place(src_place, dst_place)) {
memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
stream);
} else {
if (platform::is_same_place(ctx_place, src_place)) {
memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
stream);
platform::DeviceContextPool::Instance().Get(src.place())->Wait();
} else if (platform::is_same_place(ctx_place, dst_place)) {
platform::DeviceContextPool::Instance().Get(src.place())->Wait();
memory::Copy(dst_gpu_place, dst_ptr, src_gpu_place, src_ptr, size,
stream);
} else {
PADDLE_THROW("ctx is not belong to dst_gpu_place or src_gpu_place.");
}
}
}
#endif
}
......@@ -78,10 +93,10 @@ void TensorCopy(const Tensor& src, const platform::Place& dst_place,
Tensor* dst) {
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
const platform::DeviceContext* dev_ctx;
if (platform::is_gpu_place(src.place())) {
dev_ctx = pool.Get(src.place());
} else {
if (platform::is_gpu_place(dst_place)) {
dev_ctx = pool.Get(dst_place);
} else {
dev_ctx = pool.Get(src.place());
}
TensorCopy(src, dst_place, *dev_ctx, dst);
}
......
paddle/fluid/inference/CMakeLists.txt
浏览文件 @ 9386ac0a
......@@ -28,9 +28,10 @@ endif()
if(WITH_TESTING)
# both tests/book and analysis depends the models that generated by python/paddle/fluid/tests/book
add_subdirectory(tests/book)
add_subdirectory(analysis)
endif()
add_subdirectory(analysis)
if (TENSORRT_FOUND)
add_subdirectory(tensorrt)
endif()
paddle/fluid/inference/analysis/README.md 0 → 100644
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/inference/analysis/analyzer.cc 0 → 100644
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/inference/analysis/analyzer.h 0 → 100644
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/inference/analysis/analyzer_tester.cc 0 → 100644
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/inference/analysis/argument.h
浏览文件 @ 9386ac0a
......@@ -41,6 +41,9 @@ struct Argument {
// The original program desc.
std::unique_ptr<framework::proto::ProgramDesc> origin_program_desc;
// The processed program desc.
std::unique_ptr<framework::proto::ProgramDesc> transformed_program_desc;
};
#define UNLIKELY(condition) __builtin_expect(static_cast<bool>(condition), 0)
......
paddle/fluid/inference/analysis/helper.cc 0 → 100644
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/inference/analysis/node_attr_flags.h 0 → 100644
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/memory/malloc.cc
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/operators/adam_op.cc
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/operators/adam_op.h
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/operators/checkpoint_notify_op.cc 0 → 100644
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/operators/elementwise_add_mkldnn_op.cc 0 → 100644
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/operators/load_op.cc
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/operators/math/blas.h
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/operators/reader/create_py_reader_op.cc 0 → 100644
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/operators/save_op.cc
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/platform/dynload/mklml.cc 0 → 100644
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/platform/dynload/mklml.h 0 → 100644
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/pybind/pybind.cc
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/fluid/pybind/tensor_py.h
浏览文件 @ 9386ac0a
此差异已折叠。
paddle/scripts/paddle_build.sh
浏览文件 @ 9386ac0a
此差异已折叠。
python/paddle/dataset/mnist.py
浏览文件 @ 9386ac0a
此差异已折叠。
python/paddle/fluid/__init__.py
浏览文件 @ 9386ac0a
此差异已折叠。
python/paddle/fluid/executor.py
浏览文件 @ 9386ac0a
此差异已折叠。
python/paddle/fluid/framework.py
浏览文件 @ 9386ac0a
此差异已折叠。
python/paddle/fluid/io.py
浏览文件 @ 9386ac0a
此差异已折叠。
python/paddle/fluid/layers/io.py
浏览文件 @ 9386ac0a
此差异已折叠。
python/paddle/fluid/layers/nn.py
浏览文件 @ 9386ac0a
此差异已折叠。
python/paddle/fluid/lod_tensor.py
浏览文件 @ 9386ac0a
此差异已折叠。
python/paddle/fluid/metrics.py
浏览文件 @ 9386ac0a
此差异已折叠。
python/paddle/fluid/optimizer.py
浏览文件 @ 9386ac0a
此差异已折叠。
python/paddle/fluid/tests/unittests/test_dist_mnist.py 0 → 100644
浏览文件 @ 9386ac0a
此差异已折叠。
python/paddle/fluid/trainer.py
浏览文件 @ 9386ac0a
此差异已折叠。
python/paddle/v2/dataset/mnist.py
浏览文件 @ 9386ac0a
此差异已折叠。
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或