提交 0c273ff4 编写于 作者: C chengduoZH

fix conflict

#!/bin/bash
set -e
readonly VERSION="3.8"
version=$(clang-format -version)
if ! [[ $version == *"$VERSION"* ]]; then
echo "clang-format version check failed."
echo "a version contains '$VERSION' is needed, but get '$version'"
echo "you can install the right version, and make an soft-link to '\$PATH' env"
exit -1
fi
clang-format $@
......@@ -24,4 +24,5 @@ cmake-build-*
python/paddle/v2/framework/core.so
CMakeFiles
cmake_install.cmake
paddle/.timestamp
python/paddlepaddle.egg-info/
......@@ -19,10 +19,10 @@
- id: end-of-file-fixer
- repo: local
hooks:
- id: clang-format
- id: clang-format-with-version-check
name: clang-format
description: Format files with ClangFormat.
entry: clang-format -i
entry: bash ./.clang_format.hook -i
language: system
files: \.(c|cc|cxx|cpp|cu|h|hpp|hxx|proto)$
- repo: https://github.com/PaddlePaddle/pre-commit-golang
......
......@@ -37,8 +37,8 @@ before_install:
- if [[ "$JOB" == "check_style" ]]; then sudo ln -s /usr/bin/clang-format-3.8 /usr/bin/clang-format; fi
# Paddle is using protobuf 3.1 currently. Protobuf 3.2 breaks the compatibility. So we specify the python
# protobuf version.
- pip install numpy wheel 'protobuf==3.1' sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit requests==2.9.2 LinkChecker
- pip install rarfile nltk==3.2.2 scipy==0.19.0 recordio matplotlib Pillow
- pip install -r $TRAVIS_BUILD_DIR/python/requirements.txt
- pip install wheel sphinx==1.5.6 recommonmark sphinx-rtd-theme==0.1.9 virtualenv pre-commit LinkChecker
- curl https://glide.sh/get | bash
- eval "$(GIMME_GO_VERSION=1.8.3 gimme)"
- go get -u github.com/alecthomas/gometalinter
......
......@@ -14,8 +14,8 @@
cmake_minimum_required(VERSION 3.0)
set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
set(PROJ_ROOT ${CMAKE_CURRENT_SOURCE_DIR})
set(PROJ_BINARY_ROOT ${CMAKE_CURRENT_BINARY_DIR})
set(PADDLE_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR})
set(PADDLE_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR})
include(system)
......@@ -55,6 +55,7 @@ option(WITH_C_API "Compile PaddlePaddle with C-API(Prediction)" OFF)
option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF)
option(GLIDE_INSTALL "Download and install go dependencies " ON)
option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF)
option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF)
# CMAKE_BUILD_TYPE
if(NOT CMAKE_BUILD_TYPE)
......@@ -121,8 +122,8 @@ include(version) # set PADDLE_VERSION
include(coveralls) # set code coverage
include_directories("${PROJ_ROOT}")
include_directories("${PROJ_ROOT}/paddle/cuda/include")
include_directories("${PADDLE_SOURCE_DIR}")
include_directories("${PADDLE_SOURCE_DIR}/paddle/cuda/include")
include_directories("${CMAKE_CURRENT_BINARY_DIR}/proto")
include_directories("${CMAKE_CURRENT_BINARY_DIR}/go/pserver/client/c")
include_directories(${Boost_INCLUDE_DIRS})
......@@ -137,14 +138,14 @@ set(EXTERNAL_LIBS
)
if(WITH_GPU)
list(APPEND EXTERNAL_LIB ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
list(APPEND EXTERNAL_LIBS ${CUDA_LIBRARIES} ${CUDA_rt_LIBRARY})
if(NOT WITH_DSO)
list(APPEND EXTERNAL_LIB ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY})
list(APPEND EXTERNAL_LIBS ${CUDNN_LIBRARY} ${CUDA_CUBLAS_LIBRARIES} ${CUDA_curand_LIBRARY})
endif(NOT WITH_DSO)
endif(WITH_GPU)
if(WITH_MKLDNN)
list(APPEND EXTERNAL_LIBS ${MKLDNN_LIBRARY} ${MKLDNN_IOMP_LIB})
list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB} ${MKLDNN_IOMP_LIB})
endif()
if(USE_NNPACK)
......@@ -164,10 +165,12 @@ if(WITH_GOLANG)
add_subdirectory(go)
endif(WITH_GOLANG)
set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build")
add_subdirectory(paddle)
if(WITH_PYTHON)
add_subdirectory(python)
endif()
if(WITH_DOC)
add_subdirectory(doc)
endif()
......@@ -10,13 +10,11 @@ RUN /bin/bash -c 'if [[ -n ${UBUNTU_MIRROR} ]]; then sed -i 's#http://archive.ub
ARG WITH_GPU
ARG WITH_AVX
ARG WITH_DOC
ARG WITH_STYLE_CHECK
ENV WOBOQ OFF
ENV WITH_GPU=${WITH_GPU:-OFF}
ENV WITH_GPU=${WITH_GPU:-ON}
ENV WITH_AVX=${WITH_AVX:-ON}
ENV WITH_DOC=${WITH_DOC:-OFF}
ENV WITH_STYLE_CHECK=${WITH_STYLE_CHECK:-OFF}
ENV HOME /root
# Add bash enhancements
......@@ -34,9 +32,6 @@ RUN apt-get update && \
net-tools && \
apt-get clean -y
# paddle is using numpy.flip, which is introduced since 1.12.0
RUN pip --no-cache-dir install 'numpy>=1.12.0'
# Install Go and glide
RUN wget -qO- https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz | \
tar -xz -C /usr/local && \
......@@ -58,19 +53,23 @@ RUN localedef -i en_US -f UTF-8 en_US.UTF-8
# FIXME: due to temporary ipykernel dependency issue, specify ipykernel jupyter
# version util jupyter fixes this issue.
RUN pip install --upgrade pip && \
pip install -U 'protobuf==3.1.0' && \
pip install -U wheel pillow BeautifulSoup && \
pip install -U wheel && \
pip install -U docopt PyYAML sphinx && \
pip install -U sphinx-rtd-theme==0.1.9 recommonmark && \
pip install pre-commit 'requests==2.9.2' 'ipython==5.3.0' && \
pip install -U sphinx-rtd-theme==0.1.9 recommonmark
RUN pip install pre-commit 'ipython==5.3.0' && \
pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
pip install rarfile
pip install opencv-python
COPY ./python/requirements.txt /root/
RUN pip install -r /root/requirements.txt
# To fix https://github.com/PaddlePaddle/Paddle/issues/1954, we use
# the solution in https://urllib3.readthedocs.io/en/latest/user-guide.html#ssl-py2
RUN apt-get install -y libssl-dev libffi-dev
RUN pip install certifi urllib3[secure]
# Install woboq_codebrowser to /woboq
RUN git clone https://github.com/woboq/woboq_codebrowser /woboq && \
(cd /woboq \
......
......@@ -28,6 +28,10 @@ if(NOT WITH_TIMER)
add_definitions(-DPADDLE_DISABLE_TIMER)
endif(NOT WITH_TIMER)
if(USE_EIGEN_FOR_BLAS)
add_definitions(-DPADDLE_USE_EIGEN_FOR_BLAS)
endif(USE_EIGEN_FOR_BLAS)
if(NOT WITH_PROFILER)
add_definitions(-DPADDLE_DISABLE_PROFILER)
endif(NOT WITH_PROFILER)
......@@ -129,7 +133,7 @@ if(WITH_GOLANG)
add_custom_command(OUTPUT ${CMAKE_BINARY_DIR}/glide
COMMAND env GOPATH=${GOPATH} ${GLIDE} install
COMMAND touch ${CMAKE_BINARY_DIR}/glide
DEPENDS ${PROJ_ROOT}/go/glide.lock
DEPENDS ${PADDLE_SOURCE_DIR}/go/glide.lock
WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go"
)
......
......@@ -52,7 +52,7 @@ macro(add_style_check_target TARGET_NAME)
if(SOURCES_LIST)
add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
COMMAND "${PYTHON_EXECUTABLE}" "${PROJ_ROOT}/paddle/scripts/cpplint.py"
COMMAND "${PYTHON_EXECUTABLE}" "${PADDLE_SOURCE_DIR}/paddle/scripts/cpplint.py"
"--filter=${STYLE_FILTER}"
${SOURCES_LIST}
COMMENT "cpplint: Checking source code style"
......
......@@ -2,7 +2,7 @@ if(NOT WITH_GPU)
return()
endif()
set(CUDNN_ROOT "" CACHE PATH "CUDNN ROOT")
set(CUDNN_ROOT "/usr" CACHE PATH "CUDNN ROOT")
find_path(CUDNN_INCLUDE_DIR cudnn.h
PATHS ${CUDNN_ROOT} ${CUDNN_ROOT}/include
$ENV{CUDNN_ROOT} $ENV{CUDNN_ROOT}/include ${CUDA_TOOLKIT_INCLUDE}
......
......@@ -51,7 +51,7 @@ ExternalProject_Add(
${EXTERNAL_PROJECT_LOG_ARGS}
DEPENDS ${MKLDNN_DEPENDS}
GIT_REPOSITORY "https://github.com/01org/mkl-dnn.git"
GIT_TAG "v0.9"
GIT_TAG "v0.10"
PREFIX ${MKLDNN_SOURCES_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_INSTALL_PREFIX=${MKLDNN_INSTALL_DIR}
......
......@@ -28,7 +28,7 @@ INCLUDE(ExternalProject)
SET(MKLML_PROJECT "extern_mklml")
SET(MKLML_VER "mklml_lnx_2018.0.20170720")
SET(MKLML_URL "https://github.com/01org/mkl-dnn/releases/download/v0.9/${MKLML_VER}.tgz")
SET(MKLML_URL "https://github.com/01org/mkl-dnn/releases/download/v0.10/${MKLML_VER}.tgz")
SET(MKLML_SOURCE_DIR "${THIRD_PARTY_PATH}/mklml")
SET(MKLML_DOWNLOAD_DIR "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
SET(MKLML_DST_DIR "mklml")
......
......@@ -73,10 +73,18 @@ INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})
# linear algebra libraries for cc_library(xxx SRCS xxx.c DEPS cblas)
SET(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cblas_dummy.c)
FILE(WRITE ${dummyfile} "const char * dummy = \"${dummyfile}\";")
ADD_LIBRARY(cblas STATIC ${dummyfile})
IF(${CBLAS_PROVIDER} MATCHES MKL)
ADD_LIBRARY(cblas SHARED ${dummyfile})
ELSE()
ADD_LIBRARY(cblas STATIC ${dummyfile})
ENDIF()
TARGET_LINK_LIBRARIES(cblas ${CBLAS_LIBRARIES})
IF(NOT ${CBLAS_FOUND})
ADD_DEPENDENCIES(cblas extern_openblas)
LIST(APPEND external_project_dependencies cblas)
ELSE()
IF("${CBLAS_PROVIDER}" STREQUAL "MKLML")
ADD_DEPENDENCIES(cblas mklml)
ENDIF()
ENDIF(NOT ${CBLAS_FOUND})
......@@ -9,11 +9,6 @@ function(CheckCompilerCXX11Flag)
if(${CMAKE_CXX_COMPILER_VERSION} VERSION_LESS 4.8)
message(FATAL_ERROR "Unsupported GCC version. GCC >= 4.8 required.")
endif()
# TODO(qijun) gcc 4.9 or later versions raise SEGV due to the optimization problem.
# Use Debug mode instead for now.
if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 4.9 OR CMAKE_CXX_COMPILER_VERSION VERSION_EQUAL 4.9)
set(CMAKE_BUILD_TYPE "Debug" CACHE STRING "" FORCE)
endif()
elseif(CMAKE_CXX_COMPILER_ID STREQUAL "AppleClang" OR CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
# cmake >= 3.0 compiler id "AppleClang" on Mac OS X, otherwise "Clang"
# Apple Clang is a different compiler than upstream Clang which havs different version numbers.
......@@ -158,7 +153,7 @@ set(CUDA_PROPAGATE_HOST_FLAGS OFF)
# Release/Debug flags set by cmake. Such as -O3 -g -DNDEBUG etc.
# So, don't set these flags here.
LIST(APPEND CUDA_NVCC_FLAGS -std=c++11 --default-stream per-thread)
LIST(APPEND CUDA_NVCC_FLAGS -std=c++11)
LIST(APPEND CUDA_NVCC_FLAGS --use_fast_math)
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
......
......@@ -411,7 +411,7 @@ function(py_test TARGET_NAME)
set(multiValueArgs SRCS DEPS)
cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
add_test(NAME ${TARGET_NAME}
COMMAND env PYTHONPATH=${PADDLE_PYTHON_PACKAGE_DIR}
COMMAND env PYTHONPATH=${PADDLE_PYTHON_BUILD_DIR}/lib-python
python2 ${py_test_SRCS}
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
endif()
......
......@@ -12,7 +12,7 @@ set(CPACK_PACKAGE_DESCRIPTION "")
set(CPACK_DEBIAN_PACKAGE_DEPENDS "libpython2.7-dev, libstdc++6, python-pip, curl, libgfortran3, python-pip-whl")
set(CPACK_DEBIAN_PACKAGE_SECTION Devel)
set(CPACK_DEBIAN_PACKAGE_VERSION ${PADDLE_VERSION})
set(CPACK_DEBIAN_PACKAGE_CONTROL_EXTRA "${PROJ_ROOT}/paddle/scripts/deb/postinst")
set(CPACK_DEBIAN_PACKAGE_CONTROL_EXTRA "${PADDLE_SOURCE_DIR}/paddle/scripts/deb/postinst")
#set(CPACK_GENERATOR "DEB")
# Start cpack
include (CMakePackageConfigHelpers)
......
......@@ -141,8 +141,8 @@ endmacro()
function(create_resources res_file output_file)
add_custom_command(
OUTPUT ${output_file}
COMMAND python ARGS ${PROJ_ROOT}/cmake/make_resource.py ${res_file} ${output_file}
DEPENDS ${res_file} ${PROJ_ROOT}/cmake/make_resource.py)
COMMAND python ARGS ${PADDLE_SOURCE_DIR}/cmake/make_resource.py ${res_file} ${output_file}
DEPENDS ${res_file} ${PADDLE_SOURCE_DIR}/cmake/make_resource.py)
endfunction()
......
......@@ -4,7 +4,7 @@ set(tmp_version "HEAD")
while ("${PADDLE_VERSION}" STREQUAL "")
execute_process(
COMMAND ${GIT_EXECUTABLE} describe --tags --abbrev=0 ${tmp_version}
WORKING_DIRECTORY ${PROJ_ROOT}
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}
OUTPUT_VARIABLE GIT_TAG_NAME
RESULT_VARIABLE GIT_RESULT
ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
......
关于PaddlePaddle
================
PaddlePaddle是一个最早由百度科学家和工程师共同研发的并行分布式深度学习平台,兼备易用性、高效性、灵活性和可扩展性,目前已被百度内部多个产品线广泛使用。
PaddlePaddle目前已经开放源码, 但是远未完善,我们希望能在这个基础上不断的改进、扩展和延伸。
同时我们希望广大开发者积极提供反馈和贡献源代码,建立一个活跃的开源社区。
致谢
--------
在此,特别感谢PaddlePaddle的[所有贡献者](https://github.com/PaddlePaddle/Paddle/graphs/contributors)
ABOUT
=======
PaddlPaddle is an easy-to-use, efficient, flexible and scalable deep learning platform,
which is originally developed by Baidu scientists and engineers for the purpose of applying deep learning to many products at Baidu.
PaddlePaddle is now open source but far from complete, which is intended to be built upon, improved, scaled, and extended.
We hope to build an active open source community both by providing feedback and by actively contributing to the source code.
Credits
--------
We owe many thanks to `all contributors and developers <https://github.com/PaddlePaddle/Paddle/graphs/contributors>`_ of PaddlePaddle!
......@@ -257,6 +257,11 @@ seq_concat
.. autoclass:: paddle.v2.layer.seq_concat
:noindex:
seq_slice
---------
.. autoclass:: paddle.v2.layer.seq_slice
:noindex:
kmax_sequence_score
-------------------
.. autoclass:: paddle.v2.layer.kmax_sequence_score
......@@ -362,6 +367,11 @@ trans
.. autoclass:: paddle.v2.layer.trans
:noindex:
scale_shift
-----------
.. autoclass:: paddle.v2.layer.scale_shift
:noindex:
Sampling Layers
===============
......@@ -409,9 +419,14 @@ multi_binary_label_cross_entropy_cost
.. autoclass:: paddle.v2.layer.multi_binary_label_cross_entropy_cost
:noindex:
huber_cost
----------
.. autoclass:: paddle.v2.layer.huber_cost
huber_regression_cost
-------------------------
.. autoclass:: paddle.v2.layer.huber_regression_cost
:noindex:
huber_classification_cost
-------------------------
.. autoclass:: paddle.v2.layer.huber_classification_cost
:noindex:
lambda_cost
......
## Auto Gradient Checker Design
## Backgraound:
- Operator forward computing is easy to check if the result is right because it has a clear definition. **But** backpropagation is a notoriously difficult algorithm to debug and get right:
- 1. you should get the right backpropagation formula according to the forward computation.
- 2. you should implement it right in CPP.
- 3. it's difficult to prepare test data.
- Auto gradient check gets a numeric gradient by forward Operator and use it as a reference of the backward Operator's result. It has several advantages:
- 1. numeric gradient checker only need forward operator.
- 2. user only need to prepare the input data for forward Operator.
## Mathematical Theory
The following two document from stanford has a detailed explanation of how to get numeric gradient and why it's useful.
- [Gradient checking and advanced optimization(en)](http://deeplearning.stanford.edu/wiki/index.php/Gradient_checking_and_advanced_optimization)
- [Gradient checking and advanced optimization(cn)](http://ufldl.stanford.edu/wiki/index.php/%E6%A2%AF%E5%BA%A6%E6%A3%80%E9%AA%8C%E4%B8%8E%E9%AB%98%E7%BA%A7%E4%BC%98%E5%8C%96)
## Numeric Gradient Implementation
### Python Interface
```python
def get_numeric_gradient(op,
input_values,
output_name,
input_to_check,
delta=0.005,
local_scope=None):
"""
Get Numeric Gradient for an operator's input.
:param op: C++ operator instance, could be an network
:param input_values: The input variables. Should be an dictionary, key is
variable name. Value is numpy array.
:param output_name: The final output variable name.
:param input_to_check: The input variable need to get gradient.
:param delta: The perturbation value for numeric gradient method. The
smaller delta is, the more accurate result will get. But if that delta is
too small, it could occur numerical stability problem.
:param local_scope: The local scope used for get_numeric_gradient.
:return: The gradient array in numpy format.
"""
```
### Explaination:
- Why need `output_name`
- One Operator may have multiple Output, you can get independent gradient from each Output. So user should set one output to calculate.
- Why need `input_to_check`
- One operator may have multiple inputs. Gradient Op can calculate the gradient of these Inputs at the same time. But Numeric Gradient needs to calculate them one by one. So `get_numeric_gradient` is designed to calculate the gradient for one input. If you need to compute multiple inputs, you can call `get_numeric_gradient` multiple times.
### Core Algorithm Implementation
```python
# we only compute gradient of one element each time.
# we use a for loop to compute the gradient of every element.
for i in xrange(tensor_size):
# get one input element throw it's index i.
origin = tensor_to_check.get_float_element(i)
# add delta to it, run op and then get the sum of the result tensor.
x_pos = origin + delta
tensor_to_check.set_float_element(i, x_pos)
y_pos = get_output()
# plus delta to this element, run op and get the sum of the result tensor.
x_neg = origin - delta
tensor_to_check.set_float_element(i, x_neg)
y_neg = get_output()
# restore old value
tensor_to_check.set_float_element(i, origin)
# compute the gradient of this element and store it into a numpy array.
gradient_flat[i] = (y_pos - y_neg) / delta / 2
# reshape the gradient result to the shape of the source tensor.
return gradient_flat.reshape(tensor_to_check.get_dims())
```
## Auto Graident Checker Framework
Each Operator Kernel has three kinds of Gradient:
- 1. Numeric Gradient
- 2. CPU Operator Gradient
- 3. GPU Operator Gradient(if supported)
Numeric Gradient Only relies on forward Operator. So we use Numeric Gradient as the reference value.
- 1. calculate the numeric gradient.
- 2. calculate CPU kernel Gradient with the backward Operator and compare it with the numeric gradient.
- 3. calculate GPU kernel Gradient with the backward Operator and compare it with the numeric gradient.(if support GPU)
#### Python Interface
```python
def check_grad(self,
forward_op,
input_vars,
inputs_to_check,
output_name,
no_grad_set=None,
only_cpu=False,
max_relative_error=0.005):
"""
:param forward_op: used to create backward_op
:param input_vars: numpy value of input variable. The following
computation will use these variables.
:param inputs_to_check: inputs var names that should check gradient.
:param output_name: output name that used to
:param max_relative_error: The relative tolerance parameter.
:param no_grad_set: used when create backward ops
:param only_cpu: only compute and check gradient on cpu kernel.
:return:
"""
```
### How to check if two numpy array is close enough?
if `abs_numeric_grad` is nearly zero, then use abs error for numeric_grad, not relative
```python
numeric_grad = ...
operator_grad = numpy.array(scope.find_var(grad_var_name(name)).get_tensor())
abs_numeric_grad = numpy.abs(numeric_grad)
# if abs_numeric_grad is nearly zero, then use abs error for numeric_grad, not relative
# error.
abs_numeric_grad[abs_numeric_grad < 1e-3] = 1
diff_mat = numpy.abs(abs_numeric_grad - operator_grad) / abs_numeric_grad
max_diff = numpy.max(diff_mat)
```
#### Notes:
1,The Input data for auto gradient checker should be reasonable to avoid numeric problem.
#### Refs:
- [Gradient checking and advanced optimization(en)](http://deeplearning.stanford.edu/wiki/index.php/Gradient_checking_and_advanced_optimization)
- [Gradient checking and advanced optimization(cn)](http://ufldl.stanford.edu/wiki/index.php/%E6%A2%AF%E5%BA%A6%E6%A3%80%E9%AA%8C%E4%B8%8E%E9%AB%98%E7%BA%A7%E4%BC%98%E5%8C%96)
......@@ -54,17 +54,18 @@ The life cycle of a single task is illustrated below:
<img src="src/paddle-task-states.png"/>
1. When a new pass of training starts, all tasks will be placed in the todo queue.
1. The master server will dispatch few tasks to each trainer at a time, puts them in the pending queue and waits for completion.
1. The trainer will work on its tasks and tell the master server once a task is completed. The master server will dispatch a new task to that trainer.
1. If a task timeout. the master server will move it back to the todo queue. The timeout count will increase by one. If the timeout count is above a threshold, the task is likely to cause a trainer to crash, so it will be discarded.
1. Upon trainer requests for new task, the master server will dispatch a task from todo queue to it, put the task in the pending queue and wait for completion.
1. The trainer will work on its task and tell the master server once the task is completed and ask for new task. The master server will dispatch a new task to that trainer.
1. If a task fails for any reason in trainer, or takes longer than a specific period of time, the master server will move the task back to the todo queue. The timeout count for that task will increase by one. If the timeout count is above a threshold, the task is likely to cause a trainer to crash, then it will be discarded.
1. The master server will move completed task to the done queue. When the todo queue is empty, the master server will start a new pass by moving all tasks in the done queue to todo queue and reset the timeout counter of all tasks to zero.
### Trainer Process
The trainer process will:
- Receive tasks from the master.
- Work on the tasks: calculate and upload gradient to parameter servers, and update local model by downloading new parameters from parameter servers.
- Request tasks from the master.
- Work on the tasks
- Upload gradient to parameter servers, and update local model by downloading new parameters from parameter servers.
### Parameter Server Process
......@@ -119,8 +120,8 @@ When the master is started by the Kubernetes, it executes the following steps at
1. Grabs a unique *master* lock in etcd, which prevents concurrent master instantiations.
1. Recovers the task queues from etcd if they already exist, otherwise, the master will create them.
1. Watches the trainer prefix keys `/trainer/` on etcd to find the live trainers.
1. Starts dispatching the tasks to the trainers, and updates task queue using an etcd transaction to ensure lock is held during the update.
1. Write its ip address to */master/addr* so that trainers can discover it.
1. Listens to trainers' request of task, dispatch one upon request, and updates task queue using an etcd transaction to ensure lock is held during the update.
When the master server process is dead for any reason, Kubernetes will restart it. It will be online again with all states recovered from etcd in few minutes.
......@@ -128,13 +129,11 @@ When the master server process is dead for any reason, Kubernetes will restart i
When the trainer is started by the Kubernetes, it executes the following steps at startup:
1. Watches the available parameter server prefix keys `/ps/` on etcd and waits until the count of parameter servers reaches the desired count.
1. Generates a unique ID, and sets key `/trainer/<unique ID>` with its contact address as value. The key will be deleted when the lease expires, so the master will be aware of the trainer being online and offline.
1. Waits for tasks from the master to start training.
1. Watches the available parameter server prefix keys `/ps/` on etcd and waits until the count of parameter servers reaches the desired count */ps_desired*.
1. Finds and watches */master/addr* to get master's address.
1. Requests for tasks from the master to start training.
If trainer's etcd lease expires, it will try set key `/trainer/<unique ID>` again so that the master server can discover the trainer again.
When a trainer fails, Kuberentes would try to restart it. The recovered trainer would fetch tasks from the TODO queue and go on training.
When a trainer fails, Kuberentes would try to restart it. The recovered trainer would fetch tasks from master and go on training.
### Parameter Server Process
......
# Alalysis of large model distributed training in Paddle
***NOTE: This is only some note for how we implemeted this scheme in V1, not a new design.***
## What is it
We often encounter cases that the embedding layer parameters(sparse) are so large that we can not store it in the trainer's memory when training. So we need to put them to several servers, and fetch them row by row instead of fetch all of the parameters.
## How to use
Specify command-line argument like `--loadsave_parameters_in_pserver=true --ports_num_for_sparse=1 --use_old_updater=1` when starting the paddle trainer. And also add something like `--ports_num_for_sparse=1 --pserver_num_threads=5` when starting pserver processes.
Accrodingly, configure your embedding layers like:
```python
SPARSE_REMOTE=True
w1 = data_layer(name="w1", size=dict_size)
emb1 = embedding_layer(input=w1, size=32, param_attr=ParameterAttribute(sparse_update=SPARSE_REMOTE))
w2 = data_layer(name="w2", size=dict_size)
emb2 = embedding_layer(input=w2, size=32, param_attr=ParameterAttribute(sparse_update=SPARSE_REMOTE))
...
```
## Implementation details
```c++
enum MatType {
MAT_NORMAL,
MAT_NORMAL_SHARED,
MAT_VALUE_SHARED,
MAT_SPARSE_ROW_IDS,
MAT_SPARSE_ROW_AUTO_GROW,
MAT_CACHE_ROW,
MAT_SPARSE_ROW,
MAT_SPARSE_ROW_PREFETCH,
MAT_SPARSE_ROW_PREFETCH_FULL_SIZE,
};
```
`MAT_SPARSE_ROW_PREFETCH` is what we use when configured to fetch only row of matrix when training.
In `trainer_internal.cpp:L93 trainOneBatch`:
```c++
if (config_->getOptConfig().use_sparse_remote_updater()) {
REGISTER_TIMER("prefetch");
gradientMachine_->prefetch(inArgs);
parameterUpdater_->getParametersRemote();
}
```
When doing actual network forward and backward, at the beginning of each batch, the trainer will try to download one row of data from pserver.
In `trainer/RemoteParameterUpdater.cpp`: `parameterUpdater_->getParametersRemote();`:
```c++
if (fullSize) {
...
} else {
getParams = [&] {
parameterClient_->getParameterSparse(
/* recvParameterType= */ PARAMETER_VALUE, sendBackParameterType);
};
applyL1 = [](Parameter& para, real decayRate) {
para.getMat(PARAMETER_VALUE)->applyL1(/*lr=*/1.0f, decayRate);
};
}
```
Calling `parameterClient_->getParameterSparse` will do remote call to pserver's `getParameterSparse`:
```c++
void ParameterServer2::getParameterSparse(const SendParameterRequest& request,
std::vector<Buffer>& inputBuffers,
SendParameterResponse* response,
std::vector<Buffer>* outputBuffers) {
(void)inputBuffers;
auto& buffer = *readWriteBuffer_;
size_t numReals = 0;
for (const auto& block : request.blocks()) {
numReals += getParameterConfig(block).dims(1);
}
buffer.resize(numReals);
VLOG(3) << "pserver: getParameterSparse, numReals=" << numReals;
ReadLockGuard guard(parameterMutex_);
size_t offset = 0;
for (const auto& block : request.blocks()) {
size_t width = getParameterConfig(block).dims(1);
Buffer buf = {buffer.data() + offset, width};
int type = request.send_back_parameter_type();
sendBackParameterSparse(block, type, response, &buf, width, outputBuffers);
offset += width;
}
}
```
`getParameterConfig(block).dims(1)` returns the width of the current "parameter block"(a shard of parameter object),
then `getParameterSparse` remote call returns only one row of data to the client.
......@@ -101,6 +101,7 @@ if use_mkldnn
5.**Argument**里添加两个`MkldnnMatrixPtr`,取名为`mkldnnValue``mkldnnGrad`,用于存放`MkldnnLayer`会用到的memory buffer。 并且添加函数cvt(会修改为一个更加合适的函数名),用于处理"CPU device"和"MKL-DNN device"之间memory的相互转化。
6. 在父类`Layer`中的`getOutput`函数中添加一段逻辑,用于判断`deviceId`,并针对device在MKL-DNN和CPU之间不统一的情况,做一个前期转换。 也就是调用`Argument`的cvt函数把output统一到需要的device上。
7. 在原来的`FLAGS`中添加一个`use_mkldnn`的flag,用于选择是否使用MKL-DNN的相关功能。
8. 关于MKLDNN参数的保存。由于MKLDNN参数的格式与PaddlePaddle原有的格式存在不一样的情况,所以需要在保存参数时同时保存该格式信息。目前准备扩展[Header](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/parameter/Parameter.h#L247)里面的`int32_t version`。这个值不管是在v1还是在v2里面,一直保存的是0,所以可以充分利用这个信息,定义一个枚举处理所有MKLDNN的参数格式,从而`MKLDNNLayer`就可以从输入的参数中获取需要的格式信息。
## References
......
......@@ -68,7 +68,7 @@ As a simple example, consider the following:
1. **BLAS Dependencies(optional)**
CMake will search BLAS libraries from system. If not found, OpenBLAS will be downloaded, built and installed automatically.
CMake will search BLAS libraries from the system. If not found, OpenBLAS will be downloaded, built and installed automatically.
To utilize preinstalled BLAS, you can simply specify MKL, OpenBLAS or ATLAS via `MKL_ROOT`, `OPENBLAS_ROOT` or `ATLAS_ROOT`.
```bash
......@@ -131,9 +131,9 @@ As a simple example, consider the following:
To build GPU version, you will need the following installed:
1. a CUDA-capable GPU
2. A supported version of Linux with a gcc compiler and toolchain
2. A supported version of Linux with a GCC compiler and toolchain
3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn)
4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn)
The CUDA development environment relies on tight integration with the host development environment,
including the host compiler and C runtime libraries, and is therefore only supported on
......@@ -172,6 +172,7 @@ export PATH=<path to install>/bin:$PATH
# install PaddlePaddle Python modules.
sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
```
## <span id="centos">Build on Centos 7</span>
### Install Dependencies
......@@ -192,9 +193,9 @@ sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
To build GPU version, you will need the following installed:
1. a CUDA-capable GPU
2. A supported version of Linux with a gcc compiler and toolchain
2. A supported version of Linux with a GCC compiler and toolchain
3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn)
4. NVIDIA cuDNN Library (available at https://developer.nvidia.com/cudnn)
The CUDA development environment relies on tight integration with the host development environment,
including the host compiler and C runtime libraries, and is therefore only supported on
......@@ -222,7 +223,7 @@ mkdir build && cd build
```
Finally, you can build and install PaddlePaddle:
```bash
# you can add build option here, such as:
cmake3 .. -DCMAKE_INSTALL_PREFIX=<path to install>
......
......@@ -74,13 +74,13 @@ PaddlePaddle发布新版本的时候都会发布对应版本的生产镜像以
.. code-block:: bash
docker run -it --rm paddlepaddle/paddle:0.10.0-dev /bin/bash
docker run -it --rm -v $(pwd):/paddle paddlepaddle/paddle:0.10.0-dev /bin/bash
或者,可以以后台进程方式运行容器:
.. code-block:: bash
docker run -d -p 2202:22 -p 8888:8888 paddledev/paddle:0.10.0-dev
docker run -d -p 2202:22 -p 8888:8888 -v $(pwd):/paddle paddlepaddle/paddle:0.10.0-dev /usr/sbin/sshd -D
然后用密码 :code:`root` SSH进入容器:
......
......@@ -6,14 +6,12 @@
安装流程
++++++++
PaddlePaddle提供数个预编译的二进制来进行安装,包括Docker镜像,ubuntu的deb安装包等。我们推荐使用Docker镜像来部署环境,同时欢迎贡献更多的安装包
PaddlePaddle提供Docker镜像来部署环境
.. toctree::
:maxdepth: 1
docker_install_cn.rst
ubuntu_install_cn.rst
编译流程
......
......@@ -8,14 +8,13 @@ Install PaddlePaddle
:maxdepth: 1
docker_install_en.rst
ubuntu_install_en.rst
Build from Source
-----------------
.. warning::
Please use :code:`deb` package or :code:`docker` image to install paddle. The building guide is used for hacking or contributing PaddlePaddle source code.
Please use :code:`docker` image to install paddle. The building guide is used for hacking or contributing PaddlePaddle source code.
.. toctree::
:maxdepth: 1
......
Ubuntu部署PaddlePaddle
===================================
PaddlePaddle提供了ubuntu 14.04 deb安装包。
安装
------
安装包的下载地址是\: https://github.com/PaddlePaddle/Paddle/releases
它包含四个版本\:
* cpu版本: 支持主流x86处理器平台, 使用了avx指令集。
* cpu-noavx版本:支持主流x86处理器平台,没有使用avx指令集。
* gpu版本:支持主流x86处理器平台,支持nvidia cuda平台,使用了avx指令集。
* gpu-noavx版本:支持主流x86处理器平台,支持nvidia cuda平台,没有使用avx指令集。
下载完相关安装包后,执行:
.. code-block:: shell
sudo apt-get install gdebi
gdebi paddle-*-cpu.deb
或者:
.. code-block:: shell
dpkg -i paddle-*-cpu.deb
apt-get install -f
在 :code:`dpkg -i` 的时候如果报一些依赖未找到的错误是正常的,
在 :code:`apt-get install -f` 里会继续安装 PaddlePaddle。
安装完成后,可以使用命令 :code:`paddle version` 查看安装后的paddle 版本:
.. code-block:: shell
PaddlePaddle 0.8.0b1, compiled with
with_avx: ON
with_gpu: OFF
with_double: OFF
with_python: ON
with_rdma: OFF
with_timer: OFF
with_predict_sdk:
可能遇到的问题
--------------
libcudart.so/libcudnn.so找不到
++++++++++++++++++++++++++++++
安装完成后,运行 :code:`paddle train` 报错\:
.. code-block:: shell
0831 12:36:04.151525 1085 hl_dso_loader.cc:70] Check failed: nullptr != *dso_handle For Gpu version of PaddlePaddle, it couldn't find CUDA library: libcudart.so Please make sure you already specify its path.Note: for training data on Cpu using Gpu version of PaddlePaddle,you must specify libcudart.so via LD_LIBRARY_PATH.
原因是未设置cuda运行时环境变量。 如果使用GPU版本的PaddlePaddle,请安装CUDA 7.5 和CUDNN 5到本地环境中,并设置:
.. code-block:: shell
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:/usr/local/cuda/lib:$LD_LIBRARY_PATH
export PATH=/usr/local/cuda/bin:$PATH
Debian Package installation guide
=================================
PaddlePaddle supports :code:`deb` pacakge. The installation of this :code:`deb` package is tested in ubuntu 14.04, but it should be support other debian based linux, too.
There are four versions of debian package, :code:`cpu`, :code:`gpu`, :code:`cpu-noavx`, :code:`gpu-noavx`. And :code:`noavx` version is used to support CPU which does not contain :code:`AVX` instructions. The download url of :code:`deb` package is \: https://github.com/baidu/Paddle/releases/
After downloading PaddlePaddle deb packages, you can use :code:`gdebi` install.
.. code-block:: bash
gdebi paddle-*.deb
If :code:`gdebi` is not installed, you can use :code:`sudo apt-get install gdebi` to install it.
Or you can use following commands to install PaddlePaddle.
.. code-block:: bash
dpkg -i paddle-*.deb
apt-get install -f
And if you use GPU version deb package, you need to install CUDA toolkit and cuDNN, and set related environment variables(such as LD_LIBRARY_PATH) first. It is normal when `dpkg -i` get errors. `apt-get install -f` will continue install paddle, and install dependences.
# 编译PaddlePaddle和运行单元测试
## 需要的软硬件
为了开发PaddlePaddle,我们需要
1. 一台电脑,可以装的是 Linux, BSD, Windows 或者 MacOS 操作系统,以及
1. Docker。
不需要依赖其他任何软件了。即便是 Python 和 GCC 都不需要,因为我们会把所有编译工具都安装进一个 Docker image 里。
## 总体流程
1. 获取源码
```bash
git clone https://github.com/paddlepaddle/paddle
```
2. 安装开发工具到 Docker image 里
```bash
cd paddle; docker build -t paddle:dev .
```
请注意这个命令结尾处的 `.`;它表示 `docker build` 应该读取当前目录下的 [`Dockerfile`文件](https://github.com/PaddlePaddle/Paddle/blob/develop/Dockerfile),按照其内容创建一个名为 `paddle:dev` 的 Docker image,并且把各种开发工具安装进去。
3. 编译
以下命令启动一个 Docker container 来执行 `paddle:dev` 这个 Docker image,同时把当前目录(源码树根目录)映射为 container 里的 `/paddle` 目录,并且运行 `Dockerfile` 描述的默认入口程序 [`build.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh)。这个脚本调用 `cmake``make` 来编译 `/paddle` 里的源码,结果输出到 `/paddle/build`,也就是本地的源码树根目录里的 `build` 子目录。
```bash
docker run --rm -v $PWD:/paddle paddle:dev
```
上述命令编译出一个 CUDA-enabled 版本。如果我们只需要编译一个只支持 CPU 的版本,可以用
```bash
docker run --rm -e WITH_GPU=OFF -v $PWD:/paddle paddle:dev
```
4. 运行单元测试
用本机的第一个 GPU 来运行包括 GPU 单元测试在内的所有单元测试:
```bash
NV_GPU=0 nvidia-docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest"
```
如果编译的时候我们用了 `WITH_GPU=OFF` 选项,那么编译过程只会产生 CPU-based 单元测试,那么我们也就不需要 nvidia-docker 来运行单元测试了。我们只需要:
```bash
docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest"
```
有时候我们只想运行一个特定的单元测试,比如 `memory_test`,我们可以
```bash
nvidia-docker run --rm -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest -V -R memory_test"
```
5. 清理
有时候我们会希望清理掉已经下载的第三方依赖以及已经编译的二进制文件。此时只需要:
```bash
rm -rf build
```
## 为什么要 Docker 呀?
- 什么是 Docker?
如果您没有听说 Docker,可以把它想象为一个类似 virtualenv 的系统,但是虚拟的不仅仅是 Python 的运行环境。
- Docker 还是虚拟机?
有人用虚拟机来类比 Docker。需要强调的是:Docker 不会虚拟任何硬件,Docker container 里运行的编译工具实际上都是在本机的 CPU 和操作系统上直接运行的,性能和把编译工具安装在本机运行一样。
- 为什么用 Docker?
把工具和配置都安装在一个 Docker image 里可以标准化编译环境。这样如果遇到问题,其他人可以复现问题以便帮助。
另外,对于习惯使用Windows和MacOS的开发者来说,使用Docker就不用配置交叉编译环境了。
- 我可以选择不用Docker吗?
当然可以。大家可以用把开发工具安装进入 Docker image 一样的方式,把这些工具安装到本机。这篇文档介绍基于 Docker 的开发流程,是因为这个流程比其他方法都更简便。
- 学习 Docker 有多难?
理解 Docker 并不难,大概花十分钟看一下[这篇文章](https://zhuanlan.zhihu.com/p/19902938)。这可以帮您省掉花一小时安装和配置各种开发工具,以及切换机器时需要新安装的辛苦。别忘了 PaddlePaddle 更新可能导致需要新的开发工具。更别提简化问题复现带来的好处了。
- 我可以用 IDE 吗?
当然可以,因为源码就在本机上。IDE 默认调用 make 之类的程序来编译源码,我们只需要配置 IDE 来调用 Docker 命令编译源码即可。
很多 PaddlePaddle 开发者使用 Emacs。他们在自己的 `~/.emacs` 配置文件里加两行
```emacs
(global-set-key "\C-cc" 'compile)
(setq compile-command
"docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev")
```
就可以按 `Ctrl-C``c` 键来启动编译了。
- 可以并行编译吗?
是的。我们的 Docker image 运行一个 [Bash 脚本](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh)。这个脚本调用 `make -j$(nproc)` 来启动和 CPU 核一样多的进程来并行编译。
## 可能碰到的问题
- Docker 需要 sudo
如果用自己的电脑开发,自然也就有管理员权限(sudo)了。如果用公用的电脑开发,需要请管理员安装和配置好 Docker。此外,PaddlePaddle 项目在努力开始支持其他不需要 sudo 的集装箱技术,比如 rkt。
- 在 Windows/MacOS 上编译很慢
Docker 在 Windows 和 MacOS 都可以运行。不过实际上是运行在一个 Linux 虚拟机上。可能需要注意给这个虚拟机多分配一些 CPU 和内存,以保证编译高效。具体做法请参考[这个issue](https://github.com/PaddlePaddle/Paddle/issues/627)
- 磁盘不够
本文中的例子里,`docker run` 命令里都用了 `--rm` 参数,这样保证运行结束之后的 containers 不会保留在磁盘上。可以用 `docker ps -a` 命令看到停止后但是没有删除的 containers。`docker build` 命令有时候会产生一些中间结果,是没有名字的 images,也会占用磁盘。可以参考[这篇文章](https://zaiste.net/posts/removing_docker_containers/)来清理这些内容。
# Build PaddlePaddle from Source Code and Run Unit Test
## What Developers Need
To contribute to PaddlePaddle, you need
1. A computer -- Linux, BSD, Windows, MacOS, and
1. Docker.
Nothing else. Not even Python and GCC, because you can install all build tools into a Docker image. We run all the tools by running this image.
## General Process
1. Retrieve source code.
```bash
git clone https://github.com/paddlepaddle/paddle
```
2. Install build tools into a Docker image.
```bash
cd paddle; docker build -t paddle:dev .
```
Please be aware of the `.` at the end of the command, which refers to the [`./Dockerfile` file](https://github.com/PaddlePaddle/Paddle/blob/develop/Dockerfile). `docker build` follows instructions in this file to create a Docker image named `paddle:dev`, and installs building tools into it.
3. Build from source.
This following command starts a Docker container that executes the Docker image `paddle:dev`, mapping the current directory to `/paddle/` in the container, and runs the default entry-point [`build.sh`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh) as specified in the Dockefile. `build.sh` invokes `cmake` and `make` to build PaddlePaddle source code, which had been mapped to `/paddle`, and writes outputs to `/paddle/build`, which maps to `build` in the current source directory on the computer.
```bash
docker run -v $PWD:/paddle paddle:dev
```
Above command builds a CUDA-enabled version. If we want to build a CPU-only version, we can type
```bash
docker run -e WITH_GPU=OFF -v $PWD:/paddle paddle:dev
```
4. Run unit tests.
To run all unit tests using the first GPU of a node:
```bash
NV_GPU=0 nvidia-docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest"
```
If we used `WITH_GPU=OFF` at build time, it generates only CPU-based unit tests, and we don't need nvidia-docker to run them. We can just run
```bash
docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest"
```
Sometimes we want to run a specific unit test, say `memory_test`, we can run
```bash
nvidia-docker run -v $PWD:/paddle paddle:dev bash -c "cd /paddle/build; ctest -V -R memory_test"
```
5. Clean Build.
Sometimes, we might want to clean all thirt-party dependents and built binaries. To do so, just
```bash
rm -rf build
```
## Docker, Or Not?
- What is Docker?
If you haven't heard of it, consider it something like Python's virtualenv.
- Docker or virtual machine?
Some people compare Docker with VMs, but Docker doesn't virtualize any hardware nor running a guest OS, which means there is no compromise on the performance.
- Why Docker?
Using a Docker image of build tools standardizes the building environment, which makes it easier for others to reproduce your problems and to help.
Also, some build tools don't run on Windows or Mac or BSD, but Docker runs almost everywhere, so developers can use whatever computer they want.
- Can I choose not to use Docker?
Sure, you don't have to install build tools into a Docker image; instead, you can install them in your local computer. This document exists because Docker would make the development way easier.
- How difficult is it to learn Docker?
It takes you ten minutes to read [an introductory article](https://docs.docker.com/get-started) and saves you more than one hour to install all required build tools, configure them, especially when new versions of PaddlePaddle require some new tools. Not even to mention the time saved when other people trying to reproduce the issue you have.
- Can I use my favorite IDE?
Yes, of course. The source code resides on your local computer, and you can edit it using whatever editor you like.
Many PaddlePaddle developers are using Emacs. They add the following few lines into their `~/.emacs` configure file:
```emacs
(global-set-key "\C-cc" 'compile)
(setq compile-command
"docker run --rm -it -v $(git rev-parse --show-toplevel):/paddle paddle:dev")
```
so they could type `Ctrl-C` and `c` to build PaddlePaddle from source.
- Does Docker do parallel building?
Our building Docker image runs a [Bash script](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/scripts/docker/build.sh), which calls `make -j$(nproc)` to starts as many processes as the number of your CPU cores.
## Some Gotchas
- Docker requires sudo
An owner of a computer has the administrative privilege, a.k.a., sudo, and Docker requires this privilege to work properly. If you use a shared computer for development, please ask the administrator to install and configure Docker. We will do our best to support rkt, another container technology that doesn't require sudo.
- Docker on Windows/MacOS builds slowly
On Windows and MacOS, Docker containers run in a Linux VM. You might want to give this VM some more memory and CPUs so to make the building efficient. Please refer to [this issue](https://github.com/PaddlePaddle/Paddle/issues/627) for details.
- Not enough disk space
Examples in this article uses option `--rm` with the `docker run` command. This option ensures that stopped containers do not exist on hard disks. We can use `docker ps -a` to list all containers, including stopped. Sometimes `docker build` generates some intermediate dangling images, which also take disk space. To clean them, please refer to [this article](https://zaiste.net/posts/removing_docker_containers/).
# 如何写新的Operator
- [概念简介](#概念简介)
- [实现C++类](#实现C++类)
- [定义ProtoMaker类](#定义ProtoMaker类)
- [定义Operator类](#定义Operator类)
- [定义OpKernel类](#定义OpKernel类)
- [注册Operator](#注册Operator)
- [编译](#编译)
- [绑定Python](#绑定Python)
- [实现单元测试](#实现单元测试)
- [前向Operator单测](#前向Operator单测)
- [反向Operator单测](#反向Operator单测)
- [编译和执行](#编译和执行)
## 概念简介
简单介绍需要用到基类,详细介绍请参考设计文档。
- `framework::OperatorBase`: Operator(简写,Op)基类。
- `framework::OpKernel`: Op计算函数的基类,称作Kernel。
- `framework::OperatorWithKernel`:继承自OperatorBase,Op有计算函数,称作有Kernel。
- `class OpProtoAndCheckerMaker`:描述该Op的输入、输出、属性、注释,主要用于Python API接口生成
依据是否包含kernel,将Op分为两种:包含Kernel的Op和不包含kernel的Op,前者Op的定义继承自`OperatorBase`,后者继承自`OperatorWithKernel`。本教程主要介绍带Kernel的Op如何写,简单总结Op需要包含的内容如下:
内容 | 定义位置
-------------- | :----------------------
OpProtoMake定义 | `.cc`文件,Backward Op不需要定义OpProtoMake
Op定义 | `.cc`文件
Kernel实现 | CPU、GPU共享Kernel在`.h`文件,否则,CPU可以在`.cc`文件,GPU可在`.cu`文件。
注册Op | Op注册在`.cc`文件;Kernel注册CPU在`.cc`文件,GPU在`.cu`文件
下面以矩阵乘操作,即[MulOp](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc)为例来介绍如何写带Kernel的Operator。
## 实现C++类
### 1. 定义ProtoMaker类
矩阵乘的公式:$Out = X * Y$, 可见该计算由两个输入,一个输出组成。首先定义`ProtoMaker`来描述该Op的输入、输出及注释:
```
class MulOpMaker : public framework::OpProtoAndCheckerMaker {
public:
MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "The first input of mul op");
AddInput("Y", "The second input of mul op");
AddOutput("Out", "The output of mul op");
AddComment(R"DOC(
Two Element Mul Operator.
The equation is: Out = X * Y
)DOC");
}
};
```
[`MulOpMaker`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L43)继承自`framework::OpProtoAndCheckerMaker`,构造函数包括2个:
- `framework::OpProto` : 前者存储Op的输入输出和参数属性,将用于Python API接口的生成。
- `framework::OpAttrChecker` :后者用于检查参数属性的合法性。
构造函数里通过`AddInput`添加输入参数,通过`AddOutput`添加输出参数,通过`AddComment`添加该Op的注释,这些函数会将对应内容添加到`OpProto`中。
`MulOp`中添加两个输入`X``Y`,添加了一个输出`Out`,并解释了各自含义,该命名尽可能的规范。
再举个[`ScaleOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/scale_op.cc#L37)的例子:
```
template <typename AttrType>
class ScaleOpMaker : public framework::OpProtoAndCheckerMaker {
public:
ScaleOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "The input tensor of scale operator.").NotInGradient();
AddOutput("Out", "The output tensor of scale operator.").NotInGradient();
AddComment(R"DOC(Scale operator
The equation is: Out = scale*X
)DOC");
AddAttr<AttrType>("scale", "scale of scale operator.").SetDefault(1.0);
}
};
```
在这个例子里,两处不同:
- `AddInput("X","...").NotInGradient()` : 表示`X`这个输入不参与`ScaleOp`对应的梯度Op计算之中。
- `AddAttr<AttrType>("scale", "...").SetDefault(1.0);` : 增加`scale`系数,作为参数属性,并且设置默认值为1.0。
### 2. 定义Operator类
```c++
class MulOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(const framework::InferShapeContext &ctx) const override {
auto dim0 = ctx.Input<Tensor>("X")->dims();
auto dim1 = ctx.Input<Tensor>("Y")->dims();
PADDLE_ENFORCE_EQ(dim0.size(), 2,
"input X(%s) should be a tensor with 2 dims, a matrix",
ctx.op_.Input("X"));
PADDLE_ENFORCE_EQ(dim1.size(), 2,
"input Y(%s) should be a tensor with 2 dims, a matrix",
ctx.op_.Input("Y"));
PADDLE_ENFORCE_EQ(
dim0[1], dim1[0],
"First matrix's width must be equal with second matrix's height.");
ctx.Output<Tensor>("Out")->Resize({dim0[0], dim1[1]});
}
};
```
[`MulOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/mul_op.cc#L22)继承自`OperatorWithKernel``public`成员:
```c++
using framework::OperatorWithKernel::OperatorWithKernel;
```
这句表示使用基类`OperatorWithKernel`的构造函数,也可写成:
```c++
MulOp(const std::string &type, const framework::VariableNameMap &inputs,
const framework::VariableNameMap &outputs,
const framework::AttributeMap &attrs)
: OperatorWithKernel(type, inputs, outputs, attrs) {}
```
还需要重写`InferShape`接口。`InferShape`为const函数,不能修改Op的成员变量,参数为`const framework::InferShapeContext &ctx`,通过该参数可获取到输入输出以及属性。它的功能是:
- 1). 做检查, 尽早报错:检查输入数据维度、类型等是否合法。
- 2). 设置输出Tensor的形状。
通常`OpProtoMaker``Op`类的定义写在`.cc`文件中,和要讲到的注册函数一起放在`.cc`
### 3. 定义OpKernel类
```C++
template <typename Place, typename T>
class MulKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto* X = context.Input<Tensor>("X");
auto* Y = context.Input<Tensor>("Y");
auto* Z = context.Output<Tensor>("Out");
Z->mutable_data<T>(context.GetPlace());
auto* device_context =
const_cast<platform::DeviceContext*>(context.device_context_);
math::matmul<Place, T>(*X, false, *Y, false, 1, Z, 0, device_context);
}
};
```
`MulKernel`继承自`framework::OpKernel`,带有模板参数:
- `typename Place`: 表示设备类型,不同设备(CPU、GPU)共享同一个Kernel时,需加该模板参数,不共享则不加,一个不共享的例子是[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)
- `typename T` : 表示数据类型,如`float`, `double`等。
`MulKernel`需要重写`Compute`接口,该接口参数为`const framework::ExecutionContext& context`, `ExecutionContext`相比`InferShapeContext`增加了设备类型,同样可获取到输入输出和属性参数,`Compute`函数里写具体实现时。
注意,不同设备(CPU、GPU)共享一个Op定义,是否则共享同一个`OpKernel`,取决于`Compute`调用的函数是否支持不同设备。`MulOp`的CPU、GPU实现共享同一个`Kernel``OpKernel`不共享的例子可以参考[`OnehotCrossEntropyOpKernel`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/cross_entropy_op.h#L43)
到此前向Op实现完成,需要在`.cc`文件中注册该op和kernel。反向Op类的定义和Kernel定义与前向Op类似,这里不再重复。但注意,反向Op没有`ProtoMaker`
### 4. 注册Operator
`.cc`文件中注册前向、反向Op类,注册CPU Kernel。
```c++
namespace ops = paddle::operators;
REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(mul_grad,
ops::MulGradKernel<paddle::platform::CPUPlace, float>);
```
- `REGISTER_OP` : 注册`ops::MulOp`类,类型名为`mul`,该类的`ProtoMaker``ops::MulOpMaker`,注册`ops::MulOpGrad`,类型名为`mul_grad`
- `REGISTER_OP_WITHOUT_GRADIENT` : 用于注册没有反向的Op。
- `REGISTER_OP_CPU_KERNEL` :注册`ops::MulKernel`类,并特化模板参数为`paddle::platform::CPUPlace``float`类型,同理,注册`ops::MulKernel`类。
`.cu`文件中注册GPU Kernel。
```c++
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(mul_grad,
ops::MulGradKernel<paddle::platform::GPUPlace, float>);
```
### 5. 编译
[paddle/operators/CMakeLists.txt](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/CMakeLists.txt)文件中添加编译。
```
op_library(mul_op SRCS mul_op.cc mul_op.cu DEPS math_function)
```
下面命令可以编译:
```
make mul_op
```
## 绑定Python
- 绑定Python
[`paddle/pybind/pybind.cc
`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/pybind.cc)文件中添加该类:
```
USE_OP(mul);
```
如果只实现了CPU版本,则使用`USE_CPU_ONLY_OP`:
```
USE_CPU_ONLY_OP(gather);
```
使用`USE_OP`告知编译器需要链接该Op的目标文件,具体解释参考[代码注释](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。
- 生成库
[`paddle/pybind/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/pybind/CMakeLists.txt)文件添加类到`DEPS`中,使得该Op可以链接到生成的lib库中。
```
if(WITH_PYTHON)
cc_library(paddle_pybind SHARED
SRCS pybind.cc
DEPS pybind python backward
mul_op
minus_op)
endif(WITH_PYTHON)
```
## 实现单元测试
单测包括对比前向Op不同设备(CPU、GPU)的实现、对比反向OP不同设备(CPU、GPU)的实现、反向Op的梯度测试。下面介绍介绍[`MulOp`的单测](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py)
### 前向Operator单测
前向Op单测继承自`unittest.TestCase`,并定义元类`__metaclass__ = OpTestMeta`,具体单测流程在`OpTestMeta`里完成。需在`setUp`函数定义输入输出和属性参数,以及Python对比的输出值。
```
import unittest
import numpy as np
from gradient_checker import GradientChecker, create_op
from op_test_util import OpTestMeta
class TestMulOp(unittest.TestCase):
__metaclass__ = OpTestMeta
def setUp(self):
self.type = "mul"
self.inputs = {
'X': np.random.random((32, 84)).astype("float32"),
'Y': np.random.random((84, 100)).astype("float32")
}
self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
```
首先需要`import`必要的包,下面详细解释其他值:
- `self.type = "mul" ` : 定义类型,和注册的类型一致。
- `self.inputs` : 定义输入,类型为Numpy.array,并初始化。
- `self.outputs` : 定义输出,并得到Python结算结果。
### 反向Operator单测
反向Op单测继承自`GradientChecker`,而`GradientChecker`集成自`unittest.TestCase`,所以反向单测函数需要`test_`开头。
```
class MulGradOpTest(GradientChecker):
def test_mul(self):
op = create_op("mul")
inputs = {
'X': np.random.random((32, 84)).astype("float32"),
'Y': np.random.random((84, 100)).astype("float32")
}
self.compare_grad(op, inputs)
# mul op will enlarge the relative error
self.check_grad(
op, inputs, set(["X", "Y"]), "Out", max_relative_error=0.5)
```
- 调用`create_op("mul")`创建反向Op对应的前向Op。
- 定义输入`inputs`
- 调用`compare_grad`函数对比CPU、GPU计算结果。
- 调用`check_grad`检查梯度稳定性,这里采用数值法检测梯度正确性。
- 第一个参数`op` : 前向op。
- 第二个参数`inputs` : 输入词典,词典的Key和`ProtoMaker`定义保持一致。
- 第三个参数`set(["X", "Y"])` : 指定对输入变量`X``Y`做梯度检测。
- 第四个参数`"Out"` : 指定前向网络最终的输出目标变量`Out`
### 编译和执行
单测完成之后,在[`python/paddle/v2/framework/tests/CMakeLists.txt`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/CMakeLists.txt)里添加编译:
```
py_test(test_mul_op SRCS test_mul_op.py)
```
编译时需要打开`WITH_TESTING`, 即 `cmake paddle_dir -DWITH_TESTING=ON`,编译成功之后执行单测命令为:
```
make test ARGS="-R test_mul_op -V"
```
或者:
```
ctest -R test_mul_op
```
......@@ -19,6 +19,7 @@
.. toctree::
:maxdepth: 1
dev/build_cn.rst
dev/write_docs_cn.rst
dev/contribute_to_paddle_cn.md
......
......@@ -18,6 +18,7 @@ Development
.. toctree::
:maxdepth: 1
dev/build_en.rst
dev/new_layer_en.rst
dev/contribute_to_paddle_en.md
......
......@@ -7,4 +7,3 @@ PaddlePaddle Documentation
getstarted/index_en.rst
howto/index_en.rst
api/index_en.rst
about/index_en.rst
......@@ -13,7 +13,7 @@
# serve to show the default.
import sys
import os, subprocess
sys.path.insert(0, os.path.abspath('@PROJ_ROOT@/python'))
sys.path.insert(0, os.path.abspath('@PADDLE_SOURCE_DIR@/python'))
import shlex
from recommonmark import parser, transform
import paddle
......@@ -24,7 +24,7 @@ AutoStructify = transform.AutoStructify
# If extensions (or modules to document with autodoc) are in another directory,
# add these directories to sys.path here. If the directory is relative to the
# documentation root, use os.path.abspath to make it absolute, like shown here.
templates_path = ["@PROJ_ROOT@/doc_theme/templates"]
templates_path = ["@PADDLE_SOURCE_DIR@/doc_theme/templates"]
# -- General configuration ------------------------------------------------
......@@ -120,7 +120,7 @@ html_theme = 'sphinx_rtd_theme'
# Add any paths that contain custom static files (such as style sheets) here,
# relative to this directory. They are copied after the builtin static files,
# so a file named "default.css" will overwrite the builtin "default.css".
html_static_path = ['@PROJ_ROOT@/doc_theme/static']
html_static_path = ['@PADDLE_SOURCE_DIR@/doc_theme/static']
# Output file base name for HTML help builder.
htmlhelp_basename = project + 'doc'
......
......@@ -13,7 +13,7 @@
# serve to show the default.
import sys
import os, subprocess
sys.path.insert(0, os.path.abspath('@PROJ_ROOT@/python'))
sys.path.insert(0, os.path.abspath('@PADDLE_SOURCE_DIR@/python'))
import shlex
from recommonmark import parser, transform
import paddle
......@@ -25,7 +25,7 @@ AutoStructify = transform.AutoStructify
# If extensions (or modules to document with autodoc) are in another directory,
# add these directories to sys.path here. If the directory is relative to the
# documentation root, use os.path.abspath to make it absolute, like shown here.
templates_path = ["@PROJ_ROOT@/doc_theme/templates"]
templates_path = ["@PADDLE_SOURCE_DIR@/doc_theme/templates"]
# -- General configuration ------------------------------------------------
......@@ -120,7 +120,7 @@ html_theme = 'sphinx_rtd_theme'
# Add any paths that contain custom static files (such as style sheets) here,
# relative to this directory. They are copied after the builtin static files,
# so a file named "default.css" will overwrite the builtin "default.css".
html_static_path = ['@PROJ_ROOT@/doc_theme/static']
html_static_path = ['@PADDLE_SOURCE_DIR@/doc_theme/static']
# Output file base name for HTML help builder.
htmlhelp_basename = project + 'doc'
......
......@@ -63,13 +63,24 @@ func WithAddr(addr string) func(c *Client) error {
// WithEtcd sets the client to use etcd for master discovery.
func WithEtcd(endpoints []string, timeout time.Duration) func(*Client) error {
return func(c *Client) error {
cli, err := clientv3.New(clientv3.Config{
Endpoints: endpoints,
DialTimeout: timeout,
})
if err != nil {
var cli *clientv3.Client
f := func() error {
var err error
cli, err = clientv3.New(clientv3.Config{
Endpoints: endpoints,
DialTimeout: timeout,
})
return err
}
for {
err := f()
if err != nil {
log.Warningln(err)
} else {
break
}
time.Sleep(time.Second)
}
ch := make(chan string, 1)
a, err := GetKey(cli, DefaultAddrPath, timeout)
......@@ -101,9 +112,6 @@ func NewClient(opts ...func(*Client) error) (*Client, error) {
}
}
c.ch = make(chan record, c.bufSize)
// FIXME: connection is created asyncrosly in monitorMaster go routine,
// ensure the connection is ready for use before calling c.addClient.
time.Sleep(time.Second)
return c, nil
}
......
......@@ -17,12 +17,10 @@ def main():
# network config
x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(13))
y_predict = paddle.layer.fc(input=x,
param_attr=paddle.attr.Param(
name='w', learning_rate=1e-3),
param_attr=paddle.attr.Param(name='w'),
size=1,
act=paddle.activation.Linear(),
bias_attr=paddle.attr.Param(
name='b', learning_rate=1e-3))
bias_attr=paddle.attr.Param(name='b'))
y = paddle.layer.data(name='y', type=paddle.data_type.dense_vector(1))
cost = paddle.layer.mse_cost(input=y_predict, label=y)
......
......@@ -15,6 +15,7 @@ if(Boost_FOUND)
add_subdirectory(platform)
add_subdirectory(framework)
add_subdirectory(operators)
add_subdirectory(pybind)
endif()
if(WITH_C_API)
......
......@@ -19,9 +19,9 @@ add_library(paddle_api STATIC ${API_SOURCES})
add_dependencies(paddle_api paddle_proto paddle_trainer_lib)
INCLUDE(${SWIG_USE_FILE})
INCLUDE_DIRECTORIES(${PROJ_ROOT}/paddle)
INCLUDE_DIRECTORIES(${PADDLE_SOURCE_DIR}/paddle)
FILE(GLOB PY_PADDLE_PYTHON_FILES ${PROJ_ROOT}/paddle/py_paddle/*.py)
FILE(GLOB PY_PADDLE_PYTHON_FILES ${PADDLE_SOURCE_DIR}/paddle/py_paddle/*.py)
SET_SOURCE_FILES_PROPERTIES(Paddle.i PROPERTIES CPLUSPLUS ON)
......@@ -79,16 +79,16 @@ SWIG_LINK_LIBRARIES(swig_paddle
${START_END}
)
add_custom_command(OUTPUT ${PROJ_ROOT}/paddle/py_paddle/_swig_paddle.so
COMMAND cp ${CMAKE_CURRENT_BINARY_DIR}/swig_paddle.py ${PROJ_ROOT}/paddle/py_paddle
COMMAND cp ${CMAKE_CURRENT_BINARY_DIR}/_swig_paddle.so ${PROJ_ROOT}/paddle/py_paddle
add_custom_command(OUTPUT ${PADDLE_SOURCE_DIR}/paddle/py_paddle/_swig_paddle.so
COMMAND cp ${CMAKE_CURRENT_BINARY_DIR}/swig_paddle.py ${PADDLE_SOURCE_DIR}/paddle/py_paddle
COMMAND cp ${CMAKE_CURRENT_BINARY_DIR}/_swig_paddle.so ${PADDLE_SOURCE_DIR}/paddle/py_paddle
COMMAND ${CMAKE_COMMAND} -E touch .timestamp
WORKING_DIRECTORY ${PROJ_ROOT}/paddle
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle
DEPENDS _swig_paddle
)
# TODO(yuyang18) : make wheel name calculated by cmake
add_custom_target(python_api_wheel ALL DEPENDS ${PROJ_ROOT}/paddle/py_paddle/_swig_paddle.so)
add_custom_target(python_api_wheel ALL DEPENDS ${PADDLE_SOURCE_DIR}/paddle/py_paddle/_swig_paddle.so)
if(WITH_TESTING)
IF(NOT PY_PIP_FOUND)
......
......@@ -41,7 +41,7 @@ ParameterUpdater *ParameterUpdater::createNewRemoteUpdater(
config->m->getConfig(), pserverSpec, useEtcd));
return updater;
#else
throw UnsupportError();
throw UnsupportError("not compiled with WITH_GOLANG");
#endif
}
......
......@@ -90,6 +90,18 @@ paddle_error paddle_arguments_set_ids(paddle_arguments args,
return kPD_NO_ERROR;
}
paddle_error paddle_arguments_set_frame_shape(paddle_arguments args,
uint64_t ID,
uint64_t frameHeight,
uint64_t frameWidth) {
if (args == nullptr) return kPD_NULLPTR;
auto a = castArg(args);
if (ID >= a->args.size()) return kPD_OUT_OF_RANGE;
a->args[ID].setFrameHeight(frameHeight);
a->args[ID].setFrameWidth(frameWidth);
return kPD_NO_ERROR;
}
paddle_error paddle_arguments_set_sequence_start_pos(paddle_arguments args,
uint64_t ID,
uint32_t nestedLevel,
......
......@@ -53,7 +53,10 @@ add_custom_target(paddle_capi_whole ALL
set_target_properties(paddle_capi_whole
PROPERTIES IMPORTED_LOCATION ${CMAKE_CURRENT_BINARY_DIR}/${capi_whole_library})
set(LINK_FLAGS " -Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/export.sym -Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/export.map")
# TODO: merge mkl into paddle_capi_shared
add_library(paddle_capi_shared SHARED ${CAPI_SOURCES})
set_target_properties(paddle_capi_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
target_include_directories(paddle_capi_shared PUBLIC ${CMAKE_CURRENT_BINARY_DIR})
link_paddle_exe(paddle_capi_shared)
......
......@@ -111,6 +111,20 @@ PD_API paddle_error paddle_arguments_set_ids(paddle_arguments args,
uint64_t ID,
paddle_ivector ids);
/**
* @brief paddle_arguments_set_frame_shape Set the fram size of one argument
* in array, which index is `ID`.
* @param [in] args arguments array
* @param [in] ID array index
* @param [in] frameHeight maximum height of input images
* @param [in] frameWidth maximum width of input images
* @return paddle_error
*/
PD_API paddle_error paddle_arguments_set_frame_shape(paddle_arguments args,
uint64_t ID,
uint64_t frameHeight,
uint64_t frameWidth);
/**
* @brief PDArgsSetSequenceStartPos Set sequence start position vector of one
* argument in array, which index is `ID`.
......
......@@ -3,18 +3,21 @@
#include <stdio.h>
#include <stdlib.h>
#define CHECK(stmt) \
do { \
paddle_error __err__ = stmt; \
if (__err__ != kPD_NO_ERROR) { \
fprintf(stderr, "Invoke paddle error %d \n" #stmt, __err__); \
exit(__err__); \
} \
#define CHECK(stmt) \
do { \
paddle_error __err__ = stmt; \
if (__err__ != kPD_NO_ERROR) { \
fprintf(stderr, "Invoke paddle error %d in " #stmt "\n", __err__); \
exit(__err__); \
} \
} while (0)
void* read_config(const char* filename, long* size) {
FILE* file = fopen(filename, "r");
if (file == NULL) return NULL;
if (file == NULL) {
fprintf(stderr, "Open %s error\n", filename);
return NULL;
}
fseek(file, 0L, SEEK_END);
*size = ftell(file);
fseek(file, 0L, SEEK_SET);
......
{
global:
paddle_*;
local:
*;
};
......@@ -54,6 +54,31 @@ paddle_error paddle_gradient_machine_create_for_inference(
return kPD_NO_ERROR;
}
paddle_error paddle_gradient_machine_create_for_inference_with_parameters(
paddle_gradient_machine* machine, void* mergedModel, uint64_t size) {
if (mergedModel == nullptr) return kPD_NULLPTR;
std::istringstream is(std::string(static_cast<char*>(mergedModel), size));
int64_t modelConfigSize = 0;
is.read((char*)(&modelConfigSize), sizeof(modelConfigSize));
std::string modelConfigProtobuf;
modelConfigProtobuf.resize(modelConfigSize);
is.read(&modelConfigProtobuf[0], modelConfigSize);
paddle::TrainerConfig config;
if (!config.ParseFromString(modelConfigProtobuf) || !config.IsInitialized()) {
return kPD_PROTOBUF_ERROR;
}
auto ptr = new paddle::capi::CGradientMachine();
ptr->machine.reset(paddle::GradientMachine::create(
config.model_config(), CREATE_MODE_TESTING, {paddle::PARAMETER_VALUE}));
std::vector<paddle::ParameterPtr>& parameters = ptr->machine->getParameters();
for (auto& para : parameters) {
para->load(is);
}
*machine = ptr;
return kPD_NO_ERROR;
}
paddle_error paddle_gradient_machine_destroy(paddle_gradient_machine machine) {
delete cast(machine);
return kPD_NO_ERROR;
......@@ -121,3 +146,19 @@ paddle_error paddle_gradient_machine_randomize_param(
m->machine->randParameters();
return kPD_NO_ERROR;
}
paddle_error paddle_gradient_machine_get_layer_output(
paddle_gradient_machine machine,
const char* layerName,
paddle_arguments args) {
auto m = cast(machine);
auto out = paddle::capi::cast<paddle::capi::CArguments>(args);
if (m == nullptr || layerName == nullptr || out == nullptr ||
m->machine == nullptr) {
return kPD_NULLPTR;
}
auto layerOutput = m->machine->getLayerOutput(layerName);
out->args.push_back(layerOutput);
return kPD_NO_ERROR;
}
......@@ -36,6 +36,22 @@ typedef void* paddle_gradient_machine;
PD_API paddle_error paddle_gradient_machine_create_for_inference(
paddle_gradient_machine* machine, void* modelConfigProtobuf, int size);
/**
* @brief Create a gradient machine used for model inference, using config with
* parameters which is generated by `paddle merge_model`.
* Example:
* paddle merge_model \
* --model_dir="pass-00000" \
* --model_file="merged_model.paddle"
* @param [out] machine that used for model inference
* @param [in] mergedModel
* @param [in] size
* @return paddle_error
*/
PD_API paddle_error
paddle_gradient_machine_create_for_inference_with_parameters(
paddle_gradient_machine* machine, void* mergedModel, uint64_t size);
/**
* @brief Load parameter from disk.
* @param machine Gradient Machine.
......@@ -85,6 +101,18 @@ paddle_gradient_machine_randomize_param(paddle_gradient_machine machine);
PD_API paddle_error
paddle_gradient_machine_destroy(paddle_gradient_machine machine);
/**
* @brief Get the output of the layer named `layerName`.
* @param [in] gradient machine that have run a inference
* @param [in] layerName name of specified layer
* @param [out] args output of the specified layer
* @return paddle_error
*/
PD_API paddle_error
paddle_gradient_machine_get_layer_output(paddle_gradient_machine machine,
const char* layerName,
paddle_arguments args);
#ifdef __cplusplus
}
#endif
......
......@@ -10,5 +10,5 @@ target_include_directories(capi_test_gradientMachine PUBLIC
${PADDLE_CAPI_INC_PATH})
target_link_libraries(capi_test_gradientMachine paddle_capi)
add_test(NAME capi_test_gradientMachine
COMMAND ${PROJ_ROOT}/paddle/.set_python_path.sh -d ${PROJ_ROOT}/python ${CMAKE_CURRENT_BINARY_DIR}/capi_test_gradientMachine
WORKING_DIRECTORY ${PROJ_ROOT}/paddle/capi/tests)
COMMAND ${PADDLE_SOURCE_DIR}/paddle/.set_python_path.sh -d ${PADDLE_SOURCE_DIR}/python ${CMAKE_CURRENT_BINARY_DIR}/capi_test_gradientMachine
WORKING_DIRECTORY ${PADDLE_SOURCE_DIR}/paddle/capi/tests)
......@@ -214,7 +214,8 @@ extern void hl_conv_workspace(hl_tensor_descriptor input,
int* convBwdDataAlgo,
size_t* bwdDataLimitBytes,
int* convBwdFilterAlgo,
size_t* bwdFilterLimitBytes);
size_t* bwdFilterLimitBytes,
bool useDilation);
/**
* @brief destroy filter descriptor.
......@@ -242,7 +243,9 @@ extern void hl_create_convolution_descriptor(hl_convolution_descriptor* conv,
int padding_height,
int padding_width,
int stride_height,
int stride_width);
int stride_width,
int dilation_h = 1,
int dilation_w = 1);
/**
* @brief reset convolution descriptor.
......@@ -262,7 +265,9 @@ extern void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
int padding_height,
int padding_width,
int stride_height,
int stride_width);
int stride_width,
int dilation_h = 1,
int dilation_w = 1);
/**
* @brief destroy convolution descriptor.
......
......@@ -78,7 +78,9 @@ inline void hl_create_convolution_descriptor(hl_convolution_descriptor* conv,
int padding_height,
int padding_width,
int stride_height,
int stride_width) {}
int stride_width,
int dilation_h,
int dilation_w) {}
inline void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
hl_tensor_descriptor image,
......@@ -86,7 +88,9 @@ inline void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
int padding_height,
int padding_width,
int stride_height,
int stride_width) {}
int stride_width,
int dilation_h,
int dilation_w) {}
inline void hl_destroy_convolution_descriptor(hl_convolution_descriptor conv) {}
......@@ -99,7 +103,8 @@ inline void hl_conv_workspace(hl_tensor_descriptor input,
int* convBwdDataAlgo,
size_t* bwdDataLimitBytes,
int* convBwdFilterAlgo,
size_t* bwdFilterLimitBytes) {}
size_t* bwdFilterLimitBytes,
bool useDilation) {}
inline void hl_convolution_forward(hl_tensor_descriptor input,
real* input_data,
......
......@@ -201,7 +201,8 @@ void hl_conv_workspace(hl_tensor_descriptor input,
int* convBwdDataAlgo,
size_t* bwdDataLimitBytes,
int* convBwdFilterAlgo,
size_t* bwdFilterLimitBytes) {
size_t* bwdFilterLimitBytes,
bool useDilation) {
#if CUDNN_VERSION >= 4000
CHECK_NOTNULL(input);
......@@ -213,21 +214,60 @@ void hl_conv_workspace(hl_tensor_descriptor input,
size_t memoryLimitBytes =
(1LL << 20) * FLAGS_cudnn_conv_workspace_limit_in_mb;
// For dilation
int algo = 0;
// cudnn convolution forward configuration
cudnnTensorDescriptor_t fwd_src_desc = GET_TENSOR_DESCRIPTOR(input);
cudnnTensorDescriptor_t fwd_dest_desc = GET_TENSOR_DESCRIPTOR(output);
cudnnFilterDescriptor_t fwd_filter_desc = GET_FILTER_DESCRIPTOR(filter);
cudnnConvolutionDescriptor_t fwd_conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv);
// cudnn convolution backward data configuration
cudnnFilterDescriptor_t bwd_data_filter_desc = GET_FILTER_DESCRIPTOR(filter);
cudnnTensorDescriptor_t bwd_data_diff_desc = GET_TENSOR_DESCRIPTOR(output);
cudnnTensorDescriptor_t bwd_data_grad_desc = GET_TENSOR_DESCRIPTOR(input);
cudnnConvolutionDescriptor_t bwd_data_conv_desc =
GET_CONVOLUTION_DESCRIPTOR(conv);
// cudnn convolution backward filter configuration
cudnnTensorDescriptor_t bwd_filter_src_desc = GET_TENSOR_DESCRIPTOR(input);
cudnnTensorDescriptor_t bwd_filter_diff_desc = GET_TENSOR_DESCRIPTOR(output);
cudnnConvolutionDescriptor_t bwd_filter_conv_desc =
GET_CONVOLUTION_DESCRIPTOR(conv);
cudnnFilterDescriptor_t bwd_filter_grad_desc = GET_FILTER_DESCRIPTOR(filter);
CHECK_CUDNN(dynload::cudnnGetConvolutionForwardAlgorithm(
t_resource.cudnn_handle,
fwd_src_desc,
fwd_filter_desc,
fwd_conv_desc,
fwd_dest_desc,
CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
memoryLimitBytes,
reinterpret_cast<cudnnConvolutionFwdAlgo_t*>(convFwdAlgo)));
if (useDilation) {
convFwdAlgo = &algo;
convBwdDataAlgo = &algo;
convBwdFilterAlgo = &algo;
} else {
CHECK_CUDNN(dynload::cudnnGetConvolutionForwardAlgorithm(
t_resource.cudnn_handle,
fwd_src_desc,
fwd_filter_desc,
fwd_conv_desc,
fwd_dest_desc,
CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
memoryLimitBytes,
reinterpret_cast<cudnnConvolutionFwdAlgo_t*>(convFwdAlgo)));
CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataAlgorithm(
t_resource.cudnn_handle,
bwd_data_filter_desc,
bwd_data_diff_desc,
bwd_data_conv_desc,
bwd_data_grad_desc,
CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
memoryLimitBytes,
reinterpret_cast<cudnnConvolutionBwdDataAlgo_t*>(convBwdDataAlgo)));
CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
t_resource.cudnn_handle,
bwd_filter_src_desc,
bwd_filter_diff_desc,
bwd_filter_conv_desc,
bwd_filter_grad_desc,
CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
memoryLimitBytes,
reinterpret_cast<cudnnConvolutionBwdFilterAlgo_t*>(convBwdFilterAlgo)));
}
CHECK_CUDNN(dynload::cudnnGetConvolutionForwardWorkspaceSize(
t_resource.cudnn_handle,
......@@ -238,23 +278,6 @@ void hl_conv_workspace(hl_tensor_descriptor input,
static_cast<cudnnConvolutionFwdAlgo_t>(*convFwdAlgo),
fwdLimitBytes));
// cudnn convolution backward data configuration
cudnnFilterDescriptor_t bwd_data_filter_desc = GET_FILTER_DESCRIPTOR(filter);
cudnnTensorDescriptor_t bwd_data_diff_desc = GET_TENSOR_DESCRIPTOR(output);
cudnnTensorDescriptor_t bwd_data_grad_desc = GET_TENSOR_DESCRIPTOR(input);
cudnnConvolutionDescriptor_t bwd_data_conv_desc =
GET_CONVOLUTION_DESCRIPTOR(conv);
CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataAlgorithm(
t_resource.cudnn_handle,
bwd_data_filter_desc,
bwd_data_diff_desc,
bwd_data_conv_desc,
bwd_data_grad_desc,
CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
memoryLimitBytes,
reinterpret_cast<cudnnConvolutionBwdDataAlgo_t*>(convBwdDataAlgo)));
CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
t_resource.cudnn_handle,
bwd_data_filter_desc,
......@@ -264,23 +287,6 @@ void hl_conv_workspace(hl_tensor_descriptor input,
static_cast<cudnnConvolutionBwdDataAlgo_t>(*convBwdDataAlgo),
bwdDataLimitBytes));
// cudnn convolution backward filter configuration
cudnnTensorDescriptor_t bwd_filter_src_desc = GET_TENSOR_DESCRIPTOR(input);
cudnnTensorDescriptor_t bwd_filter_diff_desc = GET_TENSOR_DESCRIPTOR(output);
cudnnConvolutionDescriptor_t bwd_filter_conv_desc =
GET_CONVOLUTION_DESCRIPTOR(conv);
cudnnFilterDescriptor_t bwd_filter_grad_desc = GET_FILTER_DESCRIPTOR(filter);
CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
t_resource.cudnn_handle,
bwd_filter_src_desc,
bwd_filter_diff_desc,
bwd_filter_conv_desc,
bwd_filter_grad_desc,
CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
memoryLimitBytes,
reinterpret_cast<cudnnConvolutionBwdFilterAlgo_t*>(convBwdFilterAlgo)));
CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize(
t_resource.cudnn_handle,
bwd_filter_src_desc,
......@@ -603,7 +609,9 @@ void hl_create_convolution_descriptor(hl_convolution_descriptor* conv,
int padding_height,
int padding_width,
int stride_height,
int stride_width) {
int stride_width,
int dilation_h,
int dilation_w) {
CHECK_NOTNULL(conv);
cudnn_convolution_descriptor hl_conv = (cudnn_convolution_descriptor)malloc(
......@@ -625,18 +633,24 @@ void hl_create_convolution_descriptor(hl_convolution_descriptor* conv,
padding_width,
stride_height,
stride_width,
1,
1,
dilation_h,
dilation_w,
mode,
data_type));
#else
if (dilation_h > 1 || dilation_w > 1) {
LOG(FATAL)
<< "Current cuDNN version does't support for dilation convolution. "
<< "The dilation convolution requires cuDNN >= v6.0.";
}
CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(hl_conv->desc,
padding_height,
padding_width,
stride_height,
stride_width,
1,
1,
dilation_h,
dilation_w,
mode));
#endif
......@@ -659,7 +673,9 @@ void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
int padding_height,
int padding_width,
int stride_height,
int stride_width) {
int stride_width,
int dilation_h,
int dilation_w) {
CHECK_NOTNULL(conv);
CHECK_NOTNULL(image);
CHECK_NOTNULL(filter);
......@@ -678,8 +694,8 @@ void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
padding_width,
stride_height,
stride_width,
1,
1,
dilation_h,
dilation_w,
mode,
data_type));
#else
......@@ -688,8 +704,8 @@ void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
padding_width,
stride_height,
stride_width,
1,
1,
dilation_h,
dilation_w,
mode));
#endif
......
......@@ -7,7 +7,7 @@ cc_library(tensor SRCS tensor.cc DEPS ddim place paddle_memory device_context)
cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
cc_library(lod_tensor SRCS lod_tensor.cc details/lod_tensor.cc DEPS ddim place tensor)
cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor)
cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor)
cc_test(variable_test SRCS variable_test.cc)
......@@ -15,40 +15,27 @@ cc_test(variable_test SRCS variable_test.cc)
cc_library(scope SRCS scope.cc)
cc_test(scope_test SRCS scope_test.cc DEPS scope)
proto_library(attribute_proto SRCS attribute.proto)
proto_library(op_proto SRCS op_proto.proto DEPS attribute_proto)
proto_library(op_desc SRCS op_desc.proto DEPS attribute_proto)
cc_test(op_proto_test SRCS op_proto_test.cc DEPS op_proto protobuf)
cc_test(op_desc_test SRCS op_desc_test.cc DEPS op_desc protobuf)
proto_library(framework_proto SRCS framework.proto)
cc_library(attribute SRCS attribute.cc DEPS op_desc op_proto)
cc_library(operator SRCS operator.cc DEPS op_desc device_context tensor scope attribute)
cc_library(attribute SRCS attribute.cc DEPS framework_proto)
cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope)
cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS op_proto operator)
cc_library(op_registry SRCS op_registry.cc DEPS op_desc grad_op_builder)
cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS operator)
cc_library(op_registry SRCS op_registry.cc DEPS grad_op_builder)
cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
cc_test(grad_op_builder_test SRCS grad_op_builder_test.cc DEPS grad_op_builder op_registry add_op)
py_proto_compile(framework_py_proto SRCS attribute.proto op_proto.proto op_desc.proto)
py_proto_compile(framework_py_proto SRCS framework.proto)
# Generate an empty __init__.py to make framework_py_proto as a valid python module.
add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
add_dependencies(framework_py_proto framework_py_proto_init)
add_custom_command(TARGET framework_py_proto POST_BUILD
COMMAND ${CMAKE_COMMAND} -E make_directory ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/proto
COMMAND cp *.py ${PADDLE_SOURCE_DIR}/python/paddle/v2/framework/proto/
COMMENT "Copy generated python proto into directory paddle/v2/framework/proto."
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
cc_library(backward SRCS backward.cc DEPS net_op)
cc_test(backward_test SRCS backward_test.cc DEPS backward)
if(WITH_PYTHON)
cc_library(paddle_pybind SHARED
SRCS pybind.cc
DEPS pybind python backward
fc_op
sgd_op
add_op
mean_op
cross_entropy_op
recurrent_op
uniform_random_op
fill_zeros_like_op)
endif(WITH_PYTHON)
cc_test(backward_test SRCS backward_test.cc DEPS backward recurrent_op device_context)
......@@ -44,7 +44,7 @@ AttrType AttrTypeID<std::vector<std::string>>() {
return STRINGS;
}
Attribute GetAttrValue(const AttrDesc& attr_desc) {
Attribute GetAttrValue(const OpDesc::Attr& attr_desc) {
switch (attr_desc.type()) {
case paddle::framework::AttrType::INT: {
return attr_desc.i();
......
......@@ -14,16 +14,15 @@ limitations under the License. */
#pragma once
#include <boost/variant.hpp>
#include <functional>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "paddle/framework/attribute.pb.h"
#include "paddle/framework/op_desc.pb.h"
#include "paddle/framework/framework.pb.h"
#include "paddle/platform/enforce.h"
#include "paddle/platform/variant.h"
namespace paddle {
namespace framework {
......@@ -37,7 +36,7 @@ typedef std::unordered_map<std::string, Attribute> AttributeMap;
template <typename T>
AttrType AttrTypeID();
Attribute GetAttrValue(const AttrDesc& attr_desc);
Attribute GetAttrValue(const OpDesc::Attr& attr_desc);
// check whether a value(attribute) fit a certain limit
template <typename T>
......
......@@ -15,31 +15,44 @@
#include "paddle/framework/backward.h"
#include <list>
#include <memory>
#include "paddle/framework/op_registry.h"
#include "paddle/operators/net_op.h"
#include "paddle/operators/recurrent_op.h"
namespace paddle {
namespace framework {
static bool AllInSet(const std::vector<std::string>& names,
const std::string& suffix,
const std::unordered_set<std::string>& set) {
template <typename Map, typename T>
static void ForEachVarName(const Map& names, T callback) {
for (auto& name : names) {
if (set.find(name + suffix) == set.end()) {
return false;
for (auto& n : name.second) {
if (callback(n)) return;
}
}
return true;
}
static std::shared_ptr<OperatorBase> NOP() {
auto net_op = std::make_shared<operators::NetOp>();
net_op->type_ = "@NOP@";
// return whether all the names + suffixes in the set
static bool AllInSet(
const std::map<std::string, std::vector<std::string>>& names,
const std::string& suffix, const std::unordered_set<std::string>& set) {
bool all_in_set = true;
ForEachVarName(names, [&all_in_set, &set, &suffix](const std::string& n) {
all_in_set = set.find(n + suffix) != set.end();
return !all_in_set;
});
return all_in_set;
}
static std::unique_ptr<OperatorBase> NOP() {
auto net_op = new operators::NetOp();
net_op->SetType("@NOP@");
net_op->CompleteAddOp();
return net_op;
return std::unique_ptr<OperatorBase>(net_op);
}
// Get backward operator from a forward operator, recursively implementation.
// Get backward operator from a forward operator, a recursive implementation.
//
// no_grad_names the gradient variable names without gradient calculating.
//
......@@ -47,127 +60,161 @@ static std::shared_ptr<OperatorBase> NOP() {
// BackwardRecursive. use `uid = uniq_id++;` to get the unique index, and
// pass `uniq_id` through recursive calling.
//
// returns The backward operator. For simple situation, it is a simple
// operator. For complex situation, it is a NetOp.
// returns The backward operator. In a simple situation, it may be a simple
// operator, in a complex situation, it maybe a NetOp.
//
// See Backward.h for details
static std::shared_ptr<OperatorBase> BackwardRecursive(
const OperatorBase& forwardOp,
std::unordered_set<std::string>& no_grad_names, size_t& uniq_id);
std::shared_ptr<OperatorBase> BackwardRecursive(
static std::unique_ptr<OperatorBase> BackwardRecursive(
const OperatorBase& forwardOp,
std::unordered_set<std::string>& no_grad_names, size_t& uniq_id) {
// If all input gradients of forwarding operator do not need to calculate,
// just return an NOP. Not return null ptr because NOP does not take
// too much time for calculation, but it is useful for simplifying logic.
if (AllInSet(forwardOp.inputs_, kGradVarSuffix, no_grad_names)) {
if (AllInSet(forwardOp.Inputs() /*names*/, kGradVarSuffix /*suffix*/,
no_grad_names /*set*/)) {
return NOP();
}
// All output gradients of forwarding operator do not need to calculate.
// Then all input gradients cannot be computed at all, and we put them into
// `no_grad_names` set. Return an NOP.
if (AllInSet(forwardOp.outputs_, kGradVarSuffix, no_grad_names)) {
for (auto& name : forwardOp.inputs_) {
// Mark all input is not need
no_grad_names.insert(name + kGradVarSuffix);
}
if (AllInSet(forwardOp.Outputs() /*names*/, kGradVarSuffix /*suffix*/,
no_grad_names /*set*/)) {
ForEachVarName(forwardOp.Inputs(),
[&no_grad_names](const std::string& name) -> bool {
no_grad_names.insert(GradVarName(name));
return false;
});
return NOP();
}
// Returned gradient network
auto net = std::make_shared<operators::NetOp>();
auto net = std::unique_ptr<operators::NetOp>(new operators::NetOp());
if (forwardOp.IsNetOp()) {
// Because forwardOp is a net op, it can static_cast.
auto& forwardNet = static_cast<const operators::NetOp&>(forwardOp);
// Map from output gradient variable name to operator's indices in
// backward net. That operator generates that variable.
// backward net's ops_. That operator generates that variable.
std::unordered_map<std::string, std::vector<size_t>> dup_output_ops;
size_t local_op_id = 0;
// reversely travel forwardNet
// reversely travel forwardNet and collect all duplicate outputs.
for (auto it = forwardNet.ops_.rbegin(); it != forwardNet.ops_.rend();
++it, ++local_op_id) {
auto fwd = *it;
auto& fwd = *it;
auto bwd = BackwardRecursive(*fwd, no_grad_names, uniq_id);
net->AddOp(bwd);
for (auto& out : bwd->outputs_) {
dup_output_ops[out].emplace_back(local_op_id);
}
ForEachVarName(bwd->Outputs(),
[&dup_output_ops, local_op_id](const std::string& out) {
dup_output_ops[out].emplace_back(local_op_id);
return false;
});
net->AppendOp(std::move(bwd));
}
// Get unique ID for this method.
auto uid = uniq_id++;
// TODO(dzh): more comment
using Pos = std::pair<size_t, std::shared_ptr<OperatorBase>>;
// multiple operators which have the same output (y for example) may
// overwrite the same y variable when backward, special operations are token
// to handle this case. For each duplicate output, rename it to an alias
// (original name with a offset), append an `add` op for its operator,
// and finally sum all the alias variable to the final output variable y.
using Pos = std::pair<size_t, std::unique_ptr<OperatorBase>>;
std::list<Pos> insert_position;
for (auto& dup_output_op : dup_output_ops) {
const std::string& name = dup_output_op.first;
// duplicate @Empty@ don't need to be added
if (name == kEmptyVarName) continue;
auto& dup_op = dup_output_op.second;
// no duplicate output
if (dup_op.size() == 1) continue;
std::vector<std::string> dup_outputs;
// process the duplicate outputs
std::vector<std::string> dup_outputs;
for (size_t i = 0; i < dup_op.size(); ++i) {
// rename each duplicate output to an alias
auto op_offset = dup_op[i];
dup_outputs.push_back(name + "@RENAME@" + std::to_string(uid) + "@" +
std::to_string(i));
net->ops_[op_offset]->Rename(name, dup_outputs.back());
}
// collect all the offset to append `add` op for each alias
insert_position.push_back(
{dup_op.back(),
OpRegistry::CreateOp(
"add", {dup_outputs}, {name},
{{"input_format",
std::vector<int>{0, static_cast<int>(dup_outputs.size())}}})});
{dup_op.back(), OpRegistry::CreateOp("add", {{"X", {dup_outputs}}},
{{"Out", {name}}}, {})});
}
// make sure the inserted `add` ops follow the BFS order.
insert_position.sort(
[](const Pos& l, const Pos& r) { return l.first > r.first; });
for (auto& pos : insert_position) {
net->InsertOp(pos.first + 1, pos.second);
net->InsertOp(pos.first + 1, std::move(pos.second));
}
} else {
std::shared_ptr<OperatorBase> grad_op = OpRegistry::CreateGradOp(forwardOp);
for (std::string& grad_input : grad_op->inputs_) {
std::unique_ptr<OperatorBase> grad_op(OpRegistry::CreateGradOp(forwardOp));
ForEachVarName(grad_op->Inputs(), [&no_grad_names, &net, &grad_op](
const std::string& grad_input) {
if (no_grad_names.count(grad_input)) {
std::string prefix =
grad_input.substr(0, grad_input.size() - kGradVarSuffix.size());
grad_input = prefix + kZeroVarSuffix;
// +1 for \0
std::string prefix = grad_input.substr(
0, grad_input.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
grad_op->Rename(grad_input, prefix + kZeroVarSuffix);
// If part of input gradient of that operator is not calculated, fill
// zero variables to that input gradient.
net->AddOp(OpRegistry::CreateOp("fill_zeros_like", {prefix},
{grad_input}, {}));
}
}
for (std::string& grad_output : grad_op->outputs_) {
if (no_grad_names.count(grad_output)) {
grad_output = kEmptyVarName;
net->AppendOp(OpRegistry::CreateOp("fill_zeros_like",
{{"Src", {prefix}}},
{{"Dst", {grad_input}}}, {}));
}
return false;
});
ForEachVarName(grad_op->Outputs(),
[&no_grad_names, &grad_op](const std::string& grad_output) {
if (no_grad_names.count(grad_output)) {
grad_op->Rename(grad_output, kEmptyVarName);
}
return false;
});
// process recurrent gradient op as a special operator.
if (forwardOp.Type() == "recurrent_op") {
// NOTE clean up cycle call somewhere (RNN's stepnet constains itself), or
// this will result in infinite loop.
const auto& rnnop =
*static_cast<const operators::RecurrentOp*>(&forwardOp);
auto rnn_grad_op =
static_cast<operators::RecurrentGradientOp*>(grad_op.get());
const auto& stepnet_op =
*static_cast<const OperatorBase*>(&rnnop.stepnet());
// create stepnet's gradient op
rnn_grad_op->set_stepnet(
BackwardRecursive(stepnet_op, no_grad_names, uniq_id));
}
if (net->ops_.empty()) { // Current no aux op is added to network
return grad_op;
}
net->AddOp(grad_op);
net->AppendOp(std::move(grad_op));
}
net->type_ = "@GENERATED_BACKWARD@";
net->SetType("@GENERATED_BACKWARD@");
net->CompleteAddOp();
return net;
return std::unique_ptr<OperatorBase>(
static_cast<OperatorBase*>(net.release()));
}
// See header for comments
std::shared_ptr<OperatorBase> Backward(
std::unique_ptr<OperatorBase> Backward(
const OperatorBase& forwardOp,
const std::unordered_set<std::string>& no_grad_vars) {
std::unordered_set<std::string> no_grad_names;
no_grad_names.reserve(no_grad_vars.size());
no_grad_names.reserve(no_grad_vars.size() + 1);
no_grad_names.insert(kEmptyVarName + kGradVarSuffix);
no_grad_names.insert(std::string(kEmptyVarName) + kGradVarSuffix);
for (auto& name : no_grad_vars) {
no_grad_names.insert(name + kGradVarSuffix);
......
......@@ -20,7 +20,7 @@ namespace framework {
// Create the backward operator from a forward operator.
// TODO(yuyang18): Add more API reference comment.
extern std::shared_ptr<OperatorBase> Backward(
extern std::unique_ptr<OperatorBase> Backward(
const OperatorBase& forwardOp,
const std::unordered_set<std::string>& no_grad_vars);
} // namespace framework
......
## Operator/expression 's Backward
# Operator/expression 's Backward
### Motivation
## Motivation
In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the fundmental gradient operators/expressions together with chain rule . Every forward network need a backward network to construct the full computation lineage, the operator/ expression's Backward feature will generate the backward pass respect to forward pass.
In Neural Network, the backpropagation algorithm follows the chain rule, so we need to compound the fundmental gradient operators/expressions together with chain rule . Every forward network need a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass.
## Backward Operator Registry
### Implement : gradient operator registry
A backward network is built up with several backward operators. Backward operators take forward operators' inputs, outputs and output gradients and then calculate its input gradients.
| | forward operator | backward operator |
| ---------------------- | ---------------- | -------------------------------- |
| **Operator::inputs_** | Inputs | Inputs, Outputs, OutputGradients |
| **Operator::outputs_** | Outputs | InputGradients |
| | forward operator | backward operator
| ---------------------- | ---------------- |------------------------- |
| **Operator::inputs_** | Inputs | Inputs, Outputs, OutputGradients |
| **Operator::outputs_** | Outputs | InputGradients |
Inputs/Outputs means the input/output of the operator, InputGradients/OutputGradients is the gradient respect to forward opeartor. Forward operator and Backward operator are isomorphic, save their corresponding needs into member attribute.
In most cases, there is a one-to-one correspondence between forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced.
We use a global hash map record the gradient operators available, follow the philosophy of minimum core, make operator pluggable unit. Each gradient is an operator and it needs to regist itself.
For example, we have got a `mul_op`, and we can register it's information and corresponding backward operator by the following macro:
grad_op_builder(fengjiayi)
```cpp
REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
```
### Implement : Backward network
`mul` is the operator's type. `MulOp` and `MulOpMaker` are the operator class and the operator maker class respectively.
`mul_grad` is the type of backward operator, and `MulOpGrad` is its class name.
## Backward Opeartor Creating
Given a certain forward operator, we can get its corresponding backward opeartor by calling:
```cpp
OperatorBase* bwd_op = BuildGradOp(const OperatorBase* fwd_op);
```
The function `BuildGradOp` will sequentially execute following processes:
1. Get the `type_` of given forward operator, and then get the corresponding backward operator's type by looking up the `OpInfoMap`.
2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these are not necessary for gradient computing.
3. Add forward inputs' gradient variables into map `output`, adding forward outputs' gradient variables into map `input`.
4. Building backward operator with `inputs`, `outputs` and forward operator's attributes.
## Backward Network Building
A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and put them together.
In our design, the network itself is also a kind of operator. So the operators contained by a big network may be some small network.
given a forward network, it generates the backward network. We only care about the Gradients—`OutputGradients`,`InputGradients`.
1. bla bla bla (yuyang)
1. Op
when the input forward network is a Op, return its gradient Operator Immediately.
2. NetOp
when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively and ensure them done. During the process, we need to collect the `OutputGradients` name.
when the input forward network is a NetOp, it need to call the sub NetOp/Operators backward function recursively. During the process, we need to collect the `OutputGradients` name according to forward NetOp.
**shared variable**. As illustrated in the pictures, two operator's `Output` `Gradient` will overwirte their shared input variable.
<p align="center">
<img src="./images/duplicate_op.png" width="70%" ><br/>
1. shared variable in two operators.
</p>
Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator replace the overwirte links.
<p align="center">
<img src="images/duplicate_op2.png" width="90%" ><br/>
We share variable in the same scope, as a result, duplicate operator `OutputGradients` will overwirte then duplicate variable.
2. replace shared variable gradient with `Add` Operator
![./images/duplicate_op]()
</p>
Share variable between operators or same input variable used in multiple operators lead to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively, and add a generic add operator instead.
![./images/duplicate_op2]()
​ Then collect the sub graph OutputGradients/InputGradients as the NetOp's and return it.
​ Then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it.
......@@ -28,19 +28,13 @@ using OpAttrChecker = framework::OpAttrChecker;
using Scope = framework::Scope;
using DeviceContext = platform::DeviceContext;
class EmptyOp : public OperatorBase {
public:
void InferShape(const Scope &scope) const override {}
void Run(const Scope &scope, const DeviceContext &dev_ctx) const override {}
};
class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
public:
RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "Input X of Add").IgnoreGradient();
AddInput("b", "Bias of Add").IgnoreGradient();
AddOutput("Out", "Out of Add").IgnoreGradient();
AddInput("X", "Input X of Add").NotInGradient();
AddInput("b", "Bias of Add").NotInGradient();
AddOutput("Out", "Out of Add").NotInGradient();
AddComment("Add Op");
}
};
......@@ -49,8 +43,8 @@ class MulOpMaker : public OpProtoAndCheckerMaker {
public:
MulOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("A", "A");
AddInput("B", "B");
AddInput("X", "A");
AddInput("Y", "B");
AddOutput("Out", "Out");
AddComment("Mul");
}
......@@ -61,7 +55,7 @@ class SigmoidOpMaker : public OpProtoAndCheckerMaker {
SigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "X");
AddOutput("Y", "Y");
AddOutput("Out", "Y");
AddComment("Sigmoid");
}
};
......@@ -71,21 +65,25 @@ class NoGradOpMaker : public OpProtoAndCheckerMaker {
NoGradOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "X input");
AddOutput("Y", "Y output");
AddOutput("Out", "Y output");
AddComment("NoGradOp, same input output. no Grad");
}
};
class FcOp : public operators::NetOp {
public:
void Init() override {
AddOp(OpRegistry::CreateOp("mul", {Input("X"), Input("W")},
{Output("mul_result")}, {}));
auto b_name = Input("b");
FcOp(const std::string &type, const VariableNameMap &inputs,
const VariableNameMap &outputs, const AttributeMap &attrs)
: NetOp(type, inputs, outputs, attrs) {
AppendOp(OpRegistry::CreateOp("mul",
{{"X", {Input("X")}}, {"Y", {Input("W")}}},
{{"Out", {Output("mul_result")}}}, {}));
auto input_b = Inputs("b");
std::string before_act = "mul_result";
if (b_name != kEmptyVarName) {
AddOp(OpRegistry::CreateOp("rowwise_add", {Output("mul_result"), b_name},
{Output("add_result")}, {}));
if (input_b.size() != 0) {
AppendOp(OpRegistry::CreateOp(
"rowwise_add", {{"X", {Output("mul_result")}}, {"b", {input_b[0]}}},
{{"Out", {Output("add_result")}}}, {}));
before_act = "add_result";
} else {
auto out_varname = Output("add_result");
......@@ -94,8 +92,8 @@ class FcOp : public operators::NetOp {
}
}
AddOp(OpRegistry::CreateOp("sigmoid", {Output(before_act)}, {Output("Out")},
{}));
AppendOp(OpRegistry::CreateOp("sigmoid", {{"X", {Output(before_act)}}},
{{"Out", {Output("Out")}}}, {}));
CompleteAddOp(false);
}
};
......@@ -107,8 +105,8 @@ class FcOpMaker : public OpProtoAndCheckerMaker {
AddInput("X", "x");
AddInput("W", "w");
AddInput("b", "b");
AddOutput("mul_result", "").SetTemporary();
AddOutput("add_result", "").SetTemporary();
AddOutput("mul_result", "").AsIntermediate();
AddOutput("add_result", "").AsIntermediate();
AddOutput("Out", "");
AddComment("");
}
......@@ -139,7 +137,7 @@ class AddOpMaker : public OpProtoAndCheckerMaker {
public:
AddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "x").SetMultiple();
AddInput("X", "x").AsDuplicable();
AddOutput("Y", "y");
AddComment("");
}
......@@ -150,51 +148,48 @@ class AddOpMaker : public OpProtoAndCheckerMaker {
namespace f = paddle::framework;
namespace ops = paddle::operators;
using EnforceNotMet = paddle::platform::EnforceNotMet;
REGISTER_OP(rowwise_add, f::EmptyOp, f::RowWiseAddOpMaker);
REGISTER_GRADIENT_OP(rowwise_add, rowwise_add_grad, f::EmptyOp);
REGISTER_OP(mul, f::EmptyOp, f::MulOpMaker);
REGISTER_GRADIENT_OP(mul, mul_grad, f::EmptyOp);
REGISTER_OP(sigmoid, f::EmptyOp, f::SigmoidOpMaker);
REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, f::EmptyOp);
REGISTER_OP(nograd, f::EmptyOp, f::NoGradOpMaker);
REGISTER_OP(fill_zeros_like, f::EmptyOp, f::FillZeroOpMaker);
REGISTER_OP(add, f::EmptyOp, f::AddOpMaker);
REGISTER_GRADIENT_OP(add, add_grad, f::EmptyOp);
REGISTER_OP(fc, f::FcOp, f::FcOpMaker);
REGISTER_OP(many_output_op, f::EmptyOp, f::ManyOutputOpMaker);
REGISTER_GRADIENT_OP(many_output_op, many_output_op_grad, f::EmptyOp);
REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, rowwise_add_grad,
f::NOP);
REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP);
REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP);
REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker);
REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NOP, f::FillZeroOpMaker);
REGISTER_OP(add, f::NOP, f::AddOpMaker, add_grad, f::NOP);
REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker);
REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad,
f::NOP);
TEST(Backward, simple_op_grad) {
auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
auto fwd = f::OpRegistry::CreateOp(
"rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
ASSERT_NE(fwd, nullptr);
auto gop = f::OpRegistry::CreateGradOp(*fwd);
ASSERT_EQ(4UL, gop->inputs_.size());
ASSERT_EQ(f::kEmptyVarName, gop->inputs_[0]);
ASSERT_EQ("rowwise_add_grad", gop->type_);
ASSERT_EQ("X" + f::kGradVarSuffix, gop->outputs_[0]);
ASSERT_EQ("b" + f::kGradVarSuffix, gop->outputs_[1]);
ASSERT_EQ("X" + f::kGradVarSuffix, gop->Output("X" + f::kGradVarSuffix));
ASSERT_EQ(1UL, gop->Inputs().size());
ASSERT_EQ("rowwise_add_grad", gop->Type());
ASSERT_EQ(f::GradVarName("x"), gop->Output(f::GradVarName("X")));
ASSERT_EQ(f::GradVarName("b"), gop->Output(f::GradVarName("b")));
}
TEST(Backward, simple_op_not_need_grad) {
auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
auto fwd = f::OpRegistry::CreateOp(
"rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
ASSERT_NE(fwd, nullptr);
auto gop = f::Backward(*fwd, {"X"});
ASSERT_EQ(std::find(gop->outputs_.begin(), gop->outputs_.end(),
"X" + f::kGradVarSuffix),
gop->outputs_.end());
auto gop = f::Backward(*fwd, {"x"});
ASSERT_EQ(gop->Output(f::GradVarName("X")), f::kEmptyVarName);
auto no_input_gop = f::Backward(*fwd, {"X", "b"});
auto no_input_gop = f::Backward(*fwd, {"x", "b"});
ASSERT_NE(no_input_gop, nullptr);
ASSERT_TRUE(no_input_gop->IsNetOp());
ASSERT_EQ(0UL,
std::static_pointer_cast<ops::NetOp>(no_input_gop)->ops_.size());
ASSERT_EQ(0UL, static_cast<ops::NetOp *>(no_input_gop.get())->ops_.size());
}
TEST(Backward, net_fc_backward_normal) {
std::shared_ptr<f::OperatorBase> fwd = f::OpRegistry::CreateOp(
"fc", {"X", "w", "b"}, {"mul_result", "add_result", "out"}, {});
std::shared_ptr<f::OperatorBase> fwd =
f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {"b"}}},
{{"mul_result", {"mul_res"}},
{"add_result", {"add_re"}},
{"Out", {"out"}}},
{});
ASSERT_NE(fwd, nullptr);
std::shared_ptr<f::OperatorBase> gop = f::Backward(*fwd, {});
ASSERT_TRUE(gop->IsNetOp());
......@@ -205,19 +200,22 @@ TEST(Backward, net_fc_backward_normal) {
ASSERT_EQ(3UL, net->ops_.size());
f::OperatorBase &d_sigmoid = *net->ops_[0];
ASSERT_EQ("sigmoid_grad", d_sigmoid.type_);
ASSERT_EQ("sigmoid_grad", d_sigmoid.Type());
f::OperatorBase &d_add = *net->ops_[1];
ASSERT_EQ("rowwise_add_grad", d_add.type_);
ASSERT_EQ("rowwise_add_grad", d_add.Type());
f::OperatorBase &d_mul = *net->ops_[2];
ASSERT_EQ("mul_grad", d_mul.type_);
ASSERT_EQ("mul_grad", d_mul.Type());
}
TEST(Backward, net_fc_backward_not_have_b) {
std::shared_ptr<f::OperatorBase> fwd =
f::OpRegistry::CreateOp("fc", {"X", "w", f::kEmptyVarName},
{"mul_result", "add_result", "tmp"}, {});
f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {}}},
{{"mul_result", {"mul_res"}},
{"add_result", {"add_res"}},
{"Out", {"tmp"}}},
{});
ASSERT_NE(fwd, nullptr);
std::shared_ptr<f::OperatorBase> gop = f::Backward(*fwd, {});
ASSERT_TRUE(gop->IsNetOp());
......@@ -228,142 +226,176 @@ TEST(Backward, net_fc_backward_not_have_b) {
ASSERT_EQ(2UL, net->ops_.size());
f::OperatorBase &d_sigmoid = *net->ops_[0];
ASSERT_EQ("sigmoid_grad", d_sigmoid.type_);
ASSERT_EQ("sigmoid_grad", d_sigmoid.Type());
f::OperatorBase &d_mul = *net->ops_[1];
ASSERT_EQ("mul_grad", d_mul.type_);
ASSERT_EQ("mul_grad", d_mul.Type());
}
TEST(Backward, net_input_of_network_not_need_grad) {
ops::NetOp net;
net.AddOp(f::OpRegistry::CreateOp("fc", {"X", "W1", "b1"},
{"mul_tmp_0", "add_tmp_0", "hidden0"}, {}));
net.AddOp(f::OpRegistry::CreateOp("fc", {"hidden0", "W2", "b2"},
{"mul_tmp_1", "add_tmp_1", "hidden1"}, {}));
net.AppendOp(f::OpRegistry::CreateOp(
"fc", {{"X", {"x"}}, {"W", {"W1"}}, {"b", {"b1"}}},
{{"mul_result", {"mul_tmp_0"}},
{"add_result", {"add_tmp_0"}},
{"Out", {"hidden0"}}},
{}));
net.AppendOp(f::OpRegistry::CreateOp(
"fc", {{"X", {"hidden0"}}, {"W", {"W2"}}, {"b", {"b2"}}},
{{"mul_result", {"mul_tmp_1"}},
{"add_result", {"add_tmp_1"}},
{"Out", {"hidden1"}}},
{}));
net.CompleteAddOp();
auto bwd = Backward(net, {"X"}); // X@GRAD is not need.
auto bwd = Backward(net, {"x"}); // x@GRAD is not need.
ASSERT_TRUE(bwd->IsNetOp());
auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
std::unordered_set<std::string> all_output = std::unordered_set<std::string>(
bwd_net->outputs_.begin(), bwd_net->outputs_.end());
all_output.erase(f::kEmptyVarName);
auto output_vars = bwd_net->OutputVars(true);
std::unordered_set<std::string> all_outputs =
std::unordered_set<std::string>(output_vars.begin(), output_vars.end());
all_outputs.erase(f::kEmptyVarName);
for (auto &out : {"W1", "b1", "hidden0", "W2", "b2"}) {
ASSERT_NE(all_output.find(out + f::kGradVarSuffix), all_output.end());
ASSERT_NE(all_outputs.find(f::GradVarName(out)), all_outputs.end());
}
// Not Generated X
ASSERT_EQ(all_output.find("X" + f::kGradVarSuffix), all_output.end());
ASSERT_EQ(all_outputs.find(f::GradVarName("X")), all_outputs.end());
ASSERT_EQ(2UL, bwd_net->ops_.size());
ASSERT_TRUE(bwd_net->ops_[1]->IsNetOp());
auto first_fc_grad = static_cast<ops::NetOp *>(bwd_net->ops_[1].get());
ASSERT_EQ(3UL, first_fc_grad->ops_.size());
ASSERT_EQ(f::kEmptyVarName,
first_fc_grad->ops_[2]->Output("A" + f::kGradVarSuffix));
first_fc_grad->ops_[2]->Output(f::GradVarName("X")));
}
TEST(Backward, net_shared_weight) {
ops::NetOp net;
net.AddOp(f::OpRegistry::CreateOp("mul", {"X", "W"}, {"Out"}, {}));
net.AddOp(f::OpRegistry::CreateOp("mul", {"Out", "W"}, {"FinalOut"}, {}));
net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"x"}}, {"Y", {"w"}}},
{{"Out", {"out"}}}, {}));
net.AppendOp(f::OpRegistry::CreateOp("mul", {{"X", {"out"}}, {"Y", {"w"}}},
{{"Out", {"FinalOut"}}}, {}));
net.CompleteAddOp();
auto bwd = f::Backward(net, {});
ASSERT_TRUE(bwd->IsNetOp());
auto bwd_net = static_cast<ops::NetOp *>(bwd.get());
ASSERT_EQ(3UL, bwd_net->ops_.size());
ASSERT_EQ("add", bwd_net->ops_[2]->type_);
ASSERT_EQ("add", bwd_net->ops_[2]->Type());
}
TEST(Backward, op_register_grad_not_for_network) {
auto fwd = f::OpRegistry::CreateOp(
"fc", {"X", "W", "b"}, {"mul_out", "add_out", "out1"},
{{"temporary_index", std::vector<int>{0, 1}}});
auto fwd =
f::OpRegistry::CreateOp("fc", {{"X", {"x"}}, {"W", {"w"}}, {"b", {"b"}}},
{{"mul_result", {"mul_out"}},
{"add_result", {"add_out"}},
{"Out", {"out1"}}},
{{"temporary_index", std::vector<int>{0, 1}}});
ASSERT_THROW(f::OpRegistry::CreateGradOp(*fwd), EnforceNotMet);
}
TEST(Backward, op_all_input_are_not_need) {
auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
auto backward = f::Backward(*fwd, {"X", "b"});
auto fwd = f::OpRegistry::CreateOp(
"rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
auto backward = f::Backward(*fwd, {"x", "b"});
ASSERT_TRUE(backward->IsNetOp());
auto net = static_cast<ops::NetOp *>(backward.get());
ASSERT_TRUE(net->ops_.empty());
}
TEST(Backward, op_all_output_are_not_need) {
auto fwd = f::OpRegistry::CreateOp("rowwise_add", {"X", "b"}, {"Out"}, {});
auto backward = f::Backward(*fwd, {"Out"});
auto fwd = f::OpRegistry::CreateOp(
"rowwise_add", {{"X", {"x"}}, {"b", {"b"}}}, {{"Out", {"out"}}}, {});
auto backward = f::Backward(*fwd, {"out"});
ASSERT_TRUE(backward->IsNetOp());
auto net = static_cast<ops::NetOp *>(backward.get());
ASSERT_TRUE(net->ops_.empty());
}
TEST(Backward, op_part_of_output_are_not_need) {
auto fwd = f::OpRegistry::CreateOp("many_output_op", {"X"}, {"Y", "Z"}, {});
auto fwd = f::OpRegistry::CreateOp("many_output_op", {{"x", {"X"}}},
{{"y", {"Y"}}, {"z", {"Z"}}}, {});
auto backward = f::Backward(*fwd, {"Z"});
ASSERT_TRUE(backward->IsNetOp());
auto net = static_cast<ops::NetOp *>(backward.get());
ASSERT_EQ(net->ops_.size(), 2UL);
auto &fill_zero = *net->ops_[0];
ASSERT_EQ("fill_zeros_like", fill_zero.type_);
ASSERT_EQ(1UL, fill_zero.inputs_.size());
ASSERT_EQ("Z", fill_zero.inputs_[0]);
ASSERT_EQ(1UL, fill_zero.outputs_.size());
ASSERT_EQ("Z" + f::kZeroVarSuffix, fill_zero.outputs_[0]);
ASSERT_EQ("fill_zeros_like", fill_zero.Type());
ASSERT_EQ(1UL, fill_zero.Inputs("Src").size());
ASSERT_EQ("Z", fill_zero.Input("Src"));
ASSERT_EQ(1UL, fill_zero.Outputs("Dst").size());
ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.Output("Dst"));
auto &d_many_out = *net->ops_[1];
ASSERT_EQ("many_output_op_grad", d_many_out.type_);
ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.inputs_.size()); // I/O/OG
ASSERT_EQ("Z" + f::kZeroVarSuffix, d_many_out.Input("z" + f::kGradVarSuffix));
ASSERT_EQ("Y" + f::kGradVarSuffix, d_many_out.Input("y" + f::kGradVarSuffix));
ASSERT_EQ("X" + f::kGradVarSuffix,
d_many_out.Output("x" + f::kGradVarSuffix));
ASSERT_EQ("many_output_op_grad", d_many_out.Type());
ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.Inputs().size()); // I/O/OG
ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix,
d_many_out.Input(f::GradVarName("z")));
ASSERT_EQ(f::GradVarName("Y"), d_many_out.Input(f::GradVarName("y")));
ASSERT_EQ(f::GradVarName("X"), d_many_out.Output(f::GradVarName("x")));
}
TEST(Backward, op_part_of_input_are_not_need) {
auto fwd = f::OpRegistry::CreateOp("mul", {"a", "b"}, {"out"}, {});
auto fwd = f::OpRegistry::CreateOp("mul", {{"X", {"a"}}, {"Y", {"b"}}},
{{"Out", {"out"}}}, {});
auto backward = f::Backward(*fwd, {"a"});
auto &grad_mul = *backward;
ASSERT_EQ(grad_mul.type_, "mul_grad");
ASSERT_EQ(grad_mul.inputs_.size(), 2UL + 1UL + 1UL);
ASSERT_EQ(grad_mul.outputs_.size(), 2UL);
ASSERT_EQ(grad_mul.Output("A" + f::kGradVarSuffix), f::kEmptyVarName);
ASSERT_EQ(grad_mul.Output("B" + f::kGradVarSuffix), "b" + f::kGradVarSuffix);
ASSERT_EQ(grad_mul.Input("Out" + f::kGradVarSuffix),
"out" + f::kGradVarSuffix);
ASSERT_EQ(grad_mul.Input("A"), "a");
ASSERT_EQ(grad_mul.Input("B"), "b");
ASSERT_EQ(grad_mul.Type(), "mul_grad");
ASSERT_EQ(grad_mul.Inputs().size(), 2UL + 1UL + 1UL);
ASSERT_EQ(grad_mul.Outputs().size(), 2UL);
ASSERT_EQ(grad_mul.Output(f::GradVarName("X")), f::kEmptyVarName);
ASSERT_EQ(grad_mul.Output(f::GradVarName("Y")), f::GradVarName("b"));
ASSERT_EQ(grad_mul.Input(f::GradVarName("Out")), f::GradVarName("out"));
ASSERT_EQ(grad_mul.Input("X"), "a");
ASSERT_EQ(grad_mul.Input("Y"), "b");
ASSERT_EQ(grad_mul.Input("Out"), "out");
}
TEST(Backward, linear_net_intermediate_variable_has_no_grad) {
ops::NetOp net;
net.AddOp(f::OpRegistry::CreateOp("fc", {"x1", "w1", "b1"},
{"mul_out1", "add_out1", "out1"}, {}));
net.AddOp(f::OpRegistry::CreateOp("fc", {"out1", "w2", "b2"},
{"mul_out2", "tmp_out2", "out2"}, {}));
net.AddOp(f::OpRegistry::CreateOp("fc", {"out2", "w3", "b3"},
{"mul_out3", "tmp_out3", "out3"}, {}));
net.AppendOp(f::OpRegistry::CreateOp(
"fc", {{"X", {"x1"}}, {"W", {"w1"}}, {"b", {"b1"}}},
{{"mul_result", {"mul_out1"}},
{"add_result", {"add_out1"}},
{"Out", {"out1"}}},
{}));
net.AppendOp(f::OpRegistry::CreateOp(
"fc", {{"X", {"out1"}}, {"W", {"w2"}}, {"b", {"b2"}}},
{{"mul_result", {"mul_out2"}},
{"add_result", {"tmp_out2"}},
{"Out", {"out2"}}},
{}));
net.AppendOp(f::OpRegistry::CreateOp(
"fc", {{"X", {"out2"}}, {"W", {"w3"}}, {"b", {"b3"}}},
{{"mul_result", {"mul_out3"}},
{"add_result", {"tmp_out3"}},
{"Out", {"out3"}}},
{}));
net.CompleteAddOp();
auto backward = f::Backward(net, {"mul_out2", "tmp_out2", "out2"});
ASSERT_TRUE(backward->IsNetOp());
auto bwd_net = static_cast<ops::NetOp *>(backward.get());
ASSERT_EQ(bwd_net->ops_.size(), 3UL);
auto &grad_fc = *bwd_net->ops_[0];
EXPECT_EQ(grad_fc.inputs_.size(),
3UL /* external input number */
const char *all = paddle::operators::NetOp::kAll;
EXPECT_EQ(grad_fc.Inputs(all).size(),
2UL /* external input number */
+ 1UL /* external output number*/
+ 1UL /* number of gradient of external output*/
+ 2U /* internal variable number*/);
EXPECT_EQ(grad_fc.outputs_.size(), 2UL /* input number of mul*/
+ 2UL /* input number of rowwise_add */
+ 1UL /* input number of sigmod */);
EXPECT_EQ(bwd_net->ops_[1]->inputs_.size(), 0UL);
EXPECT_EQ(bwd_net->ops_[1]->outputs_.size(), 0UL);
EXPECT_EQ(bwd_net->ops_[2]->inputs_.size(), 0UL);
EXPECT_EQ(bwd_net->ops_[2]->outputs_.size(), 0UL);
EXPECT_EQ(grad_fc.Outputs(all).size(),
2UL /* input number of mul*/
+ 2UL /* input number of rowwise_add
*/
+ 1UL /* input number of sigmod */);
EXPECT_EQ(bwd_net->ops_[1]->Inputs(all).size(), 0UL);
EXPECT_EQ(bwd_net->ops_[1]->Outputs(all).size(), 0UL);
EXPECT_EQ(bwd_net->ops_[2]->Inputs(all).size(), 0UL);
EXPECT_EQ(bwd_net->ops_[2]->Outputs(all).size(), 0UL);
}
......@@ -283,6 +283,5 @@ std::ostream& operator<<(std::ostream& os, const DDim& ddim) {
DDim::DDim(std::initializer_list<int> init_list) {
*this = make_ddim(init_list);
}
} // namespace framework
} // namespace paddle
......@@ -14,13 +14,12 @@ limitations under the License. */
#pragma once
#include <boost/variant.hpp>
#include <initializer_list>
#include <stdexcept>
#include <vector>
#include "paddle/framework/dim.h"
#include "paddle/platform/enforce.h"
#include "unsupported/Eigen/CXX11/Tensor"
#include "paddle/platform/variant.h"
namespace paddle {
namespace framework {
......
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
syntax = "proto2";
package paddle.framework;
enum AttrType {
INT = 0;
FLOAT = 1;
STRING = 2;
INTS = 3;
FLOATS = 4;
STRINGS = 5;
}
// OpDesc describes an instance of a C++ framework::OperatorBase
// derived class type.
message OpDesc {
message Attr {
required string name = 1;
required AttrType type = 2;
optional int32 i = 3;
optional float f = 4;
optional string s = 5;
repeated int32 ints = 6;
repeated float floats = 7;
repeated string strings = 8;
};
message Var {
required string parameter = 1;
repeated string arguments = 2;
};
required string type = 3;
repeated Var inputs = 1;
repeated Var outputs = 2;
repeated Attr attrs = 4;
};
// OpProto describes a C++ framework::OperatorBase derived class.
message OpProto {
// VarProto describes the C++ type framework::Variable.
message Var {
required string name = 1;
required string comment = 2;
optional bool duplicable = 3 [ default = false ];
optional bool intermediate = 4 [ default = false ];
optional bool not_in_gradient = 5 [ default = false ];
}
// AttrProto describes the C++ type Attribute.
message Attr {
required string name = 1;
required AttrType type = 2;
required string comment = 3;
// If that attribute is generated, it means the Paddle third
// language binding has responsibility to fill that
// attribute. End-User should not set that attribute.
optional bool generated = 4 [ default = false ];
}
required string type = 1;
repeated Var inputs = 2;
repeated Var outputs = 3;
repeated Attr attrs = 4;
required string comment = 5;
}
......@@ -13,90 +13,45 @@ express or implied. See the License for the specific language governing
permissions and limitations under the License. */
#include "paddle/framework/grad_op_builder.h"
#include "paddle/framework/op_proto.pb.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace framework {
class OpRegistry;
using VarIndexMap = std::unordered_map<std::string, int>;
enum class OpArgType { IN, OUT };
static std::vector<int>* GetOpFormat(OperatorBase* op, const OpArgType& type) {
std::string key = type == OpArgType::IN ? "input_format" : "output_format";
return op->attrs_.count(key)
? &boost::get<std::vector<int>>(op->attrs_.at(key))
: nullptr;
}
static const std::vector<int>* GetOpFormat(const OperatorBase* op,
const OpArgType& type) {
std::string key = type == OpArgType::IN ? "input_format" : "output_format";
return op->attrs_.count(key)
? &boost::get<std::vector<int>>(op->attrs_.at(key))
: nullptr;
}
static void TransOpArg(const OperatorBase* src_op, OperatorBase* dst_op,
const OpArgType& src_type, const OpArgType& dst_type,
int& idx, bool is_grad) {
const std::vector<std::string>& src_inout =
src_type == OpArgType::IN ? src_op->inputs_ : src_op->outputs_;
const std::vector<int>* src_format = GetOpFormat(src_op, src_type);
std::vector<std::string>& dst_inout =
dst_type == OpArgType::IN ? dst_op->inputs_ : dst_op->outputs_;
std::vector<int>* dst_format = GetOpFormat(dst_op, dst_type);
const OpProto& proto = OpRegistry::protos().at(src_op->type_);
static void TransOpArg(const OperatorBase* src_op, const OpArgType& src_type,
bool is_grad, VariableNameMap* vars) {
const auto& src_inout =
src_type == OpArgType::IN ? src_op->Inputs() : src_op->Outputs();
auto& dst_inout = *vars;
auto& proto = OpInfoMap::Instance().Get(src_op->Type()).Proto();
const auto& src_arg_list =
src_type == OpArgType::IN ? proto.inputs() : proto.outputs();
for (const auto& arg : src_arg_list) {
std::string src_name = arg.name();
std::string dst_name = is_grad ? src_name + kGradVarSuffix : src_name;
(*dst_op->in_out_idxs_)[dst_name] = idx++;
int src_arg_idx = src_op->in_out_idxs_->at(src_name);
int src_begin =
src_format == nullptr ? src_arg_idx : src_format->at(src_arg_idx);
int src_end = src_format == nullptr ? src_arg_idx + 1
: src_format->at(src_arg_idx + 1);
for (int i = src_begin; i < src_end; ++i) {
std::string s =
is_grad ? src_inout[i] + kGradVarSuffix
: (arg.ignore_gradient() ? kEmptyVarName : src_inout[i]);
dst_inout.emplace_back(s);
}
if (dst_format != nullptr) {
dst_format->push_back(dst_inout.size());
if (arg.not_in_gradient() && !is_grad) continue;
const std::string src_name = arg.name();
std::string dst_name = is_grad ? GradVarName(src_name) : src_name;
dst_inout[dst_name].reserve(src_inout.at(src_name).size());
for (auto& var_name : src_inout.at(src_name)) {
std::string s = is_grad ? GradVarName(var_name) : var_name;
dst_inout[dst_name].emplace_back(s);
}
}
}
OperatorBase* BuildGradOp(const OperatorBase* op) {
std::string grad_op_type = OpRegistry::grad_ops().at(op->type_);
OperatorBase* grad_op = OpRegistry::op_creators().at(grad_op_type)();
grad_op->type_ = grad_op_type;
grad_op->attrs_ = op->attrs_;
grad_op->attrs_.erase("input_format");
grad_op->attrs_.erase("output_format");
if (GetOpFormat(op, OpArgType::IN) != nullptr) {
grad_op->attrs_["output_format"] = std::vector<int>({0});
}
if (GetOpFormat(op, OpArgType::IN) != nullptr ||
GetOpFormat(op, OpArgType::OUT) != nullptr) {
grad_op->attrs_["input_format"] = std::vector<int>({0});
}
grad_op->in_out_idxs_.reset(new VarIndexMap());
int in_idx = 0;
int out_idx = 0;
TransOpArg(op, grad_op, OpArgType::IN, OpArgType::IN, in_idx, false); // I
TransOpArg(op, grad_op, OpArgType::OUT, OpArgType::IN, in_idx, false); // G
TransOpArg(op, grad_op, OpArgType::OUT, OpArgType::IN, in_idx, true); // OG
TransOpArg(op, grad_op, OpArgType::IN, OpArgType::OUT, out_idx, true); // IG
return grad_op;
auto& info = OpInfoMap::Instance().Get(op->Type());
PADDLE_ENFORCE(info.HasGradientOp());
VariableNameMap inputs;
VariableNameMap outputs;
TransOpArg(op, OpArgType::IN, false, &inputs); // I
TransOpArg(op, OpArgType::OUT, false, &inputs); // O
TransOpArg(op, OpArgType::OUT, true, &inputs); // OG
TransOpArg(op, OpArgType::IN, true, &outputs); // IG
auto& grad_info = OpInfoMap::Instance().Get(info.grad_op_type_);
return grad_info.Creator()(info.grad_op_type_, inputs, outputs, op->Attrs());
}
} // namespace framework
......
......@@ -8,22 +8,15 @@ USE_OP(add_two);
namespace paddle {
namespace framework {
class NOP : public OperatorBase {
public:
void InferShape(const Scope &scope) const override {}
void Run(const Scope &scope,
const platform::DeviceContext &dev_ctx) const override {}
};
class MutiInOutOpMaker : public OpProtoAndCheckerMaker {
public:
MutiInOutOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("In1", "a single input");
AddInput("In2_mult", "a multiple input").SetMultiple();
AddInput("In2_mult", "a multiple input").AsDuplicable();
AddInput("In3", "another single input");
AddOutput("Out1", "a single output");
AddOutput("Out2_mult", "a multiple output").SetMultiple();
AddOutput("Out2_mult", "a multiple output").AsDuplicable();
AddComment("test op with multiple inputs and outputs");
}
};
......@@ -33,10 +26,10 @@ class IOIgnoredOpMaker : public OpProtoAndCheckerMaker {
IOIgnoredOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("In1", "a single input");
AddInput("In2_mult", "a multiple input").SetMultiple().IgnoreGradient();
AddInput("In3_mult", "another multiple input").SetMultiple();
AddOutput("Out1_mult", "a multiple output").SetMultiple();
AddOutput("Out2", "a single output").IgnoreGradient();
AddInput("In2_mult", "a multiple input").AsDuplicable().NotInGradient();
AddInput("In3_mult", "another multiple input").AsDuplicable();
AddOutput("Out1_mult", "a multiple output").AsDuplicable();
AddOutput("Out2", "a single output").NotInGradient();
AddComment("op with inputs and outputs ignored in gradient calculating");
}
};
......@@ -47,35 +40,33 @@ class IOIgnoredOpMaker : public OpProtoAndCheckerMaker {
namespace f = paddle::framework;
TEST(GradOpBuilder, AddTwo) {
std::shared_ptr<f::OperatorBase> add_op(
f::OpRegistry::CreateOp("add_two", {"x", "y"}, {"out"}, {}));
std::shared_ptr<f::OperatorBase> add_op(f::OpRegistry::CreateOp(
"add_two", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {}));
std::shared_ptr<f::OperatorBase> grad_add_op =
f::OpRegistry::CreateGradOp(*add_op);
EXPECT_EQ(static_cast<int>(grad_add_op->inputs_.size()), 4);
EXPECT_EQ(static_cast<int>(grad_add_op->outputs_.size()), 2);
EXPECT_EQ(grad_add_op->Inputs().size(), 4UL);
EXPECT_EQ(grad_add_op->Outputs().size(), 2UL);
EXPECT_EQ(grad_add_op->Input("X"), "x");
EXPECT_EQ(grad_add_op->Input("Y"), "y");
EXPECT_EQ(grad_add_op->Input("Out"), "out");
EXPECT_EQ(grad_add_op->Input("Out@GRAD"), "out@GRAD");
EXPECT_EQ(grad_add_op->Output("X@GRAD"), "x@GRAD");
EXPECT_EQ(grad_add_op->Output("Y@GRAD"), "y@GRAD");
EXPECT_EQ(grad_add_op->Input(f::GradVarName("Out")), f::GradVarName("out"));
EXPECT_EQ(grad_add_op->Output(f::GradVarName("X")), f::GradVarName("x"));
EXPECT_EQ(grad_add_op->Output(f::GradVarName("Y")), f::GradVarName("y"));
}
REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker);
REGISTER_GRADIENT_OP(mult_io, mult_io_grad, f::NOP);
REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker);
REGISTER_GRADIENT_OP(io_ignored, io_ignored_grad, f::NOP);
REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker, mult_io_grad, f::NOP);
REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker, io_ignored_grad, f::NOP);
TEST(GradOpBuilder, MutiInOut) {
f::AttributeMap attrs{{"input_format", std::vector<int>{0, 1, 4, 5}},
{"output_format", std::vector<int>{0, 1, 3}}};
std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(
"mult_io", {"in1", "in2_1", "in2_2", "in2_3", "in3"},
{"out1", "out2_1", "out2_2"}, attrs));
"mult_io", {{"In1", {"in1"}},
{"In2_mult", {"in2_1", "in2_2", "in2_3"}},
{"In3", {"in3"}}},
{{"Out1", {"out1"}}, {"Out2_mult", {"out2_1", "out2_2"}}}, {}));
std::shared_ptr<f::OperatorBase> grad_test_op =
f::OpRegistry::CreateGradOp(*test_op);
ASSERT_EQ(grad_test_op->inputs_.size(), 5UL + 3UL + 3UL);
ASSERT_EQ(grad_test_op->Inputs().size(), 3UL + 2UL + 2UL);
EXPECT_EQ(grad_test_op->Input("In1"), "in1");
EXPECT_EQ(grad_test_op->Inputs("In2_mult"),
std::vector<std::string>({"in2_1", "in2_2", "in2_3"}));
......@@ -83,55 +74,49 @@ TEST(GradOpBuilder, MutiInOut) {
EXPECT_EQ(grad_test_op->Input("Out1"), "out1");
EXPECT_EQ(grad_test_op->Inputs("Out2_mult"),
std::vector<std::string>({"out2_1", "out2_2"}));
EXPECT_EQ(grad_test_op->Input("Out1" + f::kGradVarSuffix),
"out1" + f::kGradVarSuffix);
EXPECT_EQ(grad_test_op->Inputs("Out2_mult" + f::kGradVarSuffix),
EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out1")),
f::GradVarName("out1"));
EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out2_mult")),
std::vector<std::string>(
{"out2_1" + f::kGradVarSuffix, "out2_2" + f::kGradVarSuffix}));
{f::GradVarName("out2_1"), f::GradVarName("out2_2")}));
ASSERT_EQ(grad_test_op->outputs_.size(), 5UL);
EXPECT_EQ(grad_test_op->Output("In1" + f::kGradVarSuffix),
"in1" + f::kGradVarSuffix);
EXPECT_EQ(grad_test_op->Outputs("In2_mult" + f::kGradVarSuffix),
std::vector<std::string>({"in2_1" + f::kGradVarSuffix,
"in2_2" + f::kGradVarSuffix,
"in2_3" + f::kGradVarSuffix}));
EXPECT_EQ(grad_test_op->Output("In3" + f::kGradVarSuffix),
"in3" + f::kGradVarSuffix);
ASSERT_EQ(grad_test_op->Outputs().size(), 3UL);
EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
std::vector<std::string>({f::GradVarName("in2_1"),
f::GradVarName("in2_2"),
f::GradVarName("in2_3")}));
EXPECT_EQ(grad_test_op->Output(f::GradVarName("In3")), f::GradVarName("in3"));
}
TEST(GradOpBuilder, IOIgnoredInGradient) {
f::AttributeMap attrs{{"input_format", std::vector<int>{0, 1, 3, 5}},
{"output_format", std::vector<int>{0, 2, 3}}};
std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(
"io_ignored", {"in1", "in2_1", "in2_2", "in3_1", "in3_2"},
{"out1_1", "out1_2", "out2"}, attrs));
"io_ignored", {{"In1", {"in1"}},
{"In2_mult", {"in2_1", "in2_2"}},
{"In3_mult", {"in3_1", "in3_2"}}},
{{"Out1_mult", {"out1_1", "out1_2"}}, {"Out2", {"out2"}}}, {}));
std::shared_ptr<f::OperatorBase> grad_test_op =
f::OpRegistry::CreateGradOp(*test_op);
// 'In2' and 'Out2' are ignored in gradient calculating
ASSERT_EQ(grad_test_op->inputs_.size(), 5UL + 3UL + 3UL);
ASSERT_EQ(grad_test_op->Inputs().size(), 2UL + 1UL + 2UL);
EXPECT_EQ(grad_test_op->Input("In1"), "in1");
EXPECT_EQ(grad_test_op->Inputs("In2_mult"),
std::vector<std::string>({f::kEmptyVarName, f::kEmptyVarName}));
EXPECT_EQ(grad_test_op->Inputs("In3_mult"),
std::vector<std::string>({"in3_1", "in3_2"}));
EXPECT_EQ(grad_test_op->Inputs("Out1_mult"),
std::vector<std::string>({"out1_1", "out1_2"}));
EXPECT_EQ(grad_test_op->Input("Out2"), f::kEmptyVarName);
EXPECT_EQ(grad_test_op->Inputs("Out1_mult" + f::kGradVarSuffix),
EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out1_mult")),
std::vector<std::string>(
{"out1_1" + f::kGradVarSuffix, "out1_2" + f::kGradVarSuffix}));
EXPECT_EQ(grad_test_op->Input("Out2" + f::kGradVarSuffix),
"out2" + f::kGradVarSuffix);
{f::GradVarName("out1_1"), f::GradVarName("out1_2")}));
EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out2")),
f::GradVarName("out2"));
ASSERT_EQ(grad_test_op->outputs_.size(), 5UL);
EXPECT_EQ(grad_test_op->Output("In1" + f::kGradVarSuffix),
"in1" + f::kGradVarSuffix);
EXPECT_EQ(grad_test_op->Outputs("In2_mult" + f::kGradVarSuffix),
ASSERT_EQ(grad_test_op->Outputs().size(), 3UL);
EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
std::vector<std::string>(
{"in2_1" + f::kGradVarSuffix, "in2_2" + f::kGradVarSuffix}));
EXPECT_EQ(grad_test_op->Outputs("In3_mult" + f::kGradVarSuffix),
{f::GradVarName("in2_1"), f::GradVarName("in2_2")}));
EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In3_mult")),
std::vector<std::string>(
{"in3_1" + f::kGradVarSuffix, "in3_2" + f::kGradVarSuffix}));
{f::GradVarName("in3_1"), f::GradVarName("in3_2")}));
}
......@@ -19,32 +19,59 @@
namespace paddle {
namespace framework {
LODTensor LODTensor::SliceShared(size_t level_begin, size_t level_end) const {
PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
auto new_lod = details::SliceLOD(*lod_start_pos_, level_begin, level_end);
// slice levels just need to update LOD info, each level will contains the
// whole tensor_, so no need to modify tensor_.
return LODTensor(tensor_, new_lod);
LODTensor::LOD LODTensor::LOD::SliceLevels(size_t level_begin,
size_t level_end) const {
LOD new_lod;
new_lod.reserve(level_end - level_begin);
for (size_t i = level_begin; i < level_end; i++) {
new_lod.emplace_back(at(i));
}
return new_lod;
}
LODTensor LODTensor::SliceShared(size_t level, size_t elem_begin,
size_t elem_end) const {
PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
NumLevels());
PADDLE_ENFORCE(elem_begin < NumElements(level),
"element begin [%d] out of range [%d]", elem_begin,
NumElements(level));
PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
"element end [%d] out of range [%d]", elem_end,
NumElements(level));
auto new_lod = details::SliceLOD(*lod_start_pos_, level, elem_begin, elem_end,
true /*tensor_shared*/);
// slice elements just need to update LOD info, because offsets are not
// changed, so the original tensor_ can be reused.
return LODTensor(tensor_, new_lod);
LODTensor::LOD LODTensor::LOD::SliceInLevel(size_t level, size_t elem_begin,
size_t elem_end) const {
// slice the lod.
LOD new_lod;
new_lod.reserve(size() - level);
auto start = this->at(level)[elem_begin];
auto end = this->at(level)[elem_end];
for (auto it = this->begin() + level; it != this->end(); it++) {
auto it_begin = std::find(it->begin(), it->end(), start);
auto it_end = std::find(it_begin, it->end(), end);
PADDLE_ENFORCE(it_begin != it->end(), "error in parsing lod info");
PADDLE_ENFORCE(it_end != it->end(), "error in parsing lod info");
new_lod.emplace_back(it_begin, it_end + 1);
// reset offset if tensor is copyed and sliced.
std::transform(new_lod.back().begin(), new_lod.back().end(),
new_lod.back().begin(),
[start](int v) { return v - start; });
PADDLE_ENFORCE_EQ(new_lod.back().front(), 0, "error in slice LOD");
}
PADDLE_ENFORCE_LE(new_lod.size(), this->size());
return new_lod;
}
bool operator==(const LODTensor::LOD& a, const LODTensor::LOD& b) {
if (a.size() != b.size()) {
return false;
}
for (size_t i = 0; i < a.size(); i++) {
const auto& a_level = a[i];
const auto& b_level = b[i];
if (a_level.size() != b_level.size()) {
return false;
}
for (size_t j = 0; j < a_level.size(); j++) {
if (a_level[j] != b_level[j]) {
return false;
}
}
}
return true;
}
} // namespace framework
......
......@@ -15,7 +15,7 @@
#pragma once
#include <memory>
#if (!PADDLE_ONLY_CPU)
#if !defined(PADDLE_ONLY_CPU)
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>
#endif
......@@ -31,30 +31,29 @@ namespace framework {
* LODTensor (Level of details Tensor)
* see https://en.wikipedia.org/wiki/Level_of_details for reference.
*/
class LODTensor {
class LODTensor : public Tensor {
public:
// Level save offsets of each unit.
#ifdef PADDLE_ONLY_CPU
using Level = std::vector<size_t>;
template <typename T>
using Vector = std::vector<T>;
#else
using Level = thrust::device_vector<size_t>;
template <typename T>
using Vector = thrust::host_vector<T>;
#endif
// LOD stores offsets of each level of units, the largest units level first,
// LoD stores offsets of each level of units, the largest units level first,
// then the smaller units level. Each Level stores the offsets of units in
// Tesor.
typedef std::vector<Level> LOD;
class LOD : public std::vector<Vector<size_t>> {
public:
LOD SliceLevels(size_t level_begin, size_t level_end) const;
LOD SliceInLevel(size_t level, size_t elem_begin, size_t elem_end) const;
};
LODTensor() {}
LODTensor(const std::shared_ptr<Tensor> &tensor,
const std::shared_ptr<LOD> &lod) {
Reset(tensor, lod);
}
explicit LODTensor(const LOD &lod) : lod_(lod) {}
void Reset(const std::shared_ptr<Tensor> &tensor,
const std::shared_ptr<LOD> &lod) {
tensor_ = tensor;
lod_start_pos_ = lod;
}
virtual Tensor *Clone() const { return new LODTensor(lod_); }
/*
* Get a element from LOD.
......@@ -65,16 +64,14 @@ class LODTensor {
PADDLE_ENFORCE(elem < NumElements(level),
"element begin [%d] out of range [%d]", elem,
NumElements(level));
return (*lod_start_pos_)[level][elem];
return (lod_)[level][elem];
}
/*
* Number of LODTensor's levels, each level has units of data, for example,
* in the sentence's view, article, paragraph, sentence are 3 levels.
*/
size_t NumLevels() const {
return lod_start_pos_ ? lod_start_pos_->size() : 0UL;
}
size_t NumLevels() const { return lod_.size(); }
/*
* Number of elements in a level.
*/
......@@ -82,64 +79,71 @@ class LODTensor {
PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
NumLevels());
// the last offset is the end of last element
return lod_start_pos_->at(level).size() - 1;
return lod_[level].size() - 1;
}
/*
* Slice of levels[level_begin:level_end], with tensor copied.
*/
template <typename T>
LODTensor SliceCopied(size_t level_begin, size_t level_end,
const platform::Place &dst_place) const;
/*
* Slice of levels[level_begin:level_end], with tensor shared.
*/
LODTensor SliceShared(size_t level_begin, size_t level_end) const;
/*
* Slice of elements of a level, [elem_begin: elem_end], with tensor copied.
* @note: low performance in slice lod_start_pos_.
*/
template <typename T>
LODTensor SliceCopied(size_t level, size_t elem_begin, size_t elem_end,
const platform::Place &dst_place) const;
LODTensor SliceLevels(size_t level_begin, size_t level_end) const;
/*
* Slice of elements of a level, [elem_begin: elem_end], with tensor shared.
* @note: low performance in slice lod_start_pos_.
*/
LODTensor SliceShared(size_t level, size_t elem_begin, size_t elem_end) const;
/*
* Copy other's lod_start_pos_, to share LOD info.
* @note: the LOD info should not be changed.
* @note: low performance in slice lod_.
*/
void ShareLOD(const LODTensor &other) {
lod_start_pos_ = other.lod_start_pos_;
}
template <typename T>
LODTensor SliceInLevel(size_t level, size_t elem_begin,
size_t elem_end) const;
/*
* Copy other's lod_start_pos_'s content, free to mutate.
* Copy other's lod_'s content, free to mutate.
*/
void CopyLOD(const LODTensor &other) {
lod_start_pos_ = std::make_shared<LOD>(*other.lod_start_pos_);
}
void CopyLOD(const LODTensor &other) { lod_ = other.lod_; }
/*
* Determine whether LODTensor has a valid LOD info.
*/
bool HasLOD() const { return bool(lod_start_pos_); }
LOD *lod() const { return lod_start_pos_.get(); }
const LOD &lod() const { return lod_; }
LOD *mutable_lod() { return &lod_; }
std::shared_ptr<Tensor> &tensor() { return tensor_; }
Tensor *raw_tensor() { return tensor_.get(); }
virtual ~LODTensor() {}
private:
std::shared_ptr<LOD> lod_start_pos_;
std::shared_ptr<Tensor> tensor_;
LOD lod_;
};
bool operator==(const LODTensor::LOD &a, const LODTensor::LOD &b);
template <typename T>
LODTensor LODTensor::SliceLevels(size_t level_begin, size_t level_end) const {
auto new_lod = lod_.SliceLevels(level_begin, level_end);
// slice levels just need to update LOD info, each level will contains the
// whole tensor_, so no need to modify tensor_.
LODTensor new_tensor(new_lod);
new_tensor.ShareDataWith<T>(*this);
return new_tensor;
}
template <typename T>
LODTensor LODTensor::SliceInLevel(size_t level, size_t elem_begin,
size_t elem_end) const {
PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
NumLevels());
PADDLE_ENFORCE(elem_begin < NumElements(level),
"element begin [%d] out of range [%d]", elem_begin,
NumElements(level));
PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
"element end [%d] out of range [%d]", elem_end,
NumElements(level));
auto new_lod = lod_.SliceInLevel(level, elem_begin, elem_end);
// slice elements just need to update LOD info, because offsets are not
// changed, so the original tensor_ can be reused.
LODTensor new_tensor(new_lod);
new_tensor.ShareDataWith<T>(*this);
return new_tensor;
}
} // namespace framework
} // namespace paddle
#include "paddle/framework/lod_tensor_impl.h"
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/framework/details/lod_tensor.h"
namespace paddle {
namespace framework {
template <typename T>
LODTensor LODTensor::SliceCopied(size_t level_begin, size_t level_end,
const platform::Place &dst_place) const {
PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
auto new_lod = details::SliceLOD(*lod_start_pos_, level_begin, level_end);
auto new_tensor = std::make_shared<Tensor>();
new_tensor->CopyFrom<T>(*tensor_, dst_place);
return LODTensor(new_tensor, new_lod);
}
template <typename T>
LODTensor LODTensor::SliceCopied(size_t level, size_t elem_begin,
size_t elem_end,
const platform::Place &dst_place) const {
PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
NumLevels());
PADDLE_ENFORCE(elem_begin < NumElements(level),
"element begin [%d] out of range [%d]", elem_begin,
NumElements(level));
PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
"element end [%d] out of range [%d]", elem_end,
NumElements(level));
auto new_lod = details::SliceLOD(*lod_start_pos_, level, elem_begin, elem_end,
false /*tensor_shared*/);
auto start_idx = new_lod->front().front();
auto end_idx = new_lod->front().back() - 1 /*the next element's start*/;
auto sliced_tensor = tensor_->Slice<T>(start_idx, end_idx);
auto new_tensor = std::make_shared<Tensor>();
new_tensor->CopyFrom<T>(sliced_tensor, dst_place);
return LODTensor(new_tensor, new_lod);
}
} // namespace framework
} // namespace paddle
......@@ -15,6 +15,7 @@
#include <glog/logging.h>
#include <gtest/gtest.h>
#include <algorithm>
#include <memory>
namespace paddle {
......@@ -29,22 +30,28 @@ class LODTensorTester : public ::testing::Test {
// 0 10 20
// 0 5 10 15 20
// 0 2 5 7 10 12 15 20
auto lod = std::make_shared<LODTensor::LOD>();
lod->push_back(std::vector<size_t>{0, 10, 20});
lod->push_back(std::vector<size_t>{0, 5, 10, 15, 20});
lod->push_back(std::vector<size_t>{0, 2, 5, 7, 10, 12, 15, 17, 20});
LODTensor::LOD lod;
lod.push_back(std::vector<size_t>{0, 10, 20});
lod.push_back(std::vector<size_t>{0, 5, 10, 15, 20});
lod.push_back(std::vector<size_t>{0, 2, 5, 7, 10, 12, 15, 17, 20});
auto tensor = std::make_shared<Tensor>();
tensor->Resize({20 /*batch size*/, 128 /*dim*/});
ASSERT_EQ(lod.size(), 3UL);
tensor.Resize({20 /*batch size*/, 128 /*dim*/});
// malloc memory
tensor->mutable_data<float>(place);
tensor.mutable_data<float>(place);
lod_tensor.reset(new LODTensor(lod));
lod_tensor->Resize({20 /*batch size*/, 128 /*dim*/});
lod_tensor->Reset(tensor, lod);
lod_tensor->ShareDataWith<float>(tensor);
// lod_tensor->ShareDataWith<Tensor>(tensor);
}
protected:
std::unique_ptr<LODTensor> lod_tensor;
platform::CPUPlace place;
Tensor tensor;
};
TEST_F(LODTensorTester, NumLevels) { ASSERT_EQ(lod_tensor->NumLevels(), 3UL); }
......@@ -55,110 +62,54 @@ TEST_F(LODTensorTester, NumElements) {
ASSERT_EQ(lod_tensor->NumElements(2), 8UL);
}
TEST_F(LODTensorTester, SliceShared_Level) {
// slice 1 level
for (size_t level = 0; level < 3UL; ++level) {
auto new_lod_tensor = lod_tensor->SliceShared(level, level + 1);
ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
ASSERT_EQ(new_lod_tensor.NumElements(0UL), lod_tensor->NumElements(level));
ASSERT_EQ(new_lod_tensor.tensor(), lod_tensor->tensor());
}
// slice 2 level
for (size_t level = 0; level < 2UL; ++level) {
auto new_lod_tensor = lod_tensor->SliceShared(level, level + 2);
ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor->NumElements(level));
ASSERT_EQ(new_lod_tensor.NumElements(1),
lod_tensor->NumElements(level + 1));
ASSERT_EQ(new_lod_tensor.tensor(), lod_tensor->tensor());
}
}
TEST_F(LODTensorTester, SliceCopied_Level) {
TEST_F(LODTensorTester, SliceLevels) {
// slice 1 level
for (size_t level = 0; level < 3UL; ++level) {
auto new_lod_tensor =
lod_tensor->SliceCopied<float>(level, level + 1, place);
auto new_lod_tensor = lod_tensor->SliceLevels<float>(level, level + 1);
ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
ASSERT_EQ(new_lod_tensor.NumElements(0UL), lod_tensor->NumElements(level));
// ASSERT_EQ(new_lod_tensor.tensor(), lod_tensor->tensor());
// TODO(superjom) add tensor comparation here.
// ASSERT_EQ(new_lod_tensor, *lod_tensor);
}
// slice 2 level
for (size_t level = 0; level < 2UL; ++level) {
auto new_lod_tensor =
lod_tensor->SliceCopied<float>(level, level + 2, place);
auto new_lod_tensor = lod_tensor->SliceLevels<float>(level, level + 2);
ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor->NumElements(level));
ASSERT_EQ(new_lod_tensor.NumElements(1),
lod_tensor->NumElements(level + 1));
// ASSERT_EQ(new_lod_tensor.tensor(), lod_tensor->tensor());
// TODO(superjom) add tensor comparation here.
ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor->data<float>());
}
}
TEST_F(LODTensorTester, SliceShared_Element) {
size_t level = 0;
auto new_lod_tensor = lod_tensor->SliceShared(level, 0, 2);
ASSERT_EQ(new_lod_tensor.NumLevels(), 3UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
ASSERT_EQ(new_lod_tensor.NumElements(2), 8UL);
ASSERT_EQ(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor());
level = 1;
new_lod_tensor = lod_tensor->SliceShared(level, 0, 2);
ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
ASSERT_EQ(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor());
}
TEST_F(LODTensorTester, SliceCopied_Element) {
TEST_F(LODTensorTester, SliceInLevel) {
size_t level = 0;
auto new_lod_tensor = lod_tensor->SliceCopied<float>(level, 0, 2, place);
ASSERT_EQ(new_lod_tensor.NumLevels(), 3UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
ASSERT_EQ(new_lod_tensor.NumElements(2), 8UL);
ASSERT_NE(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor());
auto new_lod_tensor = lod_tensor->SliceInLevel<float>(level, 0, 2);
EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL);
EXPECT_EQ(new_lod_tensor.NumElements(0), 2UL);
EXPECT_EQ(new_lod_tensor.NumElements(1), 4UL);
EXPECT_EQ(new_lod_tensor.NumElements(2), 8UL);
ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor->data<float>());
level = 1;
new_lod_tensor = lod_tensor->SliceCopied<float>(level, 0, 2, place);
new_lod_tensor = lod_tensor->SliceInLevel<float>(level, 0, 2);
ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
ASSERT_NE(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor());
level = 1;
// LOD is
// 0 5 10
// 0 2 5 7 10
new_lod_tensor = lod_tensor->SliceCopied<float>(level, 1, 3, place);
ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
ASSERT_EQ(new_lod_tensor.lod_element(0, 0), 0UL);
ASSERT_EQ(new_lod_tensor.lod_element(0, 1), 5UL);
ASSERT_EQ(new_lod_tensor.lod_element(1, 0), 0UL);
ASSERT_EQ(new_lod_tensor.lod_element(1, 1), 2UL);
ASSERT_EQ(new_lod_tensor.lod_element(1, 2), 5UL);
ASSERT_EQ(new_lod_tensor.lod_element(1, 3), 7UL);
// TODO(superjom) compare the content of these tensors
ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor->data<float>());
}
TEST_F(LODTensorTester, ShareLOD) {
LODTensor new_lod_tensor;
new_lod_tensor.ShareLOD(*lod_tensor);
new_lod_tensor.CopyLOD(*lod_tensor);
ASSERT_EQ(new_lod_tensor.lod(), lod_tensor->lod());
}
TEST_F(LODTensorTester, CopyLOD) {
LODTensor new_lod_tensor;
new_lod_tensor.CopyLOD(*lod_tensor);
ASSERT_NE(new_lod_tensor.lod(), lod_tensor->lod());
bool equals = std::equal(lod_tensor->lod().begin(), lod_tensor->lod().end(),
new_lod_tensor.lod().begin());
ASSERT_TRUE(equals);
}
} // namespace framework
......
......@@ -12,35 +12,18 @@
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <memory>
#include "paddle/framework/op_info.h"
namespace paddle {
namespace framework {
namespace details {
/*
* Slice levels from LOD.
*
* @lod: LOD to slice.
* @level_begin: level to begin slice.
* @level_end: level to end slice.
*/
std::shared_ptr<LODTensor::LOD> SliceLOD(const LODTensor::LOD &lod,
size_t level_begin, size_t level_end);
static OpInfoMap* g_op_info_map = nullptr;
/*
* Slice elements from a level of LOD.
*
* @lod: LOD to slice.
* @level: which level to slice.
* @elem_begin: element's index to begin slice.
* @elem_end: element's index to end slice.
*/
std::shared_ptr<LODTensor::LOD> SliceLOD(const LODTensor::LOD &lod,
size_t level, size_t elem_begin,
size_t elem_end, bool tensor_shared);
} // namespace details
OpInfoMap& OpInfoMap::Instance() {
if (g_op_info_map == nullptr) {
g_op_info_map = new OpInfoMap();
}
return *g_op_info_map;
}
} // namespace framework
} // namespace paddle
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <functional>
#include <map>
#include <string>
#include <unordered_map>
#include "paddle/framework/attribute.h"
namespace paddle {
namespace framework {
class OperatorBase;
using VariableNameMap = std::map<std::string, std::vector<std::string>>;
using OpCreator = std::function<OperatorBase*(
const std::string& /*type*/, const VariableNameMap& /*inputs*/,
const VariableNameMap& /*outputs*/, const AttributeMap& /*attrs*/)>;
struct OpInfo {
OpCreator creator_;
std::string grad_op_type_;
OpProto* proto_;
OpAttrChecker* checker_;
bool HasOpProtoAndChecker() const {
return proto_ != nullptr && checker_ != nullptr;
}
const OpProto& Proto() const {
PADDLE_ENFORCE_NOT_NULL(proto_, "Operator Proto has not been registered");
PADDLE_ENFORCE(proto_->IsInitialized(),
"Operator Proto must be initialized in op info");
return *proto_;
}
const OpAttrChecker& Checker() const {
PADDLE_ENFORCE_NOT_NULL(checker_,
"Operator Checker has not been registered");
return *checker_;
}
const OpCreator& Creator() const {
PADDLE_ENFORCE_NOT_NULL(creator_,
"Operator Creator has not been registered");
return creator_;
}
bool HasGradientOp() const { return !grad_op_type_.empty(); }
};
class OpInfoMap {
public:
static OpInfoMap& Instance();
OpInfoMap(const OpInfoMap& o) = delete;
OpInfoMap(OpInfoMap&& o) = delete;
OpInfoMap& operator=(const OpInfoMap& o) = delete;
OpInfoMap& operator=(OpInfoMap&& o) = delete;
bool Has(const std::string& op_type) const {
return map_.find(op_type) != map_.end();
}
void Insert(const std::string& type, const OpInfo& info) {
PADDLE_ENFORCE(!Has(type), "Operator %s has been registered", type);
map_.insert({type, info});
}
const OpInfo& Get(const std::string& type) const {
auto it = map_.find(type);
PADDLE_ENFORCE(it != map_.end(), "Operator %s are not found", type);
return it->second;
}
template <typename Callback>
void IterAllInfo(Callback callback) {
for (auto& it : map_) {
callback(it.first, it.second);
}
}
private:
OpInfoMap() = default;
std::unordered_map<std::string, const OpInfo> map_;
};
} // namespace framework
} // namespace paddle
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
// Protocol Message for 3rd-party language binding.
//
// Paddle Python package will use `OpProto` to generate op creation methods.
// The op creation methods take user's input and generate `OpDesc` proto
// message,
// then pass `OpDesc` to C++ side and create Op pointer.
//
syntax = "proto2";
package paddle.framework;
import "attribute.proto";
// Attribute protocol message for 3rd-party language binding.
// It will store the Op support what attribute and what type.
message AttrProto {
// Supported attribute name. e.g. `scale` for cosine op.
required string name = 1;
// Supported attribute type.
required AttrType type = 2;
// Supported attribute comments. It helps 3rd-party language generate
// doc-string.
required string comment = 3;
// If that attribute is generated, it means the Paddle third language
// binding has responsibility to fill that attribute. End-User should
// not set that attribute.
optional bool generated = 4 [ default = false ];
}
// Input or output message for 3rd-party language binding.
// It contains parameter name and its comments.
message VarProto {
// Input or output name in that op creation function.
// e.g. `cos(a, b, output, ...)`, "a", "b", "output" are names.
required string name = 1;
// The comment for that input. It helps 3rd-party language generate
// doc-string.
required string comment = 2;
// Is that input/output could be a list or not.
// If so, that Op should write a attributed named `input_format` or
// `output_format`.
//
// e.g.
// If the op is a fc op, the inputs are `X`, `W`, `b`. The `X` and `W`
// could be multiple, so the multiple of `X` and `W` is True, and OpDesc
// will hold a attribute of them.
//
// The Op desc of same fc could be
// {
// "type": "fc",
// "input": ["X1", "X2", "W1", "W2", "b"],
// "output": "fc.out",
// "attrs" : {
// "input_format": [0, 2, 4, 5]
// }
// }
//
optional bool multiple = 3 [ default = false ];
// It marks that output is a temporary output. That output is not used by
// user, but used by other op internally as input. If other op is not use
// that output, it could be optimized early.
//
// Attribute temporary_index will be set in OpDesc if there is some
// outputs are temporary.
//
// output = [ "xxx.out1", "xxx.tmp", "xxx.out2"],
// attrs = {
// "temporary_index": [1]
// }
optional bool temporary = 4 [ default = false ];
// The gradient of operator can be ignored immediately
// e.g. operator AddOp, y = x1 + x2, the gradient of dy/dx1, dy/dx2
// can be ignored for the future optimized on graph.
optional bool ignore_gradient = 6;
}
// Op protocol message for 3rd-party language binding.
// It contains all information for generating op creation method.
message OpProto {
// The input information to generate op creation method.
repeated VarProto inputs = 1;
// The output information to generate op creation method.
repeated VarProto outputs = 2;
// The attribute information to generate op creation method.
repeated AttrProto attrs = 3;
// The comments for that Op. It helps 3rd-party language generate
// doc-string. The whole documentation of that Op is generated by comment,
// inputs, outputs, attrs together.
required string comment = 4;
// The type of that Op.
required string type = 5;
}
#include <gtest/gtest.h>
#include <paddle/framework/op_proto.pb.h>
TEST(TestOpProto, ALL) {
paddle::framework::OpProto proto;
{
auto ipt = proto.mutable_inputs()->Add();
*ipt->mutable_name() = "a";
*ipt->mutable_comment() = "the one input of cosine op";
}
{
auto ipt = proto.mutable_inputs()->Add();
*ipt->mutable_name() = "b";
*ipt->mutable_comment() = "the other input of cosine op";
}
{
auto opt = proto.mutable_outputs()->Add();
*opt->mutable_name() = "output";
*opt->mutable_comment() = "the output of cosine op";
}
{
auto attr = proto.mutable_attrs()->Add();
*attr->mutable_name() = "scale";
attr->set_type(paddle::framework::AttrType::FLOAT);
*attr->mutable_comment() = "the scale attribute of cosine op";
}
proto.set_type("cos");
*proto.mutable_comment() = "cosine op, output = scale * cos(a, b)";
ASSERT_TRUE(proto.IsInitialized());
}
\ No newline at end of file
......@@ -17,5 +17,45 @@ limitations under the License. */
#include <vector>
namespace paddle {
namespace framework {} // namespace framework
namespace framework {
std::unique_ptr<OperatorBase> OpRegistry::CreateOp(
const std::string& type, const VariableNameMap& inputs,
const VariableNameMap& outputs, AttributeMap attrs) {
auto& info = OpInfoMap::Instance().Get(type);
info.Checker().Check(attrs);
auto op = info.Creator()(type, inputs, outputs, attrs);
return std::unique_ptr<OperatorBase>(op);
}
static VariableNameMap ConvertOpDescVarsToVarNameMap(
const google::protobuf::RepeatedPtrField<OpDesc::Var>& op_desc_vars) {
VariableNameMap ret_val;
for (auto& var : op_desc_vars) {
auto& var_names = ret_val[var.parameter()];
auto& var_names_in_proto = var.arguments();
var_names.reserve(static_cast<size_t>(var_names_in_proto.size()));
std::copy(var_names_in_proto.begin(), var_names_in_proto.end(),
std::back_inserter(var_names));
}
return ret_val;
}
std::unique_ptr<OperatorBase> OpRegistry::CreateOp(const OpDesc& op_desc) {
VariableNameMap inputs = ConvertOpDescVarsToVarNameMap(op_desc.inputs());
VariableNameMap outputs = ConvertOpDescVarsToVarNameMap(op_desc.outputs());
AttributeMap attrs;
for (auto& attr : op_desc.attrs()) {
attrs[attr.name()] = GetAttrValue(attr);
}
return CreateOp(op_desc.type(), inputs, outputs, attrs);
}
std::unique_ptr<OperatorBase> OpRegistry::CreateGradOp(const OperatorBase& op) {
PADDLE_ENFORCE(!op.IsNetOp(), "Use framework::Backward to get backward ops");
return std::unique_ptr<OperatorBase>(BuildGradOp(&op));
}
} // namespace framework
} // namespace paddle
此差异已折叠。
......@@ -7,6 +7,7 @@ namespace paddle {
namespace framework {
class CosineOp : public OperatorBase {
public:
using OperatorBase::OperatorBase;
void Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const override {}
void InferShape(const Scope& scope) const override {}
......@@ -27,6 +28,7 @@ class CosineOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
class MyTestOp : public OperatorBase {
public:
using OperatorBase::OperatorBase;
void InferShape(const Scope& scope) const override {}
void Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const override {}
......@@ -36,8 +38,8 @@ class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
public:
MyTestOpProtoAndCheckerMaker(OpProto* proto, OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("input", "input of cosine op").SetMultiple();
AddOutput("output", "output of cosine op").SetTemporary();
AddInput("input", "input of cosine op").AsDuplicable();
AddOutput("output", "output of cosine op").AsIntermediate();
auto my_checker = [](int i) {
PADDLE_ENFORCE(i % 2 == 0, "'test_attr' must be even!");
};
......@@ -49,16 +51,24 @@ class MyTestOpProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
} // namespace framework
} // namespace paddle
REGISTER_OP(cos_sim, paddle::framework::CosineOp,
paddle::framework::CosineOpProtoAndCheckerMaker);
REGISTER_OP(my_test_op, paddle::framework::MyTestOp,
paddle::framework::MyTestOpProtoAndCheckerMaker);
static void BuildVar(const std::string& param_name,
std::initializer_list<const char*> arguments,
paddle::framework::OpDesc::Var* var) {
var->set_parameter(param_name);
for (auto& arg_name : arguments) {
var->add_arguments(arg_name);
}
}
REGISTER_OP_WITHOUT_GRADIENT(cos_sim, paddle::framework::CosineOp,
paddle::framework::CosineOpProtoAndCheckerMaker);
REGISTER_OP_WITHOUT_GRADIENT(my_test_op, paddle::framework::MyTestOp,
paddle::framework::MyTestOpProtoAndCheckerMaker);
TEST(OpRegistry, CreateOp) {
paddle::framework::OpDesc op_desc;
op_desc.set_type("cos_sim");
op_desc.add_inputs("aa");
op_desc.add_outputs("bb");
BuildVar("input", {"aa"}, op_desc.add_inputs());
BuildVar("output", {"bb"}, op_desc.add_outputs());
float scale = 3.3;
auto attr = op_desc.mutable_attrs()->Add();
......@@ -66,8 +76,7 @@ TEST(OpRegistry, CreateOp) {
attr->set_type(paddle::framework::AttrType::FLOAT);
attr->set_f(scale);
std::shared_ptr<paddle::framework::OperatorBase> op =
paddle::framework::OpRegistry::CreateOp(op_desc);
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
paddle::framework::Scope scope;
paddle::platform::CPUDeviceContext dev_ctx;
op->Run(scope, dev_ctx);
......@@ -78,8 +87,8 @@ TEST(OpRegistry, CreateOp) {
TEST(OpRegistry, IllegalAttr) {
paddle::framework::OpDesc op_desc;
op_desc.set_type("cos_sim");
op_desc.add_inputs("aa");
op_desc.add_outputs("bb");
BuildVar("input", {"aa"}, op_desc.add_inputs());
BuildVar("output", {"bb"}, op_desc.add_outputs());
auto attr = op_desc.mutable_attrs()->Add();
attr->set_name("scale");
......@@ -103,33 +112,23 @@ TEST(OpRegistry, IllegalAttr) {
TEST(OpRegistry, DefaultValue) {
paddle::framework::OpDesc op_desc;
op_desc.set_type("cos_sim");
op_desc.add_inputs("aa");
op_desc.add_outputs("bb");
BuildVar("input", {"aa"}, op_desc.add_inputs());
BuildVar("output", {"bb"}, op_desc.add_outputs());
ASSERT_TRUE(op_desc.IsInitialized());
std::shared_ptr<paddle::framework::OperatorBase> op =
paddle::framework::OpRegistry::CreateOp(op_desc);
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
paddle::framework::Scope scope;
paddle::platform::CPUDeviceContext dev_ctx;
op->Run(scope, dev_ctx);
ASSERT_EQ(op->GetAttr<float>("scale"), 1.0);
}
static void SetInputFormat(paddle::framework::OpDesc* desc) {
auto attr = desc->add_attrs();
attr->set_name("input_format");
attr->set_type(paddle::framework::INTS);
attr->mutable_ints()->Add(0);
attr->mutable_ints()->Add(1);
}
TEST(OpRegistry, CustomChecker) {
paddle::framework::OpDesc op_desc;
op_desc.set_type("my_test_op");
op_desc.add_inputs("ii");
op_desc.add_outputs("oo");
SetInputFormat(&op_desc);
BuildVar("input", {"ii"}, op_desc.add_inputs());
BuildVar("output", {"oo"}, op_desc.add_outputs());
// attr 'test_attr' is not set
bool caught = false;
......@@ -169,7 +168,6 @@ TEST(OpRegistry, CustomChecker) {
attr->set_name("test_attr");
attr->set_type(paddle::framework::AttrType::INT);
attr->set_i(4);
SetInputFormat(&op_desc);
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
paddle::platform::CPUDeviceContext dev_ctx;
paddle::framework::Scope scope;
......
此差异已折叠。
此差异已折叠。
......@@ -23,20 +23,22 @@ static int op_run_num = 0;
class OpWithoutKernelTest : public OperatorBase {
public:
void Init() override { x = 1; }
OpWithoutKernelTest(const std::string& type, const VariableNameMap& inputs,
const VariableNameMap& outputs, const AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs), x(1) {}
void InferShape(const Scope& scope) const override {}
void Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const override {
op_run_num++;
ASSERT_EQ((int)inputs_.size(), 1);
ASSERT_EQ((int)outputs_.size(), 1);
ASSERT_EQ(scope.FindVar(inputs_[0]), nullptr);
++op_run_num;
ASSERT_EQ(static_cast<int>(inputs_.size()), 1);
ASSERT_EQ(static_cast<int>(outputs_.size()), 1);
ASSERT_EQ(scope.FindVar(inputs_.at("input")[0]), nullptr);
ASSERT_EQ(x, 1);
ASSERT_NE(scope.FindVar(outputs_[0]), nullptr);
ASSERT_NE(scope.FindVar(outputs_.at("output")[0]), nullptr);
}
public:
float x = 0;
int x{0};
};
class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
......@@ -54,14 +56,25 @@ class OpeWithoutKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
} // namespace framework
} // namespace paddle
REGISTER_OP(test_operator, paddle::framework::OpWithoutKernelTest,
paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker);
static void BuildVar(const std::string& param_name,
std::initializer_list<const char*> arguments,
paddle::framework::OpDesc::Var* var) {
var->set_parameter(param_name);
for (auto& arg_name : arguments) {
*var->mutable_arguments()->Add() = arg_name;
}
}
REGISTER_OP_WITHOUT_GRADIENT(
test_operator, paddle::framework::OpWithoutKernelTest,
paddle::framework::OpeWithoutKernelTestProtoAndCheckerMaker);
TEST(OperatorBase, all) {
paddle::framework::OpDesc op_desc;
op_desc.set_type("test_operator");
*op_desc.mutable_inputs()->Add() = "IN1";
*op_desc.mutable_outputs()->Add() = "OUT1";
BuildVar("input", {"IN1"}, op_desc.add_inputs());
BuildVar("output", {"OUT1"}, op_desc.add_outputs());
auto attr = op_desc.mutable_attrs()->Add();
attr->set_name("scale");
attr->set_type(paddle::framework::AttrType::FLOAT);
......@@ -97,6 +110,9 @@ class OpKernelTestProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
static int cpu_kernel_run_num = 0;
class OpWithKernelTest : public OperatorWithKernel {
public:
using OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(const framework::InferShapeContext& ctx) const override {}
};
......@@ -113,33 +129,15 @@ class CPUKernelTest : public OpKernel {
}
};
// multiple inputs test
class OperatorMultiInputsTest : public OperatorBase {
public:
void Init() override { x = 1; }
void InferShape(const Scope& scope) const override {}
void Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const override {
ASSERT_EQ(scope.FindVar(inputs_[0]), nullptr);
ASSERT_EQ(x, 1);
ASSERT_NE(scope.FindVar(outputs_[0]), nullptr);
ASSERT_EQ(Input("x"), "IN1");
ASSERT_EQ(Input("y"), "OUT1");
}
public:
float x = 0;
};
class OpKernelTestMultiInputsProtoAndCheckerMaker
: public OpProtoAndCheckerMaker {
public:
OpKernelTestMultiInputsProtoAndCheckerMaker(OpProto* proto,
OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("xs", "inputs of test op").SetMultiple();
AddInput("xs", "inputs of test op").AsDuplicable();
AddInput("k", "input of test op");
AddOutput("ys", "outputs of test op").SetMultiple();
AddOutput("ys", "outputs of test op").AsDuplicable();
AddAttr<float>("scale", "scale of cosine op")
.SetDefault(1.0)
.LargerThan(0.0);
......@@ -187,8 +185,9 @@ class CPUKernalMultiInputsTest : public OpKernel {
} // namespace framework
} // namespace paddle
REGISTER_OP(op_with_kernel, paddle::framework::OpWithKernelTest,
paddle::framework::OpKernelTestProtoAndCheckerMaker);
REGISTER_OP_WITHOUT_GRADIENT(
op_with_kernel, paddle::framework::OpWithKernelTest,
paddle::framework::OpKernelTestProtoAndCheckerMaker);
REGISTER_OP_CPU_KERNEL(op_with_kernel,
paddle::framework::CPUKernelTest<float, float>);
......@@ -196,8 +195,9 @@ REGISTER_OP_CPU_KERNEL(op_with_kernel,
TEST(OpKernel, all) {
paddle::framework::OpDesc op_desc;
op_desc.set_type("op_with_kernel");
*op_desc.mutable_inputs()->Add() = "IN1";
*op_desc.mutable_outputs()->Add() = "OUT1";
BuildVar("x", {"IN1"}, op_desc.add_inputs());
BuildVar("y", {"OUT1"}, op_desc.add_outputs());
auto attr = op_desc.mutable_attrs()->Add();
attr->set_name("scale");
attr->set_type(paddle::framework::AttrType::FLOAT);
......@@ -212,8 +212,9 @@ TEST(OpKernel, all) {
ASSERT_EQ(paddle::framework::cpu_kernel_run_num, 1);
}
REGISTER_OP(op_multi_inputs_with_kernel, paddle::framework::OpWithKernelTest,
paddle::framework::OpKernelTestMultiInputsProtoAndCheckerMaker);
REGISTER_OP_WITHOUT_GRADIENT(
op_multi_inputs_with_kernel, paddle::framework::OpWithKernelTest,
paddle::framework::OpKernelTestMultiInputsProtoAndCheckerMaker);
REGISTER_OP_CPU_KERNEL(op_multi_inputs_with_kernel,
paddle::framework::CPUKernalMultiInputsTest);
......@@ -223,32 +224,15 @@ TEST(OpKernel, multi_inputs) {
OpDesc op_desc;
op_desc.set_type("op_multi_inputs_with_kernel");
*op_desc.mutable_inputs()->Add() = "x0";
*op_desc.mutable_inputs()->Add() = "x1";
*op_desc.mutable_inputs()->Add() = "x2";
*op_desc.mutable_inputs()->Add() = "k0";
*op_desc.mutable_outputs()->Add() = "y0";
*op_desc.mutable_outputs()->Add() = "y1";
BuildVar("xs", {"x0", "x1", "x2"}, op_desc.add_inputs());
BuildVar("k", {"k0"}, op_desc.add_inputs());
BuildVar("ys", {"y0", "y1"}, op_desc.add_outputs());
auto attr = op_desc.mutable_attrs()->Add();
attr->set_name("scale");
attr->set_type(paddle::framework::AttrType::FLOAT);
attr->set_f(3.14);
auto attr0 = op_desc.mutable_attrs()->Add();
attr0->set_name("input_format");
attr0->set_type(paddle::framework::AttrType::INTS);
auto input_format = attr0->mutable_ints();
input_format->Add(0); // x0
input_format->Add(3); // k
input_format->Add(4); // end
auto attr1 = op_desc.mutable_attrs()->Add();
attr1->set_name("output_format");
attr1->set_type(paddle::framework::AttrType::INTS);
auto output_format = attr1->mutable_ints();
output_format->Add(0); // y0
output_format->Add(2); // y1
paddle::platform::CPUDeviceContext cpu_device_context;
paddle::framework::Scope scope;
scope.NewVar("x0")->GetMutable<Tensor>();
......@@ -261,3 +245,22 @@ TEST(OpKernel, multi_inputs) {
auto op = paddle::framework::OpRegistry::CreateOp(op_desc);
op->Run(scope, cpu_device_context);
}
class OperatorClone : public paddle::framework::OperatorBase {
public:
DEFINE_OP_CLONE_METHOD(OperatorClone);
OperatorClone(const std::string& type,
const paddle::framework::VariableNameMap& inputs,
const paddle::framework::VariableNameMap& outputs,
const paddle::framework::AttributeMap& attrs)
: OperatorBase(type, inputs, outputs, attrs) {}
void InferShape(const paddle::framework::Scope& scope) const override {}
void Run(const paddle::framework::Scope& scope,
const paddle::platform::DeviceContext& dev_ctx) const override {}
};
TEST(Operator, Clone) {
OperatorClone a("ABC", {}, {}, {});
auto b = a.Clone();
ASSERT_EQ(a.Type(), b->Type());
}
\ No newline at end of file
......@@ -79,11 +79,11 @@ class Tensor {
inline const DDim& dims() const;
/*! Resize the dimensions of the memory block. */
inline void Resize(const DDim& dims);
inline Tensor& Resize(const DDim& dims);
/*! The internal of two tensors share the same memory block. */
template <typename T>
inline void ShareDataWith(const Tensor& src);
inline Tensor& ShareDataWith(const Tensor& src);
/**
* @brief Copy the content of external tensor to a new place.
......@@ -105,6 +105,11 @@ class Tensor {
template <typename T>
inline Tensor Slice(const int& begin_idx, const int& end_idx) const;
platform::Place place() const {
PADDLE_ENFORCE_NOT_NULL(holder_, "Tensor get place() must contains holder");
return holder_->place();
}
private:
template <typename T>
inline void check_memory_size() const;
......
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册