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

revert-3824-remove_grad_op_type
superjom 8 years ago
commit b1b71eab6c

3
.gitignore vendored

@ -24,4 +24,5 @@ cmake-build-*
python/paddle/v2/framework/core.so
CMakeFiles
cmake_install.cmake
paddle/.timestamp
python/paddlepaddle.egg-info/

@ -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
- 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)
@ -36,8 +36,8 @@ include(simd)
################################ Configurations #######################################
option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" ${CUDA_FOUND})
option(WITH_AVX "Compile PaddlePaddle with AVX intrinsics" ${AVX_FOUND})
option(WITH_MKLDNN "Compile PaddlePaddle with mkl-dnn support." ${AVX_FOUND})
option(WITH_MKLML "Compile PaddlePaddle with mklml package." ${AVX_FOUND})
option(WITH_MKLDNN "Compile PaddlePaddle with mkl-dnn support." OFF)
option(WITH_MKLML "Compile PaddlePaddle with mklml package." OFF)
option(WITH_DSO "Compile PaddlePaddle with dynamic linked CUDA" ON)
option(WITH_TESTING "Compile PaddlePaddle with unit testing" ON)
option(WITH_SWIG_PY "Compile PaddlePaddle with inference api" ON)
@ -121,8 +121,8 @@ include(version) # set PADDLE_VERSION
include(coveralls) # set code coverage
include_directories("${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})
@ -144,7 +144,7 @@ if(WITH_GPU)
endif(WITH_GPU)
if(WITH_MKLDNN)
list(APPEND EXTERNAL_LIBS ${MKLDNN_LIBRARY} ${MKLDNN_IOMP_LIB})
list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB} ${MKLDNN_IOMP_LIB})
endif()
if(USE_NNPACK)
@ -164,10 +164,12 @@ if(WITH_GOLANG)
add_subdirectory(go)
endif(WITH_GOLANG)
set(PADDLE_PYTHON_BUILD_DIR "${CMAKE_CURRENT_BINARY_DIR}/python/build")
add_subdirectory(paddle)
if(WITH_PYTHON)
add_subdirectory(python)
endif()
if(WITH_DOC)
add_subdirectory(doc)
endif()

@ -28,7 +28,7 @@ RUN apt-get update && \
wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
curl sed grep graphviz libjpeg-dev zlib1g-dev \
python-matplotlib gcc-4.8 g++-4.8 \
automake locales clang-format-3.8 swig doxygen cmake \
automake locales clang-format swig doxygen cmake \
liblapack-dev liblapacke-dev libboost-dev \
clang-3.8 llvm-3.8 libclang-3.8-dev \
net-tools && \
@ -64,13 +64,28 @@ RUN pip install --upgrade pip && \
pip install -U sphinx-rtd-theme==0.1.9 recommonmark && \
pip install pre-commit 'requests==2.9.2' 'ipython==5.3.0' && \
pip install 'ipykernel==4.6.0' 'jupyter==1.0.0' && \
pip install rarfile
pip install opencv-python rarfile 'scipy>=0.19.0' 'nltk>=3.2.2'
# 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]
# TODO(qijun) The template library Eigen doesn't work well with GCC 5
# coming with the default Docker image, so we switch to use GCC 4.8
# by default. And I will check Eigen library later.
RUN ln -sf gcc-4.8 /usr/bin/gcc && \
ln -sf gcc-ar-4.8 /usr/bin/gcc-ar && \
ln -sf gcc-nm-4.8 /usr/bin/gcc-nm && \
ln -sf gcc-ranlib-4.8 /usr/bin/gcc-ranlib && \
ln -sf gcc-4.8 /usr/bin/x86_64-linux-gnu-gcc && \
ln -sf gcc-ar-4.8 /usr/bin/x86_64-linux-gnu-gcc-ar && \
ln -sf gcc-nm-4.8 /usr/bin/x86_64-linux-gnu-gcc-nm && \
ln -sf gcc-ranlib-4.8 /usr/bin/x86_64-linux-gnu-gcc-ranlib && \
ln -sf g++-4.8 /usr/bin/g++ && \
ln -sf g++-4.8 /usr/bin/x86_64-linux-gnu-g++
# Install woboq_codebrowser to /woboq
RUN git clone https://github.com/woboq/woboq_codebrowser /woboq && \
(cd /woboq \

@ -129,7 +129,7 @@ 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"

@ -9,10 +9,12 @@ 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)
if(NOT ANDROID)
# 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()
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"

@ -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)

@ -257,6 +257,11 @@ seq_concat
.. autoclass:: paddle.v2.layer.seq_concat
:noindex:
kmax_sequence_score
-------------------
.. autoclass:: paddle.v2.layer.kmax_sequence_score
:noindex:
sub_nested_seq
--------------
.. autoclass:: paddle.v2.layer.sub_nested_seq

@ -0,0 +1,146 @@
## 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
1The 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)

@ -13,22 +13,18 @@
# serve to show the default.
import sys
import os, subprocess
sys.path.insert(0, os.path.abspath('@PADDLE_SOURCE_DIR@/python'))
import shlex
from recommonmark import parser, transform
try:
import py_paddle
import paddle
import paddle.v2
except ImportError:
print("Must install paddle python package before generating documentation")
sys.exit(1)
import paddle
import paddle.v2
MarkdownParser = parser.CommonMarkParser
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 ------------------------------------------------
@ -124,7 +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'

@ -13,15 +13,11 @@
# serve to show the default.
import sys
import os, subprocess
sys.path.insert(0, os.path.abspath('@PADDLE_SOURCE_DIR@/python'))
import shlex
from recommonmark import parser, transform
try:
import py_paddle
import paddle
import paddle.v2
except ImportError:
print("Must install paddle python package before generating documentation")
sys.exit(1)
import paddle
import paddle.v2
MarkdownParser = parser.CommonMarkParser
@ -29,7 +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 ------------------------------------------------
@ -124,7 +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'

10
go/glide.lock generated

@ -1,5 +1,5 @@
hash: 1b9b07408ca7fac27a374dc2ccd2433e4bff090484008a037df967284949a582
updated: 2017-08-03T21:46:51.744995189Z
updated: 2017-08-07T23:37:48.867469328Z
imports:
- name: github.com/beorn7/perks
version: 4c0e84591b9aa9e6dcfdf3e020114cd81f89d5f9
@ -10,7 +10,7 @@ imports:
- name: github.com/cockroachdb/cmux
version: 112f0506e7743d64a6eb8fedbcff13d9979bbf92
- name: github.com/coreos/etcd
version: c31bec0f29facff13f7c3e3d948e55dd6689ed42
version: d0d1a87aa96ae14914751d42264262cb69eda170
subpackages:
- alarm
- auth
@ -24,6 +24,7 @@ imports:
- error
- etcdserver
- etcdserver/api
- etcdserver/api/etcdhttp
- etcdserver/api/v2http
- etcdserver/api/v2http/httptypes
- etcdserver/api/v3client
@ -210,11 +211,6 @@ testImports:
version: 04cdfd42973bb9c8589fd6a731800cf222fde1a9
subpackages:
- spew
- name: github.com/docker/docker
version: b6d164e6c46d8115b146e4c3ac93784e9ef8b49e
subpackages:
- pkg/ioutils
- pkg/longpath
- name: github.com/pmezard/go-difflib
version: d8ed2627bdf02c080bf22230dbb337003b7aba2d
subpackages:

@ -1,24 +1,30 @@
package master_test
import (
"io/ioutil"
"net/url"
"os"
"strings"
"testing"
"time"
"github.com/PaddlePaddle/Paddle/go/master"
"github.com/coreos/etcd/clientv3"
"github.com/coreos/etcd/embed"
"github.com/docker/docker/pkg/ioutils"
"github.com/stretchr/testify/assert"
)
func TestNewServiceWithEtcd(t *testing.T) {
// setup an embed etcd server
etcdDir, err := ioutils.TempDir("", "")
etcdDir, err := ioutil.TempDir("", "")
if err != nil {
t.Fatal(err)
}
cfg := embed.NewConfig()
lpurl, _ := url.Parse("http://localhost:0")
lcurl, _ := url.Parse("http://localhost:0")
cfg.LPUrls = []url.URL{*lpurl}
cfg.LCUrls = []url.URL{*lcurl}
cfg.Dir = etcdDir
e, err := embed.StartEtcd(cfg)
if err != nil {
@ -30,15 +36,13 @@ func TestNewServiceWithEtcd(t *testing.T) {
t.Fatal(err)
}
}()
select {
case <-e.Server.ReadyNotify():
t.Log("Server is ready!")
case <-time.After(60 * time.Second):
e.Server.Stop() // trigger a shutdown
t.Fatal("Server took too long to start!")
}
ep := []string{"127.0.0.1:2379"}
<-e.Server.ReadyNotify()
port := strings.Split(e.Clients[0].Addr().String(), ":")[1]
endpoint := "127.0.0.1:" + port
ep := []string{endpoint}
masterAddr := "127.0.0.1:3306"
store, err := master.NewEtcdClient(ep, masterAddr, master.DefaultLockPath, master.DefaultAddrPath, master.DefaultStatePath, 30)
if err != nil {

@ -90,8 +90,12 @@ func cArrayToSlice(p unsafe.Pointer, len int) []byte {
type selector bool
func (s selector) Select() bool {
return bool(s)
func (s selector) Select() (bool, error) {
return bool(s), nil
}
func (s selector) Done() error {
return nil
}
type lister []client.Server
@ -114,11 +118,10 @@ func paddle_new_pserver_client(addrs *C.char, selected int) C.paddle_pserver_cli
}
//export paddle_new_etcd_pserver_client
func paddle_new_etcd_pserver_client(etcdEndpoints *C.char, selected int) C.paddle_pserver_client {
// TODO(Longfei: use etcd lock to decide which trainer to initialize the parameters)
func paddle_new_etcd_pserver_client(etcdEndpoints *C.char) C.paddle_pserver_client {
addr := C.GoString(etcdEndpoints)
etcdClient := client.NewEtcd(addr)
c := client.NewClient(etcdClient, etcdClient.Desired(), selector(selected != 0))
c := client.NewClient(etcdClient, etcdClient.Desired(), etcdClient)
return add(c)
}
@ -136,7 +139,12 @@ func paddle_pserver_client_release(client C.paddle_pserver_client) {
//export paddle_begin_init_params
func paddle_begin_init_params(client C.paddle_pserver_client) C.int {
c := get(client)
if selected := c.BeginInitParams(); selected {
selected, err := c.BeginInitParams()
if err != nil {
panic(err)
}
if selected {
return 1
}
return 0

@ -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)

@ -27,9 +27,13 @@ import (
// TODO(helin): add RPC call retry logic
// Selector selects if the client should initialize parameter servers.
// Selector selects if the client should initialize parameters and
// reports the initialization process done.
type Selector interface {
Select() bool
// Select selects if the client should initialize parameter servers.
Select() (bool, error)
// Done indicates the initialization process is done.
Done() error
}
// Server is the identification of a parameter Server.
@ -115,7 +119,7 @@ func (c *Client) monitorPservers(l Lister, pserverNum int) {
// servers. Other trainers will be blocked until the initialization is
// done, and they need to get the initialized parameters from
// parameter servers using GetParams.
func (c *Client) BeginInitParams() bool {
func (c *Client) BeginInitParams() (bool, error) {
return c.sel.Select()
}

@ -124,8 +124,12 @@ func initEtcdClient() {
type selector bool
func (s selector) Select() bool {
return bool(s)
func (s selector) Select() (bool, error) {
return bool(s), nil
}
func (s selector) Done() error {
return nil
}
type lister []client.Server
@ -135,7 +139,11 @@ func (l lister) List() []client.Server {
}
func testClient(t *testing.T, c *client.Client) {
selected := c.BeginInitParams()
selected, err := c.BeginInitParams()
if err != nil {
t.Fatal(err)
}
if !selected {
t.Fatal("should be selected.")
}

@ -16,53 +16,60 @@ package client
import (
"context"
"errors"
"fmt"
"strconv"
"strings"
"time"
"github.com/PaddlePaddle/Paddle/go/pserver"
"github.com/coreos/etcd/clientv3"
"github.com/coreos/etcd/clientv3/concurrency"
log "github.com/sirupsen/logrus"
)
const (
defaultEtcdTimeout time.Duration = 5 * time.Second
initLockPath = "/init_ps/lock"
initDonePath = "/init_ps/done"
initDoneVal = "1"
)
// EtcdClient is used by pserver client that is a part of trainer process.
// Etcd is used by pserver client that is a part of trainer process.
// TODO:
// 1. add watcher to watch the change state of pservers)
// 1. add etcd lock)
type EtcdClient struct {
// 1. add watcher to watch the change state of pservers.
type Etcd struct {
client *clientv3.Client
timeout time.Duration
endpoints []string
lock *concurrency.Mutex
}
// Desired read ps desired number from etcd.
func (p *EtcdClient) Desired() int {
func (e *Etcd) Desired() int {
var psDesired int
for {
ctx, cancel := context.WithTimeout(context.Background(), p.timeout)
resp, err := p.client.Get(ctx, pserver.PsDesired)
ctx, cancel := context.WithTimeout(context.Background(), e.timeout)
resp, err := e.client.Get(ctx, pserver.PsDesired)
cancel()
if err != nil {
log.Errorf("Get ps dresire number failed! recnnectiong..., %v", err)
time.Sleep(p.timeout)
time.Sleep(e.timeout)
continue
}
kvs := resp.Kvs
if len(kvs) == 0 {
log.Infoln("Waiting for ps desired registered ...")
time.Sleep(p.timeout)
time.Sleep(e.timeout)
continue
}
psDesired, err = strconv.Atoi(string(resp.Kvs[0].Value))
if err != nil {
log.Errorf("psDesired %d invalid %v", psDesired, err)
time.Sleep(p.timeout)
time.Sleep(e.timeout)
continue
}
@ -73,26 +80,26 @@ func (p *EtcdClient) Desired() int {
}
// List return the pserver list read from etcd.
func (p *EtcdClient) List() []Server {
psDesired := p.Desired()
func (e *Etcd) List() []Server {
psDesired := e.Desired()
servers := make([]Server, psDesired)
for {
for i := 0; i < psDesired; i++ {
ctx, cancel := context.WithTimeout(context.Background(), p.timeout)
ctx, cancel := context.WithTimeout(context.Background(), e.timeout)
psKey := pserver.PsPath + strconv.Itoa(i)
log.Debugf("checking %s", psKey)
resp, err := p.client.Get(ctx, psKey)
resp, err := e.client.Get(ctx, psKey)
cancel()
if err != nil {
log.Infof("Get psKey= %s error, %v", psKey, err)
time.Sleep(p.timeout)
time.Sleep(e.timeout)
continue
}
kvs := resp.Kvs
if len(kvs) == 0 {
log.Infof("Waiting for ps addr registered ...")
time.Sleep(p.timeout)
time.Sleep(e.timeout)
continue
}
@ -100,7 +107,7 @@ func (p *EtcdClient) List() []Server {
// TODO(Longfei) check the ps address
if psAddr == "" {
log.Infof("Get psKey = %s, psAddr is empty", psKey)
time.Sleep(p.timeout)
time.Sleep(e.timeout)
continue
}
log.Debugf("got value (%s) for key: %s", psAddr, psKey)
@ -113,7 +120,7 @@ func (p *EtcdClient) List() []Server {
}
// NewEtcd create a etcd client to return the state of pserver on etcd.
func NewEtcd(endpoints string) *EtcdClient {
func NewEtcd(endpoints string) *Etcd {
ep := strings.Split(endpoints, ",")
var cli *clientv3.Client
var err error
@ -130,10 +137,118 @@ func NewEtcd(endpoints string) *EtcdClient {
break
}
log.Infof("Connected to etcd: %s\n", endpoints)
client := &EtcdClient{
client := &Etcd{
client: cli,
timeout: defaultEtcdTimeout,
endpoints: ep,
}
return client
}
// Select indicates if the current trainer is selected to initialize
// the pserver parameters.
func (e *Etcd) Select() (bool, error) {
sess, err := concurrency.NewSession(e.client, concurrency.WithTTL(5))
if err != nil {
return false, err
}
lock := concurrency.NewMutex(sess, initLockPath)
log.Infof("Trying to acquire lock at %s.", initLockPath)
// Do not use timeout context here, since we don't know how
// long does it take for other trainers to initialize the
// parameters.
err = lock.Lock(context.Background())
if err != nil {
return false, err
}
log.Infof("Successfully acquired lock at %s.", initLockPath)
get := clientv3.OpGet(initDonePath)
ctx, cancel := context.WithTimeout(context.Background(), e.timeout)
tresp, err := e.client.Txn(ctx).If(lock.IsOwner()).Then(get).Commit()
cancel()
if err != nil {
return false, err
}
if !tresp.Succeeded {
return false, errors.New("no longer the owner of the lock")
}
resp := tresp.Responses[0].GetResponseRange()
if len(resp.Kvs) == 0 {
// Key value not set, select current trainer.
e.lock = lock
log.Infoln("Trainer selected.")
return true, nil
}
if string(resp.Kvs[0].Value) == initDoneVal {
log.Infoln("Initialization is already done.")
ctx, cancel = context.WithTimeout(context.Background(), e.timeout)
err = lock.Unlock(ctx)
cancel()
if err != nil {
log.Errorln(err)
}
return false, nil
}
return false, fmt.Errorf("key %s have unexpected value: %v", initDonePath, resp.Kvs[0].Value)
}
// Done indicates the parameter initialization process is done.
func (e *Etcd) Done() error {
if e.lock == nil {
return errors.New("lock is nil, Done called unexpectedly")
}
put := clientv3.OpPut(initDonePath, initDoneVal)
ctx, cancel := context.WithTimeout(context.Background(), e.timeout)
tresp, err := e.client.Txn(ctx).If(e.lock.IsOwner()).Then(put).Commit()
cancel()
if err != nil {
return err
}
if !tresp.Succeeded {
return errors.New("no longer the owner of the lock")
}
ctx, cancel = context.WithTimeout(context.Background(), e.timeout)
err = e.lock.Unlock(ctx)
cancel()
if err != nil {
log.Errorln(err)
} else {
e.lock = nil
}
return nil
}
// Close closes the etcd client.
func (e *Etcd) Close() error {
var err error
if e.lock != nil {
ctx, cancel := context.WithTimeout(context.Background(), e.timeout)
err = e.lock.Unlock(ctx)
cancel()
if err == nil {
e.lock = nil
}
}
cErr := e.client.Close()
if cErr != nil {
if err != nil {
log.Errorln(cErr)
return err
}
return cErr
}
return err
}

@ -0,0 +1,106 @@
package client_test
import (
"io/ioutil"
"net/url"
"os"
"strings"
"sync"
"testing"
"github.com/PaddlePaddle/Paddle/go/pserver/client"
"github.com/coreos/etcd/embed"
)
func TestSelector(t *testing.T) {
etcdDir, err := ioutil.TempDir("", "")
if err != nil {
t.Fatal(err)
}
cfg := embed.NewConfig()
lpurl, _ := url.Parse("http://localhost:0")
lcurl, _ := url.Parse("http://localhost:0")
cfg.LPUrls = []url.URL{*lpurl}
cfg.LCUrls = []url.URL{*lcurl}
cfg.Dir = etcdDir
e, err := embed.StartEtcd(cfg)
if err != nil {
t.Fatal(err)
}
defer func() {
e.Close()
if err := os.RemoveAll(etcdDir); err != nil {
t.Fatal(err)
}
}()
<-e.Server.ReadyNotify()
port := strings.Split(e.Clients[0].Addr().String(), ":")[1]
endpoint := "127.0.0.1:" + port
var mu sync.Mutex
selectedCount := 0
var wg sync.WaitGroup
selectAndDone := func(c *client.Etcd) {
defer wg.Done()
selected, err := c.Select()
if err != nil {
panic(err)
}
if selected {
mu.Lock()
selectedCount++
mu.Unlock()
err = c.Done()
if err != nil {
t.Fatal(err)
}
}
}
c0 := client.NewEtcd(endpoint)
c1 := client.NewEtcd(endpoint)
c2 := client.NewEtcd(endpoint)
c3 := client.NewEtcd(endpoint)
wg.Add(3)
go selectAndDone(c0)
go selectAndDone(c1)
go selectAndDone(c2)
wg.Wait()
// simulate trainer crashed and restarted after the
// initialization process.
wg.Add(1)
go selectAndDone(c3)
wg.Wait()
mu.Lock()
if selectedCount != 1 {
t.Fatal("selected count wrong:", selectedCount)
}
mu.Unlock()
err = c0.Close()
if err != nil {
t.Fatal(err)
}
err = c1.Close()
if err != nil {
t.Fatal(err)
}
err = c2.Close()
if err != nil {
t.Fatal(err)
}
err = c3.Close()
if err != nil {
t.Fatal(err)
}
}

@ -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
}

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

Loading…
Cancel
Save