Merge branch 'develop' of https://github.com/paddlepaddle/paddle into voc_dataset

cblas_new
wanghaoshuang 8 years ago
commit b142a6b063

@ -4,6 +4,7 @@ cache:
- $HOME/.ccache
- $HOME/.cache/pip
- $TRAVIS_BUILD_DIR/build/third_party
- $TRAVIS_BUILD_DIR/build_android/third_party
sudo: required
dist: trusty
os:
@ -11,6 +12,7 @@ os:
env:
- JOB=build_doc
- JOB=check_style
- JOB=build_android
addons:
apt:
packages:

@ -28,7 +28,9 @@ if(NOT CMAKE_CROSSCOMPILING)
endif(NOT CMAKE_CROSSCOMPILING)
find_package(Git REQUIRED)
find_package(Threads REQUIRED)
find_package(Boost QUIET)
if(NOT ANDROID)
find_package(Boost QUIET)
endif()
include(simd)
@ -135,7 +137,8 @@ if(WITH_GPU)
endif(WITH_GPU)
if(USE_NNPACK)
list(APPEND EXTERNAL_LIBS ${NNPACK_LIB} ${PTHREADPOOL_LIB} "rt")
include(external/nnpack)
list(APPEND EXTERNAL_LIBS ${NNPACK_LIBS})
endif(USE_NNPACK)
add_subdirectory(proto)
@ -151,7 +154,9 @@ if(WITH_GOLANG)
endif(WITH_GOLANG)
add_subdirectory(paddle)
add_subdirectory(python)
if(WITH_PYTHON)
add_subdirectory(python)
endif()
if(WITH_DOC)
add_subdirectory(doc)
endif()

@ -25,7 +25,7 @@ COPY ./paddle/scripts/docker/root/ /root/
RUN apt-get update && \
apt-get install -y \
git python-pip python-dev openssh-server bison \
wget unzip tar xz-utils bzip2 gzip coreutils ntp \
wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
curl sed grep graphviz libjpeg-dev zlib1g-dev \
python-numpy python-matplotlib gcc g++ \
automake locales clang-format-3.8 swig doxygen cmake \

@ -14,6 +14,17 @@ RUN apt-get update && \
wget curl tar unzip gcc g++ locales clang-format-3.8 swig cmake && \
apt-get clean -y
# Install Go and glide
RUN wget -O go.tgz https://storage.googleapis.com/golang/go1.8.1.linux-amd64.tar.gz && \
tar -C /usr/local -xzf go.tgz && \
mkdir /root/gopath && \
mkdir /root/gopath/bin && \
mkdir /root/gopath/src && \
rm go.tgz
ENV GOROOT=/usr/local/go GOPATH=/root/gopath
# should not be in the same line with GOROOT definition, otherwise docker build could not find GOROOT.
ENV PATH=${PATH}:${GOROOT}/bin:${GOPATH}/bin
# git credential to skip password typing
RUN git config --global credential.helper store

@ -102,12 +102,19 @@ if(WITH_GOLANG)
message(FATAL_ERROR "no glide executeble found: $ENV{GOPATH}/bin/glide")
endif()
add_custom_target(go_vendor)
add_custom_command(TARGET go_vendor
# this command will only run when the file it depends is missing
# or has changed, or the output is missing.
add_custom_command(OUTPUT ${CMAKE_BINARY_DIR}/glide
COMMAND env GOPATH=${GOPATH} ${GLIDE} install
COMMAND touch ${CMAKE_BINARY_DIR}/glide
DEPENDS ${PROJ_ROOT}/go/glide.lock
WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go"
)
add_dependencies(go_vendor go_path)
)
# depends on the custom command which outputs
# ${CMAKE_BINARY_DIR}/glide, the custom command does not need to
# run every time this target is built.
add_custom_target(go_vendor DEPENDS ${CMAKE_BINARY_DIR}/glide go_path)
endif()
endif(WITH_GOLANG)

@ -27,7 +27,8 @@ set(IGNORE_PATTERN
.*cblas\\.h.*
.*\\.pb\\.txt
.*LtrDataProvider.*
.*MultiDataProvider.*)
.*MultiDataProvider.*
.*pb.*)
# add_style_check_target
#
@ -52,14 +53,13 @@ macro(add_style_check_target TARGET_NAME)
endif()
endforeach()
if(LINT MATCHES ON)
# cpplint code style
get_filename_component(base_filename ${filename} NAME)
set(CUR_GEN ${CMAKE_CURRENT_BINARY_DIR}/${base_filename}.cpplint)
add_custom_command(OUTPUT ${CUR_GEN}
PRE_BUILD
COMMAND env ${py_env} "${PYTHON_EXECUTABLE}" "${PROJ_ROOT}/paddle/scripts/cpplint.py"
"--filter=${STYLE_FILTER}"
"--write-success=${CUR_GEN}" ${filename}
DEPENDS ${filename}
add_custom_command(TARGET ${TARGET_NAME} PRE_BUILD
COMMAND "${PYTHON_EXECUTABLE}" "${PROJ_ROOT}/paddle/scripts/cpplint.py"
"--filter=${STYLE_FILTER}"
"--write-success=${CUR_GEN}" ${filename}
WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
endif()
endforeach()

@ -106,6 +106,10 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
SET(CMAKE_SYSTEM_PROCESSOR armv7-a)
ENDIF()
ENDIF()
IF(ANDROID_ABI STREQUAL "arm64-v8a")
SET(ANDROID_TOOLCHAIN_NAME aarch64-linux-android)
SET(CMAKE_SYSTEM_PROCESSOR aarch64)
ENDIF()
SET(ANDROID_TOOLCHAIN_PREFIX "${ANDROID_TOOLCHAIN_ROOT}/bin/${ANDROID_TOOLCHAIN_NAME}-")
ENDIF()
@ -162,6 +166,10 @@ IF("${CMAKE_VERSION}" VERSION_LESS "3.7.0")
ENDIF()
ENDIF()
IF(ANDROID_ABI STREQUAL "arm64-v8a")
LIST(APPEND ANDROID_COMPILER_FLAGS -march=armv8-a)
ENDIF()
STRING(REPLACE ";" " " ANDROID_COMPILER_FLAGS "${ANDROID_COMPILER_FLAGS}")
STRING(REPLACE ";" " " ANDROID_LINKER_FLAGS "${ANDROID_LINKER_FLAGS}")
@ -186,6 +194,10 @@ ELSE()
SET(CMAKE_ANDROID_STANDALONE_TOOLCHAIN ${ANDROID_STANDALONE_TOOLCHAIN})
ENDIF()
SET(CMAKE_ANDROID_ARCH_ABI ${ANDROID_ABI})
SET(CMAKE_ANDROID_ARM_MODE ${ANDROID_ARM_MODE})
SET(CMAKE_ANDROID_ARM_NEON ${ANDROID_ARM_NEON})
IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
SET(CMAKE_ANDROID_ARM_MODE ${ANDROID_ARM_MODE})
IF(ANDROID_ABI STREQUAL "armeabi-v7a")
SET(CMAKE_ANDROID_ARM_NEON ${ANDROID_ARM_NEON})
ENDIF()
ENDIF()
ENDIF()

@ -52,6 +52,7 @@ ExternalProject_Add(
ADD_LIBRARY(glog STATIC IMPORTED GLOBAL)
SET_PROPERTY(TARGET glog PROPERTY IMPORTED_LOCATION ${GLOG_LIBRARIES})
ADD_DEPENDENCIES(glog extern_glog)
ADD_DEPENDENCIES(glog extern_glog gflags)
LINK_LIBRARIES(glog gflags)
LIST(APPEND external_project_dependencies glog)

@ -7,10 +7,24 @@ set(NNPACK_ROOT $ENV{NNPACK_ROOT} CACHE PATH "Folder contains NNPACK")
find_path(NNPACK_INC_DIR nnpack.h PATHS ${NNPACK_ROOT}/include)
find_library(NNPACK_LIB NAMES nnpack PATHS ${NNPACK_ROOT}/lib)
find_library(PTHREADPOOL_LIB NAMES pthreadpool PATHS ${NNPACK_ROOT}/lib)
find_library(NNPACK_UKERNELS_LIB NAMES nnpack_ukernels PATHS ${NNPACK_ROOT}/lib)
find_library(NNPACK_CPUFEATURES_LIB NAMES cpufeatures PATHS ${NNPACK_ROOT}/lib)
if(NNPACK_INC_DIR AND NNPACK_LIB AND PTHREADPOOL_LIB)
set(NNPACK_FOUND ON)
INCLUDE_DIRECTORIES(${NNPACK_INC_DIR})
set(NNPACK_LIBS)
list(APPEND NNPACK_LIBS ${NNPACK_LIB} ${PTHREADPOOL_LIB})
if (NNPACK_UKERNELS_LIB)
list(APPEND NNPACK_LIBS ${NNPACK_UKERNELS_LIB})
endif()
if (NNPACK_CPUFEATURES_LIB)
list(APPEND NNPACK_LIBS ${NNPACK_CPUFEATURES_LIB})
endif()
if(NOT ANDROID)
list(APPEND NNPACK_LIBS "rt")
endif()
else()
message(FATAL_ERROR "Cannot find NNPACK in (${NNPACK_ROOT})")
endif()

@ -32,7 +32,12 @@ IF(NOT ${CBLAS_FOUND})
# arm_soft_fp_abi branch of OpenBLAS to support softfp
# https://github.com/xianyi/OpenBLAS/tree/arm_soft_fp_abi
SET(OPENBLAS_COMMIT "b5c96fcfcdc82945502a2303116a64d89985daf5")
SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=ARMV7 ARM_SOFTFP_ABI=1 USE_THREAD=0)
IF(ANDROID_ABI MATCHES "^armeabi(-v7a)?$")
SET(TARGET "ARMV7")
ELSEIF(ANDROID_ABI STREQUAL "arm64-v8a")
SET(TARGET "ARMV8")
ENDIF()
SET(OPTIONAL_ARGS HOSTCC=${HOST_C_COMPILER} TARGET=${TARGET} ARM_SOFTFP_ABI=1 USE_THREAD=0)
ELSEIF(RPI)
# use hardfp
SET(OPENBLAS_COMMIT "v0.2.19")

@ -90,11 +90,11 @@
# including binary directory for generated headers.
include_directories(${CMAKE_CURRENT_BINARY_DIR})
if(NOT APPLE)
if(NOT APPLE AND NOT ANDROID)
find_package(Threads REQUIRED)
link_libraries(${CMAKE_THREAD_LIBS_INIT})
set(CMAKE_CXX_LINK_EXECUTABLE "${CMAKE_CXX_LINK_EXECUTABLE} -ldl -lrt")
endif(NOT APPLE)
endif(NOT APPLE AND NOT ANDROID)
function(merge_static_libs TARGET_NAME)
set(libs ${ARGN})
@ -104,6 +104,7 @@ function(merge_static_libs TARGET_NAME)
foreach(lib ${libs})
list(APPEND libs_deps ${${lib}_LIB_DEPENDS})
endforeach()
list(REMOVE_DUPLICATES libs_deps)
if(APPLE) # Use OSX's libtool to merge archives
# To produce a library we need at least one source file.
@ -127,7 +128,7 @@ function(merge_static_libs TARGET_NAME)
# Get the file names of the libraries to be merged
set(libfiles ${libfiles} $<TARGET_FILE:${lib}>)
endforeach()
add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
COMMAND rm "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a"
COMMAND /usr/bin/libtool -static -o "${CMAKE_CURRENT_BINARY_DIR}/lib${TARGET_NAME}.a" ${libfiles})
else() # general UNIX: use "ar" to extract objects and re-add to a common lib
@ -145,11 +146,11 @@ function(merge_static_libs TARGET_NAME)
DEPENDS ${lib} ${objdir}
WORKING_DIRECTORY ${objdir})
# Empty dummy source file that goes into merged library
set(mergebase ${lib}.mergebase.c)
add_custom_command(OUTPUT ${mergebase}
COMMAND ${CMAKE_COMMAND} -E touch ${mergebase}
DEPENDS ${objlistfile})
# Empty dummy source file that goes into merged library
set(mergebase ${lib}.mergebase.c)
add_custom_command(OUTPUT ${mergebase}
COMMAND ${CMAKE_COMMAND} -E touch ${mergebase}
DEPENDS ${objlistfile})
list(APPEND mergebases "${mergebase}")
endforeach()
@ -184,6 +185,10 @@ function(cc_library TARGET_NAME)
add_dependencies(${TARGET_NAME} ${cc_library_DEPS})
target_link_libraries(${TARGET_NAME} ${cc_library_DEPS})
endif()
# cpplint code style
add_style_check_target(${TARGET_NAME} ${cc_library_SRCS})
else(cc_library_SRCS)
if (cc_library_DEPS)
merge_static_libs(${TARGET_NAME} ${cc_library_DEPS})
@ -337,7 +342,7 @@ function(go_test TARGET_NAME)
string(REPLACE "${PADDLE_GO_PATH}" "" CMAKE_CURRENT_SOURCE_REL_DIR ${CMAKE_CURRENT_SOURCE_DIR})
add_custom_target(${TARGET_NAME} ALL DEPENDS go_vendor ${go_test_DEPS})
add_custom_command(TARGET ${TARGET_NAME} POST_BUILD
COMMAND env GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} test
COMMAND env GOPATH=${GOPATH} ${CMAKE_Go_COMPILER} test -race
-c -o "${CMAKE_CURRENT_BINARY_DIR}/${TARGET_NAME}"
".${CMAKE_CURRENT_SOURCE_REL_DIR}"
WORKING_DIRECTORY "${PADDLE_IN_GOPATH}/go")

@ -474,6 +474,11 @@ prelu
.. autoclass:: paddle.v2.layer.prelu
:noindex:
gated_unit
-----------
.. autoclass:: paddle.v2.layer.gated_unit
:noindex:
Detection output Layer
======================

@ -11,6 +11,7 @@ import (
"github.com/namsral/flag"
log "github.com/sirupsen/logrus"
"github.com/topicai/candy"
"github.com/PaddlePaddle/Paddle/go/master"
"github.com/PaddlePaddle/Paddle/go/utils/networkhelper"
@ -20,11 +21,18 @@ func main() {
port := flag.Int("port", 8080, "port of the master server.")
ttlSec := flag.Int("ttl", 60, "etcd lease TTL in seconds.")
endpoints := flag.String("endpoints", "http://127.0.0.1:2379", "comma separated etcd endpoints. If empty, fault tolerance will not be enabled.")
taskTimeoutDur := flag.Duration("task_timout_dur", 20*time.Minute, "task timout duration.")
taskTimeoutMax := flag.Int("task_timeout_max", 3, "max timtout count for each task before it being declared failed task.")
chunkPerTask := flag.Int("chunk_per_task", 10, "chunk per task.")
taskTimeoutDur := flag.Duration("task-timout-dur", 20*time.Minute, "task timout duration.")
taskTimeoutMax := flag.Int("task-timeout-max", 3, "max timtout count for each task before it being declared failed task.")
chunkPerTask := flag.Int("chunk-per-task", 10, "chunk per task.")
logLevel := flag.String("log-level", "info",
"log level, possible values: debug, info, warning, error, fatal, panic")
flag.Parse()
level, e := log.ParseLevel(*logLevel)
candy.Must(e)
log.SetLevel(level)
if *endpoints == "" {
log.Warningln("-endpoints not set, fault tolerance not be enabled.")
}

@ -40,7 +40,7 @@ func main() {
idx = *index
} else {
e = pserver.NewEtcdClient(*etcdEndpoint, *numPservers, *etcdTimeout)
idx, err = e.Register()
idx, err = e.Register(*port)
candy.Must(err)
cp, err = pserver.NewCheckpointFromFile(*checkpointPath, idx, e)

@ -2,6 +2,7 @@ package master
import (
"os"
"time"
"github.com/PaddlePaddle/Paddle/go/connection"
"github.com/PaddlePaddle/recordio"
@ -36,9 +37,9 @@ func (c *Client) getRecords() {
for {
t, err := c.getTask()
if err != nil {
// TODO(helin): wait before move on with next
// getTask call.
log.Errorln(err)
log.Errorf("Get task failed, sleep 3 seconds and continue, %s", err)
time.Sleep(3 * time.Second)
continue
}

@ -215,6 +215,7 @@ func readChunks(globPaths []string) ([]Chunk, error) {
}
count := index.NumChunks()
log.Infof("readChunks: file %s has %d chunks", path, count)
for i := 0; i < count; i++ {
chunk := Chunk{
Path: path,

@ -1,5 +1,23 @@
import paddle.v2 as paddle
import paddle.v2.dataset.uci_housing as uci_housing
import paddle.v2.master as master
import os
import cPickle as pickle
etcd_ip = os.getenv("MASTER_IP", "127.0.0.1")
etcd_endpoint = "http://" + etcd_ip + ":2379"
def cloud_reader():
print "connecting to master, etcd endpoints: ", etcd_endpoint
master_client = master.client(etcd_endpoint, 5, 64)
master_client.set_dataset(
["/pfs/dlnel/public/dataset/uci_housing/uci_housing-*-of-*"])
while 1:
r, e = master_client.next_record()
if not r:
break
yield pickle.loads(r)
def main():
@ -22,13 +40,13 @@ def main():
# create optimizer of new remote updater to pserver
optimizer = paddle.optimizer.Momentum(momentum=0)
#TODO(zhihong) : replace optimizer with new OptimizerConfig
print "etcd endoint: ", etcd_endpoint
trainer = paddle.trainer.SGD(cost=cost,
parameters=parameters,
update_equation=optimizer,
is_local=False,
pserver_spec="localhost:3000")
pserver_spec=etcd_endpoint,
use_etcd=True)
# event_handler to print training and testing info
def event_handler(event):
@ -47,11 +65,11 @@ def main():
print "Test %d, %.2f" % (event.pass_id, result.cost)
# training
# NOTE: use uci_housing.train() as reader for non-paddlecloud training
trainer.train(
reader=paddle.batch(
paddle.reader.shuffle(
uci_housing.train(), buf_size=500),
batch_size=2),
cloud_reader, buf_size=500), batch_size=2),
feeding={'x': 0,
'y': 1},
event_handler=event_handler,

@ -3,11 +3,13 @@ package client_test
import (
"context"
"io/ioutil"
"math/rand"
"net"
"net/http"
"net/rpc"
"strconv"
"strings"
"sync"
"testing"
"time"
@ -100,27 +102,34 @@ func (l lister) List() []client.Server {
return l
}
func ClientTest(t *testing.T, c *client.Client) {
func testClient(t *testing.T, c *client.Client) {
selected := c.BeginInitParams()
if !selected {
t.Fatal("should be selected.")
}
const numParameter = 100
const numParameter = 1000
config, err := ioutil.ReadFile("./c/test/testdata/optimizer.pb")
if err != nil {
t.Fatalf("read optimizer proto failed")
}
var wg sync.WaitGroup
for i := 0; i < numParameter; i++ {
var p pserver.Parameter
p.Name = "p_" + strconv.Itoa(i)
p.ElementType = pserver.Float32
p.Content = make([]byte, (i+1)*100)
err := c.InitParam(pserver.ParameterWithConfig{Param: p, Config: config})
if err != nil {
t.Fatal(err)
}
wg.Add(1)
go func(i int) {
var p pserver.Parameter
p.Name = "p_" + strconv.Itoa(i)
p.ElementType = pserver.Float32
p.Content = make([]byte, (i+1)*100)
err := c.InitParam(pserver.ParameterWithConfig{Param: p, Config: config})
if err != nil {
t.Fatal(err)
}
wg.Done()
}(i)
}
wg.Wait()
err = c.FinishInitParams()
if err != nil {
@ -128,7 +137,7 @@ func ClientTest(t *testing.T, c *client.Client) {
}
var grads []pserver.Gradient
for i := 0; i < numParameter/2; i++ {
for i := 0; i < numParameter; i++ {
var g pserver.Gradient
g.Name = "p_" + strconv.Itoa(i)
g.ElementType = pserver.Float32
@ -136,9 +145,31 @@ func ClientTest(t *testing.T, c *client.Client) {
grads = append(grads, g)
}
err = c.SendGrads(grads)
if err != nil {
t.Fatal(err)
const paramPerGroup = 10
const numGroups = numParameter / paramPerGroup
// shuffle send grads order
for i := range grads {
j := rand.Intn(i + 1)
grads[i], grads[j] = grads[j], grads[i]
}
for i := 0; i < numGroups; i++ {
var gs []pserver.Gradient
if i == numGroups-1 {
gs = grads[i*paramPerGroup:]
} else {
gs = grads[i*paramPerGroup : (i+1)*paramPerGroup]
}
wg.Add(1)
go func(gs []pserver.Gradient) {
err := c.SendGrads(gs)
if err != nil {
t.Fatal(err)
}
wg.Done()
}(gs)
}
names := make([]string, numParameter)
@ -146,20 +177,35 @@ func ClientTest(t *testing.T, c *client.Client) {
names[i] = "p_" + strconv.Itoa(i)
}
params, err := c.GetParams(names)
if err != nil {
t.Fatal(err)
}
for i := 0; i < numGroups; i++ {
var ns []string
if i == numGroups-1 {
ns = names[i*paramPerGroup:]
} else {
ns = names[i*paramPerGroup : (i+1)*paramPerGroup]
}
if len(names) != len(params) {
t.Fatalf("parameter size not match, need: %d, have: %d", len(names), len(params))
}
wg.Add(1)
go func(ns []string) {
params, err := c.GetParams(ns)
if err != nil {
t.Fatal(err)
}
for i := range params {
if names[i] != params[i].Name {
t.Fatalf("order of returned parameter does not required: parameter name: %s, required name: %s", names[i], params[i].Name)
}
if len(ns) != len(params) {
t.Fatalf("parameter size not match, need: %d, have: %d", len(names), len(params))
}
for i := range params {
if ns[i] != params[i].Name {
t.Fatalf("order of returned parameter does not required: parameter name: %s, required name: %s", ns[i], params[i].Name)
}
}
wg.Done()
}(ns)
}
wg.Wait()
}
func TestNativeClient(t *testing.T) {
@ -169,13 +215,14 @@ func TestNativeClient(t *testing.T) {
servers[i] = client.Server{Index: i, Addr: ":" + strconv.Itoa(pserverClientPorts[i])}
}
c1 := client.NewClient(lister(servers), len(servers), selector(true))
ClientTest(t, c1)
testClient(t, c1)
}
// TODO: tmperary disable etcdClient test for dependency of etcd)
// EtcdClient is a disabled test, since we have not embedded etcd into
// our test.
func EtcdClient(t *testing.T) {
initEtcdClient()
etcdClient := client.NewEtcd(etcdEndpoints)
c2 := client.NewClient(etcdClient, etcdClient.Desired(), selector(true))
ClientTest(t, c2)
testClient(t, c2)
}

@ -12,6 +12,7 @@ import (
)
const (
// DefaultEtcdTimeout is the default etcd timeout
DefaultEtcdTimeout time.Duration = 5 * time.Second
)
@ -66,12 +67,12 @@ func (p *EtcdClient) List() []Server {
for {
for i := 0; i < psDesired; i++ {
ctx, cancel := context.WithTimeout(context.Background(), p.timeout)
cancel()
psKey := pserver.PsPath + strconv.Itoa(i)
log.Debugf("checking %s", psKey)
resp, err := p.client.Get(ctx, psKey)
cancel()
if err != nil {
log.Infof("Get psKey= %s error, %v", psKey, err)
log.Infof("Get psKey=%s error, %v", psKey, err)
time.Sleep(p.timeout)
continue
}

@ -49,7 +49,7 @@ func NewEtcdClient(endpoints string, numPservers int, timeout time.Duration) *Et
// Register registers the pserver on etcd
//
// Register returns the index of the current pserver.
func (e *EtcdClient) Register() (int, error) {
func (e *EtcdClient) Register(port int) (int, error) {
var err error
e.externalIP, err = networkhelper.GetExternalIP()
@ -116,7 +116,7 @@ func (e *EtcdClient) Register() (int, error) {
for {
ctx, cancel := context.WithTimeout(context.Background(), time.Second)
var err error
pserverIdx, err = e.registerPserverEtcd(ctx)
pserverIdx, err = e.registerPserverEtcd(ctx, port)
cancel()
if err != nil {
log.Warn(err)
@ -140,7 +140,7 @@ func (e *EtcdClient) initDesiredPservers(ctx context.Context, numPservers int) (
}
// registerPserverEtcd registers pserver node on etcd using transaction.
func (e *EtcdClient) registerPserverEtcd(ctx context.Context) (int, error) {
func (e *EtcdClient) registerPserverEtcd(ctx context.Context, port int) (int, error) {
var idx int
_, err := concurrency.NewSTM(e.etcdClient, func(c concurrency.STM) error {
registered := false
@ -156,8 +156,9 @@ func (e *EtcdClient) registerPserverEtcd(ctx context.Context) (int, error) {
log.Fatal(err)
}
// find the first id and write info
c.Put(psKey, e.externalIP, clientv3.WithLease(resp.ID))
log.Debugf("set pserver node %s with value %s", psKey, e.externalIP)
pserverAddr := e.externalIP + ":" + strconv.Itoa(port)
c.Put(psKey, pserverAddr, clientv3.WithLease(resp.ID))
log.Debugf("set pserver node %s with value %s", psKey, pserverAddr)
ch, kaerr := e.etcdClient.KeepAlive(context.TODO(), resp.ID)
if kaerr != nil {
log.Errorf("keepalive etcd node error: %v", kaerr)

@ -19,6 +19,7 @@ var nullPtr = unsafe.Pointer(uintptr(0))
type optimizer struct {
opt *C.struct_paddle_optimizer
elementType ElementType
contentLen int
}
func cArrayToSlice(p unsafe.Pointer, len int) []byte {
@ -37,10 +38,11 @@ func cArrayToSlice(p unsafe.Pointer, len int) []byte {
func newOptimizer(paramWithConfigs ParameterWithConfig, State []byte) *optimizer {
o := &optimizer{}
o.elementType = paramWithConfigs.Param.ElementType
o.contentLen = len(paramWithConfigs.Param.Content)
p := paramWithConfigs.Param
c := paramWithConfigs.Config
s := State
paramBufferSize := C.size_t(len(p.Content) / C.sizeof_float)
paramBufferSize := C.size_t(len(p.Content))
log.WithFields(log.Fields{
"ElementType": p.ElementType,
"ParamSize": paramBufferSize,
@ -78,7 +80,11 @@ func (o *optimizer) UpdateParameter(g Gradient) error {
return fmt.Errorf("Name: %s, parameter and gradient element type not match, parameter: %v, gradient: %v", g.Name, o.elementType, g.ElementType)
}
r := C.paddle_update_parameter(o.opt, C.paddle_element_type(g.ElementType), unsafe.Pointer(&g.Content[0]), C.int(len(g.Content))/C.sizeof_float)
if o.contentLen != len(g.Content) {
return fmt.Errorf("Name: %s, parameter and gradient does not have same content len, parameter: %d, gradient: %d", g.Name, o.contentLen, len(g.Content))
}
r := C.paddle_update_parameter(o.opt, C.paddle_element_type(g.ElementType), unsafe.Pointer(&g.Content[0]), C.int(len(g.Content)))
if r != 0 {
return fmt.Errorf("optimizer update returned error code: %d", r)
}

@ -31,7 +31,7 @@ func TestServiceFull(t *testing.T) {
err = s.InitParam(pserver.ParameterWithConfig{Param: p, Config: config}, nil)
if err != nil {
t.FailNow()
t.Fatal(err)
}
var p1 pserver.Parameter
@ -40,40 +40,40 @@ func TestServiceFull(t *testing.T) {
p1.ElementType = pserver.Float32
err = s.InitParam(pserver.ParameterWithConfig{Param: p1, Config: config}, nil)
if err != nil {
t.FailNow()
t.Fatal(err)
}
err = s.FinishInitParams(0, nil)
if err != nil {
t.FailNow()
t.Fatal(err)
}
var param pserver.Parameter
err = s.GetParam("param_b", &param)
if err != nil {
t.FailNow()
t.Fatal(err)
}
if !reflect.DeepEqual(param, p1) {
t.FailNow()
t.Fatal("not equal:", param, p1)
}
g1, g2 := pserver.Gradient(p1), pserver.Gradient(p)
err = s.SendGrad(g1, nil)
if err != nil {
t.FailNow()
t.Fatal(err)
}
err = s.SendGrad(g2, nil)
if err != nil {
t.FailNow()
t.Fatal(err)
}
var param1 pserver.Parameter
err = s.GetParam("param_a", &param1)
if err != nil {
t.FailNow()
t.Fatal(err)
}
// don't compare content, since it's already changed by
@ -82,7 +82,7 @@ func TestServiceFull(t *testing.T) {
p.Content = nil
if !reflect.DeepEqual(param1, p) {
t.FailNow()
t.Fatal("not equal:", param1, p)
}
}
@ -90,16 +90,16 @@ func TestMultipleInit(t *testing.T) {
var cp pserver.Checkpoint
s, err := pserver.NewService(0, 1, "", nil, cp)
if err != nil {
t.Error(err)
t.Fatal(err)
}
err = s.FinishInitParams(0, nil)
if err != nil {
t.FailNow()
t.Fatal(err)
}
err = s.FinishInitParams(0, nil)
if err.Error() != pserver.AlreadyInitialized {
t.FailNow()
t.Fatal(err)
}
}
@ -108,7 +108,7 @@ func TestUninitialized(t *testing.T) {
s, err := pserver.NewService(0, 1, "", nil, cp)
err = s.SendGrad(pserver.Gradient{}, nil)
if err.Error() != pserver.Uninitialized {
t.FailNow()
t.Fatal(err)
}
}
@ -154,12 +154,12 @@ func TestBlockUntilInitialized(t *testing.T) {
err = s.InitParam(pserver.ParameterWithConfig{Param: p, Config: config}, nil)
if err != nil {
t.FailNow()
t.Fatal(err)
}
err = s.FinishInitParams(0, nil)
if err != nil {
t.FailNow()
t.Fatal(err)
}
wg.Wait()

@ -14,6 +14,7 @@ if(Boost_FOUND)
add_subdirectory(memory)
add_subdirectory(platform)
add_subdirectory(framework)
add_subdirectory(operators)
add_subdirectory(pybind)
endif()

@ -64,11 +64,7 @@ ModelConfig* TrainerConfig::getModelConfig() const {
ParameterConfig::ParameterConfig() : m(new ParameterConfigPrivate()) {}
ParameterConfig::~ParameterConfig() {
if (m) {
delete m;
}
}
ParameterConfig::~ParameterConfig() { delete m; }
ParameterConfig* ParameterConfig::createParameterConfigFromParameterSharedPtr(
void* ptr) {
@ -98,11 +94,7 @@ void* ParameterConfig::getRawPtr() { return m->getConfigPtr(); }
OptimizationConfig::OptimizationConfig() : m(new OptimizationConfigPrivate()) {}
OptimizationConfig::~OptimizationConfig() {
if (m) {
delete m;
}
}
OptimizationConfig::~OptimizationConfig() { delete m; }
std::string OptimizationConfig::toProtoString() {
return m->getConfig().SerializeAsString();

@ -843,7 +843,8 @@ public:
bool useSparseUpdater);
static ParameterUpdater* createNewRemoteUpdater(
OptimizationConfig* config,
const std::string pserverSpec) throw(UnsupportError);
const std::string pserverSpec,
const bool useEtcd) throw(UnsupportError);
~ParameterUpdater();
/**

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

Loading…
Cancel
Save