Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into improve_pruning
xzl
8 years ago
commit
23b1a27483
@ -0,0 +1,21 @@
|
||||
# Design Doc: Remote Parameter Updater for Cluster Train
|
||||
|
||||
For an overview of distribute training, please refer to [distributed training design doc](README.md). In this design doc, we will discuss the parameter updater that will use parameter server cclient [The Client Library of Parameter Server Design Doc](pserver_client.md) to manage and update parameters.
|
||||
|
||||
## Parameter Updater
|
||||
|
||||
Parameter Updater is used by trainer to manage and update parameter, there are mainly two kind of parameter updater: local and remote, since this design is for cluster train, we will only discuss remote parameter updater here.
|
||||
|
||||
### Remote Parameter Updater
|
||||
|
||||
Remote Parameter Updater manage parameters through remote parameter server with the client that communicate with pserver([The Client Library of Parameter Server Design Doc](pserver_client.md))
|
||||
|
||||
In PaddlePaddle Python V2 API, trainer is implemented in python, and the trainer will hold a instance of parameter updater and call it's functions directly. In this design, we will also expose the api of RemoteParameterUpdater to python with swig.
|
||||
|
||||
#### Sparse Remote Parameter Updater
|
||||
|
||||
Since we will only implement dense parameter management new, the mechanism for sparse parameter will be discussed in next stage.
|
||||
|
||||
### Interface Design
|
||||
|
||||
TBD
|
||||
@ -0,0 +1,82 @@
|
||||
package master
|
||||
|
||||
import (
|
||||
"log"
|
||||
"time"
|
||||
|
||||
"github.com/PaddlePaddle/Paddle/go/connection"
|
||||
)
|
||||
|
||||
// Addresser provide the address of the master server.
|
||||
type Addresser interface {
|
||||
Address() string
|
||||
}
|
||||
|
||||
// Client is the client of the master server.
|
||||
type Client struct {
|
||||
conn *connection.Conn
|
||||
}
|
||||
|
||||
// NewClient creates a new Client.
|
||||
func NewClient(addr Addresser) *Client {
|
||||
c := &Client{}
|
||||
c.conn = connection.New()
|
||||
go c.monitorMaster(addr)
|
||||
return c
|
||||
}
|
||||
|
||||
func (c *Client) monitorMaster(addr Addresser) {
|
||||
lastMaster := ""
|
||||
monitor := func() {
|
||||
// get the lastest address of the master server,
|
||||
// connect to the new address once address changed.
|
||||
curMaster := addr.Address()
|
||||
if curMaster != lastMaster {
|
||||
if curMaster == "" {
|
||||
err := c.conn.Close()
|
||||
if err != nil {
|
||||
log.Println(err)
|
||||
}
|
||||
} else {
|
||||
err := c.conn.Connect(curMaster)
|
||||
if err != nil {
|
||||
log.Println(err)
|
||||
|
||||
// connect to addr failed, set
|
||||
// to last known addr in order
|
||||
// to retry next time.
|
||||
curMaster = lastMaster
|
||||
}
|
||||
|
||||
}
|
||||
}
|
||||
|
||||
lastMaster = curMaster
|
||||
}
|
||||
|
||||
monitor()
|
||||
ticker := time.NewTicker(10 * time.Second)
|
||||
for _ = range ticker.C {
|
||||
monitor()
|
||||
}
|
||||
}
|
||||
|
||||
// SetDataset set dataset for the master server to dispatch.
|
||||
//
|
||||
// SetDataset can be call multiple times from different nodes. But
|
||||
// only the first call will be honored.
|
||||
func (c *Client) SetDataset(globPaths []string) error {
|
||||
return c.conn.Call("Service.SetDataset", globPaths, nil)
|
||||
}
|
||||
|
||||
// GetTask gets a new task from the master server.
|
||||
func (c *Client) GetTask() (Task, error) {
|
||||
var t Task
|
||||
err := c.conn.Call("Service.GetTask", 0, &t)
|
||||
return t, err
|
||||
}
|
||||
|
||||
// TaskFinished tells the master server a task is finished.
|
||||
func (c *Client) TaskFinished(taskID int) error {
|
||||
return c.conn.Call("Service.TaskFinished", taskID, nil)
|
||||
}
|
||||
@ -0,0 +1,120 @@
|
||||
package master_test
|
||||
|
||||
import (
|
||||
"fmt"
|
||||
"net"
|
||||
"net/http"
|
||||
"net/rpc"
|
||||
"os"
|
||||
"strconv"
|
||||
"strings"
|
||||
"testing"
|
||||
"time"
|
||||
|
||||
log "github.com/sirupsen/logrus"
|
||||
|
||||
"github.com/PaddlePaddle/Paddle/go/master"
|
||||
"github.com/PaddlePaddle/recordio"
|
||||
)
|
||||
|
||||
const (
|
||||
totalTask = 20
|
||||
chunkPerTask = 10
|
||||
)
|
||||
|
||||
var port int
|
||||
|
||||
func init() {
|
||||
log.SetLevel(log.ErrorLevel)
|
||||
|
||||
l, err := net.Listen("tcp", ":0")
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
|
||||
ss := strings.Split(l.Addr().String(), ":")
|
||||
p, err := strconv.Atoi(ss[len(ss)-1])
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
port = p
|
||||
|
||||
go func(l net.Listener) {
|
||||
s := master.NewService(chunkPerTask, time.Second, 1)
|
||||
server := rpc.NewServer()
|
||||
err := server.Register(s)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
|
||||
mux := http.NewServeMux()
|
||||
mux.Handle(rpc.DefaultRPCPath, server)
|
||||
err = http.Serve(l, mux)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
}(l)
|
||||
}
|
||||
|
||||
type addresser string
|
||||
|
||||
func (a addresser) Address() string {
|
||||
return string(a)
|
||||
}
|
||||
|
||||
func TestClientFull(t *testing.T) {
|
||||
const p = "/tmp/master_client_test_0"
|
||||
f, err := os.Create(p)
|
||||
if err != nil {
|
||||
panic(err)
|
||||
}
|
||||
|
||||
for i := 0; i < totalTask*chunkPerTask; i++ {
|
||||
w := recordio.NewWriter(f, -1, -1)
|
||||
w.Write(nil)
|
||||
// call Close to force RecordIO writing a chunk.
|
||||
w.Close()
|
||||
}
|
||||
f.Close()
|
||||
|
||||
c := master.NewClient(addresser(fmt.Sprintf(":%d", port)))
|
||||
c.SetDataset([]string{p})
|
||||
|
||||
checkOnePass := func(i int) {
|
||||
var tasks []master.Task
|
||||
for i := 0; i < totalTask; i++ {
|
||||
task, err := c.GetTask()
|
||||
if err != nil {
|
||||
t.Fatal(i, err)
|
||||
}
|
||||
tasks = append(tasks, task)
|
||||
}
|
||||
|
||||
_, err = c.GetTask()
|
||||
if err == nil {
|
||||
t.Fatal(i, "should get error.")
|
||||
}
|
||||
|
||||
err = c.TaskFinished(tasks[0].ID)
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
tasks = tasks[1:]
|
||||
task, err := c.GetTask()
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
tasks = append(tasks, task)
|
||||
|
||||
for _, task := range tasks {
|
||||
err = c.TaskFinished(task.ID)
|
||||
if err != nil {
|
||||
t.Fatal(i, err)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
for i := 0; i < 10; i++ {
|
||||
checkOnePass(i)
|
||||
}
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
File diff suppressed because it is too large
Load Diff
@ -1,11 +1,22 @@
|
||||
cmake_minimum_required(VERSION 3.0)
|
||||
|
||||
include_directories(${CMAKE_BINARY_DIR})
|
||||
|
||||
add_executable(main main.c)
|
||||
add_dependencies(main client)
|
||||
add_dependencies(main paddle_pserver_cclient)
|
||||
add_executable(test_cclient test_cclient.c)
|
||||
add_dependencies(test_cclient paddle_pserver_cclient)
|
||||
|
||||
if(APPLE)
|
||||
set(CMAKE_EXE_LINKER_FLAGS "-framework CoreFoundation -framework Security")
|
||||
else()
|
||||
set(CMAKE_EXE_LINKER_FLAGS "-pthread")
|
||||
endif()
|
||||
target_link_libraries(main ${CMAKE_BINARY_DIR}/libclient.a)
|
||||
|
||||
if(PROJ_ROOT)
|
||||
include_directories(${CMAKE_BINARY_DIR}/go/pserver/cclient/)
|
||||
target_link_libraries(main ${CMAKE_BINARY_DIR}/go/pserver/cclient/libpaddle_pserver_cclient.a pthread)
|
||||
target_link_libraries(test_cclient ${CMAKE_BINARY_DIR}/go/pserver/cclient/libpaddle_pserver_cclient.a pthread)
|
||||
else(PROJ_ROOT)
|
||||
include_directories(${CMAKE_BINARY_DIR})
|
||||
target_link_libraries(main ${CMAKE_BINARY_DIR}/libpaddle_pserver_cclient.a pthread)
|
||||
target_link_libraries(test_cclient ${CMAKE_BINARY_DIR}/libpaddle_pserver_cclient.a pthread)
|
||||
endif(PROJ_ROOT)
|
||||
|
||||
@ -0,0 +1,117 @@
|
||||
#include <stdio.h>
|
||||
#include <stdlib.h>
|
||||
|
||||
#include "libpaddle_pserver_cclient.h"
|
||||
|
||||
typedef float real;
|
||||
|
||||
void fail() {
|
||||
// TODO(helin): fix: gtest using cmake is not working, using this
|
||||
// hacky way for now.
|
||||
printf("test failed.\n");
|
||||
exit(-1);
|
||||
}
|
||||
|
||||
void print_parameter(paddle_gradient* param) {
|
||||
if (param == NULL) {
|
||||
printf("param is NULL!!\n");
|
||||
} else {
|
||||
printf("==== parameter ====\n");
|
||||
printf("name: %s\n", param->name);
|
||||
printf("content_len: %d\n", param->content_len);
|
||||
printf("content_type: %d\n", param->element_type);
|
||||
int i;
|
||||
for (i = 0; i < param->content_len / (int)sizeof(real); ++i) {
|
||||
printf("%f ", ((float*)param->content)[i]);
|
||||
}
|
||||
printf("\n\n");
|
||||
}
|
||||
}
|
||||
|
||||
int main() {
|
||||
char addr[] = "localhost:3000";
|
||||
paddle_pserver_client c = paddle_new_pserver_client(addr, 1);
|
||||
|
||||
char* names[] = {"param_a", "param_b"};
|
||||
|
||||
retry:
|
||||
printf("init parameter to pserver:\n");
|
||||
|
||||
real param_content1[] = {0.1, 0.2, 0.3};
|
||||
real param_content2[] = {0.4, 0.5, 0.6};
|
||||
paddle_parameter** params =
|
||||
(paddle_parameter**)malloc(sizeof(paddle_parameter*) * 2);
|
||||
params[0] = (paddle_parameter*)malloc(sizeof(paddle_parameter));
|
||||
params[0]->name = names[0];
|
||||
params[0]->content = (unsigned char*)param_content1;
|
||||
params[0]->content_len = 3 * sizeof(real);
|
||||
params[0]->element_type = PADDLE_ELEMENT_TYPE_FLOAT32;
|
||||
|
||||
params[1] = (paddle_parameter*)malloc(sizeof(paddle_parameter));
|
||||
params[1]->name = names[1];
|
||||
params[1]->content = (unsigned char*)param_content2;
|
||||
params[1]->content_len = 3 * sizeof(real);
|
||||
params[1]->element_type = PADDLE_ELEMENT_TYPE_INT32;
|
||||
|
||||
if (paddle_begin_init_params(c)) {
|
||||
if (paddle_init_param(c, *params[0], NULL, 0) != 0) {
|
||||
goto retry;
|
||||
}
|
||||
if (paddle_init_param(c, *params[1], NULL, 0) != 0) {
|
||||
goto retry;
|
||||
}
|
||||
if (paddle_finish_init_params(c) != 0) {
|
||||
goto retry;
|
||||
}
|
||||
} else {
|
||||
fail();
|
||||
}
|
||||
|
||||
printf("get inited parameters from pserver:\n");
|
||||
// get parameters again by reusing the allocated parameter buffers.
|
||||
if (paddle_get_params(c, params, 2) != 0) {
|
||||
fail();
|
||||
}
|
||||
print_parameter(params[0]);
|
||||
print_parameter(params[1]);
|
||||
|
||||
printf("send gradient to pserver:\n");
|
||||
real gradient_content1[] = {0.01, 0.02, 0.03};
|
||||
real gradinet_content2[] = {0.04, 0.05, 0.06};
|
||||
|
||||
paddle_gradient** grads =
|
||||
(paddle_gradient**)malloc(sizeof(paddle_gradient*) * 2);
|
||||
grads[0] = (paddle_gradient*)malloc(sizeof(paddle_gradient));
|
||||
grads[0]->name = names[0];
|
||||
grads[0]->content = (unsigned char*)gradient_content1;
|
||||
grads[0]->content_len = 3 * sizeof(real);
|
||||
grads[0]->element_type = PADDLE_ELEMENT_TYPE_FLOAT32;
|
||||
|
||||
grads[1] = (paddle_gradient*)malloc(sizeof(paddle_gradient));
|
||||
grads[1]->name = names[1];
|
||||
grads[1]->content = (unsigned char*)gradinet_content2;
|
||||
grads[1]->content_len = 3 * sizeof(real);
|
||||
grads[1]->element_type = PADDLE_ELEMENT_TYPE_INT32;
|
||||
|
||||
printf("print gradient sent to pserver:\n");
|
||||
print_parameter(grads[0]);
|
||||
print_parameter(grads[1]);
|
||||
|
||||
if (paddle_send_grads(c, grads, 2) != 0) {
|
||||
fail();
|
||||
}
|
||||
|
||||
printf("get updated parameters from pserver:\n");
|
||||
// get parameters again by reusing the allocated parameter buffers.
|
||||
if (paddle_get_params(c, params, 2) != 0) {
|
||||
fail();
|
||||
}
|
||||
print_parameter(params[0]);
|
||||
print_parameter(params[1]);
|
||||
|
||||
if (paddle_save_model(c, "/tmp/") != 0) {
|
||||
fail();
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
@ -0,0 +1,131 @@
|
||||
import paddle.v2 as paddle
|
||||
import gzip
|
||||
|
||||
|
||||
def softmax_regression(img):
|
||||
predict = paddle.layer.fc(input=img,
|
||||
size=10,
|
||||
act=paddle.activation.Softmax())
|
||||
return predict
|
||||
|
||||
|
||||
def multilayer_perceptron(img):
|
||||
# The first fully-connected layer
|
||||
hidden1 = paddle.layer.fc(input=img, size=128, act=paddle.activation.Relu())
|
||||
# The second fully-connected layer and the according activation function
|
||||
hidden2 = paddle.layer.fc(input=hidden1,
|
||||
size=64,
|
||||
act=paddle.activation.Relu())
|
||||
# The thrid fully-connected layer, note that the hidden size should be 10,
|
||||
# which is the number of unique digits
|
||||
predict = paddle.layer.fc(input=hidden2,
|
||||
size=10,
|
||||
act=paddle.activation.Softmax())
|
||||
return predict
|
||||
|
||||
|
||||
def convolutional_neural_network(img):
|
||||
# first conv layer
|
||||
conv_pool_1 = paddle.networks.simple_img_conv_pool(
|
||||
input=img,
|
||||
filter_size=5,
|
||||
num_filters=20,
|
||||
num_channel=1,
|
||||
pool_size=2,
|
||||
pool_stride=2,
|
||||
act=paddle.activation.Tanh())
|
||||
# second conv layer
|
||||
conv_pool_2 = paddle.networks.simple_img_conv_pool(
|
||||
input=conv_pool_1,
|
||||
filter_size=5,
|
||||
num_filters=50,
|
||||
num_channel=20,
|
||||
pool_size=2,
|
||||
pool_stride=2,
|
||||
act=paddle.activation.Tanh())
|
||||
# The first fully-connected layer
|
||||
fc1 = paddle.layer.fc(input=conv_pool_2,
|
||||
size=128,
|
||||
act=paddle.activation.Tanh())
|
||||
# The softmax layer, note that the hidden size should be 10,
|
||||
# which is the number of unique digits
|
||||
predict = paddle.layer.fc(input=fc1,
|
||||
size=10,
|
||||
act=paddle.activation.Softmax())
|
||||
return predict
|
||||
|
||||
|
||||
def main():
|
||||
paddle.init(use_gpu=False, trainer_count=1)
|
||||
|
||||
# define network topology
|
||||
images = paddle.layer.data(
|
||||
name='pixel', type=paddle.data_type.dense_vector(784))
|
||||
label = paddle.layer.data(
|
||||
name='label', type=paddle.data_type.integer_value(10))
|
||||
|
||||
# Here we can build the prediction network in different ways. Please
|
||||
# choose one by uncomment corresponding line.
|
||||
predict = softmax_regression(images)
|
||||
#predict = multilayer_perceptron(images)
|
||||
#predict = convolutional_neural_network(images)
|
||||
|
||||
cost = paddle.layer.classification_cost(input=predict, label=label)
|
||||
parameters = paddle.parameters.create(cost)
|
||||
|
||||
optimizer = paddle.optimizer.Momentum(
|
||||
learning_rate=0.1 / 128.0,
|
||||
momentum=0.9,
|
||||
regularization=paddle.optimizer.L2Regularization(rate=0.0005 * 128))
|
||||
|
||||
trainer = paddle.trainer.SGD(cost=cost,
|
||||
parameters=parameters,
|
||||
update_equation=optimizer,
|
||||
is_local=False,
|
||||
pserver_spec="localhost:3000")
|
||||
|
||||
lists = []
|
||||
|
||||
def event_handler(event):
|
||||
if isinstance(event, paddle.event.EndIteration):
|
||||
if event.batch_id % 1000 == 0:
|
||||
print "Pass %d, Batch %d, Cost %f, %s" % (
|
||||
event.pass_id, event.batch_id, event.cost, event.metrics)
|
||||
|
||||
elif isinstance(event, paddle.event.EndPass):
|
||||
result = trainer.test(reader=paddle.batch(
|
||||
paddle.dataset.mnist.test(), batch_size=128))
|
||||
print "Test with Pass %d, Cost %f, %s\n" % (
|
||||
event.pass_id, result.cost, result.metrics)
|
||||
lists.append((event.pass_id, result.cost,
|
||||
result.metrics['classification_error_evaluator']))
|
||||
|
||||
trainer.train(
|
||||
reader=paddle.batch(
|
||||
paddle.reader.shuffle(
|
||||
paddle.dataset.mnist.train(), buf_size=8192),
|
||||
batch_size=128),
|
||||
event_handler=event_handler,
|
||||
num_passes=100)
|
||||
|
||||
# find the best pass
|
||||
best = sorted(lists, key=lambda list: float(list[1]))[0]
|
||||
print 'Best pass is %s, testing Avgcost is %s' % (best[0], best[1])
|
||||
print 'The classification accuracy is %.2f%%' % (100 - float(best[2]) * 100)
|
||||
|
||||
test_creator = paddle.dataset.mnist.test()
|
||||
test_data = []
|
||||
for item in test_creator():
|
||||
test_data.append((item[0], ))
|
||||
if len(test_data) == 100:
|
||||
break
|
||||
|
||||
# output is a softmax layer. It returns probabilities.
|
||||
# Shape should be (100, 10)
|
||||
probs = paddle.infer(
|
||||
output_layer=predict, parameters=parameters, input=test_data)
|
||||
print probs.shape
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
main()
|
||||
@ -0,0 +1,60 @@
|
||||
import paddle.v2 as paddle
|
||||
import paddle.v2.dataset.uci_housing as uci_housing
|
||||
|
||||
|
||||
def main():
|
||||
# init
|
||||
paddle.init(use_gpu=False, trainer_count=1)
|
||||
|
||||
# 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'),
|
||||
size=1,
|
||||
act=paddle.activation.Linear(),
|
||||
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)
|
||||
|
||||
# create parameters
|
||||
parameters = paddle.parameters.create(cost)
|
||||
|
||||
# create optimizer
|
||||
optimizer = paddle.optimizer.Momentum(momentum=0)
|
||||
|
||||
trainer = paddle.trainer.SGD(cost=cost,
|
||||
parameters=parameters,
|
||||
update_equation=optimizer,
|
||||
is_local=False,
|
||||
pserver_spec="localhost:3000")
|
||||
|
||||
# event_handler to print training and testing info
|
||||
def event_handler(event):
|
||||
if isinstance(event, paddle.event.EndIteration):
|
||||
if event.batch_id % 100 == 0:
|
||||
print "Pass %d, Batch %d, Cost %f" % (
|
||||
event.pass_id, event.batch_id, event.cost)
|
||||
|
||||
if isinstance(event, paddle.event.EndPass):
|
||||
if (event.pass_id + 1) % 10 == 0:
|
||||
result = trainer.test(
|
||||
reader=paddle.batch(
|
||||
uci_housing.test(), batch_size=2),
|
||||
feeding={'x': 0,
|
||||
'y': 1})
|
||||
print "Test %d, %.2f" % (event.pass_id, result.cost)
|
||||
|
||||
# training
|
||||
trainer.train(
|
||||
reader=paddle.batch(
|
||||
paddle.reader.shuffle(
|
||||
uci_housing.train(), buf_size=500),
|
||||
batch_size=2),
|
||||
feeding={'x': 0,
|
||||
'y': 1},
|
||||
event_handler=event_handler,
|
||||
num_passes=30)
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
main()
|
||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue