一、搭建环境
分布式paddle的环境和tensorflow的环境很相似，在我之前的博客中已经写得非常清楚了
我的是 python3.6+cuda9+cudnn7+ubuntu16.04
博客链接：https://blog.csdn.net/qq_28626909/article/details/85007363
搭建好环境知道我们开始安装paddlepaddle，
官方网站链接：http://www.paddlepaddle.org/
我的配置是NVIDIA384+cuda9.0+cudnn7.0,所以我直接pip install paddlepaddle-gpu即可
cuda8+cudnn7.0是pip install paddlepaddle-gpu==1.2.0.post87
cuda8+cudnn5.0是pip install paddlepaddle-gpu==1.2.0.post85
二，测试单机paddle
为了测试我们的paddle是否安装成功，我们这里放出代码，可以直接跑的。
#!D:/workplace/python
# -*- coding: utf-8 -*-
# @File : first_test.py
# @Author: WangYe
# @Date : 2019/1/10
# @Software: PyCharm
import paddle
import paddle.fluid as fluid
import numpy
import six
# Configure the neural network.
def net(x, y):
y_predict = fluid.layers.fc(input=x, size=1, act=None)
cost = fluid.layers.square_error_cost(input=y_predict, label=y)
avg_cost = fluid.layers.mean(cost)
return y_predict, avg_cost
# Define train function.
def train(save_dirname):
x = fluid.layers.data(name='x', shape=[13], dtype='float32')
y = fluid.layers.data(name='y', shape=[1], dtype='float32')
y_predict, avg_cost = net(x, y)
sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
sgd_optimizer.minimize(avg_cost)
train_reader = paddle.batch(
paddle.reader.shuffle(paddle.dataset.uci_housing.train(), buf_size=500),
batch_size=20)
place = fluid.CPUPlace()
exe = fluid.Executor(place)
def train_loop(main_program):
feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
exe.run(fluid.default_startup_program())
PASS_NUM = 1000
for pass_id in range(PASS_NUM):
total_loss_pass = 0
for data in train_reader():
avg_loss_value, = exe.run(
main_program, feed=feeder.feed(data), fetch_list=[avg_cost])
total_loss_pass += avg_loss_value
if avg_loss_value < 5.0:
if save_dirname is not None:
fluid.io.save_inference_model(
save_dirname, ['x'], [y_predict], exe)
return
print("Pass %d, total avg cost = %f" % (pass_id, total_loss_pass))
train_loop(fluid.default_main_program())
# Infer by using provided test data.
def infer(save_dirname=None):
place = fluid.CPUPlace()
exe = fluid.Executor(place)
inference_scope = fluid.core.Scope()
with fluid.scope_guard(inference_scope):
[inference_program, feed_target_names, fetch_targets] = (
fluid.io.load_inference_model(save_dirname, exe))
test_reader = paddle.batch(paddle.dataset.uci_housing.test(), batch_size=20)
test_data = six.next(test_reader())
test_feat = numpy.array(list(map(lambda x: x[0], test_data))).astype("float32")
test_label = numpy.array(list(map(lambda x: x[1], test_data))).astype("float32")
results = exe.run(inference_program,
feed={feed_target_names[0]: numpy.array(test_feat)},
fetch_list=fetch_targets)
print("infer results: ", results[0])
print("ground truth: ", test_label)
# Run train and infer.
if __name__ == "__main__":
save_dirname = "fit_a_line.inference.model"
train(save_dirname)
infer(save_dirname)代码的很多部分应该用过tensorflow的人都能看懂，paddlepaddle还是比tensorflow更简洁的
运行结果截图：
三、分布式paddle的测试
分布式paddle和tensorflow是有很多相似之处，我这里对官方网站的代码进行了一点改动，这样测试起来就更为方便
tensorflow把参数服务器叫ps，计算节点叫worker
padddle把参数服务器叫PSERVER,计算节点叫TRAINER
这是第一段，也就是PSERVER的代码，我把代码命名为：
dist_train_ps.py
#!D:/workplace/python
# -*- coding: utf-8 -*-
# @File : dist_train_ps.py
# @Author: WangYe
# @Date : 2019/1/11
# @Software: PyCharm
import os
import paddle
import paddle.fluid as fluid
# train reader
BATCH_SIZE = 20
EPOCH_NUM = 30
BATCH_SIZE = 8
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.uci_housing.train(), buf_size=500),
batch_size=BATCH_SIZE)
def train():
y = fluid.layers.data(name='y', shape=[1], dtype='float32')
x = fluid.layers.data(name='x', shape=[13], dtype='float32')
y_predict = fluid.layers.fc(input=x, size=1, act=None)
loss = fluid.layers.square_error_cost(input=y_predict, label=y)
avg_loss = fluid.layers.mean(loss)
opt = fluid.optimizer.SGD(learning_rate=0.001)
opt.minimize(avg_loss)
place = fluid.CPUPlace()
feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
exe = fluid.Executor(place)
# fetch distributed training environment setting
training_role = os.getenv("PADDLE_TRAINING_ROLE", "PSERVER")
port = os.getenv("PADDLE_PSERVER_PORT", "6174")
pserver_ips = os.getenv("PADDLE_PSERVER_IPS", "localhost")
trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
eplist = []
for ip in pserver_ips.split(","):
eplist.append(':'.join([ip, port]))
pserver_endpoints = ",".join(eplist)
trainers = int(os.getenv("PADDLE_TRAINERS","1"))
current_endpoint = os.getenv("PADDLE_CURRENT_IP", "localhost") + ":" + port
t = fluid.DistributeTranspiler()
t.transpile(
trainer_id = trainer_id,
pservers = pserver_endpoints,
trainers = trainers)
if training_role == "PSERVER":
pserver_prog = t.get_pserver_program(current_endpoint)
startup_prog = t.get_startup_program(current_endpoint, pserver_prog)
exe.run(startup_prog)
exe.run(pserver_prog)
elif training_role == "TRAINER":
trainer_prog = t.get_trainer_program()
exe.run(fluid.default_startup_program())
for epoch in range(EPOCH_NUM):
for batch_id, batch_data in enumerate(train_reader()):
avg_loss_value, = exe.run(trainer_prog,
feed=feeder.feed(batch_data),
fetch_list=[avg_loss])
if (batch_id + 1) % 10 == 0:
print("Epoch: {0}, Batch: {1}, loss: {2}".format(
epoch, batch_id, avg_loss_value[0]))
# destory the resource of current trainer node in pserver server node
exe.close()
else:
raise AssertionError("PADDLE_TRAINING_ROLE should be one of [TRAINER, PSERVER]")
train()下面的是第二段，也就是的TRAINER，我把代码命名为
dist_train_worker.py
#!D:/workplace/python
# -*- coding: utf-8 -*-
# @File : dist_train_ps.py
# @Author: WangYe
# @Date : 2019/1/11
# @Software: PyCharm
import os
import paddle
import paddle.fluid as fluid
# train reader
BATCH_SIZE = 20
EPOCH_NUM = 30
BATCH_SIZE = 8
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.uci_housing.train(), buf_size=500),
batch_size=BATCH_SIZE)
def train():
y = fluid.layers.data(name='y', shape=[1], dtype='float32')
x = fluid.layers.data(name='x', shape=[13], dtype='float32')
y_predict = fluid.layers.fc(input=x, size=1, act=None)
loss = fluid.layers.square_error_cost(input=y_predict, label=y)
avg_loss = fluid.layers.mean(loss)
opt = fluid.optimizer.SGD(learning_rate=0.001)
opt.minimize(avg_loss)
place = fluid.CPUPlace()
feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
exe = fluid.Executor(place)
# fetch distributed training environment setting
training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
port = os.getenv("PADDLE_PSERVER_PORT", "6174")
pserver_ips = os.getenv("PADDLE_PSERVER_IPS", "localhost")
trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
eplist = []
for ip in pserver_ips.split(","):
eplist.append(':'.join([ip, port]))
pserver_endpoints = ",".join(eplist)
trainers = int(os.getenv("PADDLE_TRAINERS","1"))
current_endpoint = os.getenv("PADDLE_CURRENT_IP", "") + ":" + port
t = fluid.DistributeTranspiler()
t.transpile(
trainer_id = trainer_id,
pservers = pserver_endpoints,
trainers = trainers)
if training_role == "PSERVER":
pserver_prog = t.get_pserver_program(current_endpoint)
startup_prog = t.get_startup_program(current_endpoint, pserver_prog)
exe.run(startup_prog)
exe.run(pserver_prog)
elif training_role == "TRAINER":
trainer_prog = t.get_trainer_program()
exe.run(fluid.default_startup_program())
for epoch in range(EPOCH_NUM):
for batch_id, batch_data in enumerate(train_reader()):
avg_loss_value, = exe.run(trainer_prog,
feed=feeder.feed(batch_data),
fetch_list=[avg_loss])
if (batch_id + 1) % 10 == 0:
print("Epoch: {0}, Batch: {1}, loss: {2}".format(
epoch, batch_id, avg_loss_value[0]))
# destory the resource of current trainer node in pserver server node
exe.close()
else:
raise AssertionError("PADDLE_TRAINING_ROLE should be one of [TRAINER, PSERVER]")
train()对于这两端代码，有一下几点说明
1.分布式测试代码用到的数据是上面第二部分单机测试的数据，它的数据是自动下载的。也就是说运行过上面的单机代码后，就自动有了分布式的数据
2.分布式代码需要分开执行，同tensorflow一样，不过我做的改动在代码里了，所以执行代码之后命令行不用添加参数，如下截图
3.可能遇到的问题：
除开我刚开始放出的链接中遇到的环境问题，我遇到了一个 段错误的一个问题，那是因为ps和worker的ip写的有问题
四、分布式paddle的关键参数
training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER")
port = os.getenv("PADDLE_PSERVER_PORT", "6174")
pserver_ips = os.getenv("PADDLE_PSERVER_IPS", "localhost")
trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
eplist = []
for ip in pserver_ips.split(","):
eplist.append(':'.join([ip, port]))
pserver_endpoints = ",".join(eplist)
trainers = int(os.getenv("PADDLE_TRAINERS","1"))
current_endpoint = os.getenv("PADDLE_CURRENT_IP", "localhost") + ":" + port我们一个一个来看，以上面的worker的代码为例（paddle中把worker成为tariner）,从上往下讲
paddle_tarining_role（str）：训练对象（trainer还是pserver）
paddle_pserver_port(int):端口（这里和tensorflow不同的是必须指定同一个端口，即所有tariner和ps都是6174）
paddle_pserver_ips(str):ps的ip地址，如果多个就用逗号隔开
paddle_trainer_id(int):tariner各自的编号，在ps中全部填写0
paddle_trainers(int):trainer的数量，即tensorflow中多少个worker
paddle_current_ip(str):当前ps节点的ip地址，如果是worker，则可以不指定
官方参数解读网站截图：
运行截图：
ok，大功告成！！
扫一扫
449
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
