Skip to content

P
PaddleRec

BaiXuePrincess / PaddleRec
与 Fork 源项目一致

Fork自 PaddlePaddle / PaddleRec

通知 1
Star 0
Fork 0
P
PaddleRec
未验证 提交 7f9869d3 编写于 作者: C Chengmo 提交者： GitHub
浏览文件
操作

update paddlecloud train (#142) 

* update

* fix

* delete ps-memory

* fix

* fix
上级 9b89d8f7
隐藏空白更改
内联 并排
Showing with 316 addition and 48 deletion
core/engine/cluster/cloud/before_hook_cpu.sh.template
浏览文件 @ 7f9869d3
echo "Run before_hook.sh ..."
wget https://paddlerec.bj.bcebos.com/whl/PaddleRec.tar.gz
wget https://paddlerec.bj.bcebos.com/whl/PaddleRec.tar.gz --no-check-certificate
tar -xf PaddleRec.tar.gz
......@@ -10,6 +10,6 @@ python setup.py install
pip uninstall -y paddlepaddle
pip install paddlepaddle-gpu==<$ PADDLEPADDLE_VERSION $> --index-url=http://pip.baidu.com/pypi/simple --trusted-host pip.baidu.com
pip install paddlepaddle==<$ PADDLEPADDLE_VERSION $> --index-url=http://pip.baidu.com/pypi/simple --trusted-host pip.baidu.com
echo "End before_hook.sh ..."
core/engine/cluster/cloud/before_hook_gpu.sh.template
浏览文件 @ 7f9869d3
echo "Run before_hook.sh ..."
wget https://paddlerec.bj.bcebos.com/whl/PaddleRec.tar.gz
wget https://paddlerec.bj.bcebos.com/whl/PaddleRec.tar.gz --no-check-certificate
tar -xf PaddleRec.tar.gz
......
core/engine/cluster/cloud/cluster.sh
浏览文件 @ 7f9869d3
......@@ -39,7 +39,12 @@ function _before_submit() {
elif [ ${DISTRIBUTE_MODE} == "COLLECTIVE_GPU_K8S" ]; then
_gen_gpu_before_hook
_gen_k8s_config
_gen_k8s_job
_gen_k8s_gpu_job
_gen_end_hook
elif [ ${DISTRIBUTE_MODE} == "PS_CPU_K8S" ]; then
_gen_cpu_before_hook
_gen_k8s_config
_gen_k8s_cpu_job
_gen_end_hook
fi
......@@ -101,6 +106,7 @@ function _gen_end_hook() {
function _gen_mpi_job() {
echo "gen mpi_job.sh"
sed -e "s#<$ GROUP_NAME $>#$GROUP_NAME#g" \
-e "s#<$ JOB_NAME $>#$OLD_JOB_NAME#g" \
-e "s#<$ AK $>#$AK#g" \
-e "s#<$ SK $>#$SK#g" \
-e "s#<$ MPI_PRIORITY $>#$PRIORITY#g" \
......@@ -109,18 +115,34 @@ function _gen_mpi_job() {
${abs_dir}/cloud/mpi_job.sh.template >${PWD}/job.sh
}
function _gen_k8s_job() {
function _gen_k8s_gpu_job() {
echo "gen k8s_job.sh"
sed -e "s#<$ GROUP_NAME $>#$GROUP_NAME#g" \
-e "s#<$ JOB_NAME $>#$OLD_JOB_NAME#g" \
-e "s#<$ AK $>#$AK#g" \
-e "s#<$ SK $>#$SK#g" \
-e "s#<$ K8S_PRIORITY $>#$PRIORITY#g" \
-e "s#<$ K8S_TRAINERS $>#$K8S_TRAINERS#g" \
-e "s#<$ K8S_CPU_CORES $>#$K8S_CPU_CORES#g" \
-e "s#<$ K8S_GPU_CARD $>#$K8S_GPU_CARD#g" \
-e "s#<$ START_CMD $>#$START_CMD#g" \
${abs_dir}/cloud/k8s_job.sh.template >${PWD}/job.sh
}
function _gen_k8s_cpu_job() {
echo "gen k8s_job.sh"
sed -e "s#<$ GROUP_NAME $>#$GROUP_NAME#g" \
-e "s#<$ JOB_NAME $>#$OLD_JOB_NAME#g" \
-e "s#<$ AK $>#$AK#g" \
-e "s#<$ SK $>#$SK#g" \
-e "s#<$ K8S_PRIORITY $>#$PRIORITY#g" \
-e "s#<$ K8S_TRAINERS $>#$K8S_TRAINERS#g" \
-e "s#<$ K8S_PS_NUM $>#$K8S_PS_NUM#g" \
-e "s#<$ K8S_PS_CORES $>#$K8S_PS_CORES#g" \
-e "s#<$ K8S_CPU_CORES $>#$K8S_CPU_CORES#g" \
-e "s#<$ START_CMD $>#$START_CMD#g" \
${abs_dir}/cloud/k8s_cpu_job.sh.template >${PWD}/job.sh
}
#-----------------------------------------------------------------------------------------------------------------
......@@ -145,6 +167,7 @@ function _submit() {
function package_hook() {
cur_time=`date +"%Y%m%d%H%M"`
new_job_name="${JOB_NAME}_${cur_time}"
export OLD_JOB_NAME=${JOB_NAME}
export JOB_NAME=${new_job_name}
export job_file_path="${PWD}/${new_job_name}"
mkdir ${job_file_path}
......
core/engine/cluster/cloud/k8s_cpu_job.sh.template 0 → 100644
浏览文件 @ 7f9869d3
#!/bin/bash
###############################################################
## 注意-- 注意--注意 ##
## K8S PS-CPU多机作业作业示例 ##
###############################################################
job_name=<$ JOB_NAME $>
# 作业参数
group_name="<$ GROUP_NAME $>"
job_version="paddle-fluid-v1.7.1"
start_cmd="<$ START_CMD $>"
wall_time="10:00:00"
k8s_priority=<$ K8S_PRIORITY $>
k8s_trainers=<$ K8S_TRAINERS $>
k8s_cpu_cores=<$ K8S_CPU_CORES $>
k8s_ps_num=<$ K8S_PS_NUM $>
k8s_ps_cores=<$ K8S_PS_CORES $>
# 你的ak/sk（可在paddlecloud web页面【个人中心】处获取）
ak=<$ AK $>
sk=<$ SK $>
paddlecloud job --ak ${ak} --sk ${sk} \
train --job-name ${job_name} \
--group-name ${group_name} \
--job-conf config.ini \
--start-cmd "${start_cmd}" \
--files ./* \
--job-version ${job_version} \
--k8s-priority ${k8s_priority} \
--wall-time ${wall_time} \
--k8s-trainers ${k8s_trainers} \
--k8s-cpu-cores ${k8s_cpu_cores} \
--k8s-ps-num ${k8s_ps_num} \
--k8s-ps-cores ${k8s_ps_cores} \
--is-standalone 0 \
--distribute-job-type "PSERVER" \
--json
\ No newline at end of file
core/engine/cluster/cloud/k8s_job.sh.template
浏览文件 @ 7f9869d3
......@@ -3,7 +3,7 @@
## 注意-- 注意--注意 ##
## K8S NCCL2多机作业作业示例 ##
###############################################################
job_name=${JOB_NAME}
job_name=<$ JOB_NAME $>
# 作业参数
group_name="<$ GROUP_NAME $>"
......@@ -13,8 +13,20 @@ wall_time="10:00:00"
k8s_priority=<$ K8S_PRIORITY $>
k8s_trainers=<$ K8S_TRAINERS $>
k8s_cpu_cores=<$ K8S_CPU_CORES $>
k8s_gpu_cards=<$ K8S_GPU_CARD $>
is_stand_alone=0
nccl="--distribute-job-type "NCCL2""
if [ ${k8s_trainers} == 1 ];then
is_stand_alone=1
nccl="--job-remark single-trainer"
if [ ${k8s_gpu_cards} == 1];then
nccl="--job-remark single-gpu"
echo "Attention: Use single GPU card for PaddleRec distributed training, please set runner class from 'cluster_train' to 'train' in config.yaml."
fi
fi
# 你的ak/sk（可在paddlecloud web页面【个人中心】处获取）
ak=<$ AK $>
sk=<$ SK $>
......@@ -27,9 +39,11 @@ paddlecloud job --ak ${ak} --sk ${sk} \
--files ./* \
--job-version ${job_version} \
--k8s-trainers ${k8s_trainers} \
--k8s-cpu-cores ${k8s_cpu_cores} \
--k8s-gpu-cards ${k8s_gpu_cards} \
--k8s-priority ${k8s_priority} \
--wall-time ${wall_time} \
--is-standalone 0 \
--distribute-job-type "NCCL2" \
--json
\ No newline at end of file
--is-standalone ${is_stand_alone} \
--json \
${nccl}
\ No newline at end of file
core/engine/cluster/cloud/mpi_job.sh.template
浏览文件 @ 7f9869d3
......@@ -3,7 +3,7 @@
## 注意--注意--注意 ##
## MPI 类型作业演示 ##
###############################################################
job_name=${JOB_NAME}
job_name=<$ JOB_NAME $>
# 作业参数
group_name=<$ GROUP_NAME $>
......
core/engine/cluster/cluster.py
浏览文件 @ 7f9869d3
......@@ -67,10 +67,10 @@ class ClusterEngine(Engine):
@staticmethod
def workspace_replace():
workspace = envs.get_runtime_environ("workspace")
remote_workspace = envs.get_runtime_environ("remote_workspace")
for k, v in os.environ.items():
v = v.replace("{workspace}", workspace)
v = v.replace("{workspace}", remote_workspace)
os.environ[k] = str(v)
def run(self):
......@@ -98,14 +98,12 @@ class ClusterEngine(Engine):
cluster_env_check_tool = PaddleCloudMpiEnv()
else:
raise ValueError(
"Paddlecloud with Mpi don't support GPU training, check your config"
"Paddlecloud with Mpi don't support GPU training, check your config.yaml & backend.yaml"
)
elif cluster_type.upper() == "K8S":
if fleet_mode == "PS":
if device == "CPU":
raise ValueError(
"PS-CPU on paddlecloud is not supported at this time, comming soon"
)
cluster_env_check_tool = CloudPsCpuEnv()
elif device == "GPU":
raise ValueError(
"PS-GPU on paddlecloud is not supported at this time, comming soon"
......@@ -115,7 +113,7 @@ class ClusterEngine(Engine):
cluster_env_check_tool = CloudCollectiveEnv()
elif device == "CPU":
raise ValueError(
"Unexpected config -> device: CPU with fleet_mode: Collective, check your config"
"Unexpected config -> device: CPU with fleet_mode: Collective, check your config.yaml"
)
else:
raise ValueError("cluster_type {} error, must in MPI/K8S".format(
......@@ -234,7 +232,7 @@ class PaddleCloudMpiEnv(ClusterEnvBase):
"config.train_data_path", "")
if self.cluster_env["TRAIN_DATA_PATH"] == "":
raise ValueError(
"No -- TRAIN_DATA_PATH -- found in your backend.yaml, please check."
"No -- TRAIN_DATA_PATH -- found in your backend.yaml, please add train_data_path in your backend yaml."
)
# test_data_path
self.cluster_env["TEST_DATA_PATH"] = self.backend_env.get(
......@@ -274,7 +272,7 @@ class PaddleCloudK8sEnv(ClusterEnvBase):
category=UserWarning,
stacklevel=2)
warnings.warn(
"The remote mount point will be mounted to the ./afs/",
"The remote afs path will be mounted to the ./afs/",
category=UserWarning,
stacklevel=2)
......@@ -293,3 +291,21 @@ class CloudCollectiveEnv(PaddleCloudK8sEnv):
"submit.k8s_gpu_card", 1)
self.cluster_env["K8S_CPU_CORES"] = self.backend_env.get(
"submit.k8s_cpu_cores", 1)
class CloudPsCpuEnv(PaddleCloudK8sEnv):
def __init__(self):
super(CloudPsCpuEnv, self).__init__()
def env_check(self):
super(CloudPsCpuEnv, self).env_check()
self.cluster_env["DISTRIBUTE_MODE"] = "PS_CPU_K8S"
self.cluster_env["K8S_TRAINERS"] = self.backend_env.get(
"submit.k8s_trainers", 1)
self.cluster_env["K8S_CPU_CORES"] = self.backend_env.get(
"submit.k8s_cpu_cores", 2)
self.cluster_env["K8S_PS_NUM"] = self.backend_env.get(
"submit.k8s_ps_num", 1)
self.cluster_env["K8S_PS_CORES"] = self.backend_env.get(
"submit.k8s_ps_cores", 2)
doc/distributed_train.md
浏览文件 @ 7f9869d3
......@@ -9,6 +9,7 @@
- [第三步：增加集群运行`backend.yaml`配置](#第三步增加集群运行backendyaml配置)
- [MPI集群的Parameter Server模式配置](#mpi集群的parameter-server模式配置)
- [K8S集群的Collective模式配置](#k8s集群的collective模式配置)
- [K8S集群的PS-CPU模式配置](#k8s集群的ps-cpu模式配置)
- [第四步：任务提交](#第四步任务提交)
- [使用PaddleCloud Client提交](#使用paddlecloud-client提交)
- [第一步：在`before_hook.sh`里手动安装PaddleRec](#第一步在before_hooksh里手动安装paddlerec)
......@@ -34,10 +35,10 @@
分布式运行首先需要更改`config.yaml`，主要调整以下内容：
- workspace: 调整为在节点运行时的工作目录
- runner_class: 从单机的"train"调整为"cluster_train"
- fleet_mode: 选则参数服务器模式，抑或GPU Collective模式
- distribute_strategy: 可选项，选择分布式训练的策略
- workspace: 调整为在远程点运行时的工作目录，一般设置为`"./"`即可
- runner_class: 从单机的"train"调整为"cluster_train"，单机训练->分布式训练（例外情况，k8s上单机单卡训练仍然为train）
- fleet_mode: 选则参数服务器模式(ps)，抑或GPU的all-reduce模式(collective)
- distribute_strategy: 可选项，选择分布式训练的策略，目前只在参数服务器模式下生效，可选项:`sync、asycn、half_async、geo`
配置选项具体参数，可以参考[yaml配置说明](./yaml.md)
......@@ -50,47 +51,56 @@
workspace: "paddlerec.models.rank.dnn"
mode: [single_cpu_train]
# config of each runner.
# runner is a kind of paddle training class, which wraps the train/infer process.
runner:
- name: single_cpu_train
class: train
# num of epochs
epochs: 4
# device to run training or infer
device: cpu
save_checkpoint_interval: 2 # save model interval of epochs
save_checkpoint_path: "increment_dnn" # save checkpoint path
init_model_path: "" # load model path
save_checkpoint_interval: 2
save_checkpoint_path: "increment_dnn"
init_model_path: ""
print_interval: 10
phases: [phase1]
dataset:
- name: dataloader_train
batch_size: 2
type: DataLoader
data_path: "{workspace}/data/sample_data/train"
sparse_slots: "click 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26"
dense_slots: "dense_var:13"
```
分布式的训练配置可以改为：
```yaml
# workspace
# 改变一：代码上传至节点后，与运行shell同在一个默认目录下
# 改变一：代码上传至节点后，在默认目录下
workspace: "./"
mode: [ps_cluster]
# config of each runner.
# runner is a kind of paddle training class, which wraps the train/infer process.
runner:
- name: ps_cluster
# 改变二：调整runner的class
class: cluster_train
# num of epochs
epochs: 4
# device to run training or infer
device: cpu
# 改变三 & 四： 指定fleet_mode 与 distribute_strategy
fleet_mode: ps
distribute_strategy: async
save_checkpoint_interval: 2 # save model interval of epochs
save_checkpoint_path: "increment_dnn" # save checkpoint path
init_model_path: "" # load model path
save_checkpoint_interval: 2
save_checkpoint_path: "increment_dnn"
init_model_path: ""
print_interval: 10
phases: [phase1]
dataset:
- name: dataloader_train
batch_size: 2
type: DataLoader
# 改变五： 改变数据的读取目录
# 通常而言，mpi模式下，数据会下载到远程节点执行目录的'./train_data'下， k8s则与挂载位置有关
data_path: "{workspace}/train_data"
sparse_slots: "click 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26"
dense_slots: "dense_var:13"
```
除此之外，还需关注数据及模型加载的路径，一般而言：
......@@ -165,7 +175,14 @@ submit:
# for k8s gpu
# k8s gpu 模式下，训练节点数，及每个节点上的GPU卡数
k8s_trainers: 2
k8s-cpu-cores: 4
k8s_gpu_card: 1
# for k8s ps-cpu
k8s_trainers: 2
k8s-cpu-cores: 4
k8s_ps_num: 2
k8s_ps_cores: 4
```
......@@ -173,18 +190,51 @@ submit:
除此之外，我们还需要关注上传到工作目录的文件(`files选项`)的路径问题，在示例中是`./*.py`，说明我们执行任务提交时，与这些py文件在同一目录。若不在同一目录，则需要适当调整files路径，或改为这些文件的绝对路径。
不建议利用`files`上传数据文件，可以通过指定`train_data_path`自动下载，或指定`afs_remote_mount_point`挂载实现数据到节点的转移。
不建议利用`files`上传过大的数据文件，可以通过指定`train_data_path`自动下载，或在k8s模式下指定`afs_remote_mount_point`挂载实现数据到节点的转移。
#### MPI集群的Parameter Server模式配置
下面是一个利用PaddleCloud提交MPI参数服务器模式任务的`backend.yaml`示例
首先调整`config.yaml`:
```yaml
workspace: "./"
mode: [ps_cluster]
dataset:
- name: dataloader_train
batch_size: 2
type: DataLoader
data_path: "{workspace}/train_data"
sparse_slots: "click 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26"
dense_slots: "dense_var:13"
runner:
- name: ps_cluster
class: cluster_train
epochs: 2
device: cpu
fleet_mode: ps
save_checkpoint_interval: 1
save_checkpoint_path: "increment_dnn"
init_model_path: ""
print_interval: 1
phases: [phase1]
phase:
- name: phase1
model: "{workspace}/model.py"
dataset_name: dataloader_train
thread_num: 1
```
再新增`backend.yaml`
```yaml
backend: "PaddleCloud"
cluster_type: mpi # k8s 可选
cluster_type: mpi
config:
# 填写任务运行的paddle官方版本号 >= 1.7.2， 默认1.7.2
paddle_version: "1.7.2"
# hdfs/afs的配置信息填写
......@@ -229,9 +279,45 @@ submit:
下面是一个利用PaddleCloud提交K8S集群进行GPU训练的`backend.yaml`示例
首先调整`config.yaml`
```yaml
workspace: "./"
mode: [collective_cluster]
dataset:
- name: dataloader_train
batch_size: 2
type: DataLoader
data_path: "{workspace}/train_data"
sparse_slots: "click 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26"
dense_slots: "dense_var:13"
runner:
- name: collective_cluster
class: cluster_train
epochs: 2
device: gpu
fleet_mode: collective
save_checkpoint_interval: 1 # save model interval of epochs
save_checkpoint_path: "increment_dnn" # save checkpoint path
init_model_path: "" # load model path
print_interval: 1
phases: [phase1]
phase:
- name: phase1
model: "{workspace}/model.py"
dataset_name: dataloader_train
thread_num: 1
```
再增加`backend.yaml`
```yaml
backend: "PaddleCloud"
cluster_type: mpi # k8s 可选
cluster_type: k8s # k8s 可选
config:
# 填写任务运行的paddle官方版本号 >= 1.7.2， 默认1.7.2
......@@ -271,9 +357,93 @@ submit:
# for k8s gpu
# k8s gpu 模式下，训练节点数，及每个节点上的GPU卡数
k8s_trainers: 2
k8s-cpu-cores: 4
k8s_gpu_card: 1
```
#### K8S集群的PS-CPU模式配置
下面是一个利用PaddleCloud提交K8S集群进行参数服务器CPU训练的`backend.yaml`示例
首先调整`config.yaml`:
```yaml
workspace: "./"
mode: [ps_cluster]
dataset:
- name: dataloader_train
batch_size: 2
type: DataLoader
data_path: "{workspace}/train_data"
sparse_slots: "click 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26"
dense_slots: "dense_var:13"
runner:
- name: ps_cluster
class: cluster_train
epochs: 2
device: cpu
fleet_mode: ps
save_checkpoint_interval: 1
save_checkpoint_path: "increment_dnn"
init_model_path: ""
print_interval: 1
phases: [phase1]
phase:
- name: phase1
model: "{workspace}/model.py"
dataset_name: dataloader_train
thread_num: 1
```
再新增`backend.yaml`
```yaml
backend: "PaddleCloud"
cluster_type: k8s # k8s 可选
config:
# 填写任务运行的paddle官方版本号 >= 1.7.2， 默认1.7.2
paddle_version: "1.7.2"
# hdfs/afs的配置信息填写
fs_name: "afs://xxx.com"
fs_ugi: "usr,pwd"
# 填任务输出目录的远程地址，如afs:/user/your/path/ 则此处填 /user/your/path
output_path: ""
# for k8s
# 填远程挂载地址，如afs:/user/your/path/ 则此处填 /user/your/path
afs_remote_mount_point: ""
submit:
# PaddleCloud 个人信息 AK 及 SK
ak: ""
sk: ""
# 任务运行优先级，默认high
priority: "high"
# 任务名称
job_name: "PaddleRec_CTR"
# 训练资源所在组
group: ""
# 节点上的任务启动命令
start_cmd: "python -m paddlerec.run -m ./config.yaml"
# 本地需要上传到节点工作目录的文件
files: ./*.py ./*.yaml
# for k8s gpu
# k8s ps-cpu 模式下，训练节点数，参数服务器节点数，及每个节点上的cpu核心数及内存限制
k8s_trainers: 2
k8s-cpu-cores: 4
k8s_ps_num: 2
k8s_ps_cores: 4
```
### 第四步：任务提交
当我们准备好`config.yaml`与`backend.yaml`，便可以进行一键任务提交，命令为：
......
models/rank/dnn/backend.yaml
浏览文件 @ 7f9869d3
......@@ -11,12 +11,8 @@
# 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.
workspace: "./"
backend: "PaddleCloud"
cluster_type: k8s # k8s 可选
cluster_type: k8s # mpi 可选
config:
fs_name: "afs://xxx.com"
......@@ -56,5 +52,12 @@ submit:
# for k8s gpu
k8s_trainers: 2
k8s_cpu_cores: 2
k8s_gpu_card: 1
# for k8s ps-cpu
k8s_trainers: 2
k8s_cpu_cores: 4
k8s_ps_num: 2
k8s_ps_cores: 4
setup.cfg 0 → 100644
浏览文件 @ 7f9869d3
[easy_install]
index_url=http://pip.baidu.com/pypi/simple
\ No newline at end of file
Markdown is supported
0% 或 .
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 或