update paddlecloud train (#142)

* update

* fix

* delete ps-memory

* fix

* fix

core/engine/cluster/cloud/before_hook_cpu.sh.template

 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

 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

@@ -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

+#!/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

@@ -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
+        --is-standalone ${is_stand_alone} \
+        --json \
+        ${nccl} 
\ No newline at end of file
+        
\ No newline at end of file

core/engine/cluster/cloud/mpi_job.sh.template

@@ -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

@@ -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

@@ -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：
@@ -165,26 +175,66 @@ 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
+  
 ```
 
 更多backend.yaml配置选项信息，可以查看[yaml配置说明](./yaml.md)
 
 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

@@ -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

+[easy_install]
+index_url=http://pip.baidu.com/pypi/simple
\ No newline at end of file