Support submit v2 API job

doc/usage_cn.md

@@ -40,7 +40,7 @@ Subcommands:
 	submit           Submit job to PaddlePaddle Cloud.
 
 
-Use "paddlecloud.darwin flags" for a list of top-level flags
+Use "paddlecloud flags" for a list of top-level flags
 ```
 
 ## 准备训练数据
@@ -120,10 +120,18 @@ scp -r my_training_package/ user@tunnel-server:/mnt/hdfs_mulan/idl/idl-dl/mypack
 
 执行下面的命令提交准备好的任务:
 
+- 提交基于V1 API的训练任务
+
 ```bash
 paddlecloud submit -jobname my-paddlecloud-job -cpu 1 -gpu 0 -memory 1Gi -parallelism 10 -pscpu 1 -pservers 3 -psmemory 1Gi -passes 1 -topology trainer_config.py /pfs/[datacenter_name]/home/[username]/ctr_demo_package
 ```
 
+- 提交基于V2 API的训练任务
+
+```bash
+paddlecloud submit -jobname my-paddlecloud-job -cpu 1 -gpu 0 -memory 1Gi -parallelism 10 -pscpu 1 -pservers 3 -psmemory 1Gi -passes 1 -entry "python trainer_config.py" /pfs/[datacenter_name]/home/[username]/ctr_demo_package
+```
+
 参数说明：
 - `jobname`：提交任务的名称，paddlecloud使用`jobname`唯一标识一个任务
 - `-cpu`：每个trainer进程使用的CPU资源，单位是“核”
@@ -134,6 +142,7 @@ paddlecloud submit -jobname my-paddlecloud-job -cpu 1 -gpu 0 -memory 1Gi -parall
 - `-pservers`：parameter server的节点个数
 - `-psmemory`：parameter server占用的内存资源，格式为“数字+单位”，单位可以是：Ki，Mi，Gi
 - `-topology`：指定PaddlePaddle v1训练的模型配置python文件
+- `-entry`: 指定PaddlePaddle v2训练程序的启动命令
 - `-passes`：执行训练的pass个数
 - `package`：HDFS 训练任务package的路径
 

docker/paddle_k8s

@@ -47,7 +47,7 @@ start_trainer() {
           --num_gradient_servers=$PADDLE_INIT_NUM_GRADIENT_SERVERS
         ;;
       "v2")
-        python ${TOPOLOGY}
+        ${ENTRY}
         ;;
       *)
         ;;

paddlecloud/paddlejob/paddle_job.py

@@ -18,6 +18,7 @@ def __init__(self,
                  pscpu,
                  psmemory,
                  topology,
+                 entry,
                  image,
                  passes,
                  gpu=0,
@@ -38,6 +39,7 @@ def __init__(self,
         self._pscpu = pscpu
         self._psmemory = psmemory
         self._topology = topology
+        self._entry = entry
         self._image = image
         self._volumes = volumes
         self._registry_secret = registry_secret
@@ -67,6 +69,7 @@ def get_env(self):
         envs.append({"name":"TRAINERS",             "value":str(self._parallelism)})
         envs.append({"name":"PSERVERS",             "value":str(self._pservers)})
         envs.append({"name":"TOPOLOGY",             "value":self._topology})
+        envs.append({"name":"ENTRY",                "value":self._entry})
         envs.append({"name":"TRAINER_PACKAGE",      "value":self._job_package})
         envs.append({"name":"PADDLE_INIT_PORT",     "value":str(DEFAULT_PADDLE_PORT)})
         envs.append({"name":"PADDLE_INIT_TRAINER_COUNT",        "value":str(self._cpu)})
@@ -97,6 +100,8 @@ def _get_pserver_entrypoint(self):
         return ["paddle_k8s", "start_pserver"]
 
     def _get_trainer_entrypoint(sefl):
+        if self._entry:
+            return ["paddle_k8s", "start_trainer", "v2"]
         return ["paddle_k8s", "start_trainer", "v1"]
 
     def _get_trainer_labels(self):

paddlecloud/paddlejob/views.py

@@ -34,8 +34,10 @@ def post(self, request, format=None):
         username = request.user.username
         namespace = notebook.utils.email_escape(username)
         obj = json.loads(request.body)
-        if not obj.get("topology"):
-            return utils.simple_response(500, "no topology specified")
+        topology = obj.get("topology", "")
+        entry = obj.get("entry", "")
+        if not topology and not entry:
+            return utils.simple_response(500, "no topology or entry specified")
         if not obj.get("datacenter"):
             return utils.simple_response(500, "no datacenter specified")
         dc = obj.get("datacenter")
@@ -72,7 +74,8 @@ def post(self, request, format=None):
             pservers = obj.get("pservers", 1),
             pscpu = obj.get("pscpu", 1),
             psmemory = obj.get("psmemory", "1Gi"),
-            topology = obj["topology"],
+            topology = topology,
+            entry = entry,
             gpu = obj.get("gpu", 0),
             image = obj.get("image", settings.JOB_DOCKER_IMAGE["image"]),
             passes = obj.get("passes", 1),