Merge pull request #10037 from putcn/fix-command-escape

fix training parameter issue

tools/aws_benchmarking/README.md

@@ -84,7 +84,8 @@ putcn/paddle_aws_client \
 --security_group_id <your security group id> \
 --docker_image myreponame/paddle_benchmark \
 --pserver_count 2 \
---trainer_count 2
+--trainer_count 2 \
+--trainer_command batch_size:20,local:no,device:CPU
 ```
 
 Now just wait until you see this:

tools/aws_benchmarking/client/cluster_launcher.py

@@ -80,7 +80,11 @@ parser.add_argument(
     use ami-1ae93962 for us-east-2")
 
 parser.add_argument(
-    '--pserver_command', type=str, default="", help="pserver start command")
+    '--pserver_command',
+    type=str,
+    default="",
+    help="pserver start command, format example: python,vgg.py,batch_size:128,is_local:yes"
+)
 
 parser.add_argument(
     '--trainer_image_id',
@@ -90,7 +94,11 @@ parser.add_argument(
     use ami-1ae93962 for us-west-2")
 
 parser.add_argument(
-    '--trainer_command', type=str, default="", help="trainer start command")
+    '--trainer_command',
+    type=str,
+    default="",
+    help="trainer start command, format example: python,vgg.py,batch_size:128,is_local:yes"
+)
 
 parser.add_argument(
     '--availability_zone',

tools/aws_benchmarking/server/cluster_master.py

@@ -19,6 +19,7 @@ import math
 import time
 import threading
 import logging
+import copy
 
 import netaddr
 import boto3
@@ -257,6 +258,8 @@ def script_to_str(file_path):
 
 
 def run_instances(image_id, instance_type, count, role, cmd=""):
+    if count == 0:
+        return []
     response = ec2client.run_instances(
         ImageId=image_id,
         InstanceType=instance_type,
@@ -334,6 +337,22 @@ def log_to_file(source, filename):
             log_file.write(line)
 
 
+def parse_command(command_raw, defaults={}):
+    if not command_raw:
+        command_raw = ""
+    commands_processed = []
+    parameter_map = copy.copy(defaults)
+    for seg in command_raw.split(","):
+        if ":" in seg:
+            parameters = seg.split(":")
+            parameter_map[parameters[0]] = parameters[1]
+        else:
+            commands_processed.append(seg)
+    for key, val in parameter_map.iteritems():
+        commands_processed.append("--" + key + " " + str(val))
+    return " ".join(commands_processed)
+
+
 def create_trainers(kickoff_cmd, pserver_endpoints_str):
     def create_and_start_trainer(trainer_index):
         logging.info("trainer " + str(trainer_index) + " is starting")
@@ -361,7 +380,7 @@ def create_trainers(kickoff_cmd, pserver_endpoints_str):
             TRAINER_INDEX=str(trainer_index),
             TASK_NAME=args.task_name,
             TRAINER_COUNT=args.trainer_count,
-            COMMAND=args.trainer_command,
+            COMMAND=parse_command(args.trainer_command, {"device": "GPU"}),
             MASTER_ENDPOINT=args.master_server_ip + ":" +
             str(args.master_server_port))
         logging.info(cmd)
@@ -476,7 +495,7 @@ def kickoff_pserver(host, pserver_endpoints_str):
             DOCKER_IMAGE=args.docker_image,
             PSERVER_PORT=args.pserver_port,
             TASK_NAME=args.task_name,
-            COMMAND=args.pserver_command,
+            COMMAND=parse_command(args.pserver_command, {"device": "CPU"}),
             TRAINER_COUNT=args.trainer_count,
             TRAINER_INDEX=0,
             # there is no way to use 0.0.0.0:port to start pserver

tools/aws_benchmarking/server/pserver.sh.template

 #!/bin/bash
-docker run --network="host" -i -e "SERVER_ENDPOINT={SERVER_ENDPOINT}" -e "MASTER_ENDPOINT={MASTER_ENDPOINT}" -e "TASK_NAME={TASK_NAME}" -e "TRAINER_INDEX={TRAINER_INDEX}" -e "TRAINING_ROLE=PSERVER" -e "TRAINER_COUNT={TRAINER_COUNT}" -e "TRAINERS={TRAINER_COUNT}" -e "PSERVER_HOSTS={PSERVER_HOSTS}" -e "PSERVERS={PSERVER_HOSTS}" {DOCKER_IMAGE} {COMMAND} --device CPU
+docker run --network="host" -i -e "SERVER_ENDPOINT={SERVER_ENDPOINT}" -e "MASTER_ENDPOINT={MASTER_ENDPOINT}" -e "TASK_NAME={TASK_NAME}" -e "TRAINER_INDEX={TRAINER_INDEX}" -e "TRAINING_ROLE=PSERVER" -e "TRAINER_COUNT={TRAINER_COUNT}" -e "TRAINERS={TRAINER_COUNT}" -e "PSERVER_HOSTS={PSERVER_HOSTS}" -e "PSERVERS={PSERVER_HOSTS}" {DOCKER_IMAGE} {COMMAND}
\ No newline at end of file

tools/aws_benchmarking/server/trainer.sh.template

 #!/bin/bash 
-nvidia-docker run --network="host" -i  -e "MASTER_ENDPOINT={MASTER_ENDPOINT}" -e "TASK_NAME={TASK_NAME}" -e "TRAINER_COUNT={TRAINER_COUNT}" -e "TRAINERS={TRAINER_COUNT}" -e "TRAINER_INDEX={TRAINER_INDEX}"  -e "PADDLE_INIT_TRAINER_ID={TRAINER_INDEX}" -e "TRAINING_ROLE=TRAINER"  -e "PSERVER_HOSTS={PSERVER_HOSTS}"  -e "PSERVERS={PSERVER_HOSTS}" {DOCKER_IMAGE} {COMMAND} --device GPU
+nvidia-docker run --network="host" -i  -e "MASTER_ENDPOINT={MASTER_ENDPOINT}" -e "TASK_NAME={TASK_NAME}" -e "TRAINER_COUNT={TRAINER_COUNT}" -e "TRAINERS={TRAINER_COUNT}" -e "TRAINER_INDEX={TRAINER_INDEX}"  -e "PADDLE_INIT_TRAINER_ID={TRAINER_INDEX}" -e "TRAINING_ROLE=TRAINER"  -e "PSERVER_HOSTS={PSERVER_HOSTS}"  -e "PSERVERS={PSERVER_HOSTS}" {DOCKER_IMAGE} {COMMAND}
\ No newline at end of file