fix paddle version bug; refine scripts (#186)

examples/NeurIPS2019-Learn-to-Move-Challenge/README.md

@@ -6,7 +6,7 @@ The **PARL** team gets the first place in NeurIPS reinforcement learning competi
 
 ## Dependencies
 - python3.6
-- [paddlepaddle==1.5.2](https://github.com/PaddlePaddle/Paddle)
+- [paddlepaddle==1.5.1](https://github.com/PaddlePaddle/Paddle)
 - [parl>=1.2.1](https://github.com/PaddlePaddle/PARL)
 - [osim-rl==3.0.11](https://github.com/stanfordnmbl/osim-rl)
 

examples/NeurIPS2019-Learn-to-Move-Challenge/evaluate.py

@@ -22,9 +22,10 @@ import numpy as np
 from actor import Actor
 from opensim_model import OpenSimModel
 from opensim_agent import OpenSimAgent
-from parl.utils import logger, ReplayMemory, tensorboard
+from parl.utils import logger, ReplayMemory, tensorboard, get_gpu_count
 from parl.utils.window_stat import WindowStat
 from parl.remote.client import get_global_client
+from parl.utils import machine_info
 from shutil import copy2
 
 ACT_DIM = 22
@@ -62,6 +63,15 @@ class ActorState(object):
 
 class Evaluator(object):
     def __init__(self, args):
+        if machine_info.is_gpu_available():
+            assert get_gpu_count() == 1, 'Only support training in single GPU,\
+                    Please set environment variable: `export CUDA_VISIBLE_DEVICES=[GPU_ID_TO_USE]` .'
+
+        else:
+            cpu_num = os.environ.get('CPU_NUM')
+            assert cpu_num is not None and cpu_num == '1', 'Only support training in single CPU,\
+                    Please set environment variable:  `export CPU_NUM=1`.'
+
         model = OpenSimModel(OBS_DIM, VEL_DIM, ACT_DIM)
         algorithm = parl.algorithms.DDPG(
             model,

examples/NeurIPS2019-Learn-to-Move-Challenge/evaluate_args.py

@@ -20,7 +20,7 @@ def get_args():
 
     parser.add_argument(
         '--cluster_address',
-        default='localhost:8081',
+        default='localhost:8010',
         type=str,
         help='cluster address of xparl.')
     parser.add_argument(

examples/NeurIPS2019-Learn-to-Move-Challenge/scripts/eval_difficulty1.sh

+# use which GPU
+export CUDA_VISIBLE_DEVICES=0
+
 python evaluate.py --actor_num 160 \
            --difficulty 1 \
            --penalty_coeff 3.0 \

examples/NeurIPS2019-Learn-to-Move-Challenge/scripts/eval_difficulty2.sh

+# use which GPU
+export CUDA_VISIBLE_DEVICES=0
+
+
 python evaluate.py --actor_num 160 \
            --difficulty 2 \
            --penalty_coeff 5.0 \

examples/NeurIPS2019-Learn-to-Move-Challenge/scripts/eval_difficulty3.sh

+# use which GPU
+export CUDA_VISIBLE_DEVICES=0
+
 python evaluate.py --actor_num 160 \
            --difficulty 3 \
            --vel_penalty_coeff 3.0 \

examples/NeurIPS2019-Learn-to-Move-Challenge/scripts/eval_difficulty3_first_target.sh

+# use which GPU
+export CUDA_VISIBLE_DEVICES=0
+
 python evaluate.py --actor_num 160 \
            --difficulty 3 \
            --vel_penalty_coeff 3.0 \

examples/NeurIPS2019-Learn-to-Move-Challenge/scripts/train_difficulty1.sh

@@ -4,6 +4,9 @@ if [ $# != 1 ]; then
     exit 0
 fi
 
+# use which GPU
+export CUDA_VISIBLE_DEVICES=0
+
 python train.py --actor_num 300 \
            --difficulty 1 \
            --penalty_coeff 3.0 \

examples/NeurIPS2019-Learn-to-Move-Challenge/scripts/train_difficulty2.sh

@@ -3,6 +3,9 @@ if [ $# != 1 ]; then
     exit 0
 fi
 
+# use which GPU
+export CUDA_VISIBLE_DEVICES=0
+
 python train.py --actor_num 300 \
            --difficulty 2 \
            --penalty_coeff 5.0 \

examples/NeurIPS2019-Learn-to-Move-Challenge/scripts/train_difficulty3.sh

@@ -3,6 +3,9 @@ if [ $# != 1 ]; then
     exit 0
 fi
 
+# use which GPU
+export CUDA_VISIBLE_DEVICES=0
+
 python train.py --actor_num 300 \
            --difficulty 3 \
            --vel_penalty_coeff 3.0 \

examples/NeurIPS2019-Learn-to-Move-Challenge/scripts/train_difficulty3_first_target.sh

@@ -3,6 +3,9 @@ if [ $# != 1 ]; then
     exit 0
 fi
 
+# use which GPU
+export CUDA_VISIBLE_DEVICES=0
+
 python train.py --actor_num 300 \
            --difficulty 3 \
            --vel_penalty_coeff 3.0 \

examples/NeurIPS2019-Learn-to-Move-Challenge/train.py

@@ -22,9 +22,10 @@ import numpy as np
 from actor import Actor
 from opensim_model import OpenSimModel
 from opensim_agent import OpenSimAgent
-from parl.utils import logger, ReplayMemory, tensorboard
+from parl.utils import logger, ReplayMemory, tensorboard, get_gpu_count
 from parl.utils.window_stat import WindowStat
 from parl.remote.client import get_global_client
+from parl.utils import machine_info
 
 ACT_DIM = 22
 VEL_DIM = 19
@@ -68,6 +69,15 @@ class ActorState(object):
 
 class Learner(object):
     def __init__(self, args):
+        if machine_info.is_gpu_available():
+            assert get_gpu_count() == 1, 'Only support training in single GPU,\
+                    Please set environment variable: `export CUDA_VISIBLE_DEVICES=[GPU_ID_TO_USE]` .'
+
+        else:
+            cpu_num = os.environ.get('CPU_NUM')
+            assert cpu_num is not None and cpu_num == '1', 'Only support training in single CPU,\
+                    Please set environment variable:  `export CPU_NUM=1`.'
+
         model = OpenSimModel(OBS_DIM, VEL_DIM, ACT_DIM)
         algorithm = parl.algorithms.DDPG(
             model,

examples/NeurIPS2019-Learn-to-Move-Challenge/train_args.py

@@ -20,7 +20,7 @@ def get_args():
 
     parser.add_argument(
         '--cluster_address',
-        default='localhost:8081',
+        default='localhost:8010',
         type=str,
         help='cluster address of xparl.')
     parser.add_argument(