def assign_env(env_config):
#print(env_config)
print("worker index is {}".format(env_config.worker_index))
if env_config.worker_index % 2 == 0:
print('in even env')
env = sonic_on_ray.make(game='Airstriker-Genesis', state='Level1')
else:
print('in odd env')
env = sonic_on_ray.make(game='Airstriker-Genesis', state='Level1')
return env
def assign_env(env_config):
#print(env_config)
print("worker index is {}".format(env_config.worker_index))
if env_config.worker_index % 2 == 0:
print('in even env')
env = sonic_on_ray.make(game='BustAMove-Snes',
state='BustAMove.1pplay.Level10')
else:
print('in odd env')
env = sonic_on_ray.make(game='BustAMove-Snes',
state='BustAMove.1pplay.Level40')
return env
def __init__(self, env_config):
# pick actual env based on worker and env indexes
#print("worker index is {}".format(env_config.worker_index))
#print("testing vector_index {}".format(env_config.vector_index))
#BustAMove.Challengeplay0
challenge_level = env_config.worker_index % 5
self.env = sonic_on_ray.make(game='BustAMove-Snes', state='BustAMove.Challengeplay{}'.format(challenge_level)) #BustAMove.1pplay.Level10
self.action_space = self.env.action_space
self.observation_space = self.env.observation_space
from ray.tune.registry import register_env
import sonic_on_ray
import random
env_name = 'bustamove_env'
# Note that the hyperparameters have been tuned for sonic, which can be used
# run by replacing the below function with:
#
# register_env(env_name, lambda config: sonic_on_ray.make(
# game='SonicTheHedgehog-Genesis',
# state='GreenHillZone.Act1'))
#
# However, to try Sonic, you have to obtain the ROM yourself (see then
# instructions at https://github.com/openai/retro/blob/master/README.md).
register_env(
env_name, lambda config: sonic_on_ray.make(game='ContraIII-Snes',
state='level1.1player.easy'))
# env_name2 = 'bustamove_env2'
# register_env(env_name2,
# lambda config: sonic_on_ray.make(game='BustAMove-Snes',
# state='BustAMove.1pplay.Level1'))
ray.init()
run_experiments({
'bustamove-ppo-random-search': {
'run': 'PPO',
'env': 'bustamove_env',
'stop': {
'timesteps_total': 200000
},
'repeat': 10,
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--save-checkpoint-dir',
help='Checkpoint dir',
default=None)
parser.add_argument('--load-checkpoint',
help='Path to existing checkpoint created by _save',
default=None)
parser.add_argument('--local', help='Use retro_contest.local.make')
args = parser.parse_args()
env_name = 'sonic_env'
# Note that the hyperparameters have been tuned for sonic, which can be used
if args.local:
game = 'SonicTheHedgehog-Genesis'
state = 'GreenHillZone.Act1'
register_env(env_name,
lambda config: sonic_on_ray.make_local(game, state))
else:
register_env(env_name, lambda config: sonic_on_ray.make())
ray.init()
config = ppo.DEFAULT_CONFIG.copy()
config.update({
'timesteps_per_batch':
40000,
'min_steps_per_task':
100,
'num_workers':
32,
'gamma':
0.99,
'lambda':
0.95,
'clip_param':
0.1,
'num_sgd_iter':
30,
'sgd_batchsize':
4096,
'sgd_stepsize':
5e-5,
'use_gae':
True,
'horizon':
4000,
'devices': [
'/gpu:0', '/gpu:1', '/gpu:2', '/gpu:3', '/gpu:4', '/gpu:5',
'/gpu:6', 'gpu:7'
],
'tf_session_args': {
'gpu_options': {
'allow_growth': True
}
}
})
alg = ppo.PPOAgent(config=config, env=env_name)
print("Created a PSqO object")
if args.load_checkpoint is not None:
print("Trying to restore from checkpoint", args.load_checkpoint)
alg.restore(args.load_checkpoint)
print("Restored state from checkpoint:", args.load_checkpoint)
#for i in range(10):
while True:
try:
print("Starting to train")
result = alg.train()
print('result = {}'.format(result))
except gre.GymRemoteError as e:
print('exception', e)
# game='SonicTheHedgehog-Genesis',
# state='GreenHillZone.Act1'))
#
# However, to try Sonic, you have to obtain the ROM yourself (see then
# instructions at https://github.com/openai/retro/blob/master/README.md).
# register_env(env_name,
# lambda config: sonic_on_ray.make(game='Airstriker-Genesis',
# state='Level1'))
# register_env(env_name,
# lambda config: sonic_on_ray.make(game='SonicTheHedgehog-Genesis',
# state='GreenHillZone.Act1'))
register_env(
env_name,
lambda config: sonic_on_ray.make(game='BustAMove-Snes',
state='BustAMove.1pplay.Level1'))
ray.init()
config = ppo.DEFAULT_CONFIG.copy()
config.update({
#'timesteps_per_batch': 40000,
'timesteps_per_batch': 8,
#'min_steps_per_task': 100,
#'num_workers': 32,
'num_workers': 1,
'gamma': 0.99,
'lambda': 0.95,
'clip_param': 0.1,
'num_sgd_iter': 3,
if __name__ == '__main__':
args = parser.parse_args()
if not args.config:
# Load configuration from file
config_dir = os.path.dirname(args.checkpoint)
config_path = os.path.join(config_dir, 'params.json')
with open(config_path) as f:
args.config = json.load(f)
ray.init()
env_name = 'sonic_env'
register_env(
env_name,
lambda config: sonic_on_ray.make(game='SonicTheHedgehog-Genesis',
state='GreenHillZone.Act1'))
cls = get_agent_class(args.run)
agent = cls(env=env_name, config=args.config)
num_steps = int(args.steps)
# This currently only works with PPO.
env = agent.local_evaluator.env
if args.out is not None:
rollouts = []
steps = 0
while steps < (num_steps or steps + 1):
if args.out is not None:
rollout = []
import ray
from ray.rllib import ppo
from ray.tune.registry import register_env
env_name = 'sonic_env'
# Note that the hyperparameters have been tuned for sonic, which can be used
# run by replacing the below function with:
#
# register_env(env_name, lambda config: sonic_on_ray.make(
# game='SonicTheHedgehog-Genesis',
# state='GreenHillZone.Act1'))
#
# However, to try Sonic, you have to obtain the ROM yourself (see then
# instructions at https://github.com/openai/retro/blob/master/README.md).
register_env(
env_name, lambda config: sonic_on_ray.make(game='Airstriker-Genesis',
state='Level1'))
ray.init()
config = ppo.DEFAULT_CONFIG.copy()
config.update({
'timesteps_per_batch':
40000,
'min_steps_per_task':
100,
'num_workers':
32,
'gamma':
0.99,
'lambda':