def test_multiple_num_agents(local_mode=False):
num_gpus = 0
initialize_ray(test_mode=True, local_mode=local_mode, num_gpus=num_gpus)
config = _get_default_test_config(
tune.grid_search([2, 3, 4]), "BipedalWalker-v2", num_gpus
)
return tune.run(
CEPPOTrainer,
local_dir=get_local_dir(),
name="DELETEME_TEST_extra_loss_ppo_trainer",
stop={"timesteps_total": 5000},
config=config
)
def test_cetd3(local_mode=False):
num_gpus = 0
initialize_ray(test_mode=True, local_mode=local_mode, num_gpus=num_gpus)
config = _get_default_test_config(
num_agents=3, env_name="BipedalWalker-v2", num_gpus=num_gpus
)
if "num_sgd_iter" in config:
config.pop("num_sgd_iter")
config.pop("sgd_minibatch_size")
config['timesteps_per_iteration'] = 80
config['pure_exploration_steps'] = 80
config['learning_starts'] = 180
tune.run(
CETD3Trainer,
local_dir=get_local_dir(),
name="DELETEME_TEST_extra_loss_ppo_trainer",
stop={"timesteps_total": 2000},
config=config
)
def _base(
trainer,
local_mode=False,
extra_config=None,
t=500,
env_name="BipedalWalker-v2",
num_agents=3
):
# num_agents = 3
num_gpus = 0
initialize_ray(test_mode=True, local_mode=local_mode, num_gpus=num_gpus)
config = _get_default_test_config(num_agents, env_name, num_gpus)
if extra_config:
config.update(extra_config)
stop = {"timesteps_total": t} if not isinstance(t, dict) else t
return tune.run(
trainer,
local_dir=get_local_dir(),
name="DELETEME_TEST_extra_loss_ppo_trainer",
stop=stop,
config=config
)
from ray.tune.registry import register_env
def make_pybullet(_=None):
import pybullet_envs
import gym
print("Successfully import pybullet and found: ",
pybullet_envs.getList())
return gym.make(env_name)
register_env(env_name, make_pybullet)
analysis = tune.run(
"PPO",
local_dir=get_local_dir(),
name=exp_name,
checkpoint_freq=10,
keep_checkpoints_num=10,
checkpoint_score_attr="episode_reward_mean",
checkpoint_at_end=True,
stop={"info/num_steps_sampled": stop}
if isinstance(stop, int) else stop,
config=walker_config,
max_failures=20,
reuse_actors=False
)
path = "{}-{}-{}ts.pkl".format(
exp_name, env_name, stop
)
def train(
extra_config,
trainer,
env_name,
stop,
exp_name,
num_agents,
num_seeds,
num_gpus,
num_cpus=None,
test_mode=False,
address=None,
redis_password=None,
clip_memory=False,
init_memory=None,
init_object_store_memory=None,
init_redis_max_memory=None,
**kwargs
):
# assert isinstance(stop, int)
if address is not None:
num_gpus = None
if clip_memory:
init_memory = int(300 * GB)
init_object_store_memory = int(100 * GB)
init_redis_max_memory = int(50 * GB)
initialize_ray(
test_mode=test_mode,
local_mode=False,
num_gpus=num_gpus,
address=address,
redis_password=redis_password,
memory=init_memory,
object_store_memory=init_object_store_memory,
redis_max_memory=init_redis_max_memory,
num_cpus=num_cpus
)
env_config = {"env_name": env_name, "num_agents": num_agents}
config = {
"seed": tune.grid_search([i * 100 for i in range(num_seeds)]),
"env": MultiAgentEnvWrapper,
"env_config": env_config,
"log_level": "DEBUG" if test_mode else "INFO"
}
if extra_config:
config.update(extra_config)
analysis = tune.run(
trainer,
local_dir=get_local_dir(),
name=exp_name,
checkpoint_freq=10,
keep_checkpoints_num=10,
checkpoint_score_attr="episode_reward_mean",
checkpoint_at_end=True,
stop={"info/num_steps_sampled": stop}
if isinstance(stop, int) else stop,
config=config,
max_failures=20,
reuse_actors=False,
**kwargs
)
path = "{}-{}-{}ts-{}agents.pkl".format(
exp_name, env_name, stop, num_agents
)
with open(path, "wb") as f:
data = analysis.fetch_trial_dataframes()
pickle.dump(data, f)
print("Result is saved at: <{}>".format(path))
return analysis