NORMALIZE_ADVANTAGE: tune.grid_search([False]),
"env": MultiAgentEnvWrapper,
"env_config": {
"env_name": args.env_name,
"num_agents": tune.grid_search([1, 3, 7, 10])
},
# should be fixed
"kl_coeff": 1.0,
"num_sgd_iter": 10,
"lr": 0.0002,
'sample_batch_size': 50,
'sgd_minibatch_size': 64,
'train_batch_size': 2048,
"num_gpus": 0.4,
"num_cpus_per_worker": 1,
"num_cpus_for_driver": 1,
"num_envs_per_worker": 5,
'num_workers': 1,
}
train(extra_config=walker_config,
trainer=DECETrainer,
env_name=walker_config['env_config']['env_name'],
stop={"timesteps_total": int(5e6) if not test else 2000},
exp_name=exp_name,
num_agents=walker_config['env_config']['num_agents'],
num_seeds=args.num_seeds,
num_gpus=args.num_gpus,
test_mode=args.test,
verbose=1)
"num_sgd_iter": 20,
"lr": 0.0001,
"horizon": 5000,
'sample_batch_size': 200,
'sgd_minibatch_size': 10000,
'train_batch_size': 100000,
"num_gpus": 1,
"num_cpus_per_worker": 1,
"num_cpus_for_driver": 1,
"num_envs_per_worker": 8,
'num_workers': 24,
"object_store_memory": int(5 * GB),
"memory": int(25 * GB)
}
train(
extra_config=walker_config,
trainer=DECETrainer,
env_name=walker_config['env_config']['env_name'],
stop={"timesteps_total": int(5e8)},
exp_name=exp_name,
num_agents=walker_config['env_config']['num_agents'],
num_seeds=args.num_seeds,
num_gpus=args.num_gpus,
test_mode=args.test,
verbose=1,
init_memory=int(70 * GB),
init_object_store_memory=int(15 * GB),
init_redis_max_memory=int(6 * GB),
)
"env_config": {
"env_name": "Walker2d-v3",
"num_agents": args.num_agents
},
# should be fixed
"kl_coeff": 1.0,
"num_sgd_iter": 20,
"lr": 0.0001,
'sample_batch_size': 256,
'sgd_minibatch_size': 4096,
'train_batch_size': 65536,
"num_gpus": 0.5,
"num_cpus_per_worker": 0.5,
"num_envs_per_worker": 16,
'num_workers': 8,
}
config = humanoid_config if not args.walker else walker_config
train(
extra_config=config,
trainer=PPOESTrainer,
env_name=config['env_config']['env_name'],
stop={"timesteps_total": int(2e8) if not args.walker else int(1e7)},
exp_name="DELETEME-TEST" if args.test else args.exp_name,
num_agents=args.num_agents if not args.test else 3,
num_seeds=args.num_seeds,
num_gpus=args.num_gpus,
test_mode=args.test,
# address=args.address if args.address else None
)
"env": MultiAgentEnvWrapper,
"env_config": {
"env_name": "Walker2d-v3",
"num_agents": tune.grid_search([3, 5, 10])
},
# should be fixed
"kl_coeff": 1.0,
"num_sgd_iter": 20,
"lr": 0.0001,
'sample_batch_size': 256,
'sgd_minibatch_size': 4096,
'train_batch_size': 65536,
"num_gpus": 0.45,
"num_cpus_per_worker": 0.5,
"num_cpus_for_driver": 0.5,
"num_envs_per_worker": 16,
'num_workers': 8,
}
train(
extra_config=walker_config,
trainer=PPOESTrainer,
env_name=walker_config['env_config']['env_name'],
stop={"timesteps_total": int(5e7)},
exp_name=exp_name,
num_agents=walker_config['env_config']['num_agents'],
num_seeds=3,
num_gpus=4,
test_mode=False
)
