def get_agent(params):
env = params['general_setting']['env']
# params['general_setting']['collector'] = BaseCollector(
# env
# )
if len(env.observation_space.shape) == 3:
params['net']['base_type'] = networks.CNNBase
if params['env']['frame_stack']:
buffer_param = params['replay_buffer']
efficient_buffer = replay_buffers.MemoryEfficientReplayBuffer(
int(buffer_param['size']))
params['general_setting']['replay_buffer'] = efficient_buffer
else:
params['net']['base_type'] = networks.MLPBase
if params['agent'] == 'sac':
pf = policies.GuassianContPolicy(
input_shape=env.observation_space.shape[0],
output_shape=2 * env.action_space.shape[0],
**params['net'])
vf = networks.Net(input_shape=env.observation_space.shape[0],
output_shape=1,
**params['net'])
qf = networks.FlattenNet(input_shape=env.observation_space.shape[0] +
env.action_space.shape[0],
output_shape=1,
**params['net'])
pretrain_pf = policies.UniformPolicyContinuous(
env.action_space.shape[0])
return SAC(pf=pf,
vf=vf,
qf=qf,
pretrain_pf=pretrain_pf,
**params['sac'],
**params['general_setting'])
if params['agent'] == 'twin_sac':
pf = policies.GuassianContPolicy(
input_shape=env.observation_space.shape[0],
output_shape=2 * env.action_space.shape[0],
**params['net'])
vf = networks.Net(input_shape=env.observation_space.shape[0],
output_shape=1,
**params['net'])
qf1 = networks.FlattenNet(input_shape=env.observation_space.shape[0] +
env.action_space.shape[0],
output_shape=1,
**params['net'])
qf2 = networks.FlattenNet(input_shape=env.observation_space.shape[0] +
env.action_space.shape[0],
output_shape=1,
**params['net'])
pretrain_pf = policies.UniformPolicyContinuous(
env.action_space.shape[0])
return TwinSAC(pf=pf,
vf=vf,
qf1=qf1,
qf2=qf2,
pretrain_pf=pretrain_pf,
**params['twin_sac'],
**params['general_setting'])
if params['agent'] == 'td3':
pf = policies.DetContPolicy(input_shape=env.observation_space.shape[0],
output_shape=env.action_space.shape[0],
**params['net'])
qf1 = networks.FlattenNet(input_shape=env.observation_space.shape[0] +
env.action_space.shape[0],
output_shape=1,
**params['net'])
qf2 = networks.FlattenNet(input_shape=env.observation_space.shape[0] +
env.action_space.shape[0],
output_shape=1,
**params['net'])
pretrain_pf = policies.UniformPolicyContinuous(
env.action_space.shape[0])
return TD3(pf=pf,
qf1=qf1,
qf2=qf2,
pretrain_pf=pretrain_pf,
**params['td3'],
**params['general_setting'])
if params['agent'] == 'ddpg':
pf = policies.DetContPolicy(input_shape=env.observation_space.shape[0],
output_shape=env.action_space.shape[0],
**params['net'])
qf = networks.FlattenNet(input_shape=env.observation_space.shape[0] +
env.action_space.shape[0],
output_shape=1,
**params['net'])
pretrain_pf = policies.UniformPolicyContinuous(
env.action_space.shape[0])
return DDPG(pf=pf,
qf=qf,
pretrain_pf=pretrain_pf,
**params['ddpg'],
**params['general_setting'])
if params['agent'] == 'dqn':
qf = networks.Net(input_shape=env.observation_space.shape,
output_shape=env.action_space.n,
**params['net'])
pf = policies.EpsilonGreedyDQNDiscretePolicy(
qf=qf, action_shape=env.action_space.n, **params['policy'])
pretrain_pf = policies.UniformPolicyDiscrete(
action_num=env.action_space.n)
params["general_setting"]["optimizer_class"] = optim.RMSprop
return DQN(pf=pf,
qf=qf,
pretrain_pf=pretrain_pf,
**params["dqn"],
**params["general_setting"])
if params['agent'] == 'bootstrapped dqn':
qf = networks.BootstrappedNet(
input_shape=env.observation_space.shape,
output_shape=env.action_space.n,
head_num=params['bootstrapped dqn']['head_num'],
**params['net'])
pf = policies.BootstrappedDQNDiscretePolicy(
qf=qf,
head_num=params['bootstrapped dqn']['head_num'],
action_shape=env.action_space.n,
**params['policy'])
pretrain_pf = policies.UniformPolicyDiscrete(
action_num=env.action_space.n)
params["general_setting"]["optimizer_class"] = optim.RMSprop
return BootstrappedDQN(pf=pf,
qf=qf,
pretrain_pf=pretrain_pf,
**params["bootstrapped dqn"],
**params["general_setting"])
if params['agent'] == 'qrdqn':
qf = networks.Net(input_shape=env.observation_space.shape,
output_shape=env.action_space.n *
params["qrdqn"]["quantile_num"],
**params['net'])
pf = policies.EpsilonGreedyQRDQNDiscretePolicy(
qf=qf, action_shape=env.action_space.n, **params['policy'])
pretrain_pf = policies.UniformPolicyDiscrete(
action_num=env.action_space.n)
return QRDQN(pf=pf,
qf=qf,
pretrain_pf=pretrain_pf,
**params["qrdqn"],
**params["general_setting"])
# On Policy Methods
act_space = env.action_space
params[params['agent']]['continuous'] = isinstance(act_space,
gym.spaces.Box)
buffer_param = params['replay_buffer']
buffer = replay_buffers.OnPolicyReplayBuffer(int(buffer_param['size']))
params['general_setting']['replay_buffer'] = buffer
if params[params['agent']]['continuous']:
pf = policies.GuassianContPolicy(
input_shape=env.observation_space.shape,
output_shape=2 * env.action_space.shape[0],
**params['net'])
else:
print(params['policy'])
print(params['net'])
# print(**params['policy'])
pf = policies.CategoricalDisPolicy(
input_shape=env.observation_space.shape,
output_shape=env.action_space.n,
**params['net'],
**params['policy'])
if params['agent'] == 'reinforce':
return Reinforce(pf=pf,
**params["reinforce"],
**params["general_setting"])
# Actor-Critic Frameworks
vf = networks.Net(input_shape=env.observation_space.shape,
output_shape=1,
**params['net'])
if params['agent'] == 'a2c':
return A2C(pf=pf, vf=vf, **params["a2c"], **params["general_setting"])
if params['agent'] == 'ppo':
return PPO(pf=pf, vf=vf, **params["ppo"], **params["general_setting"])
raise Exception("specified algorithm is not implemented")
def experiment(args):
import torch.multiprocessing as mp
mp.set_start_method('spawn')
device = torch.device("cuda:{}".format(args.device) if args.cuda else "cpu")
env = get_env( params['env_name'], params['env'])
env.seed(args.seed)
torch.manual_seed(args.seed)
np.random.seed(args.seed)
if args.cuda:
torch.backends.cudnn.deterministic=True
buffer_param = params['replay_buffer']
experiment_name = os.path.split( os.path.splitext( args.config )[0] )[-1] if args.id is None \
else args.id
logger = Logger( experiment_name , params['env_name'], args.seed, params, args.log_dir )
params['general_setting']['env'] = env
# replay_buffer = OnPolicyReplayBuffer(int(buffer_param['size']))
# example_ob = env.reset()
# example_dict = {
# "obs": example_ob,
# "next_obs": example_ob,
# "acts": env.action_space.sample(),
# "values": [0],
# "rewards": [0],
# "terminals": [False]
# }
replay_buffer = OnPolicyReplayBuffer( int(buffer_param['size']))
# replay_buffer.build_by_example(example_dict)
params['general_setting']['replay_buffer'] = replay_buffer
params['general_setting']['logger'] = logger
params['general_setting']['device'] = device
params['net']['base_type']=networks.MLPBase
pf = policies.CategoricalDisPolicy(
input_shape = env.observation_space.shape[0],
output_shape = env.action_space.n,
**params['net'],
**params['policy']
)
vf = networks.Net(
input_shape = env.observation_space.shape,
output_shape = 1,
**params['net']
)
params['general_setting']['collector'] = OnPlicyCollectorBase(
vf, env = env, pf = pf, replay_buffer = replay_buffer, device = "cuda",
train_render=False
)
# params['general_setting']['collector'] = ParallelOnPlicyCollector(
# vf, env = env, pf = pf, replay_buffer = replay_buffer, device=device, worker_nums=2
# )
params['general_setting']['save_dir'] = osp.join(logger.work_dir,"model")
agent = PPO(
pf = pf,
vf = vf,
**params["ppo"],
**params["general_setting"]
)
agent.train()
def experiment(args):
device = torch.device(
"cuda:{}".format(args.device) if args.cuda else "cpu")
env = get_vec_env(
params["env_name"],
params["env"],
args.vec_env_nums
)
env.seed(args.seed)
torch.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
if args.cuda:
torch.cuda.manual_seed_all(args.seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
buffer_param = params['replay_buffer']
experiment_name = os.path.split(
os.path.splitext(args.config)[0])[-1] if args.id is None \
else args.id
logger = Logger(
experiment_name, params['env_name'],
args.seed, params, args.log_dir, args.overwrite)
params['general_setting']['env'] = env
replay_buffer = OnPolicyReplayBuffer(
env_nums=args.vec_env_nums,
max_replay_buffer_size=int(buffer_param['size']),
time_limit_filter=buffer_param['time_limit_filter']
)
params['general_setting']['replay_buffer'] = replay_buffer
params['general_setting']['logger'] = logger
params['general_setting']['device'] = device
params['net']['base_type'] = networks.CNNBase
params['net']['activation_func'] = torch.nn.Tanh
print(env.observation_space.shape)
print(env.action_space.n)
pf = policies.CategoricalDisPolicy(
input_shape=env.observation_space.shape,
output_shape=env.action_space.n,
**params['net'],
**params['policy']
)
vf = networks.Net(
input_shape=env.observation_space.shape,
output_shape=1,
**params['net']
)
print(pf)
print(vf)
params['general_setting']['collector'] = VecOnPolicyCollector(
vf, env=env, pf=pf,
replay_buffer=replay_buffer, device=device,
train_render=False,
**params["collector"]
)
params['general_setting']['save_dir'] = osp.join(
logger.work_dir, "model")
agent = PPO(
pf=pf,
vf=vf,
**params["ppo"],
**params["general_setting"]
)
agent.train()