def experiment(args):
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']
replay_buffer = BaseReplayBuffer(int(buffer_param['size']))
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
params['general_setting']['replay_buffer'] = replay_buffer
params['general_setting']['logger'] = logger
params['general_setting']['device'] = device
params['net']['base_type'] = networks.MLPBase
# agent = get_agent( params )
# print(env)
# params['general_setting']['collector'] = BaseCollector(
# env, pf, replay_buffer
# )
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])
params['general_setting']['collector'] = BaseCollector(env,
pf,
replay_buffer,
device=device)
params['general_setting']['save_dir'] = osp.join(logger.work_dir, "model")
agent = SAC(pf=pf,
vf=vf,
qf=qf,
pretrain_pf=pretrain_pf,
**params['sac'],
**params['general_setting'])
agent.train()
def experiment(args):
device = torch.device(
"cuda:{}".format(args.device) if args.cuda else "cpu")
env = get_env(params['env_name'], params['env'])
env.train()
eval_env = get_env(params['env_name'], params['env'])
eval_env.eval()
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,
overwrite=args.overwrite
)
params['general_setting']['env'] = env
replay_buffer = OnPolicyReplayBuffer(
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.MLPBase
params['net']['activation_func'] = torch.nn.Tanh
pf = policies.GuassianContPolicyBasicBias(
input_shape=env.observation_space.shape[0],
output_shape=env.action_space.shape[0],
**params['net'],
**params['policy']
)
vf = networks.Net(
input_shape=env.observation_space.shape,
output_shape=1,
**params['net']
)
params['general_setting']['collector'] = OnPolicyCollectorBase(
vf, env=env, eval_env=eval_env,
pf=pf, replay_buffer=replay_buffer, device=device,
train_render=False,
**params["collector"]
)
print(pf)
print(vf)
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):
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_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
params['general_setting']['logger'] = logger
params['general_setting']['device'] = device
params['net']['base_type']=networks.MLPBase
import torch.multiprocessing as mp
mp.set_start_method('spawn')
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])
example_ob = env.reset()
example_dict = {
"obs": example_ob,
"next_obs": example_ob,
"acts": env.action_space.sample(),
"rewards": [0],
"terminals": [False]
}
replay_buffer = SharedBaseReplayBuffer( int(buffer_param['size']),
1
)
replay_buffer.build_by_example(example_dict)
params['general_setting']['replay_buffer'] = replay_buffer
params['general_setting']['collector'] = ParallelCollector(
env, pf, replay_buffer, device=device, worker_nums=1
)
params['general_setting']['save_dir'] = osp.join(logger.work_dir,"model")
agent = SAC(
pf = pf,
vf = vf,
qf = qf,
**params['sac'],
**params['general_setting']
)
agent.train()
def experiment(args):
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)
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)
params['general_setting']['env'] = env
replay_buffer = BaseReplayBuffer(
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.MLPBase
params['net']['activation_func'] = torch.nn.ReLU
pf = policies.GuassianContPolicy(
input_shape=env.observation_space.shape[0],
output_shape=2 * env.action_space.shape[0],
**params['net'],
**params['policy'])
qf1 = networks.QNet(
input_shape=env.observation_space.shape[0] + env.action_space.shape[0],
output_shape=1,
**params['net'])
qf2 = networks.QNet(
input_shape=env.observation_space.shape[0] + env.action_space.shape[0],
output_shape=1,
**params['net'])
print(pf)
print(qf1)
params['general_setting']['collector'] = BaseCollector(
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 = TwinSACQ(
pf=pf,
qf1=qf1,
qf2=qf2,
**params["twin_sac_q"],
**params["general_setting"]
)
agent.train()
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)
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)
params['general_setting']['env'] = env
replay_buffer = OnPolicyReplayBuffer(
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.MLPBase
params['net']['activation_func'] = nn.Tanh
pf = policies.GuassianContPolicyBasicBias(
input_shape=env.observation_space.shape[0],
output_shape=env.action_space.shape[0],
init_func=lambda x: init.orthogonal_init(
x, scale=np.sqrt(2), constant=0),
net_last_init_func=lambda x: init.orthogonal_init(
x, scale=0.01, constant=0),
**params['net'],
**params['policy'])
vf = networks.Net(input_shape=env.observation_space.shape,
output_shape=1,
init_func=lambda x: init.orthogonal_init(
x, scale=np.sqrt(2), constant=0),
net_last_init_func=lambda x: init.orthogonal_init(
x, scale=1, constant=0),
**params['net'])
params['general_setting']['collector'] = OnPlicyCollectorBase(
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 = TRPO(pf=pf, vf=vf, **params["trpo"], **params["general_setting"])
print(params["general_setting"])
print(agent.epoch_frames)
agent.train()
