def get(env, args):
"""Run training or test.
Args:
env (gym.Env): openAI Gym environment with continuous action space
args (argparse.Namespace): arguments including training settings
"""
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
if hyper_params["USE_HER"]:
state_dim *= 2
hidden_sizes_actor = hyper_params["NETWORK"]["ACTOR_HIDDEN_SIZES"]
hidden_sizes_vf = hyper_params["NETWORK"]["VF_HIDDEN_SIZES"]
hidden_sizes_qf = hyper_params["NETWORK"]["QF_HIDDEN_SIZES"]
# target entropy
target_entropy = -np.prod((action_dim, )).item() # heuristic
# create actor
actor = TanhGaussianDistParams(input_size=state_dim,
output_size=action_dim,
hidden_sizes=hidden_sizes_actor).to(device)
# create v_critic
vf = MLP(input_size=state_dim, output_size=1,
hidden_sizes=hidden_sizes_vf).to(device)
vf_target = MLP(input_size=state_dim,
output_size=1,
hidden_sizes=hidden_sizes_vf).to(device)
vf_target.load_state_dict(vf.state_dict())
# create q_critic
qf_1 = FlattenMLP(input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_qf).to(device)
qf_2 = FlattenMLP(input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_qf).to(device)
# create optimizers
actor_optim = optim.Adam(
actor.parameters(),
lr=hyper_params["LR_ACTOR"],
weight_decay=hyper_params["WEIGHT_DECAY"],
)
vf_optim = optim.Adam(
vf.parameters(),
lr=hyper_params["LR_VF"],
weight_decay=hyper_params["WEIGHT_DECAY"],
)
qf_1_optim = optim.Adam(
qf_1.parameters(),
lr=hyper_params["LR_QF1"],
weight_decay=hyper_params["WEIGHT_DECAY"],
)
qf_2_optim = optim.Adam(
qf_2.parameters(),
lr=hyper_params["LR_QF2"],
weight_decay=hyper_params["WEIGHT_DECAY"],
)
# make tuples to create an agent
models = (actor, vf, vf_target, qf_1, qf_2)
optims = (actor_optim, vf_optim, qf_1_optim, qf_2_optim)
# HER
her = LunarLanderContinuousHER() if hyper_params["USE_HER"] else None
# create an agent
return Agent(env, args, hyper_params, models, optims, target_entropy, her)
def run(env: gym.Env, args: argparse.Namespace, state_dim: int,
action_dim: int):
"""Run training or test.
Args:
env (gym.Env): openAI Gym environment with continuous action space
args (argparse.Namespace): arguments including training settings
state_dim (int): dimension of states
action_dim (int): dimension of actions
"""
hidden_sizes_actor = [400, 300]
hidden_sizes_critic = [400, 300]
# create actor
actor = MLP(
input_size=state_dim,
output_size=action_dim,
hidden_sizes=hidden_sizes_actor,
output_activation=torch.tanh,
).to(device)
actor_target = MLP(
input_size=state_dim,
output_size=action_dim,
hidden_sizes=hidden_sizes_actor,
output_activation=torch.tanh,
).to(device)
actor_target.load_state_dict(actor.state_dict())
# create critic
critic1 = FlattenMLP(
input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_critic,
).to(device)
critic2 = FlattenMLP(
input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_critic,
).to(device)
critic_target1 = FlattenMLP(
input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_critic,
).to(device)
critic_target2 = FlattenMLP(
input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_critic,
).to(device)
critic_target1.load_state_dict(critic1.state_dict())
critic_target2.load_state_dict(critic2.state_dict())
# concat critic parameters to use one optim
critic_parameters = list(critic1.parameters()) + list(critic2.parameters())
# create optimizers
actor_optim = optim.Adam(
actor.parameters(),
lr=hyper_params["LR_ACTOR"],
weight_decay=hyper_params["WEIGHT_DECAY"],
)
critic_optim = optim.Adam(
critic_parameters,
lr=hyper_params["LR_CRITIC"],
weight_decay=hyper_params["WEIGHT_DECAY"],
)
# noise instance to make randomness of action
exploration_noise = GaussianNoise(action_dim,
hyper_params["EXPLORATION_NOISE"],
hyper_params["EXPLORATION_NOISE"])
target_policy_noise = GaussianNoise(
action_dim,
hyper_params["TARGET_POLICY_NOISE"],
hyper_params["TARGET_POLICY_NOISE"],
)
# make tuples to create an agent
models = (actor, actor_target, critic1, critic2, critic_target1,
critic_target2)
optims = (actor_optim, critic_optim)
# create an agent
agent = TD3Agent(env, args, hyper_params, models, optims,
exploration_noise, target_policy_noise)
# run
if args.test:
agent.test()
else:
agent.train()
def run(env: gym.Env, args: argparse.Namespace, state_dim: int,
action_dim: int):
"""Run training or test.
Args:
env (gym.Env): openAI Gym environment with continuous action space
args (argparse.Namespace): arguments including training settings
state_dim (int): dimension of states
action_dim (int): dimension of actions
"""
hidden_sizes_actor = [256, 256]
hidden_sizes_vf = [256, 256]
hidden_sizes_qf = [256, 256]
# target entropy
target_entropy = -np.prod((action_dim, )).item() # heuristic
# create actor
actor = TanhGaussianDistParams(input_size=state_dim,
output_size=action_dim,
hidden_sizes=hidden_sizes_actor).to(device)
# create v_critic
vf = MLP(input_size=state_dim, output_size=1,
hidden_sizes=hidden_sizes_vf).to(device)
vf_target = MLP(input_size=state_dim,
output_size=1,
hidden_sizes=hidden_sizes_vf).to(device)
vf_target.load_state_dict(vf.state_dict())
# create q_critic
qf_1 = FlattenMLP(input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_qf).to(device)
qf_2 = FlattenMLP(input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_qf).to(device)
# create optimizers
actor_optim = optim.Adam(
actor.parameters(),
lr=hyper_params["LR_ACTOR"],
weight_decay=hyper_params["LAMBDA2"],
)
vf_optim = optim.Adam(vf.parameters(),
lr=hyper_params["LR_VF"],
weight_decay=hyper_params["LAMBDA2"])
qf_1_optim = optim.Adam(
qf_1.parameters(),
lr=hyper_params["LR_QF1"],
weight_decay=hyper_params["LAMBDA2"],
)
qf_2_optim = optim.Adam(
qf_2.parameters(),
lr=hyper_params["LR_QF2"],
weight_decay=hyper_params["LAMBDA2"],
)
# make tuples to create an agent
models = (actor, vf, vf_target, qf_1, qf_2)
optims = (actor_optim, vf_optim, qf_1_optim, qf_2_optim)
# create an agent
agent = SACfDAgent(env, args, hyper_params, models, optims, target_entropy)
# run
if args.test:
agent.test()
else:
agent.train()