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 = Agent(env, args, hyper_params, models, optims, exploration_noise,
target_policy_noise)
# run
if args.test:
agent.test()
else:
agent.train()
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]
hidden_sizes_actor = hyper_params["NETWORK"]["ACTOR_HIDDEN_SIZES"]
hidden_sizes_critic = hyper_params["NETWORK"]["CRITIC_HIDDEN_SIZES"]
# 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 critic1
critic1 = MLP(
input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_critic,
).to(device)
critic1_target = MLP(
input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_critic,
).to(device)
critic1_target.load_state_dict(critic1.state_dict())
# create critic2
critic2 = MLP(
input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_critic,
).to(device)
critic2_target = MLP(
input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_critic,
).to(device)
critic2_target.load_state_dict(critic2.state_dict())
# concat critic parameters to use one optim
critic_parameters = list(critic1.parameters()) + list(critic2.parameters())
# create optimizer
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
exploration_noise = GaussianNoise(
action_dim,
min_sigma=hyper_params["EXPLORATION_NOISE"],
max_sigma=hyper_params["EXPLORATION_NOISE"],
)
target_policy_noise = GaussianNoise(
action_dim,
min_sigma=hyper_params["TARGET_POLICY_NOISE"],
max_sigma=hyper_params["TARGET_POLICY_NOISE"],
)
# make tuples to create an agent
models = (actor, actor_target, critic1, critic1_target, critic2,
critic2_target)
optims = (actor_optim, critic_optim)
noises = (exploration_noise, target_policy_noise)
# create an agent
return Agent(env, args, hyper_params, models, optims, noises)
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
critic_1 = MLP(
input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_critic,
).to(device)
critic_2 = MLP(
input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_critic,
).to(device)
critic_target1 = MLP(
input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_critic,
).to(device)
critic_target2 = MLP(
input_size=state_dim + action_dim,
output_size=1,
hidden_sizes=hidden_sizes_critic,
).to(device)
critic_target1.load_state_dict(critic_1.state_dict())
critic_target2.load_state_dict(critic_2.state_dict())
# create optimizers
actor_optim = optim.Adam(
actor.parameters(),
lr=hyper_params["LR_ACTOR"],
weight_decay=hyper_params["WEIGHT_DECAY"],
)
critic_parameter = list(critic_1.parameters()) + list(critic_2.parameters())
critic_optim = optim.Adam(
critic_parameter,
lr=hyper_params["LR_CRITIC_1"],
weight_decay=hyper_params["WEIGHT_DECAY"],
)
# noise instance to make randomness of action
noise = GaussianNoise(
hyper_params["GAUSSIAN_NOISE_MIN_SIGMA"],
hyper_params["GAUSSIAN_NOISE_MAX_SIGMA"],
hyper_params["GAUSSIAN_NOISE_DECAY_PERIOD"],
)
# make tuples to create an agent
models = (actor, actor_target, critic_1, critic_2, critic_target1, critic_target2)
optims = (actor_optim, critic_optim)
# create an agent
agent = Agent(env, args, hyper_params, models, optims, noise)
# run
if args.test:
agent.test()
else:
agent.train()