def main():
# env initialization
env_name = "FCTP-v1"
env = env_generator_FCTP(env_name, 100000) # max_episode_steps
env_utils.set_env(env, args)
env.set_comp(False) # True for two component actions, False for one
# set a random seed
common_utils.set_random_seed(args.seed, env)
# run
module_path = "fctp_v1." + args.algo # Set agent type from algo argument from cmd line
example = importlib.import_module(module_path)
example.run(env, env_name, args)
def main():
"""Main."""
# env initialization
env = gym.make("LunarLanderContinuous-v2")
env_utils.set_env(env, args)
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
# set a random seed
common_utils.set_random_seed(args.seed, env)
# run
module_path = "examples.lunarlander_continuous_v2." + args.algo
example = importlib.import_module(module_path)
example.run(env, args, state_dim, action_dim)
def run(env: gym.Env, args: argparse.Namespace):
"""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
"""
# create multiple envs
# configure environment so that it works for discrete actions
env_single = env_utils.set_env(env, args, WRAPPERS)
env_gen = env_generator("Pong-v0", args, WRAPPERS)
env_multi = make_envs(env_gen, n_envs=hyper_params["N_WORKERS"])
# create a model
action_dim = env.action_space.n
hidden_sizes = [256, 256]
def get_cnn_model():
cnn_model = DuelingCNN(
cnn_layers=[
CNNLayer(
input_size=4,
output_size=32,
kernel_size=5,
pulling_fn=nn.MaxPool2d(3),
),
CNNLayer(
input_size=32,
output_size=32,
kernel_size=3,
pulling_fn=nn.MaxPool2d(3),
),
CNNLayer(
input_size=32,
output_size=64,
kernel_size=2,
pulling_fn=nn.MaxPool2d(3),
),
],
fc_layers=DuelingMLP(
input_size=256, output_size=action_dim, hidden_sizes=hidden_sizes
),
).to(device)
return cnn_model
dqn = get_cnn_model()
dqn_target = get_cnn_model()
dqn_target.load_state_dict(dqn.state_dict())
# create optimizer
dqn_optim = optim.Adam(
dqn.parameters(),
lr=hyper_params["LR_DQN"],
weight_decay=hyper_params["WEIGHT_DECAY"],
)
# make tuples to create an agent
models = (dqn, dqn_target)
# create an agent
agent = Agent(env_single, env_multi, args, hyper_params, models, dqn_optim)
agent.env_name = "Pong-v0"
# run
if args.test:
agent.test()
else:
agent.train()