def __init__(self, envs, args, device, rnn_state_size):
self.device = device
self.args = args
self.envs = envs
self.episode_rewards = deque(maxlen=100)
self.config = env_config(args.env_name)
self.is_embed = should_embed(args.env_name)
self.rollouts = RolloutStorage(self.args, self.config, rnn_state_size,
self.is_embed)
obs = self.envs.reset()
self.rollouts.obs[0].copy_(obs)
self.rollouts.to(self.device)
def __init__(self, envs, args):
self.args = args
self.envs = envs
assert (self.args.policy_file is not None)
self.episode_rewards = deque(maxlen=100)
self.config = env_config(args.env_name)
self.is_embed = should_embed(args.env_name)
self.actor_critic = torch.load(self.args.policy_file)
# We want to act in each step
self.args.num_steps = 1
self.rollouts = RolloutStorage(self.args, self.config,
self.actor_critic.rnn_state_size,
self.is_embed)
# Turn this on to save reconstructed beliefs to text files
self.log_est_belief = False
self.traj_cnt = 0
if self.log_est_belief:
self.actor_belief_file = open('ab' + str(self.traj_cnt) + ".txt",
"w")
self.critic_belief_file = open('cb' + str(self.traj_cnt) + ".txt",
"w")
obs = self.envs.reset()
state = self.envs.get_state()
belief = self.envs.get_belief()
self.rollouts.obs[0].copy_(obs)
self.rollouts.state[0].copy_(state)
self.rollouts.belief[0].copy_(belief)
self.rollouts.to('cpu')
def __init__(self, envs, args, log_dir):
self.device = args.device
self.args = args
self.envs = envs
self.episode_rewards = deque(maxlen=100)
self.config = env_config(args.env_name)
num_updates = int(
args.num_env_steps) // args.num_steps // args.num_processes
if self.args.use_linear_entropy_decay:
self.entropy_schedule = ExponentialSchedule(
args.entropy_coef, 2e-5, num_updates)
self.should_embed = should_embed(args.env_name)
self.experience_mem = []
self.model_path = os.path.join(
log_dir, self.args.algo + '.' + str(self.args.seed) + '.mdl')
# Using belief loss, saved file format is different
if self.args.algo == 'ah-ch' and self.args.belief_loss_coef > 0.0:
self.model_path = os.path.join(
log_dir,
self.args.algo + '.b' + '.' + str(self.args.seed) + '.mdl')
self.transitions_path = os.path.join(
log_dir, self.args.env_name + '.' + self.args.algo + '.b' +
'.' + str(self.args.seed) + '.exp')
else:
self.model_path = os.path.join(
log_dir, self.args.algo + '.' + str(self.args.seed) + '.mdl')
self.transitions_path = os.path.join(
log_dir, self.args.env_name + '.' + self.args.algo + '.' +
str(self.args.seed) + '.exp')
self.setup_agent()