def evaluate(self):
"""Evaluate."""
eval_env = VecFrameStack(self.env, self.frame_stack)
self.pi.eval()
misc.set_env_to_eval_mode(eval_env)
# Eval policy
os.makedirs(os.path.join(self.logdir, 'eval'), exist_ok=True)
outfile = os.path.join(self.logdir, 'eval',
self.ckptr.format.format(self.t) + '.json')
stats = rl_evaluate(eval_env, self.pi, self.eval_num_episodes, outfile,
self.device)
logger.add_scalar('eval/mean_episode_reward', stats['mean_reward'],
self.t, time.time())
logger.add_scalar('eval/mean_episode_length', stats['mean_length'],
self.t, time.time())
# Record policy
os.makedirs(os.path.join(self.logdir, 'video'), exist_ok=True)
outfile = os.path.join(self.logdir, 'video',
self.ckptr.format.format(self.t) + '.mp4')
rl_record(eval_env, self.pi, self.record_num_episodes, outfile,
self.device)
self.pi.train()
misc.set_env_to_train_mode(self.env)
self.data_manager.manual_reset()
def make_atari_env(game_name,
nenv=1,
seed=0,
sticky_actions=True,
timelimit=True,
noop=False,
frameskip=4,
episode_life=False,
clip_rewards=True,
frame_stack=1,
scale=False,
timelimit_maxsteps=None):
"""Create an Atari environment."""
id = game_name + 'NoFrameskip'
id += '-v0' if sticky_actions else '-v4'
def _env(rank):
def _thunk():
env = gym.make(id)
if not timelimit:
env = env.env
elif timelimit_maxsteps:
env = TimeLimit(env.env, timelimit_maxsteps)
assert 'NoFrameskip' in env.spec.id
if noop:
env = atari_wrappers.NoopResetEnv(env, noop_max=30)
env = atari_wrappers.MaxAndSkipEnv(env, skip=frameskip)
env = StepOnEndOfLifeEnv(env)
env = EpisodeInfo(env)
env.seed(seed + rank)
env = atari_wrappers.wrap_deepmind(env,
episode_life=episode_life,
clip_rewards=clip_rewards,
frame_stack=False,
scale=scale)
env = ImageTranspose(env)
return env
return _thunk
if nenv > 1:
env = SubprocVecEnv([_env(i) for i in range(nenv)], context='fork')
else:
env = DummyVecEnv([_env(0)])
if frame_stack > 1:
env = VecFrameStack(env, frame_stack)
return env
def __init__(self,
logdir,
env_fn,
policy_fn,
qf_fn,
nenv=1,
optimizer=torch.optim.Adam,
buffer_size=10000,
frame_stack=1,
learning_starts=1000,
update_period=1,
batch_size=256,
policy_lr=1e-3,
qf_lr=1e-3,
gamma=0.99,
target_update_period=1,
policy_update_period=1,
target_smoothing_coef=0.005,
alpha=0.2,
automatic_entropy_tuning=True,
target_entropy=None,
gpu=True,
eval_num_episodes=1,
record_num_episodes=1,
log_period=1000):
"""Init."""
self.logdir = logdir
self.ckptr = Checkpointer(os.path.join(logdir, 'ckpts'))
self.env_fn = env_fn
self.nenv = nenv
self.eval_num_episodes = eval_num_episodes
self.record_num_episodes = record_num_episodes
self.gamma = gamma
self.buffer_size = buffer_size
self.frame_stack = frame_stack
self.learning_starts = learning_starts
self.update_period = update_period
self.batch_size = batch_size
if target_update_period < self.update_period:
self.target_update_period = self.update_period
else:
self.target_update_period = target_update_period - (
target_update_period % self.update_period)
if policy_update_period < self.update_period:
self.policy_update_period = self.update_period
else:
self.policy_update_period = policy_update_period - (
policy_update_period % self.update_period)
self.target_smoothing_coef = target_smoothing_coef
self.log_period = log_period
self.device = torch.device(
'cuda:0' if gpu and torch.cuda.is_available() else 'cpu')
self.env = VecEpisodeLogger(env_fn(nenv=nenv))
eval_env = VecFrameStack(self.env, self.frame_stack)
self.pi = policy_fn(eval_env)
self.qf1 = qf_fn(eval_env)
self.qf2 = qf_fn(eval_env)
self.target_qf1 = qf_fn(eval_env)
self.target_qf2 = qf_fn(eval_env)
self.pi.to(self.device)
self.qf1.to(self.device)
self.qf2.to(self.device)
self.target_qf1.to(self.device)
self.target_qf2.to(self.device)
self.opt_pi = optimizer(self.pi.parameters(), lr=policy_lr)
self.opt_qf1 = optimizer(self.qf1.parameters(), lr=qf_lr)
self.opt_qf2 = optimizer(self.qf2.parameters(), lr=qf_lr)
self.target_qf1.load_state_dict(self.qf1.state_dict())
self.target_qf2.load_state_dict(self.qf2.state_dict())
self.buffer = BatchedReplayBuffer(
*
[ReplayBuffer(buffer_size, frame_stack) for _ in range(self.nenv)])
self.data_manager = ReplayBufferDataManager(self.buffer, self.env,
SACActor(self.pi),
self.device,
self.learning_starts,
self.update_period)
self.alpha = alpha
self.automatic_entropy_tuning = automatic_entropy_tuning
if self.automatic_entropy_tuning:
if target_entropy:
self.target_entropy = target_entropy
else:
target_entropies = nest.map_structure(
lambda space: -np.prod(space.shape).item(),
misc.unpack_space(self.env.action_space))
self.target_entropy = sum(nest.flatten(target_entropies))
self.log_alpha = torch.tensor(np.log([self.alpha]),
requires_grad=True,
device=self.device,
dtype=torch.float32)
self.opt_alpha = optimizer([self.log_alpha], lr=policy_lr)
else:
self.target_entropy = None
self.log_alpha = None
self.opt_alpha = None
self.mse_loss = torch.nn.MSELoss()
self.t = 0
def __init__(self,
logdir,
env_fn,
qf_fn,
nenv=1,
optimizer=torch.optim.RMSprop,
buffer_size=100000,
frame_stack=1,
learning_starts=10000,
update_period=1,
gamma=0.99,
huber_loss=True,
exploration_timesteps=1000000,
final_eps=0.1,
eval_eps=0.05,
target_update_period=10000,
batch_size=32,
gpu=True,
eval_num_episodes=1,
record_num_episodes=1,
log_period=10):
"""Init."""
self.logdir = logdir
self.ckptr = Checkpointer(os.path.join(logdir, 'ckpts'))
self.env_fn = env_fn
self.nenv = nenv
self.eval_num_episodes = eval_num_episodes
self.record_num_episodes = record_num_episodes
self.gamma = gamma
self.frame_stack = frame_stack
self.buffer_size = buffer_size
self.batch_size = batch_size
self.learning_starts = learning_starts
self.update_period = update_period
self.eval_eps = eval_eps
self.target_update_period = target_update_period - (
target_update_period % self.update_period)
self.log_period = log_period
self.device = torch.device(
'cuda:0' if gpu and torch.cuda.is_available() else 'cpu')
self.env = VecEpisodeLogger(env_fn(nenv=nenv))
stacked_env = VecFrameStack(env_fn(nenv=nenv), self.frame_stack)
self.qf = qf_fn(stacked_env).to(self.device)
self.qf_targ = qf_fn(stacked_env).to(self.device)
self.opt = optimizer(self.qf.parameters())
if huber_loss:
self.criterion = torch.nn.SmoothL1Loss(reduction='none')
else:
self.criterion = torch.nn.MSELoss(reduction='none')
self.eps_schedule = LinearSchedule(exploration_timesteps, final_eps,
1.0)
self._actor = EpsilonGreedyActor(self.qf, self.eps_schedule,
self.env.action_space)
self.buffer = ReplayBuffer(self.buffer_size, self.frame_stack)
self.data_manager = ReplayBufferDataManager(self.buffer, self.env,
self._actor, self.device,
self.learning_starts,
self.update_period)
self.t = 0
def __init__(self,
logdir,
env_fn,
policy_fn,
qf_fn,
nenv=1,
optimizer=torch.optim.Adam,
buffer_size=10000,
frame_stack=1,
learning_starts=1000,
update_period=1,
batch_size=256,
policy_lr=1e-4,
qf_lr=1e-3,
qf_weight_decay=0.01,
gamma=0.99,
noise_theta=0.15,
noise_sigma=0.2,
noise_sigma_final=0.01,
noise_decay_period=10000,
target_update_period=1,
target_smoothing_coef=0.005,
reward_scale=1,
gpu=True,
eval_num_episodes=1,
record_num_episodes=1,
log_period=1000):
"""Init."""
self.logdir = logdir
self.ckptr = Checkpointer(os.path.join(logdir, 'ckpts'))
self.env_fn = env_fn
self.nenv = nenv
self.eval_num_episodes = eval_num_episodes
self.record_num_episodes = record_num_episodes
self.gamma = gamma
self.buffer_size = buffer_size
self.frame_stack = frame_stack
self.learning_starts = learning_starts
self.update_period = update_period
self.batch_size = batch_size
if target_update_period < self.update_period:
self.target_update_period = self.update_period
else:
self.target_update_period = target_update_period - (
target_update_period % self.update_period)
self.reward_scale = reward_scale
self.target_smoothing_coef = target_smoothing_coef
self.log_period = log_period
self.device = torch.device(
'cuda:0' if gpu and torch.cuda.is_available() else 'cpu')
self.t = 0
self.env = VecEpisodeLogger(env_fn(nenv=nenv))
self.policy_fn = policy_fn
self.qf_fn = qf_fn
eval_env = VecFrameStack(self.env, self.frame_stack)
self.pi = policy_fn(eval_env)
self.qf = qf_fn(eval_env)
self.target_pi = policy_fn(eval_env)
self.target_qf = qf_fn(eval_env)
self.pi.to(self.device)
self.qf.to(self.device)
self.target_pi.to(self.device)
self.target_qf.to(self.device)
self.optimizer = optimizer
self.policy_lr = policy_lr
self.qf_lr = qf_lr
self.qf_weight_decay = qf_weight_decay
self.opt_pi = optimizer(self.pi.parameters(), lr=policy_lr)
self.opt_qf = optimizer(self.qf.parameters(),
lr=qf_lr,
weight_decay=qf_weight_decay)
self.target_pi.load_state_dict(self.pi.state_dict())
self.target_qf.load_state_dict(self.qf.state_dict())
self.noise_schedule = LinearSchedule(noise_decay_period,
noise_sigma_final, noise_sigma)
self._actor = DDPGActor(self.pi, self.env.action_space, noise_theta,
self.noise_schedule.value(self.t))
self.buffer = ReplayBuffer(buffer_size, frame_stack)
self.data_manager = ReplayBufferDataManager(self.buffer, self.env,
self._actor, self.device,
self.learning_starts,
self.update_period)
self.qf_criterion = torch.nn.MSELoss()
if self.env.action_space.__class__.__name__ == 'Discrete':
raise ValueError("Action space must be continuous!")
def __init__(self,
logdir,
env_fn,
policy_fn,
qf_fn,
nenv=1,
optimizer=torch.optim.Adam,
buffer_size=int(1e6),
frame_stack=1,
learning_starts=10000,
update_period=1,
batch_size=256,
lr=3e-4,
policy_update_period=2,
target_smoothing_coef=0.005,
reward_scale=1,
gamma=0.99,
exploration_noise=0.1,
policy_noise=0.2,
policy_noise_clip=0.5,
gpu=True,
eval_num_episodes=1,
record_num_episodes=1,
log_period=1000):
"""Init."""
self.logdir = logdir
self.ckptr = Checkpointer(os.path.join(logdir, 'ckpts'))
self.env_fn = env_fn
self.nenv = nenv
self.eval_num_episodes = eval_num_episodes
self.record_num_episodes = record_num_episodes
self.gamma = gamma
self.buffer_size = buffer_size
self.batch_size = batch_size
self.frame_stack = frame_stack
self.learning_starts = learning_starts
self.update_period = update_period
if policy_update_period < self.update_period:
self.policy_update_period = self.update_period
else:
self.policy_update_period = policy_update_period - (
policy_update_period % self.update_period)
self.reward_scale = reward_scale
self.target_smoothing_coef = target_smoothing_coef
self.exploration_noise = exploration_noise
self.policy_noise = policy_noise
self.policy_noise_clip = policy_noise_clip
self.log_period = log_period
self.device = torch.device(
'cuda:0' if gpu and torch.cuda.is_available() else 'cpu')
self.policy_fn = policy_fn
self.qf_fn = qf_fn
self.env = VecEpisodeLogger(env_fn(nenv=nenv))
eval_env = VecFrameStack(self.env, self.frame_stack)
self.pi = policy_fn(eval_env)
self.qf1 = qf_fn(eval_env)
self.qf2 = qf_fn(eval_env)
self.target_pi = policy_fn(eval_env)
self.target_qf1 = qf_fn(eval_env)
self.target_qf2 = qf_fn(eval_env)
self.pi.to(self.device)
self.qf1.to(self.device)
self.qf2.to(self.device)
self.target_pi.to(self.device)
self.target_qf1.to(self.device)
self.target_qf2.to(self.device)
self.optimizer = optimizer
self.lr = lr
self.opt_pi = optimizer(self.pi.parameters(), lr=lr)
self.opt_qf = optimizer(list(self.qf1.parameters()) +
list(self.qf2.parameters()),
lr=lr)
self.target_pi.load_state_dict(self.pi.state_dict())
self.target_qf1.load_state_dict(self.qf1.state_dict())
self.target_qf2.load_state_dict(self.qf2.state_dict())
self._actor = TD3Actor(self.pi, self.env.action_space,
exploration_noise)
self.buffer = ReplayBuffer(buffer_size, frame_stack)
self.data_manager = ReplayBufferDataManager(self.buffer, self.env,
self._actor, self.device,
self.learning_starts,
self.update_period)
self.qf_criterion = torch.nn.MSELoss()
if self.env.action_space.__class__.__name__ == 'Discrete':
raise ValueError("Action space must be continuous!")
self.low = torch.from_numpy(self.env.action_space.low).to(self.device)
self.high = torch.from_numpy(self.env.action_space.high).to(
self.device)
self.t = 0
def __init__(self,
logdir,
env_fn,
policy_fn,
qf_fn,
vf_fn,
nenv=1,
optimizer=torch.optim.Adam,
buffer_size=10000,
frame_stack=1,
learning_starts=1000,
update_period=1,
batch_size=256,
policy_lr=1e-3,
qf_lr=1e-3,
vf_lr=1e-3,
policy_mean_reg_weight=1e-3,
gamma=0.99,
target_update_period=1,
policy_update_period=1,
target_smoothing_coef=0.005,
automatic_entropy_tuning=True,
reparameterization_trick=True,
target_entropy=None,
reward_scale=1,
gpu=True,
eval_num_episodes=1,
record_num_episodes=1,
log_period=1000):
"""Init."""
self.logdir = logdir
self.ckptr = Checkpointer(os.path.join(logdir, 'ckpts'))
self.env_fn = env_fn
self.nenv = nenv
self.eval_num_episodes = eval_num_episodes
self.record_num_episodes = record_num_episodes
self.gamma = gamma
self.buffer_size = buffer_size
self.frame_stack = frame_stack
self.learning_starts = learning_starts
self.update_period = update_period
self.batch_size = batch_size
if target_update_period < self.update_period:
self.target_update_period = self.update_period
else:
self.target_update_period = target_update_period - (
target_update_period % self.update_period)
if policy_update_period < self.update_period:
self.policy_update_period = self.update_period
else:
self.policy_update_period = policy_update_period - (
policy_update_period % self.update_period)
self.rsample = reparameterization_trick
self.reward_scale = reward_scale
self.target_smoothing_coef = target_smoothing_coef
self.log_period = log_period
self.device = torch.device('cuda:0' if gpu and torch.cuda.is_available()
else 'cpu')
self.env = VecEpisodeLogger(env_fn(nenv=nenv))
eval_env = VecFrameStack(self.env, self.frame_stack)
self.pi = policy_fn(eval_env)
self.qf1 = qf_fn(eval_env)
self.qf2 = qf_fn(eval_env)
self.vf = vf_fn(eval_env)
self.target_vf = vf_fn(eval_env)
self.pi.to(self.device)
self.qf1.to(self.device)
self.qf2.to(self.device)
self.vf.to(self.device)
self.target_vf.to(self.device)
self.opt_pi = optimizer(self.pi.parameters(), lr=policy_lr)
self.opt_qf1 = optimizer(self.qf1.parameters(), lr=qf_lr)
self.opt_qf2 = optimizer(self.qf2.parameters(), lr=qf_lr)
self.opt_vf = optimizer(self.vf.parameters(), lr=vf_lr)
self.policy_mean_reg_weight = policy_mean_reg_weight
self.target_vf.load_state_dict(self.vf.state_dict())
self.buffer = ReplayBuffer(buffer_size, frame_stack)
self.data_manager = ReplayBufferDataManager(self.buffer,
self.env,
SACActor(self.pi),
self.device,
self.learning_starts,
self.update_period)
self.discrete = self.env.action_space.__class__.__name__ == 'Discrete'
self.automatic_entropy_tuning = automatic_entropy_tuning
if self.automatic_entropy_tuning:
if target_entropy:
self.target_entropy = target_entropy
else:
# heuristic value from Tuomas
if self.discrete:
self.target_entropy = np.log(1.5)
else:
self.target_entropy = -np.prod(
self.env.action_space.shape).item()
self.log_alpha = torch.zeros(1, requires_grad=True,
device=self.device)
self.opt_alpha = optimizer([self.log_alpha], lr=policy_lr)
else:
self.target_entropy = None
self.log_alpha = None
self.opt_alpha = None
self.qf_criterion = torch.nn.MSELoss()
self.vf_criterion = torch.nn.MSELoss()
self.t = 0
