def test_nested_ob(self):
"""Test."""
env = make_env('CartPole-v1')
qf = QFunction(
FeedForwardBase(env.observation_space, env.action_space))
env = NestedVecObWrapper(env)
buffer = ReplayBuffer(2000, 1)
data_manager = ReplayBufferDataManager(buffer,
env,
act_fn=BufferActor(qf),
device='cpu',
learning_starts=50,
update_period=2)
for _ in range(11):
data_manager.step_until_update()
assert buffer.num_in_buffer == 70
batch = data_manager.sample(32)
data_manager.act(batch['obs'])
assert batch['action'].shape == batch['reward'].shape
assert batch['action'].shape == batch['done'].shape
if isinstance(batch['obs'], list):
assert batch['obs'][0].shape == batch['next_obs'][0].shape
assert len(batch['obs'][0].shape) == 2
else:
assert batch['obs'].shape == batch['next_obs'].shape
assert len(batch['obs'].shape) == 2
assert len(batch['action'].shape) == 1
def test(self):
"""Test."""
buffer = ReplayBuffer(10, 4)
buffer = PrioritizedReplayBuffer(buffer, alpha=0.5)
env = atari_env('Pong').envs[0]
init_obs = env.reset()
idx = buffer.store_observation(init_obs)
assert np.allclose(buffer.encode_recent_observation()[:-3], 0)
for i in range(10):
ac = env.action_space.sample()
obs, r, done, _ = env.step(ac)
data = {'action': ac, 'reward': r, 'done': done, 'key1': 0}
buffer.store_effect(idx, data)
idx = buffer.store_observation(obs)
# Check sample shapes
s = buffer.sample(2, beta=1.)
assert len(s.keys()) == 8
assert s['obs'].shape == (2, 4, 84, 84)
assert s['next_obs'].shape == (2, 4, 84, 84)
assert s['key1'].shape == (2,)
s = buffer._encode_sample([4, 5])
# Check observation stacking
assert np.allclose(s['obs'][0][3], s['next_obs'][0][2])
assert np.allclose(s['obs'][0][2], s['next_obs'][0][1])
assert np.allclose(s['obs'][0][1], s['next_obs'][0][0])
# Check sequential samples
assert np.allclose(s['obs'][0][3], s['obs'][1][2])
# check priorities
buffer.update_priorities([4, 5], [0.5, 2])
assert buffer._it_sum[4] == 0.5 ** buffer._alpha
assert buffer._it_sum[5] == 1.0 ** buffer._alpha
assert buffer._it_min[4] == 0.5 ** buffer._alpha
assert buffer._it_min[5] == 1.0 ** buffer._alpha
assert buffer._max_priority == 1.0
# Check for wrap around when buffer is full
s = buffer._encode_sample([0])
assert not np.allclose(s['obs'][0][:-3], 0)
# Check saving and loading
state = buffer.state_dict()
buffer2 = ReplayBuffer(10, 4)
buffer2 = PrioritizedReplayBuffer(buffer2, alpha=0.5)
buffer2.load_state_dict(state)
s1 = buffer._encode_sample([1, 3, 5])
s2 = buffer2._encode_sample([1, 3, 5])
for k in s1:
assert np.allclose(s1[k], s2[k])
for i in range(10):
ac = env.action_space.sample()
obs, r, done, _ = env.step(ac)
data = {'action': ac, 'reward': r, 'done': done, 'key1': 0}
buffer.store_effect(idx, data)
buffer2.store_effect(idx, data)
idx = buffer.store_observation(obs)
idx2 = buffer2.store_observation(obs)
assert idx == idx2
assert buffer._max_priority == buffer2._max_priority
s1 = buffer._encode_sample([1, 3, 5])
s2 = buffer2._encode_sample([1, 3, 5])
for k in s1:
assert np.allclose(s1[k], s2[k])
buffer.update_priorities([4, 5], [0.1, 0.9])
buffer2.update_priorities([4, 5], [0.1, 0.9])
assert buffer._max_priority == buffer2._max_priority
assert buffer._it_sum[4] == buffer2._it_sum[4]
assert buffer._it_sum[5] == buffer2._it_sum[5]
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
class DQN(Algorithm):
"""DQN algorithm."""
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 _compute_target(self, rew, next_ob, done):
qtarg = self.qf_targ(next_ob).max_q
return rew + (1.0 - done) * self.gamma * qtarg
def _get_batch(self):
return self.data_manager.sample(self.batch_size)
def loss(self, batch):
"""Compute loss."""
q = self.qf(batch['obs'], batch['action']).value
with torch.no_grad():
target = self._compute_target(batch['reward'], batch['next_obs'],
batch['done'])
assert target.shape == q.shape
loss = self.criterion(target, q).mean()
if self.t % self.log_period < self.update_period:
logger.add_scalar('alg/maxq',
torch.max(q).detach().cpu().numpy(), self.t,
time.time())
logger.add_scalar('alg/loss',
loss.detach().cpu().numpy(), self.t, time.time())
logger.add_scalar('alg/epsilon',
self.eps_schedule.value(self._actor.t), self.t,
time.time())
return loss
def step(self):
"""Step."""
self.t += self.data_manager.step_until_update()
if self.t % self.target_update_period == 0:
self.qf_targ.load_state_dict(self.qf.state_dict())
self.opt.zero_grad()
loss = self.loss(self._get_batch())
loss.backward()
self.opt.step()
return self.t
def evaluate(self):
"""Evaluate."""
eval_env = VecEpsilonGreedy(VecFrameStack(self.env, self.frame_stack),
self.eval_eps)
self.qf.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.qf, 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.qf, self.record_num_episodes, outfile,
self.device)
self.qf.train()
misc.set_env_to_train_mode(self.env)
self.data_manager.manual_reset()
def save(self):
"""Save."""
state_dict = {
'qf': self.qf.state_dict(),
'qf_targ': self.qf.state_dict(),
'opt': self.opt.state_dict(),
'_actor': self._actor.state_dict(),
'env': misc.env_state_dict(self.env),
't': self.t
}
buffer_dict = self.buffer.state_dict()
state_dict['buffer_format'] = nest.get_structure(buffer_dict)
self.ckptr.save(state_dict, self.t)
# save buffer seperately and only once (because it can be huge)
np.savez(
os.path.join(self.ckptr.ckptdir, 'buffer.npz'),
**{f'{i:04d}': x
for i, x in enumerate(nest.flatten(buffer_dict))})
def load(self, t=None):
"""Load."""
state_dict = self.ckptr.load(t)
if state_dict is None:
self.t = 0
return self.t
self.qf.load_state_dict(state_dict['qf'])
self.qf_targ.load_state_dict(state_dict['qf_targ'])
self.opt.load_state_dict(state_dict['opt'])
self._actor.load_state_dict(state_dict['_actor'])
misc.env_load_state_dict(self.env, state_dict['env'])
self.t = state_dict['t']
buffer_format = state_dict['buffer_format']
buffer_state = dict(
np.load(os.path.join(self.ckptr.ckptdir, 'buffer.npz')))
buffer_state = nest.flatten(buffer_state)
self.buffer.load_state_dict(
nest.pack_sequence_as(buffer_state, buffer_format))
self.data_manager.manual_reset()
return self.t
def close(self):
"""Close environment."""
try:
self.env.close()
except Exception:
pass
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!")
class DDPG(Algorithm):
"""DDPG algorithm."""
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 loss(self, batch):
"""Loss function."""
# compute QFunction loss.
with torch.no_grad():
target_action = self.target_pi(batch['next_obs']).action
target_q = self.target_qf(batch['next_obs'], target_action).value
qtarg = self.reward_scale * batch['reward'].float() + (
(1.0 - batch['done']) * self.gamma * target_q)
q = self.qf(batch['obs'], batch['action']).value
assert qtarg.shape == q.shape
qf_loss = self.qf_criterion(q, qtarg)
# compute policy loss
action = self.pi(batch['obs'], deterministic=True).action
q = self.qf(batch['obs'], action).value
pi_loss = -q.mean()
# log losses
if self.t % self.log_period < self.update_period:
logger.add_scalar('loss/qf', qf_loss, self.t, time.time())
logger.add_scalar('loss/pi', pi_loss, self.t, time.time())
return pi_loss, qf_loss
def step(self):
"""Step optimization."""
self._actor.update_sigma(self.noise_schedule.value(self.t))
self.t += self.data_manager.step_until_update()
if self.t % self.target_update_period == 0:
soft_target_update(self.target_pi, self.pi,
self.target_smoothing_coef)
soft_target_update(self.target_qf, self.qf,
self.target_smoothing_coef)
if self.t % self.update_period == 0:
batch = self.data_manager.sample(self.batch_size)
pi_loss, qf_loss = self.loss(batch)
# update
self.opt_qf.zero_grad()
qf_loss.backward()
self.opt_qf.step()
self.opt_pi.zero_grad()
pi_loss.backward()
self.opt_pi.step()
return self.t
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 save(self):
"""Save."""
state_dict = {
'pi': self.pi.state_dict(),
'qf': self.qf.state_dict(),
'target_pi': self.target_pi.state_dict(),
'target_qf': self.target_qf.state_dict(),
'opt_pi': self.opt_pi.state_dict(),
'opt_qf': self.opt_qf.state_dict(),
'env': misc.env_state_dict(self.env),
't': self.t
}
buffer_dict = self.buffer.state_dict()
state_dict['buffer_format'] = nest.get_structure(buffer_dict)
self.ckptr.save(state_dict, self.t)
# save buffer seperately and only once (because it can be huge)
np.savez(
os.path.join(self.ckptr.ckptdir, 'buffer.npz'),
**{f'{i:04d}': x
for i, x in enumerate(nest.flatten(buffer_dict))})
def load(self, t=None):
"""Load."""
state_dict = self.ckptr.load(t)
if state_dict is None:
self.t = 0
return self.t
self.pi.load_state_dict(state_dict['pi'])
self.qf.load_state_dict(state_dict['qf'])
self.target_pi.load_state_dict(state_dict['target_pi'])
self.target_qf.load_state_dict(state_dict['target_qf'])
self.opt_pi.load_state_dict(state_dict['opt_pi'])
self.opt_qf.load_state_dict(state_dict['opt_qf'])
misc.env_load_state_dict(self.env, state_dict['env'])
self.t = state_dict['t']
buffer_format = state_dict['buffer_format']
buffer_state = dict(
np.load(os.path.join(self.ckptr.ckptdir, 'buffer.npz')))
buffer_state = nest.flatten(buffer_state)
self.buffer.load_state_dict(
nest.pack_sequence_as(buffer_state, buffer_format))
self.data_manager.manual_reset()
return self.t
def close(self):
"""Close environment."""
try:
self.env.close()
except Exception:
pass
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
class TD3(Algorithm):
"""TD3 algorithm."""
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 loss(self, batch):
"""Loss function."""
# compute QFunction loss.
with torch.no_grad():
target_action = self.target_pi(batch['next_obs']).action
noise = (torch.randn_like(target_action) *
self.policy_noise).clamp(-self.policy_noise_clip,
self.policy_noise_clip)
target_action = (target_action + noise).clamp(-1., 1.)
target_q1 = self.target_qf1(batch['next_obs'], target_action).value
target_q2 = self.target_qf2(batch['next_obs'], target_action).value
target_q = torch.min(target_q1, target_q2)
qtarg = self.reward_scale * batch['reward'].float() + (
(1.0 - batch['done']) * self.gamma * target_q)
q1 = self.qf1(batch['obs'], batch['action']).value
q2 = self.qf2(batch['obs'], batch['action']).value
assert qtarg.shape == q1.shape
assert qtarg.shape == q2.shape
qf_loss = self.qf_criterion(q1, qtarg) + self.qf_criterion(q2, qtarg)
# compute policy loss
if self.t % self.policy_update_period == 0:
action = self.pi(batch['obs'], deterministic=True).action
q = self.qf1(batch['obs'], action).value
pi_loss = -q.mean()
else:
pi_loss = torch.zeros_like(qf_loss)
# log losses
if self.t % self.log_period < self.update_period:
logger.add_scalar('loss/qf', qf_loss, self.t, time.time())
if self.t % self.policy_update_period == 0:
logger.add_scalar('loss/pi', pi_loss, self.t, time.time())
return pi_loss, qf_loss
def step(self):
"""Step optimization."""
self.t += self.data_manager.step_until_update()
batch = self.data_manager.sample(self.batch_size)
pi_loss, qf_loss = self.loss(batch)
# update
self.opt_qf.zero_grad()
qf_loss.backward()
self.opt_qf.step()
if self.t % self.policy_update_period == 0:
self.opt_pi.zero_grad()
pi_loss.backward()
self.opt_pi.step()
# update target networks
soft_target_update(self.target_pi, self.pi,
self.target_smoothing_coef)
soft_target_update(self.target_qf1, self.qf1,
self.target_smoothing_coef)
soft_target_update(self.target_qf2, self.qf2,
self.target_smoothing_coef)
return self.t
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 save(self):
"""Save."""
state_dict = {
'pi': self.pi.state_dict(),
'qf1': self.qf1.state_dict(),
'qf2': self.qf2.state_dict(),
'target_pi': self.target_pi.state_dict(),
'target_qf1': self.target_qf1.state_dict(),
'target_qf2': self.target_qf2.state_dict(),
'opt_pi': self.opt_pi.state_dict(),
'opt_qf': self.opt_qf.state_dict(),
'env': misc.env_state_dict(self.env),
't': self.t
}
buffer_dict = self.buffer.state_dict()
state_dict['buffer_format'] = nest.get_structure(buffer_dict)
self.ckptr.save(state_dict, self.t)
# save buffer seperately and only once (because it can be huge)
np.savez(
os.path.join(self.ckptr.ckptdir, 'buffer.npz'),
**{f'{i:04d}': x
for i, x in enumerate(nest.flatten(buffer_dict))})
def load(self, t=None):
"""Load."""
state_dict = self.ckptr.load(t)
if state_dict is None:
self.t = 0
return self.t
self.pi.load_state_dict(state_dict['pi'])
self.qf1.load_state_dict(state_dict['qf1'])
self.qf2.load_state_dict(state_dict['qf2'])
self.target_pi.load_state_dict(state_dict['target_pi'])
self.target_qf1.load_state_dict(state_dict['target_qf1'])
self.target_qf2.load_state_dict(state_dict['target_qf2'])
self.opt_pi.load_state_dict(state_dict['opt_pi'])
self.opt_qf.load_state_dict(state_dict['opt_qf'])
misc.env_load_state_dict(self.env, state_dict['env'])
self.t = state_dict['t']
buffer_format = state_dict['buffer_format']
buffer_state = dict(
np.load(os.path.join(self.ckptr.ckptdir, 'buffer.npz')))
buffer_state = nest.flatten(buffer_state)
self.buffer.load_state_dict(
nest.pack_sequence_as(buffer_state, buffer_format))
self.data_manager.manual_reset()
return self.t
def close(self):
"""Close environment."""
try:
self.env.close()
except Exception:
pass
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
class SAC(Algorithm):
"""SAC algorithm."""
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
def loss(self, batch):
"""Loss function."""
pi_out = self.pi(batch['obs'], reparameterization_trick=self.rsample)
if self.discrete:
new_ac = pi_out.action
ent = pi_out.dist.entropy()
else:
assert isinstance(pi_out.dist, TanhNormal), (
"It is strongly encouraged that you use a TanhNormal "
"action distribution for continuous action spaces.")
if self.rsample:
new_ac, new_pth_ac = pi_out.dist.rsample(
return_pretanh_value=True)
else:
new_ac, new_pth_ac = pi_out.dist.sample(
return_pretanh_value=True)
logp = pi_out.dist.log_prob(new_ac, new_pth_ac)
q1 = self.qf1(batch['obs'], batch['action']).value
q2 = self.qf2(batch['obs'], batch['action']).value
v = self.vf(batch['obs']).value
# alpha loss
if self.automatic_entropy_tuning:
if self.discrete:
ent_error = -ent + self.target_entropy
else:
ent_error = logp + self.target_entropy
alpha_loss = -(self.log_alpha * ent_error.detach()).mean()
self.opt_alpha.zero_grad()
alpha_loss.backward()
self.opt_alpha.step()
alpha = self.log_alpha.exp()
else:
alpha = 1
alpha_loss = 0
# qf loss
vtarg = self.target_vf(batch['next_obs']).value
qtarg = self.reward_scale * batch['reward'].float() + (
(1.0 - batch['done']) * self.gamma * vtarg)
assert qtarg.shape == q1.shape
assert qtarg.shape == q2.shape
qf1_loss = self.qf_criterion(q1, qtarg.detach())
qf2_loss = self.qf_criterion(q2, qtarg.detach())
# vf loss
q1_outs = self.qf1(batch['obs'], new_ac)
q1_new = q1_outs.value
q2_new = self.qf2(batch['obs'], new_ac).value
q = torch.min(q1_new, q2_new)
if self.discrete:
vtarg = q + alpha * ent
else:
vtarg = q - alpha * logp
assert v.shape == vtarg.shape
vf_loss = self.vf_criterion(v, vtarg.detach())
# pi loss
pi_loss = None
if self.t % self.policy_update_period == 0:
if self.discrete:
target_dist = CatDist(logits=q1_outs.qvals.detach())
pi_dist = CatDist(logits=alpha * pi_out.dist.logits)
pi_loss = pi_dist.kl(target_dist).mean()
else:
if self.rsample:
assert q.shape == logp.shape
pi_loss = (alpha*logp - q1_new).mean()
else:
pi_targ = q1_new - v
assert pi_targ.shape == logp.shape
pi_loss = (logp * (alpha * logp - pi_targ).detach()).mean()
pi_loss += self.policy_mean_reg_weight * (
pi_out.dist.normal.mean**2).mean()
# log pi loss about as frequently as other losses
if self.t % self.log_period < self.policy_update_period:
logger.add_scalar('loss/pi', pi_loss, self.t, time.time())
if self.t % self.log_period < self.update_period:
if self.automatic_entropy_tuning:
logger.add_scalar('ent/log_alpha',
self.log_alpha.detach().cpu().numpy(), self.t,
time.time())
if self.discrete:
scalars = {"target": self.target_entropy,
"entropy": ent.mean().detach().cpu().numpy().item()}
else:
scalars = {"target": self.target_entropy,
"entropy": -torch.mean(
logp.detach()).cpu().numpy().item()}
logger.add_scalars('ent/entropy', scalars, self.t, time.time())
else:
if self.discrete:
logger.add_scalar(
'ent/entropy',
ent.mean().detach().cpu().numpy().item(),
self.t, time.time())
else:
logger.add_scalar(
'ent/entropy',
-torch.mean(logp.detach()).cpu().numpy().item(),
self.t, time.time())
logger.add_scalar('loss/qf1', qf1_loss, self.t, time.time())
logger.add_scalar('loss/qf2', qf2_loss, self.t, time.time())
logger.add_scalar('loss/vf', vf_loss, self.t, time.time())
return pi_loss, qf1_loss, qf2_loss, vf_loss
def step(self):
"""Step optimization."""
self.t += self.data_manager.step_until_update()
if self.t % self.target_update_period == 0:
soft_target_update(self.target_vf, self.vf,
self.target_smoothing_coef)
if self.t % self.update_period == 0:
batch = self.data_manager.sample(self.batch_size)
pi_loss, qf1_loss, qf2_loss, vf_loss = self.loss(batch)
# update
self.opt_qf1.zero_grad()
qf1_loss.backward()
self.opt_qf1.step()
self.opt_qf2.zero_grad()
qf2_loss.backward()
self.opt_qf2.step()
self.opt_vf.zero_grad()
vf_loss.backward()
self.opt_vf.step()
if pi_loss:
self.opt_pi.zero_grad()
pi_loss.backward()
self.opt_pi.step()
return self.t
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 save(self):
"""Save."""
state_dict = {
'pi': self.pi.state_dict(),
'qf1': self.qf1.state_dict(),
'qf2': self.qf2.state_dict(),
'vf': self.vf.state_dict(),
'opt_pi': self.opt_pi.state_dict(),
'opt_qf1': self.opt_qf1.state_dict(),
'opt_qf2': self.opt_qf2.state_dict(),
'opt_vf': self.opt_vf.state_dict(),
'log_alpha': (self.log_alpha if self.automatic_entropy_tuning
else None),
'opt_alpha': (self.opt_alpha.state_dict()
if self.automatic_entropy_tuning else None),
'env': misc.env_state_dict(self.env),
't': self.t
}
buffer_dict = self.buffer.state_dict()
state_dict['buffer_format'] = nest.get_structure(buffer_dict)
self.ckptr.save(state_dict, self.t)
# save buffer seperately and only once (because it can be huge)
np.savez(os.path.join(self.ckptr.ckptdir, 'buffer.npz'),
**{f'{i:04d}': x for i, x in
enumerate(nest.flatten(buffer_dict))})
def load(self, t=None):
"""Load."""
state_dict = self.ckptr.load(t)
if state_dict is None:
self.t = 0
return self.t
self.pi.load_state_dict(state_dict['pi'])
self.qf1.load_state_dict(state_dict['qf1'])
self.qf2.load_state_dict(state_dict['qf2'])
self.vf.load_state_dict(state_dict['vf'])
self.target_vf.load_state_dict(state_dict['vf'])
self.opt_pi.load_state_dict(state_dict['opt_pi'])
self.opt_qf1.load_state_dict(state_dict['opt_qf1'])
self.opt_qf2.load_state_dict(state_dict['opt_qf2'])
self.opt_vf.load_state_dict(state_dict['opt_vf'])
if state_dict['log_alpha']:
with torch.no_grad():
self.log_alpha.copy_(state_dict['log_alpha'])
self.opt_alpha.load_state_dict(state_dict['opt_alpha'])
misc.env_load_state_dict(self.env, state_dict['env'])
self.t = state_dict['t']
buffer_format = state_dict['buffer_format']
buffer_state = dict(np.load(os.path.join(self.ckptr.ckptdir,
'buffer.npz')))
buffer_state = nest.flatten(buffer_state)
self.buffer.load_state_dict(nest.pack_sequence_as(buffer_state,
buffer_format))
self.data_manager.manual_reset()
return self.t
def close(self):
"""Close environment."""
try:
self.env.close()
except Exception:
pass
class ResidualSAC(Algorithm):
"""Residual SAC algorithm.
The changes to SAC for residual learning are:
1) delay the start of learning for the policy network.
2) output zero residual actions until policy learning starts.
"""
def __init__(self,
logdir,
env_fn,
policy_fn,
qf_fn,
policy_training_start=10000,
policy_zero_end=100000,
q_reg_weight=0.0001,
action_reg_weight=0.0001,
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.policy_training_start = policy_training_start
self.policy_zero_end = policy_zero_end
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
self.q_reg_weight = q_reg_weight
self.action_reg_weight = action_reg_weight
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))
if self.frame_stack > 1:
eval_env = VecFrameStack(self.env, self.frame_stack)
else:
eval_env = self.env
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 = ReplayBuffer(buffer_size, frame_stack)
self.actor = ResidualSACActor(self.pi, self.learning_starts,
self.policy_zero_end)
self.data_manager = ReplayBufferDataManager(self.buffer, self.env,
self.actor, 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 loss(self, batch):
"""Loss function."""
pi_out = self.pi(batch['obs'], reparameterization_trick=True)
logp = pi_out.dist.log_prob(pi_out.action)
q1 = self.qf1(batch['obs'], batch['action']).value
q2 = self.qf2(batch['obs'], batch['action']).value
# alpha loss
should_update_policy = (self.t >= self.policy_training_start
and self.t % self.policy_update_period == 0)
if self.automatic_entropy_tuning:
if should_update_policy:
ent_error = logp + self.target_entropy
alpha_loss = -(self.log_alpha * ent_error.detach()).mean()
self.opt_alpha.zero_grad()
alpha_loss.backward()
self.opt_alpha.step()
alpha = self.log_alpha.exp()
else:
alpha = self.alpha
alpha_loss = 0
# qf loss
with torch.no_grad():
next_pi_out = self.pi(batch['next_obs'])
next_ac = next_pi_out.action
# Account for the fact that we are learning about the base policy
# before we start updating the residual policy
if self.t < self.policy_training_start:
next_ac = nest.map_structure(torch.zeros_like, next_ac)
next_ac_logp = next_pi_out.dist.log_prob(next_ac)
q1_next = self.target_qf1(batch['next_obs'], next_ac).value
q2_next = self.target_qf2(batch['next_obs'], next_ac).value
qnext = torch.min(q1_next, q2_next) - alpha * next_ac_logp
qtarg = batch['reward'] + (1.0 -
batch['done']) * self.gamma * qnext
assert qtarg.shape == q1.shape
assert qtarg.shape == q2.shape
qf1_loss = self.mse_loss(q1,
qtarg) + self.q_reg_weight * (q1**2).mean()
qf2_loss = self.mse_loss(q2,
qtarg) + self.q_reg_weight * (q2**2).mean()
# pi loss
pi_loss = None
if should_update_policy:
q1_pi = self.qf1(batch['obs'], pi_out.action).value
q2_pi = self.qf2(batch['obs'], pi_out.action).value
min_q_pi = torch.min(q1_pi, q2_pi)
assert min_q_pi.shape == logp.shape
pi_loss = (alpha * logp - min_q_pi).mean()
action_reg = self.action_reg_weight * (pi_out.action**2).mean()
pi_loss = pi_loss + action_reg
# log pi loss about as frequently as other losses
if self.t % self.log_period < self.policy_update_period:
logger.add_scalar('loss/pi', pi_loss, self.t, time.time())
if self.t % self.log_period < self.update_period:
if self.automatic_entropy_tuning:
logger.add_scalar('alg/log_alpha',
self.log_alpha.detach().cpu().numpy(),
self.t, time.time())
scalars = {
"target": self.target_entropy,
"entropy": -torch.mean(logp.detach()).cpu().numpy().item()
}
logger.add_scalars('alg/entropy', scalars, self.t, time.time())
else:
logger.add_scalar(
'alg/entropy',
-torch.mean(logp.detach()).cpu().numpy().item(), self.t,
time.time())
logger.add_scalar('loss/qf1', qf1_loss, self.t, time.time())
logger.add_scalar('loss/qf2', qf2_loss, self.t, time.time())
logger.add_scalar('alg/qf1',
q1.mean().detach().cpu().numpy(), self.t,
time.time())
logger.add_scalar('alg/qf2',
q2.mean().detach().cpu().numpy(), self.t,
time.time())
return pi_loss, qf1_loss, qf2_loss
def step(self):
"""Step optimization."""
self.t += self.data_manager.step_until_update()
if self.t % self.target_update_period == 0:
soft_target_update(self.target_qf1, self.qf1,
self.target_smoothing_coef)
soft_target_update(self.target_qf2, self.qf2,
self.target_smoothing_coef)
if self.t % self.update_period == 0:
batch = self.data_manager.sample(self.batch_size)
pi_loss, qf1_loss, qf2_loss = self.loss(batch)
# update
if pi_loss:
self.opt_pi.zero_grad()
pi_loss.backward()
self.opt_pi.step()
self.opt_qf1.zero_grad()
qf1_loss.backward()
self.opt_qf1.step()
self.opt_qf2.zero_grad()
qf2_loss.backward()
self.opt_qf2.step()
return self.t
def evaluate(self):
"""Evaluate."""
if self.frame_stack > 1:
eval_env = VecFrameStack(self.env, self.frame_stack)
else:
eval_env = self.env
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 save(self):
"""Save."""
state_dict = {
'pi':
self.pi.state_dict(),
'qf1':
self.qf1.state_dict(),
'qf2':
self.qf2.state_dict(),
'target_qf1':
self.target_qf1.state_dict(),
'target_qf2':
self.target_qf2.state_dict(),
'opt_pi':
self.opt_pi.state_dict(),
'opt_qf1':
self.opt_qf1.state_dict(),
'opt_qf2':
self.opt_qf2.state_dict(),
'log_alpha':
(self.log_alpha if self.automatic_entropy_tuning else None),
'opt_alpha': (self.opt_alpha.state_dict()
if self.automatic_entropy_tuning else None),
'env':
misc.env_state_dict(self.env),
'actor':
self.actor.state_dict(),
't':
self.t
}
buffer_dict = self.buffer.state_dict()
state_dict['buffer_format'] = nest.get_structure(buffer_dict)
self.ckptr.save(state_dict, self.t)
# save buffer seperately and only once (because it can be huge)
np.savez(
os.path.join(self.ckptr.ckptdir, 'buffer.npz'),
**{f'{i:04d}': x
for i, x in enumerate(nest.flatten(buffer_dict))})
def load(self, t=None):
"""Load."""
state_dict = self.ckptr.load(t)
if state_dict is None:
self.t = 0
return self.t
self.pi.load_state_dict(state_dict['pi'])
self.qf1.load_state_dict(state_dict['qf1'])
self.qf2.load_state_dict(state_dict['qf2'])
self.target_qf1.load_state_dict(state_dict['target_qf1'])
self.target_qf2.load_state_dict(state_dict['target_qf2'])
self.opt_pi.load_state_dict(state_dict['opt_pi'])
self.opt_qf1.load_state_dict(state_dict['opt_qf1'])
self.opt_qf2.load_state_dict(state_dict['opt_qf2'])
if state_dict['log_alpha']:
with torch.no_grad():
self.log_alpha.copy_(state_dict['log_alpha'])
self.opt_alpha.load_state_dict(state_dict['opt_alpha'])
misc.env_load_state_dict(self.env, state_dict['env'])
self.actor.load_state_dict(state_dict['actor'])
self.t = state_dict['t']
if os.path.exists(os.path.join(self.ckptr.ckptdir, 'buffer.npz')):
buffer_format = state_dict['buffer_format']
buffer_state = dict(
np.load(os.path.join(self.ckptr.ckptdir, 'buffer.npz'),
allow_pickle=True))
buffer_state = nest.flatten(buffer_state)
self.buffer.load_state_dict(
nest.pack_sequence_as(buffer_state, buffer_format))
self.data_manager.manual_reset()
return self.t
def close(self):
"""Close environment."""
try:
self.env.close()
except Exception:
pass