def _init(self,
learning_starts=1000,
buffer_size=10000,
prioritized_replay=True,
prioritized_replay_alpha=0.6,
prioritized_replay_beta=0.4,
prioritized_replay_eps=1e-6,
train_batch_size=32,
sample_batch_size=4):
self.replay_starts = learning_starts
self.prioritized_replay_beta = prioritized_replay_beta
self.prioritized_replay_eps = prioritized_replay_eps
self.train_batch_size = train_batch_size
# Stats
self.update_weights_timer = TimerStat()
self.sample_timer = TimerStat()
self.replay_timer = TimerStat()
self.grad_timer = TimerStat()
self.throughput = RunningStat()
# Set up replay buffer
if prioritized_replay:
self.replay_buffer = PrioritizedReplayBuffer(
buffer_size, alpha=prioritized_replay_alpha)
else:
self.replay_buffer = ReplayBuffer(buffer_size)
assert buffer_size >= self.replay_starts
def _init(self, num_sgd_iter=1, timesteps_per_batch=1):
self.update_weights_timer = TimerStat()
self.sample_timer = TimerStat()
self.grad_timer = TimerStat()
self.throughput = RunningStat()
self.num_sgd_iter = num_sgd_iter
self.timesteps_per_batch = timesteps_per_batch
def __init__(self,
workers,
num_sgd_iter=1,
train_batch_size=1,
sgd_minibatch_size=0,
standardize_fields=frozenset([]),
aux_loss_every_k=16,
aux_loss_num_sgd_iter=9,
aux_loss_start_after_num_steps=0):
PolicyOptimizer.__init__(self, workers)
self.update_weights_timer = TimerStat()
self.standardize_fields = standardize_fields
self.sample_timer = TimerStat()
self.grad_timer = TimerStat()
self.throughput = RunningStat()
self.num_sgd_iter = num_sgd_iter
self.sgd_minibatch_size = sgd_minibatch_size
self.train_batch_size = train_batch_size
self.learner_stats = {}
self.policies = dict(
self.workers.local_worker().foreach_trainable_policy(lambda p, i:
(i, p)))
logger.debug("Policies to train: {}".format(self.policies))
self.aux_loss_every_k = aux_loss_every_k
self.aux_loss_num_sgd_iter = aux_loss_num_sgd_iter
self.aux_loss_start_after_num_steps = aux_loss_start_after_num_steps
self.memory = []
# Assert that train batch size is divisible by sgd minibatch size to make populating
# policy logits simpler.
assert train_batch_size % sgd_minibatch_size == 0, (
f"train_batch_size: {train_batch_size}"
f"sgd_minibatch_size: {sgd_minibatch_size}")
def _init(self, num_sgd_iter=1, train_batch_size=1):
self.update_weights_timer = TimerStat()
self.sample_timer = TimerStat()
self.grad_timer = TimerStat()
self.throughput = RunningStat()
self.num_sgd_iter = num_sgd_iter
self.train_batch_size = train_batch_size
self.learner_stats = {}
def testCombiningStat(self):
for shape in [(), (3, ), (3, 4)]:
li = []
rs1 = RunningStat(shape)
rs2 = RunningStat(shape)
rs = RunningStat(shape)
for _ in range(5):
val = np.random.randn(*shape)
rs1.push(val)
rs.push(val)
li.append(val)
for _ in range(9):
rs2.push(val)
rs.push(val)
li.append(val)
rs1.update(rs2)
assert np.allclose(rs.mean, rs1.mean)
assert np.allclose(rs.std, rs1.std)
def __init__(self, workers, num_sgd_iter=1, train_batch_size=1):
PolicyOptimizer.__init__(self, workers)
self.update_weights_timer = TimerStat()
self.sample_timer = TimerStat()
self.grad_timer = TimerStat()
self.throughput = RunningStat()
self.num_sgd_iter = num_sgd_iter
self.train_batch_size = train_batch_size
self.learner_stats = {}
def testRunningStat(self):
for shp in ((), (3,), (3, 4)):
li = []
rs = RunningStat(shp)
for _ in range(5):
val = np.random.randn(*shp)
rs.push(val)
li.append(val)
m = np.mean(li, axis=0)
self.assertTrue(np.allclose(rs.mean, m))
v = np.square(m) if (len(li) == 1) else np.var(li, ddof=1, axis=0)
self.assertTrue(np.allclose(rs.var, v))
def _init(self,
learning_starts=1000,
buffer_size=10000,
prioritized_replay=True,
prioritized_replay_alpha=0.6,
prioritized_replay_beta=0.4,
schedule_max_timesteps=100000,
beta_annealing_fraction=0.2,
final_prioritized_replay_beta=0.4,
prioritized_replay_eps=1e-6,
train_batch_size=32,
sample_batch_size=4):
self.replay_starts = learning_starts
# linearly annealing beta used in Rainbow paper
self.prioritized_replay_beta = LinearSchedule(
schedule_timesteps=int(schedule_max_timesteps *
beta_annealing_fraction),
initial_p=prioritized_replay_beta,
final_p=final_prioritized_replay_beta)
self.prioritized_replay_eps = prioritized_replay_eps
self.train_batch_size = train_batch_size
# Stats
self.update_weights_timer = TimerStat()
self.sample_timer = TimerStat()
self.replay_timer = TimerStat()
self.grad_timer = TimerStat()
self.throughput = RunningStat()
self.learner_stats = {}
# Set up replay buffer
if prioritized_replay:
def new_buffer():
return PrioritizedReplayBuffer(buffer_size,
alpha=prioritized_replay_alpha)
else:
def new_buffer():
return ReplayBuffer(buffer_size)
self.replay_buffers = collections.defaultdict(new_buffer)
assert buffer_size >= self.replay_starts
def __init__(self, workers, train_batch_size=10000, microbatch_size=1000):
PolicyOptimizer.__init__(self, workers)
if train_batch_size <= microbatch_size:
raise ValueError(
"The microbatch size must be smaller than the train batch "
"size, got {} vs {}".format(microbatch_size, train_batch_size))
self.update_weights_timer = TimerStat()
self.sample_timer = TimerStat()
self.grad_timer = TimerStat()
self.throughput = RunningStat()
self.train_batch_size = train_batch_size
self.microbatch_size = microbatch_size
self.learner_stats = {}
self.policies = dict(
self.workers.local_worker().foreach_trainable_policy(lambda p, i:
(i, p)))
logger.debug("Policies to train: {}".format(self.policies))
def __init__(self,
workers,
num_sgd_iter=1,
train_batch_size=1,
sgd_minibatch_size=0,
standardize_fields=frozenset([])):
PolicyOptimizer.__init__(self, workers)
self.update_weights_timer = TimerStat()
self.standardize_fields = standardize_fields
self.sample_timer = TimerStat()
self.grad_timer = TimerStat()
self.throughput = RunningStat()
self.num_sgd_iter = num_sgd_iter
self.sgd_minibatch_size = sgd_minibatch_size
self.train_batch_size = train_batch_size
self.learner_stats = {}
self.policies = dict(self.workers.local_worker()
.foreach_trainable_policy(lambda p, i: (i, p)))
logger.debug("Policies to train: {}".format(self.policies))
def _init(self,
learning_starts=1000,
buffer_size=10000,
prioritized_replay=True,
prioritized_replay_alpha=0.6,
prioritized_replay_beta=0.4,
prioritized_replay_eps=1e-6,
train_batch_size=32,
sample_batch_size=4,
clip_rewards=True):
self.replay_starts = learning_starts
self.prioritized_replay_beta = prioritized_replay_beta
self.prioritized_replay_eps = prioritized_replay_eps
self.train_batch_size = train_batch_size
# Stats
self.update_weights_timer = TimerStat()
self.sample_timer = TimerStat()
self.replay_timer = TimerStat()
self.grad_timer = TimerStat()
self.throughput = RunningStat()
# Set up replay buffer
if prioritized_replay:
def new_buffer():
return PrioritizedReplayBuffer(
buffer_size,
alpha=prioritized_replay_alpha,
clip_rewards=clip_rewards)
else:
def new_buffer():
return ReplayBuffer(buffer_size, clip_rewards)
self.replay_buffers = collections.defaultdict(new_buffer)
assert buffer_size >= self.replay_starts
def _init(self, batch_size=32):
self.update_weights_timer = TimerStat()
self.sample_timer = TimerStat()
self.grad_timer = TimerStat()
self.throughput = RunningStat()
self.batch_size = batch_size
def _init(self):
self.update_weights_timer = TimerStat()
self.sample_timer = TimerStat()
self.grad_timer = TimerStat()
self.throughput = RunningStat()
