def step(self) -> Optional[_NextValueNotReady]:
with self.queue_timer:
try:
batch, _ = self.minibatch_buffer.get()
except queue.Empty:
return _NextValueNotReady()
with self.grad_timer:
# Use LearnerInfoBuilder as a unified way to build the final
# results dict from `learn_on_loaded_batch` call(s).
# This makes sure results dicts always have the same structure
# no matter the setup (multi-GPU, multi-agent, minibatch SGD,
# tf vs torch).
learner_info_builder = LearnerInfoBuilder(num_devices=1)
multi_agent_results = self.local_worker.learn_on_batch(batch)
for pid, results in multi_agent_results.items():
learner_info_builder.add_learn_on_batch_results(results, pid)
self.learner_info = learner_info_builder.finalize()
learner_stats = {
pid: info[LEARNER_STATS_KEY]
for pid, info in self.learner_info.items()
}
self.weights_updated = True
self.num_steps += 1
self.outqueue.put((batch.count, learner_stats))
self.learner_queue_size.push(self.inqueue.qsize())
def gen_replay(_):
while True:
item = local_buffer.replay()
if item is None:
yield _NextValueNotReady()
else:
yield item
def gen_replay(_):
while True:
item = local_buffer.sample(num_items_to_replay)
if item is None:
yield _NextValueNotReady()
else:
yield item
def gen_replay(_):
while True:
item = ray.get(replay_buffer_actor.replay.remote())
if item is None:
yield _NextValueNotReady()
else:
yield item
def base_iterator(timeout=None):
while check():
try:
item = input_queue.get_nowait()
yield item
except queue.Empty:
yield _NextValueNotReady()
raise RuntimeError("Error raised reading from queue")
def base_iterator(timeout=None):
while check():
try:
item = input_queue.get(timeout=0.001)
yield item
except queue.Empty:
yield _NextValueNotReady()
raise RuntimeError("Dequeue `check()` returned False! "
"Exiting with Exception from Dequeue iterator.")
def base_iterator(timeout=None):
while check():
try:
item = input_queue.get_nowait()
yield item
except queue.Empty:
yield _NextValueNotReady()
# Sleep to give time for any error to be printed
time.sleep(3)
raise RuntimeError("Error raised reading from queue")
def step(self) -> Optional[_NextValueNotReady]:
with self.queue_timer:
try:
batch, _ = self.minibatch_buffer.get()
except queue.Empty:
return _NextValueNotReady()
with self.grad_timer:
fetches = self.local_worker.learn_on_batch(batch)
self.weights_updated = True
self.stats = get_learner_stats(fetches)
self.num_steps += 1
self.outqueue.put((batch.count, self.stats))
self.learner_queue_size.push(self.inqueue.qsize())
def gen_replay(timeout):
while True:
samples = {}
idxes = None
for policy_id, replay_buffer in replay_buffers.buffers.items():
policy_multiplier = 2 if policy_id == 'policy_team' else 1
if len(replay_buffer) >= min_size_to_learn*policy_multiplier and \
replay_buffers.steps[policy_id] >= learn_every:
idxes = replay_buffer.sample_idxes(train_batch_size)
replay_buffers.steps[policy_id] = 0
if prioritized:
(obses_t, actions, rewards, obses_tp1, dones, w, ind) \
= replay_buffer.sample_with_idxes(idxes, beta)
else:
(obses_t, actions, rewards, obses_tp1,
dones) = replay_buffer.sample_with_idxes(idxes)
weights = np.ones_like(rewards)
batch_indexes = -np.ones_like(rewards)
samples[policy_id] = SampleBatch({
"obs":
obses_t,
"actions":
actions,
"rewards":
rewards,
"new_obs":
obses_tp1,
"dones":
dones,
"weights":
weights,
"batch_indexes":
batch_indexes
})
if samples == {}:
yield _NextValueNotReady()
else:
yield MultiAgentBatch(samples, train_batch_size)
def gen_replay(timeout):
while True:
samples = {}
idxes = None
for policy_id, reservoir_buffer in reservoir_buffers.buffers.items(
):
if len(reservoir_buffer) >= min_size_to_learn and \
reservoir_buffers.steps[policy_id] >= learn_every:
# idxes = reservoir_buffer.sample_idxes(train_batch_size)
(obses_t,
actions) = reservoir_buffer.sample(train_batch_size)
samples[policy_id] = SampleBatch({
"obs": obses_t,
"actions": actions,
})
reservoir_buffers.steps[policy_id] = 0
if samples == {}:
yield _NextValueNotReady()
else:
yield MultiAgentBatch(samples, train_batch_size)
def gen_replay(timeout):
while True:
samples = {}
for policy_id, replay_buffer in replay_buffers.buffers.items():
if len(replay_buffer.replay_batches) >= min_size_to_learn and \
replay_buffers.steps[policy_id] >= learn_every:
batch = None
for x in range(train_batch_size):
if batch is None:
batch = replay_buffer.replay(
).decompress_if_needed()
else:
batch = batch.concat(
replay_buffer.replay().decompress_if_needed())
replay_buffers.steps[policy_id] = 0
samples[policy_id] = batch
if samples == {}:
yield _NextValueNotReady()
else:
yield MultiAgentBatch(samples, train_batch_size)
def __call__(self, x):
try:
self.queue.put_nowait(x)
except queue.Full:
return _NextValueNotReady()
def __call__(self, x: Any) -> Any:
try:
self.queue.put(x, timeout=0.001)
except queue.Full:
return _NextValueNotReady()
return x
