class TFMultiGPULearner(LearnerThread):
"""Learner that can use multiple GPUs and parallel loading.
This is for use with AsyncSamplesOptimizer.
"""
def __init__(self,
local_worker,
num_gpus=1,
lr=0.0005,
train_batch_size=500,
num_data_loader_buffers=1,
minibatch_buffer_size=1,
num_sgd_iter=1,
learner_queue_size=16,
learner_queue_timeout=300,
num_data_load_threads=16,
_fake_gpus=False):
"""Initialize a multi-gpu learner thread.
Arguments:
local_worker (RolloutWorker): process local rollout worker holding
policies this thread will call learn_on_batch() on
num_gpus (int): number of GPUs to use for data-parallel SGD
lr (float): learning rate
train_batch_size (int): size of batches to learn on
num_data_loader_buffers (int): number of buffers to load data into
in parallel. Each buffer is of size of train_batch_size and
increases GPU memory usage proportionally.
minibatch_buffer_size (int): max number of train batches to store
in the minibatching buffer
num_sgd_iter (int): number of passes to learn on per train batch
learner_queue_size (int): max size of queue of inbound
train batches to this thread
num_data_loader_threads (int): number of threads to use to load
data into GPU memory in parallel
"""
LearnerThread.__init__(self, local_worker, minibatch_buffer_size,
num_sgd_iter, learner_queue_size,
learner_queue_timeout)
self.lr = lr
self.train_batch_size = train_batch_size
if not num_gpus:
self.devices = ["/cpu:0"]
elif _fake_gpus:
self.devices = [
"/cpu:{}".format(i) for i in range(int(math.ceil(num_gpus)))
]
else:
self.devices = [
"/gpu:{}".format(i) for i in range(int(math.ceil(num_gpus)))
]
logger.info("TFMultiGPULearner devices {}".format(self.devices))
assert self.train_batch_size % len(self.devices) == 0
assert self.train_batch_size >= len(self.devices), "batch too small"
if set(self.local_worker.policy_map.keys()) != {DEFAULT_POLICY_ID}:
raise NotImplementedError("Multi-gpu mode for multi-agent")
self.policy = self.local_worker.policy_map[DEFAULT_POLICY_ID]
# per-GPU graph copies created below must share vars with the policy
# reuse is set to AUTO_REUSE because Adam nodes are created after
# all of the device copies are created.
self.par_opt = []
with self.local_worker.tf_sess.graph.as_default():
with self.local_worker.tf_sess.as_default():
with tf.variable_scope(DEFAULT_POLICY_ID, reuse=tf.AUTO_REUSE):
if self.policy._state_inputs:
rnn_inputs = self.policy._state_inputs + [
self.policy._seq_lens
]
else:
rnn_inputs = []
adam = tf.train.AdamOptimizer(self.lr)
for _ in range(num_data_loader_buffers):
self.par_opt.append(
LocalSyncParallelOptimizer(
adam,
self.devices,
[v for _, v in self.policy._loss_inputs],
rnn_inputs,
999999, # it will get rounded down
self.policy.copy))
self.sess = self.local_worker.tf_sess
self.sess.run(tf.global_variables_initializer())
self.idle_optimizers = queue.Queue()
self.ready_optimizers = queue.Queue()
for opt in self.par_opt:
self.idle_optimizers.put(opt)
for i in range(num_data_load_threads):
self.loader_thread = _LoaderThread(self, share_stats=(i == 0))
self.loader_thread.start()
self.minibatch_buffer = MinibatchBuffer(
self.ready_optimizers, minibatch_buffer_size,
learner_queue_timeout, num_sgd_iter)
@override(LearnerThread)
def step(self):
assert self.loader_thread.is_alive()
with self.load_wait_timer:
opt, released = self.minibatch_buffer.get()
with self.grad_timer:
fetches = opt.optimize(self.sess, 0)
self.weights_updated = True
self.stats = get_learner_stats(fetches)
if released:
self.idle_optimizers.put(opt)
self.outqueue.put(opt.num_tuples_loaded)
self.learner_queue_size.push(self.inqueue.qsize())
class TFMultiGPULearner(LearnerThread):
"""Learner that can use multiple GPUs and parallel loading.
This is for use with AsyncSamplesOptimizer.
"""
def __init__(self,
local_evaluator,
num_gpus=1,
lr=0.0005,
train_batch_size=500,
num_data_loader_buffers=1,
minibatch_buffer_size=1,
num_sgd_iter=1,
learner_queue_size=16,
num_data_load_threads=16,
_fake_gpus=False):
LearnerThread.__init__(self, local_evaluator, minibatch_buffer_size,
num_sgd_iter, learner_queue_size)
self.lr = lr
self.train_batch_size = train_batch_size
if not num_gpus:
self.devices = ["/cpu:0"]
elif _fake_gpus:
self.devices = [
"/cpu:{}".format(i) for i in range(int(math.ceil(num_gpus)))
]
else:
self.devices = [
"/gpu:{}".format(i) for i in range(int(math.ceil(num_gpus)))
]
logger.info("TFMultiGPULearner devices {}".format(self.devices))
assert self.train_batch_size % len(self.devices) == 0
assert self.train_batch_size >= len(self.devices), "batch too small"
if set(self.local_evaluator.policy_map.keys()) != {DEFAULT_POLICY_ID}:
raise NotImplementedError("Multi-gpu mode for multi-agent")
self.policy = self.local_evaluator.policy_map[DEFAULT_POLICY_ID]
# per-GPU graph copies created below must share vars with the policy
# reuse is set to AUTO_REUSE because Adam nodes are created after
# all of the device copies are created.
self.par_opt = []
with self.local_evaluator.tf_sess.graph.as_default():
with self.local_evaluator.tf_sess.as_default():
with tf.variable_scope(DEFAULT_POLICY_ID, reuse=tf.AUTO_REUSE):
if self.policy._state_inputs:
rnn_inputs = self.policy._state_inputs + [
self.policy._seq_lens
]
else:
rnn_inputs = []
adam = tf.train.AdamOptimizer(self.lr)
for _ in range(num_data_loader_buffers):
self.par_opt.append(
LocalSyncParallelOptimizer(
adam,
self.devices,
[v for _, v in self.policy._loss_inputs],
rnn_inputs,
999999, # it will get rounded down
self.policy.copy))
self.sess = self.local_evaluator.tf_sess
self.sess.run(tf.global_variables_initializer())
self.idle_optimizers = queue.Queue()
self.ready_optimizers = queue.Queue()
for opt in self.par_opt:
self.idle_optimizers.put(opt)
for i in range(num_data_load_threads):
self.loader_thread = _LoaderThread(self, share_stats=(i == 0))
self.loader_thread.start()
self.minibatch_buffer = MinibatchBuffer(self.ready_optimizers,
minibatch_buffer_size,
num_sgd_iter)
@override(LearnerThread)
def step(self):
assert self.loader_thread.is_alive()
with self.load_wait_timer:
opt, released = self.minibatch_buffer.get()
with self.grad_timer:
fetches = opt.optimize(self.sess, 0)
self.weights_updated = True
self.stats = get_learner_stats(fetches)
if released:
self.idle_optimizers.put(opt)
self.outqueue.put(opt.num_tuples_loaded)
self.learner_queue_size.push(self.inqueue.qsize())
class LearnerThread(threading.Thread):
"""Background thread that updates the local model from sample trajectories.
This is for use with AsyncSamplesOptimizer.
The learner thread communicates with the main thread through Queues. This
is needed since Ray operations can only be run on the main thread. In
addition, moving heavyweight gradient ops session runs off the main thread
improves overall throughput.
"""
def __init__(self, local_worker, minibatch_buffer_size, num_sgd_iter,
learner_queue_size, learner_queue_timeout):
"""Initialize the learner thread.
Arguments:
local_worker (RolloutWorker): process local rollout worker holding
policies this thread will call learn_on_batch() on
minibatch_buffer_size (int): max number of train batches to store
in the minibatching buffer
num_sgd_iter (int): number of passes to learn on per train batch
learner_queue_size (int): max size of queue of inbound
train batches to this thread
learner_queue_timeout (int): raise an exception if the queue has
been empty for this long in seconds
"""
threading.Thread.__init__(self)
self.learner_queue_size = WindowStat("size", 50)
self.local_worker = local_worker
self.inqueue = queue.Queue(maxsize=learner_queue_size)
self.outqueue = queue.Queue()
self.minibatch_buffer = MinibatchBuffer(inqueue=self.inqueue,
size=minibatch_buffer_size,
timeout=learner_queue_timeout,
num_passes=num_sgd_iter,
init_num_passes=num_sgd_iter)
self.queue_timer = TimerStat()
self.grad_timer = TimerStat()
self.load_timer = TimerStat()
self.load_wait_timer = TimerStat()
self.daemon = True
self.weights_updated = False
self.stats = {}
self.stopped = False
self.num_steps = 0
def run(self):
while not self.stopped:
self.step()
def step(self):
with self.queue_timer:
batch, _ = self.minibatch_buffer.get()
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.learner_queue_size.push(self.inqueue.qsize())