Fetching contributors…
Cannot retrieve contributors at this time
329 lines (268 sloc) 13.4 KB
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from collections import namedtuple
import logging
import tensorflow as tf
# Variable scope in which created variables will be placed under
logger = logging.getLogger(__name__)
class LocalSyncParallelOptimizer(object):
"""Optimizer that runs in parallel across multiple local devices.
LocalSyncParallelOptimizer automatically splits up and loads training data
onto specified local devices (e.g. GPUs) with `load_data()`. During a call
to `optimize()`, the devices compute gradients over slices of the data in
parallel. The gradients are then averaged and applied to the shared
The data loaded is pinned in device memory until the next call to
`load_data`, so you can make multiple passes (possibly in randomized order)
over the same data once loaded.
This is similar to tf.train.SyncReplicasOptimizer, but works within a
single TensorFlow graph, i.e. implements in-graph replicated training:
optimizer: Delegate TensorFlow optimizer object.
devices: List of the names of TensorFlow devices to parallelize over.
input_placeholders: List of input_placeholders for the loss function.
Tensors of these shapes will be passed to build_graph() in order
to define the per-device loss ops.
rnn_inputs: Extra input placeholders for RNN inputs. These will have
shape [BATCH_SIZE // MAX_SEQ_LEN, ...].
max_per_device_batch_size: Number of tuples to optimize over at a time
per device. In each call to `optimize()`,
`len(devices) * per_device_batch_size` tuples of data will be
processed. If this is larger than the total data size, it will be
build_graph: Function that takes the specified inputs and returns a
TF Policy Graph instance.
grad_norm_clipping: None or int stdev to clip grad norms by
def __init__(self,
self.optimizer = optimizer
self.devices = devices
self.max_per_device_batch_size = max_per_device_batch_size
self.loss_inputs = input_placeholders + rnn_inputs
self.build_graph = build_graph
# First initialize the shared loss network
with tf.name_scope(TOWER_SCOPE_NAME):
self._shared_loss = build_graph(self.loss_inputs)
shared_ops = tf.get_collection(
tf.GraphKeys.UPDATE_OPS, scope=tf.get_variable_scope().name)
# Then setup the per-device loss graphs that use the shared weights
self._batch_index = tf.placeholder(tf.int32, name="batch_index")
# Dynamic batch size, which may be shrunk if there isn't enough data
self._per_device_batch_size = tf.placeholder(
tf.int32, name="per_device_batch_size")
self._loaded_per_device_batch_size = max_per_device_batch_size
# When loading RNN input, we dynamically determine the max seq len
self._max_seq_len = tf.placeholder(tf.int32, name="max_seq_len")
self._loaded_max_seq_len = 1
# Split on the CPU in case the data doesn't fit in GPU memory.
with tf.device("/cpu:0"):
data_splits = zip(
*[tf.split(ph, len(devices)) for ph in self.loss_inputs])
self._towers = []
for device, device_placeholders in zip(self.devices, data_splits):
self._setup_device(device, device_placeholders,
avg = average_gradients([t.grads for t in self._towers])
if grad_norm_clipping:
clipped = []
for grad, _ in avg:
clipped, _ = tf.clip_by_global_norm(clipped, grad_norm_clipping)
for i, (grad, var) in enumerate(avg):
avg[i] = (clipped[i], var)
# gather update ops for any batch norm layers. TODO(ekl) here we will
# use all the ops found which won't work for DQN / DDPG, but those
# aren't supported with multi-gpu right now anyways.
self._update_ops = tf.get_collection(
tf.GraphKeys.UPDATE_OPS, scope=tf.get_variable_scope().name)
for op in shared_ops:
self._update_ops.remove(op) # only care about tower update ops
if self._update_ops:
logger.debug("Update ops to run on apply gradient: {}".format(
with tf.control_dependencies(self._update_ops):
self._train_op = self.optimizer.apply_gradients(avg)
def load_data(self, sess, inputs, state_inputs):
"""Bulk loads the specified inputs into device memory.
The shape of the inputs must conform to the shapes of the input
placeholders this optimizer was constructed with.
The data is split equally across all the devices. If the data is not
evenly divisible by the batch size, excess data will be discarded.
sess: TensorFlow session.
inputs: List of arrays matching the input placeholders, of shape
[BATCH_SIZE, ...].
state_inputs: List of RNN input arrays. These arrays have size
The number of tuples loaded per device.
feed_dict = {}
assert len(self.loss_inputs) == len(inputs + state_inputs), \
(self.loss_inputs, inputs, state_inputs)
# Let's suppose we have the following input data, and 2 devices:
# 1 2 3 4 5 6 7 <- state inputs shape
# A A A B B B C C C D D D E E E F F F G G G <- inputs shape
# The data is truncated and split across devices as follows:
# |---| seq len = 3
# |---------------------------------| seq batch size = 6 seqs
# |----------------| per device batch size = 9 tuples
if len(state_inputs) > 0:
smallest_array = state_inputs[0]
seq_len = len(inputs[0]) // len(state_inputs[0])
self._loaded_max_seq_len = seq_len
smallest_array = inputs[0]
self._loaded_max_seq_len = 1
seq_batch_size = (self.max_per_device_batch_size //
self._loaded_max_seq_len * len(self.devices))
if len(smallest_array) < seq_batch_size:
# Dynamically shrink the batch size if insufficient data
seq_batch_size = make_divisible_by(
len(smallest_array), len(self.devices))
if seq_batch_size < len(self.devices):
raise ValueError("Must load at least 1 tuple sequence per device, "
"got only {} total.".format(len(smallest_array)))
self._loaded_per_device_batch_size = (
seq_batch_size // len(self.devices) * self._loaded_max_seq_len)
if len(state_inputs) > 0:
# First truncate the RNN state arrays to the seq_batch_size
state_inputs = [
make_divisible_by(arr, seq_batch_size) for arr in state_inputs
# Then truncate the data inputs to match
inputs = [arr[:len(state_inputs[0]) * seq_len] for arr in inputs]
assert len(state_inputs[0]) * seq_len == len(inputs[0]), \
(len(state_inputs[0]), seq_batch_size, seq_len, len(inputs[0]))
for ph, arr in zip(self.loss_inputs, inputs + state_inputs):
feed_dict[ph] = arr
truncated_len = len(inputs[0])
for ph, arr in zip(self.loss_inputs, inputs + state_inputs):
truncated_arr = make_divisible_by(arr, seq_batch_size)
feed_dict[ph] = truncated_arr
truncated_len = len(truncated_arr)[t.init_op for t in self._towers], feed_dict=feed_dict)
tuples_per_device = truncated_len / len(self.devices)
assert tuples_per_device > 0, "No data loaded?"
assert tuples_per_device % self._loaded_per_device_batch_size == 0
return tuples_per_device
def optimize(self, sess, batch_index):
"""Run a single step of SGD.
Runs a SGD step over a slice of the preloaded batch with size given by
self._loaded_per_device_batch_size and offset given by the batch_index
Updates shared model weights based on the averaged per-device
sess: TensorFlow session.
batch_index: Offset into the preloaded data. This value must be
between `0` and `tuples_per_device`. The amount of data to
process is at most `max_per_device_batch_size`.
The outputs of extra_ops evaluated over the batch.
feed_dict = {
self._batch_index: batch_index,
self._per_device_batch_size: self._loaded_per_device_batch_size,
self._max_seq_len: self._loaded_max_seq_len,
for tower in self._towers:
fetches = {"train": self._train_op}
for tower in self._towers:
return, feed_dict=feed_dict)
def get_common_loss(self):
return self._shared_loss
def get_device_losses(self):
return [t.loss_graph for t in self._towers]
def _setup_device(self, device, device_input_placeholders, num_data_in):
assert num_data_in <= len(device_input_placeholders)
with tf.device(device):
with tf.name_scope(TOWER_SCOPE_NAME):
device_input_batches = []
device_input_slices = []
for i, ph in enumerate(device_input_placeholders):
current_batch = tf.Variable(
if i < num_data_in:
scale = self._max_seq_len
granularity = self._max_seq_len
scale = self._max_seq_len
granularity = 1
current_slice = tf.slice(
([self._batch_index // scale * granularity] +
[0] * len(ph.shape[1:])),
([self._per_device_batch_size // scale * granularity] +
[-1] * len(ph.shape[1:])))
graph_obj = self.build_graph(device_input_slices)
device_grads = graph_obj.gradients(self.optimizer)
return Tower(
*[batch.initializer for batch in device_input_batches]),
device_grads, graph_obj)
# Each tower is a copy of the loss graph pinned to a specific device.
Tower = namedtuple("Tower", ["init_op", "grads", "loss_graph"])
def make_divisible_by(a, n):
if type(a) is int:
return a - a % n
return a[0:a.shape[0] - a.shape[0] % n]
def average_gradients(tower_grads):
"""Averages gradients across towers.
Calculate the average gradient for each shared variable across all towers.
Note that this function provides a synchronization point across all towers.
tower_grads: List of lists of (gradient, variable) tuples. The outer
list is over individual gradients. The inner list is over the
gradient calculation for each tower.
List of pairs of (gradient, variable) where the gradient has been
averaged across all towers.
TODO(ekl): We could use NCCL if this becomes a bottleneck.
average_grads = []
for grad_and_vars in zip(*tower_grads):
# Note that each grad_and_vars looks like the following:
# ((grad0_gpu0, var0_gpu0), ... , (grad0_gpuN, var0_gpuN))
grads = []
for g, _ in grad_and_vars:
if g is not None:
# Add 0 dimension to the gradients to represent the tower.
expanded_g = tf.expand_dims(g, 0)
# Append on a 'tower' dimension which we will average over
# below.
if not grads:
# Average over the 'tower' dimension.
grad = tf.concat(axis=0, values=grads)
grad = tf.reduce_mean(grad, 0)
# Keep in mind that the Variables are redundant because they are shared
# across towers. So .. we will just return the first tower's pointer to
# the Variable.
v = grad_and_vars[0][1]
grad_and_var = (grad, v)
return average_grads