『TensorFlow』分布式训练_其二_单机多GPU并行&GPU模式设定

建议比对『MXNet』第七弹_多GPU并行程序设计

一、tensorflow GPU设置

GPU指定占用

gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.7)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))   

上面分配给tensorflow的GPU显存大小为：GPU实际显存*0.7。

GPU模式禁用

import os 
os.environ["CUDA_VISIBLE_DEVICES"]="-1"  

GPU资源申请规则

# 设置 GPU 按需增长
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)

二、单机多GPU工作原理

以一篇csdn博客（出处见水印）上的图说明多GPU工作原理：

想让 TensorFlow 在多个 GPU 上运行, 需要建立 multi-tower 结构, 在这个结构里每个 tower 分别被指配给不同的 GPU 运行，汇总工作一般交由CPU完成，示意如下，

# 新建一个 graph.
c = []
for d in ['/gpu:2', '/gpu:3']:
  with tf.device(d):
    a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3])
    b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2])
    c.append(tf.matmul(a, b))
with tf.device('/cpu:0'):
  sum = tf.add_n(c)
# 新建session with log_device_placement并设置为True.
sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
# 运行这个op.
print sess.run(sum) 

三、官方demo

多GPU分布+LR衰减+滑动平均

和MXNet不同，由于TensorFlow使用上下文指定设备，所以数据无需显示的拷贝到指定设备，在目标设备上下文中获取即可（需要调用对应节点于该设备下，如下文中的出队节点）

另一个值得注意的点在于收集来的梯度格式为List of lists of (gradient, variable) tuples，我们计算后返回的是List of (gradient, variable) tuples，variable随便指定一组gpu上的即可，这是因为和MXNet不同，MXNet是得到grad平均值后分发给各个GPU各自更新，TensorFlow实际是各个GPU使用同一套参数（tf.get_variable_scope().reuse_variables()），虽然会被拷贝到各个设备，但是彼此之间是有逻辑关系的，是共享参数，简化示意如下：

#将神经网络的优化过程跑在不同的GPU上
for i in range(N_GPU)：
    with tf.debice('/gpu:%d'%i)
        with tf.name_scope('GPU_%d'%i) as scope:
            cur_loss = get_loss(x,y_regularizer,scope)
            #tf.get_variable的命名空间
            tf.get_variable_scope().reuse_variables()
            #使用当前gpu计算所有变量的梯度
            grads= opt.compute_gradients(cur_loss)
            tower_grads.append(grads)
#计算变量的平均梯度
grads = average_gradients(tower_grads)
#使用平均梯度更新参数
apply_gradient_op = opt.apply_gradients(grads,global_step = global)

models/tutorials/image/cifar10/cifer10_multi_gpu-train.py

# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
 
"""A binary to train CIFAR-10 using multiple GPUs with synchronous updates.
Accuracy:
cifar10_multi_gpu_train.py achieves ~86% accuracy after 100K steps (256
epochs of data) as judged by cifar10_eval.py.
Speed: With batch_size 128.
System        | Step Time (sec/batch)  |     Accuracy
--------------------------------------------------------------------
1 Tesla K20m  | 0.35-0.60              | ~86% at 60K steps  (5 hours)
1 Tesla K40m  | 0.25-0.35              | ~86% at 100K steps (4 hours)
2 Tesla K20m  | 0.13-0.20              | ~84% at 30K steps  (2.5 hours)
3 Tesla K20m  | 0.13-0.18              | ~84% at 30K steps
4 Tesla K20m  | ~0.10                  | ~84% at 30K steps
Usage:
Please see the tutorial and website for how to download the CIFAR-10
data set, compile the program and train the model.
http://tensorflow.org/tutorials/deep_cnn/
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
 
from datetime import datetime
import os.path
import re
import time
 
import numpy as np
from six.moves import xrange  # pylint: disable=redefined-builtin
import tensorflow as tf
import cifar10
 
FLAGS = tf.app.flags.FLAGS
 
tf.app.flags.DEFINE_string('train_dir', '/tmp/cifar10_train',
                           """Directory where to write event logs """
                           """and checkpoint.""")
tf.app.flags.DEFINE_integer('max_steps', 1000000,
                            """Number of batches to run.""")
tf.app.flags.DEFINE_integer('num_gpus', 1,
                            """How many GPUs to use.""")
tf.app.flags.DEFINE_boolean('log_device_placement', False,
                            """Whether to log device placement.""")
 
 
def tower_loss(scope, images, labels):
  """Calculate the total loss on a single tower running the CIFAR model.
  Args:
    scope: unique prefix string identifying the CIFAR tower, e.g. 'tower_0'
    images: Images. 4D tensor of shape [batch_size, height, width, 3].
    labels: Labels. 1D tensor of shape [batch_size].
  Returns:
     Tensor of shape [] containing the total loss for a batch of data
  """
 
  # Build inference Graph.
  logits = cifar10.inference(images)
 
  # Build the portion of the Graph calculating the losses. Note that we will
  # assemble the total_loss using a custom function below.
  _ = cifar10.loss(logits, labels)
 
  # Assemble all of the losses for the current tower only.
  losses = tf.get_collection('losses', scope)
 
  # Calculate the total loss for the current tower.
  total_loss = tf.add_n(losses, name='total_loss')
 
  # Attach a scalar summary to all individual losses and the total loss; do the
  # same for the averaged version of the losses.
  for l in losses + [total_loss]:
    # Remove 'tower_[0-9]/' from the name in case this is a multi-GPU training
    # session. This helps the clarity of presentation on tensorboard.
    loss_name = re.sub('%s_[0-9]*/' % cifar10.TOWER_NAME, '', l.op.name)
    tf.summary.scalar(loss_name, l)
 
  return total_loss
 
 
def average_gradients(tower_grads):
  """Calculate the average gradient for each shared variable across all towers.
  Note that this function provides a synchronization point across all towers.
  Args:
    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.
  Returns:
     List of pairs of (gradient, variable) where the gradient has been averaged
     across all towers.
  """
  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:
      # 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.
      grads.append(expanded_g)
 
    # 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)
    average_grads.append(grad_and_var)
  return average_grads
 
 
def train():
  """Train CIFAR-10 for a number of steps."""
  with tf.Graph().as_default(), tf.device('/cpu:0'):
    # Create a variable to count the number of train() calls. This equals the
    # number of batches processed * FLAGS.num_gpus.
    global_step = tf.get_variable(
        'global_step', [],
        initializer=tf.constant_initializer(0), trainable=False)
 
    # Calculate the learning rate schedule.
    num_batches_per_epoch = (cifar10.NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN /
                             FLAGS.batch_size)
    decay_steps = int(num_batches_per_epoch * cifar10.NUM_EPOCHS_PER_DECAY)
 
    # Decay the learning rate exponentially based on the number of steps.
    lr = tf.train.exponential_decay(cifar10.INITIAL_LEARNING_RATE,
                                    global_step,
                                    decay_steps,
                                    cifar10.LEARNING_RATE_DECAY_FACTOR,
                                    staircase=True)
 
    # Create an optimizer that performs gradient descent.
    opt = tf.train.GradientDescentOptimizer(lr)
 
    # Get images and labels for CIFAR-10.
    images, labels = cifar10.distorted_inputs()
    batch_queue = tf.contrib.slim.prefetch_queue.prefetch_queue(
          [images, labels], capacity=2 * FLAGS.num_gpus)
    # Calculate the gradients for each model tower.
    tower_grads = []
    with tf.variable_scope(tf.get_variable_scope()):
      for i in xrange(FLAGS.num_gpus):
        with tf.device('/gpu:%d' % i):
          with tf.name_scope('%s_%d' % (cifar10.TOWER_NAME, i)) as scope:
            # Dequeues one batch for the GPU
            image_batch, label_batch = batch_queue.dequeue()
            # Calculate the loss for one tower of the CIFAR model. This function
            # constructs the entire CIFAR model but shares the variables across
            # all towers.
            loss = tower_loss(scope, image_batch, label_batch)
 
            # Reuse variables for the next tower.
            tf.get_variable_scope().reuse_variables()
 
            # Retain the summaries from the final tower.
            summaries = tf.get_collection(tf.GraphKeys.SUMMARIES, scope)
 
            # Calculate the gradients for the batch of data on this CIFAR tower.
            grads = opt.compute_gradients(loss)
 
            # Keep track of the gradients across all towers.
            tower_grads.append(grads)
 
    # We must calculate the mean of each gradient. Note that this is the
    # synchronization point across all towers.
    grads = average_gradients(tower_grads)
 
    # Add a summary to track the learning rate.
    summaries.append(tf.summary.scalar('learning_rate', lr))
 
    # Add histograms for gradients.
    for grad, var in grads:
      if grad is not None:
        summaries.append(tf.summary.histogram(var.op.name + '/gradients', grad))
 
    # Apply the gradients to adjust the shared variables.
    apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)
 
    # Add histograms for trainable variables.
    for var in tf.trainable_variables():
      summaries.append(tf.summary.histogram(var.op.name, var))
 
    # Track the moving averages of all trainable variables.
    variable_averages = tf.train.ExponentialMovingAverage(
        cifar10.MOVING_AVERAGE_DECAY, global_step)
    variables_averages_op = variable_averages.apply(tf.trainable_variables())
 
    # Group all updates to into a single train op.
    train_op = tf.group(apply_gradient_op, variables_averages_op)
 
    # Create a saver.
    saver = tf.train.Saver(tf.global_variables())
 
    # Build the summary operation from the last tower summaries.
    summary_op = tf.summary.merge(summaries)
################################################################################
    # Build an initialization operation to run below.
    init = tf.global_variables_initializer()
 
    # Start running operations on the Graph. allow_soft_placement must be set to
    # True to build towers on GPU, as some of the ops do not have GPU
    # implementations.
    sess = tf.Session(config=tf.ConfigProto(
        allow_soft_placement=True,
        log_device_placement=FLAGS.log_device_placement))
    sess.run(init)
 
    # Start the queue runners.
    tf.train.start_queue_runners(sess=sess)
 
    summary_writer = tf.summary.FileWriter(FLAGS.train_dir, sess.graph)
 
    for step in xrange(FLAGS.max_steps):
      start_time = time.time()
      _, loss_value = sess.run([train_op, loss])
      duration = time.time() - start_time
 
      assert not np.isnan(loss_value), 'Model diverged with loss = NaN'
 
      if step % 10 == 0:
        num_examples_per_step = FLAGS.batch_size * FLAGS.num_gpus
        examples_per_sec = num_examples_per_step / duration
        sec_per_batch = duration / FLAGS.num_gpus
 
        format_str = ('%s: step %d, loss = %.2f (%.1f examples/sec; %.3f '
                      'sec/batch)')
        print (format_str % (datetime.now(), step, loss_value,
                             examples_per_sec, sec_per_batch))
 
      if step % 100 == 0:
        summary_str = sess.run(summary_op)
        summary_writer.add_summary(summary_str, step)
 
      # Save the model checkpoint periodically.
      if step % 1000 == 0 or (step + 1) == FLAGS.max_steps:
        checkpoint_path = os.path.join(FLAGS.train_dir, 'model.ckpt')
        saver.save(sess, checkpoint_path, global_step=step)
 
 
def main(argv=None):  # pylint: disable=unused-argument
  cifar10.maybe_download_and_extract()
  if tf.gfile.Exists(FLAGS.train_dir):
    tf.gfile.DeleteRecursively(FLAGS.train_dir)
  tf.gfile.MakeDirs(FLAGS.train_dir)
  train()
 
 
if __name__ == '__main__':
  tf.app.run()

数据输入函数如下，

def distorted_inputs():
  """Construct distorted input for CIFAR training using the Reader ops.
  Returns:
    images: Images. 4D tensor of [batch_size, IMAGE_SIZE, IMAGE_SIZE, 3] size.
    labels: Labels. 1D tensor of [batch_size] size.
  Raises:
    ValueError: If no data_dir
  """
  if not FLAGS.data_dir:
    raise ValueError('Please supply a data_dir')
  data_dir = os.path.join(FLAGS.data_dir, 'cifar-10-batches-bin')
  images, labels = cifar10_input.distorted_inputs(data_dir=data_dir,
                                                  batch_size=FLAGS.batch_size)
  if FLAGS.use_fp16:
    images = tf.cast(images, tf.float16)
    labels = tf.cast(labels, tf.float16)
return images, labels

tf.contrib.slim.prefetch_queue.prefetch_queue从介绍来看就是个输入数据队列

Signature: tf.contrib.slim.prefetch_queue.prefetch_queue(tensors, capacity=8, num_threads=1, dynamic_pad=False, shared_name=None, name=None)
Docstring:
Creates a queue to prefetech tensors from `tensors`.

A queue runner for enqueing tensors into the prefetch_queue is automatically
added to the TF QueueRunners collection.

Example:
This is for example useful to pre-assemble input batches read with
`tf.train.batch()` and enqueue the pre-assembled batches.  Ops that dequeue
from the pre-assembled queue will not pay the cost of assembling the batch.

images, labels = tf.train.batch([image, label], batch_size=32, num_threads=4)
batch_queue = prefetch_queue([images, labels])
images, labels = batch_queue.dequeue()
logits = Net(images)
loss = Loss(logits, labels)

Args:
  tensors: A list or dictionary of `Tensors` to enqueue in the buffer.
  capacity: An integer. The maximum number of elements in the queue.
  num_threads: An integer.  Number of threads running the enqueue op.
  dynamic_pad: Boolean.  Whether to allow variable dimensions in input shapes.
  shared_name: (optional). If set, this queue will be shared under the given
    name across multiple sessions.
  name: (Optional) A name for the operations.

Returns:
  A queue from which you can dequeue tensors with the same type and shape
  as `tensors`.