#!/usr/bin/env python

import tensorflow as tf
import horovod.tensorflow as hvd
layers = tf.contrib.layers
learn = tf.contrib.learn

tf.logging.set_verbosity(tf.logging.INFO)


def conv_model(feature, target, mode):
    """2-layer convolution model."""
    # Convert the target to a one-hot tensor of shape (batch_size, 10) and
    # with a on-value of 1 for each one-hot vector of length 10.
    target = tf.one_hot(tf.cast(target, tf.int32), 10, 1, 0)

    # Reshape feature to 4d tensor with 2nd and 3rd dimensions being
    # image width and height final dimension being the number of color channels.
    feature = tf.reshape(feature, [-1, 28, 28, 1])

    # First conv layer will compute 32 features for each 5x5 patch
    with tf.variable_scope('conv_layer1'):
        h_conv1 = layers.conv2d(
            feature, 32, kernel_size=[5, 5], activation_fn=tf.nn.relu)
        h_pool1 = tf.nn.max_pool(
            h_conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

    # Second conv layer will compute 64 features for each 5x5 patch.
    with tf.variable_scope('conv_layer2'):
        h_conv2 = layers.conv2d(
            h_pool1, 64, kernel_size=[5, 5], activation_fn=tf.nn.relu)
        h_pool2 = tf.nn.max_pool(
            h_conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
        # reshape tensor into a batch of vectors
        h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])

    # Densely connected layer with 1024 neurons.
    h_fc1 = layers.dropout(
        layers.fully_connected(
            h_pool2_flat, 1024, activation_fn=tf.nn.relu),
        keep_prob=0.5,
        is_training=mode == tf.contrib.learn.ModeKeys.TRAIN)

    # Compute logits (1 per class) and compute loss.
    logits = layers.fully_connected(h_fc1, 10, activation_fn=None)
    loss = tf.losses.softmax_cross_entropy(target, logits)

    return tf.argmax(logits, 1), loss


def main(_):
    # Horovod: initialize Horovod.
    hvd.init()

    # Download and load MNIST dataset.
    mnist = learn.datasets.mnist.read_data_sets('MNIST-data-%d' % hvd.rank())

    # Build model...
    with tf.name_scope('input'):
        image = tf.placeholder(tf.float32, [None, 784], name='image')
        label = tf.placeholder(tf.float32, [None], name='label')
    predict, loss = conv_model(image, label, tf.contrib.learn.ModeKeys.TRAIN)

    # Horovod: adjust learning rate based on number of GPUs.
    opt = tf.train.RMSPropOptimizer(0.001 * hvd.size())

    # Horovod: add Horovod Distributed Optimizer.
    opt = hvd.DistributedOptimizer(opt)

    global_step = tf.contrib.framework.get_or_create_global_step()
    train_op = opt.minimize(loss, global_step=global_step)

    hooks = [
        # Horovod: BroadcastGlobalVariablesHook broadcasts initial variable states
        # from rank 0 to all other processes. This is necessary to ensure consistent
        # initialization of all workers when training is started with random weights
        # or restored from a checkpoint.
        hvd.BroadcastGlobalVariablesHook(0),

        # Horovod: adjust number of steps based on number of GPUs.
        tf.train.StopAtStepHook(last_step=20000 // hvd.size()),

        tf.train.LoggingTensorHook(tensors={'step': global_step, 'loss': loss},
                                   every_n_iter=10),
    ]

    # Horovod: pin GPU to be used to process local rank (one GPU per process)
    config = tf.ConfigProto()
    config.gpu_options.allow_growth = True
    config.gpu_options.visible_device_list = str(hvd.local_rank())

    # Horovod: save checkpoints only on worker 0 to prevent other workers from
    # corrupting them.
    checkpoint_dir = './checkpoints' if hvd.rank() == 0 else None

    # The MonitoredTrainingSession takes care of session initialization,
    # restoring from a checkpoint, saving to a checkpoint, and closing when done
    # or an error occurs.
    with tf.train.MonitoredTrainingSession(checkpoint_dir=checkpoint_dir,
                                           hooks=hooks,
                                           config=config) as mon_sess:
        while not mon_sess.should_stop():
            # Run a training step synchronously.
            image_, label_ = mnist.train.next_batch(100)
            mon_sess.run(train_op, feed_dict={image: image_, label: label_})


if __name__ == "__main__":
    tf.app.run()

为了使用Horovod,必须把下面的内容增加到程序中
hvd.init()

使用config.gpu_options.visible_device_list固定此进程要使用的服务器GPU.通过每个进程一个GPU的典型设置，
可以将其设置为本地排名。在这种情况下，服务器上的第一个进程将分配第一个GPU,第二个进程将分配第二个GPU,依次类推。。

扩大学习率通过workers节点的数目
同步分布式训练的有效批量大小按workers数目进行调整
学习率的提高可以补偿增加的批次大小

在hvd.DistributedOptimzer中包装优化器。分布式优化器将梯度训练委托给原始优化器，使用averages梯度或allgather渐变，然后应用这些平滑梯度。
增加hvd.BroadcastGlobalVariablesHook(0),为了广播初始化为0的出事变量状态广播到其他进程。
当训练以随机权重开始或检查点恢复时，确保workers节点一致性非常有必要。或者，如果你没有使用MonitoredTrainingSession,可以再初始化全局变量后执行hvd.broadcast_global_variables操作

修改代码保存在worker 0，以防止其他工作人员破坏他们，如果hvd.rank()!=0，可以通过checkpoint_dir=None传递给tf.train.MonitoredTrainingSession来完成此操作。

import tensorflow as tf
import horovod.tensorflow as hvd

#Initialize Horovod
hvd.init()

#Pin GPU to be used to process local rank (one GPU per process)
config = tf.ConfigProto()
config.gpu_options.visible_device_list = str(hvd.local_rank())
print(config.gpu_options.visible_device_list)
print(hvd.size())

#Build model ...
loss = ...
opt = tf.train.AdadeltaOptimizer(0.01 * hvd.size())

#Add Horovod Distributed Optimizer
opt = hvd.DistributedOptimizer(opt)

#Add hook to broadcast variables from rank 0 to all other processes during
#initialization.
hooks = [hvd.BroadcastGlobalVariablesHook(0)]

#Mke training operation
train_op = opt.minimize(loss)

#Save checkpoints only on worker 0 to prevent other workers from corrupting them.
checkpoint_dir = '/tmp/train_logs' if hvd.rank() == 0 else None

#The MonitoredTrainingSession takes care of session initialization,
#restoring from a checkpoint, saving to a checkpoint, and closing when done
# or an error occures.
with tf.train.MonitoredTrainingSession(checkpoint_dir=checkpoint_dir,
                                       config=config,
                                       hooks=hooks) as mon_sess:
    while not mon_sess.should_stop():
        #Perform synchronous training.
        mon_sess.run(train_op)

运行方式
1运行在一台机器的4个GPUs上
mpirun -np 4 \
    -H localhost:4 \
    -bind-to none -map-by slot \
    -x NCCL_DEBUG=INFO -x LD_LIBRARY_PATH -x PATH \
    -mca pml ob1 -mca btl ^openib \
    python train.py
2运行在四台机器的4个GPUs上
mpirun -np 16 \
    -H server1:4,server2:4,server3:4,server4:4 \
    -bind-to none -map-by slot \
    -x NCCL_DEBUG=INFO -x LD_LIBRARY_PATH -x PATH \
    -mca pml ob1 -mca btl ^openib \
    python train.py