Tensorflow多GPU并发训练指南

忘是亡心i 2022-04-14 04:54 437阅读 0赞

1.缘起

tensorflow使用GPU时默认占满所有可用GPU的显存，但只在第一个GPU上进行计算。下图展示了一个典型的使用GPU训练的例子，虽然机器上有两块GPU，但却只有一块真正在工作，如果不加以利用，另一块GPU就白白浪费了。我们知道，GPU是一种相对比较昂贵的计算资源，虽然正值矿难，相比之前动辄八九千一块1080Ti的价格低了不少，但也不是一般人能浪费的起的，因此如何有效提高GPU特别是tensorflow上的利用率就成为了一项重要的考量。

2.指定单GPU的简单方法

一种经常提及的方法是设置可见的GPU，方法是通过设置CUDA\_VISIBLE\_DEVICES来完成，如果在shell中运行，每次运行某个shell脚本之前，加上export CUDA\_VISIBLE\_DEVICES=0\#或者是你期望运行的GPU id（0到n-1,其中n是机器上GPU卡的数量），如果是在python脚本，还可以在所有代码之前:

import os  
    os.environ\["CUDA\_VISIBLE\_DEVICES"\]="0"

另外为缓解tensorflow一上来就占满整个GPU导致的问题，还可以在创建sess的时候传入附加的参数，更好的控制GPU的使用方法是

config = tf.ConfigProto(allow\_soft\_placement=True,allow\_grouth=True)  
    config.gpu\_options.per\_process\_gpu\_memory\_fraction = 0.9 \#占用90%显存  
    sess = tf.Session(config=config)

3.多GPU编程指南

上面的方法确实可以解决一定的问题，比如多个用户公用几块卡，每个人分配不同的卡来做实验，或者每张卡上运行不同的参数设置。但有时候我们更需要利用好已有的卡，来或得线性的加速比，以便在更短的时候获取参考结果，上面的方法就无能无力，只能自己写代码解决了，而当前怎么把单GPU代码转换为多GPU代码好像还没有一篇文章给出示例。

说到tensorflow的多GPU编程，不得不说是一部悲壮的血泪史。用过caffe的用户都知道caffe下的多GPU是多么的简单，从单GPU切换到多GPU你根本不用修改任何代码，只需要在编译选项中把USE\_NCCL打开就好，剩下的都是caffe帮你自动完成了。到了tensorflow就不一样的，所有的东西都得你自己操心，包括变量放哪，平均梯度什么的。当然好处是你对训练的过程理解的也更透彻了。

tensorflow的models里面有一个使用多GPU的例子，路径为tutorials/image/cifar10/cifar10\_multi\_gpu\_train.py,但它仅仅是一个能运行的demo而已，里面并没有解释枚举代码的含义，更别提为什么写那些代码了，乱的一比。

从网上找出一个能用且好用的例子并不像想象中的简单，我踏遍整个网络，把百度和谷歌中所有相关的文章都看了一遍，觉得还不错的是Multi-GPU Basic，里面用一个小例子对比了单GPU和多GPU的差异。

另外一个用Mnist的例子是这边文章要描述的重点，不过里面仍有一些干扰我们理解的代码，比如那个assign\_to\_device函数。

其实tensorflow使用多GPU也是非常简单的，无外乎用tf.device("/gpu:0")之类的包起来，但要注意变量存放的位置。

我们也以一个mnist单GPU训练的例子作为开始:

import tensorflow as tf
    import numpy as np
    from tensorflow.examples.tutorials.mnist import input_data
     
    mnist=input_data.read_data_sets("/tmp/mnist/",one_hot=True)
     
    num_gpus=2
    num_steps=200
    learning_rate=0.001
    batch_size=1024
    display_step=10
     
    num_input=784
    num_classes=10
    def conv_net(x,is_training):
        # "updates_collections": None is very import ,without will only get 0.10
        batch_norm_params = {"is_training": is_training, "decay": 0.9, "updates_collections": None}
        #,'variables_collections': [ tf.GraphKeys.TRAINABLE_VARIABLES ]
        with slim.arg_scope([slim.conv2d, slim.fully_connected],
                            activation_fn=tf.nn.relu,
                            weights_initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.01),
                            weights_regularizer=slim.l2_regularizer(0.0005),
                            normalizer_fn=slim.batch_norm, normalizer_params=batch_norm_params):
            with tf.variable_scope("ConvNet",reuse=tf.AUTO_REUSE):
                x = tf.reshape(x, [-1, 28, 28, 1])
                net = slim.conv2d(x, 6, [5,5], scope="conv_1")
                net = slim.max_pool2d(net, [2, 2],scope="pool_1")
                net = slim.conv2d(net, 12, [5,5], scope="conv_2")
                net = slim.max_pool2d(net, [2, 2], scope="pool_2")
                net = slim.flatten(net, scope="flatten")
                net = slim.fully_connected(net, 100, scope="fc")
                net = slim.dropout(net,is_training=is_training)
                net = slim.fully_connected(net, num_classes, scope="prob", activation_fn=None,normalizer_fn=None)
                return net
    def train_single():
        X = tf.placeholder(tf.float32, [None, num_input])
        Y = tf.placeholder(tf.float32, [None, num_classes])
        logits=conv_net(X,True)
        loss=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=Y,logits=logits))
        opt=tf.train.AdamOptimizer(learning_rate)
        train_op=opt.minimize(loss)
        logits_test=conv_net(X,False)
        correct_prediction = tf.equal(tf.argmax(logits_test, 1), tf.argmax(Y, 1))
        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
        with tf.Session() as sess:
            sess.run(tf.global_variables_initializer())
            for step in range(1,num_steps+1):
                batch_x, batch_y = mnist.train.next_batch(batch_size)
                sess.run(train_op,feed_dict={X:batch_x,Y:batch_y})
                if step%display_step==0 or step==1:
                    loss_value,acc=sess.run([loss,accuracy],feed_dict={X:batch_x,Y:batch_y})
                    print("Step:" + str(step) + ":" + str(loss_value) + " " + str(acc))
            print("Done")
            print("Testing Accuracy:",np.mean([sess.run(accuracy, feed_dict={X: mnist.test.images[i:i + batch_size],
                  Y: mnist.test.labels[i:i + batch_size]}) for i in
                  range(0, len(mnist.test.images), batch_size)]))
     
    if __name__ == "__main__":
        train_single()

在上面寥寥数行的代码中，我们下载并构建了mnist数据集，通过mnist.train.next\_batch(batch\_szie)的方式返回一个batch的数据，然后扔进conv\_net网络，计算出loss后再使用Adam优化器进行训练，最后测试了其精度。

只需要几分钟，在我的一块1080Ti卡上就能运行完毕，并且得到相当好的结果:

Step:1:2.213318 0.20703125  
Step:10:0.46338645 0.88183594  
Step:20:0.18729115 0.9550781  
Step:30:0.17860937 0.9628906  
Step:40:0.11540267 0.97558594  
Step:50:0.081396215 0.9824219  
Step:60:0.097750194 0.9746094  
Step:70:0.060169913 0.984375  
Step:80:0.059070613 0.98828125  
Step:90:0.060746174 0.9892578  
Step:100:0.057775088 0.9892578  
Step:110:0.038614694 0.98828125  
Step:120:0.0369242 0.9921875  
Step:130:0.035249908 0.9941406  
Step:140:0.03395287 0.9902344  
Step:150:0.03798459 0.98828125  
Step:160:0.052775905 0.99121094  
Step:170:0.017296169 0.99609375  
Step:180:0.026407585 0.9951172  
Step:190:0.044104658 0.9941406  
Step:200:0.025472593 0.99121094  
Done  
('Testing Accuracy:', 0.99301463)

接下来我们将以此为起点，对其进行多GPU化改造。下面的图很好的阐释了整个流程。

![68747470733a2f2f7777772e74656e736f72666c6f772e6f72672f696d616765732f506172616c6c656c69736d2e706e67][]

多GPU并行可分为模型并行和数据并行两大类，上图展示的是数据并行，这也是我们经常用到的方法，而其中数据并行又可分为同步方式和异步方式两种，由于我们一般都会配置同样的显卡，因此这儿也选择了同步方式，也就是把数据分给不同的卡，等所有的GPU都计算完梯度后进行平均，最后再更新梯度。

首先要改造的就是数据读取部分，由于现在我们有多快卡，每张卡要分到不同的数据，所以在获取batch的时候要把大小改为batch\_x,batch\_y=mnist.train.next\_batch(batch\_size\*num\_gpus)，一次取足够的数据保证每块卡都分到batch\_size大小的数据。然后我们对取到的数据进行切分，我们以i表示GPU的索引，连续的batch\_size大小的数据分给同一块GPU:

\_x=X\[i\*batch\_size:(i+1)\*batch\_size\]  
    \_y=Y\[i\*batch\_size:(i+1)\*batch\_size\]

由于我们多个GPU上共享同样的图，为了防止名字混乱，最好使用name\_scope进行区分，也就是如下的形式:

for i in range(2):  
                with tf.device("/gpu:%d"%i):  
                    with tf.name\_scope("tower\_%d"%i):  
                        \_x=X\[i\*batch\_size:(i+1)\*batch\_size\]  
                        \_y=Y\[i\*batch\_size:(i+1)\*batch\_size\]  
                        logits=conv\_net(\_x,dropout,reuse\_vars,True)

我们需要有个列表存储所有GPU上的梯度，还有就是复用变量，需要在之前定义如下两个值:

tower\_grads=\[\]  
     reuse\_vars=False

所有的准备工作都已完成，就可以计算每个GPU上的梯度了

opt = tf.train.AdamOptimizer(learning\_rate)  
                        loss=tf.reduce\_mean(tf.nn.softmax\_cross\_entropy\_with\_logits(labels=\_y,logits=logits))  
                        grads=opt.compute\_gradients(loss)  
                        reuse\_vars=True  
                        tower\_grads.append(grads)

这样tower\_grads就存储了所有GPU上所有变量的梯度，下面就是计算平均值了，这个是所有见过的函数中唯一一个几乎从没变过的代码:

def average\_gradients(tower\_grads):  
        average\_grads=\[\]  
        for grad\_and\_vars in zip(\*tower\_grads):  
            grads=\[\]  
            for g,\_ in grad\_and\_vars:  
                expend\_g=tf.expand\_dims(g,0)  
                grads.append(expend\_g)  
            grad=tf.concat(grads,0)  
            grad=tf.reduce\_mean(grad,0)  
            v=grad\_and\_vars\[0\]\[1\]  
            grad\_and\_var=(grad,v)  
            average\_grads.append(grad\_and\_var)  
        return average\_grads

tower\_grads里面保存的形式是(第一个GPU上的梯度，第二个GPU上的梯度,...第N-1个GPU上的梯度)，这里有一点需要注意的是zip(\*)，它的作用上把上面的那个列表转换成((grad0\_gpu0, var0\_gpu0), ... , (grad0\_gpuN, var0\_gpuN))的形式，也就是以列访问的方式，取到的就是某个变量在不同GPU上的值。

最后就是更新梯度了

grads=average\_gradients(tower\_grads)  
    train\_op=opt.apply\_gradients(grads)

上面的讲述略有零散，最后我们给个全代码版本方便大家测试:

import tensorflow as tf
    import numpy as np
    from tensorflow.contrib import slim
    from tensorflow.examples.tutorials.mnist import input_data
    mnist = input_data.read_data_sets("/tmp/mnist/", one_hot=True)
     
    num_gpus = 2
    num_steps = 1000
    learning_rate = 0.001
    batch_size = 1000
    display_step = 10
     
    num_input = 784
    num_classes = 10
     
    def conv_net_with_layers(x,is_training,dropout = 0.75):
        with tf.variable_scope("ConvNet", reuse=tf.AUTO_REUSE):
            x = tf.reshape(x, [-1, 28, 28, 1])
            x = tf.layers.conv2d(x, 12, 5, activation=tf.nn.relu)
            x = tf.layers.max_pooling2d(x, 2, 2)
            x = tf.layers.conv2d(x, 24, 3, activation=tf.nn.relu)
            x = tf.layers.max_pooling2d(x, 2, 2)
            x = tf.layers.flatten(x)
            x = tf.layers.dense(x, 100)
            x = tf.layers.dropout(x, rate=dropout, training=is_training)
            out = tf.layers.dense(x, 10)
            out = tf.nn.softmax(out) if not is_training else out
        return out
     
    def conv_net(x,is_training):
        # "updates_collections": None is very import ,without will only get 0.10
        batch_norm_params = {"is_training": is_training, "decay": 0.9, "updates_collections": None}
        #,'variables_collections': [ tf.GraphKeys.TRAINABLE_VARIABLES ]
        with slim.arg_scope([slim.conv2d, slim.fully_connected],
                            activation_fn=tf.nn.relu,
                            weights_initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.01),
                            weights_regularizer=slim.l2_regularizer(0.0005),
                            normalizer_fn=slim.batch_norm, normalizer_params=batch_norm_params):
            with tf.variable_scope("ConvNet",reuse=tf.AUTO_REUSE):
                x = tf.reshape(x, [-1, 28, 28, 1])
                net = slim.conv2d(x, 6, [5,5], scope="conv_1")
                net = slim.max_pool2d(net, [2, 2],scope="pool_1")
                net = slim.conv2d(net, 12, [5,5], scope="conv_2")
                net = slim.max_pool2d(net, [2, 2], scope="pool_2")
                net = slim.flatten(net, scope="flatten")
                net = slim.fully_connected(net, 100, scope="fc")
                net = slim.dropout(net,is_training=is_training)
                net = slim.fully_connected(net, num_classes, scope="prob", activation_fn=None,normalizer_fn=None)
                return net
     
    def average_gradients(tower_grads):
        average_grads = []
        for grad_and_vars in zip(*tower_grads):
            grads = []
            for g, _ in grad_and_vars:
                expend_g = tf.expand_dims(g, 0)
                grads.append(expend_g)
            grad = tf.concat(grads, 0)
            grad = tf.reduce_mean(grad, 0)
            v = grad_and_vars[0][1]
            grad_and_var = (grad, v)
            average_grads.append(grad_and_var)
        return average_grads
     
     
    def train():
        with tf.device("/cpu:0"):
            global_step=tf.train.get_or_create_global_step()
            tower_grads = []
            X = tf.placeholder(tf.float32, [None, num_input])
            Y = tf.placeholder(tf.float32, [None, num_classes])
            opt = tf.train.AdamOptimizer(learning_rate)
            with tf.variable_scope(tf.get_variable_scope()):
                for i in range(2):
                    with tf.device("/gpu:%d" % i):
                        with tf.name_scope("tower_%d" % i):
                                _x = X[i * batch_size:(i + 1) * batch_size]
                                _y = Y[i * batch_size:(i + 1) * batch_size]
                                logits = conv_net(_x, True)
                                tf.get_variable_scope().reuse_variables()
                                loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=_y, logits=logits))
                                grads = opt.compute_gradients(loss)
                                tower_grads.append(grads)
                                if i == 0:
                                    logits_test = conv_net(_x, False)
                                    correct_prediction = tf.equal(tf.argmax(logits_test, 1), tf.argmax(_y, 1))
                                    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
            grads = average_gradients(tower_grads)
            train_op = opt.apply_gradients(grads)
            with tf.Session() as sess:
                sess.run(tf.global_variables_initializer())
                for step in range(1, num_steps + 1):
                    batch_x, batch_y = mnist.train.next_batch(batch_size * num_gpus)
                    sess.run(train_op, feed_dict={X: batch_x, Y: batch_y})
                    if step % 10 == 0 or step == 1:
                        loss_value, acc = sess.run([loss, accuracy], feed_dict={X: batch_x, Y: batch_y})
                        print("Step:" + str(step) + ":" + str(loss_value) + " " + str(acc))
                print("Done")
                print("Testing Accuracy:",
                      np.mean([sess.run(accuracy, feed_dict={X: mnist.test.images[i:i + batch_size],
                                                             Y: mnist.test.labels[i:i + batch_size]}) for i in
                               range(0, len(mnist.test.images), batch_size)]))
    def train_single():
        X = tf.placeholder(tf.float32, [None, num_input])
        Y = tf.placeholder(tf.float32, [None, num_classes])
        logits=conv_net(X,True)
        loss=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=Y,logits=logits))
        opt=tf.train.AdamOptimizer(learning_rate)
        train_op=opt.minimize(loss)
        logits_test=conv_net(X,False)
        correct_prediction = tf.equal(tf.argmax(logits_test, 1), tf.argmax(Y, 1))
        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
        with tf.Session() as sess:
            sess.run(tf.global_variables_initializer())
            for step in range(1,num_steps+1):
                batch_x, batch_y = mnist.train.next_batch(batch_size)
                sess.run(train_op,feed_dict={X:batch_x,Y:batch_y})
                if step%display_step==0 or step==1:
                    loss_value,acc=sess.run([loss,accuracy],feed_dict={X:batch_x,Y:batch_y})
                    print("Step:" + str(step) + ":" + str(loss_value) + " " + str(acc))
            print("Done")
            print("Testing Accuracy:",np.mean([sess.run(accuracy, feed_dict={X: mnist.test.images[i:i + batch_size],
                  Y: mnist.test.labels[i:i + batch_size]}) for i in
                  range(0, len(mnist.test.images), batch_size)]))
     
    if __name__ == "__main__":
        #train_single()
        train()

我这边的运行结果是

Step:1:2.1529438 0.27929688  
Step:10:0.4463266 0.89941406  
Step:20:0.18885617 0.9580078  
Step:30:0.12153786 0.96484375  
Step:40:0.07257775 0.98339844  
Step:50:0.07452829 0.98535156  
Step:60:0.048265547 0.99121094  
Step:70:0.02948389 0.9873047  
Step:80:0.048876762 0.9902344  
Step:90:0.06494201 0.9902344  
Step:100:0.024681691 0.98535156  
Step:110:0.025596365 0.99316406  
Step:120:0.02779768 0.9941406  
Step:130:0.02191917 0.9980469  
Step:140:0.022235561 0.9951172  
Step:150:0.0124597605 0.9970703  
Step:160:0.011557209 0.9902344  
Step:170:0.014788041 0.99609375  
Step:180:0.015063373 0.99609375  
Step:190:0.012394376 0.99902344  
Step:200:0.017424839 0.99609375  
Done  
('Testing Accuracy:', 0.99301463)

[68747470733a2f2f7777772e74656e736f72666c6f772e6f72672f696d616765732f506172616c6c656c69736d2e706e67]: /images/20220414/21a1fe7e69bc440080d35da6f1f060ed.png