89、tensorflow使用GPU并行计算

'''
Created on May 25, 2017

@author: p0079482
'''
# 分布式深度学习模型训练模式
# 在一台机器的多个GPU上并行训练深度学习模型
from datetime import datetime
import os
import time

import tensorflow as tf
import mnist_inference

# 定义训练神经网络时需要用到的配置。
BATCH_SIZE = 100
LEARNING_RATE_BASE = 0.001
LEARNING_RATE_DECAY = 0.99
REGULARAZTION_REATE = 0.0001
TRAINING_STEPS = 1000
MOVING_AVERAGE_DECAY = 0.99
N_GPU = 4

# 定义日志和模型输出的路径
MODEL_SAVE_PATH = "/path/to/logs_and_models/"
MODEL_NAME = "model.ckpt"

# 定义数据存储的路径。因为需要为不同的GPU提供不同的训练数据，所以通过placerholder的方式
# 就需要手动准备多分数据。为了方便训练数据的获取过程，可以采用第7章中介绍的输入队列的方式
# 从TFRecord中读取数据。于是在这里提供的数据文件路径为将MNIST训练数据转化为TFRecords格式之后的路径
# 如何将MNIST数据转化为TFRecord格式在第7章中有详细介绍，这里不再赘述
DATA_PATH = "/path/to/output.tfrecords"


# 定义输入队列得到训练数据，具体细节可以参考第7章
def get_input():
    filename_queue = tf.train.string_input_producer([DATA_PATH])
    reader = tf.TFRecordReader()
    _, serialized_example = reader.read(filename_queue)
    # 定义数据解析格式
    features = tf.parse_single_example(serialized_example,
                                       features={
                                          'image_raw':tf.FixedLenFeature([], tf.string),
                                          'pixels':tf.FixedLenFeature([], tf.int64),
                                          'label':tf.FixedLenFeature([], tf.int64),
                                        })
    # 解析图片和标签信息
    decoded_image = tf.decode_raw(features['image_raw'], tf.uint8)
    reshaped_image = tf.reshape(decoded_image, [784])
    retyped_image = tf.cast(reshaped_image, tf.float32)
    label = tf.cast(features['label'], tf.int32)
    
    # 定义输入队列并返回
    min_after_dequeue = 10000
    capacity = min_after_dequeue + 3 * BATCH_SIZE
    return tf.train.shuffle_batch([retyped_image, label],
                                  batch_size=BATCH_SIZE,
                                  capacity=capacity,
                                  min_after_dequeue=min_after_dequeue)
    
# 定义损失函数。对于给定的训练数据、正则化损失计算规则和命名空间，计算在这个命名空间下的总损失
# 之所以需要给定命名空间是因为不同的GPU上计算得出的正则化损失都会加入名为loss的集合，
# 如果不通过命名空间就会将不同GPU上的正则化损失都加进来
def get_loss(x, y_, regularizer, scope, reuse_variables=None):
    # 沿用5.5节中定义的函数来计算神经网络的前向传播结果
    with tf.variable_scope(tf.get_variable_scope(), reuse=reuse_variables):
        y = mnist_inference.inference(x, regularizer)
    # 计算交叉熵损失
    cross_entropy = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y, labels=y_))
    # 计算当前GPU上计算得到的正则化损失
    regularization_loss = tf.add_n(tf.get_collection('losses', scope))
    # 计算最终的总损失
    loss = cross_entropy + regularization_loss
    return loss

def average_gradients(tower_grads):
    average_grads = []
    # 枚举所有的变量和变量在不同GPU上计算得出的梯度
    for grad_and_vars in zip(*tower_grads):
        # 计算所有GPU上的梯度平均值
        grads = []
        for g, _ in grad_and_vars:
            expanded_g = tf.expand_dims(g, 0)
            grads.append(expanded_g)
        grad = tf.concat(grads, 0)
        grad = tf.reduce_mean(grad, 0)
        
        v = grad_and_vars[0][1]
        grad_and_var = (grad, v)
        # 将变量和它的平均梯度对应起来
        average_grads.append(grad_and_var)
    # 返回所有变量的平均梯度，这将被用于变量更新
    return average_grads

# 主训练过程
def main(argv=None):
    # 将简单的运算放在CPU上，只有神经网络的训练过程放在GPU上
    with tf.Graph().as_default(), tf.device('/cpu:0'):
        # 获取训练batch
        x, y_ = get_input()
        regularizer = tf.contrib.layers.l2_regularizer(REGULARAZTION_REATE)
        
        # 定义训练轮数和指数衰减的学习率
        global_step = tf.get_variable('global_step',
                                      [],
                                      initializer=tf.constant_initializer(0),
                                      trainable=False)
        learning_rate = tf.train.exponential_decay(LEARNING_RATE_BASE,
                                                   global_step,
                                                   60000 / BATCH_SIZE,
                                                   LEARNING_RATE_DECAY)
        # 定义优化方法
        opt = tf.train.GradientDescentOptimizer(learning_rate)
        
        tower_grads = []
        reuse_variables = False
        # 将神经网络的优化过程跑在不同的GPU上
        for i in range(N_GPU):
            # 将优化过程指定在一个GPU上
            with tf.device('/gpu:%d' % i):
                with tf.name_scope('GPU_%d' % i) as scope:
                    cur_loss = get_loss(x, y_, regularizer, scope, reuse_variables)
                    # 在第一次声明变量之后，将控制变量重用的参数设置为True.这样可以
                    # 让不同的GPU更新同一组参数。注意tf.name_scope函数并不会影响
                    # tf.get_variable的命名空间
                    reuse_variables = True
                    # 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，并输出到TensorBoard日志中
        grads = average_gradients(tower_grads)
        for grad, var in grads:
            if grad is not None:
                tf.summary.histogram('gradients_on_average/%s' % var.op.name, grad)
        
        # 使用平均梯度更新参数
        apply_gradient_op = opt.apply_gradients(grads,
                                                global_step=global_step)
        for var in tf.trainable_variables():
            tf.summary.histogram(var.op.name, var)
        
        # 计算变量的滑动平均值
        variable_averages = tf.train.ExponentialMovingAverage(
                                                              MOVING_AVERAGE_DECAY,
                                                              global_step)
        variables_to_average = (tf.trainable_variables() + tf.moving_average_variables())
        variable_averages_op = variable_averages.apply(variables_to_average)
        
        # 每一轮迭代需要更新变量的取值并更新变量的滑动平均值
        train_op = tf.group(apply_gradient_op, variable_averages_op)
        saver = tf.train.Saver(tf.global_variables())
        summary_op = tf.summary.merge_all()
        init = tf.global_variables_initializer()
        
        # 训练过程
        with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
                                              log_device_placement=True)) as sess:
            # 初始化所有变量并启动队列
            init.run()
            coord = tf.train.Coordinator()
            threads = tf.train.start_queue_runners(sess=sess, coord=coord)
            summary_writer = tf.summary.FileWriter(MODEL_SAVE_PATH, sess.graph)
            for step in range(TRAINING_STEPS):
                # 执行神经网络训练操作，并记录训练操作的运行时间
                start_time = time.time()
                _, loss_value = sess.run([train_op, cur_loss])
                duration = time.time() - start_time
                
                 # 每隔一段时间展示当前的训练进度，并统计训练速度
                if step != 0 and step % 10 == 0 and duration != 0:
                    # 计算使用过的训练数据个数，因为在每一次运行训练操作时，每一个GPU
                    # 都会使用一个batch的训练数据，所以总共用到的训练数据个数为
                    # batch大小*GPU个数
                    num_examples_per_step = BATCH_SIZE * N_GPU
                    
                    # num_examples_per_step为本次迭代使用到的训练数据个数
                    # duration为运行当前训练过程使用的时间，于是平均每秒可以处理的训练数据个数为
                    # num_examples_per_step/duration
                    examples_per_sec = num_examples_per_step / duration
                    
                    # duration为运行当前训练过程使用的时间，因为在每一个训练过程中
                    # 每一个GPU都会使用一个batch的训练数据，所以在单个batch上的训练所需要时间为
                    # duration/GPU个数
                    sec_per_batch = duration / N_GPU
                    
                    # 输出训练信息
                    format_str = ('step %d, loss = %.2f (%  .1f examples/ sec; %.3f sec/batch)')
                    print(format_str % (step, loss_value, examples_per_sec, sec_per_batch))
                    
                    # 通过TensorBoard可视化训练过程
                    summary = sess.run(summary_op)
                    summary_writer.add_summary(summary, step)
                    
                    # 每隔一段时间保存当前的模型
                    if step % 1000 == 0 or (step + 1) == TRAINING_STEPS:
                        checkpoint_path = os.path.join(MODEL_SAVE_PATH, MODEL_NAME)
                        saver.save(sess, checkpoint_path, global_step=step)
            coord.request_stop()
            coord.join(threads)

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

下面是训练完的结果

step 20, loss = 29.53 ( 10362.6 examples/ sec; 0.010 sec/batch)
step 30, loss = 9.62 ( 12022.4 examples/ sec; 0.008 sec/batch)
step 40, loss = 16.63 ( 10689.3 examples/ sec; 0.009 sec/batch)
step 50, loss = 10.68 ( 11293.4 examples/ sec; 0.009 sec/batch)
step 60, loss = 14.73 ( 10895.0 examples/ sec; 0.009 sec/batch)
step 70, loss = 17.17 ( 11192.9 examples/ sec; 0.009 sec/batch)
step 80, loss = 12.43 ( 11236.8 examples/ sec; 0.009 sec/batch)
step 90, loss = 5.16 ( 11398.3 examples/ sec; 0.009 sec/batch)
step 100, loss = 8.06 ( 12466.7 examples/ sec; 0.008 sec/batch)
step 110, loss = 13.57 ( 11081.5 examples/ sec; 0.009 sec/batch)
step 120, loss = 9.43 ( 11396.2 examples/ sec; 0.009 sec/batch)
step 130, loss = 12.21 ( 13296.7 examples/ sec; 0.008 sec/batch)
step 140, loss = 6.15 ( 11868.9 examples/ sec; 0.008 sec/batch)
step 150, loss = 9.93 ( 12089.1 examples/ sec; 0.008 sec/batch)
step 160, loss = 10.42 ( 11733.5 examples/ sec; 0.009 sec/batch)
step 170, loss = 23.47 ( 11859.4 examples/ sec; 0.008 sec/batch)
step 180, loss = 2.97 ( 11358.0 examples/ sec; 0.009 sec/batch)
step 190, loss = 5.44 ( 11085.0 examples/ sec; 0.009 sec/batch)
step 200, loss = 3.98 ( 13347.3 examples/ sec; 0.007 sec/batch)
step 210, loss = 11.98 ( 10551.4 examples/ sec; 0.009 sec/batch)
step 220, loss = 9.17 ( 11115.3 examples/ sec; 0.009 sec/batch)
step 230, loss = 15.31 ( 12450.5 examples/ sec; 0.008 sec/batch)
step 240, loss = 5.92 ( 11729.5 examples/ sec; 0.009 sec/batch)
step 250, loss = 9.94 ( 10497.2 examples/ sec; 0.010 sec/batch)
step 260, loss = 2.94 ( 11398.1 examples/ sec; 0.009 sec/batch)
step 270, loss = 7.30 ( 10497.4 examples/ sec; 0.010 sec/batch)
step 280, loss = 3.98 ( 11946.0 examples/ sec; 0.008 sec/batch)
step 290, loss = 7.66 ( 11307.2 examples/ sec; 0.009 sec/batch)
step 300, loss = 2.03 ( 11968.7 examples/ sec; 0.008 sec/batch)
step 310, loss = 2.39 ( 8672.0 examples/ sec; 0.012 sec/batch)
step 320, loss = 2.07 ( 3835.6 examples/ sec; 0.026 sec/batch)
step 330, loss = 2.71 ( 12087.7 examples/ sec; 0.008 sec/batch)
step 340, loss = 2.70 ( 11907.3 examples/ sec; 0.008 sec/batch)
step 350, loss = 7.17 ( 7671.2 examples/ sec; 0.013 sec/batch)
step 360, loss = 8.36 ( 11863.6 examples/ sec; 0.008 sec/batch)
step 370, loss = 2.48 ( 11782.7 examples/ sec; 0.008 sec/batch)
step 380, loss = 2.27 ( 11081.5 examples/ sec; 0.009 sec/batch)
step 390, loss = 2.85 ( 11562.4 examples/ sec; 0.009 sec/batch)
step 400, loss = 2.99 ( 12088.9 examples/ sec; 0.008 sec/batch)
step 410, loss = 5.08 ( 12465.6 examples/ sec; 0.008 sec/batch)
step 420, loss = 2.12 ( 12869.1 examples/ sec; 0.008 sec/batch)
step 430, loss = 2.83 ( 13756.3 examples/ sec; 0.007 sec/batch)
step 440, loss = 7.56 ( 13297.8 examples/ sec; 0.008 sec/batch)
step 450, loss = 3.51 ( 12634.6 examples/ sec; 0.008 sec/batch)
step 460, loss = 2.23 ( 13297.8 examples/ sec; 0.008 sec/batch)
step 470, loss = 1.80 ( 12869.2 examples/ sec; 0.008 sec/batch)
step 480, loss = 5.92 ( 9730.3 examples/ sec; 0.010 sec/batch)
step 490, loss = 4.01 ( 12647.0 examples/ sec; 0.008 sec/batch)
step 500, loss = 2.29 ( 12466.9 examples/ sec; 0.008 sec/batch)
step 510, loss = 2.20 ( 13078.4 examples/ sec; 0.008 sec/batch)
step 520, loss = 3.70 ( 13296.5 examples/ sec; 0.008 sec/batch)
step 530, loss = 2.11 ( 13298.3 examples/ sec; 0.008 sec/batch)
step 540, loss = 1.73 ( 13296.6 examples/ sec; 0.008 sec/batch)
step 550, loss = 1.20 ( 12868.9 examples/ sec; 0.008 sec/batch)
step 560, loss = 3.44 ( 13078.6 examples/ sec; 0.008 sec/batch)
step 570, loss = 1.35 ( 11562.0 examples/ sec; 0.009 sec/batch)
step 580, loss = 3.51 ( 13205.2 examples/ sec; 0.008 sec/batch)
step 590, loss = 3.11 ( 12868.8 examples/ sec; 0.008 sec/batch)
step 600, loss = 3.40 ( 12869.1 examples/ sec; 0.008 sec/batch)
step 610, loss = 2.49 ( 13297.7 examples/ sec; 0.008 sec/batch)
step 620, loss = 2.68 ( 12620.3 examples/ sec; 0.008 sec/batch)
step 630, loss = 2.09 ( 11907.3 examples/ sec; 0.008 sec/batch)
step 640, loss = 3.82 ( 8487.3 examples/ sec; 0.012 sec/batch)
step 650, loss = 2.77 ( 11081.5 examples/ sec; 0.009 sec/batch)
step 660, loss = 2.55 ( 12089.1 examples/ sec; 0.008 sec/batch)
step 670, loss = 2.53 ( 10228.3 examples/ sec; 0.010 sec/batch)
step 680, loss = 5.17 ( 9498.5 examples/ sec; 0.011 sec/batch)
step 690, loss = 2.02 ( 10498.4 examples/ sec; 0.010 sec/batch)
step 700, loss = 0.21 ( 12088.9 examples/ sec; 0.008 sec/batch)
step 710, loss = 1.95 ( 12868.7 examples/ sec; 0.008 sec/batch)
step 720, loss = 3.90 ( 13296.2 examples/ sec; 0.008 sec/batch)
step 730, loss = 2.17 ( 9277.6 examples/ sec; 0.011 sec/batch)
step 740, loss = 1.09 ( 9730.1 examples/ sec; 0.010 sec/batch)
step 750, loss = 1.33 ( 12466.8 examples/ sec; 0.008 sec/batch)
step 760, loss = 3.17 ( 9797.9 examples/ sec; 0.010 sec/batch)
step 770, loss = 3.20 ( 13297.9 examples/ sec; 0.008 sec/batch)
step 780, loss = 4.28 ( 13756.4 examples/ sec; 0.007 sec/batch)
step 790, loss = 1.23 ( 12465.4 examples/ sec; 0.008 sec/batch)
step 800, loss = 1.78 ( 12868.8 examples/ sec; 0.008 sec/batch)
step 810, loss = 1.12 ( 12924.2 examples/ sec; 0.008 sec/batch)
step 820, loss = 2.09 ( 13297.1 examples/ sec; 0.008 sec/batch)
step 830, loss = 0.71 ( 11967.1 examples/ sec; 0.008 sec/batch)
step 840, loss = 3.03 ( 12088.8 examples/ sec; 0.008 sec/batch)
step 850, loss = 2.76 ( 12868.8 examples/ sec; 0.008 sec/batch)
step 860, loss = 1.64 ( 12087.1 examples/ sec; 0.008 sec/batch)
step 870, loss = 2.43 ( 9066.8 examples/ sec; 0.011 sec/batch)
step 880, loss = 1.73 ( 11398.2 examples/ sec; 0.009 sec/batch)
step 890, loss = 0.61 ( 12980.4 examples/ sec; 0.008 sec/batch)
step 900, loss = 3.44 ( 12868.8 examples/ sec; 0.008 sec/batch)
step 910, loss = 0.96 ( 11445.9 examples/ sec; 0.009 sec/batch)
step 920, loss = 2.95 ( 13756.3 examples/ sec; 0.007 sec/batch)
step 930, loss = 2.99 ( 12868.5 examples/ sec; 0.008 sec/batch)
step 940, loss = 0.34 ( 13752.5 examples/ sec; 0.007 sec/batch)
step 950, loss = 1.05 ( 13297.8 examples/ sec; 0.008 sec/batch)
step 960, loss = 2.34 ( 13295.7 examples/ sec; 0.008 sec/batch)
step 970, loss = 1.32 ( 13297.6 examples/ sec; 0.008 sec/batch)
step 980, loss = 2.46 ( 12466.6 examples/ sec; 0.008 sec/batch)
step 990, loss = 1.02 ( 13297.7 examples/ sec; 0.008 sec/batch)