学习笔记TF065:TensorFlowOnSpark

Hadoop生态大数据系统分为Yam、 HDFS、MapReduce计算框架。TensorFlow分布式相当于MapReduce计算框架，Kubernetes相当于Yam调度系统。TensorFlowOnSpark，利用远程直接内存访问(Remote Direct Memory Access,RDMA)解决存储功能和调度，实现深度学习和大数据融合。TensorFlowOnSpark(TFoS)，雅虎开源项目。 。支持ApacheSpark集群分布式TensorFlow训练、预测。TensorFlowOnSpark提供桥接程序，每个Spark Executor启动一个对应TensorFlow进程，通过远程进程通信(RPC)交互。


TensorFlowOnSpark架构。TensorFlow训练程序用Spark集群运行，管理Spark集群步骤：预留，在Executor执行每个TensorFlow进程保留一个端口，启动数据消息监听器。启动，在Executor启动TensorFlow主函数。数据获取，TensorFlow Readers和QueueRunners机制直接读取HDFS数据文件，Spark不访问数据；Feeding，SparkRDD 数据发送TensorFlow节点，数据通过feed_dict机制传入TensorFlow计算图。关闭，关闭Executor TensorFlow计算节点、参数服务节点。Spark Driver->Spark Executor->参数服务器->TensorFlow Core->gRPC、RDMA->HDFS数据集。 。


TensorFlowOnSpark MNIST。/wiki/GetStarted_standalone 。Standalone模式Spark集群，一台计算机。安装 Spark、Hadoop。部署Java 1.8.0 JDK。下载Spark2.1.0版 。下载Hadoop2.7.3版 。0.12.1版本支持较好。
修改配置文件，设置环境变量，启动Hadoop:$HADOOP_HOME/sbin/start-all.sh。检出TensorFlowOnSpark源代码：


    git clone --recurse-submodules .git
    cd TensorFlowOnSpark
    git submodule init
    git submodule update --force
    git submodule foreach --recursive git clean -dfx


源代码打包，提交任务使用：


    cd TensorflowOnSpark/src
    zip -r ../tfspark.zip *


设置TensorFlowOnSpark根目录环境变量：


    cd TensorFlowOnSpark
    export TFoS_HOME=$(pwd)


启动Spark主节点(master):


    $(SPARK_HOME)/sbin/start-master.sh


配置两个工作节点(worker)实例，master-spark-URL连接主节点：


    export MASTER=spark://$(hostname):7077
    export SPARK_WORKER_INSTANCES=2
    export CORES_PER_WORKER=1
    export TOTAL_CORES=$(($(CORES_PER_WORKER)*$(SPARK_WORKER_INSTANCES)))
    $(SPARK_HOME)/sbin/start-slave.sh -c $CORES_PER_WORKER -m 3G $(MASTER)


提交任务，MNIST zip文件转换为HDFS RDD 数据集：


    $(SPARK_HOME)/bin/spark-submit \
    --master $(MASTER) --conf spark.ui.port=4048 --verbose \
    $(TFoS_HOME)/examples/mnist/mnist_data_setup.py \
    --output examples/mnist/csv \
    --format csv


查看处理过的数据集：


    hadoop fs -ls hdfs://localhost:9000/user/libinggen/examples/mnist/csv


查看保存图片、标记向量：


    hadoop fs -ls hdfs://localhost:9000/user/libinggen/examples/mnist/csv/train/labels


把训练集、测试集分别保存RDD数据。
/blob/master/examples/mnist/mnist_data_setup.py 。


    from __future__ import absolute_import
    from __future__ import division
    from __future__ import print_function
    import numpy
    import tensorflow as tf
    from array import array
    from tensorflow.contrib.learn.python.learn.datasets import mnist
    def toTFExample(image, label):
      """Serializes an image/label as a TFExample byte string"""
      example = tf.train.Example(
        features = tf.train.Features(
          feature = {
            'label': tf.train.Feature(int64_list=tf.train.Int64List(value=label.astype("int64"))),
            'image': tf.train.Feature(int64_list=tf.train.Int64List(value=image.astype("int64")))
          }
        )
      )
      return example.SerializeToString()
    def fromTFExample(bytestr):
      """Deserializes a TFExample from a byte string"""
      example = tf.train.Example()
      example.ParseFromString(bytestr)
      return example
    def toCSV(vec):
      """Converts a vector/array into a CSV string"""
      return ','.join([str(i) for i in vec])
    def fromCSV(s):
      """Converts a CSV string to a vector/array"""
      return [float(x) for x in s.split(',') if len(s) > 0]
    def writeMNIST(sc, input_images, input_labels, output, format, num_partitions):
      """Writes MNIST image/label vectors into parallelized files on HDFS"""
      # load MNIST gzip into memory
      # MNIST图像、标记向量写入HDFS
      with open(input_images, 'rb') as f:
        images = numpy.array(mnist.extract_images(f))
      with open(input_labels, 'rb') as f:
        if format == "csv2":
          labels = numpy.array(mnist.extract_labels(f, one_hot=False))
        else:
          labels = numpy.array(mnist.extract_labels(f, one_hot=True))
      shape = images.shape
      print("images.shape: {0}".format(shape))          # 60000 x 28 x 28
      print("labels.shape: {0}".format(labels.shape))   # 60000 x 10
      # create RDDs of vectors
      imageRDD = sc.parallelize(images.reshape(shape[0], shape[1] * shape[2]), num_partitions)
      labelRDD = sc.parallelize(labels, num_partitions)
      output_images = output + "/images"
      output_labels = output + "/labels"
      # save RDDs as specific format
      # RDDs保存特定格式
      if format == "pickle":
        imageRDD.saveAsPickleFile(output_images)
        labelRDD.saveAsPickleFile(output_labels)
      elif format == "csv":
        imageRDD.map(toCSV).saveAsTextFile(output_images)
        labelRDD.map(toCSV).saveAsTextFile(output_labels)
      elif format == "csv2":
        imageRDD.map(toCSV).zip(labelRDD).map(lambda x: str(x[1]) + "|" + x[0]).saveAsTextFile(output)
      else: # format == "tfr":
        tfRDD = imageRDD.zip(labelRDD).map(lambda x: (bytearray(toTFExample(x[0], x[1])), None))
        # requires: --jars tensorflow-hadoop-1.0-SNAPSHOT.jar
        tfRDD.saveAsNewAPIHadoopFile(output, "org.tensorflow.hadoop.io.TFRecordFileOutputFormat",
                                    keyClass="org.apache.hadoop.io.BytesWritable",
                                    valueClass="org.apache.hadoop.io.NullWritable")
    #  Note: this creates TFRecord files w/o requiring a custom Input/Output format
    #  else: # format == "tfr":
    #    def writeTFRecords(index, iter):
    #      output_path = "{0}/part-{1:05d}".format(output, index)
    #      writer = tf.python_io.TFRecordWriter(output_path)
    #      for example in iter:
    #        writer.write(example)
    #      return [output_path]
    #    tfRDD = imageRDD.zip(labelRDD).map(lambda x: toTFExample(x[0], x[1]))
    #    tfRDD.mapPartitionsWithIndex(writeTFRecords).collect()
    def readMNIST(sc, output, format):
      """Reads/verifies previously created output"""
      output_images = output + "/images"
      output_labels = output + "/labels"
      imageRDD = None
      labelRDD = None
      if format == "pickle":
        imageRDD = sc.pickleFile(output_images)
        labelRDD = sc.pickleFile(output_labels)
      elif format == "csv":
        imageRDD = sc.textFile(output_images).map(fromCSV)
        labelRDD = sc.textFile(output_labels).map(fromCSV)
      else: # format.startswith("tf"):
        # requires: --jars tensorflow-hadoop-1.0-SNAPSHOT.jar
        tfRDD = sc.newAPIHadoopFile(output, "org.tensorflow.hadoop.io.TFRecordFileInputFormat",
                                  keyClass="org.apache.hadoop.io.BytesWritable",
                                  valueClass="org.apache.hadoop.io.NullWritable")
        imageRDD = tfRDD.map(lambda x: fromTFExample(str(x[0])))
      num_images = imageRDD.count()
      num_labels = labelRDD.count() if labelRDD is not None else num_images
      samples = imageRDD.take(10)
      print("num_images: ", num_images)
      print("num_labels: ", num_labels)
      print("samples: ", samples)
    if __name__ == "__main__":
      import argparse
      from pyspark.context import SparkContext
      from pyspark.conf import SparkConf
      parser = argparse.ArgumentParser()
      parser.add_argument("-f", "--format", help="output format", choices=["csv","csv2","pickle","tf","tfr"], default="csv")
      parser.add_argument("-n", "--num-partitions", help="Number of output partitions", type=int, default=10)
      parser.add_argument("-o", "--output", help="HDFS directory to save examples in parallelized format", default="mnist_data")
      parser.add_argument("-r", "--read", help="read previously saved examples", action="store_true")
      parser.add_argument("-v", "--verify", help="verify saved examples after writing", action="store_true")
  args = parser.parse_args()
      print("args:",args)
      sc = SparkContext(conf=SparkConf().setAppName("mnist_parallelize"))
      if not args.read:
        # Note: these files are inside the mnist.zip file
        writeMNIST(sc, "mnist/train-images-idx3-ubyte.gz", "mnist/train-labels-idx1-ubyte.gz", args.output + "/train", args.format, args.num_partitions)
        writeMNIST(sc, "mnist/t10k-images-idx3-ubyte.gz", "mnist/t10k-labels-idx1-ubyte.gz", args.output + "/test", args.format, args.num_partitions)
      if args.read or args.verify:
        readMNIST(sc, args.output + "/train", args.format)


提交训练任务，开始训练，在HDFS生成mnist_model，命令：


    ${SPARK_HOME}/bin/spark-submit \
    --master ${MASTER} \
    --py-files ${TFoS_HOME}/examples/mnist/spark/mnist_dist.py \
    --conf spark.cores.max=${TOTAL_CORES} \
    --conf spark.task.cpus=${CORES_PER_WORKER} \
    --conf spark.executorEnv.JAVA_HOME="$JAVA_HOME" \
    ${TFoS_HOME}/examples/mnist/spark/mnist_spark.py \
    --cluster_size ${SPARK_WORKER_INSTANCES} \
    --images examples/mnist/csv/train/images \
    --labels examples/mnist/csv/train/labels \
    --format csv \
    --mode train \
    --model mnist_model


mnist_dist.py 构建TensorFlow 分布式任务，定义分布式任务主函数，启动TensorFlow主函数map_fun，数据获取方式Feeding。获取TensorFlow集群和服务器实例：


    cluster, server = TFNode.start_cluster_server(ctx, 1, args.rdma)


TFNode调用tfspark.zip TFNode.py文件。


mnist_spark.py文件是训练主程序，TensorFlowOnSpark部署步骤：


    from __future__ import absolute_import
    from __future__ import division
    from __future__ import print_function
    from pyspark.context import SparkContext
    from pyspark.conf import SparkConf
    import argparse
    import os
    import numpy
    import sys
    import tensorflow as tf
    import threading
    import time
    from datetime import datetime
    from tensorflowonspark import TFCluster
    import mnist_dist
    sc = SparkContext(conf=SparkConf().setAppName("mnist_spark"))
    executors = sc._conf.get("spark.executor.instances")
    num_executors = int(executors) if executors is not None else 1
    num_ps = 1
    parser = argparse.ArgumentParser()
    parser.add_argument("-b", "--batch_size", help="number of records per batch", type=int, default=100)
    parser.add_argument("-e", "--epochs", help="number of epochs", type=int, default=1)
    parser.add_argument("-f", "--format", help="example format: (csv|pickle|tfr)", choices=["csv","pickle","tfr"], default="csv")
    parser.add_argument("-i", "--images", help="HDFS path to MNIST images in parallelized format")
    parser.add_argument("-l", "--labels", help="HDFS path to MNIST labels in parallelized format")
    parser.add_argument("-m", "--model", help="HDFS path to save/load model during train/inference", default="mnist_model")
    parser.add_argument("-n", "--cluster_size", help="number of nodes in the cluster", type=int, default=num_executors)
    parser.add_argument("-o", "--output", help="HDFS path to save test/inference output", default="predictions")
    parser.add_argument("-r", "--readers", help="number of reader/enqueue threads", type=int, default=1)
    parser.add_argument("-s", "--steps", help="maximum number of steps", type=int, default=1000)
    parser.add_argument("-tb", "--tensorboard", help="launch tensorboard process", action="store_true")
    parser.add_argument("-X", "--mode", help="train|inference", default="train")
    parser.add_argument("-c", "--rdma", help="use rdma connection", default=False)
    args = parser.parse_args()
    print("args:",args)
    print("{0} ===== Start".format(datetime.now().isoformat()))
    if args.format == "tfr":
      images = sc.newAPIHadoopFile(args.images, "org.tensorflow.hadoop.io.TFRecordFileInputFormat",
                                  keyClass="org.apache.hadoop.io.BytesWritable",
                                  valueClass="org.apache.hadoop.io.NullWritable")
      def toNumpy(bytestr):
        example = tf.train.Example()
        example.ParseFromString(bytestr)
        features = example.features.feature
        image = numpy.array(features['image'].int64_list.value)
        label = numpy.array(features['label'].int64_list.value)
        return (image, label)
      dataRDD = images.map(lambda x: toNumpy(str(x[0])))
    else:
      if args.format == "csv":
        images = sc.textFile(args.images).map(lambda ln: [int(x) for x in ln.split(',')])
        labels = sc.textFile(args.labels).map(lambda ln: [float(x) for x in ln.split(',')])
      else: # args.format == "pickle":
        images = sc.pickleFile(args.images)
        labels = sc.pickleFile(args.labels)
      print("zipping images and labels")
      dataRDD = images.zip(labels)
    #1.为在Executor执行每个TensorFlow进程保留一个端口
    cluster = TFCluster.run(sc, mnist_dist.map_fun, args, args.cluster_size, num_ps, args.tensorboard, TFCluster.InputMode.SPARK)
    #2.启动Tensorflow主函数
    cluster.start(mnist_dist.map_fun, args)
    if args.mode == "train":
      #3.训练
      cluster.train(dataRDD, args.epochs)
    else:
      #3.预测
      labelRDD = cluster.inference(dataRDD)
      labelRDD.saveAsTextFile(args.output)
    #4.关闭Executor TensorFlow计算节点、参数服务节点
    cluster.shutdown()
    print("{0} ===== Stop".format(datetime.now().isoformat()))


预测命令：


    ${SPARK_HOME}/bin/spark-submit \
    --master ${MASTER} \
    --py-files ${TFoS_HOME}/examples/mnist/spark/mnist_dist.py \
    --conf spark.cores.max=${TOTAL_CORES} \
    --conf spark.task.cpus=${CORES_PER_WORKER} \
    --conf spark.executorEnv.JAVA_HOME="$JAVA_HOME" \
    ${TFoS_HOME}/examples/mnist/spark/mnist_spark.py \
    --cluster_size ${SPARK_WORKER_INSTANCES} \
    --images examples/mnist/csv/test/images \
    --labels examples/mnist/csv/test/labels \
    --mode inference \
    --format csv \
    --model mnist_model \
    --output predictions


还可以Amazon EC2运行及在Hadoop集群采用YARN模式运行。


参考资料：
《TensorFlow技术解析与实战》


欢迎推荐上海机器学习工作机会，我的微信：qingxingfengzi