Tensorflowonspark 读取SparkRDD为输入 并进行批次训练

要点:

基于tensroflowonspark实现基础的回归分析
数据的输入来自spark RDD
batch训练

代码

主程序代码main.py

from pyspark.context import SparkContext
from pyspark.conf import SparkConf
from tensorflowonspark import TFCluster,TFNode
from pyspark.sql import SparkSession
import tensor_test
if __name__ == "__main__":
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument("--rdma",help="use rdma connection",default=False)
    args = parser.parse_args()
    spark = SparkSession.builder \
            .getOrCreate()
    #conf = SparkConf().setAppName("ceshi")
    #sc = SparkContext(conf=conf)
    sc = spark.sparkContext
    sc.setLogLevel("WARN")
    import numpy as np
    x_data = np.linspace(-1,1,300)[:,np.newaxis]
    noise = np.random.normal(0,0.05,x_data.shape)
    y_data = np.square(x_data)  + noise
    x_data_rdd = sc.parallelize(x_data)
    y_data_rdd = sc.parallelize(y_data)
    in_rdd = x_data_rdd.zip(y_data_rdd)
    num_executors = 4
    #num_executors = int(sc._conf.get("spark.executor.instances"))
    #num_executors = int(executors) if executors is not None else 1
    tensorboard = False
    num_ps = 1
    cluster = TFCluster.run(sc,tensor_test.main_func,args,num_executors,num_ps,tensorboard,TFCluster.InputMode.SPARK)
    cluster.train(in_rdd,1)   #1 代表epochs
    print ("Done===")
    cluster.shutdown()

自定义Tensorflow任务tensor_test.py

def main_func(args,ctx):
    import numpy as np
    import  tensorflow as tf
    import sys
    job_name = ctx.job_name
    task_index = ctx.task_index
    cluster,server = ctx.start_cluster_server(1,args.rdma)
    tf_feed = ctx.get_data_feed()
    def add_layer(inputs,insize,outsize,activation_func=None):
      Weights = tf.Variable(tf.random_normal([insize,outsize]))
      bias = tf.Variable(tf.zeros([1,outsize])+0.1)
      wx_plus_b = tf.matmul(inputs,Weights) + bias
      if activation_func:
          return activation_func(wx_plus_b)
      else:
          return wx_plus_b
    def rdd_generator():
        while not tf_feed.should_stop():
            batch = tf_feed.next_batch(1)
            if len(batch) == 0:
                return
            row = batch[0]
            x = np.array(row[0]).astype(np.float32)
            y = np.array(row[1]).astype(np.int64)
            yield (x,y)
    #x_data = np.linspace(-1,1,300)[:,np.newaxis]
    #noise = np.random.normal(0,0.05,x_data.shape)
    #y_data = np.square(x_data)  + noise
    if job_name == "ps":
        server.join()
    elif job_name == "worker":
        ds = tf.data.Dataset.from_generator(rdd_generator,(tf.float32,tf.float32)).batch(100)
        iterator = ds.make_one_shot_iterator()
        x,y_ = iterator.get_next()
        l1 = add_layer(x,1,10,activation_func=tf.nn.relu)
        preds = add_layer(l1,10,1,activation_func=None)
        
        global_step = tf.train.get_or_create_global_step()
        loss = tf.reduce_mean(tf.reduce_sum(tf.square(y_ - preds),reduction_indices=[1]))
        train = tf.train.GradientDescentOptimizer(0.05).minimize(loss,global_step=global_step)
        #Test trained labels
        saver = tf.train.Saver()
        init_op = tf.global_variables_initializer()
    logdir = ctx.absolute_path("/home/devops/test/TensorFlowOnSpark/examples/mnist/my/curve/log")
    #hooks = [tf.train.StopAtStepHook(last_step=2)]
    hooks = []
    with tf.train.MonitoredTrainingSession() as sess:
        sess.run(init_op)
        step = 0
        while not sess.should_stop() and not tf_feed.should_stop():
            _,preds_val,step =sess.run([train,preds,global_step])
            #if (step % 1 == 0) and (not sess.should_stop()):
    print ("{} step of Values of predictions are:{}".format(step,preds_val))
    if step >= 10 or sess.should_stop():
        tf_feed.terminate()

「点点赞赏,手留余香」

    还没有人赞赏,快来当第一个赞赏的人吧!