mnist_spark.py

# Adapted from: https://www.tensorflow.org/beta/tutorials/distribute/multi_worker_with_keras

from __future__ import absolute_import, division, print_function, unicode_literals


def main_fun(args, ctx):
  import numpy as np
  import tensorflow as tf
  from tensorflowonspark import TFNode

  strategy = tf.distribute.experimental.MultiWorkerMirroredStrategy()

  def build_and_compile_cnn_model():
    model = tf.keras.Sequential([
        tf.keras.layers.Conv2D(32, 3, activation='relu', input_shape=(28, 28, 1)),
        tf.keras.layers.MaxPooling2D(),
        tf.keras.layers.Flatten(),
        tf.keras.layers.Dense(64, activation='relu'),
        tf.keras.layers.Dense(10, activation='softmax')
    ])
    model.compile(
        loss=tf.keras.losses.sparse_categorical_crossentropy,
        optimizer=tf.keras.optimizers.SGD(learning_rate=0.001),
        metrics=['accuracy'])
    return model

  # single node
  # single_worker_model = build_and_compile_cnn_model()
  # single_worker_model.fit(x=train_datasets, epochs=3)

  tf_feed = TFNode.DataFeed(ctx.mgr, False)

  def rdd_generator():
    while not tf_feed.should_stop():
      batch = tf_feed.next_batch(1)
      if len(batch) > 0:
        example = batch[0]
        image = np.array(example[0]).astype(np.float32) / 255.0
        image = np.reshape(image, (28, 28, 1))
        label = np.array(example[1]).astype(np.float32)
        label = np.reshape(label, (1,))
        yield (image, label)
      else:
        return

  ds = tf.data.Dataset.from_generator(rdd_generator, (tf.float32, tf.float32), (tf.TensorShape([28, 28, 1]), tf.TensorShape([1])))
  ds = ds.batch(args.batch_size)

  # this fails
  # callbacks = [tf.keras.callbacks.ModelCheckpoint(filepath=args.model_dir)]
  tf.io.gfile.makedirs(args.model_dir)
  filepath = args.model_dir + "/weights-{epoch:04d}"
  callbacks = [tf.keras.callbacks.ModelCheckpoint(filepath=filepath, verbose=1, save_weights_only=True)]

  with strategy.scope():
    multi_worker_model = build_and_compile_cnn_model()

  # Note: MultiWorkerMirroredStrategy (CollectiveAllReduceStrategy) is synchronous,
  # so we need to ensure that all workers complete training before any of them run out of data from the RDD.
  # And given that Spark RDD partitions (and partition sizes) can be non-evenly divisible by num_workers,
  # we'll just stop training at 90% of the total expected number of steps.
  steps_per_epoch = 60000 / args.batch_size
  steps_per_epoch_per_worker = steps_per_epoch / ctx.num_workers
  max_steps_per_worker = steps_per_epoch_per_worker * 0.9

  multi_worker_model.fit(x=ds, epochs=args.epochs, steps_per_epoch=max_steps_per_worker, callbacks=callbacks)

  from tensorflow_estimator.python.estimator.export import export_lib
  export_dir = export_lib.get_timestamped_export_dir(args.export_dir)
  multi_worker_model.save(export_dir, save_format='tf')

  # terminating feed tells spark to skip processing further partitions
  tf_feed.terminate()


if __name__ == '__main__':
  import argparse
  from pyspark.context import SparkContext
  from pyspark.conf import SparkConf
  from tensorflowonspark import TFCluster

  sc = SparkContext(conf=SparkConf().setAppName("mnist_keras"))
  executors = sc._conf.get("spark.executor.instances")
  num_executors = int(executors) if executors is not None else 1

  parser = argparse.ArgumentParser()
  parser.add_argument("--batch_size", help="number of records per batch", type=int, default=64)
  parser.add_argument("--cluster_size", help="number of nodes in the cluster", type=int, default=num_executors)
  parser.add_argument("--epochs", help="number of epochs", type=int, default=3)
  parser.add_argument("--images_labels", help="path to MNIST images and labels in parallelized format")
  parser.add_argument("--model_dir", help="path to save checkpoint", default="mnist_model")
  parser.add_argument("--export_dir", help="path to export saved_model", default="mnist_export")
  parser.add_argument("--tensorboard", help="launch tensorboard process", action="store_true")

  args = parser.parse_args()
  print("args:", args)

  # create RDD of input data
  def parse(ln):
    vec = [int(x) for x in ln.split(',')]
    return (vec[1:], vec[0])

  images_labels = sc.textFile(args.images_labels).map(parse)

  cluster = TFCluster.run(sc, main_fun, args, args.cluster_size, num_ps=0, tensorboard=args.tensorboard, input_mode=TFCluster.InputMode.SPARK, master_node='chief')
  # Note: need to feed extra data to ensure that each worker receives sufficient data to complete epochs
  # to compensate for variability in partition sizes and spark scheduling
  cluster.train(images_labels, args.epochs)
  cluster.shutdown()