Spark TensorFlow Distributor

This package helps users do distributed training with TensorFlow on their Spark clusters.

Installation

This package requires Python 3.6+, tensorflow>=2.1.0 and pyspark>=3.0.0 to run. To install spark-tensorflow-distributor, run:

pip install spark-tensorflow-distributor

The installation does not install PySpark because for most users, PySpark is already installed. In addition, tensorflow not installed so that users may choose between regular and CPU-only installation via pip install tensorflow and pip install tensorflow-cpu. If you do not have PySpark installed, you can install it directly:

pip install pyspark>=3.0.*

Note also that in order to use many features of this package, you must set up Spark custom resource scheduling for GPUs on your cluster. See the Spark docs for this.

Running Tests

For integration tests, first build the master and worker images and then run the test script.

docker-compose build --build-arg PYTHON_INSTALL_VERSION=3.7
./tests/integration/run.sh

For linting, run the following.

./tests/lint.sh

To use the autoformatter, run the following.

yapf --recursive --in-place spark_tensorflow_distributor

Examples

Run following example code in pyspark shell:

from spark_tensorflow_distributor import MirroredStrategyRunner

# Adapted from https://www.tensorflow.org/tutorials/distribute/multi_worker_with_keras
def train():
    import tensorflow as tf
    import uuid

    BUFFER_SIZE = 10000
    BATCH_SIZE = 64

    def make_datasets():
        (mnist_images, mnist_labels), _ = \
            tf.keras.datasets.mnist.load_data(path=str(uuid.uuid4())+'mnist.npz')

        dataset = tf.data.Dataset.from_tensor_slices((
            tf.cast(mnist_images[..., tf.newaxis] / 255.0, tf.float32),
            tf.cast(mnist_labels, tf.int64))
        )
        dataset = dataset.repeat().shuffle(BUFFER_SIZE).batch(BATCH_SIZE)
        return dataset

    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

    train_datasets = make_datasets()
    options = tf.data.Options()
    options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.DATA
    train_datasets = train_datasets.with_options(options)
    multi_worker_model = build_and_compile_cnn_model()
    multi_worker_model.fit(x=train_datasets, epochs=3, steps_per_epoch=5)

MirroredStrategyRunner(num_slots=8).run(train)