MultiWorker Mirrored Strategy MNIST with TensorFlow
MultiWorkerMirroredStrategy on Hopsworks
Tested with TensorFlow 2.4.0
Machine Learning on Hopsworks
The hops
python module
hops
is a helper library for Hops that facilitates development by hiding the complexity of running applications and iteracting with services.
Have a feature request or encountered an issue? Please let us know on github.
Using the experiment
module
To be able to run your Machine Learning code in Hopsworks, the code for the whole program needs to be provided and put inside a wrapper function. Everything, from importing libraries to reading data and defining the model and running the program needs to be put inside a wrapper function.
The experiment
module provides an api to Python programs such as TensorFlow, Keras and PyTorch on a Hopsworks on any number of machines and GPUs.
An Experiment could be a single Python program, which we refer to as an Experiment.
Grid search or genetic hyperparameter optimization such as differential evolution which runs several Experiments in parallel, which we refer to as Parallel Experiment.
ParameterServerStrategy, CollectiveAllReduceStrategy and MultiworkerMirroredStrategy making multi-machine/multi-gpu training as simple as invoking a function for orchestration. This mode is referred to as Distributed Training.
Using the tensorboard
module
The tensorboard
module allow us to get the log directory for summaries and checkpoints to be written to the TensorBoard we will see in a bit. The only function that we currently need to call is tensorboard.logdir()
, which returns the path to the TensorBoard log directory. Furthermore, the content of this directory will be put in as a Dataset in your project’s Experiments folder.
The directory could in practice be used to store other data that should be accessible after the experiment is finished.
# Use this module to get the TensorBoard logdir
from hops import tensorboard
tensorboard_logdir = tensorboard.logdir()
Using the hdfs
module
The hdfs
module provides a method to get the path in HopsFS where your data is stored, namely by calling hdfs.project_path()
. The path resolves to the root path for your project, which is the view that you see when you click Data Sets
in HopsWorks. To point where your actual data resides in the project you to append the full path from there to your Dataset. For example if you create a mnist folder in your Resources Dataset, the path to the mnist data would be hdfs.project_path() + 'Resources/mnist'
# Use this module to get the path to your project in HopsFS, then append the path to your Dataset in your project
from hops import hdfs
project_path = hdfs.project_path()
# Downloading the mnist dataset to the current working directory
from hops import hdfs
mnist_hdfs_path = hdfs.project_path() + "Resources/mnist"
local_mnist_path = hdfs.copy_to_local(mnist_hdfs_path)
Documentation
See the following links to learn more about running experiments in Hopsworks
- Learn more about experiments
- Building End-To-End pipelines
- Give us a star, create an issue or a feature request on Hopsworks github
Managing experiments
Experiments service provides a unified view of all the experiments run using the experiment
module.
As demonstrated in the gif it provides general information about the experiment and the resulting metric. Experiments can be visualized meanwhile or after training in a TensorBoard.
def multi_worker_mirrored_training():
import sys
import numpy as np
import tensorflow as tf
from hops import tensorboard
from hops import devices
from hops import hdfs
import pydoop.hdfs as pydoop
log_dir = tensorboard.logdir()
# Define distribution strategy
strategy = tf.distribute.experimental.MultiWorkerMirroredStrategy() #communication=tf.distribute.experimental.CollectiveCommunication.NCCL
# Define per device batch size
batch_size_per_replica = 8
# Define global batch size
batch_size = batch_size_per_replica * strategy.num_replicas_in_sync
# Define model hyper parameters
epochs = 10
steps_per_epoch = 5
validation_steps = 2
kernel = 2
pool = 2
dropout = 0.01
num_classes = 10
# Input image dimensions
img_rows, img_cols = 28, 28
input_shape = (28, 28, 1)
train_filenames = [hdfs.project_path() + "TourData/mnist/train/train.tfrecords"]
validation_filenames = [hdfs.project_path() + "TourData/mnist/validation/validation.tfrecords"]
# Create an iterator over the dataset
def data_input(filenames, batch_size=128, shuffle=False, repeat=None):
def parser(serialized_example):
"""Parses a single tf.Example into image and label tensors."""
features = tf.io.parse_single_example(
serialized_example,
features={
'image_raw': tf.io.FixedLenFeature([], tf.string),
'label': tf.io.FixedLenFeature([], tf.int64),
})
image = tf.io.decode_raw(features['image_raw'], tf.uint8)
image.set_shape([28 * 28])
# Normalize the values of the image from the range [0, 255] to [-0.5, 0.5]
image = tf.cast(image, tf.float32) / 255 - 0.5
label = tf.cast(features['label'], tf.int32)
# Reshape the tensor
image = tf.reshape(image, [img_rows, img_cols, 1])
# Create a one hot array for your labels
label = tf.one_hot(label, num_classes)
return image, label
# Import MNIST data
dataset = tf.data.TFRecordDataset(filenames)
# Map the parser over dataset, and batch results by up to batch_size
dataset = dataset.map(parser)
if shuffle:
dataset = dataset.shuffle(buffer_size=128)
dataset = dataset.batch(batch_size, drop_remainder=True)
dataset = dataset.repeat(repeat)
options = tf.data.Options()
options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.OFF
return dataset.with_options(options)
# construct model under distribution strategy scope
with strategy.scope():
model = tf.keras.Sequential()
model.add(tf.keras.layers.Conv2D(32, kernel_size=(kernel, kernel),
activation='relu',
input_shape=input_shape))
model.add(tf.keras.layers.Conv2D(64, (kernel, kernel), activation='relu'))
model.add(tf.keras.layers.MaxPooling2D(pool_size=(pool, pool)))
model.add(tf.keras.layers.Dropout(dropout))
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(128, activation='relu'))
model.add(tf.keras.layers.Dropout(dropout))
model.add(tf.keras.layers.Dense(num_classes, activation='softmax'))
opt = tf.keras.optimizers.Adadelta(1.0)
model.compile(loss=tf.keras.losses.categorical_crossentropy,
optimizer=opt,
metrics=['accuracy'])
callbacks = [
tf.keras.callbacks.TensorBoard(log_dir=log_dir),
tf.keras.callbacks.ModelCheckpoint(filepath=log_dir),
]
model.fit(data_input(train_filenames, batch_size),
verbose=0,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
validation_data=data_input(validation_filenames, batch_size),
validation_steps=validation_steps,
callbacks=callbacks
)
score = model.evaluate(data_input(validation_filenames, batch_size), steps=1)
metrics = {'accuracy': score[1]}
return metrics
Starting Spark application
ID | YARN Application ID | Kind | State | Spark UI | Driver log |
---|---|---|---|---|---|
13 | application_1600264891477_0018 | pyspark | idle | Link | Link |
SparkSession available as 'spark'.
from hops import experiment
experiment.mirrored(multi_worker_mirrored_training, name='mnist model', metric_key='accuracy')
Finished Experiment
('hdfs://rpc.namenode.service.consul:8020/Projects/demo_deep_learning_admin000/Experiments/application_1600264891477_0018_1', {'accuracy': 0.5, 'log': 'Experiments/application_1600264891477_0018_1/chief_0_output.log'})