TensorFlow I/O

TensorFlow I/O is a collection of file systems and file formats that are not available in TensorFlow's built-in support. A full list of supported file systems and file formats by TensorFlow I/O can be found here.

The use of tensorflow-io is straightforward with keras. Below is an example to Get Started with TensorFlow with the data processing aspect replaced by tensorflow-io:

import tensorflow as tf
import tensorflow_io as tfio

# Read the MNIST data into the IODataset.
dataset_url = "https://storage.googleapis.com/cvdf-datasets/mnist/"
d_train = tfio.IODataset.from_mnist(
    dataset_url + "train-images-idx3-ubyte.gz",
    dataset_url + "train-labels-idx1-ubyte.gz",
)

# Shuffle the elements of the dataset.
d_train = d_train.shuffle(buffer_size=1024)

# By default image data is uint8, so convert to float32 using map().
d_train = d_train.map(lambda x, y: (tf.image.convert_image_dtype(x, tf.float32), y))

# prepare batches the data just like any other tf.data.Dataset
d_train = d_train.batch(32)

# Build the model.
model = tf.keras.models.Sequential(
    [
        tf.keras.layers.Flatten(input_shape=(28, 28)),
        tf.keras.layers.Dense(512, activation=tf.nn.relu),
        tf.keras.layers.Dropout(0.2),
        tf.keras.layers.Dense(10, activation=tf.nn.softmax),
    ]
)

# Compile the model.
model.compile(
    optimizer="adam", loss="sparse_categorical_crossentropy", metrics=["accuracy"]
)

# Fit the model.
model.fit(d_train, epochs=5, steps_per_epoch=200)

In the above MNIST example, the URL's to access the dataset files are passed directly to the tfio.IODataset.from_mnist API call. This is due to the inherent support that tensorflow-io provides for HTTP/HTTPS file system, thus eliminating the need for downloading and saving datasets on a local directory.

NOTE: Since tensorflow-io is able to detect and uncompress the MNIST dataset automatically if needed, we can pass the URL's for the compressed files (gzip) to the API call as is.

Please check the official documentation for more detailed and interesting usages of the package.

Installation

Python Package

The tensorflow-io Python package can be installed with pip directly using:

$ pip install tensorflow-io

People who are a little more adventurous can also try our nightly binaries:

$ pip install tensorflow-io-nightly

To ensure you have a version of TensorFlow that is compatible with TensorFlow-IO, you can specify the tensorflow extra requirement during install:

pip install tensorflow-io[tensorflow]

Similar extras exist for the tensorflow-gpu, tensorflow-cpu and tensorflow-rocm packages.

Docker Images

In addition to the pip packages, the docker images can be used to quickly get started.

For stable builds:

$ docker pull tfsigio/tfio:latest
$ docker run -it --rm --name tfio-latest tfsigio/tfio:latest

For nightly builds:

$ docker pull tfsigio/tfio:nightly
$ docker run -it --rm --name tfio-nightly tfsigio/tfio:nightly

R Package

Once the tensorflow-io Python package has been successfully installed, you can install the development version of the R package from GitHub via the following:

if (!require("remotes")) install.packages("remotes")
remotes::install_github("tensorflow/io", subdir = "R-package")

TensorFlow Version Compatibility

To ensure compatibility with TensorFlow, it is recommended to install a matching version of TensorFlow I/O according to the table below. You can find the list of releases here.

Performance Benchmarking

We use github-pages to document the results of API performance benchmarks. The benchmark job is triggered on every commit to master branch and facilitates tracking performance w.r.t commits.

Contributing

Tensorflow I/O is a community led open source project. As such, the project depends on public contributions, bug-fixes, and documentation. Please see:

• contribution guidelines for a guide on how to contribute.
• development doc for instructions on the development environment setup.
• tutorials for a list of tutorial notebooks and instructions on how to write one.

Build Status and CI

Build	Status
Linux CPU Python 2
Linux CPU Python 3
Linux GPU Python 2
Linux GPU Python 3

Because of manylinux2010 requirement, TensorFlow I/O is built with Ubuntu:16.04 + Developer Toolset 7 (GCC 7.3) on Linux. Configuration with Ubuntu 16.04 with Developer Toolset 7 is not exactly straightforward. If the system have docker installed, then the following command will automatically build manylinux2010 compatible whl package:

#!/usr/bin/env bash

ls dist/*
for f in dist/*.whl; do
  docker run -i --rm -v $PWD:/v -w /v --net=host quay.io/pypa/manylinux2010_x86_64 bash -x -e /v/tools/build/auditwheel repair --plat manylinux2010_x86_64 $f
done
sudo chown -R $(id -nu):$(id -ng) .
ls wheelhouse/*

It takes some time to build, but once complete, there will be python 3.5, 3.6, 3.7 compatible whl packages available in wheelhouse directory.

On macOS, the same command could be used. However, the script expects python in shell and will only generate a whl package that matches the version of python in shell. If you want to build a whl package for a specific python then you have to alias this version of python to python in shell. See .github/workflows/build.yml Auditwheel step for instructions how to do that.

Note the above command is also the command we use when releasing packages for Linux and macOS.

TensorFlow I/O uses both GitHub Workflows and Google CI (Kokoro) for continuous integration. GitHub Workflows is used for macOS build and test. Kokoro is used for Linux build and test. Again, because of the manylinux2010 requirement, on Linux whl packages are always built with Ubuntu 16.04 + Developer Toolset 7. Tests are done on a variatiy of systems with different python3 versions to ensure a good coverage:

TensorFlow I/O has integrations with many systems and cloud vendors such as Prometheus, Apache Kafka, Apache Ignite, Google Cloud PubSub, AWS Kinesis, Microsoft Azure Storage, Alibaba Cloud OSS etc.

We tried our best to test against those systems in our continuous integration whenever possible. Some tests such as Prometheus, Kafka, and Ignite are done with live systems, meaning we install Prometheus/Kafka/Ignite on CI machine before the test is run. Some tests such as Kinesis, PubSub, and Azure Storage are done through official or non-official emulators. Offline tests are also performed whenever possible, though systems covered through offine tests may not have the same level of coverage as live systems or emulators.

References for emulators:

• Official PubSub Emulator by Google Cloud for Cloud PubSub.
• Official Azurite Emulator by Azure for Azure Storage.
• None-official LocalStack emulator by LocalStack for AWS Kinesis.

Community

• SIG IO Google Group and mailing list: io@tensorflow.org
• SIG IO Monthly Meeting Notes
• Gitter room: tensorflow/sig-io

Additional Information

• Streaming Machine Learning with Tiered Storage and Without a Data Lake - Kai Waehner
• TensorFlow with Apache Arrow Datasets - Bryan Cutler
• How to build a custom Dataset for Tensorflow - Ivelin Ivanov
• TensorFlow on Apache Ignite - Anton Dmitriev

License

Apache License 2.0