datasets's Introduction

TensorFlow Datasets

Note: tensorflow_datasets is not yet released. Follow the release tracking issue to be notified of release.

TensorFlow Datasets provides many public datasets as tf.data.Datasets.

Installation

pip install tensorflow-datasets

# Requires tensorflow or tensorflow-gpu to be installed
# Some datasets require additional libraries; see setup.py extras_require

Usage

import tensorflow_datasets as tfds

# See available datasets
print(tfds.list_builders())

# Construct a tf.data.Dataset
dataset = tfds.load(name="mnist",
                    split=tfds.Split.TRAIN,
                    data_dir="~/tfdata",
                    download=True)

# Build your input pipeline
dataset = dataset.shuffle(1000).batch(128).prefetch(1)
features = dataset.make_one_shot_iterator().get_next()
image, label = features["input"], features["target"]

`DatasetBuilder`

All datasets are implemented as subclasses of DatasetBuilder.

import tensorflow_datasets as tfds

# The following is the equivalent of the `load` call above.

# You can fetch the DatasetBuilder class by string
mnist_builder = tfds.builder("mnist")(data_dir="~/tfdata")

# Download the dataset
mnist_builder.download_and_prepare()
# Construct a tf.data.Dataset
dataset = mnist_builder.as_dataset(split=tfds.Split.TRAIN)

Non-TensorFlow Usage

All datasets are usable outside of TensorFlow with the numpy_iterator method, which takes the same arguments as as_dataset.

import tensorflow_datasets as tfds

mnist_builder = tfds.builder("mnist")(data_dir="~/tfdata")
mnist_builder.download_and_prepare()
for element in mnist_builder.numpy_iterator(split=tfds.Split.TRAIN):
  numpy_image, numpy_label = element["input"], element["target"]