Pix3D is a large-scale dataset of diverse image-shape pairs with pixel-level 2D-3D alignment.

file_format EXPERIMENTAL, may change at any time; Format of the record files in which dataset will be read/written to. If None, defaults to tfrecord.
**kwargs Arguments passed to DatasetBuilder.

builder_config tfds.core.BuilderConfig for this builder.



info tfds.core.DatasetInfo for this builder.



versions Versions (canonical + availables), in preference order.



Constructs a

Callers must pass arguments as keyword arguments.

The output types vary depending on the parameters. Examples:

builder = tfds.builder('imdb_reviews')

# Default parameters: Returns the dict of
ds_all_dict = builder.as_dataset()
assert isinstance(ds_all_dict, dict)
print(ds_all_dict.keys())  # ==> ['test', 'train', 'unsupervised']

assert isinstance(ds_all_dict['test'],
# Each dataset (test, train, unsup.) consists of dictionaries
# {'label': <tf.Tensor: .. dtype=int64, numpy=1>,
#  'text': <tf.Tensor: .. dtype=string, numpy=b"I've watched the movie ..">}
# {'label': <tf.Tensor: .. dtype=int64, numpy=1>,
#  'text': <tf.Tensor: .. dtype=string, numpy=b'If you love Japanese ..'>}

# With as_supervised: only contains (feature, label) tuples
ds_all_supervised = builder.as_dataset(as_supervised=True)
assert isinstance(ds_all_supervised, dict)
print(ds_all_supervised.keys())  # ==> ['test', 'train', 'unsupervised']

assert isinstance(ds_all_supervised['test'],
# Each dataset (test, train, unsup.) consists of tuples (text, label)
# (<tf.Tensor: ... dtype=string, numpy=b"I've watched the movie ..">,
#  <tf.Tensor: ... dtype=int64, numpy=1>)
# (<tf.Tensor: ... dtype=string, numpy=b"If you love Japanese ..">,
#  <tf.Tensor: ... dtype=int64, numpy=1>)

# Same as above plus requesting a particular split
ds_test_supervised = builder.as_dataset(as_supervised=True, split='test')
assert isinstance(ds_test_supervised,
# The dataset consists of tuples (text, label)
# (<tf.Tensor: ... dtype=string, numpy=b"I've watched the movie ..">,
#  <tf.Tensor: ... dtype=int64, numpy=1>)
# (<tf.Tensor: ... dtype=string, numpy=b"If you love Japanese ..">,
#  <tf.Tensor: ... dtype=int64, numpy=1>)

split Which split of the data to load (e.g. 'train', 'test', ['train', 'test'], 'train[80%:]',...). See our split API guide. If None, will return all splits in a Dict[Split,].
batch_size int, batch size. Note that variable-length features will be 0-padded if batch_size is set. Users that want more custom behavior should use batch_size=None and use the API to construct a custom pipeline. If batch_size == -1, will return feature dictionaries of the whole dataset with tf.Tensors instead of a
shuffle_files bool, whether to shuffle the input files. Defaults to False.
decoders Nested dict of Decoder objects which allow to customize the decoding. The structure should match the feature structure, but only customized feature keys need to be present. See the guide for more info.
read_config tfds.ReadConfig, Additional options to configure the input pipeline (e.g. seed, num parallel reads,...).
as_supervised bool, if True, the returned will have a 2-tuple structure (input, label) according to If False, the default, the returned will have a dictionary with all the features.

Returns, or if split=None, dict<key: tfds.Split, value:>.

If batch_size is -1, will return feature dictionaries containing the entire dataset in tf.Tensors instead of a


Returns the DatasetInfo using given kwargs and config files.

Sub-class should call this and add information not present in config files using kwargs directly passed to tfds.core.DatasetInfo object.

If information is present both in passed arguments and config files, config files will prevail.

**kwargs kw args to pass to DatasetInfo directly.


Downloads and prepares dataset for reading.

download_dir directory where downloaded files are stored. Defaults to "~/tensorflow-datasets/downloads".
download_config, further configuration for downloading and preparing dataset.
file_format optional str or file_adapters.FileFormat, format of the record files in which the dataset will be written.

IOError if there is not enough disk space available.
RuntimeError when the config cannot be found.


Returns the default builder config if there is one.

Note that for dataset builders that cannot use the cls.BUILDER_CONFIGS, we need a method that uses the instance to get BUILDER_CONFIGS and DEFAULT_BUILDER_CONFIG_NAME.

the default builder config if there is one


Returns metadata (README, CITATIONS, ...) specified in config files.

The config files are read from the same package where the DatasetBuilder has been defined, so those metadata might be wrong for legacy builders.


Returns a reference to the dataset produced by this dataset builder.

Includes the config if specified, the version, and the data_dir that should contain this dataset.

namespace if this dataset is a community dataset, and therefore has a namespace, then the namespace must be provided such that it can be set in the reference. Note that a dataset builder is not aware that it is part of a namespace.

a reference to this instantiated builder.

CLASS_INDEX ['background', 'bed', 'bookcase', 'chair', 'desk', 'misc', 'sofa', 'table', 'tool', 'wardrobe']



TEST_SPLIT_IDX '/tmpfs/venv/lib/python3.9/site-packages/tensorflow_graphics/datasets/pix3d/splits/pix3d_test.npy'
TRAIN_SPLIT_IDX '/tmpfs/venv/lib/python3.9/site-packages/tensorflow_graphics/datasets/pix3d/splits/pix3d_train.npy'
VERSION Instance of tensorflow_datasets.core.utils.version.Version
builder_config_cls None



code_path Instance of etils.epath.gpath.PosixGPath
default_builder_config None
name 'pix3d'
pkg_dir_path None
url_infos None