Main classes

DatasetInfo

class datasets.DatasetInfo(description: str = <factory>, citation: str = <factory>, homepage: str = <factory>, license: str = <factory>, features: Optional[datasets.features.Features] = None, post_processed: Optional[datasets.info.PostProcessedInfo] = None, supervised_keys: Optional[datasets.info.SupervisedKeysData] = None, builder_name: Optional[str] = None, config_name: Optional[str] = None, version: Optional[Union[str, datasets.utils.version.Version]] = None, splits: Optional[dict] = None, download_checksums: Optional[dict] = None, download_size: Optional[int] = None, post_processing_size: Optional[int] = None, dataset_size: Optional[int] = None, size_in_bytes: Optional[int] = None)[source]

Information about a dataset.

DatasetInfo documents datasets, including its name, version, and features. See the constructor arguments and properties for a full list.

Note: Not all fields are known on construction and may be updated later.

classmethod from_directory(dataset_info_dir: dict) → datasets.info.DatasetInfo[source]

Create DatasetInfo from the JSON file in dataset_info_dir.

This function updates all the dynamically generated fields (num_examples, hash, time of creation,…) of the DatasetInfo.

This will overwrite all previous metadata.

Parameters

dataset_info_dirstr The directory containing the metadata file. This should be the root directory of a specific dataset version.

write_to_directory(dataset_info_dir)[source]

Write DatasetInfo as JSON to dataset_info_dir. Also save the license separately in LICENCE.

Dataset

The base class datasets.Dataset implements a Dataset backed by an Apache Arrow table.

class datasets.Dataset(arrow_table: pyarrow.lib.Table, data_files: Optional[List[dict]] = None, info: Optional[datasets.info.DatasetInfo] = None, split: Optional[datasets.splits.NamedSplit] = None, indices_table: Optional[pyarrow.lib.Table] = None, indices_data_files: Optional[List[dict]] = None, fingerprint: Optional[str] = None, inplace_history: Optional[List[dict]] = None)[source]

A Dataset backed by an Arrow table or Record Batch.

__getitem__(key: Union[int, slice, str]) → Union[Dict, List][source]

Can be used to index columns (by string names) or rows (by integer index or iterable of indices or bools)

__iter__()[source]

Iterate through the examples. If a formatting is set with datasets.Dataset.set_format() rows will be returned with the selected format.

__len__()[source]

Number of rows in the dataset

add_elasticsearch_index(column: str, index_name: Optional[str] = None, host: Optional[str] = None, port: Optional[int] = None, es_client: Optional[elasticsearch.Elasticsearch] = None, es_index_name: Optional[str] = None, es_index_config: Optional[dict] = None)[source]

Add a text index using ElasticSearch for fast retrieval. This is done in-place.

Parameters
  • column (str) – The column of the documents to add to the index.

  • index_name (Optional str) – The index_name/identifier of the index. This is the index name that is used to call datasets.Dataset.get_nearest_examples() or datasets.Dataset.search(). By default it corresponds to column.

  • documents (Union[List[str], datasets.Dataset]) – The documents to index. It can be a datasets.Dataset.

  • es_client (elasticsearch.Elasticsearch) – The elasticsearch client used to create the index.

  • es_index_name (Optional str) – The elasticsearch index name used to create the index.

  • es_index_config (Optional dict) – The configuration of the elasticsearch index. Default config is:

Config:

{
    "settings": {
        "number_of_shards": 1,
        "analysis": {"analyzer": {"stop_standard": {"type": "standard", " stopwords": "_english_"}}},
    },
    "mappings": {
        "properties": {
            "text": {
                "type": "text",
                "analyzer": "standard",
                "similarity": "BM25"
            },
        }
    },
}

Example:

es_client = elasticsearch.Elasticsearch()
ds = datasets.load_dataset('crime_and_punish', split='train')
ds.add_elasticsearch_index(column='line', es_client=es_client, es_index_name="my_es_index")
scores, retrieved_examples = ds.get_nearest_examples('line', 'my new query', k=10)
add_faiss_index(column: str, index_name: Optional[str] = None, device: Optional[int] = None, string_factory: Optional[str] = None, metric_type: Optional[int] = None, custom_index: Optional[faiss.Index] = None, train_size: Optional[int] = None, faiss_verbose: bool = False, dtype=<class 'numpy.float32'>)[source]

Add a dense index using Faiss for fast retrieval. By default the index is done over the vectors of the specified column. You can specify device if you want to run it on GPU (device must be the GPU index). You can find more information about Faiss here:

Parameters
  • column (str) – The column of the vectors to add to the index.

  • index_name (Optional str) – The index_name/identifier of the index. This is the index_name that is used to call datasets.Dataset.get_nearest_examples() or datasets.Dataset.search(). By default it corresponds to column.

  • device (Optional int) – If not None, this is the index of the GPU to use. By default it uses the CPU.

  • string_factory (Optional str) – This is passed to the index factory of Faiss to create the index. Default index class is IndexFlat.

  • metric_type (Optional int) – Type of metric. Ex: faiss.faiss.METRIC_INNER_PRODUCT or faiss.METRIC_L2.

  • custom_index (Optional faiss.Index) – Custom Faiss index that you already have instantiated and configured for your needs.

  • train_size (Optional int) – If the index needs a training step, specifies how many vectors will be used to train the index.

  • faiss_verbose (bool, defaults to False) – Enable the verbosity of the Faiss index.

  • dtype (data-type) – The dtype of the numpy arrays that are indexed. Default is np.float32.

Example:

ds = datasets.load_dataset('crime_and_punish', split='train')
ds_with_embeddings = ds.map(lambda example: {'embeddings': embed(example['line']}))
ds_with_embeddings.add_faiss_index(column='embeddings')
# query
scores, retrieved_examples = ds_with_embeddings.get_nearest_examples('embeddings', embed('my new query'), k=10)
# save index
ds_with_embeddings.save_faiss_index('embeddings', 'my_index.faiss')

ds = datasets.load_dataset('crime_and_punish', split='train')
# load index
ds.load_faiss_index('embeddings', 'my_index.faiss')
# query
scores, retrieved_examples = ds.get_nearest_examples('embeddings', embed('my new query'), k=10)
add_faiss_index_from_external_arrays(external_arrays: numpy.array, index_name: str, device: Optional[int] = None, string_factory: Optional[str] = None, metric_type: Optional[int] = None, custom_index: Optional[faiss.Index] = None, train_size: Optional[int] = None, faiss_verbose: bool = False, dtype=<class 'numpy.float32'>)[source]

Add a dense index using Faiss for fast retrieval. The index is created using the vectors of external_arrays. You can specify device if you want to run it on GPU (device must be the GPU index). You can find more information about Faiss here: - For string factory

Parameters
  • external_arrays (np.array) – If you want to use arrays from outside the lib for the index, you can set external_arrays. It will use external_arrays to create the Faiss index instead of the arrays in the given column.

  • index_name (str) – The index_name/identifier of the index. This is the index_name that is used to call datasets.Dataset.get_nearest_examples() or datasets.Dataset.search().

  • device (Optional int) – If not None, this is the index of the GPU to use. By default it uses the CPU.

  • string_factory (Optional str) – This is passed to the index factory of Faiss to create the index. Default index class is IndexFlat.

  • metric_type (Optional int) – Type of metric. Ex: faiss.faiss.METRIC_INNER_PRODUCT or faiss.METRIC_L2.

  • custom_index (Optional faiss.Index) – Custom Faiss index that you already have instantiated and configured for your needs.

  • train_size (Optional int) – If the index needs a training step, specifies how many vectors will be used to train the index.

  • faiss_verbose (bool, defaults to False) – Enable the verbosity of the Faiss index.

  • dtype (numpy.dtype) – The dtype of the numpy arrays that are indexed. Default is np.float32.

property cache_files

The cache file containing the Apache Arrow table backing the dataset.

cast_(features: datasets.features.Features)[source]

Cast the dataset to a new set of features.

You can also remove a column using Dataset.map() with feature but cast_() is in-place (doesn’t copy the data to a new dataset) and is thus faster.

Parameters

features (datasets.Features) – New features to cast the dataset to. The name of the fields in the features must match the current column names. The type of the data must also be convertible from one type to the other. For non-trivial conversion, e.g. string <-> ClassLabel you should use map() to update the Dataset.

cleanup_cache_files()[source]

Clean up all cache files in the dataset cache directory, excepted the currently used cache file if there is one. Be carefull when running this command that no other process is currently using other cache files.

Returns

Number of removed files

property column_names

Names of the columns in the dataset.

property data

The Apache Arrow table backing the dataset.

drop_index(index_name: str)

Drop the index with the specified column.

Parameters

index_name (str) – The index_name/identifier of the index.

export(filename: str, format: str = 'tfrecord')[source]

Writes the Arrow dataset to a TFRecord file.

The dataset must already be in tensorflow format. The records will be written with keys from dataset._format_columns.

Parameters
  • filename (str) – The filename, including the .tfrecord extension, to write to.

  • (Optional[str], default (format) – “tfrecord”): The type of output file. Currently this is a no-op, as TFRecords are the only option. This enables a more flexible function signature later.

filter(function: Optional[Callable] = None, with_indices=False, input_columns: Optional[Union[str, List[str]]] = None, batch_size: Optional[int] = 1000, remove_columns: Optional[List[str]] = None, keep_in_memory: bool = False, load_from_cache_file: bool = True, cache_file_name: Optional[str] = None, writer_batch_size: Optional[int] = 1000, fn_kwargs: Optional[dict] = None, num_proc: Optional[int] = None, suffix_template: str = '_{rank:05d}_of_{num_proc:05d}', new_fingerprint: Optional[str] = None) → datasets.arrow_dataset.Dataset[source]

Apply a filter function to all the elements in the table in batches and update the table so that the dataset only includes examples according to the filter function.

Parameters
  • function (callable) – with one of the following signature: - function(example: Union[Dict, Any]) -> bool if with_indices=False - function(example: Union[Dict, Any], indices: int) -> bool if with_indices=True If no function is provided, default to an always True function: lambda x: True

  • with_indices (bool, defaults to False) – Provide example indices to function. Note that in this case the signature of function should be def function(example, idx): ….

  • input_columns (Optional[Union[str, List[str]]], defaults to None) – The columns to be passed into function as positional arguments. If None, a dict mapping to all formatted columns is passed as one argument.

  • batch_size (Optional[int], defaults to 1000) – Number of examples per batch provided to function if batched=True batch_size <= 0 or batch_size == None: Provide the full dataset as a single batch to function

  • remove_columns (Optional[List[str]], defaults to None) – Remove a selection of columns while doing the mapping. Columns will be removed before updating the examples with the output of function, i.e. if function is adding columns with names in remove_columns, these columns will be kept.

  • keep_in_memory (bool, defaults to False) – Keep the dataset in memory instead of writing it to a cache file.

  • load_from_cache_file (bool, defaults to True) – If a cache file storing the current computation from function can be identified, use it instead of recomputing.

  • cache_file_name (Optional[str], defaults to None) – Provide the name of a cache file to use to store the results of the computation instead of the automatically generated cache file name.

  • writer_batch_size (int, defaults to 1000) – Number of rows per write operation for the cache file writer. Higher value gives smaller cache files, lower value consume less temporary memory while running .map().

  • fn_kwargs (Optional[Dict], defaults to None) – Keyword arguments to be passed to function

  • num_proc (Optional[int], defaults to None) – Number of processes for multiprocessing. By default it doesn’t use multiprocessing.

  • (str, defaults to "_{rank (suffix_template) – 05d}_of_{num_proc:05d}”): If cache_file_name is specified, then this suffix will be added at the end of the base name of each. For example, if cache_file_name is “processed.arrow”, then for rank=1 and num_proc=4, the resulting file would be “processed_00001_of_00004.arrow” for the default suffix.

  • new_fingerprint (Optional[str], defaults to None) – the new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments

flatten_(max_depth=16)[source]

Flatten the Table. Each column with a struct type is flattened into one column per struct field. Other columns are left unchanged.

formatted_as(type: Optional[str] = None, columns: Optional[List] = None, output_all_columns: bool = False, **format_kwargs)[source]

To be used in a with statement. Set __getitem__ return format (type and columns)

Parameters
  • type (Optional str) – output type selected in [None, ‘numpy’, ‘torch’, ‘tensorflow’, ‘pandas’] None means __getitem__ returns python objects (default)

  • columns (Optional List[str]) – columns to format in the output None means __getitem__ returns all columns (default)

  • output_all_columns (bool default to False) – keep un-formatted columns as well in the output (as python objects)

  • format_kwargs – keywords arguments passed to the convert function like np.array, torch.tensor or tensorflow.ragged.constant.

classmethod from_buffer(buffer: pyarrow.lib.Buffer, info: Optional[datasets.info.DatasetInfo] = None, split: Optional[datasets.splits.NamedSplit] = None, indices_buffer: Optional[pyarrow.lib.Buffer] = None) → datasets.arrow_dataset.Dataset[source]

Instantiate a Dataset backed by an Arrow buffer

classmethod from_dict(mapping: dict, features: Optional[datasets.features.Features] = None, info: Optional[Any] = None, split: Optional[Any] = None) → datasets.arrow_dataset.Dataset[source]

Convert :obj:dict to a “obj”pyarrow.Table to create a :obj:datasets.Dataset.

Parameters
  • mapping (:obj:mapping) – A mapping of strings to Arrays or Python lists.

  • features (:obj:datasets.Features, optional, defaults to :obj:None) – If specified, the features types of the dataset

  • info (:obj:datasets.DatasetInfo, optional, defaults to :obj:None) – If specified, the dataset info containing info like description, citation, etc.

  • split (:obj:datasets.NamedSplit, optional, defaults to :obj:None) – If specified, the name of the dataset split.

classmethod from_file(filename: str, info: Optional[datasets.info.DatasetInfo] = None, split: Optional[datasets.splits.NamedSplit] = None, indices_filename: Optional[str] = None) → datasets.arrow_dataset.Dataset[source]

Instantiate a Dataset backed by an Arrow table at filename

classmethod from_pandas(df: pandas.core.frame.DataFrame, features: Optional[datasets.features.Features] = None, info: Optional[datasets.info.DatasetInfo] = None, split: Optional[datasets.splits.NamedSplit] = None) → datasets.arrow_dataset.Dataset[source]

Convert :obj:pandas.DataFrame to a “obj”pyarrow.Table to create a :obj:datasets.Dataset.

The column types in the resulting Arrow Table are inferred from the dtypes of the pandas.Series in the DataFrame. In the case of non-object Series, the NumPy dtype is translated to its Arrow equivalent. In the case of object, we need to guess the datatype by looking at the Python objects in this Series.

Be aware that Series of the object dtype don’t carry enough information to always lead to a meaningful Arrow type. In the case that we cannot infer a type, e.g. because the DataFrame is of length 0 or the Series only contains None/nan objects, the type is set to null. This behavior can be avoided by constructing explicit features and passing it to this function.

Parameters
  • df (:obj:pandas.DataFrame) – the dataframe that contains the dataset.

  • features (:obj:datasets.Features, optional, defaults to :obj:None) – If specified, the features types of the dataset

  • info (:obj:datasets.DatasetInfo, optional, defaults to :obj:None) – If specified, the dataset info containing info like description, citation, etc.

  • split (:obj:datasets.NamedSplit, optional, defaults to :obj:None) – If specified, the name of the dataset split.

get_index(index_name: str) → datasets.search.BaseIndex

List the index_name/identifiers of all the attached indexes.

get_nearest_examples(index_name: str, query: Union[str, numpy.array], k: int = 10) → datasets.search.NearestExamplesResults

Find the nearest examples in the dataset to the query.

Parameters
  • index_name (str) – The index_name/identifier of the index.

  • query (Union[str, np.ndarray]) – The query as a string if index_name is a text index or as a numpy array if index_name is a vector index.

  • k (int) – The number of examples to retrieve.

Ouput:

scores (List[float]): The retrieval scores of the retrieved examples. examples (dict): The retrieved examples.

get_nearest_examples_batch(index_name: str, queries: Union[List[str], numpy.array], k: int = 10) → datasets.search.BatchedNearestExamplesResults

Find the nearest examples in the dataset to the query.

Parameters
  • index_name (str) – The index_name/identifier of the index.

  • queries (Union[List[str], np.ndarray]) – The queries as a list of strings if index_name is a text index or as a numpy array if index_name is a vector index.

  • k (int) – The number of examples to retrieve per query.

Ouput:

total_scores (List[List[float]): The retrieval scores of the retrieved examples per query. total_examples (List[dict]): The retrieved examples per query.

property info

datasets.DatasetInfo object containing all the metadata in the dataset.

list_indexes() → List[str]

List the colindex_nameumns/identifiers of all the attached indexes.

load_faiss_index(index_name: str, file: str, device: Optional[int] = None)

Load a FaissIndex from disk. If you want to do additional configurations, you can have access to the faiss index object by doing .get_index(index_name).faiss_index to make it fit your needs

Parameters
  • index_name (str) – The index_name/identifier of the index. This is the index_name that is used to call .get_nearest or .search.

  • file (str) – The path to the serialized faiss index on disk.

  • device (Optional int) – If not None, this is the index of the GPU to use. By default it uses the CPU.

static load_from_disk(dataset_path: str) → datasets.arrow_dataset.Dataset[source]

Load the dataset from a dataset directory

Parameters

dataset_path (str) – path of the dataset directory where the dataset will be loaded from

map(function: Optional[Callable] = None, with_indices: bool = False, input_columns: Optional[Union[str, List[str]]] = None, batched: bool = False, batch_size: Optional[int] = 1000, drop_last_batch: bool = False, remove_columns: Optional[List[str]] = None, keep_in_memory: bool = False, load_from_cache_file: bool = True, cache_file_name: Optional[str] = None, writer_batch_size: Optional[int] = 1000, features: Optional[datasets.features.Features] = None, disable_nullable: bool = False, fn_kwargs: Optional[dict] = None, num_proc: Optional[int] = None, suffix_template: str = '_{rank:05d}_of_{num_proc:05d}', new_fingerprint: Optional[str] = None) → datasets.arrow_dataset.Dataset[source]

Apply a function to all the elements in the table (individually or in batches) and update the table (if function does updated examples).

Parameters
  • function (callable) – with one of the following signature: - function(example: Union[Dict, Any]) -> Union[Dict, Any] if batched=False and with_indices=False - function(example: Union[Dict, Any], indices: int) -> Union[Dict, Any] if batched=False and with_indices=True - function(batch: Union[Dict[List], List[Any]]) -> Union[Dict, Any] if batched=True and with_indices=False - function(batch: Union[Dict[List], List[Any]], indices: List[int]) -> Union[Dict, Any] if batched=True and with_indices=True If no function is provided, default to identity function: lambda x: x

  • with_indices (bool, defaults to False) – Provide example indices to function. Note that in this case the signature of function should be def function(example, idx): ….

  • input_columns (Optional[Union[str, List[str]]], defaults to None) – The columns to be passed into function as positional arguments. If None, a dict mapping to all formatted columns is passed as one argument.

  • batched (bool, defaults to False) – Provide batch of examples to function

  • batch_size (Optional[int], defaults to 1000) – Number of examples per batch provided to function if batched=True batch_size <= 0 or batch_size == None: Provide the full dataset as a single batch to function

  • (bool, default (drop_last_batch) – False): Whether a last batch smaller than the batch_size should be dropped instead of being processed by the function.

  • remove_columns (Optional[List[str]], defaults to None) – Remove a selection of columns while doing the mapping. Columns will be removed before updating the examples with the output of function, i.e. if function is adding columns with names in remove_columns, these columns will be kept.

  • keep_in_memory (bool, defaults to False) – Keep the dataset in memory instead of writing it to a cache file.

  • load_from_cache_file (bool, defaults to True) – If a cache file storing the current computation from function can be identified, use it instead of recomputing.

  • cache_file_name (Optional[str], defaults to None) – Provide the name of a cache file to use to store the results of the computation instead of the automatically generated cache file name.

  • writer_batch_size (int, defaults to 1000) – Number of rows per write operation for the cache file writer. Higher value gives smaller cache files, lower value consume less temporary memory while running .map().

  • features (Optional[datasets.Features], defaults to None) – Use a specific Features to store the cache file instead of the automatically generated one.

  • disable_nullable (bool, defaults to True) – Disallow null values in the table.

  • fn_kwargs (Optional[Dict], defaults to None) – Keyword arguments to be passed to function

  • num_proc (Optional[int], defaults to None) – Number of processes for multiprocessing. By default it doesn’t use multiprocessing.

  • (str, defaults to "_{rank (suffix_template) – 05d}_of_{num_proc:05d}”): If cache_file_name is specified, then this suffix will be added at the end of the base name of each. For example, if cache_file_name is “processed.arrow”, then for rank=1 and num_proc=4, the resulting file would be “processed_00001_of_00004.arrow” for the default suffix.

  • new_fingerprint (Optional[str], defaults to None) – the new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments

property num_columns

Number of columns in the dataset.

property num_rows

Number of rows in the dataset (same as datasets.Dataset.__len__()).

remove_columns_(column_names: Union[str, List[str]])[source]

Remove one or several column(s) in the dataset and the features associated to them.

You can also remove a column using Dataset.map() with remove_columns but the present method is in-place (doesn’t copy the data to a new dataset) and is thus faster.

Parameters

column_names (Union[str, List[str]]) – Name of the column(s) to remove.

rename_column_(original_column_name: str, new_column_name: str)[source]

Rename a column in the dataset and move the features associated to the original column under the new column name.

You can also rename a column using Dataset.map() with remove_columns but the present method:
  • takes care of moving the original features under the new column name.

  • doesn’t copy the data to a new dataset and is thus much faster.

Parameters
  • original_column_name (str) – Name of the column to rename.

  • new_column_name (str) – New name for the column.

reset_format()[source]

Reset __getitem__ return format to python objects and all columns.

Same as self.set_format()

save_faiss_index(index_name: str, file: str)

Save a FaissIndex on disk

Parameters
  • index_name (str) – The index_name/identifier of the index. This is the index_name that is used to call .get_nearest or .search.

  • file (str) – The path to the serialized faiss index on disk.

save_to_disk(dataset_path: str)[source]

Save the dataset in a dataset directory

Parameters

dataset_path (str) – path of the dataset directory where the dataset will be saved to

search(index_name: str, query: Union[str, numpy.array], k: int = 10) → datasets.search.SearchResults

Find the nearest examples indices in the dataset to the query.

Parameters
  • index_name (str) – The name/identifier of the index.

  • query (Union[str, np.ndarray]) – The query as a string if index_name is a text index or as a numpy array if index_name is a vector index.

  • k (int) – The number of examples to retrieve.

Ouput:

scores (List[List[float]): The retrieval scores of the retrieved examples. indices (List[List[int]]): The indices of the retrieved examples.

search_batch(index_name: str, queries: Union[List[str], numpy.array], k: int = 10) → datasets.search.BatchedSearchResults

Find the nearest examples indices in the dataset to the query.

Parameters
  • index_name (str) – The index_name/identifier of the index.

  • queries (Union[List[str], np.ndarray]) – The queries as a list of strings if index_name is a text index or as a numpy array if index_name is a vector index.

  • k (int) – The number of examples to retrieve per query.

Ouput:

total_scores (List[List[float]): The retrieval scores of the retrieved examples per query. total_indices (List[List[int]]): The indices of the retrieved examples per query.

select(indices: collections.abc.Iterable, keep_in_memory: bool = False, indices_cache_file_name: Optional[str] = None, writer_batch_size: Optional[int] = 1000, new_fingerprint: Optional[str] = None) → datasets.arrow_dataset.Dataset[source]

Create a new dataset with rows selected following the list/array of indices.

Parameters
  • indices (sequence, iterable, ndarray or Series) – List or 1D-array of integer indices for indexing.

  • (bool, default (keep_in_memory) – False): Keep the indices mapping in memory instead of writing it to a cache file.

  • (Optional[str], default (indices_cache_file_name) – None): Provide the name of a cache file to use to store the indices mapping instead of the automatically generated cache file name.

  • (int, default (writer_batch_size) – 1000): Number of rows per write operation for the cache file writer. Higher value gives smaller cache files, lower value consume less temporary memory while running .map().

  • new_fingerprint (Optional[str], defaults to None) – the new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments

set_format(type: Optional[str] = None, columns: Optional[List] = None, output_all_columns: bool = False, **format_kwargs)[source]

Set __getitem__ return format (type and columns)

Parameters
  • type (Optional str) – output type selected in [None, ‘numpy’, ‘torch’, ‘tensorflow’, ‘pandas’] None means __getitem__ returns python objects (default)

  • columns (Optional List[str]) – columns to format in the output None means __getitem__ returns all columns (default)

  • output_all_columns (bool default to False) – keep un-formatted columns as well in the output (as python objects)

  • format_kwargs – keywords arguments passed to the convert function like np.array, torch.tensor or tensorflow.ragged.constant.

property shape

Shape of the dataset (number of columns, number of rows).

shard(num_shards: int, index: int, contiguous: bool = False, keep_in_memory: bool = False, indices_cache_file_name: Optional[str] = None, writer_batch_size: Optional[int] = 1000) → datasets.arrow_dataset.Dataset[source]

Return the index-nth shard from dataset split into num_shards pieces.

This shards deterministically. dset.shard(n, i) will contain all elements of dset whose index mod n = i.

dset.shard(n, i, contiguous=True) will instead split dset into contiguous chunks, so it can be easily concatenated back together after processing. If n % i == l, then the first l shards will have length (n // i) + 1, and the remaining shards will have length (n // i). datasets.concatenate([dset.shard(n, i, contiguous=True) for i in range(n)]) will return a dataset with the same order as the original.

Be sure to shard before using any randomizing operator (such as shuffle). It is best if the shard operator is used early in the dataset pipeline.

Parameters
  • num_shards (int) – How many shards to split the dataset into.

  • index (int) – Which shard to select and return.

  • contiguous – (bool, defaults to False): Whether to select contiguous blocks of indices for shards.

  • keep_in_memory (bool, defaults to False) – Keep the dataset in memory instead of writing it to a cache file.

  • load_from_cache_file (bool, defaults to True) – If a cache file storing the current computation from function can be identified, use it instead of recomputing.

  • indices_cache_file_name (Optional[str], defaults to None) – Provide the name of a cache file to use to store the indices of each shard instead of the automatically generated cache file name.

  • writer_batch_size (int, defaults to 1000) – Number of rows per write operation for the cache file writer. Higher value gives smaller cache files, lower value consume less temporary memory while running .map().

shuffle(seed: Optional[int] = None, generator: Optional[numpy.random._generator.Generator] = None, keep_in_memory: bool = False, load_from_cache_file: bool = True, indices_cache_file_name: Optional[str] = None, writer_batch_size: Optional[int] = 1000, new_fingerprint: Optional[str] = None) → datasets.arrow_dataset.Dataset[source]

Create a new Dataset where the rows are shuffled.

Currently shuffling uses numpy random generators. You can either supply a NumPy BitGenerator to use, or a seed to initiate NumPy’s default random generator (PCG64).

Parameters
  • seed (Optional int) – A seed to initialize the default BitGenerator if generator=None. If None, then fresh, unpredictable entropy will be pulled from the OS. If an int or array_like[ints] is passed, then it will be passed to SeedSequence to derive the initial BitGenerator state.

  • generator (Optional np.random.Generator) – Numpy random Generator to use to compute the permutation of the dataset rows. If generator=None (default), uses np.random.default_rng (the default BitGenerator (PCG64) of NumPy).

  • keep_in_memory (bool, defaults to False) – Keep the shuffled indices in memory instead of writing it to a cache file.

  • load_from_cache_file (bool, defaults to True) – If a cache file storing the shuffled indices can be identified, use it instead of recomputing.

  • indices_cache_file_name (Optional[str], defaults to None) – Provide the name of a cache file to use to store the shuffled indices instead of the automatically generated cache file name.

  • writer_batch_size (int, defaults to 1000) – Number of rows per write operation for the cache file writer. Higher value gives smaller cache files, lower value consume less temporary memory while running .map().

  • new_fingerprint (Optional[str], defaults to None) – the new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments

sort(column: str, reverse: bool = False, kind: str = None, keep_in_memory: bool = False, load_from_cache_file: bool = True, indices_cache_file_name: Optional[str] = None, writer_batch_size: Optional[int] = 1000, new_fingerprint: Optional[str] = None) → datasets.arrow_dataset.Dataset[source]

Create a new dataset sorted according to a column.

Currently sorting according to a column name uses numpy sorting algorithm under the hood. The column should thus be a numpy compatible type (in particular not a nested type). This also means that the column used for sorting is fully loaded in memory (which should be fine in most cases).

Parameters
  • column (str) – column name to sort by.

  • reverse – (bool, defaults to False): If True, sort by descending order rather then ascending.

  • kind (Optional str) – Numpy algorithm for sorting selected in {‘quicksort’, ‘mergesort’, ‘heapsort’, ‘stable’}, The default is ‘quicksort’. Note that both ‘stable’ and ‘mergesort’ use timsort under the covers and, in general, the actual implementation will vary with data type. The ‘mergesort’ option is retained for backwards compatibility.

  • keep_in_memory (bool, defaults to False) – Keep the sorted indices in memory instead of writing it to a cache file.

  • load_from_cache_file (bool, defaults to True) – If a cache file storing the sorted indices can be identified, use it instead of recomputing.

  • indices_cache_file_name (Optional[str], defaults to None) – Provide the name of a cache file to use to store the sorted indices instead of the automatically generated cache file name.

  • writer_batch_size (int, defaults to 1000) – Number of rows per write operation for the cache file writer. Higher value gives smaller cache files, lower value consume less temporary memory.

  • new_fingerprint (Optional[str], defaults to None) – the new fingerprint of the dataset after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments

property split

datasets.DatasetInfo object containing all the metadata in the dataset.

train_test_split(test_size: Optional[Union[float, int]] = None, train_size: Optional[Union[float, int]] = None, shuffle: bool = True, seed: Optional[int] = None, generator: Optional[numpy.random._generator.Generator] = None, keep_in_memory: bool = False, load_from_cache_file: bool = True, train_indices_cache_file_name: Optional[str] = None, test_indices_cache_file_name: Optional[str] = None, writer_batch_size: Optional[int] = 1000, train_new_fingerprint: Optional[str] = None, test_new_fingerprint: Optional[str] = None) → DatasetDict[source]

Return a dictionary (datasets.DatsetDict) with two random train and test subsets (train and test Dataset splits). Splits are created from the dataset according to test_size, train_size and shuffle.

This method is similar to scikit-learn train_test_split with the omission of the stratified options.

Parameters
  • test_size (Optional np.random.Generator) – Size of the test split If float, should be between 0.0 and 1.0 and represent the proportion of the dataset to include in the test split. If int, represents the absolute number of test samples. If None, the value is set to the complement of the train size. If train_size is also None, it will be set to 0.25.

  • train_size (Optional np.random.Generator) – Size of the train split If float, should be between 0.0 and 1.0 and represent the proportion of the dataset to include in the train split. If int, represents the absolute number of train samples. If None, the value is automatically set to the complement of the test size.

  • shuffle (Optional bool, defaults to True) – Whether or not to shuffle the data before splitting.

  • seed (Optional int) – A seed to initialize the default BitGenerator if generator=None. If None, then fresh, unpredictable entropy will be pulled from the OS. If an int or array_like[ints] is passed, then it will be passed to SeedSequence to derive the initial BitGenerator state.

  • generator (Optional np.random.Generator) – Numpy random Generator to use to compute the permutation of the dataset rows. If generator=None (default), uses np.random.default_rng (the default BitGenerator (PCG64) of NumPy).

  • keep_in_memory (bool, defaults to False) – Keep the splits indices in memory instead of writing it to a cache file.

  • load_from_cache_file (bool, defaults to True) – If a cache file storing the splits indices can be identified, use it instead of recomputing.

  • train_cache_file_name (Optional[str], defaults to None) – Provide the name of a cache file to use to store the train split indices instead of the automatically generated cache file name.

  • test_cache_file_name (Optional[str], defaults to None) – Provide the name of a cache file to use to store the test split indices instead of the automatically generated cache file name.

  • writer_batch_size (int, defaults to 1000) – Number of rows per write operation for the cache file writer. Higher value gives smaller cache files, lower value consume less temporary memory while running .map().

  • train_new_fingerprint (Optional[str], defaults to None) – the new fingerprint of the train set after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments

  • test_new_fingerprint (Optional[str], defaults to None) – the new fingerprint of the test set after transform. If None, the new fingerprint is computed using a hash of the previous fingerprint, and the transform arguments

unique(column: str) → List[Any][source]

Return a list of the unique elements in a column.

This is implemented in the low-level backend and as such, very fast.

Parameters

column (str) – column name (list all the column names with datasets.Dataset.column_names())

Returns: list of unique elements in the given column.

datasets.concatenate_datasets(dsets: List[datasets.arrow_dataset.Dataset], info: Optional[Any] = None, split: Optional[Any] = None)[source]

Converts a list of :obj:datasets.Dataset with the same schema into a single :obj:datasets.Dataset.

Parameters
  • dsets (:obj:List[datasets.Dataset]) – A list of Datasets to concatenate

  • info (:obj:datasets.DatasetInfo, optional, defaults to :obj:None) – If specified, the dataset info containing info like description, citation, etc.

  • split (:obj:datasets.NamedSplit, optional, defaults to :obj:None) – If specified, the name of the dataset split.

DatasetDict

Dictionary with split names as keys (‘train’, ‘test’ for example), and datasets.Dataset objects as values. It also has dataset transform methods like map or filter, to process all the splits at once.

class datasets.DatasetDict[source]

A dictionary (dict of str: datasets.Dataset) with dataset transforms methods (map, filter, etc.)

property cache_files

The cache files containing the Apache Arrow table backing each split.

cast_(features: datasets.features.Features)[source]

Cast the dataset to a new set of features. The transformation is applied to all the datasets of the dataset dictionary.

You can also remove a column using Dataset.map() with feature but cast_() is in-place (doesn’t copy the data to a new dataset) and is thus faster.

Parameters

features (datasets.Features) – New features to cast the dataset to. The name and order of the fields in the features must match the current column names. The type of the data must also be convertible from one type to the other. For non-trivial conversion, e.g. string <-> ClassLabel you should use map() to update the Dataset.

cleanup_cache_files() → Dict[str, int][source]

Clean up all cache files in the dataset cache directory, excepted the currently used cache file if there is one. Be carefull when running this command that no other process is currently using other cache files.

Returns

Dict with the number of removed files for each split

property column_names

Names of the columns in each split of the dataset.

property data

The Apache Arrow tables backing each split.

filter(function, with_indices=False, input_columns: Optional[Union[str, List[str]]] = None, batch_size: Optional[int] = 1000, remove_columns: Optional[List[str]] = None, keep_in_memory: bool = False, load_from_cache_file: bool = True, cache_file_names: Optional[Dict[str, str]] = None, writer_batch_size: Optional[int] = 1000, fn_kwargs: Optional[dict] = None, num_proc: Optional[int] = None) → datasets.dataset_dict.DatasetDict[source]

Apply a filter function to all the elements in the table in batches and update the table so that the dataset only includes examples according to the filter function. The transformation is applied to all the datasets of the dataset dictionary.

Parameters
  • function (callable) – with one of the following signature: - function(example: Dict) -> bool if with_indices=False - function(example: Dict, indices: int) -> bool if with_indices=True

  • with_indices (bool, defaults to False) – Provide example indices to function. Note that in this case the signature of function should be def function(example, idx): ….

  • input_columns (Optional[Union[str, List[str]]], defaults to None) – The columns to be passed into function as positional arguments. If None, a dict mapping to all formatted columns is passed as one argument.

  • batch_size (Optional[int], defaults to 1000) – Number of examples per batch provided to function if batched=True batch_size <= 0 or batch_size == None: Provide the full dataset as a single batch to function

  • remove_columns (Optional[List[str]], defaults to None) – Remove a selection of columns while doing the mapping. Columns will be removed before updating the examples with the output of function, i.e. if function is adding columns with names in remove_columns, these columns will be kept.

  • keep_in_memory (bool, defaults to False) – Keep the dataset in memory instead of writing it to a cache file.

  • load_from_cache_file (bool, defaults to True) – If a cache file storing the current computation from function can be identified, use it instead of recomputing.

  • cache_file_names (Optional[Dict[str, str]], defaults to None) – Provide the name of a cache file to use to store the results of the computation instead of the automatically generated cache file name. You have to provide one cache_file_name per dataset in the dataset dictionary.

  • writer_batch_size (int, defaults to 1000) – Number of rows per write operation for the cache file writer. Higher value gives smaller cache files, lower value consume less temporary memory while running .map().

  • fn_kwargs (Optional[Dict], defaults to None) – Keyword arguments to be passed to function

  • num_proc (Optional[int], defaults to None) – Number of processes for multiprocessing. By default it doesn’t use multiprocessing.

flatten_(max_depth=16)[source]

Flatten the Apache Arrow Table of each split (nested features are flatten). Each column with a struct type is flattened into one column per struct field. Other columns are left unchanged.

formatted_as(type: Optional[str] = None, columns: Optional[List] = None, output_all_columns: bool = False, **format_kwargs)[source]

To be used in a with statement. Set __getitem__ return format (type and columns) The transformation is applied to all the datasets of the dataset dictionary.

Parameters
  • type (Optional str) – output type selected in [None, ‘numpy’, ‘torch’, ‘tensorflow’, ‘pandas’] None means __getitem__ returns python objects (default)

  • columns (Optional List[str]) – columns to format in the output None means __getitem__ returns all columns (default)

  • output_all_columns (bool default to False) – keep un-formatted columns as well in the output (as python objects)

  • format_kwargs – keywords arguments passed to the convert function like np.array, torch.tensor or tensorflow.ragged.constant.

static load_from_disk(dataset_dict_path: str) → datasets.dataset_dict.DatasetDict[source]

Load the dataset dict from a dataset dict directory

Parameters

dataset_dict_path (str) – path of the dataset dict directory where the dataset dict will be loaded from

map(function, with_indices: bool = False, input_columns: Optional[Union[str, List[str]]] = None, batched: bool = False, batch_size: Optional[int] = 1000, remove_columns: Optional[List[str]] = None, keep_in_memory: bool = False, load_from_cache_file: bool = True, cache_file_names: Optional[Dict[str, str]] = None, writer_batch_size: Optional[int] = 1000, features: Optional[datasets.features.Features] = None, disable_nullable: bool = False, fn_kwargs: Optional[dict] = None, num_proc: Optional[int] = None) → datasets.dataset_dict.DatasetDict[source]

Apply a function to all the elements in the table (individually or in batches) and update the table (if function does updated examples). The transformation is applied to all the datasets of the dataset dictionary.

Parameters
  • function (callable) – with one of the following signature: - function(example: Dict) -> Union[Dict, Any] if batched=False and with_indices=False - function(example: Dict, indices: int) -> Union[Dict, Any] if batched=False and with_indices=True - function(batch: Dict[List]) -> Union[Dict, Any] if batched=True and with_indices=False - function(batch: Dict[List], indices: List[int]) -> Union[Dict, Any] if batched=True and with_indices=True

  • with_indices (bool, defaults to False) – Provide example indices to function. Note that in this case the signature of function should be def function(example, idx): ….

  • input_columns (Optional[Union[str, List[str]]], defaults to None) – The columns to be passed into function as positional arguments. If None, a dict mapping to all formatted columns is passed as one argument.

  • batched (bool, defaults to False) – Provide batch of examples to function

  • batch_size (Optional[int], defaults to 1000) – Number of examples per batch provided to function if batched=True batch_size <= 0 or batch_size == None: Provide the full dataset as a single batch to function

  • remove_columns (Optional[List[str]], defaults to None) – Remove a selection of columns while doing the mapping. Columns will be removed before updating the examples with the output of function, i.e. if function is adding columns with names in remove_columns, these columns will be kept.

  • keep_in_memory (bool, defaults to False) – Keep the dataset in memory instead of writing it to a cache file.

  • load_from_cache_file (bool, defaults to True) – If a cache file storing the current computation from function can be identified, use it instead of recomputing.

  • cache_file_names (Optional[Dict[str, str]], defaults to None) – Provide the name of a cache file to use to store the results of the computation instead of the automatically generated cache file name. You have to provide one cache_file_name per dataset in the dataset dictionary.

  • writer_batch_size (int, defaults to 1000) – Number of rows per write operation for the cache file writer. Higher value gives smaller cache files, lower value consume less temporary memory while running .map().

  • features (Optional[datasets.Features], defaults to None) – Use a specific Features to store the cache file instead of the automatically generated one.

  • disable_nullable (bool, defaults to True) – Disallow null values in the table.

  • fn_kwargs (Optional[Dict], defaults to None) – Keyword arguments to be passed to function

  • num_proc (Optional[int], defaults to None) – Number of processes for multiprocessing. By default it doesn’t use multiprocessing.

property num_columns

Number of columns in each split of the dataset.

property num_rows

Number of rows in each split of the dataset (same as datasets.Dataset.__len__()).

remove_columns_(column_names: Union[str, List[str]])[source]

Remove one or several column(s) from each split in the dataset and the features associated to the column(s).

The transformation is applied to all the splits of the dataset dictionary.

You can also remove a column using Dataset.map() with remove_columns but the present method is in-place (doesn’t copy the data to a new dataset) and is thus faster.

Parameters

column_names (Union[str, List[str]]) – Name of the column(s) to remove.

rename_column_(original_column_name: str, new_column_name: str)[source]

Rename a column in the dataset and move the features associated to the original column under the new column name. The transformation is applied to all the datasets of the dataset dictionary.

You can also rename a column using Dataset.map() with remove_columns but the present method:
  • takes care of moving the original features under the new column name.

  • doesn’t copy the data to a new dataset and is thus much faster.

Parameters
  • original_column_name (str) – Name of the column to rename.

  • new_column_name (str) – New name for the column.

reset_format()[source]

Reset __getitem__ return format to python objects and all columns. The transformation is applied to all the datasets of the dataset dictionary.

Same as self.set_format()

save_to_disk(dataset_dict_path: str)[source]

Save the dataset dict in a dataset dict directory.

Parameters

dataset_dict_path (str) – path of the dataset dict directory where the dataset dict will be saved to

set_format(type: Optional[str] = None, columns: Optional[List] = None, output_all_columns: bool = False, **format_kwargs)[source]

Set __getitem__ return format (type and columns) The transformation is applied to all the datasets of the dataset dictionary.

Parameters
  • type (Optional str) – output type selected in [None, ‘numpy’, ‘torch’, ‘tensorflow’, ‘pandas’] None means __getitem__ returns python objects (default)

  • columns (Optional List[str]) – columns to format in the output None means __getitem__ returns all columns (default)

  • output_all_columns (bool default to False) – keep un-formatted columns as well in the output (as python objects)

  • format_kwargs – keywords arguments passed to the convert function like np.array, torch.tensor or tensorflow.ragged.constant.

property shape

Shape of each split of the dataset (number of columns, number of rows).

shuffle(seeds: Optional[Dict[str, int]] = None, generators: Optional[Dict[str, numpy.random._generator.Generator]] = None, keep_in_memory: bool = False, load_from_cache_file: bool = True, indices_cache_file_names: Optional[Dict[str, str]] = None, writer_batch_size: Optional[int] = 1000)[source]

Create a new Dataset where the rows are shuffled. The transformation is applied to all the datasets of the dataset dictionary.

Currently shuffling uses numpy random generators. You can either supply a NumPy BitGenerator to use, or a seed to initiate NumPy’s default random generator (PCG64).

Parameters
  • seeds (Optional Dict[str, int]) – A seed to initialize the default BitGenerator if generator=None. If None, then fresh, unpredictable entropy will be pulled from the OS. If an int or array_like[ints] is passed, then it will be passed to SeedSequence to derive the initial BitGenerator state. You have to provide one seed per dataset in the dataset dictionary.

  • generators (Optional Dict[str, np.random.Generator]) – Numpy random Generator to use to compute the permutation of the dataset rows. If generator=None (default), uses np.random.default_rng (the default BitGenerator (PCG64) of NumPy). You have to provide one generator per dataset in the dataset dictionary.

  • keep_in_memory (bool, defaults to False) – Keep the dataset in memory instead of writing it to a cache file.

  • load_from_cache_file (bool, defaults to True) – If a cache file storing the current computation from function can be identified, use it instead of recomputing.

  • (Optional[Dict[str, str]], default (indices_cache_file_names) – None): Provide the name of a cache file to use to store the indices mappings instead of the automatically generated cache file name. You have to provide one cache_file_name per dataset in the dataset dictionary.

  • writer_batch_size (int, defaults to 1000) – Number of rows per write operation for the cache file writer. Higher value gives smaller cache files, lower value consume less temporary memory while running .map().

sort(column: str, reverse: bool = False, kind: str = None, keep_in_memory: bool = False, load_from_cache_file: bool = True, indices_cache_file_names: Optional[Dict[str, str]] = None, writer_batch_size: Optional[int] = 1000) → datasets.dataset_dict.DatasetDict[source]

Create a new dataset sorted according to a column. The transformation is applied to all the datasets of the dataset dictionary.

Currently sorting according to a column name uses numpy sorting algorithm under the hood. The column should thus be a numpy compatible type (in particular not a nested type). This also means that the column used for sorting is fully loaded in memory (which should be fine in most cases).

Parameters
  • column (str) – column name to sort by.

  • reverse – (bool, defaults to False): If True, sort by descending order rather then ascending.

  • kind (Optional str) – Numpy algorithm for sorting selected in {‘quicksort’, ‘mergesort’, ‘heapsort’, ‘stable’}, The default is ‘quicksort’. Note that both ‘stable’ and ‘mergesort’ use timsort under the covers and, in general, the actual implementation will vary with data type. The ‘mergesort’ option is retained for backwards compatibility.

  • keep_in_memory (bool, defaults to False) – Keep the dataset in memory instead of writing it to a cache file.

  • load_from_cache_file (bool, defaults to True) – If a cache file storing the current computation from function can be identified, use it instead of recomputing.

  • indices_cache_file_names (Optional[Dict[str, str]], defaults to None) – Provide the name of a cache file to use to store the indices mapping instead of the automatically generated cache file name. You have to provide one cache_file_name per dataset in the dataset dictionary.

  • writer_batch_size (int, defaults to 1000) – Number of rows per write operation for the cache file writer. Higher value gives smaller cache files, lower value consume less temporary memory while running .map().

unique(column: str) → Dict[str, List[Any]][source]

Return a list of the unique elements in a column for each split.

This is implemented in the low-level backend and as such, very fast.

Parameters

column (str) – column name (list all the column names with datasets.Dataset.column_names())

Returns: Dict[:obj: str, list] of unique elements in the given column.

Features

class datasets.Features[source]
copy() → a shallow copy of D[source]
class datasets.Sequence(feature: Any, length: int = - 1, id: Optional[str] = None)[source]

Construct a list of feature from a single type or a dict of types. Mostly here for compatiblity with tfds.

class datasets.ClassLabel(num_classes: int = None, names: List[str] = None, names_file: str = None, id: Optional[str] = None)[source]

Handle integer class labels. Here for compatiblity with tfds.

There are 3 ways to define a ClassLabel, which correspond to the 3 :param * num_classes: create 0 to (num_classes-1) labels :param * names: a list of label strings :param * names_file: a file containing the list of labels.

Note: On python2, the strings are encoded as utf-8.

Parameters
  • num_classesint, number of classes. All labels must be < num_classes.

  • nameslist<str>, string names for the integer classes. The order in which the names are provided is kept.

  • names_filestr, path to a file with names for the integer classes, one per line.

int2str(values: Union[int, collections.abc.Iterable])[source]

Conversion integer => class name string.

str2int(values: collections.abc.Iterable)[source]

Conversion class name string => integer.

class datasets.Value(dtype: str, id: Optional[str] = None)[source]

Encapsulate an Arrow datatype for easy serialization.

class datasets.Translation(languages: List[str], id: Optional[str] = None)[source]

FeatureConnector for translations with fixed languages per example. Here for compatiblity with tfds.

Input: The Translate feature accepts a dictionary for each example mapping

string language codes to string translations.

Output: A dictionary mapping string language codes to translations as Text

features.

Example:

# At construction time:

datasets.features.Translation(languages=['en', 'fr', 'de'])

# During data generation:

yield {
        'en': 'the cat',
        'fr': 'le chat',
        'de': 'die katze'
}
class datasets.TranslationVariableLanguages(languages: List = None, num_languages: int = None, id: Optional[str] = None)[source]

FeatureConnector for translations with variable languages per example. Here for compatiblity with tfds.

Input: The TranslationVariableLanguages feature accepts a dictionary for each

example mapping string language codes to one or more string translations. The languages present may vary from example to example.

Output:
language: variable-length 1D tf.Tensor of tf.string language codes, sorted

in ascending order.

translation: variable-length 1D tf.Tensor of tf.string plain text

translations, sorted to align with language codes.

Example:

# At construction time:

datasets.features.Translation(languages=['en', 'fr', 'de'])

# During data generation:

yield {
        'en': 'the cat',
        'fr': ['le chat', 'la chatte,']
        'de': 'die katze'
}

# Tensor returned :

{
        'language': ['en', 'de', 'fr', 'fr'],
        'translation': ['the cat', 'die katze', 'la chatte', 'le chat'],
}
class datasets.Array2D(shape: tuple, dtype: str, id: Union[str, NoneType] = None)[source]
class datasets.Array3D(shape: tuple, dtype: str, id: Union[str, NoneType] = None)[source]
class datasets.Array4D(shape: tuple, dtype: str, id: Union[str, NoneType] = None)[source]
class datasets.Array5D(shape: tuple, dtype: str, id: Union[str, NoneType] = None)[source]

MetricInfo

class datasets.MetricInfo(description: str, citation: str, features: datasets.features.Features, inputs_description: str = <factory>, homepage: str = <factory>, license: str = <factory>, codebase_urls: List[str] = <factory>, reference_urls: List[str] = <factory>, streamable: bool = False, format: Optional[str] = None, metric_name: Optional[str] = None, config_name: Optional[str] = None, experiment_id: Optional[str] = None)[source]

Information about a metric.

MetricInfo documents a metric, including its name, version, and features. See the constructor arguments and properties for a full list.

Note: Not all fields are known on construction and may be updated later.

classmethod from_directory(metric_info_dir) → datasets.info.MetricInfo[source]

Create MetricInfo from the JSON file in metric_info_dir.

Parameters

metric_info_dirstr The directory containing the metadata file. This should be the root directory of a specific dataset version.

write_to_directory(metric_info_dir)[source]

Write MetricInfo as JSON to metric_info_dir. Also save the license separately in LICENCE.

Metric

The base class Metric implements a Metric backed by one or several datasets.Dataset.

class datasets.Metric(config_name: Optional[str] = None, keep_in_memory: bool = False, cache_dir: Optional[str] = None, num_process: int = 1, process_id: int = 0, seed: Optional[int] = None, experiment_id: Optional[str] = None, max_concurrent_cache_files: int = 10000, timeout: Union[int, float] = 100, **kwargs)[source]

A Metrics is the base class and common API for all metrics.

Parameters
  • config_name (str) – This is used to define a hash specific to a metrics computation script and prevents the metric’s data to be overridden when the metric loading script is modified.

  • keep_in_memory (bool) – keep all predictions and references in memory. Not possible in distributed settings.

  • cache_dir (str) – Path to a directory in which temporary prediction/references data will be stored. The data directory should be located on a shared file-system in distributed setups.

  • num_process (int) – specify the total number of nodes in a distributed settings. This is useful to compute metrics in distributed setups (in particular non-additive metrics like F1).

  • process_id (int) – specify the id of the current process in a distributed setup (between 0 and num_process-1) This is useful to compute metrics in distributed setups (in particular non-additive metrics like F1).

  • seed (Optional int) – If specified, this will temporarily set numpy’s random seed when datasets.Metric.compute() is run.

  • experiment_id (str) – A specific experiment id. This is used if several distributed evaluations share the same file system. This is useful to compute metrics in distributed setups (in particular non-additive metrics like F1).

  • max_concurrent_cache_files (int) – Max number of concurrent metrics cache files (default 10000).

  • timeout (Union[int, float]) – Timeout in second for distributed setting synchronization.

add(*, prediction=None, reference=None)[source]

Add one prediction and reference for the metric’s stack.

add_batch(*, predictions=None, references=None)[source]

Add a batch of predictions and references for the metric’s stack.

compute(*args, **kwargs) → Optional[dict][source]

Compute the metrics.

Parameters
  • disallow the usage of positional arguments to prevent mistakes (We) –

  • predictions (Optional list/array/tensor) – predictions

  • references (Optional list/array/tensor) – references

  • **kwargs (Optional other kwargs) – will be forwared to the metrics _compute() method (see details in the docstring)

Returns

Dictionnary with the metrics if this metric is run on the main process (process_id == 0) None if the metric is not run on the main process (process_id != 0)

download_and_prepare(download_config: Optional[datasets.utils.file_utils.DownloadConfig] = None, dl_manager: Optional[datasets.utils.download_manager.DownloadManager] = None, **download_and_prepare_kwargs)[source]

Downloads and prepares dataset for reading.

Parameters
  • (Optional datasets.DownloadConfig (download_config) – specific download configuration parameters.

  • dl_manager (Optional datasets.DownloadManager) – specific Download Manger to use