deeplake.core.dataset

Dataset

class deeplake.core.dataset.Dataset
add_creds_key(creds_key: str, managed: bool = False)

Adds a new creds key to the dataset. These keys are used for tensors that are linked to external data.

Examples

>>> # create/load a dataset
>>> ds = deeplake.empty("path/to/dataset")
>>> # add a new creds key
>>> ds.add_creds_key("my_s3_key")
Parameters:
  • creds_key (str) – The key to be added.

  • managed (bool) –

    • If True, the creds corresponding to the key will be fetched from Activeloop platform.

    • Defaults to False.

Raises:

ValueError – If the dataset is not connected to Activeloop platform and managed is True.

Note

managed parameter is applicable only for datasets that are connected to Activeloop platform.

append(sample: Dict[str, Any], skip_ok: bool = False, append_empty: bool = False)

Append samples to mutliple tensors at once. This method expects all tensors being updated to be of the same length.

Parameters:
  • sample (dict) – Dictionary with tensor names as keys and samples as values.

  • skip_ok (bool) – Skip tensors not in sample if set to True.

  • append_empty (bool) – Append empty samples to tensors not specified in sample if set to True. If True, skip_ok is ignored.

Raises:
  • KeyError – If any tensor in the dataset is not a key in sample and skip_ok is False.

  • TensorDoesNotExistError – If tensor in sample does not exist.

  • ValueError – If all tensors being updated are not of the same length.

  • NotImplementedError – If an error occurs while writing tiles.

  • Exception – Error while attempting to rollback appends.

  • SampleAppendingError – Error that occurs when someone tries to append a tensor value directly to the dataset without specifying tensor name.

Examples

>>> ds = deeplake.empty("../test/test_ds")
>>> ds.create_tensor('data')
Tensor(key='data')
>>> ds.create_tensor('labels')
Tensor(key='labels')
>>> ds.append({"data": [1, 2, 3, 4], "labels":[0, 1, 2, 3]})
property branch: str

The current branch of the dataset

property branches

Lists all the branches of the dataset.

Returns:

List of branches.

change_creds_management(creds_key: str, managed: bool)

Changes the management status of the creds key.

Parameters:
  • creds_key (str) – The key whose management status is to be changed.

  • managed (bool) – The target management status. If True, the creds corresponding to the key will be fetched from activeloop platform.

Raises:
  • ValueError – If the dataset is not connected to activeloop platform.

  • KeyError – If the creds key is not present in the dataset.

Examples

>>> # create/load a dataset
>>> ds = deeplake.dataset("path/to/dataset")
>>> # add a new creds key
>>> ds.add_creds_key("my_s3_key")
>>> # Populate the name added with creds dictionary
>>> # These creds are only present temporarily and will have to be repopulated on every reload
>>> ds.populate_creds("my_s3_key", {})
>>> # Change the management status of the key to True. Before doing this, ensure that the creds have been created on activeloop platform
>>> # Now, this key will no longer use the credentials populated in the previous step but will instead fetch them from activeloop platform
>>> # These creds don't have to be populated again on every reload and will be fetched every time the dataset is loaded
>>> ds.change_creds_management("my_s3_key", True)
checkout(address: str, create: bool = False) str | None

Checks out to a specific commit_id or branch. If create = True, creates a new branch with name address.

Parameters:
  • address (str) – The commit_id or branch to checkout to.

  • create (bool) – If True, creates a new branch with name as address.

Returns:

The commit_id of the dataset after checkout.

Return type:

Optional[str]

Raises:

Exception – If the dataset is a filtered view.

Examples

>>> ds = deeplake.empty("../test/test_ds")
>>> ds.create_tensor("abc")
Tensor(key='abc')
>>> ds.abc.append([1, 2, 3])
>>> first_commit = ds.commit()
>>> ds.checkout("alt", create=True)
'firstdbf9474d461a19e9333c2fd19b46115348f'
>>> ds.abc.append([4, 5, 6])
>>> ds.abc.numpy()
array([[1, 2, 3],
       [4, 5, 6]])
>>> ds.checkout(first_commit)
'firstdbf9474d461a19e9333c2fd19b46115348f'
>>> ds.abc.numpy()
array([[1, 2, 3]])

Note

Checkout from a head node in any branch that contains uncommitted data will lead to an automatic commit before the checkout.

clear_cache()
  • Flushes (see Dataset.flush()) the contents of the cache layers (if any) and then deletes contents of all the layers of it.

  • This doesn’t delete data from the actual storage.

  • This is useful if you have multiple datasets with memory caches open, taking up too much RAM.

  • Also useful when local cache is no longer needed for certain datasets and is taking up storage space.

property client

Returns the client of the dataset.

commit(message: str | None = None, allow_empty=False) str

Stores a snapshot of the current state of the dataset.

Parameters:
  • message (str, Optional) – Used to describe the commit.

  • allow_empty (bool) – If True, commit even if there are no changes.

Returns:

the commit id of the saved commit that can be used to access the snapshot.

Return type:

str

Raises:
  • Exception – If dataset is a filtered view.

  • EmptyCommitError – if there are no changes and user does not forced to commit unchanged data.

Note

  • Commiting from a non-head node in any branch, will lead to an automatic checkout to a new branch.

  • This same behaviour will happen if new samples are added or existing samples are updated from a non-head node.

property commit_id: str | None

The lasted committed commit id of the dataset. If there are no commits, this returns None.

property commits: List[Dict]

Lists all the commits leading to the current dataset state.

Returns:

List of dictionaries containing commit information.

connect(creds_key: str, dest_path: str | None = None, org_id: str | None = None, ds_name: str | None = None, token: str | None = None)

Connect a Deep Lake cloud dataset through a deeplake path.

Examples

>>> # create/load an s3 dataset
>>> s3_ds = deeplake.dataset("s3://bucket/dataset")
>>> ds = s3_ds.connect(dest_path="hub://my_org/dataset", creds_key="my_managed_credentials_key")
>>> # or
>>> ds = s3_ds.connect(org_id="my_org", creds_key="my_managed_credentials_key")
Parameters:
  • creds_key (str) – The managed credentials to be used for accessing the source path.

  • dest_path (str, optional) – The full path to where the connected Deep Lake dataset will reside. Can be: a Deep Lake path like hub://organization/dataset

  • org_id (str, optional) – The organization to where the connected Deep Lake dataset will be added.

  • ds_name (str, optional) – The name of the connected Deep Lake dataset. Will be infered from dest_path or src_path if not provided.

  • token (str, optional) – Activeloop token used to fetch the managed credentials.

Returns:

The connected Deep Lake dataset.

Return type:

Dataset

Raises:
  • InvalidSourcePathError – If the dataset’s path is not a valid s3 or gcs path.

  • InvalidDestinationPathError – If dest_path, or org_id and ds_name do not form a valid Deep Lake path.

copy(dest: str | Path, tensors: List[str] | None = None, overwrite: bool = False, creds=None, token=None, num_workers: int = 0, scheduler='threaded', progressbar=True, public: bool = False)

Copies this dataset or dataset view to dest. Version control history is not included.

Parameters:
  • dest (str, pathlib.Path) – Destination dataset or path to copy to. If a Dataset instance is provided, it is expected to be empty.

  • tensors (List[str], optional) – Names of tensors (and groups) to be copied. If not specified all tensors are copied.

  • overwrite (bool) – If True and a dataset exists at destination, it will be overwritten. Defaults to False.

  • creds (dict, Optional) – creds required to create / overwrite datasets at dest.

  • token (str, Optional) – token used to for fetching credentials to dest.

  • num_workers (int) – The number of workers to use for copying. Defaults to 0. When set to 0, it will always use serial processing, irrespective of the scheduler.

  • scheduler (str) – The scheduler to be used for copying. Supported values include: ‘serial’, ‘threaded’, ‘processed’ and ‘ray’. Defaults to ‘threaded’.

  • progressbar (bool) – Displays a progress bar If True (default).

  • public (bool) – Defines if the dataset will have public access. Applicable only if Deep Lake cloud storage is used and a new Dataset is being created. Defaults to False.

Returns:

New dataset object.

Return type:

Dataset

Raises:

DatasetHandlerError – If a dataset already exists at destination path and overwrite is False.

create_group(name: str, exist_ok=False) Dataset

Creates a tensor group. Intermediate groups in the path are also created.

Parameters:
  • name – The name of the group to create.

  • exist_ok – If True, the group is created if it does not exist. If False, an error is raised if the group already exists. Defaults to False.

Returns:

The created group.

Raises:

TensorGroupAlreadyExistsError – If the group already exists and exist_ok is False.

Examples

>>> ds.create_group("images")
>>> ds['images'].create_tensor("cats")
>>> ds.create_groups("images/jpg/cats")
>>> ds["images"].create_tensor("png")
>>> ds["images/jpg"].create_group("dogs")
create_tensor(name: str, htype: str = 'unspecified', dtype: str | dtype = 'unspecified', sample_compression: str = 'unspecified', chunk_compression: str = 'unspecified', hidden: bool = False, create_sample_info_tensor: bool = True, create_shape_tensor: bool = True, create_id_tensor: bool = True, verify: bool = True, exist_ok: bool = False, **kwargs)

Creates a new tensor in the dataset.

Examples

>>> # create dataset
>>> ds = deeplake.dataset("path/to/dataset")
>>> # create tensors
>>> ds.create_tensor("images", htype="image", sample_compression="jpg")
>>> ds.create_tensor("videos", htype="video", sample_compression="mp4")
>>> ds.create_tensor("data")
>>> ds.create_tensor("point_clouds", htype="point_cloud")
>>> # append data
>>> ds.images.append(np.ones((400, 400, 3), dtype='uint8'))
>>> ds.videos.append(deeplake.read("videos/sample_video.mp4"))
>>> ds.data.append(np.zeros((100, 100, 2)))
Parameters:
  • name (str) – The name of the tensor to be created.

  • htype (str) –

    • The class of data for the tensor.

    • The defaults for other parameters are determined in terms of this value.

    • For example, htype="image" would have dtype default to uint8.

    • These defaults can be overridden by explicitly passing any of the other parameters to this function.

    • May also modify the defaults for other parameters.

  • dtype (str) – Optionally override this tensor’s dtype. All subsequent samples are required to have this dtype.

  • sample_compression (str) – All samples will be compressed in the provided format. If None, samples are uncompressed. For link[] tensors, sample_compression is used only for optimizing dataset views.

  • chunk_compression (str) – All chunks will be compressed in the provided format. If None, chunks are uncompressed. For link[] tensors, chunk_compression is used only for optimizing dataset views.

  • hidden (bool) – If True, the tensor will be hidden from ds.tensors but can still be accessed via ds[tensor_name].

  • create_sample_info_tensor (bool) – If True, meta data of individual samples will be saved in a hidden tensor. This data can be accessed via tensor[i].sample_info.

  • create_shape_tensor (bool) – If True, an associated tensor containing shapes of each sample will be created.

  • create_id_tensor (bool) – If True, an associated tensor containing unique ids for each sample will be created. This is useful for merge operations.

  • verify (bool) – Valid only for link htypes. If True, all links will be verified before they are added to the tensor. verify is always True even if specified as False if create_shape_tensor or create_sample_info_tensor is True.

  • exist_ok (bool) – If True, the group is created if it does not exist. if False, an error is raised if the group already exists.

  • **kwargs

    • htype defaults can be overridden by passing any of the compatible parameters.

    • To see all htypes and their correspondent arguments, check out Htypes.

Returns:

The new tensor, which can be accessed by dataset[name] or dataset.name.

Return type:

Tensor

Raises:
create_tensor_like(name: str, source: Tensor, unlink: bool = False) Tensor

Copies the source tensor’s meta information and creates a new tensor with it. No samples are copied, only the meta/info for the tensor is.

Examples

>>> ds.create_tensor_like("cats", ds["images"])
Parameters:
  • name (str) – Name for the new tensor.

  • source (Tensor) – Tensor who’s meta/info will be copied. May or may not be contained in the same dataset.

  • unlink (bool) – Whether to unlink linked tensors.

Returns:

New Tensor object.

Return type:

Tensor

delete(large_ok=False)

Deletes the entire dataset from the cache layers (if any) and the underlying storage. This is an IRREVERSIBLE operation. Data once deleted can not be recovered.

Parameters:

large_ok (bool) – Delete datasets larger than 1 GB. Defaults to False.

Raises:

DatasetTooLargeToDelete – If the dataset is larger than 1 GB and large_ok is False.

delete_group(name: str, large_ok: bool = False)

Delete a tensor group from the dataset.

Examples

>>> ds.delete_group("images/dogs")
Parameters:
  • name (str) – The name of tensor group to be deleted.

  • large_ok (bool) – Delete tensor groups larger than 1 GB. Disabled by default.

Returns:

None

Raises:

TensorGroupDoesNotExistError – If tensor group of name name does not exist in the dataset.

delete_tensor(name: str, large_ok: bool = False)

Delete a tensor from the dataset.

Examples

>>> ds.delete_tensor("images/cats")
Parameters:
  • name (str) – The name of tensor to be deleted.

  • large_ok (bool) – Delete tensors larger than 1 GB. Disabled by default.

Returns:

None

Raises:
  • TensorDoesNotExistError – If tensor of name name does not exist in the dataset.

  • TensorTooLargeToDelete – If the tensor is larger than 1 GB and large_ok is False.

delete_view(id: str)

Deletes the view with given view id.

Parameters:

id (str) – Id of the view to delete.

Raises:

KeyError – if view with given id does not exist.

diff(id_1: str | None = None, id_2: str | None = None, as_dict=False) Dict | None

Returns/displays the differences between commits/branches.

For each tensor this contains information about the sample indexes that were added/modified as well as whether the tensor was created.

Parameters:
  • id_1 (str, Optional) – The first commit_id or branch name.

  • id_2 (str, Optional) – The second commit_id or branch name.

  • as_dict (bool, Optional) – If True, returns a dictionary of the differences instead of printing them. This dictionary will have two keys - “tensor” and “dataset” which represents tensor level and dataset level changes, respectively. Defaults to False.

Returns:

Optional[Dict]

Raises:

ValueError – If id_1 is None and id_2 is not None.

Note

  • If both id_1 and id_2 are None, the differences between the current state and the previous commit will be calculated. If you’re at the head of the branch, this will show the uncommitted changes, if any.

  • If only id_1 is provided, the differences between the current state and id_1 will be calculated. If you’re at the head of the branch, this will take into account the uncommitted changes, if any.

  • If only id_2 is provided, a ValueError will be raised.

  • If both id_1 and id_2 are provided, the differences between id_1 and id_2 will be calculated.

Note

A dictionary of the differences between the commits/branches is returned if as_dict is True.

  • If id_1 and id_2 are None, a dictionary containing the differences between the current state and the previous commit will be returned.

  • If only id_1 is provided, a dictionary containing the differences in the current state and id_1 respectively will be returned.

  • If only id_2 is provided, a ValueError will be raised.

  • If both id_1 and id_2 are provided, a dictionary containing the differences in id_1 and id_2 respectively will be returned.

None is returned if as_dict is False.

Example of a dict returned:

>>> {
...    "image": {"data_added": [3, 6], "data_updated": {0, 2}, "created": False, "info_updated": False, "data_transformed_in_place": False},
...    "label": {"data_added": [0, 3], "data_updated": {}, "created": True, "info_updated": False, "data_transformed_in_place": False},
...    "other/stuff" : {"data_added": [3, 3], "data_updated": {1, 2}, "created": True, "info_updated": False, "data_transformed_in_place": False},
... }
  • Here, “data_added” is a range of sample indexes that were added to the tensor.

    • For example [3, 6] means that sample 3, 4 and 5 were added.

    • Another example [3, 3] means that no samples were added as the range is empty.

  • “data_updated” is a set of sample indexes that were updated.

    • For example {0, 2} means that sample 0 and 2 were updated.

  • “created” is a boolean that is True if the tensor was created.

  • “info_updated” is a boolean that is True if the info of the tensor was updated.

  • “data_transformed_in_place” is a boolean that is True if the data of the tensor was transformed in place.

extend(samples: Dict[str, Any], skip_ok: bool = False)

Appends multiple rows of samples to mutliple tensors at once. This method expects all tensors being updated to be of the same length.

Parameters:
  • samples (Dict[str, Any]) – Dictionary with tensor names as keys and samples as values.

  • skip_ok (bool) – Skip tensors not in samples if set to True.

Raises:
  • KeyError – If any tensor in the dataset is not a key in samples and skip_ok is False.

  • TensorDoesNotExistError – If tensor in samples does not exist.

  • ValueError – If all tensors being updated are not of the same length.

  • NotImplementedError – If an error occurs while writing tiles.

  • Exception – Error while attempting to rollback appends.

filter(function: Callable | str, num_workers: int = 0, scheduler: str = 'threaded', progressbar: bool = True, save_result: bool = False, result_path: str | None = None, result_ds_args: dict | None = None)

Filters the dataset in accordance of filter function f(x: sample) -> bool

Parameters:
  • function (Callable, str) – Filter function that takes sample as argument and returns True / False if sample should be included in result. Also supports simplified expression evaluations. See deeplake.core.query.query.DatasetQuery for more details.

  • num_workers (int) – Level of parallelization of filter evaluations. 0 indicates in-place for-loop evaluation, multiprocessing is used otherwise.

  • scheduler (str) – Scheduler to use for multiprocessing evaluation. “threaded” is default.

  • progressbar (bool) – Display progress bar while filtering. True is default.

  • save_result (bool) – If True, result of the filter will be saved to a dataset asynchronously.

  • result_path (Optional, str) – Path to save the filter result. Only applicable if save_result is True.

  • result_ds_args (Optional, dict) – Additional args for result dataset. Only applicable if save_result is True.

Returns:

View of Dataset with elements that satisfy filter function.

Example

Following filters are identical and return dataset view where all the samples have label equals to 2.

>>> dataset.filter(lambda sample: sample.labels.numpy() == 2)
>>> dataset.filter('labels == 2')
flush()

Necessary operation after writes if caches are being used. Writes all the dirty data from the cache layers (if any) to the underlying storage. Here dirty data corresponds to data that has been changed/assigned and but hasn’t yet been sent to the underlying storage.

get_commit_details(commit_id) Dict

Get details of a particular commit.

Parameters:

commit_id (str) – commit id of the commit.

Returns:

Dictionary of details with keys - commit, author, time, message.

Return type:

Dict

Raises:

KeyError – If given commit_id is was not found in the dataset.

get_creds_keys() List[str]

Returns the list of creds keys added to the dataset. These are used to fetch external data in linked tensors

get_view(id: str) ViewEntry

Returns the dataset view corresponding to id.

Examples

>>> # save view
>>> ds[:100].save_view(id="first_100")
>>> # load view
>>> first_100 = ds.get_view("first_100").load()
>>> # 100
>>> print(len(first_100))

See Dataset.save_view() to learn more about saving views.

Parameters:

id (str) – id of required view.

Returns:

ViewEntry

Raises:

KeyError – If no such view exists.

get_views(commit_id: str | None = None) List[ViewEntry]

Returns list of views stored in this Dataset.

Parameters:

commit_id (str, optional) –

  • Commit from which views should be returned.

  • If not specified, views from current commit is returned.

  • If not specified, views from the currently checked out commit will be returned.

Returns:

List of ViewEntry instances.

Return type:

List[ViewEntry]

property groups: Dict[str, Dataset]

All sub groups in this group

property has_head_changes

Returns True if currently at head node and uncommitted changes are present.

property info

Returns the information about the dataset.

property is_view: bool

Returns True if this dataset is a view and False otherwise.

load_view(id: str, optimize: bool | None = False, tensors: List[str] | None = None, num_workers: int = 0, scheduler: str = 'threaded', progressbar: bool | None = True)

Loads the view and returns the Dataset by id. Equivalent to ds.get_view(id).load().

Parameters:
  • id (str) – id of the view to be loaded.

  • optimize (bool) – If True, the dataset view is optimized by copying and rechunking the required data before loading. This is necessary to achieve fast streaming speeds when training models using the dataset view. The optimization process will take some time, depending on the size of the data.

  • tensors (Optional, List[str]) – Tensors to be copied if optimize is True. By default all tensors are copied.

  • num_workers (int) – Number of workers to be used for the optimization process. Only applicable if optimize=True. Defaults to 0.

  • scheduler (str) – The scheduler to be used for optimization. Supported values include: ‘serial’, ‘threaded’, ‘processed’ and ‘ray’. Only applicable if optimize=True. Defaults to ‘threaded’.

  • progressbar (bool) – Whether to use progressbar for optimization. Only applicable if optimize=True. Defaults to True.

Returns:

The loaded view.

Return type:

Dataset

Raises:

KeyError – if view with given id does not exist.

log()

Displays the details of all the past commits.

property max_len

Return the maximum length of the tensor

merge(target_id: str, conflict_resolution: str | None = None, delete_removed_tensors: bool = False, force: bool = False)

Merges the target_id into the current dataset.

Parameters:
  • target_id (str) – The commit_id or branch to merge.

  • conflict_resolution (str, Optional) –

    • The strategy to use to resolve merge conflicts.

    • Conflicts are scenarios where both the current dataset and the target id have made changes to the same sample/s since their common ancestor.

    • Must be one of the following
      • None - this is the default value, will raise an exception if there are conflicts.

      • ”ours” - during conflicts, values from the current dataset will be used.

      • ”theirs” - during conflicts, values from target id will be used.

  • delete_removed_tensors (bool) – If True, deleted tensors will be deleted from the dataset.

  • force (bool) –

    • Forces merge.

    • force=True will have these effects in the following cases of merge conflicts:
      • If tensor is renamed on target but is missing from HEAD, renamed tensor will be registered as a new tensor on current branch.

      • If tensor is renamed on both target and current branch, tensor on target will be registered as a new tensor on current branch.

      • If tensor is renamed on target and a new tensor of the new name was created on the current branch, they will be merged.

Raises:
  • Exception – if dataset is a filtered view.

  • ValueError – if the conflict resolution strategy is not one of the None, “ours”, or “theirs”.

property meta: DatasetMeta

Returns the metadata of the dataset.

property min_len

Return the minimum length of the tensor

property num_samples: int

Returns the length of the smallest tensor. Ignores any applied indexing and returns the total length.

property parent

Returns the parent of this group. Returns None if this is the root dataset.

property pending_commit_id: str

The commit_id of the next commit that will be made to the dataset. If you’re not at the head of the current branch, this will be the same as the commit_id.

pop(index: int | None = None)

Removes a sample from all the tensors of the dataset. For any tensor if the index >= len(tensor), the sample won’t be popped from it.

Parameters:

index (int, Optional) – The index of the sample to be removed. If it is None, the index becomes the length of the longest tensor - 1.

Raises:

IndexError – If the index is out of range.

populate_creds(creds_key: str, creds: dict)

Populates the creds key added in add_creds_key with the given creds. These creds are used to fetch the external data. This needs to be done everytime the dataset is reloaded for datasets that contain links to external data.

Examples

>>> # create/load a dataset
>>> ds = deeplake.dataset("path/to/dataset")
>>> # add a new creds key
>>> ds.add_creds_key("my_s3_key")
>>> # populate the creds
>>> ds.populate_creds("my_s3_key", {"aws_access_key_id": "my_access_key", "aws_secret_access_key": "my_secret_key"})
pytorch(transform: Callable | None = None, tensors: Sequence[str] | None = None, tobytes: bool | Sequence[str] = False, num_workers: int = 1, batch_size: int = 1, drop_last: bool = False, collate_fn: Callable | None = None, pin_memory: bool = False, shuffle: bool = False, buffer_size: int = 2048, use_local_cache: bool = False, use_progress_bar: bool = False, return_index: bool = True, pad_tensors: bool = False, transform_kwargs: Dict[str, Any] | None = None)

Converts the dataset into a pytorch Dataloader.

Parameters:
  • transform (Callable, Optional) – Transformation function to be applied to each sample.

  • tensors (List, Optional) – Optionally provide a list of tensor names in the ordering that your training script expects. For example, if you have a dataset that has “image” and “label” tensors, if tensors=[“image”, “label”], your training script should expect each batch will be provided as a tuple of (image, label).

  • tobytes (bool) – If True, samples will not be decompressed and their raw bytes will be returned instead of numpy arrays. Can also be a list of tensors, in which case those tensors alone will not be decompressed.

  • num_workers (int) – The number of workers to use for fetching data in parallel.

  • batch_size (int) – Number of samples per batch to load. Default value is 1.

  • drop_last (bool) – Set to True to drop the last incomplete batch, if the dataset size is not divisible by the batch size. if False and the size of dataset is not divisible by the batch size, then the last batch will be smaller. Default value is False. Read torch.utils.data.DataLoader docs for more details.

  • collate_fn (Callable, Optional) – merges a list of samples to form a mini-batch of Tensor(s). Used when using batched loading from a map-style dataset. Read torch.utils.data.DataLoader docs for more details.

  • pin_memory (bool) – If True, the data loader will copy Tensors into CUDA pinned memory before returning them. Default value is False. Read torch.utils.data.DataLoader docs for more details.

  • shuffle (bool) – If True, the data loader will shuffle the data indices. Default value is False. Details about how Deep Lake shuffles data can be found at https://docs.activeloop.ai/how-hub-works/shuffling-in-ds.pytorch

  • buffer_size (int) – The size of the buffer used to shuffle the data in MBs. Defaults to 2048 MB. Increasing the buffer_size will increase the extent of shuffling.

  • use_local_cache (bool) – If True, the data loader will use a local cache to store data. The default cache location is ~/.activeloop/cache, but it can be changed by setting the LOCAL_CACHE_PREFIX environment variable. This is useful when the dataset can fit on the machine and we don’t want to fetch the data multiple times for each iteration. Default value is False

  • use_progress_bar (bool) – If True, tqdm will be wrapped around the returned dataloader. Default value is True.

  • return_index (bool) – If True, the returned dataloader will have a key “index” that contains the index of the sample(s) in the original dataset. Default value is True.

  • pad_tensors (bool) – If True, shorter tensors will be padded to the length of the longest tensor. Default value is False.

  • transform_kwargs (optional, Dict[str, Any]) – Additional kwargs to be passed to transform.

Returns:

A torch.utils.data.DataLoader object.

Raises:

EmptyTensorError – If one or more tensors being passed to pytorch are empty.

Note

Pytorch does not support uint16, uint32, uint64 dtypes. These are implicitly type casted to int32, int64 and int64 respectively. This spins up it’s own workers to fetch data.

query(query_string: str)

Returns a sliced Dataset with given query results.

It allows to run SQL like queries on dataset and extract results. See supported keywords and the Tensor Query Language documentation here.

Parameters:

query_string (str) – An SQL string adjusted with new functionalities to run on the given Dataset object

Returns:

A Dataset object.

Return type:

Dataset

Examples

Query from dataset all the samples with lables other than 5

>>> import deeplake
>>> ds = deeplake.load('hub://activeloop/fashion-mnist-train')
>>> query_ds = ds.query("select * where labels != 5")

Query from dataset first appeard 1000 samples where the categories is car and 1000 samples where the categories is motorcycle

>>> ds_train = deeplake.load('hub://activeloop/coco-train')
>>> query_ds_train = ds_train.query("(select * where contains(categories, 'car') limit 1000) union (select * where contains(categories, 'motorcycle') limit 1000)")
property read_only

Returns True if dataset is in read-only mode and False otherwise.

rechunk(tensors: str | List[str] | None = None, num_workers: int = 0, scheduler: str = 'threaded', progressbar: bool = True)

Rewrites the underlying chunks to make their sizes optimal. This is usually needed in cases where a lot of updates have been made to the data.

Parameters:
  • tensors (str, List[str], Optional) – Name/names of the tensors to rechunk. If None, all tensors in the dataset are rechunked.

  • num_workers (int) – The number of workers to use for rechunking. Defaults to 0. When set to 0, it will always use serial processing, irrespective of the scheduler.

  • scheduler (str) – The scheduler to be used for rechunking. Supported values include: ‘serial’, ‘threaded’, ‘processed’ and ‘ray’. Defaults to ‘threaded’.

  • progressbar (bool) – Displays a progress bar If True (default).

rename(path: str | Path)

Renames the dataset to path.

Example

>>> ds = deeplake.load("hub://username/dataset")
>>> ds.rename("hub://username/renamed_dataset")
Parameters:

path (str, pathlib.Path) – New path to the dataset.

Raises:

RenameError – If path points to a different directory.

rename_group(name: str, new_name: str) None

Renames group with name name to new_name

Parameters:
  • name (str) – Name of group to be renamed.

  • new_name (str) – New name of group.

Raises:
rename_tensor(name: str, new_name: str) Tensor

Renames tensor with name name to new_name

Parameters:
  • name (str) – Name of tensor to be renamed.

  • new_name (str) – New name of tensor.

Returns:

Renamed tensor.

Return type:

Tensor

Raises:
reset()

Resets the uncommitted changes present in the branch.

Note

The uncommitted data is deleted from underlying storage, this is not a reversible operation.

property root

Returns the root dataset of a group.

property sample_indices

Returns all the indices pointed to by this dataset view.

save_view(message: str | None = None, path: str | Path | None = None, id: str | None = None, optimize: bool = False, tensors: List[str] | None = None, num_workers: int = 0, scheduler: str = 'threaded', verbose: bool = True, **ds_args) str

Saves a dataset view as a virtual dataset (VDS)

Examples

>>> # Save to specified path
>>> vds_path = ds[:10].save_view(path="views/first_10", id="first_10")
>>> vds_path
views/first_10
>>> # Path unspecified
>>> vds_path = ds[:100].save_view(id="first_100", message="first 100 samples")
>>> # vds_path = path/to/dataset
>>> # Random id
>>> vds_path = ds[:100].save_view()
>>> # vds_path = path/to/dataset/.queries/92f41922ed0471ec2d27690b7351fc96bea060e6c5ee22b14f7ffa5f291aa068

See Dataset.get_view() to learn how to load views by id. These virtual datasets can also be loaded from their path like normal datasets.

Parameters:
  • message (Optional, str) – Custom user message.

  • path (Optional, str, pathlib.Path) –

    • The VDS will be saved as a standalone dataset at the specified path.

    • If not specified, the VDS is saved under .queries subdirectory of the source dataset’s storage.

    • If the user doesn’t have write access to the source dataset and the source dataset is a Deep Lake cloud dataset, then the VDS is saved is saved under the user’s Deep Lake account and can be accessed using deeplake.load(f"hub://{username}/queries/{query_hash}").

  • id (Optional, str) – Unique id for this view. Random id will be generated if not specified.

  • optimize (bool) –

    • If True, the dataset view will be optimized by copying and rechunking the required data. This is necessary to achieve fast streaming speeds when training models using the dataset view. The optimization process will take some time, depending on the size of the data.

    • You can also choose to optimize the saved view later by calling its ViewEntry.optimize() method.

  • tensors (List, optional) – Names of tensors (and groups) to be copied. If not specified all tensors are copied.

  • num_workers (int) – Number of workers to be used for optimization process. Applicable only if optimize=True. Defaults to 0.

  • scheduler (str) – The scheduler to be used for optimization. Supported values include: ‘serial’, ‘threaded’, ‘processed’ and ‘ray’. Only applicable if optimize=True. Defaults to ‘threaded’.

  • verbose (bool) – If True, logs will be printed. Defaults to True.

  • ds_args (dict) – Additional args for creating VDS when path is specified. (See documentation for deeplake.dataset())

Returns:

Path to the saved VDS.

Return type:

str

Raises:

Note

Specifying path makes the view external. External views cannot be accessed using the parent dataset’s Dataset.get_view(), Dataset.load_view(), Dataset.delete_view() methods. They have to be loaded using deeplake.load().

size_approx()

Estimates the size in bytes of the dataset. Includes only content, so will generally return an under-estimate.

summary()

Prints a summary of the dataset.

tensorflow(tensors: Sequence[str] | None = None, tobytes: bool | Sequence[str] = False, fetch_chunks: bool = True)

Converts the dataset into a tensorflow compatible format.

See https://www.tensorflow.org/api_docs/python/tf/data/Dataset

Parameters:
  • tensors (List, Optional) – Optionally provide a list of tensor names in the ordering that your training script expects. For example, if you have a dataset that has “image” and “label” tensors, if tensors=["image", "label"], your training script should expect each batch will be provided as a tuple of (image, label).

  • tobytes (bool) – If True, samples will not be decompressed and their raw bytes will be returned instead of numpy arrays. Can also be a list of tensors, in which case those tensors alone will not be decompressed.

  • fetch_chunks – See fetch_chunks argument in deeplake.core.tensor.Tensor.numpy()

Returns:

tf.data.Dataset object that can be used for tensorflow training.

property tensors: Dict[str, Tensor]

All tensors belonging to this group, including those within sub groups. Always returns the sliced tensors.

property token

Get attached token of the dataset

update_creds_key(old_creds_key: str, new_creds_key: str)

Replaces the old creds key with the new creds key. This is used to replace the creds key used for external data.

visualize(width: int | str | None = None, height: int | str | None = None)

Visualizes the dataset in the Jupyter notebook.

Parameters:
  • width – Union[int, str, None] Optional width of the visualizer canvas.

  • height – Union[int, str, None] Optional height of the visualizer canvas.

Raises:

Exception – If the dataset is not a Deep Lake cloud dataset and the visualization is attempted in colab.

DeepLakeCloudDataset

class deeplake.core.dataset.DeepLakeCloudDataset

Bases: Dataset

Subclass of Dataset. Deep Lake cloud datasets are those datasets which are stored on Activeloop servers, their paths look like: hub://username/dataset_name.

add_creds_key(creds_key: str, managed: bool = False)

Adds a new creds key to the dataset. These keys are used for tensors that are linked to external data.

Examples

>>> # create/load a dataset
>>> ds = deeplake.dataset("hub://username/dataset")
>>> # add a new creds key
>>> ds.add_creds_key("my_s3_key")
Parameters:
  • creds_key (str) – The key to be added.

  • managed (bool) – If True, the creds corresponding to the key will be fetched from activeloop platform. Note, this is only applicable for datasets that are connected to activeloop platform. Defaults to False.

change_creds_management(creds_key: str, managed: bool)

Changes the management status of the creds key.

Parameters:
  • creds_key (str) – The key whose management status is to be changed.

  • managed (bool) – The target management status. If True, the creds corresponding to the key will be fetched from activeloop platform.

Raises:
  • ValueError – If the dataset is not connected to activeloop platform.

  • KeyError – If the creds key is not present in the dataset.

Examples

>>> # create/load a dataset
>>> ds = deeplake.dataset("hub://username/dataset")
>>> # add a new creds key
>>> ds.add_creds_key("my_s3_key")
>>> # Populate the name added with creds dictionary
>>> # These creds are only present temporarily and will have to be repopulated on every reload
>>> ds.populate_creds("my_s3_key", {})
>>> # Change the management status of the key to True. Before doing this, ensure that the creds have been created on activeloop platform
>>> # Now, this key will no longer use the credentials populated in the previous step but will instead fetch them from activeloop platform
>>> # These creds don't have to be populated again on every reload and will be fetched every time the dataset is loaded
>>> ds.change_creds_management("my_s3_key", True)
property client

Returns the client of the dataset.

connect(*args, **kwargs)

Connect a Deep Lake cloud dataset through a deeplake path.

Examples

>>> # create/load an s3 dataset
>>> s3_ds = deeplake.dataset("s3://bucket/dataset")
>>> ds = s3_ds.connect(dest_path="hub://my_org/dataset", creds_key="my_managed_credentials_key")
>>> # or
>>> ds = s3_ds.connect(org_id="my_org", creds_key="my_managed_credentials_key")
Parameters:
  • creds_key (str) – The managed credentials to be used for accessing the source path.

  • dest_path (str, optional) – The full path to where the connected Deep Lake dataset will reside. Can be: a Deep Lake path like hub://organization/dataset

  • org_id (str, optional) – The organization to where the connected Deep Lake dataset will be added.

  • ds_name (str, optional) – The name of the connected Deep Lake dataset. Will be infered from dest_path or src_path if not provided.

  • token (str, optional) – Activeloop token used to fetch the managed credentials.

Returns:

The connected Deep Lake dataset.

Return type:

Dataset

Raises:
  • InvalidSourcePathError – If the dataset’s path is not a valid s3 or gcs path.

  • InvalidDestinationPathError – If dest_path, or org_id and ds_name do not form a valid Deep Lake path.

delete(large_ok=False)

Deletes the entire dataset from the cache layers (if any) and the underlying storage. This is an IRREVERSIBLE operation. Data once deleted can not be recovered.

Parameters:

large_ok (bool) – Delete datasets larger than 1 GB. Defaults to False.

Raises:

DatasetTooLargeToDelete – If the dataset is larger than 1 GB and large_ok is False.

property is_actually_cloud: bool

Datasets that are connected to Deep Lake cloud can still technically be stored anywhere. If a dataset is in Deep Lake cloud but stored without hub:// prefix, it should only be used for testing.

rename(path)

Renames the dataset to path.

Example

>>> ds = deeplake.load("hub://username/dataset")
>>> ds.rename("hub://username/renamed_dataset")
Parameters:

path (str, pathlib.Path) – New path to the dataset.

Raises:

RenameError – If path points to a different directory.

property token

Get attached token of the dataset

update_creds_key(old_creds_key: str, new_creds_key: str)

Replaces the old creds key with the new creds key. This is used to replace the creds key used for external data.

visualize(width: int | str | None = None, height: int | str | None = None)

Visualizes the dataset in the Jupyter notebook.

Parameters:
  • width – Union[int, str, None] Optional width of the visualizer canvas.

  • height – Union[int, str, None] Optional height of the visualizer canvas.

Raises:

Exception – If the dataset is not a Deep Lake cloud dataset and the visualization is attempted in colab.

ViewEntry

class deeplake.core.dataset.ViewEntry

Represents a view saved inside a dataset.

delete()

Deletes the view.

property id: str

Returns id of the view.

load(verbose=True)

Loads the view and returns the Dataset.

Parameters:

verbose (bool) – If True, logs will be printed. Defaults to True.

Returns:

Loaded dataset view.

Return type:

Dataset

property message: str

Returns the message with which the view was saved.

optimize(tensors: List[str] | None = None, unlink=True, num_workers=0, scheduler='threaded', progressbar=True)

Optimizes the dataset view by copying and rechunking the required data. This is necessary to achieve fast streaming speeds when training models using the dataset view. The optimization process will take some time, depending on the size of the data.

Example

>>> # save view
>>> ds[:10].save_view(view_id="first_10")
>>> # optimize view
>>> ds.get_view("first_10").optimize()
>>> # load optimized view
>>> ds.load_view("first_10")
Parameters:
  • tensors (List[str]) – Tensors required in the optimized view. By default all tensors are copied.

  • unlink (bool) –

    • If True, this unlinks linked tensors (if any) by copying data from the links to the view.

    • This does not apply to linked videos. Set deeplake.constants._UNLINK_VIDEOS to True to change this behavior.

  • num_workers (int) – Number of workers to be used for the optimization process. Defaults to 0.

  • scheduler (str) – The scheduler to be used for optimization. Supported values include: ‘serial’, ‘threaded’, ‘processed’ and ‘ray’. Only applicable if optimize=True. Defaults to ‘threaded’.

  • progressbar (bool) – Whether to display a progressbar.

Returns:

ViewEntry