Getting Started with Hub


What is Hub?

The fastest way to access and manage datasets for PyTorch and TensorFlow

Hub provides fast access to the state-of-the-art datasets for Deep Learning, enabling data scientists to manage them, build scalable data pipelines and connect to Pytorch and Tensorflow.


Your project must have the “Internet access” upgrade enabled in order to connect to an online service like Activeloop’s hub from within your project!


This is user-contributed content. Credits go to Activeloop!


  1. Install Hub pip3 install hub

  2. Register and authenticate to upload datasets to Activeloop store.

    activeloop register
    activeloop login

    Alternatively, add username and password as arguments (use on platforms like Kaggle).

    activeloop login -u username -p password
  3. Load a dataset

    import hub
    ds = hub.Dataset("activeloop/cifar10_train")
    print(ds["label", :10].compute())
    print(ds["id", 1234].compute())
    print(ds["image", 4321].compute())
  4. Create a dataset

    import numpy as np
    import hub from hub.schema import ClassLabel, Image
    my_schema = { "image": Image((28, 28)),
                "label": ClassLabel(num_classes=10), }
    url = "./data/examples/quickstart" # write your {username}/{dataset_name} to make it remotely accessible
    ds = hub.Dataset(url, shape=(1000,), schema=my_schema)
    for i in range(len(ds)):
        ds["image", i] = np.ones((28, 28), dtype="uint8")
        ds["label", i] = 3
    print(ds["image", 5].compute())
    print(ds["label", 100:110].compute())

This code creates dataset with 1000 samples in “./data/examples/new_api_intro” folder with overwrite mode. Once the dataset is ready, you may read, write and loop over it.

You can also transfer a dataset from TFDS (as below) and convert it from/to Tensorflow or PyTorch.

import hub
import tensorflow as tf

out_ds = hub.Dataset.from_tfds('mnist', split='test+train', num=1000)
res_ds ="username/mnist") # res_ds is now a usable hub dataset

Data Storage

The first positional argument to declare a Hub dataset is url.


If url parameter has the form of username/dataset, the dataset will be stored in our cloud storage.

url = 'username/dataset'
ds = hub.Dataset(url, shape=(1000,), schema=my_schema)

You can also create or load a dataset locally or in S3, MinIO, Google Cloud Storage and Azure. In case you choose other remote storage platforms, you will need to provide the corresponding credentials as a token argument during Dataset creation or loading. It can be a filepath to your credentials or a dict.

Local storage

To store datasets locally, let the url parameter be a local path.

url = './datasets/'
ds = hub.Dataset(url, shape=(1000,), schema=my_schema)


python url = 's3://new_dataset'  # your s3 path
ds = hub.Dataset(url, shape=(1000,), schema=my_schema, token={"aws_access_key_id": "...",                                                               "aws_secret_access_key": "...",                                                               ...})``


url = 's3://new_dataset'  # minio also uses *s3://* prefix
ds = hub.Dataset(url, shape=(1000,), schema=my_schema, token={"aws_access_key_id": "your_minio_access_key",
                                                                "aws_secret_access_key": "your_minio_secret_key",
                                                                "endpoint_url": "your_minio_url:port",

Google Cloud Storage

url = 'gcs://new_dataset' # your google storage (gs://) path
ds = hub.Dataset(url, shape=(1000,), schema=my_schema, token="/path/to/credentials")


url = '' # Azure link
ds = hub.Dataset(url, shape=(1000,), schema=my_schema, token="/path/to/credentials")


Schema is a required attribute that describes what a dataset consists of. This is how you can create a simple schema:

from hub.schema import ClassLabel, Image, BBox, Text

my_schema = {
    'kind': ClassLabel(names=["cows", "horses"]),
    'animal': Image(shape=(512, 256, 3)),
    'eyes': BBox(),
    'description': Text(max_shape=(100,))


Shape is another required attribute of a dataset. It simply specifies how large a dataset is. The rules associated with shapes are derived from numpy.

Dataset Access, Modification and Deletion

In order to access the data from the dataset, you should use .compute() on a portion of the dataset: ds['key', :5].compute().

You can modify the data to the dataset with a regular assignment operator or by performing more sophisticated transforms.

You can delete your dataset with .delete() or through Activeloop’s app on in a dataset overview tab.

Flush, Commit and Close

Hub Datasets have three methods to push the final changes to the selected storage.

The most fundamental method, .flush() saves changes from cache to the dataset final storage and does not invalidate dataset object. It means that you can continue working on your data and pushing it later on.

.commit() saves the changes into a new version of a dataset that you may go back to later on if you want to.

In rare cases, you may also use .close() to invalidate the dataset object after saving the changes.

If you prefer flushing to be taken care for you, wrap your operations on the dataset with the with statement in this fashion:

with hub.Dataset(...) as ds:

Other information

For more information see Hub documentation .

Join our Slack community for help from Activeloop team and other users as well as dataset management/preprocessing tips and tricks.

For feature requests or bug reports, please open a new GitHub issue.