I was using Hub, a dataset format for AI that allows data streaming to GPUs without sacrificing performance.
I have been using Hub for image datasets and would like to try to use the product for other data types.
How would Hub work for different data types such as 3D objects, audio, video, etc?
The following Activeloop Hub doc has an example of how to upload image datasets to Hub and I am using a similar approach for working with my image dataset.
You can currently upload any type of data as uncompressed arrays. Simply don't assign the Hub type (htype) which is a property of each tensor that gives info to the Hub and Activeloop Platform about how to optimally store, parse, and visualize Hub datasets.
If you do this you'll be able to .append any array you like, as long as len(ds.tensor[i].numpy().shape) is equal for all samples. So, the samples don't have to have the same shape, but they need to have the same number of dimensions.
Going forward, Activeloop Hub will be adding support for application-specific multi-dimensional htypes such as video, image, and audio, etc. These will include an appropriate compression.
Hub has htype docs that can help you understand dimensionality & conventions for each type.
