Tub

This is the standard donkey data store. The "data" folder is what we call a "tub"

Accepted Types

The following datatypes are supported.

  • str
  • int
  • float / np.float
  • image_arrays and arrays (np.ndarray)
  • image (jpeg / png)

The Tub is an append only format, that is optimized for reads (to speed up training models). It maintains indexes for records, and uses memory mapped files.

The Tub exposes an Iterator that can be used to read records. These iterators can be further used by Pipelines to do arbitrary transformations of data prior to training (for data augumentation).

Example

from donkeycar.parts.tub_v2 import Tub

# Here we define records that have a single `input` of type `int`.
inputs = ['input']
types = ['int']
tub = Tub(path, inputs, types)