Custom dataset handles

The custom dataset handles are meant to handle file stored image datasets, and are all derived from BaseDataset. Available custom handles are:

coco.keypoints_dataset.KeypointsDataset	Handler for (a subset of) a COCO keypoints dataset.
coco.mask_dataset.ConceptDataset	Data handle for ground truth segmentations of visual concepts (body parts) generated from COCO keypoints.
broden.BrodenHandle	Handle to collect a sub-dataset of a dataset following Broden format.
fasseg.FASSEGHandle	Handle for FASSEG-like datasets.

Useful wrappers are:

ActivationDatasetWrapper

Wrapper for image datasets that will generate and yield activation maps.

Tips and tricks: Merge transformations

Also worth noting are the Merge dict transformations that allow boolean combination of masks and (boolean) classification labels. For details see their class documentation and the example given in Boolean Label Combinations.

Tips and tricks: Caching

All custom handles offer the option to specify a transforms_cache handle. This can be used to cache costly transformation or loading operations, e.g. in-memory or by dumping intermediate results to files. For details on available cache types and combination options have a look at the caching module. See also Dataset Caching

Examples

In the following, usage examples for some handles are shown. For details about the dataset formats have a look at the class documentations.

• MS COCO
• Broden
• Activation Map Storage and Handling
• Dataset Caching