Purpose

Expand the parse and export suite with more classification formats that are accepted by FiftyOne. These include three new formats:

• Flat collection of folders, each folder corresponding to a class name: before, classification_directory_parser was able to do this if these sub-folders were divided into train/valid/test/ subdirectories, now this works with a flat structure. The splits are generated at parse-time randomly.
• train/validation/test subdirectories with a data folder and a labels.json residing in each
• Flat structure: without the splitting of train/validation/test, just a flat structure with a data folder where the images reside and a labels.json file. Again the splits are randomly assigned

I did not have to implement the mechanism that randomly assigns splits for flat directory structure because this is handled by the LuxonisDataset make_splits() method which only assigns splits to unassigned ids. Here the default ratio is {"train": 0.8, "val": 0.1, "test": 0.1}

• Implemented the corresponding ClassificationDirectoryExporter

Specification

Dependencies & Potential Impact

Deployment Plan

Testing & Validation

After parsing then exporting, then opening a session with the exported datasets:

MNIST:

Caltech101:

Imagenet-sample: