AUKUS Container

To support users that require tools for an ML T&E workflow, we define a container that would accept an input dataset and apply perturbations to the entire dataset. The perturbed images should be saved to disk and then the container will shut down. In order to support this workflow, the AUKUS container was created.

Given a COCO dataset and a configuration file, the AUKUS container is able to generate perturbed images for each image in the dataset. Each perturbed image will be saved to a given output directory as a new COCO dataset. Once all perturbed images are saved, the container will terminate.

How to Use

To run the AUKUS container, use the following command: docker run -v /path/to/input:/root/input/:ro -v /path/to/output:/root/output/ nrtk-jatic This will mount the inputs to the correct locations and use the nrtk-perturber CLI script with the default args.

Default Behavior

COCO dataset loaded from the dataset directory. Must be in the directory mounted to /root/input/. The annotations file is expected to be named annotations.json at the root of the dataset directory. Optionally, image_metadata.json, also at the root of the dataset directory, containing a list of per-image dictionaries of metadata may be provided. If a perturber requires a specific piece of metadata, it should be specified via this file.
Configuration file loaded from nrtk_config.json. Must be in the directory mounted to /root/input/. The configuration file specifies the PerturbImageFactory parameters for image perturbations and should have the following structure:
```
{
  "type": "one-of-the-keys-below",
  "PerturbImageFactoryImpl1": {
    "param1": "val1",
    "param2": "val2"
  },
  "PerturbImageFactoryImpl2": {
    "p1": 4.5,
    "p2": null
  }
}
```
The “type” key is considered a special key that should always be present and it specifies one of the other keys within the same dictionary. Each other key in the dictionary should be the name of a Configurable inheriting class type. In this case, the “type” key specifies which PerturbImageFactory configuration to use. Each subdictionary represents the configuration for that specific PerturbImageFactory implementation.
For each perturber in the perturber factory provided by the configuration file, a sub-directory with a new COCO dataset is created with the corresponding perturbed images and detections in /root/output.

Customizing Behavior

If the user wants to use different input/output paths, the container expects the following arguments:

dataset_dir : Input COCO dataset

output_dir : Directory to store the perturbed dataset

config_file : JSON configuration file path

Note: The values for dataset_dir and config_file should be written from the perspective of the container (i.e. /path/on/container/dataset_dir/ instead of /path/on/local/machine/dataset_dir/)

Limitations

Currently, the AUKUS container supports the loading of only COCO datasets. Any existing dataset must be converted to a COCO dataset before using the AUKUS container. Please see KWCOCO documentation for more information on COCO datasets.