AUKUS Container

A user of T&E tools requested a container that would accept an input image to generate saliency maps. Once the maps are generated, the user expects the maps to be saved to disk and the container will shut down. In order to fulfill this request, the AUKUS container was created.

Given an input image, a XAITK saliency configuration file, and a HuggingFace model name, the AUKUS container is able to find detections and generate saliency maps for the input image. Each saliency map will be saved to a given output directory as individual images. Once all saliency maps 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/ xaitk-jatic This will mount the inputs to the correct locations and use the default args. The default args will attempt to load an image named example_img.jpeg, save saliency maps to /root/output, load a config file named config.json, and load the facebook/detr-resnet-50 model. The example_img.jpeg and config.json must be in the directory mounted to /root/input/.

If the user wants to use different arguments, the container expects the following arguments:

  • image_file: input image

  • output_dir: directory to write saliency maps to

  • config_file: configuration file specifying the GenerateObjectDetectorBlackboxSaliency for saliency map generation

  • hugging_face_model_name: name of HuggingFace model to use

Please note the values for image_file and config_file should be written from the perspective of the container (i.e. /path/on/container/image_file.jpeg instead of /path/on/local/machine/image_file.jpeg)

Limitations

Currently, the main limitation of the AUKUS container is only being able to use HuggingFace models for object detections. While this allows for some freedom of choice in model, users will not only be limited to use HuggingFace models, but also need to access HuggingFace models during execution.