Add support for pre-computed fairness majority masks

[f4str]

Modify the poisoning scenario to accept the majority_masks field which will attempt to read a pre-computed npz file which contains the majority masks for the train and test data. This will override the explanatory model if provided.

The npz file must be placed in the root of the armory directory (under armory/). So the provided path:

"adhoc": {
    ...
    "compute_fairness_metrics": true,
    "majority_masks": "audio_majority_masks.npz",
    ...
}

will look for armory/audio_majority_masks.npz. The file will have two keyword arg arrays: train and test which are boolean arrays the same length as the train and test sets, respectively.

This works on both docker and non-docker mode.

[f4str]

A zip file containing the majority masks for the speech commands dataset and a sample config file for testing: majority_masks.zip

[swsuggs]

Thank you Farhan, this is great. I'm just adding the npz file to its own directory with a bit of extra code to facilitate loading it.