Debugging XAI - Can XAI methods detect confounding variables?

1. Setup

Install packages from the environment.yml file - we use conda, but pipenv or venv will work too.

conda env create --name envname --file=environment.yml

2. Data generation

Currently the data can be downloaded internally in the correct structure on MS Teams, or also from Kaggle, albeit in a different file structure (we combine all of the training and testing images from the original source into the structure ./images/dog/ or ./images/cat/ and do the train/val/test splits afterwards).

To generate the data for experiments, use the following scripts:

Watermark experiments

python generate_watermarks.py {split} {rescaled}

where split is an integer [0..10] describing which 'shuffling' of the data should be used for generation (via random seeds). rescaled is a string, either 'rescaled' to rescale the data to the range [-1,1] or blank/anything else to use the default min-max scaling to the range [0,1].

For variable-position watermarks, use:

python generate_watermarks_variable.py {split} {rescaled}

where the same parameterisation applies as above.

COCO Lightness Experiments

Requires the COCO 2017 train and val data as well as annotations placed in the directory structure as

project  
│
└───coco
│   │
│   └─── annotations
│       │   captions_train2017.json
│       │   ...
│   └─── train2017
│       │   000000000009.jpg
│       │   ...
│   └─── val2017
│       │   000000000139.jpg
│       │   ...

Then, run the generation script as:

python generate_coco_splits.py {split} {rescaled}

Where the same parameterisation for split (shuffling of the data via random seeds) and rescaling apply. Here, the rescaling to [-1,1] (if specified) is done to the HLS lightness channel only.

3. Model Training

Watermark experiments

Currently this is only pushed to git for the non-rescaled variants, but this will be updated in due course.

python train_server_splits.py {split} {all|confounder|suppressor|no_watermark}

or

python train_server_variable.py {split} {all|confounder|suppressor|no_watermark}

where the first argument is, as before, the split of data to train over, and the second argument refers to whether to train all three model variants, or just one of confounder, suppressor, or no_watermark. train_server_variable.py is for the variable-position watermark, though in the future this should be unified to one script.

COCO lightness experiments

python coco_train_server_splits.py {split} {conf|sup|norm}

where conf is for training the confounder model, sup for the suppressor, or norm for the normalised brightness model.

Experiments and results

All XAI method execution and resulting energy metric calculations are run here.

Watermark experiments

Currently this is only pushed to git for the non-rescaled variants, but this will be updated in due course.

python calculate_watermarks_energy.py {split} {model_ind} {variable} Where split is defined as before, the model_ind is also an integer [0..4] for which trained model to use, and variable makes use of the variable-position watermark data if set.

COCO lightness experiments

python calculate_coco_energy.py {split} {model_ind}