2005 generate adversarial dataset

[treubig26]

What does this MR do and why?

Add engine to support creation of adversarial datasets (datasets with an evasion attack pre-applied to all samples for the ability to perform quicker real-time demonstrations).

Also refactors the batches from tuples of (x, y) to dictionaries (where the x_key and y_key are provided so armory knows how to obtain x/y from the dictionary). This is done to avoid reducing or dropping information from the original dataset.

Related issues

#2005

How to set up and validate locally

In the examples/src/charmory_examples/image_classification/eurosat_precomputed_pgd/ folder:

• Run create_adversarial_eurosat.py [convnext|swin|vit] to create an adversarial dataset
• Run run_against_pregenerated.py [convnext|swin|vit] [convnext|swin|vit] to perform evaluation using the adversarial dataset
  • first argument is the name of the model to be evaluated
  • second argument is the name of the model used to generate the adversarial dataset (i.e., what was used with create_adversarial_eurosat.py)
• Run run_with_attack.py [convnext|swin|vit] to perform inline attack evaluation (typical armory usage)

MR approval & acceptance process

Areas of focus for the review:

• Briefly look over changes to examples introduced by LightningEngine rename and evaluation.Dataset changes
• Tuple vs dict batch changes in src/charmory/data.py and src/charmory/tasks/base.py
• New adversarial dataset engine in src/charmory/engine/adversarial_dataset.py
• EuroSAT example in examples/src/charmory_examples/image_classification/eurosat_precomputed_pgd/
• New exports of target info in src/charmory/tasks/image_classification.py and src/charmory/tasks/object_detection.py
  • Not strictly necessary, but was put in place when the initial plan was to rely on mlflow exports and is still useful

Closes #2005

[mwartell]

There are two deprecation warnings on vit at least.

/home/msw/.virtualenvs/cdao/lib/python3.10/site-packages/transformers/models/vit/feature_extraction_vit.py:28: FutureWarning: The class ViTFeatureExtractor is deprecated and will be removed in version 5 of Transformers. Please use ViTImageProcessor instead.
/home/msw/.virtualenvs/cdao/lib/python3.10/site-packages/datasets/load.py:2086: FutureWarning: 'task' was deprecated in version 2.13.0 and will be removed in 3.0.0.

[treubig26]

There are two deprecation warnings on vit at least.

/home/msw/.virtualenvs/cdao/lib/python3.10/site-packages/transformers/models/vit/feature_extraction_vit.py:28: FutureWarning: The class ViTFeatureExtractor is deprecated and will be removed in version 5 of Transformers. Please use ViTImageProcessor instead. /home/msw/.virtualenvs/cdao/lib/python3.10/site-packages/datasets/load.py:2086: FutureWarning: 'task' was deprecated in version 2.13.0 and will be removed in 3.0.0.

Yeah, those are coming right out of HuggingFace because of the model's built-in configuration/metadata. I've seen similar deprecation warnings on a lot of the models picked at random from the HF hub (even from popular models). If HF goes through with the deprecated feature removal in version 3.0, there will be a lot of broken/unusuable models in their hub.

[mwartell]

A PR of Treubig cleanliness. Thanks.

I am not waiting for the dataset to load before merging. I'll report errors post hoc should I happen to see one.