Occlusion

Bases: MultitaskExplainerMixin, AttributionExplainer

The Occlusion method is a perturbation-based approach to interpret model behavior by analyzing the impact of removing or masking parts of the input text. The principle is simple: by systematically occluding (i.e., masking, deleting, or replacing) specific tokens or spans in the input and observing how the model's output changes, one can infer the relative importance of each part of the input to the model's behavior.

Reference: Zeiler and Fergus (2014). Visualizing and understanding convolutional networks. Paper

Examples:

>>> from interpreto import Granularity, Occlusion
>>> from interpreto.attributions import InferenceModes
>>> method = Occlusion(model, tokenizer, batch_size=4,
>>>                    inference_mode=InferenceModes.SOFTMAX,
>>>                    granularity=Granularity.WORD)
>>> explanations = method(text)

Parameters:

Name	Type	Description	Default
model	PreTrainedModel	model to explain	required
tokenizer	PreTrainedTokenizer	Hugging Face tokenizer associated with the model	required
batch_size	int	batch size for the attribution method	4
granularity	Granularity	granularity level of the perturbations (token, word, sentence, etc.)	WORD
inference_mode	Callable[[Tensor], Tensor]	The mode used for inference. It can be either one of LOGITS, SOFTMAX, or LOG_SOFTMAX. Use InferenceModes to choose the appropriate mode.	LOGITS
device	device	device on which the attribution method will be run	None