Explainer [source]

The Explainer module serves as a basis for all explainers implemented within the framework. When it is necessary to implement a custom explainer, the base class needs to be extended to enable the support of the framework's functionality. The method accumulates crucial attributes and methods, making it visible to the Recommender, and Optimizer modules.

List of available explainers:

Explainer	Supported Modules	Target Modalities
GradCam	Convolution	Vision, Time Series
GuidedGradCam	Convolution	Vision, Time Series
Gradient	Linear, Convolution, Recurrent, Transformer	Vision, Language, Time Series
GradientXInput	Linear, Convolution, Recurrent, Transformer	Vision, Language, Time Series
SmoothGrad	Linear, Convolution, Recurrent, Transformer	Vision, Language, Time Series
VarGrad	Linear, Convolution, Recurrent, Transformer	Vision, Language, Time Series
IntegratedGradients	Linear, Convolution, Recurrent, Transformer	Vision, Language, Time Series
LRPUniformEpsilon	Linear, Convolution, Recurrent, Transformer	Vision, Language, Time Series
LRPEpsilonPlus	Linear, Convolution, Recurrent, Transformer	Vision, Language, Time Series
LRPEpsilonGammaBox	Linear, Convolution, Recurrent, Transformer	Vision, Language, Time Series
LRPEpsilonAlpha2Beta1	Linear, Convolution, Recurrent, Transformer	Vision, Language, Time Series
RAP	Linear, Convolution, Recurrent, Transformer	Vision, Language, Time Series
KernelShap	Linear, Convolution, Recurrent, Transformer, Decision Trees	Vision, Language, Structured Data, Time Series
Lime	Linear, Convolution, Recurrent, Transformer, Decision Trees	Vision, Language, Structured Data, Time Series
AttentionRollout	Transformer	Vision, Language
TransformerAttribution	Transformer	Vision, Language

Usage

import torch
from torch.utils.data import DataLoader

from pnpxai.utils import set_seed
from pnpxai.explainers import LRPUniformEpsilon

from helpers import get_imagenet_dataset, get_torchvision_model

set_seed(seed=0)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# load model, dataset, and explainer
model, transform = get_torchvision_model("resnet18")
model = model.to(device)
explainer = LRPUniformEpsilon(model=model, epsilon=1e-6, n_classes=1000)

dataset = get_imagenet_dataset(transform=transform, subset_size=8)
loader = DataLoader(dataset, batch_size=8)
inputs, targets = next(iter(loader))
inputs, targets = inputs.to(device), targets.to(device)

# make explanation
attrs = explainer.attribute(inputs, targets)

print(attrs)