torchxai-tools 0.1.3

TorchXAI is a PyTorch-based toolkit for evaluating machine learning models using explainability techniques.

TorchXAI

TorchXAI is a lightweight PyTorch toolkit for evaluating machine learning models using explainability techniques. It wraps Captum attribution methods and adds multi-target attribution — explain multiple output classes in a single forward pass — plus ready-to-use metrics for quantifying explanation quality.

• Captum-compatible — works alongside the Captum explainers you already use
• Multi-target — compute attributions for all targets at once, not one at a time
• Batch & scalable — built for dataset-scale evaluation across many inputs and explainers

Installation

pip install torchxai-tools

The PyPI distribution is named torchxai-tools; the import name is torchxai.

from torchxai.explainers import SaliencyExplainer   # import name is torchxai

Quick start

Generating explanations

import torch
import torch.nn as nn
from torchxai.explainers import SaliencyExplainer, IntegratedGradientsExplainer
from torchxai.data_types import SingleTargetAcrossBatch

model = nn.Sequential(nn.Linear(10, 5), nn.ReLU(), nn.Linear(5, 3))
model.eval()
inputs = torch.randn(1, 10)

# Single target
explainer = SaliencyExplainer(model)
attrs = explainer.explain(inputs=inputs, target=SingleTargetAcrossBatch(index=0))
print(attrs.shape)   # (1, 10)

# All three classes in one call
explainer_mt = SaliencyExplainer(model, multi_target=True)
targets = [SingleTargetAcrossBatch(index=i) for i in range(3)]
attrs_list = explainer_mt.explain(inputs=inputs, target=targets)
print(len(attrs_list), attrs_list[0].shape)   # 3, (1, 10)

With a baseline (IntegratedGradients, DeepLift, …)

from torchxai.explainers import IntegratedGradientsExplainer

baseline = torch.zeros_like(inputs)
explainer = IntegratedGradientsExplainer(model)
attrs = explainer.explain(
    inputs=inputs,
    baselines=baseline,
    target=SingleTargetAcrossBatch(index=0),
)

Evaluating explanation quality

from torchxai.metrics.axiomatic import completeness
from captum.attr import Saliency

net = ...   # your model
saliency = Saliency(net)
input = torch.randn(2, 3, 32, 32, requires_grad=True)
baselines = torch.zeros(2, 3, 32, 32)

attribution = saliency.attribute(input, target=3)
score = completeness(net, input, attribution, baselines)
print("Completeness:", score)

Supported explainers

Explainer	Requires baseline	Notes
SaliencyExplainer	✗
InputXGradientExplainer	✗
GuidedBackpropExplainer	✗	Not compatible with transformers
RandomExplainer	✗	Baseline for sanity-checking
IntegratedGradientsExplainer	✓
DeepLiftExplainer	✓	Not compatible with transformers
InputXBaselineGradientExplainer	✓
DeepLiftShapExplainer	✓ distribution	Not compatible with transformers
GradientShapExplainer	✓ distribution
FeatureAblationExplainer	✗	Optional feature_mask
LimeExplainer	✗	Optional feature_mask
KernelShapExplainer	✗	Optional feature_mask
OcclusionExplainer	✗	Requires sliding_window_shapes

Documentation

Full documentation including per-explainer API reference and end-to-end examples (image classification, BERT sequence classification, NER):

saifullah3396.github.io/torchxai

License

MIT — see LICENSE.txt.