Unified API

The SignXAI unified API provides a framework-agnostic interface for explainable AI methods.

Overview

The unified API automatically detects whether you’re using PyTorch or TensorFlow models and routes your requests to the appropriate backend implementation.

Main Interface

Main Function

explain(model, input_data, method, **kwargs)

Get explanations for a model’s predictions using any supported XAI method.

param model:

The neural network model (PyTorch or TensorFlow)

param input_data:

Input data for which to generate explanations

param method:

The explanation method to use

param kwargs:

Additional method-specific parameters

return:

Explanation array with same shape as input

Framework Detection

get_framework(model)

Automatically detect the framework of a given model.

param model:

The model to check

return:

‘pytorch’ or ‘tensorflow’

Common Parameters

All methods support these common parameters:

neuron_selection: Target neuron/class for explanation
batchsize: Batch size for processing (PyTorch only)
postprocess: Post-processing function to apply

Method-Specific Parameters

Different methods accept additional parameters:

Gradient-based methods:

postprocess: ‘abs’, ‘square’, or custom function
mu: SIGN threshold parameter (for SIGN variants)

LRP methods:

epsilon: Stabilization parameter for LRP-ε
alpha/beta: Parameters for LRP-α/β rule
layer_rule: Custom rules for specific layers

Integrated Gradients:

reference: Reference/baseline input
steps: Number of integration steps

SmoothGrad:

noise: Noise level for sampling
samples: Number of samples

Grad-CAM:

last_conv: Name/index of last convolutional layer

Usage Examples

Basic usage with automatic framework detection:

from signxai import explain, list_methods

# List all available methods
available_methods = list_methods()
print(f"Available methods: {available_methods}")

# TensorFlow model
explanation_tf = explain(
    tf_model,
    input_data,
    method='gradient'
)

# PyTorch model
explanation_pt = explain(
    torch_model,
    input_data,
    method='gradient'
)

Using method-specific parameters:

# LRP with epsilon
explanation = explain(
    model,
    input_data,
    method='lrp_epsilon',
    epsilon=0.01
)

# Input × Gradient method
explanation = explain(
    model,
    input_data,
    method='input_t_gradient'
)

# SIGN variant with custom mu
explanation = explain(
    model,
    input_data,
    method='gradient_x_sign_mu',
    mu=0.5
)

Error Handling

The unified API provides consistent error messages across frameworks:

UnsupportedMethodError: Method not available for the detected framework
InvalidParameterError: Invalid parameter for the chosen method
FrameworkDetectionError: Unable to determine model framework