zennit.rules

Rules based on Hooks

Classes

`AlphaBeta`	AlphaBeta LRP rule.
`Epsilon`	Epsilon LRP rule.
`Flat`	This is the Flat LRP rule.
`Gamma`	Gamma LRP rule.
`Norm`	Normalize and weigh relevance by input contribution.
`Pass`	If the rule of a layer shall not be any other, is elementwise and shall not be the gradient, the Pass rule simply passes upper layer relevance through to the lower layer.
`ReLUDeconvNet`
`ReLUGuidedBackprop`
`WSquare`	This is the WSquare LRP rule.
`ZBox`	ZBox LRP rule for input pixel space.
`ZPlus`	ZPlus (or alpha=1, beta=0) LRP rule.

class zennit.rules.AlphaBeta(alpha=2.0, beta=1.0)[source]

Bases: BasicHook

AlphaBeta LRP rule.

Parameters

alpha (float, optional) – Multiplier for the positive output term.
beta (float, optional) – Multiplier for the negative output term.

class zennit.rules.Epsilon(epsilon=1e-06)[source]

Bases: BasicHook

Epsilon LRP rule.

Parameters: epsilon (float, optional) – Stabilization parameter.

class zennit.rules.Flat[source]

Bases: BasicHook

This is the Flat LRP rule. It is essentially the same as the WSquare Rule, but with all parameters set to ones.

class zennit.rules.Gamma(gamma=0.25)[source]

Bases: BasicHook

Gamma LRP rule.

Parameters: gamma (float, optional) – Multiplier for added positive weights.

class zennit.rules.Norm[source]

Bases: BasicHook

Normalize and weigh relevance by input contribution. This is essentially the same as the LRP Epsilon Rule with a fixed epsilon only used as a stabilizer, and without the need of the attached layer to have parameters weight and bias.

class zennit.rules.Pass[source]

Bases: Hook

If the rule of a layer shall not be any other, is elementwise and shall not be the gradient, the Pass rule simply passes upper layer relevance through to the lower layer.

backward(module, grad_input, grad_output)[source]: Pass through the upper gradient, skipping the one for this layer.

class zennit.rules.ReLUDeconvNet[source]

Bases: Hook

backward(module, grad_input, grad_output)[source]: Modify ReLU gradient according to DeconvNet.

class zennit.rules.ReLUGuidedBackprop[source]

Bases: Hook

backward(module, grad_input, grad_output)[source]: Modify ReLU gradient according to GuidedBackprop.

class zennit.rules.WSquare[source]

Bases: BasicHook

This is the WSquare LRP rule.

class zennit.rules.ZBox(low, high)[source]

Bases: BasicHook

ZBox LRP rule for input pixel space.

Parameters

low (obj:torch.Tensor) – Lowest pixel values of input.
high (obj:torch.Tensor) – Highest pixel values of input.

class zennit.rules.ZPlus[source]

Bases: BasicHook

ZPlus (or alpha=1, beta=0) LRP rule.

Notes

Note that the original deep Taylor Decomposition (DTD) specification of the ZPlus Rule (https://doi.org/10.1016/j.patcog.2016.11.008) only considers positive inputs, as they are used in ReLU Networks. This implementation is effectively alpha=1, beta=0, where negative inputs are allowed.