Docs: How-To: Write custom rules - Githubissues

chr5tphr / zennit

Zennit is a high-level framework in Python using PyTorch for explaining/exploring neural networks using attribution methods like LRP.

Other

188 stars 32 forks source link

Docs: How-To: Write custom rules #131

Closed chr5tphr closed 2 years ago

chr5tphr commented 2 years ago

Add a How-To on writing custom rules, specifically using

zennit.core.Hook for general Hooks, and
zennit.core.BasicHook for LRP-based Hooks

For zennit.core.Hook:

add explanations for backward and forward and zennit.core.Hook
give a short formal demonstration for the parts of the gradient in the backward function of Hook
introduce copying mechanics and Hooks with arguments

For zennit.core.BasicHook:

give a short description of the different functions used by BasicHook
give a rough formal definition of the computation done by BasicHook using the different functions
give a few examples, with a simple instance, and subclassing of BasicHook
explain the default values and mechanics for the BasicHook argument functions